cc-ci-orchestrator/cc-ci-plan/plan-phase-lvl5-lint-rung.md

# Phase `lvl5` — level-system changes: 5th rung (`abra recipe lint`) + remove "capping"

**Mission (operator-specified, two changes):**

1. **New top rung — Level 5 = `abra recipe lint` passes against the exact recipe ref
   under test (the PR head on PR builds)**, after the existing four rungs (install,
   upgrade, backup/restore, functional). The existing four rungs' meanings are UNCHANGED.

2. **Remove the "capping" concept entirely.** The operator finds cap/cap_reason confusing.
   New level semantics (operator-decided 2026-06-11, explicit Q&A + refinement):
   **level = the highest rung that PASSED, where every rung below it is "pass" or an
   INTENTIONAL skip.** A real FAIL blocks. A not-applicable rung is split in two:
   - **Intentional skip (declared/structural):** the rung genuinely does not apply to
     this recipe — e.g. not backup-capable (declared), only one published version exists
     (no upgrade target). These are SKIPPED: they do not stop the climb.
   - **Unintentional N/A (unverified):** the rung SHOULD have run but didn't — infra
     error, missing tool (abra absent), harness exception, prior-stage abort, timeout.
     **The level cannot rise above an unverified rung** — it blocks exactly like a fail
     (the rungs below it still count).
   Formally, with statuses {pass, fail, skip (intentional), unver (unintentional)}:
   `level = max i such that rung_i == "pass" and all j < i have status in {"pass","skip"}`;
   0 if no such i.
   - install ✔, upgrade ✘, backup ✔, functional ✔, lint ✔ → **level 1** (fail blocks)
   - install ✔, upgrade ✔, backup skip(not capable), functional ✔, lint ✔ → **level 5**
     (intentional skip — previously this capped at 2; that was the confusing part)
   - install ✔, upgrade ✔, backup UNVER (harness error), functional ✔, lint ✔ →
     **level 2** (unverified blocks — we cannot claim what we didn't check)
   - all four ✔, lint unver (abra missing) → **level 4** (an unverified top rung is
     simply not earned)
   **Every N/A source in derive_rungs must be explicitly classified** intentional vs
   unintentional, from declared recipe meta / structural facts vs runtime failure —
   the classification table goes in DECISIONS.md and is Adversary-reviewed. Default for
   an unclassifiable N/A is UNINTENTIONAL (conservative).
   The words "cap", "capped", "cap_reason" disappear from code, schema, card, dashboard,
   and docs. The per-rung table remains everywhere it exists today (now distinguishing
   ✔ / ✘ / intentional-skip / unverified), so nothing is hidden — the table is the SOLE
   carrier of "why isn't this level higher". Record the whole semantics change in
   DECISIONS.md (it deliberately replaces the Phase-3 "N/A caps" stance; the
   unverified-blocks rule preserves its honest core: never claim what wasn't checked).

State files (phase-namespaced): `STATUS-lvl5.md`, `BACKLOG-lvl5.md`, `REVIEW-lvl5.md`,
`JOURNAL-lvl5.md`. DECISIONS.md shared (append).

---

## 1. Current system (file map — verified 2026-06-11)

- `runner/harness/level.py` (120 lines, PURE) — `RUNGS = ("install", "upgrade",
  "backup_restore", "functional")`, `RUNG_LABEL` 1–4, `compute_level()` (gap caps; N/A
  caps with distinct reason), `tier_to_rung`, `backup_restore_status`.
- `runner/harness/results.py:137-218` — `derive_rungs()` builds the rung dict from tier
  results; `compute_level` → results.json `level` + `cap_reason` + `capped`.
- `runner/harness/card.py` — `LEVEL_COLOR` map (line ~58), cap line hardcodes
  *"full clean climb — top level (4)"* (line ~246).
- `dashboard/dashboard.py` — `_LEVEL_COLOR` (line ~81), corner level badge `_level_pill`,
  `/badge/<recipe>.svg`.
- `tests/unit/test_level.py`, `test_results.py`, `test_card.py`, `test_dashboard.py`.
- **Lint today:** abra's pinned (non-chaos) deploy runs `abra recipe lint` internally and
  FATALs on R014 — see `runner/harness/abra.py:109-114`: the CI's mirror-origin repointing
  tripped a go-git path, which is why chaos deploys (the PR-testing path!) deliberately
  SKIP lint. So lint currently runs implicitly on some paths and not at all on others —
  the new rung makes it explicit, uniform, and visible.

## 2. Design requirements

1. **New rung `lint` appended after `functional`** → ladder install(1) upgrade(2)
   backup_restore(3) functional(4) **lint(5)**. `RUNG_LABEL[5] = "lint (abra recipe lint)"`
   (or similar). Full clean climb is now 5.
2. **What is linted:** the EXACT recipe tree/ref the run deployed (PR head on
   `!testme`/PR builds; the tested ref otherwise) — never some other branch. Run
   `abra recipe lint <recipe>` (the abra on the CI host) against the run's own checkout
   context. Capture rc + full output into the run artifacts (e.g. `lint.txt`), and put a
   pass/fail + short excerpt in results.json.
3. **Mirror-plumbing must not pollute recipe lint results (CRITICAL, see abra.py:109-114):**
   the R014/go-git failure caused by CI's origin-repointing is a HARNESS artifact, not a
   recipe defect. The lint rung must evaluate the recipe's content. Solve it properly
   (e.g. lint in a context where origin looks canonical, or pre-step the same
   stash/revert dance abra.py already does) — do NOT blanket-ignore lint rules to make
   the plumbing pass, and document exactly what (if anything) is filtered and why. Any
   filtering is a named, unit-tested, Adversary-reviewed decision.
4. **Verdict semantics UNCHANGED:** lint is a level rung, not a run gate. A lint failure
   caps the level at 4 with `cap_reason "L5 ... FAILED"`; it must NOT fail/flip the run
   verdict, and must be time-bounded + best-effort in execution (a wedged lint can never
   hang a run — hard timeout, ~60s class).
5. **No N/A escape hatch by default:** every recipe can be linted, so the rung is
   pass/fail in practice. When lint can't run (abra binary missing, timeout) the status
   is **unver** + loud log — never silently "pass", never an intentional skip; the rung
   is not earned and the level stays ≤ 4.
6. **De-cap implementation (mission item 2):** `compute_level()` reimplemented to the
   new formal rule; `cap_reason`/`capped` deleted from level.py, derive_rungs/results.json
   schema, card (the "capped: …"/"full clean climb — top level (4)" line at card.py:~246
   is replaced by the rung table alone or a neutral "level N of 5"), dashboard fields,
   and docs. **The top-corner level badge in particular** (the big LEVEL badge on the
   card, the dashboard `_level_pill`, and `/badge/<recipe>.svg`) shows ONLY the level —
   number + color, zero capping/cap-reason annotation (operator-specified 2026-06-11).
   The INLINE detail keeps the intentional-skip information: the per-rung table still
   marks each rung ✔ / ✘ / intentional-skip / unverified, so "why isn't it higher" lives
   there and only there. Unit tests rewritten to the new semantics, INCLUDING the worked
   examples from the mission and the old N/A-cases (single-published-version recipe,
   non-backup-capable recipe) now climbing past their former caps.
7. **All consumers updated coherently:** RUNGS/labels, results schema, card (color map +
   hardcoded top-level line), dashboard pills/badge SVG/legend text, docs
   (testing.md / results docs / recipe-customization.md §levels if it references L4 as
   top), and every unit test that assumes 4 is the ceiling or asserts cap_reason.
8. **History compatibility:** old results.json artifacts (level ≤ 4, lint rung absent,
   cap_reason PRESENT) must still render correctly in dashboard/card history views — no
   KeyErrors, no retroactive relabeling of old runs; renderers tolerate both schemas.
9. **Expected level shifts are findings, not regressions:** recipes previously capped by
   an N/A rung will legitimately jump levels (e.g. a non-backup-capable recipe with
   passing functional goes 2 → 4/5). P3/P4 must produce a before/after level table for
   ALL enrolled recipes so the Adversary can check every shift against the new rule —
   any shift NOT explained by the rule change is a real regression.

## 3. Work plan

**P1 — Ladder + plumbing.** level.py: new rung + the de-capped `compute_level()` per
§2 item 2 (delete cap_reason/capped); lint executor (new `harness/lint.py` or a clean
home in abra.py) with timeout, artifact capture, mirror-context handling per §2.3;
derive_rungs wiring (incl. the intentional-vs-unintentional classification of every N/A
source, §2 item 2); results schema. Unit tests for: full climb = 5; fail-blocks
(upgrade ✘ → L1 even with higher passes); intentional-skip climbs (backup skip +
functional ✔ → L4+); unverified-blocks (backup unver + functional ✔ → L2); lint unver →
stays 4; unclassifiable N/A defaults to unver; old-artifact rendering; mirror-filter
decision (if any).

**P2 — Presentation.** Card, dashboard, badge, docs — a level-5 color/legend that reads
as "above functional". Regenerate anything generated.

**P3 — Reality pass over all enrolled recipes.** Run the lint rung against every enrolled
recipe's current main/HEAD (cheap — lint only, no deploys, respects the shared-checkout
rule below). Matrix in BACKLOG-lvl5.md: pass/fail + rule hits per recipe. For failing
recipes: the rung correctly caps them at ≤4 (that is WORKING AS DESIGNED, not a phase
blocker). Where the fix is mechanical and safe, open a PR against the recipe mirror
(NEVER push main / NEVER merge — operator decides), else file the finding in
DEFERRED.md with the rule output.

**P4 — Real-CI proof.** Full-stage runs on enough recipes to prove the changes
end-to-end: ≥1 recipe reaching a genuine L5 (all five rungs green), ≥1 recipe blocked at
L4 by a lint failure (real or synthesized — a throwaway branch with a deliberate lint
violation is fine), ≥1 recipe demonstrating the N/A-skip (formerly capped by an N/A rung,
now climbing past it), ≥2 runs via the drone `!testme` path showing the rung on a real
PR, plus the canary suite green — the bad canaries must still land at their designed
levels under the NEW formula (re-derive what those designed levels are; backup/restore
fail still blocks). The unverified-blocks rule is proven by unit test + one synthesized
run (induce a tier abort and check the level stays below the unverified rung). Card +
dashboard visually verified (Read the PNGs/SVG output), and the §2 item 9 before/after
level table completed for all enrolled recipes.

## 4. Gates

**M1 — Implementation complete (pre-merge).** Branch with P1+P2; Adversary cold-runs unit
suite + lint from a clean checkout; reviews the mirror-filter decision (§2.3) explicitly;
confirms verdict-neutrality by code inspection + a targeted test. PASS required before
merge to main.

**M2 — Proven in real CI.** P3 matrix complete for ALL enrolled recipes; P4 runs done
(L5 achieved, L4-capped demonstrated, drone path ×2, canaries green); old artifacts still
render; run durations not materially inflated (lint adds ≤ ~60s). Fresh Adversary PASS →
Builder writes `## DONE` to STATUS-lvl5.md.

## 5. Guardrails (binding)

- **Rung statuses stay honest** (the Phase-3 rule, adapted to the new semantics): no
  rung is ever silently "pass" — lint errors/timeouts/missing abra are fail/na, never
  pass. The level FORMULA is the operator-decided rule in §2 item 2 (N/A skips); the
  per-rung table and the run verdict must always remain visible so completeness is
  never hidden, and the verdict logic itself is untouched by this phase.
- **No gate weakening; no verdict changes.** Existing tests/assertions untouchable except
  where the L4-ceiling assumption itself must change — those edits are mechanical and the
  Adversary checks each one against the old intent.
- **Recipe mirrors:** lint findings in recipes → PRs or DEFERRED entries only; NEVER push
  recipe-mirror main, NEVER merge their PRs.
- **Shared checkout race:** NEVER git-checkout `~/.abra/recipes/<recipe>` on cc-ci while
  that recipe's CI build is running — the harness deploys from that tree. P3's lint-only
  sweep must use its own scratch clones or run when the recipe is not mid-build.
- Real-CI etiquette: ≤2-3 concurrent live deploys; teardown every dev deploy on every
  exit path; no secrets in logs/commits. Commit author `autonomic-bot
  <autonomic-bot@noreply.git.autonomic.zone>`; push after every commit.
- CI host has no python3 on default PATH for remote one-liners — use shell or the
  harness venv (`cc-ci-run`).

## 6. Definition of Done

The new level system live on main and visible end-to-end (results.json → card →
dashboard → badge): L5 = abra recipe lint on the tested ref, capping concept fully
removed (level = highest passed rung with all lower rungs pass-or-intentional-skip;
declared skips climb, fails AND unverified rungs block; no cap/cap_reason anywhere),
all enrolled recipes linted and dispositioned with
the before/after level table adversary-checked, ≥1 real L5 + ≥1 lint-blocked L4 + ≥1
N/A-skip climb demonstrated through real CI including the drone path, old artifacts
unharmed, M1+M2 fresh Adversary PASSes, no verdict or duration regressions.