# REVIEW — cc-ci Adversary (append-only)

This file is owned by the **Adversary** loop (§6.1). The Builder seeds this stub at bootstrap and
does not edit it afterward. Adversary appends milestone/D-item verdicts (`<id>: PASS @<ts>` +
evidence, or `FAIL` + a finding in `BACKLOG.md ## Adversary findings`), and may write `## VETO`.

<!-- Adversary verdicts below -->

## M0 — Foundations: PASS @2026-05-26T21:35Z

Verified cold (fresh shell, own clone `/srv/cc-ci/cc-ci-adv`, isolated host build dir
`/root/cc-ci-advverify`, no reuse of Builder's `/root/cc-ci`).

Acceptance — "`systemctl is-system-running` healthy after a rebuild from the repo" + Builder's
sops claim:
- **Repo rebuilds cc-ci:** synced M0 commit `deb4a0f` (git-archive, no .git) to host, ran
  `nixos-rebuild build --flake .#cc-ci` → `BUILD EXIT 0`, produced
  `…-nixos-system-nixos-24.11.20250630.50ab793`. Current HEAD also builds clean.
- **System health:** `systemctl is-system-running` → `running`; `systemctl --failed` → 0 units.
- **sops decrypt:** `/run/secrets/test_secret` present, mode `400 root:root`, 41 bytes, value
  begins `cc-c…` (matches claimed generated `cc-ci-m0-…`). `secrets/secrets.yaml` is genuinely
  encrypted (2× `ENC[…]` + sops metadata block).
- **D6 leak probe (early):** the decrypted plaintext value appears **0 times** across *all* git
  history (`git grep -F over git rev-list --all`) and 0× in plaintext in `secrets.yaml`. No leak.

Note (not a finding; context for the M1 gate): the *running* system is already ahead of M0 — its
closure includes docker, `unit-swarm-init`, and **traefik** units (`traefik.yml`,
`traefik-stack.yml`, `unit-traefik-deploy`) that are **not yet committed** (HEAD `ab839ae` is
swarm-only, no traefik). Expected mid-M1 churn, but the Traefik config must be committed to the
repo before M1 is claimed or it fails D8 reproducibility — will check at the M1 gate.

## M1 — Swarm + abra target: PASS @2026-05-26T22:20Z

Verified cold from own clone; deployed my **own** probe recipe via abra (not trusting the Builder's
hand-test). Acceptance "a recipe deployed via abra is reachable over HTTPS at
`*.ci.commoninternet.net`, then fully torn down leaving no volumes" + orchestrator's M1 checklist
(a–d).

- **(a) Real coop-cloud/traefik recipe (not hand-rolled):** `docker service ls` →
  `traefik_…_app` (`traefik:v3.6.15`) + `…_socket-proxy` (lscr.io socket-proxy) — the canonical
  recipe layout, deployed via abra (`scripts/deploy-proxy.sh`). `modules/traefik.nix` is deleted.
- **(b) Wildcard on web-secure + proxy overlay:** static `traefik.yml` has `web-secure: :443`
  (web→web-secure 301 redirect, verified live). File provider `/etc/traefik/file-provider.yml`:
  `tls.certificates: [{certFile:/run/secrets/ssl_cert, keyFile:/run/secrets/ssl_key}]`; swarm
  secrets `…_ssl_cert_v1`/`…_ssl_key_v1` mounted (2909 B / 227 B = the pre-issued cert). My probe
  app `advm1probe_…_app` was attached to the `proxy` overlay.
- **E2E (cold deploy):** `abra app new custom-html -D advm1probe.ci.commoninternet.net` (forced
  `LETS_ENCRYPT_ENV=""`) → `deploy succeeded 🟢`. Via SOCKS proxy: **HTTP 200**; served cert
  `subject: CN=*.ci.commoninternet.net`, SAN-matched, `SSL certificate verify ok`, issuer LE E8 —
  i.e. the **pre-issued wildcard**, NOT a per-host ACME cert.
- **(c) No Gandi/DNS token, no ACME credential:** repo (all history) clean; on host the only
  gandi/dns-challenge strings are **commented-out** recipe-template options (`#GANDI_…`,
  `#SECRET_GANDIV5_…`) holding no value. Active traefik env = `LETS_ENCRYPT_ENV=` (empty),
  `WILDCARDS_ENABLED=1`, `compose.wildcard.yml`. `staging`/`production` certResolvers are *defined*
  in traefik.yml (stock template) but **referenced by no router**; both acme.json are **0 bytes**;
  **0 ACME lines in traefik logs**. No ACME ever fires. (Hardening risk filed — see findings.)
- **(d) Manual renewal documented:** DECISIONS.md — operator re-issues at same paths, then
  `abra app secret rm … ssl_cert` + re-insert at bumped version; install.md "Renewed out-of-band;
  never ACME here."
- **Teardown:** `abra app undeploy` + `volume remove` → post-teardown services/containers/volumes/
  secrets for the probe **all 0**. Also independently confirmed the Builder's `cchtml1` test left 0
  runtime resources (only its inert `.env` config file remains, harmless).

Verdict: **M1 PASS.** Not a hard fail on (c) — no token/credential exists and no ACME fires — but
the inert ACME resolvers + test-app default `LETS_ENCRYPT_ENV=production` are a latent hazard that
goes live when the harness deploys apps; filed as `[adversary]` for M4.

<!-- M2 live-trigger probe @2026-05-26T23:30Z: this push should create Drone build #4 -->

## M2 — Drone online: PASS @2026-05-26T23:32Z

Verified cold from own clone. Acceptance: "push to cc-ci triggers a visible green Drone build."

- **Drone server healthy:** `https://drone.ci.commoninternet.net/healthz` → HTTP 200 via gateway.
  Exec runner (`drone-runner-exec.service`) active, `polling the remote server capacity=2 type=exec`.
- **Repo wired:** in Drone's DB the `recipe-maintainers/cc-ci` repo is `repo_active=1`,
  `repo_config=.drone.yml`. Gitea↔Drone OAuth proven by the in-pipeline `clone` step succeeding
  against the private repo (build can't clone without working OAuth/repo token).
- **Push→green, independently triggered:** I pushed my own commit `91a8e8d` (a REVIEW.md change) →
  Drone created **build #4**, `build_event=push`, `build_trigger=@hook` (Gitea webhook), and it ran
  **`success`**: stage `self-test` exit 0, steps `clone`+`hello` both exit 0. Builds #1–#3 (Builder
  commits) likewise all `success` via `@hook`. (My earlier M0/M1 review pushes predate the
  `.drone.yml`, so correctly produced no builds.)
- **Visible logs (D7 precondition):** `logs` table holds per-step log blobs for every build; Drone
  UI/API serve them. Full D7 UX is M8.

Verdict: **M2 PASS.** No new findings.

## M3 — Comment bridge: PRE-CLAIM PROGRESS (not yet PASS) @2026-05-26T23:48Z

M3 is **Blocked** in STATUS (Gitea not delivering webhooks), so not a gate verdict yet. But the
bridge is deployed and I independently hammered its auth/filter logic — the part I can verify
regardless of the delivery leg (and which survives a pivot to API polling). Probes were live POSTs
to `https://ci.commoninternet.net/hook` via the SOCKS proxy, with HMAC signatures I computed from
the on-host secret (read with root; value never printed/committed):

| probe | expect | got |
|---|---|---|
| no `X-Gitea-Signature` | 401 | **401** |
| bad signature | 401 | **401** |
| valid sig, event=`ping` (not issue_comment) | 204 | **204** |
| valid sig, `!testmexyz` on a real PR | 204 (no trigger) | **204** |
| valid sig, `!testme` but issue is not a PR | 204 | **204** |
| valid sig, `!testme` on PR, action=`edited` | 204 | **204** |
| valid sig, `!testme` on real PR, **non-collaborator** | 403 | **403** |

So: HMAC fail-closed + timing-safe (`compare_digest`, verified before body parse), `!testmexyz`
correctly ignored (exact trimmed match), non-PR ignored, and a non-collaborator is rejected (403;
collaborator status re-checked via Gitea API, not trusted from the signed payload). Source review
of `bridge/bridge.py` found no auth bypass.

**Blocker independently corroborated (operator-side):** the bridge hook *is* registered + active on
`recipe-maintainers/cc-ci` (id 210, events `[issue_comment]` → `ci.commoninternet.net/hook`), and
the bot is not a Gitea site-admin (`GET /admin/hooks` → 403) nor org owner, so it genuinely cannot
inspect/change Gitea's `[webhook] ALLOWED_HOST_LIST`. Endorse STATUS `## Blocked`: needs operator
allowlisting or the documented poll-the-API fallback.

**Still UNVERIFIED for an M3 PASS:** (1) the positive path — a valid collaborator `!testme` actually
starts a build + posts the PR comment end-to-end; (2) real Gitea→bridge delivery (or the polling
pivot). Will complete both when M3 is claimed.

**Noted for M7 (not a finding yet):** the Drone-managed Gitea webhook (id 209) carries its webhook
secret as a `?secret=` query param in the hook URL (Drone default; admin-only in Gitea, not in cc-ci
git / CI logs / dashboard). Will adjudicate against D6 at M7.

## M4 — Harness + install stage: VERIFICATION IN PROGRESS (no verdict yet) @2026-05-27T00:35Z

M4 is CLAIMED. Code review done; runtime checks so far:
- **A1 CLOSED** (see BACKLOG): harness forces `LETS_ENCRYPT_ENV=""` every deploy; live app
  `cust-c95a69` served the wildcard cert, 0 ACME lines, no certresolver.
- **Happy-path teardown works:** a prior run's app `cust-e084bd` was fully torn down (gone) — not
  an orphan; earlier ambiguity was a run cycling apps.
- **Two teardown-robustness defects filed (A2, A3):** janitor's `-pr` filter is dead code under the
  `cust-<hex>` naming (no crash-orphan reaping); teardown is best-effort/unverified and deletes the
  `.env` even on failed undeploy (silent orphan, run still green).
- **Deferred to next idle tick (a Builder harness run is active now; sequential-only):** my own
  cold install run (green install + Playwright + clean teardown verification) and the §6 kill-mid-run
  probe to test A3 empirically. Verdict (PASS/FAIL) follows that.

## M4 — Harness + install stage: PASS @2026-05-27T01:05Z

Verified by my **own** cold harness run (`RECIPE=custom-html REF=advcold… cc-ci-run
runner/run_recipe_ci.py`, app `cust-cfeb6a`, isolated from a Builder run that happened to run
concurrently as `cust-3c1970` — no collision, distinct domains/volumes/secrets):
- **Install stage green:** `test_install.py` → 2 passed (27s): `test_http_reachable` (HTTPS 200 via
  gateway) + `test_playwright_page` (real Chromium loads the live app, status 200, served HTML).
- **Guaranteed teardown:** after the run, `cust-cfeb6a` left **0** services / volumes / secrets /
  containers / `.env` — fully clean. Infra (traefik/drone/bridge/backups) untouched.
- A1 closed (no-ACME enforced). **Open robustness findings A2 (dead `-pr` janitor) + A3 (unverified
  best-effort teardown)** concern the *crash* path (finalizer-skipped), not this happy-path run;
  they don't block M4's literal acceptance but must be resolved before DONE (D2 teardown guarantee).
  Kill-mid-run probe to substantiate A2/A3 deferred until the host is idle.

Verdict: **M4 PASS.**

## M5 — Upgrade + backup/restore stages: PASS @2026-05-27T01:05Z

Same cold run, stages 2 and 3 — both genuine end-to-end (no mocks; assertions reviewed in source
and not softened):
- **Upgrade green:** `test_upgrade.py` → 1 passed (41s). Deploys the **previous published version**
  (`previous_version` = `recipe_versions[-2]`), writes a marker into the volume-backed html dir,
  upgrades to latest (`abra upgrade`), then asserts HTTP 200 **and** the marker survives — a real
  version change with data persistence across the volume (`cust-…_content`), not a no-op.
- **Backup/restore green:** `test_backup.py` → 1 passed (37s). Writes `original`, `abra backup`,
  mutates to `mutated` (asserted), `abra restore`, then asserts the served content is back to
  `original` ("restore did not return the pre-mutation state"). Real backup→mutate→restore cycle
  via backup-bot-two.
- Teardown clean (same `cust-cfeb6a` 0-remnant check above covers all three stages — same domain
  reused per stage).

Verdict: **M5 PASS.**

## M6 — Recipe-local tests + second recipe: VERIFICATION IN PROGRESS (no verdict yet) @2026-05-27T01:48Z

M6 CLAIMED. Host has been continuously busy (Builder M6.5 ramp), so deploy-based checks are
deferred to an idle window; static + evidence review so far:
- **custom-html 3-stage:** already verified cold by me (see M5 PASS) — green + clean teardown.
- **D4 recipe-local discovery — code genuine:** `run_recipe_ci.snapshot_recipe_tests` copies the
  recipe-shipped `tests/` before abra re-checkouts to a version tag, then `run_recipe_local` deploys
  the app and runs those tests against the LIVE app via `CCCI_BASE_URL`/`CCCI_APP_DOMAIN`, merged as
  a separate stage with guaranteed teardown. Demo branch `recipe-maintainers/custom-html@
  ci/d4-recipe-local` confirmed to ship `tests/test_recipe_local.py` (Gitea API). Will run it cold to
  confirm the stage executes+passes.
- **keycloak (#2) install — test genuine:** `/realms/master` 200 health + real Playwright admin
  console login (waits for the username field). `recipe_meta.py` (HEALTH_PATH/timeouts) confirms D5
  "no harness surgery". Empirical keycloak reproduction deferred (heavy deploy; idle window).
- **Filed [adversary] A4** (concurrency): same-recipe concurrent runs share `~/.abra/recipes/<recipe>`
  with no isolation/lock/concurrency-cap — a collision vector for the §6 concurrency check; to
  confirm empirically.

Pending for idle host: cold D4 run, keycloak reproduce, A2/A3 kill-probe re-test, A4 concurrency test.

## D6/M7 — preliminary leak scan of published Drone logs (PASS so far; M7 not yet claimed) @2026-05-27T02:05Z

Host-safe probe while the host was busy. Pulled Drone's `database.sqlite`, dumped all 42 `logs`
rows (~25.5k chars of published per-step build output), scanned:
- **Known infra secrets — 0 leaks:** webhook HMAC (64), drone token (32), gitea token (40) each
  appear **0×** in the logs (exact `grep -F`).
- **No value patterns:** 0 matches for `password|secret|token = <value>`.
- The only long hex/base64 hits are **git commit SHAs** in `git clone/merge` output — benign.
Caveat: current Drone logs are hello-world + self-test; the full M7/D6 test must also cover
app-generated secrets (e.g. keycloak DB passwords) in recipe-run logs AND the dashboard (M8). This
is a clean baseline, not the final D6 verdict. (DB copy was scanned off-box and deleted; no secret
value printed or committed.)