Repository navigation
merge queue: checking #1336 on main (2745ccd), stacked on 4 pull requests ahead - #1347
Closed
mergify[bot] wants to merge 30 commits into
Closed
mergify[bot] wants to merge 30 commits into
mergify[bot] wants to merge 30 commits into
Conversation
Agents kept narrating by-model health and patching charts before splitting volume, fill, resolve, and filter — lock that check into standing defaults. Co-authored-by: Cursor <cursoragent@cursor.com> Change-Id: I8efc336daf9878751865cf3c4bac4d55cd50733f
Phrase-judge rejected leads that named only the trigger phrase; state the split and the stop-first catch as the lesson, with the triggers as examples. Co-authored-by: Cursor <cursoragent@cursor.com> Change-Id: I7dad9e54f9d280bb4d8f41ac06662293e39c9e05
…mple phrasing. Co-authored-by: Cursor <cursoragent@cursor.com> Change-Id: Idebc0f0ba74e47864bf479446a8e27b0574cdcde
…lesson. Co-authored-by: Cursor <cursoragent@cursor.com> Change-Id: Iaa84683fcf5a66e8f70a0dbf99ebcf19acc1d7a7
The dashboard-health defaults name that audit flag in backticks; the skill-existence check must treat it as a process token, not a skill path. Co-authored-by: Cursor <cursoragent@cursor.com> Change-Id: I1afa1d44a91d2886300ab5f17843b8e583dbda0c
Rule-scope rejected a lead that named automate-mandate mid thrash; keep named flags only as examples after the same-type-catch rule. Co-authored-by: Cursor <cursoragent@cursor.com> Change-Id: Ieedc8823a35a6be02d6b392d5c704f1b817d7ea1 Co-authored-by: Cursor <cursoragent@cursor.com>
tests/test_install.py and engine/hooks/_runner/tests/test_wrap_installed.py asserted one settings entry per hook, so every per-hook guarantee they check failed once CATSTACK_HOOK_DISPATCHER (#1122) collapsed an event to a single _runner/dispatch.py entry: 70 failures and 1 error on #1122's head, all in these two files. load_config() in tests/test_install.py now reads an installed config and expands a collapsed event back into one entry per hook, resolving each hook's registration the way dispatch.py itself does (dispatch.load_event_hooks: the cursor registry, or the hook's own manifest for claude and codex). Each assertion therefore sees the identical shape in both layouts and keeps checking the same guarantee -- right event, right harness, Stop mirrored to SubagentStop or opted out, dead entries pruned, non-catstack entries untouched. test_wrap_installed.py's fixture pins the flag off, since it asserts the settings bytes install writes rather than a hook's registration. #1122's own test_default_is_off_and_byte_identical_to_no_flag_at_all now compares the default install against an install that sets the flag to whatever flags.py resolves as the default, and clears the key from its own environment first, so the flag-on run of the full suite cannot leak into the "default" side of the comparison. The CATSTACK_HOOK_DISPATCHER default is left off. Flipping it on is gated on both full-suite runs passing, and the source branch plan/hook-backlog-1c-install-tests-accept-the-one-entry-per-event-layout-default-on-only-if-both-suites-pass never changed engine/hooks/_flags/flags.py, so that gate was not met there and this commit does not claim it. Cherry-picked from 6330abf on that branch. The branch's two other content commits (engine/hooks/_runner/doctor.py, scripts/test/run_skill_scenarios.py) are unrelated to this claim and are left behind.
…ew claim: tests.test_install and test_wrap_installed accept both install layouts (one entry per hook with CATSTACK_HOOK_DISPATCHER off, one dispatcher entry per event with it on), and the flag defaults to on only if bash scripts/test/run_all_tests.sh passes with the flag on and off. Review lane: behavior Safety invariant: If either suite run fails, the default stays off. Per-hook guarantees stay checked in the dispatcher layout. Files: tests/test_install.py, tests/test_wrap_installed.py Change types: test Acceptance criteria: This workflow publishes one pull request for the install-layout claim only. Exit code: 0 Invoker-Finalize-Id: f459f9c9-a623-4fd7-9708-21cc3ae4cbda
…s/g0.t0.a-a464d403f-4908b6a4 — Review claim: tests.test_install and test_wrap_installed accept both install layouts (one entry per hook with CATSTACK_HOOK_DISPATCHER off, one dispatcher entry per event with it on), and the flag defaults to on only if bash scripts/test/run_all_tests.sh passes with the flag on and off. Review lane: behavior Safety invariant: If either suite run fails, the default stays off. Per-hook guarantees stay checked in the dispatcher layout. Files: tests/test_install.py, tests/test_wrap_installed.py Change types: test Acceptance criteria: This workflow publishes one pull request for the install-layout claim only.
Review claim: - An opt-in payload recorder in the runner plus a replay script can feed recorded hook-event payloads through installed hook commands and produce a verdict-equivalence report (stdout/context/block/exit) per detector. Review lane: behavior Safety invariant: - Recording is off unless CATSTACK_HOOK_PAYLOAD_DIR is set; when on it never alters hook stdin/stdout/exit and caps each payload (default 256KB) and total dir (default 2000 files, oldest evicted). - Replay runs against a scratch metrics/state dir, never live settings or the live runs.jsonl. Goal: - Runner: when env var set, append the raw stdin payload to $CATSTACK_HOOK_PAYLOAD_DIR/<payload_hash>-<hook>.json before spawning the hook child, where <payload_hash> is sha256(stdin)[:12]. - scripts/test/replay_hook_payloads.py: for each recorded payload, run the hook command under a scratch env, capture stdout/stderr/exit/ findings; emit an equivalence report JSON + table comparing two command sets (fleet vs dispatcher args taken as parameters). Motivation: - The catastrophic regression risk of the dispatcher is a detector that silently stops speaking; the only credible gate is byte-identical verdicts on real payloads. Cherry-picked from origin/plan/hook-fleet-dispatcher-measure-the-herd-then-dispatch-it (e356e8c), scoped to only this claim's files. Dropped this slice's dependency on the separate metrics-event-uid claim (17ef151, not on this branch): the payload recorder computes its own local hash instead of reusing a published `event_uid` metrics-row field, and the README wording and test imports were adjusted to match. Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
… payload recorder in the runner plus a replay script can feed recorded hook-event payloads through installed hook commands and produce a verdict-equivalence report per detector. Review lane: behavior Safety invariant: Hook stdout and exit codes stay the same. This slice does not change the other four claims. Files: engine/hooks/_runner/run.py, scripts/test/replay_hook_payloads.py Change types: bugfix Acceptance criteria: The branch contains only this claim, and the review step publishes one pull request. Exit code: 0 Invoker-Finalize-Id: 61e213a0-a6cd-4950-ba23-0acc17742997
…-1bc38490 — Review claim: An opt-in payload recorder in the runner plus a replay script can feed recorded hook-event payloads through installed hook commands and produce a verdict-equivalence report per detector. Review lane: behavior Safety invariant: Hook stdout and exit codes stay the same. This slice does not change the other four claims. Files: engine/hooks/_runner/run.py, scripts/test/replay_hook_payloads.py Change types: bugfix Acceptance criteria: The branch contains only this claim, and the review step publishes one pull request.
The fleet self-saturates: one SubagentStop burst spawned 15-30 processes that each re-parsed the same transcript, measured at 15.1s p50 wall idle against 1.55s for a single process, with per-hook durations escalating in launch order. dispatch.py now runs opted-in detectors in-process, sequentially, parsing the event once and passing the same parsed payload to every detector. A hook opts in with subagent_stop.dispatch in its manifest, naming a MODULE:FUNCTION entry point (default detect:detect). A detector with no resolvable entry point still runs as a subprocess and its metrics row is marked subprocess_fallback. Each detector keeps its own verdict, its own runs.jsonl row under its own hook name, and its own rule_ids; a detector that raises, exits, or overruns its 4s soft budget is caught and recorded without stopping its siblings. The event budget defaults to 9s for SubagentStop. mirror_stop_hooks_to_subagent_stop.py emits one dispatcher entry for the opted-in hooks and leaves every other hook's direct entry alone. (cherry picked from commit 531f5c8 on plan/hook-fleet-dispatcher-measure-the-herd-then-dispatch-it)
…op hooks dispatch through one process that parses the event once and shares one transcript parse across all opted-in detectors, emitting the same per-detector metrics rows and verdicts as the subprocess fleet. Review lane: behavior Safety invariant: Hook stdout and exit codes stay the same. This slice does not change the other four claims. Files: engine/hooks/_runner/dispatch.py Change types: bugfix Acceptance criteria: The branch contains only this claim, and the review step publishes one pull request. Exit code: 0 Invoker-Finalize-Id: aaa80940-f2d5-4d79-bb40-79c68fe333ad
…e5bede34 — Review claim: SubagentStop hooks dispatch through one process that parses the event once and shares one transcript parse across all opted-in detectors, emitting the same per-detector metrics rows and verdicts as the subprocess fleet. Review lane: behavior Safety invariant: Hook stdout and exit codes stay the same. This slice does not change the other four claims. Files: engine/hooks/_runner/dispatch.py Change types: bugfix Acceptance criteria: The branch contains only this claim, and the review step publishes one pull request.
Subprocess isolation was the hook fleet's free safety property, so the dispatcher has to prove it reproduced that property rather than assume it. _run_sdk_one caught a raising detector but left exit_code at 0, so outcome.classify saw a zero exit with a catstack-hook-error stderr line and recorded caught_error instead of crashed; the except branch now sets exit code 1. Four chaos fixtures under engine/hooks/_runner/tests/chaos (raise, sleep past the soft budget, exit 3, and a sibling that speaks) drive a dispatcher test asserting each failure gets its own row with the right outcome and exit code while both sibling detectors still emit their verdicts and rows. The fixtures are test-only: they live under tests/ and are registered in a scratch manifest, so the mirror script's name allowlist can never install them into a live settings.json. (cherry picked from commit b57f27c on plan/hook-fleet-dispatcher-measure-the-herd-then-dispatch-it)
…tectors that raise, hang past the soft budget, and exit non-zero each produce their own failure row while every sibling detector's verdict and row is still emitted. Review lane: behavior Safety invariant: Hook stdout and exit codes stay the same. This slice does not change the other four claims. Files: engine/hooks/_runner/tests/chaos Change types: bugfix Acceptance criteria: The branch contains only this claim, and the review step publishes one pull request. Exit code: 0 Invoker-Finalize-Id: 7ca2543e-d062-46fa-bbce-58264f6ae48f
…02624463 — Review claim: Fixture detectors that raise, hang past the soft budget, and exit non-zero each produce their own failure row while every sibling detector's verdict and row is still emitted. Review lane: behavior Safety invariant: Hook stdout and exit codes stay the same. This slice does not change the other four claims. Files: engine/hooks/_runner/tests/chaos Change types: bugfix Acceptance criteria: The branch contains only this claim, and the review step publishes one pull request.
The SubagentStop dispatcher records an in-process exception as exit 0. This pull request records that same exception as a crashed row, so a sibling can still speak while the failure stays visible. Co-authored-by: Cursor <cursoragent@cursor.com> Change-Id: If063e8ba2ce12336b8a515be81a05bb1885d2ba9
2 tasks done
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.
✨ Pull requests batch (#1332, #1314, #1309) which was ahead in the queue failed. The pull request #1336 has been requeued. ✨
#1336 is queued for merge on branch main (2745ccd).
Stacked behind 4 pull requests queued ahead of this batch, not part of it. These checks run on a tip that also carries their commits, so a failure here can come from them as much as from #1336.
Queued ahead of this batch:
This pull request has been created by Mergify to speculatively check the mergeability of #1336.
You don't need to do anything. Mergify will close this pull request automatically when it is complete.
Required conditions of queue rule
admin-bypassfor merge:check-success = testcheck-success = lintcheck-success = validateRequired conditions to stay in the queue:
-draftbase=mainlabel=admin-bypass