fix(qa): harness reliability — 4 gate-fidelity + #623 recovery fixes - #966
Merged
Conversation
…ry fixes The QA measurement harness was the meta-blocker: noisy/flaky signals that conflate harness failures with product quality. Four root-caused, TARGETED fixes (each preserves its integrity gate — risk-mitigations enforced + tested): 1. party_traveled false-cap (assert_behavioral.py): a COMPLETE single-location drama (clock advanced AND a quest resolved AND >=8 beats) no longer RED-caps every lens <=2.5 like a frozen stall. Conjunctive in-place-progression exception; a frozen opening (day==1/morning, no quest) still fails >=2 conjuncts and stays RED. Additive — can only let a run PASS. 2. persona rc!=0 mis-counted as a quality fail (sweep_v2.sh + ui_playtest_app.sh + release_readiness.py): a NON-quota player-process CRASH (player_rc!=0) is now retried once, then stamped part_b.harness_error → surfaced as a cross_persona_sat EVIDENCE GAP (harness_contaminated, still blocks release as "re-measure"), NOT a score_pass quality fail. Discriminator is the process exit, never the score — a played low-score run exits rc=0 and still fails the quality gate. 3. Mac-handoff PART_A_DEADLINE drift (ui_playtest_app.sh): the flat 420s mint deadline had drifted BELOW the DM cold-open's own 500/550s timeout → healthy-but-slow mints abandoned (flaky scripted/codex legs). Now derived from the cold-open tier + 90s margin, with a liveness-bounded grace (extends only while dm.combined.jsonl streams / the DM proc is alive; a dead cold-open fails promptly). can_act:true+minted_run PASS condition unchanged. 4. #623 silent-drop recycle + hang escalation (dm_beat_mark.py + lib_beat_driver.sh + play*.sh): (a) an empty pre-beat mark on a CONTINUING beat (first==0, a mark-write bug) no longer fails-open and stamps recycled prose "genuine" (masked dead beat); a true first-prose cold open (first==1) is unaffected. (b) a deadline-killed (rc=124) ROUTINE beat no longer escalates the retry to the cold-open tier (the ~14-15min single-beat hang) — it routes to the visible failure path; cold opens + non-124 transient errors still retry. Tests: assert_behavioral 38, release_readiness 47, dm_beat_mark 8 (new file), fast_gate 222 — all green.
|
Caution Review failedThe pull request is closed. ℹ️ Recent review info⚙️ Run configurationConfiguration used: Organization UI Review profile: CHILL Plan: Pro Plus Run ID: 📒 Files selected for processing (2)
📝 WalkthroughWalkthroughThe PR hardens the QA automation pipeline across four areas: (1) adds a ChangesQA Harness Hardening
Sequence Diagram(s)sequenceDiagram
participant sweep as sweep_v2.sh run_persona
participant playtest as ui_playtest_app.sh
participant release as release_readiness.py
sweep->>playtest: run Part A + Part B
playtest->>playtest: coldopen_is_live() liveness gate (Part A)
playtest->>playtest: player_rc != 0 AND no quota/429 text?
alt harness error (non-quota)
playtest->>playtest: PART_B_HARNESS_ERROR=true
playtest->>sweep: write run.json {part_b: {harness_error: true}}
sweep->>playtest: retry once (non-quota path)
playtest->>sweep: rc (after retry)
else quota / session limit
playtest->>sweep: short-circuit to quota path (no retry)
end
sweep->>release: aggregate run.json results
release->>release: part_b_harness_error → exclude from score_pass_failed_personas
release->>release: add to cross_persona_sat evidence_gaps
release->>release: exclude from arc_completed failed_part_b
release-->>sweep: harness_contaminated=true, harness_error_personas=[...]
Estimated code review effort🎯 4 (Complex) | ⏱️ ~60 minutes Possibly related issues
Possibly related PRs
Poem
🚥 Pre-merge checks | ✅ 4 | ❌ 1❌ Failed checks (1 warning)
✅ Passed checks (4 passed)
✏️ Tip: You can configure your own custom pre-merge checks in the settings. ✨ Finishing Touches📝 Generate docstrings
Comment |
![coderabbitai[bot]](https://avatars.githubusercontent.com/in/347564?s=60&v=4){kind=small}
There was a problem hiding this comment.
Actionable comments posted: 1
🧹 Nitpick comments (2)
qa/ui_playtest_app.sh (1)
1119-1121: 💤 Low valueMinor:
-rflag is unnecessary for named files.
grep -ris for recursive directory search. When targeting specific files ("$RUNDIR/backend.log" "$PLAYERDIR/player.err"), the-rflag has no effect but adds slight confusion. Consider using just-qiE.This doesn't affect correctness since grep handles it gracefully.
🤖 Prompt for AI Agents
Verify each finding against current code. Fix only still-valid issues, skip the rest with a brief reason, keep changes minimal, and validate. In `@qa/ui_playtest_app.sh` around lines 1119 - 1121, The grep command on lines 1119-1121 includes the `-r` flag, which is unnecessary when searching specific named files like "$RUNDIR/backend.log" and "$PLAYERDIR/player.err". The `-r` flag is only needed for recursive directory searches and adds unnecessary confusion here. Remove the `-r` flag from the grep invocation, keeping only the `-qiE` flags with the specific file paths.qa/test_assert_behavioral.py (1)
698-736: 💤 Low valueConsider adding test for clock-not-advanced with arc-resolved.
The four tests thoroughly cover the happy path and three failure modes. One edge case not explicitly tested:
day=1,tod="morning"(clock NOT advanced) withquest_completed=True(arc resolved). This would confirm thatclock_advancedis independently required, not just correlated witharc_resolved.Since the implementation uses a strict AND, the current tests provide adequate coverage—but an explicit test would guard against future refactors that might accidentally broaden the exception to arc-only.
🤖 Prompt for AI Agents
Verify each finding against current code. Fix only still-valid issues, skip the rest with a brief reason, keep changes minimal, and validate. In `@qa/test_assert_behavioral.py` around lines 698 - 736, Add a new test function after test_party_traveled_still_red_when_arc_resolved_but_too_few_beats that explicitly tests the case where the clock has not advanced (day=1, tod="morning") but the arc is resolved (quest_completed=True). This test should verify that party_traveled fails because clock_advanced is independently required by the AND condition, even when arc_resolved is true. Create test events with beats >= 8, set up state with day=1, tod="morning", quest_completed=True, and visited_count=1, then assert that the output contains "[FAIL] party_traveled" and the return code is 1, confirming that both clock advancement and arc resolution are strictly required by the exception logic.
🤖 Prompt for all review comments with AI agents
Verify each finding against current code. Fix only still-valid issues, skip the
rest with a brief reason, keep changes minimal, and validate.
Inline comments:
In `@qa/vm/sweep_v2.sh`:
- Around line 123-124: The retry gating condition at line 123 in the sweep_v2.sh
file uses the quota_tripped function with only the backend log (_bl variable) to
detect quota-related failures. However, the upstream classifier in
qa/ui_playtest_app.sh checks for quota/429 errors across both backend.log and
player/player.err files. Update the quota_tripped call to align with the
upstream behavior by including both log sources in the quota detection logic,
ensuring that quota-triggered aborts are correctly identified and do not consume
the single retry allocation.
---
Nitpick comments:
In `@qa/test_assert_behavioral.py`:
- Around line 698-736: Add a new test function after
test_party_traveled_still_red_when_arc_resolved_but_too_few_beats that
explicitly tests the case where the clock has not advanced (day=1,
tod="morning") but the arc is resolved (quest_completed=True). This test should
verify that party_traveled fails because clock_advanced is independently
required by the AND condition, even when arc_resolved is true. Create test
events with beats >= 8, set up state with day=1, tod="morning",
quest_completed=True, and visited_count=1, then assert that the output contains
"[FAIL] party_traveled" and the return code is 1, confirming that both clock
advancement and arc resolution are strictly required by the exception logic.
In `@qa/ui_playtest_app.sh`:
- Around line 1119-1121: The grep command on lines 1119-1121 includes the `-r`
flag, which is unnecessary when searching specific named files like
"$RUNDIR/backend.log" and "$PLAYERDIR/player.err". The `-r` flag is only needed
for recursive directory searches and adds unnecessary confusion here. Remove the
`-r` flag from the grep invocation, keeping only the `-qiE` flags with the
specific file paths.
🪄 Autofix (Beta)
Fix all unresolved CodeRabbit comments on this PR:
- Push a commit to this branch (recommended)
- Create a new PR with the fixes
ℹ️ Review info
⚙️ Run configuration
Configuration used: Organization UI
Review profile: CHILL
Plan: Pro Plus
Run ID: 9c1b3513-a2e6-43a9-8de2-5d4be760b9c2
📒 Files selected for processing (11)
qa/assert_behavioral.pyqa/dm_beat_mark.pyqa/lib_beat_driver.shqa/release_readiness.pyqa/test_assert_behavioral.pyqa/test_dm_beat_mark.pyqa/test_release_readiness.pyqa/ui_playtest_app.shqa/vm/sweep_v2.shscripts/play.shscripts/play_party.sh
Comment on lines
+123
to
+124
| if [ "$rc" -ne 0 ] && ! quota_tripped "$_bl"; then | ||
| note " $persona rc=$rc (non-quota harness crash) — retrying ONCE on a clean store" |
There was a problem hiding this comment.
Align retry quota detection with the upstream non-quota predicate.
At Line 123, retry gating checks only backend.log. The upstream Part B classifier (qa/ui_playtest_app.sh) treats quota/429 based on both backend.log and player/player.err; this mismatch can consume the one retry on quota-aborted runs.
Suggested fix
-quota_tripped(){ # $1 = a run dir or a log path; rc 0 if a session-limit/429 is present
- grep -qriE "session limit|HTTP 429|hit your (session|usage) limit" "$1" 2>/dev/null
+quota_tripped(){ # $@ = run dir(s)/log path(s); rc 0 if a session-limit/429 is present
+ grep -qriE "session limit|HTTP 429|hit your (session|usage) limit" "$@" 2>/dev/null
}
@@
- local _bl="qa/ui_playtest_runs/vm2-$persona/backend.log"
- if [ "$rc" -ne 0 ] && ! quota_tripped "$_bl"; then
+ local _bl="qa/ui_playtest_runs/vm2-$persona/backend.log"
+ local _pe="qa/ui_playtest_runs/vm2-$persona/player/player.err"
+ if [ "$rc" -ne 0 ] && ! quota_tripped "$_bl" "$_pe"; then🤖 Prompt for AI Agents
Verify each finding against current code. Fix only still-valid issues, skip the
rest with a brief reason, keep changes minimal, and validate.
In `@qa/vm/sweep_v2.sh` around lines 123 - 124, The retry gating condition at line
123 in the sweep_v2.sh file uses the quota_tripped function with only the
backend log (_bl variable) to detect quota-related failures. However, the
upstream classifier in qa/ui_playtest_app.sh checks for quota/429 errors across
both backend.log and player/player.err files. Update the quota_tripped call to
align with the upstream behavior by including both log sources in the quota
detection logic, ensuring that quota-triggered aborts are correctly identified
and do not consume the single retry allocation.
…tatus-blind tool count
Adversarial verification of FIX 4 found a real gate-weakening: arc_resolved used
coverage_from_tool_counts' quest_resolved bucket, which counts the set_quest_status TOOL NAME
regardless of the status argument. A frozen DM could pass party_traveled (and dodge the
behavioral RED-cap) on a dead scene with one advance_time + one set_quest_status(status="active").
Reproduced GREEN on a frozen single-scene run.
Fix: arc_resolved now requires an ACTUAL snapshot quest at status=="completed" (which a
set_quest_status("active"/"failed") does not produce). + a regression test that a frozen scene
with a set_quest_status tool call present stays RED. The legitimate single-scene-progression
case (a quest reaching status=="completed") still passes. assert_behavioral 39/39 green.
This was referenced Jun 16, 2026
100yenadmin
added a commit
that referenced
this pull request
Jun 16, 2026
…ness-aware contract (#623) (#970) test_coldopen_part_a_poll_window_outlasts_max_effort_coldopen pinned the literal `PART_A_DEADLINE="${WOS_APP_PART_A_DEADLINE:-420}"` and the old `seq 1 "$part_a_polls"` for-loop — both of which FIX 3 (#623, shipped in #966) intentionally replaced. #966's fast_gate doesn't run test_adversarial_release.py, so the stale test slipped through and left main RED. The production code is correct: 420 was SHORTER than the cold-open's own model-aware timeout (clawdnd_dm_timeout 1 = 500/550s), so a healthy-but-slow mint was abandoned ~80s early — the exact flake the test NAME warns about. FIX 3 derives the deadline FROM that tier + a positive margin and makes the poll liveness-aware. Rewrite the test to assert that *contract* semantically (extract the margin via regex, assert it's > 0 = the real "outlasts" guarantee; assert the hard-deadline + coldopen_is_live grace; keep the no-hardcoded-`seq 1 70` guard) — so a regression to a flat / too- short / non-positive-margin deadline now fails here, but the intended derivation passes. Test-only; 18/18 in test_adversarial_release.py green; fast_gate 222 intact. Co-authored-by: Eva <arncalso@gmail.com>
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Root-caused via a 4-surface analysis workflow; each fix is TARGETED and preserves its integrity gate (risk-mitigations tested). (1) party_traveled no longer false-caps a complete single-scene drama; (2) a persona player crash (rc!=0) is retried+reclassified as harness-inconclusive (evidence gap), not a quality fail; (3) PART_A_DEADLINE derived from the cold-open tier + liveness grace (fixes flaky handoff legs); (4) #623 empty-mark recycle + hang-escalation closed. Tests: assert_behavioral 38 / release_readiness 47 / dm_beat_mark 8 / fast_gate 222 green.
Summary by CodeRabbit
Release Notes
Bug Fixes
Tests