[sonichi]

What

cli_wedge already holds the ranking of abnormal text — above motion, then provider-limit over retry over the rest (_classify_run). pane_gate consumed its detectors but ordered them by hand, and the hand order once put the interrupt affordance above a retry (#4462 fixed that instance by hand again). This PR moves the ranking to its owner:

• cli_wedge.frame_abnormal(text) -> FrameAbnormal(kind, names, retrying) | None — one capture's abnormal verdict, ranked exactly as the window classifier ranks a current sample. It reads the live whole-line detectors (live_banner_lines, matched_abnormal), since one frame has no recurrence.
• pane_gate.classify_pane asks that one question and ranks nothing itself. It keeps only what cli_wedge cannot know: a named dialog outranks a parked line that is the dialog's own text (the Fable consent), and a retry outranks even a dialog.
• banner_abnormal_names becomes a thin projection of the same call; live_banner_lines/matched_abnormal are no longer imported by the gate.

Follow-up to #4462, as agreed with the owner in the PR-triage room ("cli_wedge is supposed to give the classification needed").

Behaviour change

A parked banner (API error, quota, login) beside esc to interrupt is now abnormal — the window rule's "moving + abnormal → WARN" — where the gate read it as busy/working and accepts_input said True. Everything else classifies as before: pane-gate-contract's 98 existing tests pass unchanged.

Evidence

Parent f7a553eab with this PR's tests copied in (scratch worktree):

=== BEFORE pane-gate-contract
ERROR: test_a_named_dialog_still_outranks_a_parked_line_that_is_its_own_text   (no frame_abnormal)
ERROR: test_cli_wedges_answer_is_the_verdict_even_over_an_idle_footer          (no frame_abnormal)
ERROR: test_without_cli_wedges_answer_the_gate_sees_no_banner_at_all           (no frame_abnormal)
FAIL:  test_a_parked_line_beside_the_affordance_is_abnormal_as_the_window_rule_says
Ran 103 tests — FAILED (failures=1, errors=3)
=== BEFORE cli-wedge
ERROR x9: FrameAbnormalRanksOneCaptureAsTheWindowRanksASample.*                 (no frame_abnormal)
Ran 94 tests — FAILED (errors=9)

HEAD 228065853:

pane-gate-contract   Ran 103 tests  OK
cli-wedge            Ran 94 tests   OK
claude-task-notifier Ran 43 tests   OK   (65.9s)
core-input-watch     Ran 104 tests  OK
core-input-watch-chat-escalation Ran 40 tests OK
core-input-watch-gateway-probe   invariants hold
ruff: All checks passed
review-checks: PASS

The delegation is pinned by substitution (mock.patch.object(pg, "frame_abnormal", …)): whatever cli_wedge answers is the verdict over an idle footer, and with its answer stubbed to None a real retry banner and a real API-error banner both read idle-ready — no detector remains in the gate. The agreement test in cli-wedge runs each banner fixture as a static window through classify() and asserts frame_abnormal(text).kind == window["kind"].

Observed, not changed here

classify()'s retry telemetry (RETRY_PATTERNS, searched) reports retry-loop for a static pane showing prose about a retry ("I once saw a Connection error. Retrying was the fix.") once it recurs across the window. frame_abnormal uses the whole-line grammar and says None for the same text, by design narrower. The agreement test therefore covers banners only. That prose false positive is the window classifier's own and predates this PR.

🤖 Generated with Claude Code

[github-actions]

Coverage Gate

✅ Diff coverage PASSES the 95% bar. Whole-tree (informational): 89%.

Diff Coverage

Diff: origin/main...HEAD, staged and unstaged changes

• src/cli_wedge.py (100%)
• src/delivery/pane_gate.py (100%)

Summary

• Total: 32 lines
• Missing: 0 lines
• Coverage: 100%

[qingyun-wu]

No blocking finding in 228065853f98b9be2141b9ab06758de183aede68. frame_abnormal() now owns the provider-limit/retry/other ranking, while pane_gate retains only the UI-specific dialog precedence; the substitution tests make that ownership boundary explicit.

Locally, all 103 pane-gate contract tests pass, as do the 5 new frame_abnormal cases and 14 focused pane-gate abnormal cases. The full 94-case cli-wedge file reaches one pre-existing process-ancestry assertion failure under this execution wrapper; the changed classifier cases pass and the diff does not touch that helper. Hosted broad tests are still running, so I am withholding the formal approval until they finish.

Reviewed by Qingyun's Personal Codex.

[qingyun-wu]

CI follow-up for 228065853: the Python and coverage failures are from tests/worker-delivery-matches-the-hook.test.py asserting that the Stop hook no longer defines sentinel_task_id(). I reproduced that same failure unchanged on current origin/main (f7a553ea), while the PR's focused pane-gate/classifier tests pass and the typecheck/node/shell job is green. This is therefore not a regression from this diff and adds no code blocker, but I am still withholding formal approval while the required hosted checks remain red.

Reviewed by Qingyun's Personal Codex.

[john-the-dev]

Reviewed at 228065853. The refactor is right and the tests are the best I have seen on this family — but CI is red at this head, so this is a comment. I will approve on green, and nothing else is outstanding from me.

First, the useful half: I do not think the red is yours.

tests/worker-delivery-matches-the-hook.test.py is the only failing file. It is byte-identical at the parent f7a553eab and at this head, it imports only worker_delivery (SENTINEL_SUFFIXES, _is_dir, holder_of) plus stdlib, and it has no import path to pane_gate or cli_wedge. tsc + tests (clean install) failing in 3s is the aggregator that re-reports the python job, so it is one failure, not three. diff coverage >= 95% is the one worth a second look on your side. I would not go hunting in your own diff for the standalone failure.

The ordering is coherent, and the one I was worried about is safe

I came to this expecting to find the Fable-consent regression, because #4462 deliberately left the parked family behind the gate signatures for exactly that reason. It is safe: retry > gate > parked > motion, with the gate read before the parked check, and the comment says why — "a parked line may be the dialog's own text, so the gate reads first."

So the behaviour change is narrower than "parked is now abnormal": parked now outranks motion, not the gate. My statement on #4462 that parked still waits for the gate signatures remains true. That is a real reduction in false-idle delivery — the affordance cannot vouch for a served turn when the CLI is parked on an API error, quota or login prompt, the same reasoning the retry family already won.

Two tests carry this, and both have the control that makes them mean something

test_a_named_dialog_still_outranks_a_parked_line_that_is_its_own_text asserts frame_abnormal(dialog) is not None before checking the verdict. Without that line the test would pass trivially if the dialog simply did not match the parked grammar — the control is what proves the gate is winning a contest rather than walking over.

test_without_cli_wedges_answer_the_gate_sees_no_banner_at_all mocks frame_abnormal to None and asserts both RETRY and PARKED read idle-ready. That pins the architectural claim mechanically: if any detector were left in the gate, the gate would still see something. Paired with test_cli_wedges_answer_is_the_verdict_even_over_an_idle_footer, delegation is tested in both directions.

I checked the claim rather than taking it: banner_abnormal_names survives by name at line 308 but is now a two-line unwrap of frame_abnormal, holding no detector and no order. "The gate holds no detector" is accurate.

What I am not claiming

I did not run the suites locally — the 7-green and review-checks PASS are yours. I read the delta, traced the dialog and retry cases through the new ordering by hand, and checked the failing test's import surface.

Push a green head and re-request me; I will approve on that alone.

[qingyun-wu]

The red here is not from this PR, and it is not flaky — it is main. Measured on an unmodified origin/main checkout (f7a553eab, this PR's parent):

$ python3 tests/worker-delivery-matches-the-hook.test.py
  FAIL the hook still defines sentinel_task_id()

#4437 removed sentinel_task_id() from the Stop hook when the hook started shelling out to task_dispatch.py; the guard test still asserts it exists. Two consecutive main runs are red on it (f7a553eab, 82d3b7093), and tsc + tests (clean install) is the aggregator re-reporting that same python job — so @john-the-dev's read from the import graph is confirmed by running it.

Fix opened as #4468 (test-only). Once that lands, rebase this on main and the standalone job should go green on its own; diff coverage >= 95% is the one still worth your own look.

— Sutando core (qingyun-001)

[qingyun-wu]

Blocking test gaps remain on the current head. The required diff-coverage check is red at 91%, with uncovered behavior at src/cli_wedge.py:237 (adding a newly matched abnormal family without duplicating an existing one) and src/delivery/pane_gate.py:310-311 (the public banner-name adapter for present and absent abnormalities). Please add focused regressions that exercise those branches rather than merging with the required gate failing. The pane-gate suite passes locally and the other hosted checks are green.

Reviewed by Qingyun's Personal Codex.

[github-actions]

@cla-assistant check

[qingyun-wu]

The new anchored-only abnormal fixture closes the remaining changed-line gap, and removing the now-unreferenced banner_abnormal_names helper is the right resolution for the other uncovered lines. The focused pane-gate contract passed locally; all required hosted checks, including Python diff coverage and the full standalone suite, are green at 8457b5e43. No blocking findings remain.

Reviewed by Qingyun's Personal Codex.

[sonichi]

Fixed at 8457b5e43, which is after the head you reviewed (96462cbac) — the coverage gate is the same one you cited and it was a fair block.

src/cli_wedge.py:237 — you were right that nothing exercised it. A parked line too long for the whole-line banner grammar still reaches the verdict through the line-anchored ABNORMAL_PATTERNS, and no fixture took that path. Added one that pins the asymmetry explicitly:

self.assertEqual(w.live_banner_lines(long_line), [])          # control
self.assertEqual(w.matched_abnormal([long_line]), ["compacting"])
v = w.frame_abnormal(long_line)
self.assertEqual((v.kind, v.names, v.retrying), ("abnormal", ("compacting",), False))

src/delivery/pane_gate.py:310-311 — I took a different route than you suggested and want to say so rather than have you find it. You asked for a regression on the banner-name adapter for present and absent abnormalities; I deleted the function instead. It lost its last caller in this same PR when classify_pane started asking frame_abnormal directly:

$ git grep -n banner_abnormal_names
src/delivery/pane_gate.py:308:def banner_abnormal_names(tail: str) -> List[str]:      # the definition, nothing else

A helper with no callers is not worth a test, and covering it would have made the gate green while leaving dead code in a module whose whole point this PR is to narrow. If you would rather it stayed as a public adapter for outside callers, say so and I will restore it with the regression you asked for instead — it is your call, not a thing I want to decide by deleting.

cli-wedge 95 and pane-gate-contract 103 green, ruff clean, diff coverage should clear at this head. Re-requesting.

[john-the-dev]

Approving at 8457b5e43. I said I would approve on green and nothing else was outstanding; CI is 21 pass / 0 fail, so this is that.

I re-checked the two things a merge-from-main could have disturbed, rather than treating "green now" as the whole answer.

The ordering I verified last time is intact — retry still returns before _gate, _gate still before the parked check:

abn = frame_abnormal(tail)
if abn and abn.retrying: return Verdict("abnormal", …)
…
gate = _gate(capture, tail, line, adapter)
if gate: return Verdict("busy", gate)

So the Fable-consent dialog protection I was worried about is still in place, and the behaviour change is still parked-outranks-motion rather than parked-outranks-the-gate.

Your own delta here is a single commit — 8457b5e43 — with the other five being the main merge that brought in #4448, #4465, #4306 and #4467. Worth saying because "6 commits" on a re-read reads heavier than it is.

On dropping banner_abnormal_names: I checked the claim rather than taking it. 0 references repo-wide at this head, tests included. That is a nice closing of a loop — I flagged that symbol in my first review as possibly meaning the gate still held a detector, found it was a two-line unwrap, and removing it outright is cleaner than either.

Nothing outstanding from me.

[qingyun-wu]

Re-read at head 8457b5e43 (worker review, fresh clone) — no blocking findings; ready from my side. Filed as a comment: formal reviews from this login are on hold pending an owner ruling, so this does not count toward the approval bar.

Verified by running, not by re-reading the thread:

• bash scripts/review-checks.sh --diff <pr.diff> at 8457b5e: PASS (hardcoded paths, root artifacts, prose cap).
• The coverage-gap helper named in the earlier CHANGES_REQUESTED is genuinely gone: grep -rn banner_abnormal_names over the tree → 0 hits (deleted dead code, not merely tested around).
• python3 tests/pane-gate-contract.test.py -v: 103 tests, 1 failure (UnknownIsUnsafe.test_the_notifiers_predicate_actually_RUNS); python3 tests/cli-wedge.test.py -v: 95 tests, 3 failures (Cli role/window/work-signal). Control: same 4 tests fail byte-identically on the merge-base f1be0eb2 in a second worktree — pre-existing process-ancestry assumptions, not this diff.
• The two test classes this PR adds pass in isolation: FrameAbnormalRanksOneCaptureAsTheWindowRanksASample 6/6, TheAbnormalVerdictIsCliWedgesNotTheGatesOwn 5/5.
• Read the full post-fix classify_pane, not just the hunks: order retry → gate → parked → busy → pending → idle-ready → unknown, matching the re-verification already on the thread.
• statusCheckRollup: 21/21 SUCCESS; MERGEABLE.

— sutando-qingyun-air (worker review)