orchestrator-host: symlink the opencode installer's binary into ~/.local/bin; README: nginx file perms, first-rebuild wrapper, pre-cutover checks

Found on the first activation of #cc-ci on 195.201.88.249: the opencode installer lands in ~/.opencode/bin (opencode-web then failed at EXEC), and nginx's pre-start config check runs as the nginx user, so the staged cert/htpasswd must be root:nginx 0640, not root:root 0600. Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01FqkQq3CDmFWcQ7u1LzoyRz
nix/hosts/cc-ci: real hardware + networking from the box; document the infect fixes
2026-09-07 20:24:47 +00:00 · 2026-09-07 20:15:16 +00:00 · 2026-09-07 19:58:33 +00:00 · 2026-08-31 20:47:10 +00:00 · 2026-08-31 20:46:55 +00:00 · 2026-08-31 20:31:59 +00:00
66 changed files with 3392 additions and 258 deletions
@@ -0,0 +1,97 @@
+---
+name: cc-ci-cleanup
+description: Tidy the fleet's open recipe PRs. Reconciles every mirror from TRUE upstream first (which alone closes PRs upstream already merged), then surveys every open PR deterministically, CLOSES the ones that can no longer be merged or were never meant to be (CI sweep artifacts, obsolete bumps, superseded duplicates) with a reason, and reports prioritised action items for the ones that SHOULD merge — what specifically is blocking each. NEVER merges a recipe PR. Invoke as /cc-ci-cleanup [recipe ...] [--dry-run].
+---
+
+# cc-ci-cleanup
+
+Open recipe PRs accumulate and rot. Some were never meant to merge (CI sweep artifacts), some were
+overtaken (upstream merged the same change, or a newer PR supersedes them), and some genuinely should
+land but are quietly blocked. Left alone the list becomes noise, and a real CVE fix hides in it.
+
+This skill separates those three, acts on the first two, and hands you a short list for the third.
+
+**Boundaries.** It **CLOSES** irrelevant PRs and **NEVER MERGES** any recipe PR — those change what
+deploys on other people's infrastructure, so a human merges them (see AGENTS.md). Closing is the only
+write it performs, always with a comment saying why.
+
+## Arguments
+- `<recipe> …` — limit to these recipes (else every recipe in `cc-ci-plan/used-recipes.md`).
+- `--dry-run` — classify and report, close nothing.
+
+## Procedure
+
+### 1. Reconcile every mirror from TRUE upstream — MANDATORY, FIRST
+```
+cc-ci-plan/reconcile-upstream.sh --all        # or: reconcile-upstream.sh <recipe>...
+```
+**Do not skip this and do not reorder it.** Every signal in step 2 is measured against the mirror's
+`main`; against a stale mirror they are all wrong. This step also does a chunk of the cleanup by
+itself — it closes any PR whose changes upstream has already merged.
+
+> On the first real run (2026-08-11) this alone closed **three** PRs that looked pending and were
+> already merged upstream: discourse #6 (carrying **140 CVEs**), keycloak #6 (**12 CVEs**), n8n #5.
+> All three had been reported to the operator as outstanding work. mailu #6 went the same way earlier
+> the same day. Reconciling is not hygiene, it is how you avoid recommending work that is already done.
+
+### 2. Survey every open PR (deterministic — no judgement yet)
+```
+python3 cc-ci-plan/pr-survey.py [recipe ...]          # add --json for the raw facts
+```
+Per PR it measures: `behind_main`, `ahead`, `mergeable`, `diff_files`, the images it **adds**, which
+of those are **already in main**, `obsolete`, the newest `!testme` verdict + build, `branch_kind`,
+and age/idle days. It decides nothing — that is this skill's job.
+
+### 3. Classify
+
+**CLOSE — cannot merge, or was never meant to.** Each needs a *positive* reason, not an absence:
+
+| signal | why it is closeable |
+|---|---|
+| `branch_kind: ci-artifact` (`ci/*`) | regall/cfold sweeps and `!testme` probes — harness artifacts, never intended to merge |
+| `obsolete: true` | every image it adds is **already pinned in main** — it has nothing left to contribute |
+| superseded | a newer PR on the same recipe makes the same bump (name both numbers in the comment) |
+| `diff_files: 0` | genuinely empty diff — nothing to merge |
+
+**NEVER close on:**
+- `DIFF-UNREADABLE` — the diff could not be fetched, which is NOT an empty diff. gitea #4 reads that
+  way (force-pushed branch) while being a verified, green, needed fix.
+- any field that came back `null`/unknown.
+- a PR that carries a **CVE fix** and is the only thing carrying it, even if it looks stale — report it
+  instead. Losing a security fix to tidiness is far worse than a long PR list.
+- `--dry-run`.
+
+**NEEDS WORK — should merge, something blocks it.** Give the *specific* next action:
+| signal | action item |
+|---|---|
+| `mergeable: false` | conflicts — rebase the branch on `main` and re-run `!testme` |
+| `behind_main > 0` | out of date — rebase, then re-verify (a green from before main moved proves nothing) |
+| `ci: failed` | diagnose via `/ci-test-review`; classify recipe-bug vs stale test |
+| `ci: never-run` | run `!testme` |
+| blocked on the operator | say exactly what is needed (a secret, an upstream release, a decision) |
+
+**READY — green, current, no conflicts.** Action item is simply: review and merge.
+
+### 4. Close the CLOSE set (skip entirely under `--dry-run`)
+Comment first, then close. The comment must say **which signal** made it closeable and **what to do
+if that is wrong** ("reopen if …"), so a wrong call is cheap to undo. Never close silently.
+
+### 5. Report
+Order by what deserves attention, not by recipe name:
+
+1. **CVE-carrying PRs that should merge** — most severe first, with the CVE ids.
+2. Other **READY** PRs (green + current).
+3. **NEEDS WORK**, each with its one specific action.
+4. **CLOSED this run**, with the reason for each.
+5. Anything **deliberately left alone** despite looking stale, and why.
+
+End with a one-line summary: `N open → C closed, R ready to merge, W need work`.
+
+## Guardrails
+- **Never merge a recipe PR.** Create/verify/close only; the operator merges.
+- **Reconcile first, always.** Judging a PR against a stale mirror is how you close good work or
+  recommend work that is already done.
+- **Close only on a positive signal**, never on "looks old". Age alone is not a reason — several
+  60-day-old PRs here are green and mergeable.
+- **Never close a lone CVE fix.** Report it, however stale.
+- Every close gets a comment with its reason and a reopen hint.
@@ -79,12 +79,29 @@ For each real (non-flaky) finding, write the actual fix and open a PR. **Never m
  it handles the mirror to `git.autonomic.zone/recipe-maintainers/<recipe>` (upstream
  `git.coopcloud.tech`). Keep the change **bounded** to the diagnosed root cause; don't rewrite the
  recipe.
+- **Before editing any test, read `tests/STYLE.md` in the cc-ci repo.** It encodes the rules a test
+  change must satisfy — set state up through the app's interface rather than its database, gate on
+  version instead of branching, correct the fixture/wait but NEVER the assertion, and diagnose from
+  the app's own telemetry before concluding a test is stale.
 - **CI-server-side fix → cc-ci PR.** Branch the cc-ci product repo
  (`recipe-maintainers/cc-ci`), apply the fix, and open the PR via the Gitea API (use the
  `GITEA_*` creds from `/srv/cc-ci/.testenv`). **Single-writer discipline:** work on a dedicated
  branch in a SEPARATE clone — **never push `main`, never touch the build loops' working clones**
  (`/cc-ci`, `/cc-ci-adv`) or their in-flight state.

+> ### ⚠️ RECONCILE FROM UPSTREAM FIRST — always, before any PR work or upgrade check
+> ```
+> cc-ci-plan/reconcile-upstream.sh <recipe>...   # or --all
+> ```
+> Deterministic, idempotent, and safe (recipe work lives in branches, never on mirror `main`). It
+> force-syncs each mirror to coopcloud's **default branch — resolved from the API, `main` OR
+> `master`** — and closes any mirror PR whose changes upstream already merged. Skipping it has cost
+> us three distinct ways: mailu #6 was reported as the fix for two internet-facing CVEs while
+> upstream had already merged AND released it; a stale mirror makes a survey report "no upgrades
+> available" so the recipe drops out of the weekly run; and reading the wrong branch on a recipe with
+> a stale `main` beside a live `master` (gitea) manufactures a false "three releases behind, missing
+> two CVSS-9.8 RCEs" finding.
+
 ### 5. VERIFY each PR on the CI server (deterministic; still never merge)
 A PR is only "working" once **cc-ci verifies it green** (operator rule) — dogfood the CI that found
 the bug. Verification is deterministic (the harness), not an AI judgement.
@@ -46,12 +46,19 @@ This is `/recipe-upgrade` step 1's research, stopping before it implements anyth
 >    `git -C ~/.abra/recipes/<r> remote set-url origin "https://$GITEA_USERNAME:$GITEA_PASSWORD@git.autonomic.zone/recipe-maintainers/<r>.git"`
 > 3. **dirty worktree** — usually just the untracked cc-ci overlay; `git stash -u` before, `stash pop`
 >    after. Only a genuinely dirty TRACKED tree is a skip.
-> 4. **tag+digest pins abra cannot parse** — abra FATAs and aborts the WHOLE recipe (immich). This is
->    **not** "not fetchable": enumerate the compose's `image:` refs yourself and check the upstream
->    registry directly for the ones abra could not read, picking the newest tag the app version
->    supports rather than the numerically highest.
+> 4. **tag+digest pins abra cannot parse** — abra FATAs and aborts the WHOLE recipe (immich). Do not
+>    hand-check the registry; run the resolver, which is abra-independent and covers every image:
+>    ```
+>    python3 /srv/cc-ci/cc-ci-plan/resolve-images.py <recipe> --ssh cc-ci --table
+>    ```
+>    It reports, per image, `newest_within_major` (the compatibility-safe pick) and
+>    `newest_same_shape` (the newest of that tag's form). **Use `newest_within_major` unless you have
+>    checked the app supports the major jump** — immich's postgres tag encodes the pg major plus the
+>    vectorchord/pgvectors versions immich-server is built against, so taking the newest would break
+>    the deploy. `all_resolved: false` means an image could NOT be resolved — that is a `?`, never a 0.

-**Reconcile the mirror from true upstream FIRST — ALWAYS, no exceptions.** This is the same reconcile
+**Reconcile the mirror from true upstream FIRST — ALWAYS, no exceptions** — one command,
+`cc-ci-plan/reconcile-upstream.sh <recipe>... | --all`. This is the same reconcile
 `/upgrade-all` does. Do not skip it in the name of keeping the sweep read-only: skipping it makes you
 research a stale checkout, and on the first real run that produced **two recipes with no survey output
 at all**, which is indistinguishable from "no upgrades" unless you check. It is safe — recipe work
@@ -66,6 +73,12 @@ nothing; it also auto-closes mirror PRs whose changes upstream has already merge
 > API (`/api/v1/repos/coop-cloud/<recipe>` → `default_branch`) before reading any file, and never
 > `git reset --hard origin/main` on a checkout that tracks `master`.

+**Cross-check abra with the resolver.** abra is the primary source, but it silently contributes
+nothing for images it cannot parse, and it reported "no new versions" for images that did have them
+(mumble v1.6.870-0 → -4). Run `resolve-images.py` for every recipe and take the UNION of the two: on
+the first real sweep the resolver found upgrades abra missed entirely in five recipes, one of which
+(plausible's clickhouse) carried four CVEs.
+
 Then read versions:
 ```
 set -a; . /srv/cc-ci/.testenv; set +a
@@ -91,6 +104,27 @@ CRITICAL came from, and an image with no window is not counted at all.
  distinction was the difference between two false zeros and the truth (both recipes turned out fine,
  but nothing in the survey said so).

+### 2c. Know which recipes CANNOT see CVEs at all
+```
+python3 cc-ci-plan/audit-sources.py --security-sources
+```
+A recipe whose sources yield **no CVE data at all** cannot produce a meaningful `0` — nothing was
+measured, the same way a missing registry file cannot. Render those as **`?`**, not `0`.
+
+**The fleet is currently at zero such recipes.** The last two — `mattermost-lts` (empty advisory
+feed, client-side-rendered bulletins) and `mumble` (nothing published anywhere) — were fixed by
+declaring an NVD CPE in their registry:
+```
+- nvd-cpe: mattermost-team-edition = cpe:2.3:a:mattermost:mattermost_server:*:*:*:*:*:*:*:*
+```
+**If this sweep ever reports a blind recipe again, that is the fix**: find the product's CPE at
+nvd.nist.gov and add the line. Prefer a real advisory feed or an attributable changelog when one
+exists — NVD lags the vendor — but a lagging source beats no source, and it turns a `?` into a
+number.
+
+An *unparseable page* is NOT the same thing: it is harmless when the same project also publishes an
+advisory feed (redis, gitea, minio, clickhouse all do). Only "no usable source for this image" counts.
+
 ### 3. Run the advisory scan over that window
 ```
 python3 /srv/cc-ci/cc-ci-plan/advisory-scan.py <recipe> --from <old-app> --to <new-app> \
@@ -312,6 +312,12 @@ test change, and a test change is **gated by `--with-tests`**:
  Do **NOT** modify any test. Report `SUCCESS-PENDING-TESTS` (recipe PR open; `!testme` red on a
  stale test; operator to decide).
 - **`--with-tests` — open + verify a cc-ci test PR.** Make it the `ci-test-review` way:
+  0. **READ `tests/STYLE.md` in the cc-ci repo FIRST.** It is the rulebook for changing a test, and
+     it is written against the failures this pipeline has actually produced. The two that matter most
+     here: **set state up through the app's own interface, never its database** (a plausible fixture
+     that INSERTed rows passed on v2 and silently broke on v3, holding the recipe RED for six weeks),
+     and **gate on version rather than writing a fixture that supports both** — old-version tests can
+     simply be deleted, since the older version is only exercised through the upgrade tier.
  1. Branch `recipe-maintainers/cc-ci` in a **separate clone** (single-writer: never push `main`,
     never touch the build loops' `/cc-ci` `/cc-ci-adv` clones); update the test/overlay.
  2. **Verify the recipe upgrade WITH the updated test applied.** `!testme` on the recipe PR uses the
@@ -18,7 +18,7 @@ RECIPE="${1:?usage: testme-on-pr.sh <recipe> <pr-index>}"
 PRIDX="${2:?usage: testme-on-pr.sh <recipe> <pr-index>}"
 TESTENV="${TESTENV:-/srv/cc-ci/.testenv}"
 set -a; . "$TESTENV"; set +a
-: "${GITEA_USERNAME:?}"; : "${GITEA_PASSWORD:?}"; : "${GITEA_URL:?}"
+: "${GITEA_USERNAME:?}"; : "${GITEA_PASSWORD:?}"; : "${GITEA_URL:=git.autonomic.zone}"
 NS="${GITEA_NAMESPACE:-recipe-maintainers}"
 API="https://${GITEA_URL}/api/v1"; AUTH=(-u "${GITEA_USERNAME}:${GITEA_PASSWORD}")
 POST="${POST:-1}"; MAX_WAIT="${MAX_WAIT:-480}"; INTERVAL="${INTERVAL:-30}"
@@ -73,6 +73,29 @@ done
 # 5) Stray exited containers (debug one-shots) — best-effort prune.
 docker container prune -f >/dev/null 2>&1 || true

+# 6) Unused IMAGES — the one that actually took CI down. Every run pulls each recipe's images and
+#    nothing ever removed the old ones: on 2026-08-11 they had grown to 72GB (63GB of it unused),
+#    the root filesystem hit 100% under two concurrent runs, and the harness died at startup with
+#    `OSError: [Errno 28] No space left on device: '/var/lib/cc-ci-runs/<build>'`. Every !testme
+#    from build 1236 to 1242 failed that way — with no results.json, so the PR badges just read
+#    "failure" and looked like recipe regressions.
+#
+#    Only prune above a threshold, so a healthy host keeps its layer cache and runs stay fast.
+#    `image prune -a` removes only images no container references, so anything deployed (infra +
+#    warm-* canonicals) is untouched; anything else is re-pulled on demand.
+#
+#    Volumes are deliberately NOT pruned here — see the KEEP_RE guard in (3): warm-* canonicals are
+#    data-warm and their volumes are legitimately dangling between runs.
+DISK_PRUNE_PCT="${DISK_PRUNE_PCT:-60}"
+used_pct="$(df --output=pcent / 2>/dev/null | tail -1 | tr -dc '0-9')"
+if [ -n "$used_pct" ] && [ "$used_pct" -ge "$DISK_PRUNE_PCT" ]; then
+  echo "  disk ${used_pct}% >= ${DISK_PRUNE_PCT}% -> pruning unused images"
+  freed="$(docker image prune -af 2>/dev/null | awk '/Total reclaimed space/ {print $4, $5}')"
+  echo "  reclaimed: ${freed:-0B}; disk now $(df -h / | tail -1 | awk '{print $5" used, "$4" free"}')"
+else
+  echo "  disk ${used_pct:-?}% < ${DISK_PRUNE_PCT}% -> keeping image cache"
+fi
+
 if [ "$removed" -eq 0 ]; then
  echo "== orphan sweep: clean (nothing to remove) =="
 else
@@ -13,8 +13,8 @@ a health gate, not as silent drift.

 > **Two hosts, two flakes — don't confuse them.** This skill updates the **orchestrator** host:
 > the machine this session runs on (`cc-ci-orchestrator-1`, Hetzner cpx22 **server 134487234**,
-> tailnet `100.84.190.30`, public `168.119.126.100`), flake checkout **`/srv/cc-ci-orch`** (repo
-> `recipe-maintainers/cc-ci-orchestrator`), target **`.#cc-ci-orchestrator-hetzner`**. The **cc-ci
+> tailnet `cc-ci`, public `195.201.88.249` — the SAME host as the cc-ci CI server since 2026-09), flake checkout **`/srv/cc-ci-orch`** (repo
+> `recipe-maintainers/cc-ci-orchestrator`), target **`.#cc-ci`** (which now also rebuilds the CI server half, from the cc-ci repo flake input). The **cc-ci
 > CI server** (`ssh cc-ci`, repo `recipe-maintainers/cc-ci`, target `.#cc-ci`) is a different
 > machine — that's `/cc-ci-server-update`, NOT this skill.

@@ -72,7 +72,7 @@ deploy).
 ### 3. Build (catch errors before any activation)

 ```
-cd /srv/cc-ci-orch && nixos-rebuild build --flake .#cc-ci-orchestrator-hetzner 2>&1 | tail -15
+cd /srv/cc-ci-orch && nixos-rebuild build --flake .#cc-ci 2>&1 | tail -15
 readlink -f result
 ```
 Build failure → fix on the branch (option renames etc.) before going further. Never activate a
@@ -85,7 +85,7 @@ activation breaks the host (cf. the cc-ci server's 2026-08-03 no-default-route o
 reboot — Hetzner API power-cycle on server **134487234** if SSH is gone (see
 `hetzner-server-recovery`) — lands back on the last-known-good generation.
 ```
-cd /srv/cc-ci-orch && setsid nohup nixos-rebuild test --flake .#cc-ci-orchestrator-hetzner \
+cd /srv/cc-ci-orch && setsid nohup nixos-rebuild test --flake .#cc-ci \
  > /tmp/orchestrator-test-switch.log 2>&1 < /dev/null & echo launched
 # after it settles (poll; tailscaled/sshd may blip):
 readlink /run/current-system            # should be the new store path
@@ -100,7 +100,7 @@ switch.
 ### 5. Switch (make permanent — only after 4 is healthy)

 ```
-cd /srv/cc-ci-orch && nixos-rebuild switch --flake .#cc-ci-orchestrator-hetzner 2>&1 | tail -10
+cd /srv/cc-ci-orch && nixos-rebuild switch --flake .#cc-ci 2>&1 | tail -10
 ```
 (If it fails with "Unit nixos-rebuild-switch-to-configuration.service was already loaded", the
 detached test's transient unit is still running — wait or `systemctl stop` it, then retry.)
@@ -127,7 +127,7 @@ git add flake.lock                                # flake.nix too if the channel
 git commit -m "flake: bump nixpkgs (nixos-26.05, $(date -u +%Y-%m-%d))

 nixpkgs: <old-rev[:8]> -> <new-rev[:8]> (nixos-26.05 tip)
-Deployed to cc-ci-orchestrator-hetzner: build + test + switch + health gate green."
+Deployed to the cc-ci host (.#cc-ci): build + test + switch + health gate green."
 git push -u origin HEAD
 ```
 Open the PR on `recipe-maintainers/cc-ci-orchestrator` (Gitea API with the `GITEA_*` creds from
@@ -78,8 +78,38 @@ ssh cc-ci 'systemctl --failed --no-legend; df -h / | tail -1; docker service ls
 systemctl --failed --no-legend; df -h / | tail -1; tmux ls
 ```
 - Failed units, core swarm services not 1/1 (warm-* spares flapping is a known benign pattern —
-  note, don't page), disk >80% (server) / >85% (orchestrator) → findings. Server unreachable →
+  note, don't page), disk **>65% (server)** / >85% (orchestrator) → findings. Server unreachable →
  HIGH: recommend `hetzner-server-recovery`.
+  > **65%, not 80%, on the server — it is not a steady-state measure.** Two concurrent recipe runs
+  > pull images and write volumes worth tens of GB, so a host sitting at 73% still hits 100% mid-run.
+  > That is exactly what happened on 2026-08-11: 63GB of unused images had accumulated (nothing ever
+  > pruned them), the filesystem filled during a run, and the harness died at startup with
+  > `OSError: [Errno 28] No space left on device`. Remedy: `docker image prune -af` on cc-ci — it
+  > spares anything a container references, so infra and warm-* canonicals are untouched. Do NOT
+  > `docker volume prune`: warm-* canonical volumes are data-warm and legitimately dangling.
+- **!testme actually produces results** (the check that would have caught the above days earlier):
+  the newest few `/var/lib/cc-ci-runs/<build>/` dirs must each contain `results.json`. A build that
+  dies before the harness writes one leaves an EMPTY dir — and the PR badge still says "failure", so
+  it reads as a recipe regression rather than a sick host. Builds 1236–1242 all failed that way.
+  Finding: *"N recent builds produced no results.json — the harness is dying at startup, check disk
+  and the drone step log"*. The step log lives in drone's sqlite
+  (`/var/lib/docker/volumes/drone_ci_commoninternet_net_data/_data/database.sqlite`) — copy it and
+  read `logs.log_data` for the failing `steps.step_id`; the bridge's drone token is not extractable
+  (distroless container, swarm secret).
+  > **If the error is ENOSPC but the disk is fine**, it is not disk. Seen 2026-08-11: builds 1244-1249
+  > died on `mkdir /var/lib/cc-ci-runs/<build>` with **110GB free and 16% inodes**, while the identical
+  > mkdir succeeded as root over ssh, inside the runner's own mount namespace, and 61/61 times in a
+  > stress loop — and the same harness run by hand with a numeric run id worked fine. Restarting
+  > `drone-runner-exec` did NOT help, and neither did recreating the runs directory with a fresh
+  > inode (it recurred afterwards — that apparent fix was coincidence).
+  >
+  > **It is INTERMITTENT and tracks concurrent activity**, which is the useful signal: every failure
+  > landed while a second run or a manual deploy was in flight (1252 was triggered while 1251 was
+  > still finishing), and every build on a quiet host succeeded (1243, 1250, 1251, 1253). Free space
+  > never moved during a failing build. So on ENOSPC-with-free-disk: **wait for the host to go quiet
+  > and re-trigger** before treating it as a recipe failure. Root cause is still NOT established;
+  > `DRONE_RUNNER_CAPACITY=2` allows the overlap, so lowering it to 1 is the obvious next experiment
+  > if it becomes disruptive.
 - **Bridge / !testme path**: `docker service ls` shows `ccci-bridge_app 1/1` AND the bridge log
  has no auth errors (`docker service logs --since 24h ccci-bridge_app 2>&1 | grep -ci "401\|user does not exist"` == 0).
  A silently-401ing bridge drops `!testme` (seen 2026-08-03, stale rotated Gitea secret) →
@@ -113,7 +143,8 @@ minutes, no PRs). If it is instead that a known CVE is sitting unpatched, recomm
 `/cve-check` over waiting for the next weekly run whenever the question is "are we exposed?".

 `ALL HEALTHY` requires: recent successful weekly run + published report, no stale tests, no
-CVE PR open >14 days, both hosts <30 days behind their channel, zero failed units, disk under
+CVE PR open >14 days, both hosts <30 days behind their channel, zero failed units, recent builds all
+producing results.json, disk under
 thresholds, bridge clean, maintained-set consistent. Anything else is a finding — even minor
 ones get a recommended next step. Order findings by priority (CVE/unreachable-host first).

@@ -83,8 +83,29 @@ failure (AI — this is the `ci-test-review` step-3 diagnosis):
  changed upstream, what the test currently asserts.
 - **FLAKY** → re-run once or twice; if it passes, drop it (not stale, just flaky).

+> ### ⚠️ RECONCILE FROM UPSTREAM FIRST — always, before any PR work or upgrade check
+> ```
+> cc-ci-plan/reconcile-upstream.sh <recipe>...   # or --all
+> ```
+> Deterministic, idempotent, and safe (recipe work lives in branches, never on mirror `main`). It
+> force-syncs each mirror to coopcloud's **default branch — resolved from the API, `main` OR
+> `master`** — and closes any mirror PR whose changes upstream already merged. Skipping it has cost
+> us three distinct ways: mailu #6 was reported as the fix for two internet-facing CVEs while
+> upstream had already merged AND released it; a stale mirror makes a survey report "no upgrades
+> available" so the recipe drops out of the weekly run; and reading the wrong branch on a recipe with
+> a stale `main` beside a live `master` (gitea) manufactures a false "three releases behind, missing
+> two CVSS-9.8 RCEs" finding.
+
 ### 2. For each stale test — author the minimal test update (AI; never weaken)

+> **Read `tests/STYLE.md` in the cc-ci repo before writing the update.** It is the rulebook for test
+> changes, written from failures this pipeline actually produced. Most load-bearing: set state up
+> through the app's **own interface, never its database** (a plausible fixture that INSERTed rows
+> passed on v2 and silently broke on v3 — 202 acks, rows in postgres, nothing ingested — and held the
+> recipe RED for six weeks), **gate on version rather than supporting both** (old-version tests can be
+> deleted; the older version is only exercised via the upgrade tier), and correct the fixture or the
+> wait but **never the assertion**.
+
 Work on **one recipe at a time** (serialize — each verification deploys a recipe on the shared
 Swarm). For each `STALE_TESTS` entry:

@@ -31,6 +31,12 @@ Then present the roster grouped as follows, and close with the situation guide.
  PR). `--with-tests` also fixes that recipe's stale test.
 - **/recipe-report** — (re)generate the weekly report page for report.ci.commoninternet.net.

+**Keeping the PR list honest**
+- **/cc-ci-cleanup** — reconciles every mirror from true upstream (which alone closes PRs upstream
+  already merged), then closes the open recipe PRs that can no longer merge or were never meant to
+  (CI sweep artifacts, obsolete bumps, superseded duplicates) and reports what is actually blocking
+  the ones that should land. Never merges.
+
 **Security (CVEs)**
 - **/cve-check** — fleet-wide CVE sweep with **no upgrading**: for every recipe, work out what
  upgrade is available (per image, sidecars included), scan it for CVEs, and publish a CVE report.
@@ -81,6 +87,7 @@ ARM skills never touch cc-ci infra. After a submodule bump run `scripts/gen-ccte
 | "Run the weekly upgrades now" | `/upgrade-all` (or `systemctl start cc-ci-upgrade-all.service`) |
 | "Upgrade just <recipe>" | `/recipe-upgrade <recipe>` |
 | "The report site is stale/missing a week" | `/recipe-report` |
+| "The open PR list is a mess / what should I merge?" | `/cc-ci-cleanup` |
 | "What CVEs are we exposed to right now?" | `/cve-check` (read-only, no PRs) |
 | "A CVE just dropped — check and patch it" | `/cve-check-and-upgrade` (add `--min-severity high` to skip the noise) |
 | "Is <recipe> vulnerable?" | `/cve-check <recipe>` |
@@ -30,15 +30,18 @@ the orchestrator watches from outside.

 Reboot resilience is handled by **`cc-ci-loops.service`** (system unit): on boot it logs the reboot
 to `REBOOTS.md` (boot_id-gated) and runs `launch.sh start` with `RESUME_PHASE=1`, so the loops +
-watchdog auto-resume the saved phase. The orchestrator session itself is NOT auto-started — the
-operator reconnects to it (that's why the startup notification matters). The orchestrator now runs on
-a **Hetzner `cpx22`** cloud server (`cc-ci-orchestrator-1`, tailnet `100.84.190.30`, public
-`168.119.126.100`, flake host `cc-ci-orchestrator-hetzner`) — see
-`cc-ci-plan/plan-orchestrator-hetzner-migration.md`. The earlier Pi→Incus-VM move is the historical
-`cc-ci-plan/plan-orchestrator-migration.md`. Rebuild this host with
-`nixos-rebuild switch --flake .#cc-ci-orchestrator-hetzner` from `/srv/cc-ci-orch` — but **always
+watchdog auto-resume the saved phase. The orchestrator session itself is relaunched by
+`cc-ci-orchestrator.service` (`agents.py up orchestrator`) — the operator reconnects to it (that's
+why the startup notification matters). Since 2026-09 the orchestrator runs on the **same Hetzner
+host as the cc-ci CI server** (`cc-ci`, public `195.201.88.249`, tailnet `cc-ci`), declared by
+`nixosConfigurations.cc-ci` in this repo's `flake.nix`, which imports the CI server from the cc-ci
+repo's `nixosModules.cc-ci-server`. `ssh cc-ci` from the loops user therefore goes to loopback.
+The full provisioning + deploy guide is `README.md`; the move is recorded in
+`cc-ci-plan/plan-cc-ci-combined-host.md`; the previous hosts (Pi → Incus VM → Hetzner `cpx22`
+shared with notplants) are in `archive/`. Rebuild this host with
+`nixos-rebuild switch --flake .#cc-ci` from `/srv/cc-ci-orch` — but **always
 `nixos-rebuild test` the same flake target first and verify the host is still healthy/reachable
-before the `switch`** (general policy for nix deploys to this host and the cc-ci server: `test`
+before the `switch`** (general policy for nix deploys to this host: `test`
 leaves the bootloader and system profile untouched, so a reboot always recovers to the
 last-known-good generation; the 2026-08-03 cc-ci 26.05 bump outage is the cautionary tale, see
 `.cc-ci-logs/server-update-2026-08-03.md`).
@@ -115,3 +118,29 @@ When the orchestrator, Builder, or assistant makes intentional repository change
 promptly and push them to `git.autonomic.zone` in append-only fashion (never force-push). Match the
 existing commit author and message style in this repo. Do not bundle unrelated worktree changes you
 did not make; stage only the intended files.
+
+## Ship as PRs, merge them yourself, operator reviews retrospectively
+
+**This applies to the two INFRASTRUCTURE repos — `recipe-maintainers/cc-ci-orchestrator` (here) and
+`recipe-maintainers/cc-ci` (the CI product).** For work in either:
+
+1. Branch, don't commit straight to `main`.
+2. Open a PR with a description written to be read **after** the fact: what changed, why, and what
+   evidence says it works (test output, a verified run, a before/after number). The PR *is* the
+   review artifact and the historical record.
+3. **Merge it yourself once it is verified** — do not wait for review. The invocation is the
+   authorization; blocking on review would stall the pipeline these repos exist to run.
+4. The operator reviews **retrospectively**, from the PR.
+
+So the PR is not a gate — it is how the work stays legible. A PR that merely says "fix scanner" has
+failed at its only job.
+
+> ### This does NOT extend to RECIPE repos
+> Recipe PRs — any `coop-cloud/<recipe>` or its `recipe-maintainers/<recipe>` mirror — are
+> **created and verified but NEVER merged by an agent**. Those change what deploys on other people's
+> infrastructure, so a human merges them. The split is deliberate: agents own the tooling, the
+> operator owns the recipes.
+
+If work has already landed on `main` without a PR, do not rewrite published history to fix it.
+Create a branch pinned at the pre-work commit and open the PR against that, so the diff is still
+reviewable and merging only advances the pointer (see PRs #2-#5, 2026-08-11).
@@ -1,47 +1,344 @@
 # cc-ci-orchestrator

-Orchestrator workspace for building the **cc-ci** Co-op Cloud recipe CI server. The plan, launch
-tooling, and loop prompts live in [`cc-ci-plan/`](cc-ci-plan/); see [`AGENTS.md`](AGENTS.md) for the
-roles and operating model. Secrets (`.testenv`) are gitignored — never commit them.
+The **cc-ci orchestrator**: the agent loops that built the cc-ci Co-op Cloud recipe CI server,
+the operator's steering session, and the weekly autonomous recipe-upgrade run — plus the NixOS
+host they run on. Since 2026-09 that host is **the same Hetzner server as the CI server itself**:
+one `nixos-rebuild` from this repo builds both, because this flake imports the CI server as a
+module from the [cc-ci](https://git.autonomic.zone/recipe-maintainers/cc-ci) repo.

-## Run the orchestrator in tmux (survives disconnects + closing your laptop)
+| | where |
+|---|---|
+| Orchestrator loops, timers (weekly upgrader, hourly supervisor) | `nix/modules/cc-ci.nix` → `nixosModules.cc-ci-orchestrator` |
+| The host contract those need (loops user, claude/opencode CLIs, opencode web UI) | `nix/modules/orchestrator-host.nix` → `nixosModules.orchestrator-host` |
+| The CI server (swarm, traefik, drone, runner, `!testme` bridge, dashboard, reports, acme-dns) | cc-ci repo `nix/modules/` → `nixosModules.cc-ci-server` (flake input `cc-ci`) |
+| The machine: hardware, networking, tailscale, root keys | `nix/hosts/cc-ci/` → `nixosConfigurations.cc-ci` |
+| Plans, launch tooling, loop prompts, journal | `cc-ci-plan/` (see `AGENTS.md` for roles) |
+| Skills the orchestrator runs (`/upgrade-all`, `/recipe-upgrade`, `/cc-ci-status`, …) | `.claude/skills/`, `.opencode/skills/` |
+| How it used to be built (Pi → Incus VM → shared Hetzner box) | `archive/` |

-Keep this supervising session alive on the host with tmux, and use `--remote-control` so you can
-watch/steer it from **claude.ai/code** (or the mobile app).
+Secrets (`.testenv`, `upgrader.env`, `.sops/`, everything under `/secrets`) are gitignored — never
+commit them.
+
+---
+
+# Deploying a cc-ci host from scratch
+
+This is the whole path from "nothing" to a working CI server + orchestrator on one Hetzner
+server. It was last done on 2026-09-07 for `195.201.88.249` and is written so a person or an LLM
+can repeat it. Read it once before starting; the order matters.
+
+## 0. What you need in hand
+
+- A **Hetzner Cloud** project you can create servers in (console login or an API token).
+- **SSH keys**: yours, and the orchestrator's own key so the automation can reach the box. The
+  public keys that get root are tracked in `nix/hosts/cc-ci/ssh-keys` (one per line).
+- Read access to `recipe-maintainers/cc-ci`, `recipe-maintainers/cc-ci-orchestrator` (both public
+  read) and the **private** `recipe-maintainers/cc-ci-secrets` (the `autonomic-bot` deploy key,
+  `autonomic-bot-gitea-ed25519`, has it).
+- The out-of-band secrets listed in §4. If you are migrating, they come from the old host; if
+  you are starting fresh you create them (each row says how).
+- Control of the DNS zone (Gandi for `commoninternet.net`) for the cutover in §7.
+
+## 1. Provision the server on Hetzner (Debian image)
+
+In the Hetzner Cloud console (or with `hcloud server create`):
+
+| setting | value | why |
+|---|---|---|
+| Image | **Debian 13** (any recent Debian/Ubuntu works with nixos-infect) | it is replaced by NixOS in §2 |
+| Type | **x86**, **8 GB RAM**, 4 vCPU — e.g. `cpx32` (dedicated AMD) or `cx33`. **Never `cax*`** (ARM): the flakes are `x86_64-linux`. | swarm + recipe deploys + 3–6 agent sessions; 4 GB is too small |
+| Disk | the type's default 150+ GB NVMe | docker layers alone are ~60 GB after a few weeks |
+| Network | public **IPv4** required; IPv6 optional (leave enabled or not, NixOS config ignores it) | cc-ci serves 80/443 and DNS on 53 publicly |
+| SSH keys | add every key from `nix/hosts/cc-ci/ssh-keys` you want to log in with, at least the orchestrator's | nixos-infect carries `/root/.ssh/authorized_keys` over |
+| Name | `cc-ci` | becomes the hostname |
+| Firewall | if a Hetzner Cloud Firewall is attached it must allow **22/tcp, 80/tcp, 443/tcp, 53/tcp, 53/udp** in, and ICMP | the NixOS firewall is separate and is configured by the flake |
+
+Check you can log in: `ssh root@<ip> hostname`.
+
+## 2. Convert Debian → NixOS with nixos-infect
+
+[nixos-infect](https://github.com/elitak/nixos-infect) installs NixOS over the running Debian and
+reboots. Run it detached so the SSH session dropping does not kill it:

 ```bash
-# 0. Exit any running orchestrator session first — a conversation can't be resumed while it's live:
-#      /exit        (inside Claude)   or Ctrl-D
-
-# 1. Start a detachable tmux session on this host
-tmux new -s orchestrator
-
-# 2. Inside tmux, resume the orchestrator conversation WITH remote control:
-claude --resume autonomous-orchestrator \
-  --remote-control "autonomous-orchestrator" \
-  --dangerously-skip-permissions
-#    - If name-resume opens a picker instead of resuming directly, choose "autonomous-orchestrator".
-#    - Or resume by the stable session id (more deterministic in a fresh pane):
-#        claude --resume 34a80a99-b37e-4809-b8da-ccc9fafe785e \
-#          --remote-control "autonomous-orchestrator" --dangerously-skip-permissions
-
-# 3. Detach — the process keeps running:  press Ctrl-b, then d
+ssh root@<ip> 'cat > /root/infect.sh <<"EOF"
+#!/usr/bin/env bash
+set -euo pipefail
+# Pinned nixos-infect revision (same one that built the previous cc-ci hosts).
+INFECT_SHA="40f62a680bb0e8f2f607d79abfaaecd99d59401c"
+export NIX_CHANNEL="nixos-26.05"   # must match the nixpkgs channel in flake.nix
+export PROVIDER="hetznercloud"     # GRUB + Hetzner networking
+export NIXOS_IMPORT=""             # the real config comes from the flake in §5
+# The Debian 13 cloud image mounts /tmp as tmpfs; nixos-infect makes a temporary swapfile
+# there and swapon fails with "Invalid argument". 8 GB RAM needs no extra swap: skip it.
+export NO_SWAP=true
+curl -fsSL "https://raw.githubusercontent.com/elitak/nixos-infect/${INFECT_SHA}/nixos-infect" | bash -x
+EOF
+chmod +x /root/infect.sh
+nohup /root/infect.sh > /var/log/nixos-infect.log 2>&1 &'
 ```

-**Reconnect later**
- On this host: `tmux attach -t orchestrator`
- From anywhere: **claude.ai/code** → the `autonomous-orchestrator` session
+It downloads Nix, builds a NixOS system (5–10 min; follow with
+`ssh root@<ip> tail -f /var/log/nixos-infect.log`), then reboots. The SSH host key changes:
+`ssh-keygen -R <ip>` and confirm `ssh root@<ip> nixos-version` prints a 26.05 version.

-**Why it survives:** tmux keeps the `claude` process alive across SSH disconnects and your laptop
-closing; remote-control runs *outbound* from this host to Anthropic, so it stays connected
-regardless of the viewer. After a host reboot, re-run steps 1–2.
+### 2a. What went wrong on 2026-09-07, and the fixes (Debian 13 image, NixOS 26.05)

-> Two different "names": `--resume <name|id>` selects the *conversation* to restore (shown in the
-> `/resume` picker); the `--remote-control "<name>"` value is only the web display label and resumes
-> nothing. Resuming reuses the same session id each time (stays `34a8…`) — don't pass
-> `--fork-session` unless you intend to branch a new conversation.
->
-> Already inside a live session and just want the web surface? Run `/remote-control` — no exit/resume.
+All three bit on the first attempt; the script above and §3 already include the fixes, this is
+so you recognise them if they come back in another form.
+
+1. **`swapon: /tmp/nixos-infect.XXXX.swp: Invalid argument`** right at the start, script exits.
+   The Debian 13 cloud image mounts `/tmp` as tmpfs and a swapfile cannot live there.
+   Fix: `NO_SWAP=true` (in the script above). An 8 GB box does not need the temporary swap.
+2. **The box never comes back after the reboot: it boots NixOS, but nearly every unit fails**
+   (`dbus`, `systemd-logind`, `sshd`, networking …) with
+   `Could not start dynamically linked executable: /usr/bin/dbus-daemon` in the journal.
+   nixos-infect leaves the old Debian root in place and relies on NixOS's first boot to move it
+   to `/old-root` (`/etc/NIXOS_LUSTRATE`). With NixOS 26.05's systemd-based initrd that
+   lustration did not happen, so Debian's `/etc/systemd/system/*.service` files shadowed the
+   NixOS units and started Debian binaries. Fix, from Hetzner **rescue mode**
+   (`enable_rescue` + `reset` in the API/console, ssh in, `mount /dev/sda1 /mnt/root`):
+   move everything except `nix`, `boot`, `swapfile`, `lost+found`, `var/log`, `var/empty`,
+   `etc/nixos`, `etc/resolv.conf`, `etc/NIXOS`, `etc/machine-id`, `etc/ssh/ssh_host_*`,
+   `root/.nix-*`, `root/.ssh` into `/mnt/root/old-root`, delete `etc/NIXOS_LUSTRATE`, unmount,
+   `disable_rescue`, `reset`. (`/old-root`, ~1 GB, can be deleted once the host is in service.)
+3. **Boots, units fine, but no network.** The generated `networking.nix` has
+   `defaultGateway = "172.31.1.1";` — a bare string. Since NixOS 25.05 that yields no default
+   route. Fix: `defaultGateway = { address = "172.31.1.1"; interface = "eth0"; };` (this is what
+   `nix/hosts/cc-ci/networking.nix` carries). To apply it from rescue mode, chroot into the
+   mounted root and rebuild the boot entry — the nix sandbox cannot `pivot_root` inside a chroot,
+   so turn it off for that one build:
+   ```bash
+   for d in proc sys dev dev/pts; do mount --bind /$d /mnt/root/$d; done
+   mount -t tmpfs tmpfs /mnt/root/run; cp -L /etc/resolv.conf /mnt/root/etc/resolv.conf
+   chroot /mnt/root /nix/var/nix/profiles/system/sw/bin/bash -c '
+     export PATH=/nix/var/nix/profiles/system/sw/bin NIX_REMOTE= HOME=/root
+     export NIX_PATH=nixos-config=/etc/nixos/configuration.nix:nixpkgs=/root/.nix-defexpr/channels/nixos
+     ln -sfn /nix/var/nix/profiles/system /run/current-system
+     nixos-rebuild boot --option sandbox false'
+   ```
+   The `journalctl -D /mnt/root/var/log/journal -b 0` trick (reading the dead system's journal
+   from rescue mode) is what told these apart.
+
+> Rescue mode without a console: `POST /servers/<id>/actions/enable_rescue` with your ssh key
+> id, then `…/actions/reset`; afterwards `disable_rescue` **and check `rescue_enabled` is false
+> before** the next `reset`, or it boots the rescue image again. `scripts/recovery/hetzner.py`
+> wraps these (token in `/srv/cc-ci/.hcloud-token`).
+
+## 3. Capture the machine-specific config into this repo
+
+nixos-infect wrote `/etc/nixos/{hardware-configuration,networking,configuration}.nix`. Only the
+first two matter; the flake replaces `configuration.nix`.
+
+```bash
+scp root@<ip>:/etc/nixos/hardware-configuration.nix nix/hosts/cc-ci/hardware.nix
+scp root@<ip>:/etc/nixos/networking.nix            nix/hosts/cc-ci/networking.nix
+```
+
+Then in `nix/hosts/cc-ci/`:
+
+- `hardware.nix`: keep as generated (GRUB EFI with `efiInstallAsRemovable`, `/boot/efi` by UUID,
+  `/dev/sda1` root). Do not copy another host's file — the UUIDs are per machine.
+- `networking.nix`: keep the static IPv4 + Hetzner gateway `172.31.1.1`. Make sure
+  `networking.defaultGateway` has **both** `address` and `interface = "eth0"` (§2a item 3). If
+  the generated IPv6 block has an empty address, delete the IPv6 parts; a real global address
+  (as on the 2026-09 box) can stay.
+- `configuration.nix`: set `cc-ci.publicIPv4` to the server's IPv4 and check `system.stateVersion`
+  is the release you installed (never change it later).
+- `ssh-keys`: the root keys.
+
+Commit on a branch; the rebuild in §5 can use the local checkout before the PR merges.
+
+## 4. Stage the workspace and secrets on the new host
+
+Everything in this section is **outside git**. Do it as root over SSH, in this order.
+
+### 4a. Tailscale
+
+```bash
+# a reusable (or fresh) tailnet auth key from the tailscale admin console
+install -m600 /dev/stdin /etc/ts-auth-key <<<'tskey-auth-…'
+```
+
+### 4b. The CI server's checkout and its one out-of-band secret
+
+```bash
+# root's deploy key for the private cc-ci-secrets submodule
+install -d -m700 /root/.ssh
+install -m600 <autonomic-bot-gitea-ed25519> /root/.ssh/autonomic-bot-gitea-ed25519
+cat > /root/.ssh/config <<'EOF'
+Host git.autonomic.zone
+    Port 2222
+    User git
+    IdentityFile /root/.ssh/autonomic-bot-gitea-ed25519
+    IdentitiesOnly yes
+EOF
+# the deployed checkout: nightly-sweep runs from it, sops reads secrets/secrets.yaml from it
+git clone --recursive https://git.autonomic.zone/recipe-maintainers/cc-ci.git /etc/cc-ci
+# the master (recovery) age key — the only sops recipient a fresh host can be
+install -d -m700 /var/lib/sops-nix
+install -m600 <master-age.txt> /var/lib/sops-nix/key.txt
+```
+
+`/etc/cc-ci/secrets/secrets.yaml` is encrypted to the master key and the *old* host's SSH host
+key. That is enough to deploy. Afterwards (optional, tidier) add the new host as a recipient:
+`ssh-to-age < /etc/ssh/ssh_host_ed25519_key.pub`, add it to `secrets/.sops.yaml` in cc-ci-secrets,
+`sops updatekeys secrets.yaml`, push, `git -C /etc/cc-ci submodule update --remote`.
+
+### 4c. The orchestrator's workspace (as the `loops` user — it exists after the first rebuild, so
+run §5 once first if this is a fresh host, then come back)
+
+```bash
+sudo -iu loops
+git clone --recursive https://git.autonomic.zone/recipe-maintainers/cc-ci-orchestrator.git /srv/cc-ci-orch
+sudo ln -sfn /srv/cc-ci-orch /srv/cc-ci          # every script and unit says /srv/cc-ci
+cd /srv/cc-ci-orch
+git clone https://git.autonomic.zone/recipe-maintainers/cc-ci.git cc-ci        # Builder clone
+git clone https://git.autonomic.zone/recipe-maintainers/cc-ci.git cc-ci-adv    # Adversary clone
+mkdir -p .cc-ci-logs .sops
+```
+
+Then the files below (`install -m600 -o loops -g users`):
+
+| file | what | source |
+|---|---|---|
+| `/srv/cc-ci/.testenv` | `TS_AUTH_KEY`, `GITEA_PASSWORD` (autonomic-bot), `DOCKERHUB_USERNAME/TOKEN`, model API keys | old host `/secrets/files/cc-ci.testenv`; fresh: create each credential |
+| `/srv/cc-ci/upgrader.env` | `LOOP_TIER`, `LOOP_MODEL`, `REPORT_MODEL` for the weekly run (no secrets) | old host, or copy the example in `AGENTS.md` |
+| `/srv/cc-ci/.sops/master-age.txt` | the same master age key as 4b (skills that re-key secrets use it) | old host |
+| `~loops/.ssh/cc-ci-root-ed25519` (+`.pub`) | `ssh cc-ci` as root — to loopback on this host | old host; fresh: `ssh-keygen -t ed25519` and add the pub to `nix/hosts/cc-ci/ssh-keys` |
+| `~loops/.ssh/autonomic-bot-gitea-ed25519` (+`.pub`) | pushes recipe branches / PRs as `autonomic-bot` | old host; fresh: new key added to the bot's Gitea account |
+| `~loops/.ssh/tangled-ed25519` | optional, tangled.org mirrors | old host |
+| `~loops/.claude/` | Claude Code auth + settings + the orchestrator session history | old host (`rsync -a`); fresh: `claude auth login` as loops (device code, interactive) |
+| `~loops/.local/share/opencode/auth.json`, `~loops/.config/opencode/` | opencode provider auth (the weekly upgrader runs on opencode) | old host; fresh: `opencode auth login` |
+| `/etc/nginx/oc-selfsigned.{crt,key}`, `/etc/nginx/oc-htpasswd` | the tailnet-only opencode UI; **nginx refuses to start without them**, and its config check runs as the `nginx` user, so: `root:nginx`, crt `0644`, key + htpasswd `0640` (the `nginx` group exists after the first rebuild — fix ownership then and `systemctl restart nginx`) | old host, or generate (commands in `nix/modules/orchestrator-host.nix`) |
+
+`~loops/.ssh/config` is written by the activation script on first rebuild (`Host cc-ci` →
+`127.0.0.1`, `git.autonomic.zone`, `tangled.org`); it is not overwritten if present.
+
+## 5. Build and activate
+
+From the checkout with the §3 commit (root can build from the loops-owned checkout via sudo):
+
+```bash
+# as root, detached (the activation restarts sshd/tailscale; a dropped session must not kill it).
+# Three things the FIRST rebuild on a bare infect system needs, none of which the converged
+# host needs afterwards: `git` on PATH (nix's flake fetcher shells out to it and the infect
+# system has none — hence nix-shell), HOME=/root (so root's `git config --global
+# safe.directory '*'` applies to the loops-owned checkout), and a login shell (`bash -l`, for
+# NIX_SSL_CERT_FILE and friends from /etc/set-environment).
+git config --global --add safe.directory '*'
+systemd-run --unit=ccci-rebuild --collect -E HOME=/root -p WorkingDirectory=/srv/cc-ci-orch \
+  bash -lc 'nix-shell -p git --run "nixos-rebuild test --flake /srv/cc-ci-orch#cc-ci"'
+journalctl -fu ccci-rebuild          # ~10 min the first time (image pulls + two OCI image builds)
+```
+
+`test` first, always: it activates WITHOUT touching the bootloader, so if the activation breaks
+networking or sshd a reboot from the Hetzner console lands on the last known-good generation.
+Later rebuilds are simply `sudo nixos-rebuild test|switch --flake .#cc-ci` from the checkout.
+
+The first activation takes a while: it pulls the traefik/drone/keycloak images, builds the bridge
+and dashboard OCI images with Nix, initialises the swarm and runs the serialized reconcile
+oneshots (`swarm-init → deploy-proxy → deploy-drone → deploy-bridge → deploy-dashboard →
+deploy-reports`, `deploy-backupbot`, `warm-keycloak`). Verify:
+
+```bash
+systemctl is-system-running        # running — or list-units --failed and read journalctl -u <unit>
+tailscale status | head -3
+docker service ls                  # traefik app+socket-proxy, drone, bridge, dashboard, reports, backups: 1/1
+systemctl status cc-ci-loops cc-ci-orchestrator opencode-web nginx acme-dns
+systemctl list-timers 'cc-ci-*' nightly-sweep
+sudo -iu loops tmux ls             # cc-ci-orchestrator (+ loops sessions if a phase is active)
+# the CI front doors, before DNS points here (expect 200 / 200 / 303 and ssl_verify=0 once
+# /var/lib/acme is restored or a cert has been issued):
+curl -s --resolve ci.commoninternet.net:443:127.0.0.1 -o /dev/null -w '%{http_code} %{ssl_verify_result}\n' https://ci.commoninternet.net/
+curl -s --resolve report.ci.commoninternet.net:443:127.0.0.1 -o /dev/null -w '%{http_code}\n' https://report.ci.commoninternet.net/
+curl -s --resolve drone.ci.commoninternet.net:443:127.0.0.1 -o /dev/null -w '%{http_code}\n' https://drone.ci.commoninternet.net/
+dig +short @<ip> ns-acme.commoninternet.net   # acme-dns answering on the public 53
+```
+
+Seen on 2026-09-07: `tailscaled-autoconnect` failed with `invalid key: API key does not exist` —
+the reusable auth key had been revoked. Generate a fresh one in the tailscale admin console, put
+it in `/etc/ts-auth-key`, `systemctl restart tailscaled-autoconnect`. Nothing else depends on it
+during the install; the box is reachable on its public IP throughout.
+
+When it is healthy: `sudo nixos-rebuild switch --flake .#cc-ci` (same config, now also the boot
+default). **If you are migrating from another host, do §6 before letting it serve anything**: right
+after the first activation scale the `!testme` bridge to 0 and mask the two orchestrator timers so
+the new box does not process PR comments or start a second weekly run while the old host is live:
+
+```bash
+docker service scale ccci-bridge_app=0
+systemctl mask --now cc-ci-upgrade-all.timer cc-ci-upgrade-supervisor.timer
+```
+
+## 6. Migrating: restore state from the previous host
+
+Over tailscale (`rsync -aHAX --numeric-ids root@<old>:<path> <path>`), with the matching service
+stopped on the new host while its directory is copied:
+
+| path | holds | notes |
+|---|---|---|
+| `/var/lib/cc-ci-reports` | the published weekly report pages (`report.ci…`) | |
+| `/var/lib/cc-ci-runs` | per-run artifacts the dashboard shows | |
+| `/var/lib/ci-warm` | warm-canonical state + alerts | recipe warm *volumes* are caches: not copied, rebuilt by the Sunday sweep / first use |
+| `/var/lib/acme` | the Let's Encrypt cert + account **and `acme-dns-accounts.json`** — the account the permanent `_acme-challenge` CNAME points at | without it a fresh registration + a new CNAME at Gandi is needed (registration is disabled in `acme-dns.nix`) |
+| `/var/lib/acme-dns` | the acme-dns zone DB | |
+| `/var/lib/ci-certs` | the copy traefik is handed | then `systemctl restart cc-ci-acme-traefik-handoff` |
+| `/root/.abra` | abra's per-app env files for the deployed stacks | |
+| Drone data volume `/var/lib/docker/volumes/drone_ci_commoninternet_net_data` | Drone's DB: the Gitea OAuth grant, repo activation, build history | `docker service scale drone_ci_commoninternet_net_app=0` on the new host, copy, scale back to 1. Otherwise run `scripts/bootstrap-drone-oauth.sh` (cc-ci repo) with the bot password and re-activate repos |
+| `/srv/cc-ci-orch/.cc-ci-logs`, `/srv/cc-ci-orch/cc-ci-plan/upstream/`, `REBOOTS.md`, `JOURNAL.md` | orchestrator history, the upgrader's per-recipe release-note registry | as loops; do the final sync after stopping the orchestrator on the old host |
+
+## 7. Cutover and verification
+
+1. **DNS** (operator, Gandi zone `commoninternet.net`): A records `ci`, `*.ci` and `ns-acme` →
+   the new IPv4. `acme NS ns-acme` and `_acme-challenge.ci CNAME <account>.acme…` stay as they
+   are. Wait for propagation (`dig +short ci.commoninternet.net`).
+2. Check the new host answers on the new IP before DNS moves: `dig @<new-ip> ns-acme.commoninternet.net`
+   (acme-dns), `curl --resolve ci.commoninternet.net:443:<new-ip> https://ci.commoninternet.net/`
+   (dashboard, valid cert), same for `report.ci` and `drone.ci`.
+3. Old host: `docker service scale ccci-bridge_app=0 drone_ci_commoninternet_net_app=0`;
+   `systemctl disable --now cc-ci-upgrade-all.timer cc-ci-upgrade-supervisor.timer` on the old
+   orchestrator. New host: `docker service scale ccci-bridge_app=1`;
+   `systemctl unmask cc-ci-upgrade-all.timer cc-ci-upgrade-supervisor.timer && systemctl start` both.
+4. End to end: post `!testme` on an open recipe PR and watch it turn green on the new Drone;
+   open `https://ci.commoninternet.net` and `https://report.ci.commoninternet.net`.
+5. The orchestrator: as loops on the new host `cd /srv/cc-ci-orch && python3 cc-ci-plan/agents.py up orchestrator`
+   (or just `systemctl restart cc-ci-orchestrator`), attach with `claude --resume` or from
+   claude.ai/code. Its startup routine (AGENTS.md) reports phase + reboot count.
+6. Keep the old host as a cold standby for a week, then delete it and its tailnet node.
+
+## 8. Day 2
+
+- **Update the host** (nixpkgs bump for both halves): `/cc-ci-orchestrator-update`, which is
+  `nix flake update` → `nixos-rebuild test` → verify → `switch` → PR. The `cc-ci` input follows
+  this flake's nixpkgs, so the CI server is rebuilt on the same nixpkgs.
+- **Update only cc-ci's code** (harness/tests/modules): merge in the cc-ci repo, then
+  `nix flake update cc-ci` here and rebuild; also `git -C /etc/cc-ci pull --recurse-submodules`
+  so the deployed checkout the sweep runs from matches.
+- **Something is down**: `systemctl --failed`, `journalctl -u deploy-<x>`, `docker service ps <svc>`;
+  the cc-ci repo's `docs/runbook.md`. Host unreachable: Hetzner console → reboot lands on the last
+  `switch`ed generation; rescue mode + `nixos-enter` for anything worse (skill
+  `/hetzner-server-recovery`).
+
+---
+
+# Operating the orchestrator session
+
+The steering session is a long-lived interactive Claude Code session under tmux with
+`--remote-control`, so it can be watched and steered from **claude.ai/code** (or the mobile app).
+`cc-ci-orchestrator.service` relaunches it on boot via `cc-ci-plan/agents.py up orchestrator`
+(backend + model in `cc-ci-plan/agents.toml`).
+
+```bash
+# attach on the host
+sudo -iu loops tmux attach -t cc-ci-orchestrator
+# or resume the conversation by hand in a fresh tmux pane
+claude --resume autonomous-orchestrator --remote-control "autonomous-orchestrator" --dangerously-skip-permissions
+# already inside a live session and just want the web surface? /remote-control
+```
+
+`--resume <name|id>` selects the *conversation* to restore; the `--remote-control "<name>"` value is
+only the web display label. Don't pass `--fork-session` unless you mean to branch.

 ## Kick off / supervise the loops

@@ -53,5 +350,5 @@ cd /srv/cc-ci/cc-ci-plan
 ./launch.sh stop
 ```

-Full supervision guide, credential map, and the Incus VM fallback are in
-[`cc-ci-plan/kickoff.md`](cc-ci-plan/kickoff.md) and [`cc-ci-plan/plan.md`](cc-ci-plan/plan.md) §1.5.
+Full supervision guide, credential map and history are in `cc-ci-plan/kickoff.md`,
+`cc-ci-plan/plan.md` §1.5 and `cc-ci-plan/JOURNAL.md`.
@@ -0,0 +1,22 @@
+# archive/ — how cc-ci and its orchestrator were built and moved, before the combined host
+
+Historical record only. Nothing in here is deployed or evaluated. It was moved out of the live
+tree on 2026-09-07 when the CI server and the orchestrator were consolidated onto one Hetzner
+host (`nixosConfigurations.cc-ci` in `../flake.nix`; deploy guide in `../README.md`; the plan
+that did it is `../cc-ci-plan/plan-cc-ci-combined-host.md`).
+
+| path | what it was |
+|---|---|
+| `nix/configuration-incus-vm.nix` | Channel-based NixOS config of the first orchestrator VM on b1 (Incus, 2 GB). Ran the loops as root; hard-coded the dead Incus cc-ci IP. Replaced by the Hetzner host 2026-05-31. |
+| `nix/README.md` | The README for that Incus VM config. |
+| `nix/cc-ci-orchestrator-hetzner/` | The orchestrator's own Hetzner `cpx22` host (`168.119.126.100`, tailnet `cc-ci-orchestrator-1`), 2026-05-31 → 2026-09. From 2026-08-20 the live copy of this config was `notplants-nix`'s `notplants-orchestrator` host (the box became a shared agent host for several projects); this one had drifted and still carried lichen/project-orchestrator units. Superseded by `../nix/hosts/cc-ci` + `../nix/modules/orchestrator-host.nix`. |
+| `nix/atproto-likes.nix` | A notplants (not cc-ci) service that lived on the shared box; kept by notplants-nix. |
+| `terraform/` | OpenTofu for the `cpx22` orchestrator server (Debian 12 → nixos-infect at `nixos-24.11`). The combined host was provisioned by hand instead; the README documents that path. Note its `user-data.sh` would fail on the Debian 13 image (nixos-infect's temp swapfile on a tmpfs `/tmp`) — see the README's `NO_SWAP=true` note. |
+| `plans/plan-orchestrator-migration.md` | Pi → Incus VM move of the orchestrator (2026-05). |
+| `plans/plan-orchestrator-hetzner-migration.md` | Incus VM → Hetzner `cpx22` move of the orchestrator (2026-05-31). Has the reboot-resilience design (`cc-ci-loops.service`). |
+| `plans/plan-migrate-cc-ci-to-hetzner.md`, `plans/plan-cc-ci-hetzner-migration.md`, `plans/plan-cc-ci-hetzner-terraform.md` | The CI server's own move from the `cc-nix-test` Incus VM to Hetzner `cpx32` (`91.98.47.73`, 2026-05-31), and the terraform that provisioned it (lives in the cc-ci repo). |
+| `plans/plan-repo-consolidation.md` | The earlier repo layout consolidation. |
+
+The cc-ci server's own history (machine-docs, decisions, the clean-room rebuild that proved
+"two repos + one age key + one `nixos-rebuild switch`") is in the cc-ci repo under
+`machine-docs/` and `docs/`.
@@ -14,7 +14,11 @@
  # with ACME, and opens 80/443 (previously only 22 was public). Canonical source of
  # the module is the project repo; ../atproto-likes.nix is a copy kept in this tree
  # because pure evaluation cannot import a path outside the flake.
-  imports = [ ../../atproto-likes.nix ];
+  # cc-ci.nix holds everything that exists FOR cc-ci (the loops, the cc-ci orchestrator session,
+  # the weekly recipe-upgrade run and its hourly supervisor). Split out 2026-08-20 so this host —
+  # a general agent box that also serves unrelated projects — can be reasoned about separately
+  # from cc-ci. Also exported as `nixosModules.cc-ci` from this repo's flake for other hosts.
+  imports = [ ../../atproto-likes.nix ../../modules/cc-ci.nix ];

  services.openssh = {
    enable = true;
@@ -127,7 +131,8 @@ SSHCFG

  # opencode web server — one shared instance; agent sessions attach to it for web visibility.
  # Serves the web UI at http://oc.commoninternet.net (via nginx below, tailscale-only).
-  # Provider creds are read from /srv/cc-ci/.testenv at startup.
+  # Public provider configuration is tracked in the cc-ci repo; credentials
+  # remain in /srv/cc-ci/.testenv.
  systemd.services.opencode-web = {
    description = "opencode web server for cc-ci agents";
    wantedBy = [ "multi-user.target" ];
@@ -137,7 +142,10 @@ SSHCFG
      Type = "simple";
      User = "loops"; Group = "users";
      WorkingDirectory = "/srv/cc-ci-orch/cc-ci";
-      EnvironmentFile = "/srv/cc-ci/.testenv";
+      EnvironmentFile = [
+        "-/srv/cc-ci/cc-ci/.env.public"
+        "/srv/cc-ci/.testenv"
+      ];
      ExecStartPre = "${pkgs.coreutils}/bin/rm -rf /tmp/opencode";
      ExecStart = "/home/loops/.local/bin/opencode serve --hostname 127.0.0.1 --port 4096";
      Restart = "on-failure";
@@ -225,72 +233,7 @@ SSHCFG
    };
  };

-  # cc-ci-loops supervisor — workspace staged 2026-05-31, so ENABLED for reboot-resilience.
-  systemd.services.cc-ci-loops = {
-    description = "cc-ci Builder/Adversary loops + watchdog (launch.sh start)";
-    wantedBy = [ "multi-user.target" ];  # enabled after workspace staged (Hetzner cutover)
-    after = [ "network-online.target" "tailscaled.service" "claude-install.service" ];
-    wants = [ "network-online.target" ];
-    serviceConfig = {
-      # KillMode=process: this unit only LAUNCHES the tmux server, it does not own it. With the
-      # default (control-group) systemd kills every leftover process in the cgroup when the unit
-      # stops — and since one tmux server hosts every agent session on this host, a rebuild that
-      # merely touched this unit wiped all of them (operator 2026-08-01). Only the (already
-      # exited) main process is killed now; `systemctl stop` therefore does NOT tear down agents.
-      KillMode = "process";
-      Type = "oneshot"; RemainAfterExit = true;
-      User = "loops"; Group = "users";
-      WorkingDirectory = "/srv/cc-ci/cc-ci";
-      # Append one line to REBOOTS.md per genuine reboot (boot_id-gated; not on manual restart).
-      ExecStartPre = "${pkgs.bash}/bin/bash /srv/cc-ci/cc-ci-plan/reboot-log.sh";
-    };
-    # CLAUDE_BIN points at the standalone CLI installed by claude-install.service; the loops
-    # backend defaults to claude (persisted in .loop-backend). Without this, launch.py's preflight
-    # `which(claude)` fails because the systemd `path` below has no /home/loops/.local/bin.
-    environment = { RESUME_PHASE = "1"; HOME = "/home/loops"; CLAUDE_BIN = "/home/loops/.local/bin/claude"; };
-    path = [ pkgs.bash pkgs.tmux pkgs.git pkgs.python3 pkgs.openssh pkgs.nettools ];
-    script = ''
-      # Put the standalone claude/opencode binaries on PATH. On a cold boot this is the env the
-      # tmux server (and thus every agent session) inherits, so bare `claude` resolves everywhere.
-      export PATH="/home/loops/.local/bin:$PATH"
-      [ -x /srv/cc-ci/cc-ci-plan/launch.sh ] && /srv/cc-ci/cc-ci-plan/launch.sh start || \
-        echo "workspace not staged yet — skipping loop start"
-    '';
-  };

-  # cc-ci-orchestrator supervisor — the operator's steering session. Same shape as
-  # lichen-orchestrator / project-orchestrator above: this unit only LAUNCHES the orchestrator's
-  # tmux session via the agent-orchestrator harness (cc-ci-plan/agents.py); it does not own the
-  # session or the tmux server. The orchestrator agent is declared in cc-ci-plan/agents.toml on
-  # the OPencode backend (backend = "opencode", model = "opencode/glm-5.2"), so on boot it
-  # attaches to the shared opencode web server (opencode-web.service below) and is reachable for
-  # Remote Control at https://oc.commoninternet.net under the /srv/cc-ci-orch project. The harness
-  # watchdog (started by `agents.py up`) keeps it alive: heal-only (no stall reboots — a persistent
-  # supervisor must not be killed just for idling). Added 2026-08-03 to give the cc-ci orchestrator
-  # the same reboot-resilience the other two orchestrators already have.
-  systemd.services.cc-ci-orchestrator = {
-    description = "cc-ci orchestrator (operator steering session) — agents.py up orchestrator, opencode backend";
-    wantedBy = [ "multi-user.target" ];
-    after = [ "network-online.target" "tailscaled.service" "opencode-web.service" ];
-    wants = [ "network-online.target" ];
-    serviceConfig = {
-      # KillMode=process: see the note on cc-ci-loops — a rebuild that merely touches this unit
-      # must not tear down the (shared) tmux server and every agent session with it.
-      KillMode = "process";
-      Type = "oneshot"; RemainAfterExit = true;
-      User = "loops"; Group = "users";
-      WorkingDirectory = "/srv/cc-ci-orch";
-    };
-    environment = { HOME = "/home/loops"; };
-    path = [ pkgs.bash pkgs.tmux pkgs.git pkgs.python3 pkgs.openssh pkgs.nettools ];
-    script = ''
-      export PATH="/home/loops/.local/bin:$PATH"
-      proj="/srv/cc-ci-orch"
-      echo "$(cat /proc/sys/kernel/random/boot_id) boot $(date -u +%FT%TZ) — cc-ci-orchestrator up" \
-        >> "$proj/cc-ci-plan/.ao-boot.log" 2>/dev/null || true
-      cd "$proj" && python3 cc-ci-plan/agents.py up orchestrator || echo "cc-ci orchestrator agents.py up failed"
-    '';
-  };

  # p-lichen-orchestrator supervisor — the SEPARATE lichen.page testing/hardening orchestrator
  # (distinct from cc-ci-loops above). Reboot-resilience: on boot, resume the orchestrator's Remote
@@ -356,76 +299,7 @@ SSHCFG
    '';
  };

-  # Weekly recipe upgrade — runs /upgrade-all over every enrolled recipe (opens recipe PRs
-  # verified by !testme, never merges). Replaces the boot-fragile busybox-crond-in-tmux from
-  # phase 5 §4 with a reboot-safe systemd timer. The service is timer-triggered only (NOT
-  # wantedBy multi-user.target) so it never runs on boot/activation — only on the schedule.
-  systemd.services.cc-ci-upgrade-all = {
-    description = "cc-ci weekly /upgrade-all run (recipe upgrade survey + PRs, never merges)";
-    after = [ "network-online.target" "tailscaled.service" "claude-install.service" ];
-    wants = [ "network-online.target" ];
-    serviceConfig = {
-      Type = "oneshot";  # launch-upgrader.py spawns the cc-ci-upgrader tmux session and returns
-      User = "loops"; Group = "users";
-      WorkingDirectory = "/srv/cc-ci";
-      # Optional per-run overrides for backend/model (LOOP_BACKEND, LOOP_MODEL, OPENCODE_SHARE,
-      # UPGRADER_ARGS, …). The leading "-" makes it optional: absent file → claude/sonnet defaults
-      # (current behavior). To run the weekly job on e.g. opencode-go/glm-5.2, drop a file with
-      #   LOOP_BACKEND=opencode
-      #   LOOP_MODEL=opencode-go/glm-5.2
-      # No rebuild needed to switch — the env file is read at each timer fire. Holds no secrets
-      # (the opencode-go API key lives in ~/.local/share/opencode/auth.json, mode 600).
-      EnvironmentFile = "-/srv/cc-ci/upgrader.env";
-    };
-    environment = { HOME = "/home/loops"; CLAUDE_BIN = "/home/loops/.local/bin/claude"; };
-    path = [ pkgs.bash pkgs.tmux pkgs.git pkgs.python3 pkgs.openssh pkgs.nettools ];
-    script = ''
-      export PATH="/home/loops/.local/bin:$PATH"
-      python3 /srv/cc-ci/cc-ci-plan/launch-upgrader.py start >> /srv/cc-ci/.cc-ci-logs/upgrader-cron.log 2>&1
-    '';
-  };

-  systemd.timers.cc-ci-upgrade-all = {
-    description = "Weekly trigger for cc-ci-upgrade-all (Thursdays 22:00 America/New_York — Boston 10pm)";
-    wantedBy = [ "timers.target" ];
-    timerConfig = {
-      # 10pm Thursday Boston time — DST-aware (EDT→02:00 UTC, EST→03:00 UTC) via the tz in OnCalendar.
-      OnCalendar = "Thu *-*-* 22:00:00 America/New_York";
-      Persistent = true;  # if the box was down at the scheduled time, run once on next boot
-    };
-  };

-  # Hourly SUPERVISOR — a glm-5.2 orchestrator wake-up that keeps the weekly run on track. The
-  # log-idle/429 watchdog only handles opencode-go usage-limit stalls; it does NOT cover a host
-  # disk-full crash (which killed the 2026-07-03 run) or any other environmental wedge. This is a
-  # CHEAP deterministic gate: if the weekly run is complete or actively progressing it does NOTHING
-  # (zero model tokens). Only when a run has stalled/died before completing does it launch a
-  # short-lived glm-5.2 agent that diagnoses the blockage and drives the run to a clean DONE.
-  systemd.services.cc-ci-upgrade-supervisor = {
-    description = "cc-ci hourly weekly-run supervisor (glm-5.2 — drives a stalled /upgrade-all to completion)";
-    after = [ "network-online.target" "tailscaled.service" ];
-    wants = [ "network-online.target" ];
-    serviceConfig = {
-      Type = "oneshot";  # launch-supervisor.py check: gate now, spawn the agent into tmux, return
-      User = "loops"; Group = "users";
-      WorkingDirectory = "/srv/cc-ci";
-      # Shares the weekly run's optional override file (e.g. SUPERVISOR_MODEL=…); "-" = optional.
-      EnvironmentFile = "-/srv/cc-ci/upgrader.env";
-    };
-    environment = { HOME = "/home/loops"; };
-    path = [ pkgs.bash pkgs.tmux pkgs.git pkgs.python3 pkgs.openssh pkgs.nettools ];
-    script = ''
-      export PATH="/home/loops/.local/bin:$PATH"
-      python3 /srv/cc-ci/cc-ci-plan/launch-supervisor.py check >> /srv/cc-ci/.cc-ci-logs/supervisor-cron.log 2>&1
-    '';
-  };

-  systemd.timers.cc-ci-upgrade-supervisor = {
-    description = "Hourly trigger for cc-ci-upgrade-supervisor (weekly-run health check + drive)";
-    wantedBy = [ "timers.target" ];
-    timerConfig = {
-      OnCalendar = "*-*-* *:07:00";  # every hour at :07 (offset from the weekly :00 fire)
-      Persistent = false;            # a missed hourly check is moot — the next hour re-checks
-    };
-  };
 }
@@ -867,6 +867,129 @@ session cc-ci-orchestrator-stale can be killed; recipe-mirrors org still private
  (/srv/cc-ci-orch/cc-ci), and task-tool subagents inherit their parent session's directory. The
  config now lives in the cc-ci repo at that path. VERIFIED end-to-end with the launcher's exact
  invocation: parent=glm-5.2, subagent=deepseek-v4-pro read back from the session DB.
-  LESSON: `opencode debug config` proves resolution, NOT binding — only a live subagent's recorded
-  modelID proves binding. First attempt was a false pass because the probe passed --dir (unlike the
-  real launcher) and landed in a different project.
+   LESSON: `opencode debug config` proves resolution, NOT binding — only a live subagent's recorded
+   modelID proves binding. First attempt was a false pass because the probe passed --dir (unlike the
+   real launcher) and landed in a different project.
+
+## Session 2026-08-15 19:25 UTC — opencode glm-5.2
+
+**Left off:** Recovered the stalled 2026-08-14 weekly /upgrade-all run. Killed a supervisor that had
+been relaunching hourly for ~40h (balance exhausted), then started a FRESH scoped upgrader. Run is now
+progressing (surveying the 9 remaining recipes). Watching it through to completion.
+
+**What happened (the stall):**
+- The 2026-08-14 /upgrade-all run (session ses_00200382fffeYIGl2sc3mO9JId) stalled at 03:18 Aug 14
+  mid-`lasuite-drive` with `Error: Insufficient balance` (opencode zen workspace balance ran out). It
+  had already done bluesky-pds, ghost, gitea, hedgedoc (PRs) + immich, lasuite-docs (SKIPPED up-to-date)
+  alphabetically; lasuite-drive had a plan + partial PR #6 but no RESULT/verify.
+- The supervisor cron (glm-5.2, opencode-go tier) relaunched an hourly one-shot supervisor ~40 times
+  to "drive it to completion", but each was also balance-walled (and later, just spinning). The run sat
+  INCOMPLETE + not progressing for 40h. No weekly summary, no report published for week of Aug 14.
+
+**What I did this session:**
+- Diagnosed: the opencode zen endpoint is NOW healthy (direct probe `say OK` → HTTP 200 in 1.35s —
+  balance is restored). But resuming the ORIGINAL giant session is impossible: it's 2.58M tokens
+  (267K input + 2.3M cache) and `opencode run -s … --continue` sits idle on `do_epoll_wait` with zero
+  I/O — the inference endpoint silently drops the oversized request (matches the supervisor's
+  `socket connection was closed unexpectedly` errors). A fresh small `opencode run` works fine. So the
+  giant session is unresumable; a fresh start is the only path.
+- Killed the stuck supervisor (tmux `cc-ci-supervisor`, proc 377329).
+- `UPGRADER_ARGS="lasuite-drive lasuite-meet mailu matrix-synapse mattermost-lts mumble n8n plausible
+  wordpress --sequential" python3 /srv/cc-ci/cc-ci-plan/launch-upgrader.py fresh` — this killed the
+  stuck resume, archived the old giant session (`archive-cc-ci-upgrader — 2026-08-14`), reclaimed 10GB
+  stale images on cc-ci (disk 29%), and started a FRESH small session
+  `ses_ff920cf39ffeoogwXHTajp94cr` (zen/glm-5.2) scoped to the 9 recipes not yet done this week
+  (positions 13-21 alphabetically; positions 1-12 were already surveyed — 6 PRs + 6 up-to-date). A
+  fresh watchdog is watching the new session. The skill is idempotent (reuses existing PRs incl.
+  lasuite-drive #6, never duplicates), so scoping is safe.
+- Confirmed the fresh run is progressing: pane shows it surveying the 9 recipes (verified all present
+  in abra + all `weekly` tier; currently probing plausible/wordpress tags). Proc alive, log growing.
+
+**Phase / loop state:**
+- Build/adversary loops: STOPPED (whole sequence completed 2026-08-01; phase ghost DONE).
+- Weekly upgrader: RUNNING (fresh session ses_ff920cf39, scoped 9 recipes, --sequential, watchdog up).
+- cc-ci server: healthy (disk 29%, runner active).
+
+**Open items for next session:**
+- **Monitor the fresh upgrader to completion.** It will survey the 9 recipes, /recipe-upgrade the
+  upgradeable ones (subagents, !testme verify, open/extend PRs — NEVER merge), write the weekly summary
+  to `/srv/cc-ci/.cc-ci-logs/upgrades/`, then `launch-report.py fresh` (the upgrade-all skill does this
+  itself per SKILL.md §5), print `UPGRADE RUN COMPLETE`, and go idle. If it stalls on a usage limit,
+  the watchdog auto-resumes the SAME (small) session — that works now.
+- **Do NOT try to resume the archived giant session ses_00200382** — it's unresumable (endpoint drops
+  the 2.58M-token request). It's archived; leave it.
+- After the run completes + report publishes, operator review queue = this week's recipe PRs.
+- The supervisor cron (hourly at XX:07) should now leave the run alone once it's progressing; if a
+  supervisor fires while the run is mid-flight, its guardrails say to hand back to the resumed run, not
+  double-write. No action needed unless it interferes.
+
+**Notes:**
+- Root cause of the 40h silence was the same BUG 1 from 2026-08-10 (supervisor progress gate) partly:
+  the supervisor kept firing because the run never reached "progressing". Now that balance is restored
+  and a fresh small session is running, the gate should see progress and stand down.
+- Lesson: when a weekly run dies mid-flight on a giant context, do NOT resume the original session —
+  start fresh and scope to the remaining recipes. The /upgrade-all skill is idempotent so this is safe.
+
+## Session 2026-08-31 17:17 UTC — restricted acme-dns renewal staged
+
+**What happened:** Implemented and deployed the cc-ci restricted ACME DNS-01 design in cc-ci PRs
+#24–#27 (all merged). The initial `sqlite3` backend setting was detected during staging bootstrap
+and corrected to acme-dns 2.x `sqlite` before any account registration succeeded.
+
+**Live state:**
+- `acme-dns` is authoritative for `acme.commoninternet.net` on `91.98.47.73`, with public UDP/TCP
+  53 verified directly; it serves the apex NS and `ns-acme.commoninternet.net A 91.98.47.73`.
+- Its HTTP update API is bound solely to `127.0.0.1:8080`; registration is disabled (local POST
+  returns 403). The one restricted Lego account is stored mode `0600` under `/var/lib/acme/`, outside
+  Git and the Nix store. No Gandi credentials were used or stored.
+- Nix ACME is in Let's Encrypt staging mode for `ci.commoninternet.net` and its wildcard; its renewal
+  timer is enabled. Traefik handoff remains production-gated so staging cannot replace the live cert.
+- Every server deployment used `nixos-rebuild test`, reachability/service verification, then
+  `nixos-rebuild switch` on `#cc-ci-hetzner`.
+
+**Manual Gandi handoff (only blocker):** Add the parent-zone A+NS delegation and permanent CNAME:
+`ns-acme.commoninternet.net. A 91.98.47.73`,
+`acme.commoninternet.net. NS ns-acme.commoninternet.net.`, and
+`_acme-challenge.ci.commoninternet.net. CNAME
+9d3f0428-eb97-4b14-bb07-da2f0f6c2073.acme.commoninternet.net.` Once propagated, resume with a
+staging issuance, then production cutover/handoff verification. Do not obtain or add a Gandi token.
+
+## Session 2026-08-31 19:05 UTC — ACME production issuance complete
+
+**Completed:** The operator added the narrow Gandi delegation and permanent CNAME. Staging then
+validated both `ci.commoninternet.net` and `*.ci.commoninternet.net`; production issuance validated
+the same SAN pair and produced a Let's Encrypt certificate valid through 2026-11-29. cc-ci PRs #28–#30
+were merged to promote production, make Traefik roll certificate-secret changes even when recipe
+upgrades are held, and remove the retired SOPS wildcard-file injection.
+
+**Live cc-ci state:**
+- `/var/lib/acme/ci.commoninternet.net` is the sole source of the production cert/key; the runtime
+  pair matches it, is root-only for the key, and Traefik now references matching versioned Swarm
+  secrets. Reboot-safe `nixos-rebuild switch` completed at cc-ci `7147d77` after the required test
+  activations and health checks.
+- acme-dns remains authoritative only for the delegated subzone, its update API is localhost-only,
+  registration is disabled, and the restricted-account JSON is mode `0600`. The Nix ACME renewal
+  timer is active. No Gandi credential was added.
+
+**External blocker, not cc-ci-owned:** Direct TLS to `91.98.47.73:443` now presents the new valid
+certificate. The operator-managed public gateway at `143.244.213.108` currently accepts TCP then
+closes TLS before presenting a certificate, so its passthrough to cc-ci needs repair/verification.
+The cc-ci plan explicitly marks the gateway as operator infrastructure; do not change the DNS
+delegation or add a Gandi token to work around it.
+
+## Session 2026-08-31 20:45 UTC — weekly DeepSeek Flash run and report complete
+
+**Completed:** Started the weekly upgrader explicitly on `opencode/deepseek-v4-flash`; its parent
+and every recipe subagent were confirmed with that persisted model. All nine eligible recipe PRs
+completed `!testme` successfully. The public report was then restarted before publication when the
+first handoff was found to use the old GLM model; the replacement report session was confirmed
+`opencode/deepseek-v4-flash` and published successfully at
+`https://report.ci.commoninternet.net/week-2026-08-31.html` (HTTP 200).
+
+**Follow-up fixes:** Orchestrator PR #17 (`1b75d98`) made `testme-on-pr.sh` apply the documented
+public Gitea-host default. cc-ci PR #31 (`769fd29`) adds that public hostname to `.env.public`.
+Both commits were scanned clean and contain no coauthor trailers. No recipe PR was merged.
+
+**Security note:** A subagent briefly enabled shell tracing while debugging the verifier, exposing
+runtime credentials in its private agent trace. No values were committed or put in this journal,
+but rotate the affected `/srv/cc-ci/.testenv` credentials as a precaution.
@@ -39,6 +39,8 @@ keeps landing in pass 2, the fix is a new deterministic method in pass 1. §4c i
 ```
 advisory-scan.py <recipe> [--from <version>] [--to <version>]
                          [--image <name>=<from>:<to>]... [--adjudicate] [--json] [--registry DIR]
+
+advisory-scan.py <recipe> --compose-to <URL> [--compose-from <URL>]   # windows derived, not typed
 ```

 | Input | Meaning |
@@ -46,6 +48,8 @@ advisory-scan.py <recipe> [--from <version>] [--to <version>]
 | `<recipe>` | Recipe name; selects `cc-ci-plan/upstream/<recipe>.md` (the per-recipe URL registry) |
 | `--from` / `--to` | The **primary app image's** version window being upgraded across |
 | `--image NAME=FROM:TO` | A **sidecar image and the versions it moved between** (repeatable, all in ONE call). `NAME` is substring-matched against source repo names. Malformed values warn on stderr and are skipped. Without it that image's advisories stay unclassified. |
+| `--compose-to URL` | **Derive every window by diffing this compose against its baseline**, instead of typing `--from/--to/--image`. Point it at a PR's `compose.yml`. |
+| `--compose-from URL` | Baseline for the above. Default: the same repo's **default branch, resolved from the API** — never assumed to be `main`. |
 | `--adjudicate` | Run pass 2: append the evidence dossier for judgement |
 | `--registry` | Registry dir; also `CCCI_UPSTREAM_REGISTRY` |
 | `GITHUB_TOKEN` / `GITHUB_TOKEN_FILE` | Read-only token; **rate limit only** (60/hr anonymous → 5000/hr). Default file `/srv/cc-ci/.github-token`, mode 600. Public advisories need **no scopes**. |
@@ -103,10 +107,43 @@ URLs containing `<`, `>`, `{`, `}`, `VERSION`, or `vX.Y.Z` are **skipped as temp
 human documentation (`…/changelog/v<VERSION>/`), not fetchable, and counting them as failures is wrong.

 This is the source that would have caught gitea: the vendor blog names both CVEs, the GitHub release
-page names neither. A CVE found **only** here carries no version data, so pass 1 cannot place it — it
-goes to pass 2 (§6).
+page names neither.

-### 2c. OSV.dev — supplementary
+**When the page is a changelog organised by release, each CVE is attributed to the release heading it
+appears under** (`Changes with nginx 1.31.3`, `## v1.31.3`, …) and that becomes its fixed-in version.
+Without this, a project that publishes no advisory feed can never contribute a CVE:
+
+> **nginx publishes NO GitHub security advisories.** Every nginx CVE we can see comes from
+> `nginx.org/en/CHANGES`. Scraping ids out of it without attributing them to a release left them with
+> no patched version, so they were never classifiable — and every nginx bump in the fleet reported
+> **0** forever. nginx is a sidecar in most recipes. Measured: `1.31.1 → 1.31.3` fixes **six** CVEs
+> (three in .2, three in .3); lasuite-docs#7 went 0 → 6 and lasuite-drive#6 went 0 → 3 on this alone.
+
+A changelog CVE is tied to a window by the **image name appearing in the page URL** (window `nginx` ↔
+`nginx.org/...`). A CVE found on a vendor page with no attributable release still has no version data,
+so pass 1 cannot place it — it goes to pass 2 (§6).
+
+### 2c. NVD by CPE — the fallback for projects that publish nothing
+
+Declared per recipe in the registry as `nvd-cpe: <image-key> = <cpe:2.3:...>`.
+
+> **Why it exists.** Two recipes could not see CVEs *at all*: `mattermost-lts` (empty GitHub advisory
+> feed, security bulletins rendered client-side so a text sweep finds nothing) and `mumble` (nothing
+> published anywhere the registry points). Their scans returned `?` — nothing measured. NVD is
+> CPE-indexed and carries structured ranges, so it answers where the vendor does not: mattermost
+> 10.5.0 → 10.12.4 now scores **165**, and mumble finds `CVE-2025-71264` (fixed 1.6.870).
+
+Two range forms, both used:
+
+| NVD field | meaning | how it is judged |
+|---|---|---|
+| `versionEndExcluding X` | fixed in X exactly | a normal patched version (§4a) |
+| `versionEndIncluding X` | affected **up to and including** X; fix version unpublished | fixed when the upgrade crosses X, i.e. `from ≤ X < to` |
+
+**NVD lags the vendor** — it had neither gitea CVSS-9.8 RCE at publication — so this is a fallback,
+never a replacement for 2a/2b. Unauthenticated calls are rate-limited (~5/30s), hence the retry.
+
+### 2d. OSV.dev — supplementary

 Only when the recipe has an entry in `OSV_PACKAGES` (ecosystem + package) and a version is given.

@@ -140,8 +177,44 @@ Two invariants govern this step, both learned from a wrong answer in production.
 > `null` / `UNKNOWN`, never `0`. A `0` in a security column asserts safety. Equally, an advisory that
 > cannot be judged is **indeterminate** (§4d) — never silently counted as "not fixed".

+### 3b. Deriving the windows from a compose diff (`--compose-to`)
+
+Typing `--from/--to/--image` by hand means someone has to remember that the recipe also bumped its
+redis. That is how sidecar CVEs went uncounted for months. This mode reads the windows off the diff:
+
+1. Fetch both compose files (baseline = the repo's **default branch from the API**, since several
+   recipes keep a stale `main` beside a live `master`).
+2. Parse `{service: (image-repo, tag)}` — keyed by **service, not image repo**, because an upgrade
+   may change the repo itself (plausible moved `plausible/analytics` →
+   `ghcr.io/plausible/community-edition`; keyed by repo that reads as one image vanishing and an
+   unrelated one appearing, losing the app window entirely).
+3. Every service whose tag or repo changed becomes a window. The `app` service drives `--from/--to`
+   (coop-cloud convention: it is the recipe's primary image); the rest become `--image` windows.
+   Unchanged images produce no window — inventing one would be a false count.
+4. The derived windows are printed to stderr before the scan, so the inputs are auditable.
+
+Image names are matched against advisory sources **both ways** — an image name is often longer than
+its source repo (`clickhouse/clickhouse-server` vs `ClickHouse/ClickHouse`) and sometimes shorter
+(`redis` vs `redis/redis`).
+
+Verified on plausible PR #5: from the compose URL alone it derives `v2.0.0 → v3.2.1` plus
+`clickhouse-server 23.4.2.11-alpine → 24.12-alpine`, and reports **6** — identical to the
+hand-specified args.
+
+`--from/--to/--image` remain available for finer-grained checks (scanning a window that is not a
+literal compose diff, e.g. "what would the compatibility-safe target fix?").
+
 ### 4a. By patched version (preferred — exact)

+**A fix on the line you are upgrading FROM was already yours.** Projects that maintain several lines
+patch them all at once: mattermost fixed `CVE-2025-11794` in 10.11.4, 10.12.1 *and* 10.5.12. An
+upgrade 10.11.22 → 10.12.4 crosses 10.12.1, so a naive window test counts it — but 10.11.22 is
+already past 10.11.4, so the deployment had the fix before the upgrade. Counting it credits the
+upgrade with work it did not do. This check is **skipped for placeholder versions** (`7.4.X` parses
+to a bare `7.4`, which would read as "already fixed at 7.4" and silently drop a real fix — exactly
+how redis `CVE-2024-46981` was lost when the rule was first added).
+
+
 `patched_versions` is a **range expression** (`">= 2.18.1"`), possibly several joined by `;`. Extract
 every version-looking token; the advisory is **fixed-by-this-upgrade** if **any** patched version `p`
 satisfies `from < p <= to` — exclusive lower (a fix already in the version you were on is not this
@@ -44,7 +44,9 @@ import json
 import os
 import re
 import sys
+import time
 import urllib.error
+import urllib.parse
 import urllib.request

 REGISTRY_DIR = os.environ.get("CCCI_UPSTREAM_REGISTRY", "/srv/cc-ci/cc-ci-plan/upstream")
@@ -164,6 +166,42 @@ def _vkey(v: str | None) -> tuple:
    return tuple(out)


+def _already_fixed_on_from_line(kf: tuple, cands: list[tuple]) -> bool:
+    """Was it ALREADY fixed on the line we are upgrading FROM?
+
+    The mirror image of _superseded_on_target_line, and just as necessary. mattermost fixes each CVE
+    across several maintained lines at once — CVE-2025-11794 is patched in 10.11.4, 10.12.1 and
+    10.5.12. Upgrading 10.11.22 -> 10.12.4 crosses 10.12.1, so a naive window test counts it; but
+    10.11.22 is already past 10.11.4, so the deployment HAD the fix before the upgrade. Counting it
+    credits the upgrade with work it did not do."""
+    if len(kf) < 2:
+        return False
+    line = kf[:2]
+    for c in cands:
+        if len(c) < 2 or c[:2] != line:
+            continue
+        n = max(len(kf), len(c))
+        pad = lambda z: z + (0,) * (n - len(z))
+        if pad(c) <= pad(kf):
+            return True
+    return False
+
+
+def _superseded_on_target_line(kt: tuple, cands: list[tuple]) -> bool:
+    """Does a patched version on the TARGET's own release line sit ABOVE the target?
+
+    Projects maintain several branches at once and backport per branch, so "some patched version is
+    inside the numeric window" is not the same as "the version we land on has the fix". ClickHouse
+    fixed CVE-2023-48704 in 23.9.6.20 AND 23.10.5.20; an upgrade landing on 23.10.4.25 crosses the
+    23.9 fix numerically but is still BELOW its own line's fix, so it does NOT have it. When the
+    advisory names a fix on the target's own line and the target is older than it, that is proof of
+    absence and outranks any other candidate."""
+    if len(kt) < 2:
+        return False
+    line = kt[:2]
+    return any(c[:2] == line and c > kt for c in cands if len(c) >= 2)
+
+
 def _within(kf: tuple, kt: tuple, c: tuple) -> bool:
    """Is patched-version `c` inside the window (kf, kt] — exclusive lower, inclusive upper?

@@ -257,6 +295,51 @@ def github_advisories(urls: list[str]) -> list[dict]:
    return results


+# Release headings in a vendor changelog. nginx's CHANGES uses "Changes with nginx 1.31.3", most
+# markdown changelogs use "## 1.31.3" / "## v1.31.3".
+_HEADING_RE = re.compile(
+    r"^\s*(?:#{1,4}\s*)?(?:Changes with\s+\S+\s+|Version\s+|Release\s+)?v?(\d+\.\d+(?:\.\d+)*)\s*$"
+    r"|^\s*Changes with\s+\S+\s+(\d+\.\d+(?:\.\d+)*)", re.I)
+
+
+def _changelog_versions(text: str) -> dict:
+    """{cve: version} for a changelog that is ORGANISED BY RELEASE.
+
+    Why this exists: nginx publishes NO GitHub security advisories. Every nginx CVE we can see comes
+    from nginx.org/en/CHANGES, and scraping ids out of it without attributing them to a release
+    leaves them with no patched version — so they can never be classified, and an nginx bump reports
+    0 CVEs forever. nginx 1.31.1 -> 1.31.3 in fact fixes SIX (three in .2, three in .3), and nginx is
+    a sidecar in most of the fleet, so that was a fleet-wide blind spot.
+
+    Attributes each CVE to the nearest PRECEDING release heading — the release that fixed it.
+    """
+    plain = re.sub(r"<[^>]+>", " ", text)
+    out, cur = {}, None
+    for line in plain.splitlines():
+        m = _HEADING_RE.match(line)
+        if m:
+            cur = m.group(1) or m.group(2)
+            continue
+        if cur:
+            for cve in CVE_RE.findall(line):
+                out.setdefault(cve, cur)
+    return out
+
+
+_BLOB_RE = re.compile(r"^https://github\.com/([^/]+)/([^/]+)/blob/(.+)$")
+
+
+def _raw_if_blob(url: str) -> str:
+    """A GitHub *blob* URL is an HTML viewer, not the file.
+
+    The registry pointed ONLYOFFICE's CHANGELOG.md at its blob page. Fetching that returns 636KB of
+    markup in which the release headings do not survive HTML-stripping, so 24 CVEs were visible and
+    NONE attributable to a release — the same shape of blind spot as nginx. The raw URL attributes
+    all 24. Normalising here fixes every registry entry at once, present and future."""
+    m = _BLOB_RE.match(url)
+    return f"https://raw.githubusercontent.com/{m.group(1)}/{m.group(2)}/{m.group(3)}" if m else url
+
+
 def vendor_pages(urls: list[str]) -> list[dict]:
    """Fetch each registry URL and regex out CVE ids, with a little surrounding context."""
    out = []
@@ -269,20 +352,93 @@ def vendor_pages(urls: list[str]) -> list[dict]:
            # correct; counting them as failures would wrongly mark the recipe's count unreliable.
            out.append({"source": u, "status": "skipped: template URL (not fetchable)", "cves": [], "context": {}})
            continue
-        entry = {"source": u, "status": "ok", "cves": [], "context": {}}
+        entry = {"source": u, "status": "ok", "cves": [], "context": {}, "fixed_in": {}}
        try:
-            text = _fetch(u)
+            text = _fetch(_raw_if_blob(u))
            plain = re.sub(r"<[^>]+>", " ", text)
            for cve in sorted(set(CVE_RE.findall(plain))):
                entry["cves"].append(cve)
                i = plain.find(cve)
                entry["context"][cve] = re.sub(r"\s+", " ", plain[max(0, i - 160) : i + 200]).strip()
+            entry["fixed_in"] = _changelog_versions(text)
        except Exception as e:  # noqa: BLE001
            entry["status"] = f"error: {type(e).__name__}: {e}"
        out.append(entry)
    return out


+NVD_API = "https://services.nvd.nist.gov/rest/json/cves/2.0"
+NVD_CPE_RE = re.compile(r"^\s*[-*]?\s*nvd-cpe:\s*(\S+)\s*=\s*(cpe:2\.3:[^\s`]+)", re.M | re.I)
+
+
+def registry_cpes(recipe: str, registry_dir: str) -> list[tuple[str, str]]:
+    """[(image-key, cpe)] declared in the recipe's registry as `nvd-cpe: <key> = <cpe>`."""
+    path = os.path.join(registry_dir, f"{recipe}.md")
+    try:
+        return [(m.group(1), m.group(2)) for m in NVD_CPE_RE.finditer(open(path).read())]
+    except OSError:
+        return []
+
+
+def nvd_advisories(cpe: str, key: str) -> dict:
+    """CVEs for a CPE from NVD, with the version data the classifier needs.
+
+    THE FALLBACK FOR PROJECTS THAT PUBLISH NOTHING MACHINE-READABLE. mattermost's GitHub advisory
+    feed is empty and its security bulletins are client-side rendered; mumble publishes neither. Both
+    scanned as `?` — nothing measured — until here. NVD is CPE-indexed and carries structured ranges:
+
+      versionEndExcluding X  -> fixed in X exactly            (a patched version)
+      versionEndIncluding X  -> affected up to and INCLUDING X, fixed in some later release. The
+                                exact fix version is unknown, but the upgrade fixes it whenever it
+                                crosses X — recorded as `affected_max` and judged in the classifier.
+
+    NVD LAGS the vendor (it had neither gitea CVSS-9.8 RCE at publication), so this is a fallback,
+    never a replacement for 2a/2b. Unauthenticated calls are rate-limited to ~5/30s, hence the retry.
+    """
+    entry = {"source": f"nvd:{key}", "status": "ok", "advisories": []}
+    url = f"{NVD_API}?resultsPerPage=2000&virtualMatchString={urllib.parse.quote(cpe)}"
+    data = None
+    for attempt in range(3):
+        try:
+            data = json.loads(_fetch(url))
+            break
+        except Exception as e:  # noqa: BLE001
+            if attempt == 2:
+                entry["status"] = f"error: {type(e).__name__}"
+                return entry
+            time.sleep(8)
+    for v in (data or {}).get("vulnerabilities", []):
+        c = v.get("cve") or {}
+        cid = c.get("id")
+        if not cid:
+            continue
+        fixed, affected_max = set(), set()
+        for cfg in c.get("configurations", []):
+            for node in cfg.get("nodes", []):
+                for m in node.get("cpeMatch", []):
+                    if m.get("versionEndExcluding"):
+                        fixed.add(m["versionEndExcluding"])
+                    elif m.get("versionEndIncluding"):
+                        affected_max.add(m["versionEndIncluding"])
+        sev = None
+        for mk in ("cvssMetricV31", "cvssMetricV30", "cvssMetricV2"):
+            got = (c.get("metrics") or {}).get(mk) or []
+            if got:
+                sev = (got[0].get("cvssData") or {}).get("baseSeverity")
+                break
+        entry["advisories"].append({
+            "cve": cid, "ghsa": None, "severity": (sev or "").lower() or None,
+            "summary": next((d.get("value") for d in c.get("descriptions", [])
+                             if d.get("lang") == "en"), "")[:200],
+            "vulnerable_range": None,
+            "patched": "; ".join(sorted(fixed)) or None,
+            "affected_max": "; ".join(sorted(affected_max)) or None,
+            "url": f"https://nvd.nist.gov/vuln/detail/{cid}",
+            "published_at": c.get("published"), "description": None, "cvss": None,
+        })
+    return entry
+
+
 def osv(recipe: str, version: str | None) -> dict | None:
    pkg = OSV_PACKAGES.get(recipe)
    if not pkg or not version:
@@ -580,7 +736,8 @@ def scan(recipe: str, v_from: str | None, v_to: str | None, registry_dir: str,
        e = report["cves"].setdefault(cve, {"sources": [], "severity": None, "ghsa": None,
                                            "vulnerable_range": None, "patched": None,
                                            "context": None, "published_at": None,
-                                            "description": None, "url": None, "cvss": None})
+                                            "description": None, "url": None, "cvss": None,
+                                            "changelog_fixed_in": None, "affected_max": None})
        if src not in e["sources"]:
            e["sources"].append(src)
        for k, v in extra.items():
@@ -597,11 +754,22 @@ def scan(recipe: str, v_from: str | None, v_to: str | None, registry_dir: str,
                       context=a.get("summary"), published_at=a.get("published_at"),
                       description=a.get("description"), url=a.get("url"), cvss=a.get("cvss"))

+    for key, cpe in registry_cpes(recipe, registry_dir):
+        entry = nvd_advisories(cpe, key)
+        report["sources"].append({"source": entry["source"], "status": entry["status"],
+                                  "found": len(entry.get("advisories") or [])})
+        for a in entry.get("advisories", []):
+            record(a["cve"], entry["source"], severity=a.get("severity"),
+                   patched=a.get("patched"), affected_max=a.get("affected_max"),
+                   context=a.get("summary"), published_at=a.get("published_at"),
+                   url=a.get("url"))
+
    for entry in vendor_pages(urls):
        report["sources"].append({"source": entry["source"], "status": entry["status"],
                                  "found": len(entry.get("cves", []))})
        for cve in entry.get("cves", []):
-            record(cve, entry["source"], context=entry["context"].get(cve))
+            record(cve, entry["source"], context=entry["context"].get(cve),
+                   changelog_fixed_in=(entry.get("fixed_in") or {}).get(cve))

    for version in filter(None, (v_from, v_to)):
        o = osv(recipe, version)
@@ -636,17 +804,39 @@ def scan(recipe: str, v_from: str | None, v_to: str | None, registry_dir: str,
    #
    # A source with no window is not classified: its advisories are listed as unclassified so they
    # stay visible without inflating the count.
-    gh_sources = [x["source"] for x in report["sources"] if x["source"].startswith("github-advisories:")]
+    gh_sources = [x["source"] for x in report["sources"]
+                  if x["source"].startswith(("github-advisories:", "nvd:"))]
    primary = gh_sources[0] if (gh_sources and (v_from or v_to)) else None
    report["primary_source"] = primary

    windows = {}                                    # source name -> (from, to)
+    window_key = {}                                 # source name -> the image name it covers
    if primary:
        windows[primary] = (v_from, v_to)
+        window_key[primary] = primary.split("/")[-1]
+        # The app's window must also cover its NVD entry. NVD sources are keyed by IMAGE name
+        # (`mattermost-team-edition`) while the advisory feed is keyed by REPO (`mattermost/
+        # mattermost`), so without this the fallback source that exists precisely because the feed
+        # is empty would itself go unwindowed — and mumble/mattermost would still report nothing.
+        pname = primary.split("/")[-1].lower()
+        for src in gh_sources:
+            if src.startswith("nvd:") and src not in windows:
+                k = src.split(":", 1)[1].lower()
+                if pname in k or k in pname:
+                    windows[src] = (v_from, v_to)
+                    window_key[src] = k
    for key, wf, wt in (images or []):
        for src in gh_sources:
-            if key.lower() in src.lower() and src not in windows:
+            if src in windows:
+                continue
+            # Match BOTH ways: an image name is often longer than its source repo
+            # (`clickhouse/clickhouse-server` vs source `ClickHouse/ClickHouse`) and sometimes
+            # shorter (`redis` vs `redis/redis`). One-directional matching silently dropped the
+            # clickhouse window when the key was derived from a compose file.
+            k, name = key.lower(), src.split("/")[-1].lower()
+            if k in src.lower() or name in k:
                windows[src] = (wf, wt)
+                window_key[src] = key
    report["windows"] = {k: {"from": f, "to": t} for k, (f, t) in windows.items()}

    def _classify_window(src, wf, wt):
@@ -665,8 +855,31 @@ def scan(recipe: str, v_from: str | None, v_to: str | None, registry_dir: str,
                    continue
                patched = e.get("patched") or ""
                cands = [_vkey(t) for t in re.findall(r"\d+(?:\.\d+)*", patched)]
+                # NEVER on a placeholder: "7.4.X" parses to the bare 7.4, which then reads as
+                # "already fixed at 7.4" and silently drops a real fix (redis CVE-2024-46981).
+                # A placeholder means the fix version is unknown — that is the indeterminate path.
+                if (kf and kt and not PLACEHOLDER_RE.search(patched)
+                        and _already_fixed_on_from_line(kf, cands)):
+                    # already had it before the upgrade
+                    e.setdefault("classification", "outside-window")
+                    continue
+                if kf and kt and _superseded_on_target_line(kt, cands):
+                    # The target's own line got the fix LATER than the target: not fixed here.
+                    e.setdefault("classification", "outside-window")
+                    continue
                if kf and kt and any(_within(kf, kt, c) for c in cands):
                    got.add(cve)
+                elif kf and kt and e.get("affected_max"):
+                    # NVD's `versionEndIncluding X`: affected up to and INCLUDING X, fixed in some
+                    # later release. The exact fix version is unpublished, but the upgrade delivers
+                    # it whenever it crosses X — i.e. from <= X < to.
+                    for t in re.findall(r"\d+(?:\.\d+)*", e["affected_max"]):
+                        x = _vkey(t)
+                        n = max(len(kf), len(kt), len(x))
+                        pad = lambda z: z + (0,) * (n - len(z))
+                        if x and pad(kf) <= pad(x) < pad(kt):
+                            got.add(cve)
+                            break
                elif not patched or PLACEHOLDER_RE.search(patched):
                    # No fix version published ("TBD") or only a placeholder ("7.4.X" — which could
                    # be 7.4.1, inside the window). We cannot say either way, so say so.
@@ -700,6 +913,32 @@ def scan(recipe: str, v_from: str | None, v_to: str | None, registry_dir: str,
            report["cves"][cve]["classification"] = f"fixed-by-this-upgrade ({method}) via {src}"
            fixed_set.add(cve)

+    # A CVE seen only in a vendor CHANGELOG has no advisory feed behind it, but the changelog says
+    # which release fixed it (see _changelog_versions). Tie it to a window by the image name
+    # appearing in the page URL — nginx's window is `nginx`, and its changelog is nginx.org/... .
+    # Without this, projects that publish no GitHub advisories (nginx being the big one) can never
+    # contribute a CVE, and every nginx bump in the fleet silently reports 0.
+    from_changelog = {}
+    for cve, e in report["cves"].items():
+        if cve in fixed_set or not e.get("changelog_fixed_in"):
+            continue
+        for src, (wf, wt) in windows.items():
+            key = (window_key.get(src) or "").lower()
+            if not key:
+                continue
+            if not any(key in s_.lower() for s_ in e["sources"] if s_.startswith("http")):
+                continue
+            kf, kt = _vkey(wf), _vkey(wt)
+            cand = _vkey(e["changelog_fixed_in"])
+            if kf and kt and cand and _within(kf, kt, cand):
+                e["classification"] = (f"fixed-by-this-upgrade (named under {e['changelog_fixed_in']} "
+                                       f"in the vendor changelog) via {src}")
+                fixed_set.add(cve)
+                from_changelog[cve] = e["changelog_fixed_in"]
+            break
+    if from_changelog:
+        report["resolved_by_changelog"] = from_changelog
+
    unknown = []
    for cve, e in report["cves"].items():
        if cve in fixed_set:
@@ -778,8 +1017,19 @@ def scan(recipe: str, v_from: str | None, v_to: str | None, registry_dir: str,
    report["unclassified"] = sorted(unknown)
    # NEVER report 0 for something we could not determine — a 0 asserts safety. If ANY requested
    # window could not be ordered at all, the total is UNKNOWN rather than a partial number.
-    report["count_known"] = not unresolved_any
-    report["cve_count_fixed"] = len(fixed_set) if not unresolved_any else None
+    # A scan with NO usable source has not measured anything, so it must not report a number —
+    # least of all 0, which asserts safety. mumble had no cc-ci-plan/upstream/mumble.md at all and
+    # still produced "0 identified", which was then published as a clean 0 in a CVE report.
+    usable_sources = [
+        s for s in report["sources"]
+        if s["status"] == "ok" or s["status"].startswith("no-advisories-published")
+    ]
+    no_sources = not usable_sources
+    if no_sources:
+        report["no_usable_sources"] = True
+    report["count_known"] = not unresolved_any and not no_sources
+    report["cve_count_fixed"] = (len(fixed_set)
+                                 if (not unresolved_any and not no_sources) else None)
    report["cve_count_total_seen"] = len(report["cves"])
    # Only GENUINE failures make a count unreliable. "no-advisories-published" (404: the repo has
    # no advisory feed) and "skipped: template URL" are benign and must not degrade the verdict.
@@ -802,10 +1052,16 @@ def markdown(rep: dict) -> str:
         f"{rep.get('from') or '?'} → {rep.get('to') or '?'}"]
    if not rep.get("count_known", True):
        L.append("\n**CVEs fixed by this upgrade: UNKNOWN — the scan could NOT determine a count.**")
-        L.append("\n⚠ This is NOT zero. A version-scheme change (e.g. semver → calver) makes numeric "
-                 "ordering meaningless across this jump, so no advisory could be classified. Render "
-                 "this recipe's cve cell as `?`, never `0`. Read the vendor's release notes for the "
-                 "jump and count by hand.")
+        if rep.get("no_usable_sources"):
+            L.append("\n⚠ This is NOT zero. **No usable source was checked at all** — the registry "
+                     "file `cc-ci-plan/upstream/<recipe>.md` is missing or every source failed, so "
+                     "nothing was measured. Render this recipe's cve cell as `?`, never `0`, and add "
+                     "the registry file.")
+        else:
+            L.append("\n⚠ This is NOT zero. A version-scheme change (e.g. semver → calver) makes "
+                     "numeric ordering meaningless across this jump, so no advisory could be "
+                     "classified. Render this recipe's cve cell as `?`, never `0`. Read the vendor's "
+                     "release notes for the jump and count by hand.")
        if rep["unclassified"]:
            L.append(f"\nAdvisories seen but unclassifiable ({len(rep['unclassified'])}) — includes "
                     f"other images in this recipe: " + ", ".join(rep["unclassified"][:12]))
@@ -857,6 +1113,118 @@ def markdown(rep: dict) -> str:
    return "\n".join(L)


+def _gitea_auth(url: str) -> dict:
+    """Basic auth for the private mirror, using public config plus .testenv.
+
+    Sent as a HEADER, never embedded in the URL: in-URL credentials leak into shell history, process
+    lists and error messages, and urllib mis-parses a password containing a colon."""
+    host = re.sub(r"^https?://", "", url).split("/")[0]
+    env = {}
+    for path in (
+        "/srv/cc-ci/cc-ci/.env.public",
+        os.environ.get("CCCI_TESTENV", "/srv/cc-ci/.testenv"),
+    ):
+        try:
+            lines = open(path)
+        except OSError:
+            continue
+        for ln in lines:
+            if "=" in ln and not ln.strip().startswith("#"):
+                k, v = ln.strip().split("=", 1)
+                env[k] = v.strip().strip("\"'")
+    if host != env.get("GITEA_URL", "git.autonomic.zone"):
+        return {}
+    u, pw = env.get("GITEA_USERNAME"), env.get("GITEA_PASSWORD")
+    if not (u and pw):
+        return {}
+    import base64 as _b64
+    return {"Authorization": "Basic " + _b64.b64encode(f"{u}:{pw}".encode()).decode()}
+
+
+def _compose_images(url: str) -> dict[str, tuple[str, str]]:
+    """{service: (image-repo, tag)} for a compose file.
+
+    Keyed by SERVICE, not by image repo, because an upgrade may change the repo itself: plausible
+    moved `plausible/analytics` -> `ghcr.io/plausible/community-edition`. Keyed by repo that reads
+    as one image vanishing and an unrelated one appearing, and the app's version window is lost —
+    which is exactly the upgrade most worth scanning."""
+    txt = _fetch(url, _gitea_auth(url))
+    out, svc = {}, None
+    in_services = False
+    for line in txt.splitlines():
+        if re.match(r"^services:\s*$", line):
+            in_services = True
+            continue
+        if in_services and re.match(r"^\S", line):
+            in_services = False
+        if not in_services:
+            continue
+        m = re.match(r"^  (\S+):\s*$", line)
+        if m:
+            svc = m.group(1)
+            continue
+        m = re.match(r"^\s+image:\s*[\"']?([^\"'\s]+)", line)
+        if m and svc:
+            ref = m.group(1).split("@", 1)[0]
+            if "${" in ref or "$(" in ref:
+                continue
+            repo, _, tag = ref.rpartition(":")
+            if repo and tag:
+                out[svc] = (repo, tag)
+    return out
+
+
+def _default_branch_compose(url: str) -> str | None:
+    """Same repo as `url`, but its DEFAULT branch — resolved from the API, never assumed.
+
+    Several coopcloud recipes keep a stale `main` beside the real default `master` (gitea's `main`
+    is 1.24.2-rootless while `master` has 1.27.1-rootless), so guessing the branch produces a
+    confidently wrong baseline."""
+    m = re.match(r"(https?://[^/]+)/([^/]+)/([^/]+)/(?:raw|src)/branch/[^/]+/(.*)$", url)
+    if not m:
+        return None
+    host, owner, repo, path = m.groups()
+    try:
+        meta = json.loads(_fetch(f"{host}/api/v1/repos/{owner}/{repo}", _gitea_auth(host)))
+        br = meta.get("default_branch")
+    except Exception:  # noqa: BLE001
+        return None
+    return f"{host}/{owner}/{repo}/raw/branch/{br}/{path}" if br else None
+
+
+def windows_from_compose(to_url: str, from_url: str | None = None) -> tuple[list, str | None]:
+    """Derive the scan's version windows by DIFFING two compose files.
+
+    This is the deterministic alternative to a human (or a model) deciding which `--image` args a
+    given upgrade needs. Point it at a PR's compose and it reads the windows straight off the diff:
+    every image whose tag changed becomes a window, every image that did not change is correctly
+    left out, and nothing depends on anyone remembering that the recipe also bumped its redis.
+
+    Returns (windows, note) where windows is [(image-name, from, to)].
+    """
+    if from_url is None:
+        from_url = _default_branch_compose(to_url)
+        if not from_url:
+            raise SystemExit("could not resolve a baseline compose; pass --compose-from explicitly")
+    new, old = _compose_images(to_url), _compose_images(from_url)
+    app, others = None, []
+    for svc, (repo, tag) in sorted(new.items()):
+        if svc not in old:
+            continue
+        prev_repo, prev_tag = old[svc]
+        if prev_tag == tag and prev_repo == repo:
+            continue
+        # The `app` service is the recipe's primary image by coop-cloud convention; its window drives
+        # --from/--to so the scan's primary advisory source is judged against it. Everything else is
+        # a sidecar window keyed by its image name.
+        if svc == "app":
+            app = (repo.split("/")[-1], prev_tag, tag)
+        else:
+            others.append((repo.split("/")[-1], prev_tag, tag))
+    wins = ([app] if app else []) + others
+    return wins, f"baseline {from_url}"
+
+
 def main() -> int:
    ap = argparse.ArgumentParser(description=__doc__, formatter_class=argparse.RawDescriptionHelpFormatter)
    ap.add_argument("recipe")
@@ -868,6 +1236,13 @@ def main() -> int:
                         "fix version published), fetch their full text + references and append a "
                         "block for the agent to judge. Additive: it never changes the count above.")
    ap.add_argument("--registry", default=REGISTRY_DIR)
+    ap.add_argument("--compose-to", default=None, metavar="URL",
+                    help="derive the windows by DIFFING this compose against its baseline, instead "
+                         "of passing --from/--to/--image by hand. Point it at a PR's compose.yml "
+                         "(e.g. .../raw/branch/<pr-branch>/compose.yml).")
+    ap.add_argument("--compose-from", default=None, metavar="URL",
+                    help="baseline compose for --compose-to. Default: the same repo's DEFAULT "
+                         "branch, resolved from the API (never assumed to be `main`).")
    ap.add_argument("--image", action="append", default=[], metavar="NAME=FROM:TO",
                    help="a sidecar image and the versions it moved between, e.g. "
                         "--image redis=7.4:8.10 (repeatable). NAME matches a source repo name; "
@@ -875,6 +1250,20 @@ def main() -> int:
                         "being left unclassified.")
    a = ap.parse_args()
    images = []
+    if a.compose_to:
+        wins, note = windows_from_compose(a.compose_to, a.compose_from)
+        if not wins:
+            print(f"### Advisory scan — {a.recipe}\n\n**No image versions changed between the two "
+                  f"compose files, so this upgrade fixes no CVEs by definition.**\n\n_{note}_")
+            return 0
+        print(f"_derived from compose diff ({note}):_", file=sys.stderr)
+        for n_, f_, t_ in wins:
+            print(f"_  {n_}: {f_} → {t_}_", file=sys.stderr)
+        # The `app` service (first entry when present) drives --from/--to; the rest are --image
+        # windows. Passing every window as --image too is harmless: each is matched by name against
+        # the advisory sources, and an unmatched name is simply ignored.
+        a.v_from, a.v_to = a.v_from or wins[0][1], a.v_to or wins[0][2]
+        images = list(wins[1:])
    for spec in a.image:
        name, _, rng = spec.partition('=')
        vf, _, vt = rng.partition(':')
@@ -34,7 +34,7 @@ supports_resume = false
 prompt_delivery = "ping"       # send after the TUI connects
 connect_delay   = 12
 submit_key      = "C-m"
-preamble        = "set -a; . /srv/cc-ci/.testenv; set +a"
+preamble        = "set -a; . /srv/cc-ci/cc-ci-plan/load-env.sh; set +a"
 stall_idle      = 900
 active_re = "esc interrupt|thinking|inferring|running tool|tool call|preparing patch|reading|searching"
 limit_re  = "spend limit|usage limit|limit reached|reached your .*limit|out of (credits|tokens)"
@@ -0,0 +1,314 @@
+#!/usr/bin/env python3
+"""audit-sources — are we still looking in the right place for each recipe's updates?
+
+A recipe tracks an image repo and a set of registry URLs. Upstreams move: they rename the image,
+switch registry, archive the GitHub repo, or split a community edition out of the original. When that
+happens nothing errors — the old repo simply stops receiving tags, and the recipe looks "up to date"
+forever while real releases happen somewhere else.
+
+plausible is the worked example. It tracked `plausible/analytics` on Docker Hub; upstream moved to
+`ghcr.io/plausible/community-edition`. The old repo still exists and still serves v2.0.0, so every
+survey said "no upgrades available" while v3 shipped elsewhere.
+
+This reports the signals that catch that, per image and per registry URL:
+
+  * IMAGE GONE QUIET     — newest tag is older than --quiet-days (default 365). The single strongest
+                           signal that releases moved somewhere else.
+  * DEPRECATION WORDING  — the registry description says deprecated / moved / no longer maintained.
+  * GITHUB REPO ARCHIVED — upstream archived it.
+  * GITHUB REPO RENAMED  — the API redirects to a different owner/name than we ask for.
+  * GITHUB REPO GONE     — 404.
+
+Everything is a SIGNAL, not a verdict: a genuinely stable image (mumble, custom-html) can be quiet
+for good reason. The output is for a human to judge, so each finding says what was measured.
+
+    audit-sources.py [recipe ...] [--ssh HOST] [--quiet-days N] [--json]
+"""
+
+from __future__ import annotations
+
+import argparse
+import importlib.util
+import json
+import os
+import re
+import sys
+import urllib.error
+import urllib.request
+from datetime import datetime, timezone
+
+HERE = os.path.dirname(os.path.abspath(__file__))
+_spec = importlib.util.spec_from_file_location("resolve_images", os.path.join(HERE, "resolve-images.py"))
+RI = importlib.util.module_from_spec(_spec)
+_spec.loader.exec_module(RI)
+
+# advisory-scan supplies the source-fetching + changelog-attribution used by --security-sources
+_aspec = importlib.util.spec_from_file_location("advisory_scan", os.path.join(HERE, "advisory-scan.py"))
+A = importlib.util.module_from_spec(_aspec)
+_aspec.loader.exec_module(A)
+
+REGISTRY_DIR = os.environ.get("CCCI_UPSTREAM_REGISTRY", os.path.join(HERE, "upstream"))
+USED_RECIPES = os.path.join(HERE, "used-recipes.md")
+DEPRECATION_RE = re.compile(
+    r"\b(deprecat|no longer maintain|unmaintained|superseded|moved to|migrated to|"
+    r"has moved|discontinued|end.of.life|archived)\b", re.I)
+
+
+def _days_since(iso: str | None) -> int | None:
+    if not iso:
+        return None
+    try:
+        d = datetime.fromisoformat(iso.replace("Z", "+00:00"))
+    except ValueError:
+        return None
+    return (datetime.now(timezone.utc) - d).days
+
+
+def hub_repo_meta(repo: str) -> dict:
+    """Docker Hub repo metadata: when it was last pushed to, and how it describes itself."""
+    try:
+        d = RI._json(f"https://hub.docker.com/v2/repositories/{repo}", RI._hub_auth())
+    except urllib.error.HTTPError as e:
+        return {"status": f"HTTP {e.code}"}
+    except Exception as e:  # noqa: BLE001
+        return {"status": f"{type(e).__name__}"}
+    text = f"{d.get('description') or ''}\n{d.get('full_description') or ''}"
+    m = DEPRECATION_RE.search(text)
+    return {"status": "ok", "last_updated": d.get("last_updated"),
+            "deprecation_hint": (m.group(0) if m else None),
+            "archived": bool(d.get("is_archived") or d.get("status") == "inactive")}
+
+
+def github_repo_meta(owner: str, repo: str) -> dict:
+    """GitHub repo state — archived, renamed (the API answers with the CURRENT full_name), or gone."""
+    hdrs = {"Accept": "application/vnd.github+json"}
+    tok = RI._gh_token()
+    if tok:
+        hdrs["Authorization"] = f"Bearer {tok}"
+    try:
+        d = RI._json(f"https://api.github.com/repos/{owner}/{repo}", hdrs)
+    except urllib.error.HTTPError as e:
+        return {"status": f"HTTP {e.code}"}
+    except Exception as e:  # noqa: BLE001
+        return {"status": f"{type(e).__name__}"}
+    asked, got = f"{owner}/{repo}".lower(), (d.get("full_name") or "").lower()
+    return {"status": "ok", "archived": bool(d.get("archived")), "pushed_at": d.get("pushed_at"),
+            "renamed_to": (d.get("full_name") if got and got != asked else None),
+            "description": d.get("description") or ""}
+
+
+def newest_tag_date(registry: str, repo: str, tag: str) -> str | None:
+    """When was the repo's newest same-shape tag pushed? Docker Hub only (it dates its tags)."""
+    if registry not in ("docker.io", "registry-1.docker.io"):
+        return None
+    try:
+        d = RI._json(f"https://hub.docker.com/v2/repositories/{repo}/tags"
+                     f"?page_size=100&ordering=last_updated", RI._hub_auth())
+    except Exception:  # noqa: BLE001
+        return None
+    want = RI.shape(tag)
+    for row in d.get("results", []):
+        if RI.shape(row.get("name") or "") == want:
+            return row.get("last_updated")
+    return (d.get("results") or [{}])[0].get("last_updated")
+
+
+def security_source_audit(recipe: str) -> list[dict]:
+    """Per source: are its CVEs USABLE, or merely visible?
+
+    The nginx lesson. nginx publishes no GitHub advisories; all its CVEs live in nginx.org/en/CHANGES.
+    The scan saw them and could do nothing with them, because nothing said which release fixed which
+    CVE — so every nginx bump in the fleet reported 0. Attribution (advisory-scan §2b) fixed that for
+    changelogs organised by release, but a page that lists CVEs with NO release structure is still a
+    blind spot: visible, uncountable. This finds those.
+
+    Per source: `advisory-feed` (structured, best), `changelog` (CVEs attributable to a release),
+    `unattributable` (CVEs present but no release structure — BLIND), or `no-cve-data`.
+    """
+    urls, _ = _registry_urls(recipe)
+    out = []
+    # NVD CPE entries are a first-class source: for projects publishing nothing machine-readable
+    # (mattermost, mumble) they are the ONLY structured source, and omitting them here made two
+    # recipes look permanently blind after they had been fixed.
+    for key, cpe in A.registry_cpes(recipe, REGISTRY_DIR):
+        e = A.nvd_advisories(cpe, key)
+        n = len(e.get("advisories") or [])
+        out.append({"source": e["source"] + f"  ({cpe.split(':')[4]}/{cpe.split(':')[3]})",
+                    "kind": "advisory-feed" if n else "no-cve-data",
+                    "status": e["status"], "cves": n, "usable": n})
+    for entry in A.github_advisories(urls):
+        out.append({"source": entry["source"], "kind": "advisory-feed",
+                    "status": entry["status"], "cves": len(entry.get("advisories") or []),
+                    "usable": len(entry.get("advisories") or [])})
+    for entry in A.vendor_pages(urls):
+        if entry["status"].startswith("skipped"):
+            continue
+        n = len(entry.get("cves") or [])
+        attributed = len(entry.get("fixed_in") or {})
+        kind = ("no-cve-data" if n == 0 else
+                "changelog" if attributed else "unattributable")
+        out.append({"source": entry["source"], "kind": kind, "status": entry["status"],
+                    "cves": n, "usable": attributed})
+    return out
+
+
+def audit_recipe(recipe: str, ssh: str | None, quiet_days: int) -> dict:
+    out = {"recipe": recipe, "findings": [], "images": [], "sources": []}
+    try:
+        refs = (RI.compose_images_ssh(recipe, ssh, "~/.abra/recipes") if ssh
+                else RI.compose_images(recipe, RI.RECIPE_DIR))
+    except Exception as e:  # noqa: BLE001
+        out["findings"].append({"level": "error", "what": f"could not read compose: {e}"})
+        return out
+
+    for ref in refs:
+        if "${" in ref:
+            continue
+        info = RI.parse_ref(ref)
+        row = {"ref": ref, "registry": info["registry"], "repo": info["repo"], "tag": info["tag"]}
+        if info["registry"] in ("docker.io", "registry-1.docker.io"):
+            meta = hub_repo_meta(info["repo"])
+            row.update(meta)
+            newest = newest_tag_date(info["registry"], info["repo"], info["tag"])
+            row["newest_tag_pushed"] = newest
+            age = _days_since(newest)
+            row["newest_tag_age_days"] = age
+            if age is not None and age > quiet_days:
+                out["findings"].append({
+                    "level": "warn", "what": "image has gone quiet",
+                    "detail": f"{info['repo']}: newest {RI.shape(info['tag'])}-shaped tag pushed "
+                              f"{age} days ago — releases may have moved elsewhere"})
+            if meta.get("deprecation_hint"):
+                out["findings"].append({
+                    "level": "warn", "what": "registry text suggests deprecation",
+                    "detail": f"{info['repo']}: says {meta['deprecation_hint']!r}"})
+            if meta.get("archived"):
+                out["findings"].append({"level": "warn", "what": "registry repo archived/inactive",
+                                        "detail": info["repo"]})
+        out["images"].append(row)
+
+    urls, reg_path = ([], None)
+    try:
+        urls, reg_path = _registry_urls(recipe)
+    except Exception:  # noqa: BLE001
+        pass
+    if reg_path is None:
+        out["findings"].append({"level": "warn", "what": "no upstream registry file",
+                                "detail": f"cc-ci-plan/upstream/{recipe}.md is missing — the advisory "
+                                          f"scan has nowhere to look"})
+    seen = set()
+    for u in urls:
+        m = re.match(r"https?://github\.com/([^/]+)/([^/#?]+)", u)
+        if not m:
+            continue
+        owner, repo = m.group(1), m.group(2).removesuffix(".git")
+        if (owner, repo) in seen:
+            continue
+        seen.add((owner, repo))
+        meta = github_repo_meta(owner, repo)
+        row = {"repo": f"{owner}/{repo}", **meta}
+        age = _days_since(meta.get("pushed_at"))
+        row["pushed_age_days"] = age
+        out["sources"].append(row)
+        if meta.get("status") != "ok":
+            out["findings"].append({"level": "warn", "what": "registry source unreachable",
+                                    "detail": f"{owner}/{repo}: {meta['status']}"})
+            continue
+        if meta.get("renamed_to"):
+            out["findings"].append({"level": "alert", "what": "GitHub repo has MOVED",
+                                    "detail": f"{owner}/{repo} now answers as {meta['renamed_to']}"})
+        if meta.get("archived"):
+            out["findings"].append({"level": "alert", "what": "GitHub repo is ARCHIVED",
+                                    "detail": f"{owner}/{repo} — upstream development has stopped here"})
+        if age is not None and age > quiet_days:
+            out["findings"].append({"level": "warn", "what": "GitHub repo quiet",
+                                    "detail": f"{owner}/{repo}: last push {age} days ago"})
+    return out
+
+
+def _registry_urls(recipe: str):
+    path = os.path.join(REGISTRY_DIR, f"{recipe}.md")
+    if not os.path.exists(path):
+        return [], None
+    text = open(path).read()
+    urls = []
+    for u in re.findall(r"https?://[^\s)|\]]+", text):
+        u = u.rstrip("`'\"*.,;:>)")
+        if u and u not in urls:
+            urls.append(u)
+    return urls, path
+
+
+def all_recipes() -> list[str]:
+    out = []
+    for ln in open(USED_RECIPES):
+        ln = ln.strip()
+        if not ln or ln.startswith("#") or ln.startswith("`"):
+            continue
+        parts = ln.split()
+        if len(parts) >= 2 and parts[1] in ("weekly", "external"):
+            out.append(parts[0])
+    return out
+
+
+def main() -> int:
+    ap = argparse.ArgumentParser(description=__doc__,
+                                 formatter_class=argparse.RawDescriptionHelpFormatter)
+    ap.add_argument("recipes", nargs="*")
+    ap.add_argument("--ssh", default=None)
+    ap.add_argument("--quiet-days", type=int, default=365)
+    ap.add_argument("--json", action="store_true")
+    ap.add_argument("--security-sources", action="store_true",
+                    help="audit whether each recipe's CVE sources are USABLE (structured advisory "
+                         "feed / release-attributable changelog) or merely visible")
+    a = ap.parse_args()
+
+    recipes = a.recipes or all_recipes()
+    if a.security_sources:
+        # What matters is whether the RECIPE can see CVEs at all — not whether some individual page
+        # is unparseable. A page with no release structure is harmless when the same project also
+        # publishes an advisory feed (redis, gitea, minio, clickhouse all do); it is only a blind
+        # spot when nothing else covers that project.
+        blind_recipes, noisy = [], 0
+        for r in recipes:
+            rows = security_source_audit(r)
+            feeds = [x for x in rows if x["kind"] == "advisory-feed" and x["cves"] > 0]
+            logs = [x for x in rows if x["kind"] == "changelog"]
+            unattr = [x for x in rows if x["kind"] == "unattributable"]
+            noisy += len(unattr)
+            usable = len(feeds) + len(logs)
+            if usable == 0:
+                blind_recipes.append(r)
+                print(f"!! {r}: NO USABLE CVE SOURCE — {len(unattr)} unparseable page(s), "
+                      f"0 advisory feeds, 0 attributable changelogs")
+                for x in rows:
+                    print(f"      {x['kind']:15} {x['source'][:64]} ({x['cves']} CVEs)")
+            else:
+                print(f"OK {r}: {len(feeds)} advisory-feed(s), {len(logs)} changelog(s)"
+                      + (f", {len(unattr)} unparseable page(s) (redundant — covered by a feed)"
+                         if unattr else ""))
+                for x in logs:
+                    print(f"      changelog  {x['source'][:62]} ({x['usable']}/{x['cves']})")
+        print(f"\n{len(recipes)} recipes · {len(blind_recipes)} with NO usable CVE source"
+              + (f": {', '.join(blind_recipes)}" if blind_recipes else "")
+              + f" · {noisy} unparseable page(s) elsewhere (harmless where a feed covers them)")
+        return 0
+    reports = [audit_recipe(r, a.ssh, a.quiet_days) for r in recipes]
+    if a.json:
+        print(json.dumps(reports, indent=2))
+        return 0
+    alerts = 0
+    for rep in reports:
+        fs = rep["findings"]
+        mark = "OK " if not fs else ("!! " if any(f["level"] == "alert" for f in fs) else " ? ")
+        print(f"{mark} {rep['recipe']}")
+        for f in fs:
+            alerts += f["level"] == "alert"
+            print(f"      [{f['level']}] {f['what']}: {f.get('detail','')}")
+    print(f"\n{len(reports)} recipes audited · "
+          f"{sum(len(r['findings']) for r in reports)} findings · {alerts} alerts")
+    return 0
+
+
+if __name__ == "__main__":
+    sys.exit(main())
@@ -112,7 +112,7 @@ def start(mode="resume"):
        if not Path(OPENCODE_BIN).exists():
            die(f"opencode not found at {OPENCODE_BIN}")
        cmd = (
-            f"set -a; . /srv/cc-ci/.testenv; set +a; "
+            f"set -a; . /srv/cc-ci/cc-ci-plan/load-env.sh; set +a; "
            f"NO_COLOR=1 {OPENCODE_BIN} attach {OPENCODE_SERVER} --dir {WORKDIR}"
        )
        log(f"starting {SESSION} (backend=opencode, model={LOOP_MODEL or 'default'})")
@@ -154,7 +154,7 @@ def start(mode="resume"):
                f"OPENCODE_CONFIG_CONTENT={shlex.quote(json.dumps({'model': LOOP_MODEL}))} "
            )
        cmd = (
-            f"set -a; . /srv/cc-ci/.testenv; set +a; "
+            f"set -a; . /srv/cc-ci/cc-ci-plan/load-env.sh; set +a; "
            f"{model_env}NO_COLOR=1 {OPENCODE_BIN} attach {OPENCODE_SERVER} --dir {shlex.quote(WORKDIR)}"
        )
        log(f"starting {SESSION} (backend=opencode, model={LOOP_MODEL or 'default'})")
@@ -109,7 +109,7 @@ def start(mode, date):
        cwd = "/srv/cc-ci-orch/cc-ci"
        share_flag = "--share" if OPENCODE_SHARE else ""
        cmd = (
-            f"set -a; . /srv/cc-ci/.testenv; set +a; "
+            f"set -a; . /srv/cc-ci/cc-ci-plan/load-env.sh; set +a; "
            f"{OPENCODE_BIN} run {model_flag} {share_flag} --attach '{OPENCODE_SERVER}' "
            f"--title '{SESSION}' --dir {cwd} \"$(cat '{kf}')\""
        )
@@ -107,7 +107,7 @@ def spawn_supervisor(sid, reason):
    kf = Path(LOG_DIR) / f".kickoff-{SUP_SESSION}.txt"
    kf.write_text(build_kickoff(sid, reason))
    share = "--share" if OPENCODE_SHARE else ""
-    cmd = (f"set -a; . {WORKDIR}/.testenv; set +a; {OPENCODE_BIN} run --model '{MODEL}' {share} "
+    cmd = (f"set -a; . /srv/cc-ci/cc-ci-plan/load-env.sh; set +a; {OPENCODE_BIN} run --model '{MODEL}' {share} "
           f"--attach '{OPENCODE_SERVER}' --title '{SUP_SESSION}' --dir {WORKDIR} \"$(cat '{kf}')\"")
    _sh(["tmux", "new-session", "-d", "-s", SUP_SESSION, "-c", WORKDIR, cmd])
    _sh(["tmux", "pipe-pane", "-o", "-t", SUP_SESSION, f"cat >> '{LOG_DIR}/{SUP_SESSION}.log'"])
@@ -216,7 +216,7 @@ def start(mode="use-or-create"):
        # so they must come AFTER `run` (a global `opencode --model X run` is ignored).
        share_flag = "--share" if OPENCODE_SHARE else ""
        cmd = (
-            f"set -a; . /srv/cc-ci/.testenv; set +a; "
+            f"set -a; . /srv/cc-ci/cc-ci-plan/load-env.sh; set +a; "
            f"{OPENCODE_BIN} run {model_flag} {share_flag} --attach '{OPENCODE_SERVER}' "
            f"--title '{SESSION}' \"$(cat '{kf}')\""
        )
@@ -521,7 +521,7 @@ def resume(reason="manual"):
            "write+push the weekly summary, then `python3 /srv/cc-ci/cc-ci-plan/launch-report.py fresh`, print "
            "'" + DONE_MARKER + "', and go idle.")
    share = "--share" if OPENCODE_SHARE else ""
-    cmd = (f"set -a; . /srv/cc-ci/.testenv; set +a; {OPENCODE_BIN} run -s {sid} --continue "
+    cmd = (f"set -a; . /srv/cc-ci/cc-ci-plan/load-env.sh; set +a; {OPENCODE_BIN} run -s {sid} --continue "
           f"--model '{MODEL}' {share} --attach '{OPENCODE_SERVER}' --dir '{WORKDIR}' \"$(cat '{kf}')\"")
    subprocess.run(["tmux", "new-session", "-d", "-s", SESSION, "-c", WORKDIR, cmd])
    subprocess.run(["tmux", "pipe-pane", "-o", "-t", SESSION, f"cat >> '{LOG_FILE}'"])
@@ -323,7 +323,7 @@ def start_agent(role, session, workdir):
        # absolute repo paths for builder/adversary work.
        session_cwd = "/srv/cc-ci-orch/cc-ci"
        cmd = (
-            f"set -a; . /srv/cc-ci/.testenv; set +a; "
+            f"set -a; . /srv/cc-ci/cc-ci-plan/load-env.sh; set +a; "
            f"NO_COLOR=1 {OPENCODE_BIN} attach {OPENCODE_SERVER} --dir {session_cwd}"
        )
        log(f"starting {session} (backend=opencode, phase={pid}, model={model or 'default'})")
@@ -0,0 +1,17 @@
+#!/usr/bin/env bash
+# Load cc-ci runtime configuration. Source this file; do not execute it.
+#
+# Callers enable `set -a` so both files become exported environment variables.
+# Public configuration is loaded first, then the secret file. This preserves a
+# local secret override only where an emergency migration explicitly needs it.
+
+if [ -r /srv/cc-ci/cc-ci/.env.public ]; then
+  . /srv/cc-ci/cc-ci/.env.public
+fi
+
+if [ ! -r /srv/cc-ci/.testenv ]; then
+  echo "cc-ci secret environment is missing: /srv/cc-ci/.testenv" >&2
+  return 1
+fi
+
+. /srv/cc-ci/.testenv
@@ -71,7 +71,7 @@ log "writing report $REPORT ..."
  if [ -f "$SUMMARY" ]; then cat "$SUMMARY"; else echo "(no dated summary at $SUMMARY — /upgrade-all may have stalled on the usage limit or errored; check the cc-ci-upgrader session + $LOG)"; fi
  echo
  echo "## Open PRs per recipe (post-run)"
-  set -a; . /srv/cc-ci/.testenv 2>/dev/null; set +a
+  set -a; . /srv/cc-ci/cc-ci-plan/load-env.sh; set +a
  G="https://${GITEA_USERNAME}:${GITEA_PASSWORD}@${GITEA_URL}/api/v1"
  for r in $(curl -fsS "$G/orgs/recipe-maintainers/repos?limit=100" 2>/dev/null | python3 -c "import sys,json;print('\n'.join(sorted(x['name'] for x in json.load(sys.stdin))))" 2>/dev/null); do
    case "$r" in cc-ci|cc-ci-orchestrator|cc-ci-secrets|archived-*) continue;; esac
@@ -0,0 +1,102 @@
+# Plan — one Hetzner host for cc-ci (CI server) + cc-ci-orchestrator
+
+**Status:** IN PROGRESS (started 2026-09-07). Operator request: move the cc-ci CI server AND the
+cc-ci orchestrator onto one new Hetzner server (`195.201.88.249`, 8 GB, 150 GB, Debian 13 image),
+cleanly split off from the shared `notplants-orchestrator` box, which keeps everything else
+(lichen, project-orchestrator, notplants agents). Nix config ownership: cc-ci's config in
+`recipe-maintainers/cc-ci`, the orchestrator's in `recipe-maintainers/cc-ci-orchestrator`, and the
+orchestrator flake **includes** cc-ci's module so one `nixos-rebuild` produces the combined host.
+Last step (separate, after everything works on the current names): move both to `autonomic.zone`
+subdomains.
+
+## Facts (2026-09-07)
+
+| | old cc-ci server | old orchestrator host (stays, becomes notplants-only) | **new combined host** |
+|---|---|---|---|
+| public IP | 91.98.47.73 (fsn1, Hetzner 134485294) | 168.119.126.100 (nbg1, Hetzner 134487234) | **195.201.88.249** |
+| tailnet | `cc-ci` 100.95.31.88 | `cc-ci-orchestrator-1` 100.84.190.30 | `cc-ci` (new node) |
+| RAM / disk | 8 GB / 150 GB (83 GB used, 59 GB docker) | 4 GB + 4 GB swap / 75 GB + 250 GB `/mnt/data` | 8 GB / 150 GB, one disk |
+| built by | `cc-ci` flake `#cc-ci` (nixpkgs 26.05 rev 531670d) | `notplants-nix` flake `#notplants-orchestrator` (26.05 channel), importing `cc-ci-orchestrator`'s `nixosModules.cc-ci` | `cc-ci-orchestrator` flake `#cc-ci` importing `cc-ci`'s `nixosModules.cc-ci-server` |
+| DNS | `ci.`, `*.ci.`, `ns-acme.commoninternet.net` → 91.98.47.73 (Gandi, direct, no gateway) | `oc.commoninternet.net` → 100.84.190.30 | operator repoints at cutover |
+
+Data on the old cc-ci server that must move: `/var/lib/cc-ci-reports` (published reports),
+`/var/lib/cc-ci-runs` (dashboard artifacts, 1.7 G), `/var/lib/ci-warm` (1.4 G), `/var/lib/acme`
+(LE cert valid to 2026-11-29 + **acme-dns account json** that the `_acme-challenge` CNAME points at),
+`/var/lib/acme-dns` (the authoritative zone DB), `/var/lib/ci-certs`, `/root/.abra` (app env files),
+`/etc/cc-ci` (deployed checkout the Sunday sweep runs from), Drone's `drone_ci_commoninternet_net_data`
+volume (Gitea OAuth grant + repo activation + build history). Warm recipe volumes are caches and get
+rebuilt on first use / the Sunday sweep. Docker swarm secrets/configs cannot be copied; the reconcile
+oneshots recreate them from sops.
+
+Out-of-band secrets the new host needs (never in git): `/var/lib/sops-nix/key.txt` (= the master age
+key, `/srv/cc-ci/.sops/master-age.txt` here — the new host's SSH host key is not a sops recipient),
+`/etc/ts-auth-key`, `/srv/cc-ci/.testenv`, `/srv/cc-ci/upgrader.env`, `/srv/cc-ci/.sops/master-age.txt`,
+`~loops/.ssh/{cc-ci-root,autonomic-bot-gitea,tangled}-ed25519`, `/etc/nginx/oc-*` (self-signed cert +
+htpasswd for the opencode UI), claude/opencode/codex auth under `~loops`.
+
+## Design
+
+**cc-ci repo** (`feat/nixos-module-export`):
+- `nixosModules.cc-ci-server` = `nix/modules/default.nix`: imports all service modules + the
+  host-generic cc-ci settings that used to sit in the host file (UTC, docker/swarm firewall 80/443,
+  `environment.systemPackages = ccciRuntimeTools`, allowUnfree). No hardware, no networking, no
+  tailscale, no root keys, no stateVersion — the host supplies those.
+- New options under `cc-ci.*`: `publicIPv4` (acme-dns listen + the `ns-acme` A record),
+  `sopsFile` (absolute path to the decrypted-at-activation `secrets.yaml`, default the submodule
+  path so `#cc-ci` keeps working), `repoPath` (`/etc/cc-ci`, used by nightly-sweep).
+- `nixosConfigurations.cc-ci` (old host) keeps building unchanged via the same module.
+
+**cc-ci-orchestrator repo** (`feat/combined-cc-ci-host`):
+- flake input `cc-ci` (https, public) with `nixpkgs`/`sops-nix` `follows` so one nixpkgs + one sops-nix.
+- `nixosModules.cc-ci-orchestrator` (the existing `nix/modules/cc-ci.nix`, kept exported as
+  `nixosModules.cc-ci` too so notplants-nix keeps evaluating until it drops the input) — the loops,
+  orchestrator session and the weekly/hourly timers.
+- `nix/modules/orchestrator-host.nix`: the host contract the module assumes — `loops` user + sudo,
+  nix-ld, claude/opencode/codex installers, `opencode-web`, the tailnet-only nginx `oc.` vhost
+  (on the tailscale IP, port **8443**, because traefik owns 80/443), tool packages, PATH.
+- `nixosConfigurations.cc-ci` = `nix/hosts/cc-ci/{configuration,hardware,networking}.nix` importing
+  both modules. `/srv` is a plain directory (no `/mnt/data`), 8 GB swapfile, root keys, tailscale
+  `--hostname=cc-ci`, firewall 22 (+ what cc-ci-server opens: 80, 443, 53).
+- `loops`' ssh config `Host cc-ci` → `127.0.0.1` so every `ssh cc-ci …` in skills/scripts keeps working.
+- `archive/`: the retired Incus/Hetzner-orchestrator host config, old terraform, historical plans.
+- `README.md`: provisioning (Hetzner Debian → nixos-infect → NixOS), secrets staging, the one
+  `nixos-rebuild`, data restore, cutover, verification — written so a person or an LLM can redo it.
+
+**notplants-nix** (`chore/drop-cc-ci`, after cutover): remove the `cc-ci` input, module import, the
+four cc-ci units' mount gating, `loopsSshConfig`, `opencode-web` + the `oc.` vhost (unless something
+notplants-side uses it), tailscale hostname → `notplants-orchestrator`.
+
+## Steps
+
+1. [ ] nixos-infect the new box (`NIX_CHANNEL=nixos-26.05 PROVIDER=hetzner`); capture
+   `hardware-configuration.nix` + `networking.nix`.
+2. [ ] cc-ci: module export + options; verify `#cc-ci` still evaluates; PR.
+3. [ ] cc-ci-orchestrator: input + host + modules + archive/ + README + terraform refresh; verify
+   `#cc-ci` evaluates; PR.
+4. [ ] Stage secrets + clones on the new host; `nixos-rebuild test` → verify → `switch`.
+   Immediately after: scale the new `ccci-bridge_app` to 0 and mask the two cc-ci timers so the
+   new host does not double-process `!testme` or run a second weekly upgrade before cutover.
+5. [ ] Copy data (rsync over tailscale): reports, runs, ci-warm, acme, acme-dns, ci-certs,
+   /root/.abra, /etc/cc-ci; Drone volume with Drone scaled to 0 during the copy.
+6. [ ] Pre-cutover verification on the new IP (`curl --resolve`, port 53, dashboard, reports,
+   drone, one direct `cc-ci-run` on custom-html-tiny).
+7. [ ] Operator: Gandi A records `ci`, `*.ci`, `ns-acme` → 195.201.88.249. Then: old bridge +
+   drone + timers off, new bridge up, one real `!testme` end-to-end, a `!testme`-driven report page.
+8. [ ] Move the orchestrator: stop cc-ci units here, final rsync of `/srv/cc-ci-orch` + agent
+   state, enable on the new host, operator reconnects there; notplants-nix PR removing cc-ci.
+9. [ ] Old cc-ci server: cold standby ~1 week, then operator deletes it and the stale tailnet node.
+10. [ ] Domain move to `autonomic.zone` — separate plan, after 1–9 are proven.
+
+## Log
+
+- 2026-09-07 19:40 UTC — recon done, plan written, ssh to the new box verified as root with
+  `notplants-orchestrator-ed25519`.
+- 2026-09-07 20:05 UTC — nixos-infect started on 195.201.88.249 (rev 40f62a6, nixos-26.05,
+  PROVIDER=hetznercloud). Two false starts: the Debian 13 image has /tmp on tmpfs, so
+  nixos-infect's temp swapfile fails `swapon: Invalid argument`; fixed with `NO_SWAP=true`.
+  Build ran, box rebooted ~20:11 UTC and has not answered ping/ssh since (>25 min) — needs the
+  Hetzner console (no API token for that project on this host).
+- 2026-09-07 20:40 UTC — cc-ci branch `feat/nixos-module-export` (9b99f81) pushed: the standalone
+  `#cc-ci` drv is byte-identical before/after. Orchestrator branch `feat/combined-cc-ci-host`:
+  `#cc-ci` evaluates (gcnwq4fy…-nixos-system-cc-ci-26.05.20260803.531670d.drv) with PROVISIONAL
+  hardware/networking copied from the old CI server — to be replaced by the infect output.
@@ -83,9 +83,10 @@ repo, a commit, a log, or the dashboard** (§9) — reference locations only.
 | What | Where | How to use |
 |---|---|---|
 | **cc-ci SSH (root)** | private key `~/.ssh/cc-ci-root-ed25519`; `Host cc-ci` in `~/.ssh/config` (HostName `100.90.116.4`, no ProxyCommand) | Just run `ssh cc-ci` (logs in as **root**). The orchestrator VM is a direct tailnet peer — direct route, no proxy. Pubkey already in cc-ci's `/root/.ssh/authorized_keys`. |
-| **Gitea bot account** | `/srv/cc-ci/.testenv` → `GITEA_USERNAME` (`autonomic-bot`), `GITEA_PASSWORD`, `GITEA_URL` (`git.autonomic.zone`) | Basic-auth to the Gitea API, or mint a scoped token: `POST https://$GITEA_URL/api/v1/users/$GITEA_USERNAME/tokens`. Used to push the `cc-ci` project repo, read recipe repos, comment on PRs, and poll for `!testme` (read-level; the bot does not register webhooks). |
+| **Gitea bot account** | `/srv/cc-ci/cc-ci/.env.public` → `GITEA_USERNAME` (`autonomic-bot`); `/srv/cc-ci/.testenv` → `GITEA_PASSWORD`; `GITEA_URL` defaults to `git.autonomic.zone` | Basic-auth to the Gitea API, or mint a scoped token: `POST https://$GITEA_URL/api/v1/users/$GITEA_USERNAME/tokens`. Used to push the `cc-ci` project repo, read recipe repos, comment on PRs, and poll for `!testme` (read-level; the bot does not register webhooks). |

-Load them in a shell with: `set -a; . /srv/cc-ci/.testenv; set +a` (don't echo the values).
+Load runtime configuration in a shell with: `set -a; . /srv/cc-ci/cc-ci-plan/load-env.sh; set +a`
+(don't echo secret values). The helper loads tracked public configuration first and then `.testenv`.

 ### The Tailscale connection (how `ssh cc-ci` works)

@@ -0,0 +1,234 @@
+#!/usr/bin/env python3
+"""pr-survey — deterministic facts about every open recipe PR, for /cc-ci-cleanup to judge.
+
+Open recipe PRs rot in specific, detectable ways. This gathers the evidence; it does NOT decide
+anything — closing a PR is a judgement the skill makes, with these facts in hand.
+
+RUN `reconcile-upstream.sh --all` FIRST. Every signal below is measured against the mirror's `main`,
+and an unreconciled mirror makes all of them wrong: on 2026-08-11 three PRs (discourse #6 carrying
+140 CVEs, keycloak #6 carrying 12, n8n #5) looked pending against a stale mirror while upstream had
+already merged them. This tool refuses to guess about that — see `reconciled_recently`.
+
+Per PR:
+  behind_main      commits on main not in the branch — the "out of date" measure
+  ahead            commits on the branch not on main
+  mergeable        gitea's own verdict (false = conflicts, needs a rebase)
+  diff_files       files the PR touches (0 = nothing left to merge)
+  adds_images      the `+ image:` lines it introduces
+  already_in_main  those `+ image:` lines ALREADY present in main -> the bump landed another way
+  obsolete         true when every image it adds is already in main (nothing to contribute)
+  ci               newest `!testme` verdict + build number parsed from the PR comments
+  branch_kind      upgrade / fix / ci-artifact (`ci/*` sweep + probe branches) / other
+  age_days, stale_days (since last update)
+
+    pr-survey.py [recipe ...] [--json]
+"""
+
+from __future__ import annotations
+
+import argparse
+import base64
+import json
+import os
+import re
+import sys
+import urllib.error
+import urllib.parse
+import urllib.request
+from datetime import datetime, timezone
+
+HERE = os.path.dirname(os.path.abspath(__file__))
+USED_RECIPES = os.path.join(HERE, "used-recipes.md")
+TESTENV = os.environ.get("CCCI_TESTENV", "/srv/cc-ci/.testenv")
+PUBLIC_ENV = "/srv/cc-ci/cc-ci/.env.public"
+NS = "recipe-maintainers"
+
+
+def _env() -> dict:
+    e = {}
+    for path in (PUBLIC_ENV, TESTENV):
+        try:
+            lines = open(path)
+        except OSError:
+            continue
+        for ln in lines:
+            ln = ln.strip()
+            if "=" in ln and not ln.startswith("#"):
+                k, v = ln.split("=", 1)
+                e[k] = v.strip().strip('"').strip("'")
+    return e
+
+
+ENV = _env()
+GITEA = os.environ.get("GITEA_URL") or ENV.get("GITEA_URL", "git.autonomic.zone")
+_AUTH = base64.b64encode(
+    f"{os.environ.get('GITEA_USERNAME') or ENV.get('GITEA_USERNAME','')}:"
+    f"{os.environ.get('GITEA_PASSWORD') or ENV.get('GITEA_PASSWORD','')}".encode()
+).decode()
+
+
+def _get(path: str, raw: bool = False):
+    req = urllib.request.Request(
+        f"https://{GITEA}{path}",
+        headers={"Authorization": f"Basic {_AUTH}", "User-Agent": "cc-ci-pr-survey"},
+    )
+    with urllib.request.urlopen(req, timeout=60) as r:
+        body = r.read()
+    return body.decode(errors="replace") if raw else json.loads(body)
+
+
+def _days(iso: str | None) -> int | None:
+    if not iso:
+        return None
+    try:
+        d = datetime.fromisoformat(iso.replace("Z", "+00:00"))
+    except ValueError:
+        return None
+    return (datetime.now(timezone.utc) - d).days
+
+
+def _branch_kind(ref: str) -> str:
+    if ref.startswith("ci/"):
+        return "ci-artifact"          # regall/cfold sweeps + testme probes; never meant to merge
+    if ref.startswith("upgrade"):
+        return "upgrade"
+    if re.match(r"^(fix|feat|chore|revert)", ref):
+        return "fix"
+    return "other"
+
+
+def _main_images(recipe: str) -> set[str]:
+    """Image refs pinned on the mirror's main — the baseline a PR is judged against."""
+    out = set()
+    for f in ("compose.yml",):
+        try:
+            txt = _get(f"/{NS}/{recipe}/raw/branch/main/{f}", raw=True)
+        except Exception:  # noqa: BLE001
+            continue
+        for m in re.finditer(r"^\s*image:\s*[\"']?([^\"'\s]+)", txt, re.M):
+            out.add(m.group(1))
+    return out
+
+
+def _ci_verdict(recipe: str, number: int) -> dict:
+    """Newest cc-ci !testme outcome recorded on the PR."""
+    try:
+        cs = _get(f"/api/v1/repos/{NS}/{recipe}/issues/{number}/comments?limit=100")
+    except Exception:  # noqa: BLE001
+        return {"verdict": "unknown", "build": None}
+    for c in reversed(cs):
+        b = c.get("body") or ""
+        if "cc-ci:testme" not in b:
+            continue
+        m = re.search(r"/cc-ci/(\d+)", b)
+        if "✅" in b or "passed" in b:
+            return {"verdict": "passed", "build": m.group(1) if m else None}
+        if "❌" in b or "failure" in b:
+            return {"verdict": "failed", "build": m.group(1) if m else None}
+        if "⏳" in b or "in progress" in b:
+            return {"verdict": "running", "build": m.group(1) if m else None}
+    return {"verdict": "never-run", "build": None}
+
+
+def survey_pr(recipe: str, pr: dict, main_images: set[str]) -> dict:
+    n = pr["number"]
+    head = pr["head"]["ref"]
+    row = {
+        "recipe": recipe, "number": n, "title": pr.get("title", ""), "head": head,
+        "url": pr.get("html_url"), "branch_kind": _branch_kind(head),
+        "age_days": _days(pr.get("created_at")), "stale_days": _days(pr.get("updated_at")),
+        "mergeable": pr.get("mergeable"),
+    }
+    try:
+        row["behind_main"] = _get(
+            f"/api/v1/repos/{NS}/{recipe}/compare/{urllib.parse.quote(head, safe='')}...main"
+        ).get("total_commits", 0)
+        row["ahead"] = _get(
+            f"/api/v1/repos/{NS}/{recipe}/compare/main...{urllib.parse.quote(head, safe='')}"
+        ).get("total_commits", 0)
+    except Exception:  # noqa: BLE001
+        row["behind_main"], row["ahead"] = None, None
+    # A FAILED diff fetch must never look like an empty diff: gitea#4 404s on .diff (force-pushed
+    # branch) and would otherwise be flagged EMPTY-DIFF and closed — while being a verified, green,
+    # needed fix. Unknown is its own state.
+    diff = None
+    try:
+        body = _get(f"/{NS}/{recipe}/pulls/{n}.diff", raw=True)
+        if body.lstrip().startswith(("diff --git", "From ")) or not body.strip():
+            diff = body
+    except Exception:  # noqa: BLE001
+        diff = None
+    row["diff_files"] = None if diff is None else len(re.findall(r"^diff --git ", diff, re.M))
+    adds = re.findall(r"^\+\s*image:\s*[\"']?([^\"'\s]+)", diff or "", re.M)
+    row["adds_images"] = sorted(set(adds))
+    row["already_in_main"] = sorted({i for i in set(adds) if i in main_images})
+    # Nothing left to contribute: it touches files but every image it introduces is already pinned.
+    # Only claim obsolete when the diff was actually READ. No diff, no verdict.
+    row["obsolete"] = diff is not None and bool(adds) and set(adds).issubset(main_images)
+    row["ci"] = _ci_verdict(recipe, n)
+    return row
+
+
+def all_recipes() -> list[str]:
+    out = []
+    for ln in open(USED_RECIPES):
+        p = ln.split()
+        if len(p) >= 2 and not ln.startswith(("#", "`")) and p[1] in ("weekly", "external"):
+            out.append(p[0])
+    return out
+
+
+def main() -> int:
+    ap = argparse.ArgumentParser(description=__doc__,
+                                 formatter_class=argparse.RawDescriptionHelpFormatter)
+    ap.add_argument("recipes", nargs="*")
+    ap.add_argument("--json", action="store_true")
+    a = ap.parse_args()
+
+    rows = []
+    for r in (a.recipes or all_recipes()):
+        try:
+            prs = _get(f"/api/v1/repos/{NS}/{r}/pulls?state=open&limit=50")
+        except urllib.error.HTTPError as e:
+            rows.append({"recipe": r, "error": f"HTTP {e.code}"})
+            continue
+        if not prs:
+            continue
+        mi = _main_images(r)
+        for pr in prs:
+            rows.append(survey_pr(r, pr, mi))
+
+    if a.json:
+        print(json.dumps(rows, indent=2))
+        return 0
+
+    print(f"{len(rows)} open PR(s)\n")
+    for x in sorted(rows, key=lambda z: (z.get("recipe", ""), z.get("number", 0))):
+        if x.get("error"):
+            print(f"  {x['recipe']}: {x['error']}")
+            continue
+        flags = []
+        if x["obsolete"]:
+            flags.append("OBSOLETE(images already in main)")
+        if x["branch_kind"] == "ci-artifact":
+            flags.append("CI-ARTIFACT")
+        if x["diff_files"] == 0:
+            flags.append("EMPTY-DIFF")
+        if x["diff_files"] is None:
+            flags.append("DIFF-UNREADABLE(do not close on this)")
+        if x["mergeable"] is False:
+            flags.append("CONFLICTS")
+        if (x["behind_main"] or 0) > 0:
+            flags.append(f"BEHIND-{x['behind_main']}")
+        print(f"  {x['recipe']}#{x['number']:<3} {x['title'][:52]}")
+        print(f"      {x['branch_kind']:12} age={x['age_days']}d idle={x['stale_days']}d "
+              f"ci={x['ci']['verdict']}({x['ci']['build'] or '-'}) files={x['diff_files'] if x['diff_files'] is not None else '?'}")
+        if x["adds_images"]:
+            print(f"      adds: {', '.join(i.split('/')[-1] for i in x['adds_images'][:4])}")
+        if flags:
+            print(f"      >> {' | '.join(flags)}")
+    return 0
+
+
+if __name__ == "__main__":
+    sys.exit(main())
@@ -12,7 +12,7 @@ Subcommands (the /recipe-report agent runs them around its own review/classifica
  publish OUT.html DATE [KIND]  copy to cc-ci:/var/lib/cc-ci-reports/<KIND>-DATE.html and regen the
                           archive index. KIND is `week` (default, the weekly /recipe-report) or `cve`
                           (a /cve-check advisory sweep). BOTH kinds appear in the SAME archive index,
-                           newest first, each row suffixed "full report" or "CVE check"; the distinct
+                           newest first, each row suffixed "full" or "CVE check"; the distinct
                           filename prefix just stops a sweep overwriting a weekly edition.

 Page order: short lead → the full wire table (priority-sorted, CVEs column) → Addendum → Security
@@ -46,24 +46,27 @@ from datetime import datetime, timezone

 LOGDIR = "/srv/cc-ci/.cc-ci-logs"
 TESTENV = "/srv/cc-ci/.testenv"
+PUBLIC_ENV = "/srv/cc-ci/cc-ci/.env.public"
 INFRA = {"cc-ci", "cc-ci-orchestrator", "cc-ci-secrets"}
 HOST_REPORTS = "/var/lib/cc-ci-reports"
 # Both kinds live in ONE archive, distinguished by a suffix on a common title.
 # prefix -> (page title, index label)
-KINDS = {"week": ("The Recipe Report", "Week of {d} — full report"),
+KINDS = {"week": ("The Recipe Report", "Week of {d} — full"),
         "cve":  ("The Recipe Report — CVE check", "{d} — CVE check")}


 def _env():
    e = {}
-    try:
-        for ln in open(TESTENV):
+    for path in (PUBLIC_ENV, TESTENV):
+        try:
+            lines = open(path)
+        except FileNotFoundError:
+            continue
+        for ln in lines:
            ln = ln.strip()
            if "=" in ln and not ln.startswith("#"):
                k, v = ln.split("=", 1)
                e[k] = v.strip().strip('"').strip("'")
-    except FileNotFoundError:
-        pass
    return e


@@ -0,0 +1,63 @@
+#!/usr/bin/env bash
+# reconcile-upstream — sync recipe mirrors from TRUE upstream. Run this FIRST, always.
+# ----------------------------------------------------------------------------------
+# Every recipe we maintain is a MIRROR of a coopcloud recipe. Work done against a stale
+# mirror is wasted or wrong, in three ways we have actually hit:
+#
+#   1. A PR whose changes upstream ALREADY MERGED. mailu #6 (2024.06.57 + redis 8.10,
+#      two internet-facing Roundcube CVEs) sat open and was reported as the fix for
+#      those CVEs — while upstream had merged and released it as 3.1.3+2024.06.57. The
+#      work was done; only our mirror was behind.
+#   2. A survey that reads the stale mirror and reports "no upgrades available", so a
+#      recipe silently drops out of the weekly run.
+#   3. Reading the WRONG BRANCH. Several coopcloud recipes keep a stale `main` beside
+#      the real default `master` — gitea's `main` is at 1.24.2-rootless while `master`
+#      has 1.27.1-rootless plus the merged PRs. Reading `main` there says the recipe is
+#      three releases behind and missing two CVSS-9.8 RCE fixes, which reads exactly
+#      like a real finding. open-recipe-pr.sh resolves the default branch itself
+#      (main OR master) — never hand-pick one.
+#
+# This is deterministic: it force-syncs each mirror's `main` to upstream's default
+# branch and closes any mirror PR whose changes are already upstream. No AI judgement.
+#
+#   reconcile-upstream.sh <recipe>...     # specific recipes
+#   reconcile-upstream.sh --all           # every recipe in used-recipes.md
+#
+# Safe to run repeatedly; a mirror already in sync is a no-op. Recipe work lives in
+# BRANCHES, never on mirror `main`, so force-syncing `main` discards nothing.
+set -o errexit -o nounset -o pipefail
+
+HERE="$(cd "$(dirname "${BASH_SOURCE[0]}")" && pwd)"
+ORCH="$(dirname "$HERE")"
+SSH="${SSH:-cc-ci}"
+RECONCILE="${RECONCILE:-$ORCH/.claude/skills/recipe-upgrade/open-recipe-pr.sh}"
+USED_RECIPES="${USED_RECIPES:-$HERE/used-recipes.md}"
+
+[ -f "$RECONCILE" ] || { echo "ERROR: reconcile helper not found: $RECONCILE" >&2; exit 1; }
+set -a; . /srv/cc-ci/cc-ci-plan/load-env.sh; set +a
+: "${GITEA_USERNAME:?}"; : "${GITEA_PASSWORD:?}"; : "${GITEA_URL:=git.autonomic.zone}"
+
+if [ "${1:-}" = "--all" ]; then
+  mapfile -t RECIPES < <(awk '!/^[[:space:]]*#/ && ($2=="weekly" || $2=="external") {print $1}' "$USED_RECIPES")
+else
+  [ "$#" -gt 0 ] || { echo "usage: reconcile-upstream.sh <recipe>... | --all" >&2; exit 2; }
+  RECIPES=("$@")
+fi
+
+synced=0; closed=0; failed=0
+for r in "${RECIPES[@]}"; do
+  echo "── $r"
+  if out="$(ssh "$SSH" "GITEA_USERNAME='$GITEA_USERNAME' GITEA_PASSWORD='$GITEA_PASSWORD' GITEA_URL='$GITEA_URL' bash -s $r --reconcile-only" < "$RECONCILE" 2>&1)"; then
+    printf '%s\n' "$out" | grep -E "Force-syncing|already in sync|closed PR|still open|✓" | sed 's/^/   /' || true
+    synced=$((synced + 1))
+    closed=$((closed + $(printf '%s' "$out" | grep -c "closed PR" || true)))
+  else
+    printf '%s\n' "$out" | tail -3 | sed 's/^/   /'
+    echo "   ✗ FAILED — do NOT proceed against this mirror until it reconciles"
+    failed=$((failed + 1))
+  fi
+done
+
+echo
+echo "reconcile-upstream: ${synced} mirror(s) synced, ${closed} already-upstream PR(s) closed, ${failed} failed"
+[ "$failed" -eq 0 ]
@@ -0,0 +1,459 @@
+#!/usr/bin/env python3
+"""resolve-images — what version is each of a recipe's images on, and what is newest?
+
+An abra-independent version resolver. `abra recipe upgrade` is the normal path, but it has a hard
+failure mode: an image pinned with BOTH a tag and a digest makes it FATA and abandon the WHOLE
+recipe — even images it already parsed. immich pins two that way:
+
+    ghcr.io/immich-app/postgres:14-vectorchord0.4.3-pgvectors0.2.0@sha256:bcf6…
+    docker.io/valkey/valkey:9@sha256:3acc…
+
+so immich contributes NO version data at all and silently drops out of every survey. That is
+indistinguishable from "up to date" unless a human notices the missing row — which is exactly how it
+kept getting skipped, and why a CVE sweep reported it as unknown.
+
+This reads the compose files directly and queries the registries itself, so a digest pin is just a
+digest pin. Output is JSON (default) or a table.
+
+    resolve-images.py <recipe> [--ssh HOST] [--recipe-dir DIR] [--table] [--only IMAGE]
+
+The cc-ci host has no python3, so `--ssh cc-ci` reads the compose files from that host's checkout
+over ssh and does the resolving locally. That keeps the source of truth the SAME tree abra and CI
+use, rather than a second copy that can drift.
+
+TAG SHAPES. Registries mix wildly different tag conventions in one repo, so "newest" is meaningless
+without a shape. Each tag is reduced to a signature by replacing digit runs with '#':
+
+    v3.1.0                              -> v#.#.#
+    1.27.1-rootless                     -> #.#.#-rootless
+    8.10-alpine                         -> #.#-alpine
+    14-vectorchord0.4.3-pgvectors0.2.0  -> #-vectorchord#.#.#-pgvectors#.#.#
+
+Only tags sharing the CURRENT pin's shape are candidates. That keeps `-alpine` on `-alpine`, and
+stops a `latest`/`release`/`sha-…` tag from ever being proposed as an upgrade.
+
+TWO ANSWERS, NOT ONE. It reports `newest_same_shape` AND `newest_within_major` (same leading number).
+For a plain app image they usually agree. For a compatibility-pinned sidecar they do not, and taking
+the max would be wrong: immich's postgres tag encodes the pg major plus the vectorchord/pgvectors
+versions that immich-server is built against, so jumping pg major because a newer tag exists breaks
+the deployment. The caller picks; this tool refuses to guess and shows both.
+"""
+
+from __future__ import annotations
+
+import argparse
+import glob
+import shlex
+import subprocess
+import gzip
+import json
+import os
+import re
+import sys
+import time
+import urllib.error
+import urllib.parse
+import urllib.request
+
+UA = "cc-ci-resolve-images (+https://git.autonomic.zone/recipe-maintainers/cc-ci)"
+TIMEOUT = int(os.environ.get("RESOLVE_IMAGES_TIMEOUT", "45"))
+RECIPE_DIR = os.environ.get("ABRA_RECIPE_DIR", os.path.expanduser("~/.abra/recipes"))
+MAX_TAG_PAGES = int(os.environ.get("RESOLVE_IMAGES_MAX_PAGES", "40"))
+
+IMAGE_RE = re.compile(r"""^\s*image:\s*["']?([^"'\s]+)["']?\s*$""", re.M)
+
+
+RETRIES = int(os.environ.get("RESOLVE_IMAGES_RETRIES", "4"))
+
+
+def _fetch(url: str, headers: dict | None = None) -> bytes:
+    """GET with backoff on rate limits.
+
+    Docker Hub throttles anonymous clients hard, and a sweep re-reads the same popular repos
+    (nginx, redis, postgres) for recipe after recipe. A 429 mid-sweep used to surface as
+    'unresolved', which is indistinguishable from a real lookup failure — so retry, and let the
+    per-repo cache below remove most of the requests entirely."""
+    h = {"User-Agent": UA, "Accept-Encoding": "gzip"}
+    h.update(headers or {})
+    delay = 2.0
+    for attempt in range(RETRIES):
+        try:
+            with urllib.request.urlopen(urllib.request.Request(url, headers=h), timeout=TIMEOUT) as r:
+                raw = r.read()
+                if r.headers.get("Content-Encoding") == "gzip":
+                    raw = gzip.decompress(raw)
+                return raw
+        except urllib.error.HTTPError as e:
+            if e.code in (429, 503) and attempt < RETRIES - 1:
+                time.sleep(delay)
+                delay *= 2
+                continue
+            raise
+    raise RuntimeError("unreachable")
+
+
+_HUB_JWT: list = []
+
+
+def _hub_auth() -> dict:
+    """Authenticated Docker Hub calls get a far higher rate limit than anonymous ones.
+
+    Credentials come from /srv/cc-ci/.testenv (DOCKERHUB_USERNAME / DOCKERHUB_TOKEN), the same pair
+    the CI host already uses. Absent creds are fine — the sweep just runs anonymous and slower."""
+    if _HUB_JWT:
+        return _HUB_JWT[0]
+    env = {}
+    try:
+        for ln in open(os.environ.get("CCCI_TESTENV", "/srv/cc-ci/.testenv")):
+            if "=" in ln and not ln.strip().startswith("#"):
+                k, v = ln.strip().split("=", 1)
+                env[k] = v.strip().strip("\"'")
+    except OSError:
+        pass
+    u = os.environ.get("DOCKERHUB_USERNAME") or env.get("DOCKERHUB_USERNAME")
+    t = os.environ.get("DOCKERHUB_TOKEN") or env.get("DOCKERHUB_TOKEN")
+    hdrs = {}
+    if u and t:
+        try:
+            body = json.dumps({"username": u, "password": t}).encode()
+            req = urllib.request.Request("https://hub.docker.com/v2/users/login",
+                                         data=body, method="POST",
+                                         headers={"Content-Type": "application/json", "User-Agent": UA})
+            with urllib.request.urlopen(req, timeout=TIMEOUT) as r:
+                tokj = json.load(r).get("token")
+            if tokj:
+                hdrs = {"Authorization": f"JWT {tokj}"}
+        except Exception:  # noqa: BLE001 — anonymous is a valid fallback
+            hdrs = {}
+    _HUB_JWT.append(hdrs)
+    return hdrs
+
+
+def _json(url: str, headers: dict | None = None):
+    return json.loads(_fetch(url, headers))
+
+
+def shape(tag: str) -> str:
+    """Signature of a tag with every digit run replaced by '#'. See module docstring."""
+    return re.sub(r"\d+", "#", tag)
+
+
+def vkey(tag: str) -> tuple:
+    """Ordering key: every number in the tag, in order. '1.27.10' > '1.27.9'; text ignored."""
+    return tuple(int(x) for x in re.findall(r"\d+", tag))
+
+
+def parse_ref(ref: str) -> dict:
+    """Split an image reference into registry / repo / tag / digest."""
+    digest = None
+    if "@" in ref:
+        ref, _, digest = ref.partition("@")
+    host, repo, tag = "docker.io", ref, "latest"
+    # A leading component is a REGISTRY only when there is a path after it. Without the slash test,
+    # a bare `postgres:15.18` looks like host "postgres:15.18" because of the tag's colon — which
+    # silently sent every library image to a nonexistent registry.
+    if "/" in ref:
+        first = ref.split("/")[0]
+        if "." in first or ":" in first or first == "localhost":
+            host, _, repo = ref.partition("/")
+    if ":" in repo.split("/")[-1]:
+        repo, _, tag = repo.rpartition(":")
+    if host == "docker.io" and "/" not in repo:
+        repo = f"library/{repo}"          # bare `redis` is really `library/redis`
+    return {"registry": host, "repo": repo, "tag": tag, "digest": digest}
+
+
+HUB_RECENT_PAGES = int(os.environ.get("RESOLVE_IMAGES_HUB_PAGES", "10"))
+
+
+def _hub_tag_exists(repo: str, tag: str) -> bool:
+    try:
+        _json(f"https://hub.docker.com/v2/repositories/{repo}/tags/{tag}", _hub_auth())
+        return True
+    except Exception:  # noqa: BLE001
+        return False
+
+
+def _hub_tags(repo: str) -> list[str]:
+    """Recently-pushed tags, newest first.
+
+    Popular Docker Hub repos carry many thousands of tags, so a full enumeration is impractical —
+    but it is also unnecessary: a tag NEWER than the one we run must have been pushed AFTER it, so
+    ordering by last_updated and reading a bounded recent window is sufficient to find any upgrade.
+    (ghcr offers no ordering, which is why that path needs a different strategy.)
+    """
+    tags, url = [], (f"https://hub.docker.com/v2/repositories/{repo}/tags"
+                     f"?page_size=100&ordering=last_updated")
+    auth = _hub_auth()
+    for _ in range(HUB_RECENT_PAGES):
+        d = _json(url, auth)
+        tags += [r["name"] for r in d.get("results", [])]
+        url = d.get("next")
+        if not url:
+            break
+    return tags
+
+
+def _oci_bearer(host: str, repo: str) -> dict:
+    """Token for an OCI registry, discovered from its own auth challenge.
+
+    Registries do NOT share a token endpoint. ghcr answers at /token?scope=…&service=ghcr.io, but
+    lscr.io and dock.mau.dev advertise different realms, and assuming ghcr's shape made both 401 —
+    which then read as "could not resolve" rather than "asked the wrong URL". The registry tells us
+    where to go in its WWW-Authenticate header; use that."""
+    try:
+        urllib.request.urlopen(
+            urllib.request.Request(f"https://{host}/v2/{repo}/tags/list?n=1",
+                                   headers={"User-Agent": UA}), timeout=TIMEOUT)
+        return {}                                     # no auth needed
+    except urllib.error.HTTPError as e:
+        if e.code != 401:
+            return {}
+        chal = e.headers.get("WWW-Authenticate", "") or ""
+    except Exception:  # noqa: BLE001
+        return {}
+    if not chal.lower().startswith("bearer"):
+        return {}
+    parts = dict(re.findall(r'(\w+)="([^"]*)"', chal))
+    realm = parts.get("realm")
+    if not realm:
+        return {}
+    q = {"service": parts.get("service", host), "scope": parts.get("scope", f"repository:{repo}:pull")}
+    url = realm + ("&" if "?" in realm else "?") + urllib.parse.urlencode(q)
+    try:
+        tok = (_json(url) or {}).get("token") or (_json(url) or {}).get("access_token")
+        return {"Authorization": f"Bearer {tok}"} if tok else {}
+    except Exception:  # noqa: BLE001
+        return {}
+
+
+def _oci_tags(host: str, repo: str) -> list[str]:
+    """Tags from any OCI/v2 registry, with challenge-derived auth and Link pagination.
+
+    ghcr paginates hard — immich-server has >40,000 tags — and a truncated listing silently hides
+    the newest release line, so follow the cursor and let the caller's integrity check catch a read
+    that never reached the current pin."""
+    hdrs = _oci_bearer(host, repo)
+    tags, url = [], f"https://{host}/v2/{repo}/tags/list?n=1000"
+    for _ in range(MAX_TAG_PAGES):
+        req = urllib.request.Request(url, headers={"User-Agent": UA, **hdrs})
+        with urllib.request.urlopen(req, timeout=TIMEOUT) as r:
+            tags += (json.load(r) or {}).get("tags") or []
+            link = r.headers.get("Link", "") or ""
+        m = re.search(r'<([^>]+)>;\s*rel="next"', link)
+        if not m:
+            break
+        nxt = m.group(1)
+        url = f"https://{host}{nxt}" if nxt.startswith("/") else nxt
+    return tags
+
+
+def _gh_token() -> str | None:
+    tok = os.environ.get("GITHUB_TOKEN")
+    if tok:
+        return tok.strip()
+    try:
+        return open(os.environ.get("GITHUB_TOKEN_FILE", "/srv/cc-ci/.github-token")).read().strip() or None
+    except OSError:
+        return None
+
+
+def github_release_tags(owner: str, repo: str, max_pages: int = 4) -> list[str]:
+    """Release tag names for a GitHub repo, newest first.
+
+    FALLBACK for registries whose tag listing cannot be enumerated. ghcr has no ordering and no
+    server-side filter, and immich-machine-learning carries >40,000 tags — a full read is impractical
+    and a partial read silently hides the newest release line. The project's RELEASES are ordered,
+    small, and authoritative: container tags track them. (The GitHub Packages API would answer this
+    directly but needs a scoped token; this scan's token deliberately has none.)
+    """
+    hdrs = {"Accept": "application/vnd.github+json"}
+    tok = _gh_token()
+    if tok:
+        hdrs["Authorization"] = f"Bearer {tok}"
+    out = []
+    for page in range(1, max_pages + 1):
+        try:
+            rows = _json(f"https://api.github.com/repos/{owner}/{repo}/releases"
+                         f"?per_page=100&page={page}", hdrs)
+        except Exception:  # noqa: BLE001
+            break
+        if not rows:
+            break
+        out += [r.get("tag_name") or "" for r in rows]
+    return [t for t in out if t]
+
+
+def _release_fallback_repos(registry: str, repo: str) -> list[tuple[str, str]]:
+    """Candidate GitHub repos whose releases track this image's tags."""
+    if "ghcr.io" not in registry:
+        return []
+    parts = repo.split("/")
+    if len(parts) < 2:
+        return []
+    owner, name = parts[0], parts[-1]
+    cands = [(owner, name)]
+    # ghcr.io/immich-app/immich-machine-learning is built from immich-app/immich.
+    if name.startswith(owner.split("-")[0]):
+        cands.append((owner, owner.split("-")[0]))
+    return cands
+
+
+_TAG_CACHE: dict[tuple[str, str], tuple[list[str], str | None]] = {}
+
+
+def list_tags(registry: str, repo: str) -> tuple[list[str], str | None]:
+    if (registry, repo) in _TAG_CACHE:
+        return _TAG_CACHE[(registry, repo)]
+    res = _list_tags_uncached(registry, repo)
+    _TAG_CACHE[(registry, repo)] = res
+    return res
+
+
+def _list_tags_uncached(registry: str, repo: str) -> tuple[list[str], str | None]:
+    try:
+        return (_hub_tags(repo) if registry in ("docker.io", "registry-1.docker.io")
+                else _oci_tags(registry, repo)), None
+    except urllib.error.HTTPError as e:
+        return [], f"HTTP {e.code}"
+    except Exception as e:  # noqa: BLE001
+        return [], f"{type(e).__name__}: {e}"
+
+
+def resolve(ref: str) -> dict:
+    """Current pin -> newest same-shape tag, and newest within the current major."""
+    if "${" in ref or "$(" in ref:
+        # The tag is a compose variable (ghost pins `ghost:${IMAGE_VERSION}-alpine`). Its real value
+        # lives in .env, not here. Report it as skipped, never as a failed lookup.
+        return {**parse_ref(ref), "ref": ref, "shape": None, "candidates": 0,
+                "newest_same_shape": None, "newest_within_major": None,
+                "upgrade_available": False,
+                "status": "skipped: templated ref (tag comes from a compose variable)"}
+    info = parse_ref(ref)
+    out = {**info, "ref": ref, "shape": shape(info["tag"]), "status": "ok",
+           "newest_same_shape": None, "newest_within_major": None, "candidates": 0,
+           "upgrade_available": False}
+    tags, err = list_tags(info["registry"], info["repo"])
+    if err:
+        out["status"] = f"error: {err}"
+        return out
+    out["tags_seen"] = len(set(tags))
+    # INTEGRITY CHECK: the tag we are currently running MUST appear in the listing. If it does not,
+    # the listing is incomplete and any "newest" derived from it is a guess — ghcr paginates to tens
+    # of thousands of tags and a truncated read silently hides whole release lines. immich's
+    # machine-learning image is pinned v3.1.0, which EXISTS, yet a short read reported v1.134.0 as
+    # newest; without this check that becomes a confident, wrong answer.
+    if info["tag"] not in set(tags):
+        # Docker Hub: the window is recency-ordered, so the pin being outside it just means the pin
+        # is old — which is fine, because anything NEWER is necessarily inside the window. Confirm
+        # the pin genuinely exists (so a typo is still caught) and carry on.
+        if info["registry"] in ("docker.io", "registry-1.docker.io") and _hub_tag_exists(info["repo"], info["tag"]):
+            out["source"] = f"docker-hub:recent-{HUB_RECENT_PAGES * 100}"
+            tags = list(tags) + [info["tag"]]
+        else:
+            for owner, name in _release_fallback_repos(info["registry"], info["repo"]):
+                rel = github_release_tags(owner, name)
+                if info["tag"] in rel:
+                    tags = rel
+                    out["source"] = f"github-releases:{owner}/{name}"
+                    out["tags_seen"] = len(set(rel))
+                    break
+            else:
+                out["status"] = ("error: tag listing incomplete — the current pin "
+                                 f"{info['tag']!r} is absent from {len(set(tags))} registry tags "
+                                 f"and from the project's GitHub releases")
+                return out
+    want, cur = out["shape"], vkey(info["tag"])
+    same = [t for t in set(tags) if shape(t) == want and vkey(t)]
+    out["candidates"] = len(same)
+    if not same:
+        # Not a failure: digest-only pins and `latest`/`stable` have no comparable siblings.
+        out["status"] = "no comparable tags (shape has no numeric siblings)"
+        return out
+    newest = max(same, key=vkey)
+    out["newest_same_shape"] = newest
+    if cur:
+        within = [t for t in same if vkey(t)[:1] == cur[:1]]
+        if within:
+            out["newest_within_major"] = max(within, key=vkey)
+    out["upgrade_available"] = bool(cur and vkey(newest) > cur)
+    return out
+
+
+def compose_images_ssh(recipe: str, host: str, recipe_dir: str) -> list[str]:
+    """Same as compose_images, but the recipe tree lives on another host (cc-ci has no python3)."""
+    # NB: no shell-quoting of the directory — it may legitimately start with ~ or $HOME, and
+    # quoting it stops the remote shell expanding it, which yields an empty (and silent) result.
+    d = f"{recipe_dir}/{shlex.quote(recipe)}".replace("~", "$HOME")
+    cmd = (f'for f in {d}/compose*.yml; do case "$f" in *compose.ccci.yml) continue;; esac; '
+           f'[ -f "$f" ] && {{ cat "$f"; echo; }}; done; exit 0')
+    out = subprocess.run(["ssh", host, cmd], capture_output=True, text=True, timeout=120)
+    if out.returncode != 0:
+        raise RuntimeError(f"ssh {host}: {(out.stderr.strip() or 'no output')[:200]}")
+    if not out.stdout.strip():
+        raise RuntimeError(f"ssh {host}: no compose files found under {d}")
+    refs = []
+    for m in IMAGE_RE.finditer(out.stdout):
+        if m.group(1) not in refs:
+            refs.append(m.group(1))
+    return refs
+
+
+def compose_images(recipe: str, recipe_dir: str) -> list[str]:
+    """Every `image:` ref in the recipe's own compose files (the cc-ci overlay is NOT the recipe)."""
+    refs, base = [], os.path.join(recipe_dir, recipe)
+    for path in sorted(glob.glob(os.path.join(base, "compose*.yml"))):
+        if os.path.basename(path) == "compose.ccci.yml":
+            continue
+        try:
+            for m in IMAGE_RE.finditer(open(path).read()):
+                if m.group(1) not in refs:
+                    refs.append(m.group(1))
+        except OSError:
+            continue
+    return refs
+
+
+def main() -> int:
+    ap = argparse.ArgumentParser(description=__doc__,
+                                 formatter_class=argparse.RawDescriptionHelpFormatter)
+    ap.add_argument("recipe")
+    ap.add_argument("--ssh", default=None, metavar="HOST",
+                    help="read the recipe's compose files from HOST over ssh (e.g. --ssh cc-ci); "
+                         "resolving still happens locally")
+    ap.add_argument("--recipe-dir", default=RECIPE_DIR)
+    ap.add_argument("--table", action="store_true", help="human-readable table instead of JSON")
+    ap.add_argument("--only", default=None, help="resolve just the images whose ref contains this")
+    a = ap.parse_args()
+
+    rdir = a.recipe_dir if a.recipe_dir != RECIPE_DIR or not a.ssh else "~/.abra/recipes"
+    refs = (compose_images_ssh(a.recipe, a.ssh, rdir) if a.ssh
+            else compose_images(a.recipe, a.recipe_dir))
+    if a.only:
+        refs = [r for r in refs if a.only in r]
+    results = [resolve(r) for r in refs]
+    report = {
+        "recipe": a.recipe,
+        "images": results,
+        "upgrades_available": [r["ref"] for r in results if r["upgrade_available"]],
+        "unresolved": [r["ref"] for r in results if r["status"].startswith("error")],
+        # The whole point: distinguish "checked, current" from "could not check".
+        "all_resolved": not any(r["status"].startswith("error") for r in results),
+    }
+    if not a.table:
+        print(json.dumps(report, indent=2))
+        return 0
+    print(f"{a.recipe} — {len(results)} images")
+    for r in results:
+        flag = "UPGRADE" if r["upgrade_available"] else ("ERROR" if r["status"].startswith("error") else "current")
+        print(f"  [{flag:7}] {r['repo']}:{r['tag']}" + ("  (digest-pinned)" if r["digest"] else ""))
+        print(f"            shape={r['shape']}  candidates={r['candidates']}"
+              f"  newest_same_shape={r['newest_same_shape']}"
+              f"  newest_within_major={r['newest_within_major']}")
+        if r["status"] != "ok":
+            print(f"            status: {r['status']}")
+    return 0
+
+
+if __name__ == "__main__":
+    sys.exit(main())
@@ -503,6 +503,195 @@ class TestReleaseNoteResolution(unittest.TestCase):
        self.assertNotIn("CVE-OK", rep.get("resolved_by_release_notes") or {})


+class TestReleaseLineSemantics(unittest.TestCase):
+    """A fix inside the numeric window is not a fix on the branch you actually land on."""
+
+    def test_fix_later_on_the_targets_own_line_is_not_counted(self):
+        # ClickHouse fixed CVE-2023-48704 in 23.9.6.20 AND 23.10.5.20. Landing on 23.10.4.25 crosses
+        # the 23.9 fix numerically but is BELOW its own line's fix, so it does not have it.
+        rep = run_scan([gh("ClickHouse/ClickHouse",
+                           [adv("CVE-2023-48704", patched="v23.10.5.20; v23.9.6.20; v23.8.8.20")])],
+                       v_from="23.4.2.11", v_to="23.10.4.25",
+                       urls=["https://github.com/ClickHouse/ClickHouse"])
+        self.assertEqual(rep["fixed_by_this_upgrade"], [])
+
+    def test_fix_earlier_on_the_targets_own_line_is_counted(self):
+        rep = run_scan([gh("ClickHouse/ClickHouse",
+                           [adv("CVE-2023-47118", patched="v23.10.2.13; v23.8.6.16")])],
+                       v_from="23.4.2.11", v_to="23.10.4.25",
+                       urls=["https://github.com/ClickHouse/ClickHouse"])
+        self.assertEqual(rep["fixed_by_this_upgrade"], ["CVE-2023-47118"])
+
+    def test_fix_exactly_at_the_target_is_counted(self):
+        rep = run_scan([gh("ClickHouse/ClickHouse",
+                           [adv("CVE-2023-48298", patched="v23.10.4.25; v23.9.5.29")])],
+                       v_from="23.4.2.11", v_to="23.10.4.25",
+                       urls=["https://github.com/ClickHouse/ClickHouse"])
+        self.assertEqual(rep["fixed_by_this_upgrade"], ["CVE-2023-48298"])
+
+    def test_no_fix_on_the_target_line_falls_back_to_the_window(self):
+        # redis fixes 7.4.6/8.0.4/8.2.2 with no 8.10.x entry; landing on 8.10 still has them,
+        # because nothing on the 8.10 line is named as a LATER fix.
+        rep = run_scan([gh("redis/redis", [adv("CVE-2025-49844", patched="7.4.6; 8.0.4; 8.2.2")])],
+                       v_from="7.4", v_to="8.10", urls=["https://github.com/redis/redis"])
+        self.assertEqual(rep["fixed_by_this_upgrade"], ["CVE-2025-49844"])
+
+
+class TestAlreadyFixedOnFromLine(unittest.TestCase):
+    """A fix that landed on the line we upgrade FROM was already ours before the upgrade."""
+
+    def test_backport_to_our_own_line_is_not_credited(self):
+        # mattermost patches every maintained line at once. 10.11.22 -> 10.12.4 crosses 10.12.1, but
+        # 10.11.22 is already past 10.11.4, so the deployment HAD the fix. Counting it credits the
+        # upgrade with work it did not do.
+        rep = run_scan([gh("mattermost/mattermost",
+                           [adv("CVE-1", patched="10.11.4; 10.12.1; 10.5.12")])],
+                       v_from="10.11.22", v_to="10.12.4",
+                       urls=["https://github.com/mattermost/mattermost"])
+        self.assertEqual(rep["fixed_by_this_upgrade"], [])
+
+    def test_a_fix_ABOVE_our_position_on_the_same_line_still_counts(self):
+        rep = run_scan([gh("mattermost/mattermost", [adv("CVE-2", patched="10.11.30; 10.12.1")])],
+                       v_from="10.11.22", v_to="10.12.4",
+                       urls=["https://github.com/mattermost/mattermost"])
+        self.assertEqual(rep["fixed_by_this_upgrade"], ["CVE-2"])
+
+    def test_placeholders_never_feed_this_rule(self):
+        # "7.4.X" parses to a bare 7.4, which would read as "already fixed at 7.4" and silently drop
+        # a real fix — this is exactly how redis CVE-2024-46981 was lost when the rule was added.
+        rep = run_scan([gh("redis/redis", [adv("CVE-3", patched="6.2.X, 7.2.X, 7.4.X")])],
+                       v_from="7.4", v_to="8.10", urls=["https://github.com/redis/redis"])
+        self.assertIn("CVE-3", rep["indeterminate"])
+        self.assertEqual(rep["fixed_by_this_upgrade"], [])
+
+
+class TestChangelogAttribution(unittest.TestCase):
+    """Projects that publish no advisory feed still say which release fixed what — in their changelog."""
+
+    CHANGES = """
+Changes with nginx 1.31.3        11 Aug 2026
+    *) Security: a flaw ... (CVE-2026-60005)
+    *) Security: another ... (CVE-2026-56434)
+
+Changes with nginx 1.31.2        04 Aug 2026
+    *) Security: something ... (CVE-2026-48142)
+
+Changes with nginx 1.31.1        21 Jul 2026
+    *) Security: older ... (CVE-2026-9256)
+
+Changes with nginx 1.20.0        01 Jan 2021
+    *) Security: ancient ... (CVE-2013-2028)
+"""
+
+    def test_each_cve_is_attributed_to_the_release_that_fixed_it(self):
+        got = A._changelog_versions(self.CHANGES)
+        self.assertEqual(got["CVE-2026-60005"], "1.31.3")
+        self.assertEqual(got["CVE-2026-48142"], "1.31.2")
+        self.assertEqual(got["CVE-2026-9256"], "1.31.1")
+        self.assertEqual(got["CVE-2013-2028"], "1.20.0")
+
+    def _scan(self, wfrom, wto):
+        # nginx publishes NO GitHub advisories — the feed is empty and the changelog is everything.
+        return run_scan(
+            [gh("nginx/nginx", [])],
+            [{"source": "https://nginx.org/en/CHANGES", "status": "ok",
+              "cves": sorted(A._changelog_versions(self.CHANGES)),
+              "context": {}, "fixed_in": A._changelog_versions(self.CHANGES)}],
+            images=[("nginx", wfrom, wto)], urls=["https://github.com/nginx/nginx"])
+
+    def test_window_counts_only_the_releases_it_crosses(self):
+        rep = self._scan("1.31.1", "1.31.3")   # 1.31.1 is the FROM, so its CVE is already fixed
+        self.assertEqual(set(rep["fixed_by_this_upgrade"]),
+                         {"CVE-2026-48142", "CVE-2026-56434", "CVE-2026-60005"})
+
+    def test_a_narrower_window_counts_fewer(self):
+        rep = self._scan("1.31.2", "1.31.3")
+        self.assertEqual(set(rep["fixed_by_this_upgrade"]), {"CVE-2026-56434", "CVE-2026-60005"})
+
+    def test_ancient_entries_are_not_swept_in(self):
+        # The changelog lists the project's whole history; only the crossed releases may count.
+        rep = self._scan("1.31.1", "1.31.3")
+        self.assertNotIn("CVE-2013-2028", rep["fixed_by_this_upgrade"])
+
+    def test_evidence_is_recorded(self):
+        rep = self._scan("1.31.1", "1.31.3")
+        self.assertEqual(rep["resolved_by_changelog"]["CVE-2026-60005"], "1.31.3")
+
+
+class TestComposeDerivedWindows(unittest.TestCase):
+    """Windows read off a compose diff, so nobody has to remember which --image args an upgrade needs."""
+
+    OLD = """
+services:
+  app:
+    image: "plausible/analytics:v2.0.0"
+  db:
+    image: pgautoupgrade/pgautoupgrade:18-alpine
+  plausible_events_db:
+    image: clickhouse/clickhouse-server:23.4.2.11-alpine
+volumes:
+  data:
+"""
+    NEW = """
+services:
+  app:
+    image: "ghcr.io/plausible/community-edition:v3.2.1"
+  db:
+    image: pgautoupgrade/pgautoupgrade:18-alpine
+  plausible_events_db:
+    image: clickhouse/clickhouse-server:24.12-alpine
+volumes:
+  data:
+"""
+
+    def _windows(self, old=None, new=None):
+        pages = {"to": new if new is not None else self.NEW,
+                 "from": old if old is not None else self.OLD}
+        with unittest.mock.patch.object(A, "_fetch", lambda u, h=None: pages["to" if "to" in u else "from"]), \
+             unittest.mock.patch.object(A, "_gitea_auth", lambda u: {}):
+            return A.windows_from_compose("http://x/to", "http://x/from")[0]
+
+    def test_app_service_leads_and_sidecars_follow(self):
+        w = self._windows()
+        self.assertEqual(w[0], ("community-edition", "v2.0.0", "v3.2.1"))
+        self.assertIn(("clickhouse-server", "23.4.2.11-alpine", "24.12-alpine"), w)
+
+    def test_unchanged_images_are_not_windows(self):
+        # pgautoupgrade is identical in both; inventing a window for it would be a false count.
+        self.assertNotIn("pgautoupgrade", [n for n, _, _ in self._windows()])
+
+    def test_a_changed_image_REPO_is_still_the_same_service(self):
+        # plausible/analytics -> ghcr.io/plausible/community-edition. Keyed by image repo this reads
+        # as one image vanishing and another appearing, and the app window is lost entirely.
+        w = self._windows()
+        self.assertTrue(any(n == "community-edition" and f == "v2.0.0" for n, f, _ in w))
+
+    def test_no_change_yields_no_windows(self):
+        self.assertEqual(self._windows(old=self.NEW, new=self.NEW), [])
+
+    def test_templated_tags_are_skipped(self):
+        new = self.NEW.replace('ghcr.io/plausible/community-edition:v3.2.1', 'ghost:${IMAGE_VERSION}')
+        self.assertNotIn("ghost", [n for n, _, _ in self._windows(new=new)])
+
+
+class TestImageNameMatching(unittest.TestCase):
+    """An image name and its advisory source rarely spell each other exactly."""
+
+    def test_matches_when_the_image_name_is_LONGER_than_the_source(self):
+        # clickhouse/clickhouse-server vs source ClickHouse/ClickHouse — one-directional matching
+        # dropped this window silently when the key came from a compose file.
+        rep = run_scan([gh("ClickHouse/ClickHouse", [adv("CVE-1", patched="23.10.2.13")])],
+                       images=[("clickhouse-server", "23.4.2.11", "24.12")],
+                       urls=["https://github.com/ClickHouse/ClickHouse"])
+        self.assertIn("github-advisories:ClickHouse/ClickHouse", rep["windows"])
+        self.assertEqual(rep["cve_count_fixed"], 1)
+
+    def test_matches_when_the_image_name_is_SHORTER_than_the_source(self):
+        rep = run_scan([gh("redis/redis", [adv("CVE-2", patched="7.4.1")])],
+                       images=[("redis", "7.4", "8.10")], urls=["https://github.com/redis/redis"])
+        self.assertEqual(rep["cve_count_fixed"], 1)
+
+
 class TestAdjudicationEvidenceAssembly(unittest.TestCase):
    """Pass 2's JUDGEMENT is a model's and not testable; what IS testable is what it gets shown."""

@@ -624,16 +813,22 @@ class TestHistoricReportNumbers(unittest.TestCase):
        for cve in delta:
            self.assertIn("redis", both["cves"][cve]["sources"][0])

-    def test_mailu_scan_finds_zero_and_says_so_knowingly(self):
-        # Published as 2 via the UNION with release-note reading; the scan's own contribution is 0,
-        # and 0 here must mean "checked, none", not "could not check".
+    def test_mailu_finds_the_roundcube_pair_without_an_agent(self):
+        # Published as 2 on 2026-08-07, but only because an agent read the release notes; the scan
+        # itself contributed 0. It now reaches 2 deterministically: the CVEs appear only on
+        # github.com/Mailu/Mailu/releases, and release 2024.06.56 (inside the window) names them.
        rep = self._count("mailu", "2024.06.55", "2024.06.57", [("redis", "8.8.0", "8.10.0")])
-        self.assertEqual(rep["cve_count_fixed"], 0)
-        self.assertTrue(rep["count_known"])
+        self.assertEqual(rep["cve_count_fixed"], 2)
+        self.assertEqual(set(rep["fixed_by_this_upgrade"]), {"CVE-2026-54432", "CVE-2026-54433"})
+        # The redis bump fixes nothing new — every advisory it crosses was fixed at or before 8.6.3.
+        self.assertEqual(rep["cve_count_indeterminate"], 0)

-    def test_keycloak_26_7_0_to_26_7_1_is_7(self):
+    def test_keycloak_26_7_0_to_26_7_1_is_12(self):
+        # Was 7 while only the GHSA feed was consulted. keycloak lists five more CVEs in the 26.7.1
+        # release notes' fixed-issues section that it never filed as advisories — the gitea pattern.
        rep = self._count("keycloak", "26.7.0", "26.7.1")
-        self.assertEqual(rep["cve_count_fixed"], 7)
+        self.assertEqual(rep["cve_count_fixed"], 12)
+        self.assertEqual(len(rep.get("resolved_by_release_notes") or {}), 5)


 def _main():
@@ -18,6 +18,11 @@
 - AUTO_MIGRATIONS=true means DB migrations run automatically on backend startup. No manual step needed.
 - Minio tag uses a date-based RELEASE.YYYY-MM-DDTHH-MM-SSZ format — abra cannot parse it for upgrades;
  check manually on https://github.com/minio/minio/releases.
+- **2026-08-14: Minio stopped publishing Docker images after RELEASE.2025-09-07T16-13-09Z.**
+  GitHub has a newer release (`RELEASE.2025-10-15T17-29-55Z`, published 2025-10-16, with CVE fix
+  GHSA-jjjj-jwhf-8rgr), but the Docker image was never pushed to Docker Hub (returns 404; release
+  notes say "clone the source and build the latest container"). quay.io checked — only 2022-era
+  tags. As of this date, `RELEASE.2025-09-07T16-13-09Z` IS the newest available Docker image.
 - v5.2.0 adds two optional new env vars: DOCUMENT_ALL_ENDPOINT_ENABLED and OIDC_OP_USER_ENDPOINT_FORMAT.
  Both are backward-compatible (no action required for existing deployments).
 - Recipe version label convention: 0.X.Y+vA.B.C where A.B.C is the impress version.
@@ -6,6 +6,14 @@
 | postgres | postgres                                  | https://github.com/postgres/postgres              | https://www.postgresql.org/docs/release/                          |

 ## Standing notes
+  - **2026-08-28 re-check** (endoflife.date/api/mattermost.json 2026-08-28; Docker Hub; GitHub
+    releases): **11.7.10** (released 2026-08-26, "various bug fixes", not prerelease) is the newest
+    11.7.x ESR patch — the ESR/LTS line has NOT moved, still **11.7** (EOL **2027-05-15**). This run
+    bumps open PR #2 **11.7.9 → 11.7.10** (ESR patch within the same line). 11.8/11.9/11.10 remain
+    innovation releases (EOL 2026-09-15 / 10-15 / 11-15), NOT ESR — do NOT target; 10.12 expired.
+    postgres **15-alpine HELD** (DB-major 16/17/18 out of scope, plain-postgres recipe w/o
+    pgautoupgrade: operator dump/restore needed). git.autonomic.zone mirror main reconciled to
+    coopcloud (86e0d0b5 = 10.11.22, upstream still pins the EXPIRED 10.11 ESR).
 - mattermost-lts tracks the **11.7 ESR (Extended Support Release)** = Mattermost's LTS line. The "LTS"
  concept in Mattermost is **ESR**; there is NO "10.x LTS line". Release calendar (endoflife.date/mattermost):
  - **10.11** = previous ESR, security support **ends 2026-08-15** (~7 weeks) — migrated away 2026-06-19
@@ -75,3 +83,32 @@
  recreate DB, reimport dump. `DROP DATABASE WITH (FORCE)` requires PostgreSQL 13+ — safe on postgres:15-alpine.
  The previous inline-label approach (no restore hook) was a defect: raw PGDATA restore without a reload
  was a silent no-op. Fixed in PR #2 (restore fix cherry-picked from PR #1 ci/pg-restore).
+
+  - **2026-08-21 re-check** (endoflife.date/api/mattermost.json 2026-08-17; Mattermost docs; GitHub):
+    **11.7 ESR is the only current ESR/LTS** — latest patch **11.7.9** (2026-08-13, "Low to High
+    security fixes"), EOL 2027-05-15. **10.11 ESR ENDED 2026-08-15** — upstream main still pins
+    10.11.22 (now on an EXPIRED ESR). 11.10/11.9/11.8 are innovation (EOL 2026-11-15/10-15/09-15),
+    NOT LTS. The `-lts` recipe must track 11.7. This run **redirects open PR #2 away from the
+    innovation 11.10.0 onto 11.7.9** (the correct LTS/ESR; ESR→ESR 10.11→11.7 is "fully supported
+    and tested" per Mattermost docs — no staging hop). Refer to the 2026-08-15 report where PR #2
+    was on 11.10.0 (operator chose innovation line that week); operator should now decide the
+    `-lts` line. postgres 15-alpine still HELD (DB-major out of scope).
+
+- **2026-08-31 re-check** (endoflife.date/api/mattermost.json 2026-08-31; Docker Hub tag list;
+    GitHub release `v11.7.10`): **11.7 ESR line UNCHANGED** — latest patch still **11.7.10**
+    (released 2026-08-26, "various bug fixes"), EOL 2027-05-15. PR #2 already carries 11.7.9 →
+    11.7.10; the wildcard cert is RENEWED (valid to 2026-11-29), so this run just RE-VERIFIES PR #2
+    via `!testme` (the 2026-08-28 run's `!testme` was infra-blocked). 11.8.5 / 11.9.1 / 11.10.1
+    remain innovation (EOL 2026-09-15 / 10-15 / 11-15), NOT ESR — do NOT target; 10.11 ESR ended
+    2026-08-15 (upstream main still pins 10.11.22 = EXPIRED ESR → the 10→11 ESR move PR #2 carries
+    remains required; Mattermost docs: ESR→ESR is "fully supported and tested"). postgres 15-alpine
+    still HELD (DB-major out of scope, operator dump/pg_upgrade).
+
+## NVD CPE fallback
+This project publishes nothing machine-readable we can reach — no GitHub advisory feed,
+no release-attributable changelog — so its CVE count was `?` (nothing measured). NVD is
+CPE-indexed and carries structured version ranges, so it can answer where the vendor
+cannot. It LAGS the vendor, so it is a fallback, never the primary source.
+
+- nvd-cpe: mattermost-team-edition = cpe:2.3:a:mattermost:mattermost_server:*:*:*:*:*:*:*:*
+- nvd-cpe: postgres = cpe:2.3:a:postgresql:postgresql:*:*:*:*:*:*:*:*
@@ -0,0 +1,29 @@
+# Upstream sources — mumble
+
+| service | image | source repo | releases / changelog |
+|---------|-------|-------------|----------------------|
+| app | mumblevoip/mumble-server | https://github.com/mumble-voip/mumble | https://github.com/mumble-voip/mumble/releases |
+| web | rankenstein/mumble-web | https://github.com/rankenstein/mumble-web | https://github.com/rankenstein/mumble-web/releases |
+
+## Standing notes
+- This file was **missing entirely** until 2026-08-11. Without it the advisory scan had no source to
+  query, and still printed "0 identified by the deterministic scan" — which was then published as a
+  clean `0` in the 2026-08-11 CVE check. The scan now refuses to emit a count when it has no usable
+  source (it reports UNKNOWN), and `audit-sources.py` flags a missing registry file directly.
+- `mumblevoip/mumble-server` tracks the upstream server releases and DOES publish GitHub security
+  advisories, so it is the recipe's primary CVE source.
+- `rankenstein/mumble-web` is a **fork** of the original `Johni0702/mumble-web`, which has been
+  dormant since 2023-05. The fork itself last pushed 2023-07 and its Docker tag `0.5` was last built
+  well over five years ago. Neither is archived, but treat the web client as effectively unmaintained:
+  if a CVE lands there, expect no upstream fix and plan a replacement rather than an upgrade.
+- The server image tag is `v<version>-<build>` (e.g. `v1.6.870-4`); the trailing number is the image
+  build, not an app version, and moves independently of upstream releases — `abra recipe upgrade`
+  reports "no new versions" for it, so use `resolve-images.py` to see those bumps.
+
+## NVD CPE fallback
+This project publishes nothing machine-readable we can reach — no GitHub advisory feed,
+no release-attributable changelog — so its CVE count was `?` (nothing measured). NVD is
+CPE-indexed and carries structured version ranges, so it can answer where the vendor
+cannot. It LAGS the vendor, so it is a fallback, never the primary source.
+
+- nvd-cpe: mumble-server = cpe:2.3:a:mumble:mumble:*:*:*:*:*:*:*:*
@@ -50,3 +50,90 @@
 - 2026-08-07 run: operator directed 2.33.3 -> 2.34.2 (the newest). The whole 2.34.x line is still
  marked Pre-release on GitHub (2.33.5 holds the Latest badge); flagged in the PR body. No breaking
  changes across 2.33.3 -> 2.34.2; rolling upgrade safe (TypeORM migrations auto-run on boot).
+- 2.35.0 (2026-08-11, Pre-release): major feature release — self-hosted AI Assistant onboarding,
+  Simplified Custom Auth credentials, Agent Builder test runs + HITL, Discord agent chat channel,
+  local agent token counting, **VM expression engine now the default** (was opt-in), MCP SDK v2
+  migration + MCP 2026-07-28 discovery handshake, Kafka Node v2, Salesforce OAuth2 JWT, GitHub
+  dispatch timeout, X/Twitter Node OAuth2/API migrated to x.com, Azure Key Vault configurable
+  endpoints, Postgres-version startup warning, workflow review improvements (diffs, metadata, version
+  descriptions), and numerous core/editor bugfixes. No breaking compose/config/migration changes.
+- 2.35.1 (2026-08-12, Pre-release): 2 core bugfixes — data-tables resume scope, TLS options per hop
+  through a proxy.
+- 2.35.2 (2026-08-13, Pre-release): 1 core bugfix — report real activation mode for triggers via
+  publication outbox. **Deployed on cc-ci 2026-08-15**: 2.34.4→2.35.2, TypeORM migrations clean,
+  editor served HTTP 200. No breaking changes, no N8N_* env renames, no required operator action.
+- 2.35.3 (2026-08-14, Pre-release): bugfixes (Google Ads v21→v25 API migration, MS Teams OAuth scope
+  restore, workflow publication outbox abort deadline) + feature (skip update approval for workflows
+  from same Instance AI session). Not deployed (2.35.2 was the survey target).
+- 2.36.0 (2026-08-18, Pre-release; 2.36.1 was briefly tagged then withdrawn — 2.36.2 compares
+  2.36.0...2.36.2): a large feature minor. Core/ai-builder/editor bugfixes + new features (Agent
+  sandboxes + writable workspaces, MCP folder/move/read-data-table tools, Schedule Trigger missed-
+  execution grace period + catch-up coalescing, Confluence Cloud OAuth2 + page get sub-tree, MiniMax
+  M3/H3, LangSmith session debug export, role-mapping/roles API endpoints, n8n User Auth GA, etc.).
+  No breaking compose/env/migration changes, no `N8N_*` env renames; the deprecation notices
+  (Array.merge→Array.mergeIntoObject in expressions, workflow tags API migration) are API-level only.
+  2.36.2 (2026-08-19, patch): 3 fixes (decorator-route request schema in /discover, end-user
+  credential resolution in node parameters, editor agent-preview alignment). 2.36.3 (2026-08-20,
+  patch): 4 core/editor fixes (task-runners only restart when slow, expression engine init on
+  expression commands, test-webhook isolate release after teardown, Edit Fields type selector styles).
+- 2026-08-21 run: planning 2.34.4 → 2.36.3 (newest tag abra lists). Whole 2.36.x line still
+  Pre-release on GitHub (2.33.5/Latest was superseded by 2.35.x line). Rolling upgrade safe, sqlite
+  mode, TypeORM migrations auto-run on boot. No operator action required.
+- 2.36.4 (2026-08-21, Pre-release): 3 core fixes (env normalization before schema parsing; chat and
+  MCP trigger auth-mode validation for end-user credentials; wrap trigger closeFunction in an
+  expression isolate).
+- 2.36.5 (2026-08-21, Pre-release): 1 core fix (raise AI Assistant model verification token limit).
+- 2.36.6 (2026-08-24): 1 editor fix (personalization survey / community registration modal on
+  Instance AI landing page).
+- 2.36.7 (2026-08-25, **current Stable / Latest badge**; `stable` tag points here): API fix (adapt
+  binary data endpoint), API cross-backport (require a JSON content type on decorator routes that
+  take a body), core fixes (reconcile license auto-renewal on startup leader takeover; route OAuth2
+  token requests through environment proxies), feature (remove offset query param from list workflow
+  history endpoint). This is the newest stable-line release; the 2.37.x line is still pre-release.
+- 2.37.0 (2026-08-25, Pre-release): a large feature minor. Agent Builder / AIA / core/editor
+  bugfixes + features (Slack Agent view, prompt caching for Anthropic, SharePoint node v2 = default,
+  form-trigger OAuth2 (added then reverted), role-mapping rule CRUD endpoints, execution
+  startedAfter/startedBefore filters, Agent/community-package hardening: serialize community
+  package installs + non-destructive downloads + cold-boot storm reduction, MCP tool schemas as JSON
+  Schema 2020-12, distroless n8n image variant). **Deprecation: the "Any workflow" caller policy**
+  (API/config-level — flag for operators using execution-caller policies). Two enforced API behavior
+  changes: **require JSON content type on decorator body routes** and **"adapt binary data endpoint"**
+  (HTTP clients sending a body to those routes must send `Content-Type: application/json`). No
+  breaking compose/env/migration changes, no `N8N_*` env renames.
+- 2.37.1 (2026-08-25, Pre-release): 1 core fix (route OAuth2 token requests through env proxies in
+  every process).
+- 2.37.3 (2026-08-27, Pre-release; 2.37.2 was withdrawn): 2 runner-image fixes (copy glibc into
+  distroless runners image as a matched pair; stage libatomic into the distroless runners image) —
+  relevant if using n8n's community/distroless runner image.
+- 2026-08-28 run: PR #7 extended 2.34.4 → **2.37.3** (newest tag abra lists; 2.37.2 withdrawn, that
+  tag does not exist). 2.36.7 holds the Stable/Latest badge; 2.37.x remains Pre-release on GitHub,
+  consistent with the tracking-the-newest-tag precedent (2.34.2/2.35.2/2.36.3 were pre-release when
+  taken). Rolling upgrade safe (sqlite, TypeORM auto-migrate on boot). Operator flag: the two 2.37.0
+  API behavior changes (JSON content-type on decorator body routes; binary-data endpoint adapt) only
+  affect HTTP API callers. Recommended release: `-y` (feature minor).
+  NOTE: PR #7's `!testme` on this date was BLOCKED — `*.ci.commoninternet.net` wildcard cert expired
+  (bridge: `CERTIFICATE_VERIFY_FAILED certificate has expired`). UNVERIFIED; operator renewed the cert
+  (now valid to 2026-11-29). The 2.37.3 work must be re-verified by a fresh `!testme`.
+- 2.34.5 (2026-08-12, patch): 1 core fix (apply TLS options per hop through a proxy).
+- 2.34.6 (2026-08-14, patch): 3 fixes (core `continueErrorOutput` `details` field; Google Ads sunset
+  v21→v25 API migration; MS Teams Restore `Group.ReadWrite.All` OAuth2 scope — matches the 2.35.3
+  in-progress backports).
+- 2.35.4/2.35.5/2.35.6/2.35.7 (2026-08-19..21, patches): API schema for decorator routes in /discover;
+  Google Ads v25 view metric rename; task-runners not restarted when only slow; expression engine
+  init on expression commands; test-webhook isolate release after teardown; env normalization before
+  schema parsing; end-user credential resolution; trigger closeFunction isolate; AI Assistant token
+  limit raise. All patch bugfixes, no breaking changes.
+- 2.36.8 (2026-08-28, patch): 1 core fix (domain-restricted credential usable in its own node) —
+  cross-backport of the 2.37.4 fix into the stable line.
+- 2.36.9 (2026-08-31, patch; **now the Stable/Latest badge** — `stable` tag): 1 core fix (apply proxy
+  environment variables consistently across packages and processes, cross-backport of 2.37.1/2.37.3).
+- 2.37.4 (2026-08-28, Pre-release): 1 core fix (allow a domain-restricted credential to work in its
+  own node).
+- 2.37.5 (2026-08-31): **withdrawn** — no plain `2.37.5` tag on Docker Hub (only partial
+  `2.37.5-amd64/-arm64/-pc` artifact tags, no manifest). Skip entirely, like 2.37.2/2.36.1.
+- 2.37.6 (2026-08-31, Pre-release; **newest 2.37.x tag**): 1 core fix (thread execution id through
+  dynamic-credential storage).
+- 2026-08-31 run: PR #7 extended 2.34.4 → **2.37.6** (newest tag abra lists = 2.37.6/2.37.4/2.37.3/…;
+  2.37.5 withdrawn). 2.36.9 holds the Stable/Latest badge; 2.37.x remains Pre-release on GitHub
+  (consistent precedent). Re-verified 2.37.3→2.37.6 (pure core bugfixes), no breaking changes beyond
+  the already-flagged 2.37.0 API behavior pair. Rolling upgrade safe. Recommended release: `-y`.
@@ -4,6 +4,7 @@
 |---------|-------|-------------|----------------------|
 | app | wordpress | https://github.com/WordPress/WordPress | https://wordpress.org/news/category/releases/ (image: https://hub.docker.com/_/wordpress) |
 | db | mariadb | https://github.com/MariaDB/server | https://mariadb.com/kb/en/release-notes/ (image: https://hub.docker.com/_/mariadb) |
+| ftp (optional overlay) | atmoz/sftp | https://github.com/atmoz/sftp | https://github.com/atmoz/sftp/releases (image: https://hub.docker.com/r/atmoz/sftp) |

 ## Standing notes
 - Enrolled 2026-08-03 (operator request). Recipe = coopcloud `wordpress`; mirror
@@ -22,3 +23,7 @@
  image). mariadb major bumps (e.g. 12.x→13.x) need the usual dump/restore caution if the
  recipe ever pins a non-`mariadb`-auto-upgrading setup; the official image handles minor
  bumps in place.
+- `atmoz/sftp` is pinned as the floating `alpine` tag (updated 2026-07-28, current as of
+  2026-08-21) — not a semver abra can evaluate, so it never contributes an upgrade candidate;
+  verified current by direct Docker Hub tag check. Only bump if a real newer supported tag
+  exists and the app needs it.
@@ -31,7 +31,7 @@ matrix-synapse    weekly
 mattermost-lts    weekly
 mumble            weekly
 n8n               weekly
-plausible         weekly
+plausible         external   # maintained elsewhere — used/tested by cc-ci but NOT weekly-upgraded
 uptime-kuma       external   # maintained elsewhere — used/tested by cc-ci but NOT weekly-upgraded
 wordpress         weekly
 ```
@@ -1,5 +1,28 @@
 {
  "nodes": {
+    "cc-ci": {
+      "inputs": {
+        "nixpkgs": [
+          "nixpkgs"
+        ],
+        "sops-nix": [
+          "sops-nix"
+        ]
+      },
+      "locked": {
+        "lastModified": 1788812004,
+        "narHash": "sha256-Vc7RSeqFHCwlVRhnEEjavuJoyIID+RQdJSygHNA8s8Y=",
+        "ref": "refs/heads/main",
+        "rev": "f6dbfa368995f4d45de09f4052631fd433c87d5b",
+        "revCount": 1533,
+        "type": "git",
+        "url": "https://git.autonomic.zone/recipe-maintainers/cc-ci.git"
+      },
+      "original": {
+        "type": "git",
+        "url": "https://git.autonomic.zone/recipe-maintainers/cc-ci.git"
+      }
+    },
    "nixpkgs": {
      "locked": {
        "lastModified": 1785734586,
@@ -18,6 +41,7 @@
    },
    "root": {
      "inputs": {
+        "cc-ci": "cc-ci",
        "nixpkgs": "nixpkgs",
        "sops-nix": "sops-nix"
      }
@@ -1,32 +1,52 @@
 {
-  description = "cc-ci-orchestrator — NixOS host for the cc-ci loops runtime (Builder/Adversary/Watchdog)";
+  description = "cc-ci-orchestrator — the cc-ci orchestrator (loops, steering session, weekly upgrader) and the NixOS host it shares with the cc-ci CI server";

  inputs = {
-    # Follow the current stable release channel (operator 2026-08-01), was a hard rev pin at
-    # nixpkgs 24.11 (50ab7937, 2025-06-30) kept "the same as the cc-ci server". This host runs
-    # agents/tmux/nginx/docker, not recipe CI, so it does not need to match that server — and a
-    # frozen rev only accrues unpatched CVEs. `nix flake update` now actually moves.
+    # Stable release channel (operator 2026-08-01). `nix flake update` moves it; the cc-ci input
+    # below FOLLOWS it, so one nixpkgs builds the whole combined host and CVEs get patched once.
    nixpkgs.url = "github:NixOS/nixpkgs/nixos-26.05";

-    # sops-nix follows nixpkgs below, so it no longer needs its own matching pin.
    sops-nix.url = "github:Mic92/sops-nix";
    sops-nix.inputs.nixpkgs.follows = "nixpkgs";
+
+    # The cc-ci CI server, as a NixOS module (`nixosModules.cc-ci-server`). HTTPS, anonymous read:
+    # nix evaluates every input for every output, so the input must be fetchable without
+    # credentials. The private secrets submodule is deliberately NOT fetched through this input —
+    # the host reads the deployed --recursive checkout's secrets.yaml at activation instead
+    # (`cc-ci.sopsFile`). Both `follows` are REQUIRED: without them cc-ci's own nixpkgs/sops-nix
+    # pins would produce a second sops-nix module tree and a second nixpkgs in one system.
+    cc-ci.url = "git+https://git.autonomic.zone/recipe-maintainers/cc-ci.git";
+    cc-ci.inputs.nixpkgs.follows = "nixpkgs";
+    cc-ci.inputs.sops-nix.follows = "sops-nix";
  };

-  outputs = { nixpkgs, sops-nix, ... }:
+  outputs = { self, nixpkgs, sops-nix, cc-ci, ... }:
    let
      system = "x86_64-linux";
    in
    {
-      # Hetzner cpx11 host (nixos-infect generated hardware.nix + orchestrator config).
-      # Provision with terraform/ then run Stage 2 per terraform/README.md.
-      nixosConfigurations.cc-ci-orchestrator-hetzner = nixpkgs.lib.nixosSystem {
-        inherit system;
-        modules = [
-          sops-nix.nixosModules.sops
-          ./nix/hosts/cc-ci-orchestrator-hetzner/hardware.nix
-          ./nix/hosts/cc-ci-orchestrator-hetzner/configuration.nix
-        ];
+      nixosModules = {
+        # The orchestrator itself: loops supervisor, steering session, weekly/hourly timers.
+        cc-ci-orchestrator = ./nix/modules/cc-ci.nix;
+        # The host contract those units assume: loops user, claude/opencode CLIs, opencode web
+        # server + tailnet UI, nix-ld, tool set, `ssh cc-ci` config.
+        orchestrator-host = ./nix/modules/orchestrator-host.nix;
+        # Old name of cc-ci-orchestrator, kept while notplants-nix still imports it (2026-09).
+        cc-ci = ./nix/modules/cc-ci.nix;
+      };
+
+      nixosConfigurations = {
+        # THE live host: cc-ci CI server + cc-ci orchestrator on one Hetzner cpx32-class box
+        # (195.201.88.249, since 2026-09). README.md is the deploy guide.
+        cc-ci = nixpkgs.lib.nixosSystem {
+          inherit system;
+          modules = [
+            cc-ci.nixosModules.cc-ci-server
+            self.nixosModules.cc-ci-orchestrator
+            self.nixosModules.orchestrator-host
+            ./nix/hosts/cc-ci/configuration.nix
+          ];
+        };
      };
    };
 }
@@ -10,9 +10,10 @@ metadata:
 The cc-ci orchestrator (loops + watchdog + this session) runs on a **Hetzner cpx22** as of
 2026-05-31, replacing the Incus VM (100.116.55.106).

- Hetzner server **134487234**, public **168.119.126.100**, tailnet **cc-ci-orchestrator-1** @
-  **100.84.190.30**. Flake host **cc-ci-orchestrator-hetzner**.
- Rebuild: `sudo nixos-rebuild switch --flake .#cc-ci-orchestrator-hetzner` from `/srv/cc-ci-orch`
+- Since 2026-09-07: ONE Hetzner host for CI server + orchestrator, public **195.201.88.249**,
+  tailnet **cc-ci**, flake host **`.#cc-ci`** (this repo). Before: orchestrator on Hetzner
+  134487234 (168.119.126.100 / 100.84.190.30, `cc-ci-orchestrator-hetzner`), shared with notplants.
+- Rebuild: `sudo nixos-rebuild switch --flake .#cc-ci` from `/srv/cc-ci-orch`
  (`/srv/cc-ci` is a symlink to it). The Bash tool runs as user **loops** (uid 1000, passwordless
  sudo) — plain `nixos-rebuild switch` fails on the profile symlink; use `sudo`.
 - Reboot-resilience: `cc-ci-loops.service` is **enabled** (wantedBy multi-user.target); ExecStartPre
@@ -23,4 +24,4 @@ The cc-ci orchestrator (loops + watchdog + this session) runs on a **Hetzner cpx
  identity unknown". Set per-repo to match prior commits: `autonomic-bot
  <autonomic-bot@git.autonomic.zone>`.

-Full record: `cc-ci-plan/plan-orchestrator-hetzner-migration.md`.
+Full record: `archive/plans/plan-orchestrator-hetzner-migration.md`.
@@ -0,0 +1,68 @@
+# cc-ci — ONE Hetzner Cloud host running both the cc-ci CI server and the cc-ci orchestrator.
+#
+# This file is only what is physical or identity about the machine: hardware, networking, the
+# tailscale node, root SSH keys, swap, stateVersion. Everything functional comes from modules:
+#   cc-ci.nixosModules.cc-ci-server        recipe-maintainers/cc-ci — swarm, traefik, drone,
+#                                          runner, bridge, dashboard, reports, acme-dns, harness
+#   self.nixosModules.cc-ci-orchestrator   nix/modules/cc-ci.nix — loops, orchestrator, timers
+#   self.nixosModules.orchestrator-host    nix/modules/orchestrator-host.nix — loops user, CLIs
+# See README.md for provisioning (Hetzner Debian → nixos-infect → this flake) and staging.
+{ lib, pkgs, ... }:
+{
+  imports = [
+    ./hardware.nix
+    ./networking.nix
+  ];
+
+  networking.hostName = "cc-ci";
+
+  # ---- cc-ci server identity --------------------------------------------------------------
+  # Public address: acme-dns binds to it and publishes it as the `ns-acme` glue record; the
+  # Gandi A records for ci / *.ci / ns-acme .commoninternet.net point here.
+  cc-ci.publicIPv4 = "195.201.88.249";
+  # cc-ci is a plain flake input here (no private submodule), so the sops file is the one in
+  # the deployed --recursive checkout the weekly sweep runs from (README "Stage the workspace").
+  cc-ci.sopsFile = "/etc/cc-ci/secrets/secrets.yaml";
+
+  # ---- orchestrator identity --------------------------------------------------------------
+  # The CI server is this very host, so `ssh cc-ci` goes to loopback (the module default).
+  cc-ci-orchestrator.ciSshHost = "127.0.0.1";
+
+  # ---- tailscale — auth key staged out of band at /etc/ts-auth-key -----------------------
+  services.tailscale = {
+    enable = true;
+    authKeyFile = "/etc/ts-auth-key";
+    extraUpFlags = [ "--hostname=cc-ci" ];
+  };
+
+  # ---- ssh ----------------------------------------------------------------------------------
+  services.openssh = {
+    enable = true;
+    settings.PermitRootLogin = "yes";
+  };
+  # Root keys: PUBLIC keys, tracked deliberately in ./ssh-keys (one per line, blank lines ok).
+  users.users.root.openssh.authorizedKeys.keys =
+    builtins.filter (s: s != "") (lib.splitString "\n" (builtins.readFile ./ssh-keys));
+  # The loops user can also be reached directly (same keys) — handy for rsync of its workspace.
+  users.users.loops.openssh.authorizedKeys.keys =
+    builtins.filter (s: s != "") (lib.splitString "\n" (builtins.readFile ./ssh-keys));
+
+  # ---- firewall -------------------------------------------------------------------------------
+  # 80/443 (traefik) and 53 (acme-dns) are opened by the cc-ci-server module. The tailscale
+  # interface is trusted, which is what makes the opencode UI on 8443 tailnet-only.
+  networking.firewall = {
+    enable = true;
+    trustedInterfaces = [ "tailscale0" ];
+    allowedTCPPorts = [ 22 ];
+  };
+  networking.nameservers = [ "1.1.1.1" "8.8.8.8" ];
+
+  # ---- memory: 8 GB RAM shared by the swarm (recipe deploys) and 3–6 agent sessions ---------
+  swapDevices = [ { device = "/swapfile"; size = 8192; } ];
+
+  # ssh client for root (the orchestrator's `ssh cc-ci` goes through the loops user's own config).
+  environment.systemPackages = [ pkgs.openssh ];
+
+  # Fresh NixOS 26.05 install (nixos-infect, 2026-09-07). Never change this on an existing host.
+  system.stateVersion = "26.05";
+}
@@ -0,0 +1,19 @@
+# Generated by nixos-infect on this machine (2026-09-07), captured verbatim per README §3.
+# The ESP UUID is specific to THIS server; a new server gets a new file.
+{ modulesPath, ... }:
+{
+  imports = [ (modulesPath + "/profiles/qemu-guest.nix") ];
+  boot.loader = {
+    efi.efiSysMountPoint = "/boot/efi";
+    grub = {
+      efiSupport = true;
+      efiInstallAsRemovable = true;
+      device = "nodev";
+    };
+  };
+  fileSystems."/boot/efi" = { device = "/dev/disk/by-uuid/E079-7D41"; fsType = "vfat"; };
+  boot.initrd.availableKernelModules = [ "ata_piix" "uhci_hcd" "xen_blkfront" "vmw_pvscsi" ];
+  boot.initrd.kernelModules = [ "nvme" ];
+  fileSystems."/" = { device = "/dev/sda1"; fsType = "ext4"; };
+  
+}
@@ -0,0 +1,38 @@
+# Generated by nixos-infect on this machine (2026-09-07), captured per README §3, with ONE edit:
+# `defaultGateway` as an attrset WITH `interface = "eth0"`. The generated bare-string form leaves
+# NixOS ≥25.05 without a default route (the host boots and is unreachable) — see README §2.
+{ lib, ... }: {
+  # This file was populated at runtime with the networking
+  # details gathered from the active system.
+  networking = {
+    nameservers = [ "2a01:4ff:ff00::add:2"
+ "2a01:4ff:ff00::add:1"
+ "185.12.64.2"
+ ];
+    defaultGateway = { address = "172.31.1.1"; interface = "eth0"; };
+    defaultGateway6 = {
+      address = "fe80::1";
+      interface = "eth0";
+    };
+    dhcpcd.enable = false;
+    usePredictableInterfaceNames = lib.mkForce false;
+    interfaces = {
+      eth0 = {
+        ipv4.addresses = [
+          { address="195.201.88.249"; prefixLength=32; }
+        ];
+        ipv6.addresses = [
+          { address="2a01:4f8:1c1c:a9b::1"; prefixLength=64; }
+{ address="fe80::2ff8:e3ea:bbb8:aa39"; prefixLength=64; }
+        ];
+        ipv4.routes = [ { address = "172.31.1.1"; prefixLength = 32; } ];
+        ipv6.routes = [ { address = "fe80::1"; prefixLength = 128; } ];
+      };
+      
+    };
+  };
+  services.udev.extraRules = ''
+    ATTR{address}=="92:00:09:d5:ec:0d", NAME="eth0"
+    
+  '';
+}
@@ -0,0 +1,10 @@
+ssh-ed25519 AAAAC3NzaC1lZDI1NTE5AAAAIGZGp/DQTFuD1GvsyTzCVBUTmoWqcb5T+Z7zZo5nYLXO
+ssh-rsa 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 trav@trav480sweet
+ssh-rsa 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 aadil@t480
+ssh-ed25519 AAAAC3NzaC1lZDI1NTE5AAAAIMyHSi12R0+HCVBz7+d9fyOBnoJi8Nsj5D7vQ9UQO8a5
+ssh-ed25519 AAAAC3NzaC1lZDI1NTE5AAAAIJVlfoLBPseQ9fA9534KmRg2KWcksKZGzAJIpHJ2JpsI mfowler.email@protonmail.com
+ssh-ed25519 AAAAC3NzaC1lZDI1NTE5AAAAIAQFuqUB2qNZSDNjDsjjhVA/WnnQNVAMmsUscW6OgMDN
+ssh-ed25519 AAAAC3NzaC1lZDI1NTE5AAAAIHOcLo0YBa0UYi7i/l8K/Y/7cF2OclmDqSTlAsHM0dOS notplants-orchestrator
+ssh-ed25519 AAAAC3NzaC1lZDI1NTE5AAAAIMniNzAzuI527bfk/EipqFILFayUCwYXDoZ3R7+QgYq6
+ssh-ed25519 AAAAC3NzaC1lZDI1NTE5AAAAIOk8NaeBdPbS2gfUvbny8h0AkZlVjGYHzx4QPXSJ38gd claude@claude-vm
+ssh-ed25519 AAAAC3NzaC1lZDI1NTE5AAAAIAcyTGb/wVgdhg5oBCZZvBaR1RuUQRY/3WHnOQpNDCsp claude-cc-ci-sandbox@20260526
@@ -0,0 +1,157 @@
+# cc-ci.nix — the cc-ci ORCHESTRATOR: the Builder/Adversary loops supervisor, the operator's
+# steering session, and the weekly-upgrade + hourly-supervisor timers. Nothing else.
+#
+# Exported from this repo's flake as `nixosModules.cc-ci-orchestrator` (and, for the host that
+# used to import it under the old name, `nixosModules.cc-ci`). Split out of the shared agent
+# host config on 2026-08-20; since 2026-09 it runs on the same Hetzner host as the CI server
+# itself (`#cc-ci` in flake.nix), next to recipe-maintainers/cc-ci's `nixosModules.cc-ci-server`.
+#
+# All of it assumes the cc-ci workspaces exist on the host:
+#   /srv/cc-ci        the loops workspace (+ .cc-ci-logs, upgrader.env) — a symlink to
+#   /srv/cc-ci-orch   this repo (the orchestrator's own working dir), with cc-ci/ checked out
+# and that a `loops` user, tmux, python3 and the standalone claude/opencode CLIs are present —
+# those are host concerns, provided by nix/modules/orchestrator-host.nix, not by this module.
+{ config, pkgs, lib, ... }:
+{
+  # cc-ci-loops supervisor — workspace staged 2026-05-31, so ENABLED for reboot-resilience.
+  systemd.services.cc-ci-loops = {
+    description = "cc-ci Builder/Adversary loops + watchdog (launch.sh start)";
+    wantedBy = [ "multi-user.target" ];  # enabled after workspace staged (Hetzner cutover)
+    after = [ "network-online.target" "tailscaled.service" "claude-install.service" ];
+    wants = [ "network-online.target" ];
+    serviceConfig = {
+      # KillMode=process: this unit only LAUNCHES the tmux server, it does not own it. With the
+      # default (control-group) systemd kills every leftover process in the cgroup when the unit
+      # stops — and since one tmux server hosts every agent session on this host, a rebuild that
+      # merely touched this unit wiped all of them (operator 2026-08-01). Only the (already
+      # exited) main process is killed now; `systemctl stop` therefore does NOT tear down agents.
+      KillMode = "process";
+      Type = "oneshot"; RemainAfterExit = true;
+      User = "loops"; Group = "users";
+      WorkingDirectory = "/srv/cc-ci/cc-ci";
+      # Append one line to REBOOTS.md per genuine reboot (boot_id-gated; not on manual restart).
+      ExecStartPre = "${pkgs.bash}/bin/bash /srv/cc-ci/cc-ci-plan/reboot-log.sh";
+    };
+    # CLAUDE_BIN points at the standalone CLI installed by claude-install.service; the loops
+    # backend defaults to claude (persisted in .loop-backend). Without this, launch.py's preflight
+    # `which(claude)` fails because the systemd `path` below has no /home/loops/.local/bin.
+    environment = { RESUME_PHASE = "1"; HOME = "/home/loops"; CLAUDE_BIN = "/home/loops/.local/bin/claude"; };
+    path = [ pkgs.bash pkgs.tmux pkgs.git pkgs.python3 pkgs.openssh pkgs.nettools ];
+    script = ''
+      # Put the standalone claude/opencode binaries on PATH. On a cold boot this is the env the
+      # tmux server (and thus every agent session) inherits, so bare `claude` resolves everywhere.
+      export PATH="/home/loops/.local/bin:$PATH"
+      [ -x /srv/cc-ci/cc-ci-plan/launch.sh ] && /srv/cc-ci/cc-ci-plan/launch.sh start || \
+        echo "workspace not staged yet — skipping loop start"
+    '';
+  };
+
+  # cc-ci-orchestrator supervisor — the operator's steering session. Same shape as
+  # lichen-orchestrator / project-orchestrator above: this unit only LAUNCHES the orchestrator's
+  # tmux session via the agent-orchestrator harness (cc-ci-plan/agents.py); it does not own the
+  # session or the tmux server. The orchestrator agent is declared in cc-ci-plan/agents.toml
+  # (backend/model chosen there — Claude Code under Remote Control since 2026-09-07; before that
+  # opencode/glm-5.2 attached to the shared opencode web server, opencode-web.service in
+  # orchestrator-host.nix, which the upgrader still uses). The harness watchdog (started by
+  # `agents.py up`) keeps it alive: heal-only (no stall reboots — a persistent supervisor must not
+  # be killed just for idling). Added 2026-08-03 for reboot-resilience.
+  systemd.services.cc-ci-orchestrator = {
+    description = "cc-ci orchestrator (operator steering session) — agents.py up orchestrator";
+    wantedBy = [ "multi-user.target" ];
+    after = [ "network-online.target" "tailscaled.service" "opencode-web.service" ];
+    wants = [ "network-online.target" ];
+    serviceConfig = {
+      # KillMode=process: see the note on cc-ci-loops — a rebuild that merely touches this unit
+      # must not tear down the (shared) tmux server and every agent session with it.
+      KillMode = "process";
+      Type = "oneshot"; RemainAfterExit = true;
+      User = "loops"; Group = "users";
+      WorkingDirectory = "/srv/cc-ci-orch";
+    };
+    environment = { HOME = "/home/loops"; };
+    path = [ pkgs.bash pkgs.tmux pkgs.git pkgs.python3 pkgs.openssh pkgs.nettools ];
+    script = ''
+      export PATH="/home/loops/.local/bin:$PATH"
+      proj="/srv/cc-ci-orch"
+      echo "$(cat /proc/sys/kernel/random/boot_id) boot $(date -u +%FT%TZ) — cc-ci-orchestrator up" \
+        >> "$proj/cc-ci-plan/.ao-boot.log" 2>/dev/null || true
+      cd "$proj" && python3 cc-ci-plan/agents.py up orchestrator || echo "cc-ci orchestrator agents.py up failed"
+    '';
+  };
+
+  # Weekly recipe upgrade — runs /upgrade-all over every enrolled recipe (opens recipe PRs
+  # verified by !testme, never merges). Replaces the boot-fragile busybox-crond-in-tmux from
+  # phase 5 §4 with a reboot-safe systemd timer. The service is timer-triggered only (NOT
+  # wantedBy multi-user.target) so it never runs on boot/activation — only on the schedule.
+  systemd.services.cc-ci-upgrade-all = {
+    description = "cc-ci weekly /upgrade-all run (recipe upgrade survey + PRs, never merges)";
+    after = [ "network-online.target" "tailscaled.service" "claude-install.service" ];
+    wants = [ "network-online.target" ];
+    serviceConfig = {
+      Type = "oneshot";  # launch-upgrader.py spawns the cc-ci-upgrader tmux session and returns
+      User = "loops"; Group = "users";
+      WorkingDirectory = "/srv/cc-ci";
+      # Optional per-run overrides for backend/model (LOOP_BACKEND, LOOP_MODEL, OPENCODE_SHARE,
+      # UPGRADER_ARGS, …). The leading "-" makes it optional: absent file → claude/sonnet defaults.
+      # Current config (as of 2026-08-16): the upgrader + report run on tinfoil/deepseek-v4-pro
+      # (LOOP_MODEL + REPORT_MODEL in the env file); the hourly SUPERVISOR stays on glm-5.2
+      # (SUPERVISOR_MODEL defaults to opencode-go/glm-5.2 in launch-supervisor.py, NOT overridden
+      # here). Subagents bind deepseek via the cc-ci repo's opencode config. LOOP_TIER=zen is kept
+      # so the tier check passes; the watchdog's usage-limit probe sends the deepseek model name to
+      # the zen endpoint, which returns 200 (not 429) → resume immediately (correct: tinfoil has no
+      # rolling usage limit to wait out). No rebuild needed to switch — the env file is read at each
+      # timer fire. Holds no secrets (the tinfoil API key lives in the opencode config / auth.json).
+      EnvironmentFile = "-/srv/cc-ci/upgrader.env";
+    };
+    environment = { HOME = "/home/loops"; CLAUDE_BIN = "/home/loops/.local/bin/claude"; };
+    path = [ pkgs.bash pkgs.tmux pkgs.git pkgs.python3 pkgs.openssh pkgs.nettools ];
+    script = ''
+      export PATH="/home/loops/.local/bin:$PATH"
+      python3 /srv/cc-ci/cc-ci-plan/launch-upgrader.py start >> /srv/cc-ci/.cc-ci-logs/upgrader-cron.log 2>&1
+    '';
+  };
+
+  systemd.timers.cc-ci-upgrade-all = {
+    description = "Weekly trigger for cc-ci-upgrade-all (Thursdays 22:00 America/New_York — Boston 10pm)";
+    wantedBy = [ "timers.target" ];
+    timerConfig = {
+      # 10pm Thursday Boston time — DST-aware (EDT→02:00 UTC, EST→03:00 UTC) via the tz in OnCalendar.
+      OnCalendar = "Thu *-*-* 22:00:00 America/New_York";
+      Persistent = true;  # if the box was down at the scheduled time, run once on next boot
+    };
+  };
+
+  # Hourly SUPERVISOR — a glm-5.2 orchestrator wake-up that keeps the weekly run on track. The
+  # log-idle/429 watchdog only handles opencode-go usage-limit stalls; it does NOT cover a host
+  # disk-full crash (which killed the 2026-07-03 run) or any other environmental wedge. This is a
+  # CHEAP deterministic gate: if the weekly run is complete or actively progressing it does NOTHING
+  # (zero model tokens). Only when a run has stalled/died before completing does it launch a
+  # short-lived glm-5.2 agent that diagnoses the blockage and drives the run to a clean DONE.
+  systemd.services.cc-ci-upgrade-supervisor = {
+    description = "cc-ci hourly weekly-run supervisor (glm-5.2 — drives a stalled /upgrade-all to completion)";
+    after = [ "network-online.target" "tailscaled.service" ];
+    wants = [ "network-online.target" ];
+    serviceConfig = {
+      Type = "oneshot";  # launch-supervisor.py check: gate now, spawn the agent into tmux, return
+      User = "loops"; Group = "users";
+      WorkingDirectory = "/srv/cc-ci";
+      # Shares the weekly run's optional override file (e.g. SUPERVISOR_MODEL=…); "-" = optional.
+      EnvironmentFile = "-/srv/cc-ci/upgrader.env";
+    };
+    environment = { HOME = "/home/loops"; };
+    path = [ pkgs.bash pkgs.tmux pkgs.git pkgs.python3 pkgs.openssh pkgs.nettools ];
+    script = ''
+      export PATH="/home/loops/.local/bin:$PATH"
+      python3 /srv/cc-ci/cc-ci-plan/launch-supervisor.py check >> /srv/cc-ci/.cc-ci-logs/supervisor-cron.log 2>&1
+    '';
+  };
+
+  systemd.timers.cc-ci-upgrade-supervisor = {
+    description = "Hourly trigger for cc-ci-upgrade-supervisor (weekly-run health check + drive)";
+    wantedBy = [ "timers.target" ];
+    timerConfig = {
+      OnCalendar = "*-*-* *:07:00";  # every hour at :07 (offset from the weekly :00 fire)
+      Persistent = false;            # a missed hourly check is moot — the next hour re-checks
+    };
+  };
+}
@@ -0,0 +1,203 @@
+# orchestrator-host.nix — the host contract that nix/modules/cc-ci.nix (the orchestrator's
+# loops/timers) silently assumes, made explicit and reusable: the `loops` user the agents run as,
+# the standalone claude/opencode CLIs, the shared opencode web server and its tailnet-only UI,
+# nix-ld so foreign binaries run on NixOS, and the tool set agents reach for.
+#
+# Exported from flake.nix as `nixosModules.orchestrator-host`. A host imports this together with
+# `nixosModules.cc-ci-orchestrator`; the combined CI-server + orchestrator host (`#cc-ci`) also
+# imports recipe-maintainers/cc-ci's `nixosModules.cc-ci-server`.
+#
+# History: until 2026-09 this lived (twice, drifting) in nix/hosts/cc-ci-orchestrator-hetzner/
+# configuration.nix here and in notplants-nix's hosts/notplants-orchestrator/configuration.nix,
+# the shared agent box that also ran lichen + project-orchestrator. The cc-ci half moved to its
+# own host; this file is that half.
+{ config, lib, pkgs, ... }:
+let
+  cfg = config.cc-ci-orchestrator;
+in
+{
+  options.cc-ci-orchestrator = {
+    ciSshHost = lib.mkOption {
+      type = lib.types.str;
+      default = "127.0.0.1";
+      example = "100.95.31.88";
+      description = ''
+        Where `ssh cc-ci` (used by every skill and script that drives the CI server) connects to,
+        as root with ~loops/.ssh/cc-ci-root-ed25519. On the combined host the CI server IS this
+        machine, so the default is loopback; a standalone orchestrator points it at the CI
+        server's tailnet address.
+      '';
+    };
+
+    opencodeUiPort = lib.mkOption {
+      type = lib.types.port;
+      default = 8443;
+      description = ''
+        TLS port of the nginx front door for the opencode web UI. Not 443: on the combined host
+        Traefik (docker swarm) owns 80/443. The port is not opened in the firewall, so it is
+        reachable only over the trusted tailscale interface.
+      '';
+    };
+
+    opencodeUiHost = lib.mkOption {
+      type = lib.types.str;
+      default = "oc.commoninternet.net";
+      description = "nginx server_name for the opencode web UI (self-signed, basic auth).";
+    };
+  };
+
+  config = {
+    # ---- the loops user -------------------------------------------------------------------
+    # claude sessions run as non-root (--dangerously-skip-permissions is refused for root).
+    users.users.loops = {
+      isNormalUser = true;
+      uid = 1000;                 # fixed: workspace files are rsynced between hosts by uid
+      home = "/home/loops";
+      shell = pkgs.bash;
+      extraGroups = [ "wheel" "docker" ];
+    };
+    security.sudo.wheelNeedsPassword = false;
+    security.sudo.extraRules = [{
+      users = [ "loops" ];
+      commands = [{ command = "ALL"; options = [ "NOPASSWD" ]; }];
+    }];
+
+    # /home/loops/.local/bin holds the standalone claude + opencode binaries; it must be first on
+    # every PATH (interactive shells, tmux, the systemd units in cc-ci.nix prepend it too).
+    environment.variables.PATH = lib.mkForce
+      "/home/loops/.local/bin:/run/current-system/sw/bin:/run/wrappers/bin:/usr/bin:/bin";
+
+    # ---- nix-ld: the standalone Claude Code / opencode CLIs are foreign dynamic ELF binaries ---
+    programs.nix-ld.enable = true;
+    programs.nix-ld.libraries = with pkgs; [ stdenv.cc.cc.lib zlib openssl curl glibc ];
+
+    # ---- the toolbox every agent on this box gets ----------------------------------------
+    # Bar for adding something: an agent doing ordinary work would otherwise waste a turn
+    # discovering it is absent.
+    environment.systemPackages = with pkgs; [
+      git tmux python3 jq curl cacert
+      gnused gawk coreutils gnugrep findutils util-linux nettools openssh
+      age sops ssh-to-age
+      wget gnutar gzip unzip zip xz
+      ripgrep fd tree file less which
+      procps psmisc htop lsof strace ncdu
+      dnsutils socat netcat-gnu iproute2 iputils
+      openssl gnumake gcc pkg-config
+      yq-go diffutils patch rsync bubblewrap
+    ];
+
+    # ---- ssh config for the loops user: `ssh cc-ci` = the CI server (root) -----------------
+    # Written only if absent so a manual customisation survives rebuilds.
+    system.activationScripts.loopsSshConfig = ''
+      mkdir -p /home/loops/.ssh && chown loops:users /home/loops/.ssh && chmod 700 /home/loops/.ssh
+      if [ ! -f /home/loops/.ssh/config ]; then
+        cat > /home/loops/.ssh/config <<'SSHCFG'
+Host cc-ci
+    HostName ${cfg.ciSshHost}
+    User root
+    IdentityFile /home/loops/.ssh/cc-ci-root-ed25519
+    IdentitiesOnly yes
+    StrictHostKeyChecking accept-new
+    ServerAliveInterval 30
+
+Host git.autonomic.zone
+    HostName git.autonomic.zone
+    Port 2222
+    User git
+    IdentityFile /home/loops/.ssh/autonomic-bot-gitea-ed25519
+    IdentitiesOnly yes
+
+Host tangled.org
+    IdentityFile /home/loops/.ssh/tangled-ed25519
+    IdentitiesOnly yes
+SSHCFG
+        chmod 600 /home/loops/.ssh/config
+        chown loops:users /home/loops/.ssh/config
+      fi
+    '';
+
+    # ---- standalone CLIs (idempotent installers; re-run on every activation, no-op if present) --
+    systemd.services.claude-install = {
+      description = "Install Claude Code CLI for loops user (idempotent)";
+      wantedBy = [ "multi-user.target" ];
+      after = [ "network-online.target" ];
+      wants = [ "network-online.target" ];
+      serviceConfig = { Type = "oneshot"; RemainAfterExit = true; User = "loops"; Group = "users"; };
+      environment = { HOME = "/home/loops"; };
+      path = [ pkgs.curl pkgs.bash pkgs.coreutils pkgs.gnutar pkgs.gzip ];
+      script = ''
+        if [ ! -x "$HOME/.local/bin/claude" ]; then
+          echo "installing Claude Code CLI for loops user..."
+          curl -fsSL https://claude.ai/install.sh | bash || echo "install failed — retry on next activation"
+        fi
+      '';
+    };
+
+    systemd.services.opencode-install = {
+      description = "Install opencode CLI for loops user (idempotent)";
+      wantedBy = [ "multi-user.target" ];
+      after = [ "network-online.target" ];
+      wants = [ "network-online.target" ];
+      serviceConfig = { Type = "oneshot"; RemainAfterExit = true; User = "loops"; Group = "users"; };
+      environment = { HOME = "/home/loops"; };
+      path = [ pkgs.curl pkgs.bash pkgs.coreutils pkgs.gnutar pkgs.gzip pkgs.unzip ];
+      script = ''
+        if [ ! -x "$HOME/.local/bin/opencode" ]; then
+          echo "installing opencode CLI for loops user..."
+          curl -fsSL https://opencode.ai/install | bash || echo "install failed — retry on next activation"
+          # The installer puts the binary in ~/.opencode/bin; every unit here expects ~/.local/bin.
+          if [ -x "$HOME/.opencode/bin/opencode" ]; then
+            mkdir -p "$HOME/.local/bin" && ln -sfn "$HOME/.opencode/bin/opencode" "$HOME/.local/bin/opencode"
+          fi
+        fi
+      '';
+    };
+
+    # ---- opencode web server: one shared instance the opencode-backed agents attach to -------
+    # Provider creds come from /srv/cc-ci/.testenv (out of band, see README).
+    systemd.services.opencode-web = {
+      description = "opencode web server for cc-ci agents";
+      wantedBy = [ "multi-user.target" ];
+      after = [ "network-online.target" "tailscaled.service" "opencode-install.service" ];
+      wants = [ "network-online.target" ];
+      serviceConfig = {
+        Type = "simple";
+        User = "loops"; Group = "users";
+        WorkingDirectory = "/srv/cc-ci-orch/cc-ci";
+        EnvironmentFile = [ "-/srv/cc-ci/cc-ci/.env.public" "/srv/cc-ci/.testenv" ];
+        ExecStartPre = "${pkgs.coreutils}/bin/rm -rf /tmp/opencode";
+        ExecStart = "/home/loops/.local/bin/opencode serve --hostname 127.0.0.1 --port 4096";
+        Restart = "on-failure";
+        RestartSec = "5s";
+      };
+      environment = {
+        HOME = "/home/loops";
+        PATH = lib.mkForce "/run/wrappers/bin:/home/loops/.local/bin:/run/current-system/sw/bin:/usr/bin:/bin:/etc/profiles/per-user/loops/bin:/nix/var/nix/profiles/default/bin";
+      };
+      path = [ pkgs.bash pkgs.coreutils pkgs.git pkgs.python3 pkgs.openssh pkgs.tmux pkgs.nettools ];
+    };
+
+    # ---- tailnet-only nginx front door for the opencode UI -------------------------------
+    # Self-signed cert + basic auth, both created out of band (a store path would be world
+    # readable) — see README "Secrets to stage". nginx FAILS TO START if they are missing.
+    #   /etc/nginx/oc-selfsigned.crt  root:nginx 0644
+    #   /etc/nginx/oc-selfsigned.key  root:nginx 0640
+    #   /etc/nginx/oc-htpasswd        root:nginx 0640   (`oc:<bcrypt>`; plaintext in /secrets)
+    services.nginx = {
+      enable = true;
+      recommendedProxySettings = true;
+      virtualHosts.${cfg.opencodeUiHost} = {
+        listen = [ { addr = "0.0.0.0"; port = cfg.opencodeUiPort; ssl = true; } ];
+        # onlySSL flags the vhost as SSL so the module renders ssl_certificate for the listener.
+        onlySSL = true;
+        sslCertificate = "/etc/nginx/oc-selfsigned.crt";
+        sslCertificateKey = "/etc/nginx/oc-selfsigned.key";
+        basicAuthFile = "/etc/nginx/oc-htpasswd";
+        locations."/" = {
+          proxyPass = "http://127.0.0.1:4096";
+          proxyWebsockets = true;
+        };
+      };
+    };
+  };
+}