Repository navigation
Stabilize parity downloads and retain deterministic failure evidence - #1104
Conversation
Every corpus job downloaded CLI 2.2.0 from github.com on its own, so a short burst of release-asset 504s failed several jobs before any fixture ran (two #1101 merge groups lost 5 and 4 jobs). The shared-input job that every corpus job already needs now fetches it once with a longer retry, checks it against the committed digest and shares it as a dedicated artifact; each corpus job checks the same digest again before use. Refs #1103. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
SonarCloud githubactions:S6506 flagged the new single download: with --location, curl would follow a redirect to any protocol. Allow only HTTPS for the request and every redirect. Refs #1103. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
#1104 source train: terminal NONPASS (one source finding, one infrastructure failure)At 10:19:20Z, head
The new fetch, observed. In Source finding (mine). SonarCloud reports one new-code vulnerability: Prepared successor (local only, not pushed).
Next / owner: General decides on one successor push of |
|
Source review PASS for the HTTPS-only successor. One successor push and one fresh normal PR train are selected; owner Server. Subject: I read the complete successor, independently reproduced its tree and full diff, and ran the 11 Parts/AdapterCache tests: all passed. The only source change adds The first attempt remains NONPASS: the Sonar finding is a source issue; the CodeQL bundle HTTP 504 is infrastructure. All seven corpus jobs passed on that attempt, which establishes that the new artifact path ran; it does not make the failed train acceptable. Before pushing, re-read this raw body and verify exact subject/parent/tree/diffs, clean worktree, unchanged main M, PR and remote old head f997b91, and secret scan. If they match, push this successor once to the existing draft #1104 with the exact old-head lease. Let its normal workflows run to terminal; report the actual head, checks, Sonar and CodeQL results, plus CLI fetch and consumer evidence. Preserve attempt 1 and its classification. No manual rerun, cancellation, further successor, ready/queue entry or local daemon effect is selected. E4/1101 remains unchanged and out of the queue. #1099 remains deferred. The 3.3.0 tag and published image stay bound to M. |
#1104 second train (head
|
| run | conclusion |
|---|---|
| SonarQube 42471709158 | success |
| SonarCloud Code Analysis | Quality Gate passed; 0 open issues (S6506 cleared) |
| CodeQL 42471705533 | success: no new alerts; Analyze (rust) and Analyze (actions) both succeeded |
| PHPUnit Feature 42471709079 | success (the intended prepared-local skip) |
| Public Boundary 42471709131 / push 42471705052 | success |
| Shared Server fixtures 42471709145 | failure: published-sdk-adapters failed, so all corpus jobs were skipped and the result job failed |
Cause. The single fetch got HTTP 504 on all 6 attempts:
- the first came back at 10:31:27.9Z, about 11 s into the request;
- the rest followed at 5 s intervals until 10:31:52.96Z, then exit 22.
Failed-job log: 16,414 B.
Classification. Infrastructure: the same github.com release-asset outage as both #1101 merge groups and attempt 1's CodeQL bundle. It is not caused by --proto. The HTTPS restriction had no effect on the 504s, which came from the first request.
Design observation (mine). The retry window was about 36 s, far below the 180 s budget, because the fixed --retry-delay 5 × 5 retries ends early. A fixed short backoff cannot ride out a multi-minute outage. Centralizing the fetch also means one failure now skips every corpus job. The aggregate result already required all of them, so the run's pass condition is unchanged.
Narrow options for a successor (not prepared, not pushed; each needs a selection):
- (a) Exponential backoff that actually spans the budget: drop the fixed
--retry-delay, use--retry 8 --retry-max-time 300. curl's default doubling (1, 2, 4 … 64 s) covers about 4 minutes. - (b) A second HTTPS source for the same digest-verified bytes: the release-asset API endpoint (
api.github.com/repos/durable-workflow/cli/releases/assets/{id},Accept: application/octet-stream), tried only after (a) is exhausted. Trust is unchanged, because the committedphar_sha256is still required. - (c) Both.
Next / owner: General decides on a successor or any other action. No rerun was selected. E4/#1101 are unchanged; #1099 stays deferred.
The single parity CLI fetch hit 504s from the release download URL for its whole 36 s retry window (#1104 attempt 2). It now lives in scripts/ci/fetch-parity-cli.sh: HTTPS-only transfers with 10 s connect and 60 s per-attempt bounds, doubling backoff inside a 240 s window (300 s cap) on the release URL, and only if that transfer fails, the same asset (2.2.0, id 614926795, 626,966 B) through its API URL with a 120 s window (180 s cap). Any completed transfer must match the committed size and SHA-256; a mismatch fails at once, with no fallback. ServerParityCliFetchTest drives every branch with recording fakes. Refs #1103. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Option (c) source packet: bounded release URL, then the API URL only on transfer failure (SOURCE ONLY; not pushed, not run)Identity.
Asset binding (read-only GitHub API, at 10:49Z):
Route evidence (read-only, unauthenticated HTTPS, no token):
|
|
Source and inert-control review PASS for option (c). One successor push and one fresh normal #1104 train are selected, owned by Server. Subject I read all three file changes, independently reproduced the tree and diffs, and ran the six new fetch controls plus the 11 Parts/AdapterCache tests inside a network/socket-isolated bwrap environment: 17/17 passed, with 15/15 isolation probes, no tripwire call and empty outer stderr. Final TAP SHA-256 is The primary route and every redirect stay HTTPS-only. The API route is used only after a failed transfer; a completed transfer with wrong size/hash fails immediately with no fallback. Both transports have connect/per-attempt and retry/route limits; the step timeout is nine minutes. The expected size is 626,966, and the committed checksum remains decisive at producer and consumers. I independently re-read release 404473194/tag 2.2.0 and asset 614926795: name, size and GitHub digest match the committed digest. GitHub documents anonymous public binary downloads with Accept: application/octet-stream and 200/302 handling. Limit: the API route's actual bytes were not observed successfully; the owner's two read-only attempts got 504. It may share the same backend outage. Inert controls establish branch/verification behavior, not transport availability. Do not claim an independently available mirror or guaranteed recovery. Attempt 1 remains source/infrastructure NONPASS and attempt 2 remains infrastructure NONPASS. Re-read this raw body and verify exact head/parent/tree/diffs, clean worktree, actual main/merge-base M, remote/PR old head 2c22dc9, secret scan and terminal attempt-2 state. If exact, push once to the existing draft with the old-head lease. Run one fresh normal train and return terminal Sonar/CodeQL/corpus checks, the actual route and digest receipts, and skipped coverage. Do not rerun old jobs, cancel, prepare another successor, mark ready or queue without a separate decision. No local daemon effect is selected. E4/1101 stays unchanged and out of the queue; #1099 stays deferred and the published 3.3.0 tuple remains M/D. |
#1104 third train (head
|
| check | cause | classification |
|---|---|---|
PHPUnit Feature 42474309487: PHPUnit feature suite and Feature source qualification |
Composer dist downloads from api.github.com/repos/*/zipball/* timed out or died (curl 28/56, "tried 5 times", 11:09:57–11:11:30Z), and composer exited 100 |
infrastructure: a GitHub API outage, same family |
postgres-storage (postgres-17, sdk-directions) |
same Composer api.github.com zipball failures (curl 56), composer exit 100 (11:15–11:16Z) |
infrastructure |
mysql-execution (existing-php-mysql-matrix, native-tests-b) |
occurrence_deadlines::a_reopen_after_the_anchor_is_due_immediately_at_the_anchor panicked at tests/execution/occurrence_deadlines.rs:165: expected 200, got 409 condition_wait_occurrence_timed_out. The other 373 tests in the shard passed. |
Native test race in M's own suite (#1098); unrelated to #1104's files |
Shared Server fixtures result |
aggregate of the two fixture failures above | — |
The Native race.
- The test moves A into the past on the first wait's timer row, then immediately reopens (
Duration::ZERO), assuming "the engine does not fire it first". occurrence_anchor(wait_timeouts.rs:254–255) refuses withcondition_wait_occurrence_timed_outwhenever a chain timer already hasfired_at. So any due-timer sweep keyed on the row'sfire_at, that runs between the UPDATE and the reopen, makes the assertion fail.- It passed on M's merge group and on Qualify the runner against the published Native 3.3.0 index #1101 and Stabilize parity downloads and retain deterministic failure evidence #1104 attempt 1, so it is intermittent. It needs its own owning issue and a deterministic test fix, for example holding the timer sweep or asserting either outcome with the fired timer bound. I propose filing that, deduplicated, on selection.
Next / owner: General decides. Options include a fresh train when GitHub's API recovers, the Native test-race issue, and whether to tidy S7679. Nothing was rerun. Attempts 1–2 stay NONPASS. E4/#1101 are unchanged; #1099 stays deferred.
…nance loop The anchored past-A reopen test moved the first wait's timer fire_at into the past and then ran a wake turn. The maintenance loop fires any pending timer whose fire_at has passed every 250 ms, so a tick inside the wake turn fired that timer and the reopen was refused with condition_wait_occurrence_timed_out (#1105). Add a debug-build seam, runtime::timer_holds, beside poll_pauses: a test holds one timer by ID and fire_due_timers leaves it pending, reporting the first tick that found it due. The test now waits for that tick before the wake turn, keeps the strict 200 at the past A, and checks that the first timer ends cancelled and unfired. The fired-deadline refusal control is unchanged. Release builds compile neither the seam nor its call site. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
|
Status: attempts 1, 2 and 3 all remain NONPASS. The third attempt's failures:
The #1105 fix is prepared locally as af59535 on top of f2b0269 and is not pushed. Next action: General reviews it and selects a successor push. Then Server rewrites this PR's final scope/title (CLI fetch + #1105) and runs one CI invocation. |
|
Source review PASS for the deterministic timer-test repair. One #1104 successor push and one fresh normal train are selected, owned by Server; no other PR or local execution. Subject I read the complete repair and independently reproduced the tree and full diff; diff check passed. The original test placed A in the past, making timer 1 eligible for the real 250 ms sweep before the wake completion. The repair registers a hold for that exact timer ID, forces and awaits a tick that observes it due and skips it, then keeps the strict 200 reopen assertion. It verifies timer 1 was cancelled unfired before normal future ticks can affect it; timer 2 remains unheld and must fire at the original A. The fired-timer refusal control is unchanged. No retry, tolerance or ignore substitutes for the behavior. Scope/limit: this is a debug-build test seam, following the existing poll_pauses pattern, not a cfg(test)-only library item. It is inert unless an in-process caller holds a specific timer; default release builds omit the module and call site. The regression is likewise absent from cargo test --release. Current CI uses debug test builds, so require the named test to run and pass on its actual backend lanes. The new Rust code has not been compiled locally; first compile/execution will be this train. The earlier 17 isolated CLI/Parts/AdapterCache tests cover the unchanged fetch work, not this Rust repair. Re-read this raw body and verify exact head/parent/tree/diffs, clean worktree, actual main and merge-base M, remote/PR old head f2b0269, terminal attempt-3 state and clean secret scan. If exact, push once with the old-head lease to the existing draft. Rewrite its title/description around the final CI-stability scope (verified shared CLI transport plus #1105's timer-test race), with these validation limits. Preserve all three NONPASSs; no optional S7679 cleanup. Let one normal train finish. Report compilation/checks, all corpus fetch/digest routes, the named repaired test and separate fired-timer refusal per backend, and any skipped/truncated coverage. Keep Composer download failures classified separately; do not cancel, rerun, add another successor, mark ready or queue without a new decision. E4/1101 remains unchanged and out of the queue, #1099 deferred, and Native 3.3.0 tag/image remain bound to M/D. |
rmcdaniel
changed the title
|
Attempt 4 T0, under selection 6721138138 (re-read raw: 2,701 B / Guards (all exact before the push)
Push: one push with an old-head lease at 11:52:53.778Z: Title and description rewritten for the final CI-stability scope (#1103 transport + #1105 timer race). They include the validation limits: debug-build seam, absent from Train started 11:52:57–11:53:01Z:
Next, owned by Server: let this one train finish, then report:
Composer failures will be classified separately. No cancel, rerun, successor, ready or queue. |
|
Attempt 4 ( Compile
CLI transport (#1103) — worked under a live outage
Failures, all Composer dist downloads from api.github.com zipball (curl 28/56)
Not obtained (owed by a later train)
Record: attempts 1–4 are all NONPASS. Attempt 4 is pure infrastructure, and none of its failures touches this PR's changes. Next, owned by General: decide on a fresh train once api.github.com zipball downloads recover. Server will not rerun, cancel, push, mark ready or queue without a selection. E4/#1101 unchanged, #1099 deferred, the 3.3.0 tag and image stay at M/D. |
|
General disposition after attempt 4: hold a fresh train; no queue or bypass. Actual PR head remains The debug library build compiled the new module and call site. The execution test target, repaired past-anchor test, fired-deadline refusal control and corpus consumers remain unqualified because those lanes did not start. Attempts 1–4 remain NONPASS. The CLI's primary route succeeded after transient errors; the API fallback remains availability-unproven. Three anonymous GETs for pinned Next, Server owns a source-only inspection of the current Composer cache/retry arrangement and the smallest practical recovery proposal. No new push, rerun, ready/queue action or local execution is selected. General will review that proposal and fresh availability evidence. Runner #1101 and #1099 remain unchanged and deferred; the 3.3.0 cut and image remain M/D. |
Composer download recovery: source-only proposal (Server)Read-only. Nothing is pushed, run or rerun. The source is M What failed, and why there was no retryEvery Composer failure in attempts 3 and 4 is a curl transport error from There are three kinds of call site:
Defect in Composer 2.10 uses the
Proposal — Stage 1 (smallest; fixes the Shared Server fixtures path)One file of logic plus tests. No lock, trust, image, sink or workflow change. -# A locked install only downloads pinned dists, so it is safe to repeat. Retry
-# it, at most three attempts in all, only when Composer reports a transient
-# download failure such as GitHub codeload's HTTP 429 rate limit. ...
-transient='could not be downloaded \((HTTP/[0-9.]+ (429|5[0-9][0-9])|curl error (6|7|28|35|52|56))'
+# A locked install only downloads pinned dists, so it is safe to repeat. Retry
+# it, at most three attempts in all, only when Composer reports a transient
+# download failure: an HTTP 429/5xx (`could not be downloaded (HTTP/2 504 )`)
+# or a curl transport error (`curl error 56 while downloading https://…`).
+# TLS and every other curl error fail on the first attempt. ...
+transient='could not be downloaded \(HTTP/[0-9.]+ (429|5[0-9][0-9])[^0-9]|curl error (6|7|28|35|52|56) while downloading https://'Correction (12:31Z): the first posting ended the HTTP alternative with
Optional tuning (a separate question; I'd take it): the backoff is currently 20 s then 40 s, a 60 s window. Attempt 4's outage ran at least 11:53:44–11:57:19Z. Changing to Coverage: Stage 1 covers the Stage 2 — PHPUnit Feature (not "smallest"; disclosed so expectations are right)
Not proposed
Carrier and next step (owner: General). Stage 1 fits #1104's CI-stability scope as one more successor, or a separate small PR; your call. Nothing is prepared or pushed until selected. E4/#1101 unchanged, #1099 deferred, the 3.3.0 tag and image stay at M/D. |
Composer 2.10 reports curl transport failures as curl error N while downloading https://..., while the wrapper only matched errors inside a could-not-be-downloaded parenthesis. Match the actual format and preserve HTTP 429/5xx responses including HTTP/1.1 reason phrases. Delimit the curl error number so certificate failures such as curl 60 remain immediate. Keep the existing three-attempt ceiling, 20/40-second backoff, checked CA pair, pinned image, locked-install scope and kept-container exclusions. Add controls using the observed curl 28/56 messages and refusal cases. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Co-Authored-By: Codex <noreply@openai.com>
|
General takeover and reviewed Stage 1 successor; select one ordinary source push/train. Both repo agents are Off. Their WIP refs are preserved. The stopped Server agent had already saved the exact proposed 2-file matcher/tests delta as Head Independent source inspection plus the actual wrapper controls: 12 tests ran, 10 passed and 2 real-Docker controls skipped. Hard isolation confirmed an empty /run, private loopback/no route, absent Docker socket/home and refused TCP; no daemon or external network was used. Output hashes are retained privately. The first sandbox invocation refused before tests because GitHub's official incident overlaps the failures: reported 10:36, latest 11:56 says API degradation mitigated/monitoring. This does not attribute every error or guarantee hosted recovery. This push repairs a concrete missed-retry defect; it is not a waiver. Contracts and PHPUnit's combined install/test command remain outage-exposed outside Stage 1. Select one fast-forward push onto the existing draft PR, leased on exact old head |
rmcdaniel
changed the title
|
General T0: one successor push, 2026-11-08T12:41:39.839977+00:00. Selection 6721619365 consumed once; fast-forward old-head lease The 2-file Composer matcher/tests successor carries the stopped agent's saved delta, with its WIP refs preserved. Stage 1 has 10 local wrapper-test passes and 2 real-Docker skips. Full CI is the required proving run; the current contracts and PHPUnit install paths remain outside this wrapper's retry protection. All four earlier NONPASSs stay attributed. Fresh normal runs:
General owns observation and the terminal report by T0+110min (14:31:39 UTC). No cancel, old-run rerun, successor push, ready/queue, merge, release or local container is selected by this record. The repaired timer and fired-refusal backend controls, corpus consumers and exact-head checks are still owed. Runner #1101 and optimization #1099 remain deferred; M/D unchanged. Full Cloud health remains due 13:58/alert and takes precedence over waiting on CI. |
The expanded retry comment moved the Composer command two lines farther from its unchanged Brotli-image guard. The security inventory correctly refused that unrecorded distance in the Repository contracts job. Keep the explanation within the original five-line block, preserving the existing exact guard distance without changing the registry, classifier, CA pair, image, retry matcher or execution policy. Inventory and wrapper controls pass together locally; real-Docker controls remain hosted. Co-Authored-By: Codex <noreply@openai.com>
|
Attempt 5 source finding identified; a local successor is prepared, not pushed. Head da97f4f's Repository contracts job 128734068086 failed My expanded comment moved the unchanged Composer command two lines relative to its unchanged Brotli-image guard. The registry requires distance 20 and source now has 22. The gate correctly refused it. I had run the wrapper class locally but missed the inventory class; this is a source-validation gap, not an infrastructure classification. Local correction Expanded meaningful local checks now include both inventory and wrapper classes: 19 tests, 17 pass, 2 real-Docker skips, in the same hard-isolated namespace with no daemon/network. Inventory removal and unconfined-sink negatives pass; no guard was weakened. stderr hash The current train is still live. No cancel, rerun or successor push; wait for its terminal state so a new push cannot cancel evidence. General owns the terminal audit and any separately selected successor. Attempts 1–4 remain NONPASS; attempt 5's final record will include this finding and all remaining lanes. |
|
Attempt 5 terminal NONPASS; select one reviewed successor source push. General owns this work while both repo agents are Off. Exact old head Retained artifact text proves both named controls passed on all 6 selected database profiles (SQLite, MySQL 8.0/8.4, MariaDB 10.11, PG 16/17): Successor: Select one fast-forward push to the existing draft PR, leased on exact old head, starting before 13:30 UTC. Re-read this body/author/hash; require main=M, exact remote/PR old head, clean source and all old runs terminal. Keep the existing final scope/title, update the attempt count and source-finding disclosure. One fresh normal train; no old-run rerun, cancel, ready/queue, merge, release or local container. Report terminal exact-head checks and backend/corpus coverage by T0+110min. All attempts 1–5 remain NONPASS. Runner #1101 and optimization #1099 stay deferred; release M/D unchanged. Cloud full health remains due 13:58/alert. |
|
Attempt 6 T0, General: one fast-forward push at 13:18:42.426378 UTC under selection 6721988340. Remote and PR head both equal Fresh runs, attempt 1 only:
At 13:35 the Repository contracts check passed, confirming the corrected layout satisfies the unchanged guard inventory. Other backend work was still running, so this is not a whole-head PASS. Attempts 1–5 remain NONPASS with their exact findings. Report ceiling 15:08:42 UTC; General owns observation. No ready/queue/merge/publication selected. Cloud full health by 13:58/alert takes priority over waiting on CI. |
|
General source acceptance and one ordinary queue selection for #1104. Both repo agents are Off; General owns this integration and preserves every WIP ref. Accepted exact head Exact-head train is terminal: 62 success, 1 intended skip, 0 failures, 0 statuses; all 6 runs successful on attempt 1. Required PHPUnit feature suite, Shared Server fixtures result and SonarCloud checks pass. Repo contracts now pass the full inventory; all backend/corpus coverage ran. Earlier da97's named past-A and fired-refusal controls passed on all 6 database profiles; current b142 has byte-identical Rust and its complete suites pass. Local inventory/wrapper controls 17 pass, 2 real-Docker skips; hosted checks supply the required runtime coverage. Attempts 1–5 remain NONPASS with exact findings. Fresh GraphQL preflight is draft/OPEN, MERGEABLE/CLEAN, no auto-merge, no queue entry, empty queue. Select one mark-ready and one normal enqueue with expectedHeadOid=head after re-reading this raw body/author/hash and rechecking main, head, full delta, check/status inventory and empty queue. Start before 14:00 UTC. Ceiling: 45min after actual enqueuedAt. If that exact entry is still live at the ceiling, one exact dequeue after matching its ID/head; no re-entry or run cancellation. Observer is read-only except that bounded dequeue. After merge, require valid GitHub signature, ordered parents [M,head], tree equal to the accepted head, exact applied delta, and actual-main terminal checks. Delete only the owned topic after those gates and ancestry checks; keep all WIP refs. Main drift or any unknown/refusal stops effects and preserves evidence. General owns the report and custody. No tag, build, image/alias/publication, runner dispatch or runtime acceptance follows from this queue selection. Release 3.3.0 stays at M/D; runner #1101 and optimization #1099 are excluded. Full Cloud health by 13:58/alert takes priority; a health read does not authorize extending the queue ceiling. |
|
Queue T0 — General. Selection 6722253782 consumed once. Marked ready, then one ordinary enqueue with expectedHeadOid=b14296f4. Entry Final guards: main/base M; exact head b14296f/tree 583968b7; 63 checks all terminal (62 success,1 intended skip), 0 statuses; three required checks green; MERGEABLE/CLEAN; no auto-merge; empty queue before entry. All prior NONPASSs remain attributed. General owns watch-only custody, with the one exact dequeue at the ceiling only if this same entry is still live. No re-entry, run cancellation, tag or publication. Runner #1101 and optimization #1099 excluded. Cloud full health by 13:58 remains the next timed obligation. |
|
General — #1104 ordinary queue entry is terminal NONPASS. GitHub removed it at 14:07:43 UTC ( Merge group 43edf47 has ordered parents [2a15179, b14296f] and the accepted tree 583968b7. Shared run 42486430061 failed only The failed job is 128750154928. At 13:52:38 the published PHP reference recorder threw Artifact 13619991995 binds ZIP sha256 e967386a852b73fba542f83b0c6e7574416cd064be4d1c005fbea016b92eb1f6. All existing source-train results and earlier NONPASSs remain attributed. The CLI producer and consumers, six native-test backends, and the remaining corpus jobs passed in this merge-group invocation. Next: retain bounded status/refusal/task-shape diagnostics at this existing poll assertion, with focused valid-claim and refusal controls; then review the exact source successor before a fresh normal train. No assertion, fixture, retry limit, reference image or Native runtime is weakened. #1101 and #1099 remain deferred; the 3.3.0 cut/tag/image remain bound to M. |
The frozen PHP reference can finish the large-command-batches poll without a task. Preserve that assertion and include the final HTTP status, refusal shape and retries attributable to this poll, without headers or payloads. The previous merge-group failure did not retain those facts, so its cause remains unproved. No fixture, retry budget or runtime behavior changes. Co-Authored-By: Codex <noreply@openai.com>
|
General selects one fresh normal source train on #1104 for head 7697b46 / tree e18bc4ad963f09877d1817bbe2881a53153b8ed4, sole parent b14296f. The successor is exactly two conformance-diagnostic files: 6674 B / sha256 a9862c805d87ca1fbfcd1ea8a33792a895750713beabe2aa2ee4be6b6bccedcd. Full delta from M: 34443 B / 2f37b133fcf02fb1dd1fb6086c7e5d1ee9adc8438eeaf6158b52cbaa91b1ff25. The ordinary queue entry ended NONPASS at 14:07:43 (record 6722556394). The frozen PHP recorder failed before Native comparison; its final poll response was not retained, so the cause stays unproved. This successor preserves the existing is-array claim assertion and captures only bounded HTTP/refusal/shape facts plus retries for that specific poll. It changes no fixture, retry budget, reference image or Native behavior. Local PHP 8.4.27 checks: both files lint; two valid claims return unchanged; five refusal shapes retain the required facts without payload/header values, with messages below 1 KiB. One cached network-none, read-only, UID-1000 tooling container exited 0 with empty stderr; exact ID/name/run-label absence is proven. Actual PHPUnit execution remains hosted. Before the push: require main and merge-base M, clean exact head/tree/delta, remote and PR old head b14296f, and all prior source runs terminal. Mark the PR draft and push once with the exact old-head lease. Keep all prior NONPASSs and the queue failure. No rerun/cancel, ready/queue, release, runner change or #1099 effect is selected. Return per-check terminal evidence; the new diagnostic does not itself qualify a previously failed assertion. |
rmcdaniel
changed the title
|
General — one successor push completed at 14:17:47 UTC under selection 6722598138 (raw-body sha256 cffb470805b683326b4ad60fc004d95da6812a27271678ad6f599af87b492ca4). Exact old-head lease b14296f → 7697b46; plain fast-forward. Remote and PR now agree on head 7697b46 / tree e18bc4ad, draft and out of the queue. Main remains M. The first immediate PR readback lagged behind the pushed branch; a read-only follow-up confirmed the expected head and draft state. No second push occurred. Successor delta: 6,674 B / a9862c805d87ca1fbfcd1ea8a33792a895750713beabe2aa2ee4be6b6bccedcd. Full delta from M: 34,443 B / 2f37b133fcf02fb1dd1fb6086c7e5d1ee9adc8438eeaf6158b52cbaa91b1ff25. Two conformance-diagnostic files only. All prior NONPASSs remain unchanged. Normal runs: Shared fixtures 42489107309; PHPUnit Feature 42489107293; SonarQube 42489107187; CodeQL 42489104065; Boundary 42489107167 plus push 42489102742. No queue, cancellation, workflow rerun, release or runner-ref effect. General will classify terminal results before integration. |
The feature job selects its Unit tests explicitly. Include the new diagnostics class so its four tests execute in normal PR, merge-group and main verification. The previous green run did not execute that class. Co-Authored-By: Codex <noreply@openai.com>
The inventory correctly detects a changed full shell line, although its only semantic change is one extra PHPUnit class after Composer finishes. Re-review that line with the same sink count, guard, distance and trust confinement. Do not change the classifier or any execution permission. Co-Authored-By: Codex <noreply@openai.com>
|
General selects one wiring successor push on the existing draft #1104. Head d8d40e5, tree fa4163bca3c244450a55ad2ff44b45f10fba307d; two commits on old remote 7697b46: dd2affe and d8d40e5. The successor is 9212 B / 664c9e833eebd5609354d71658b0663d181ad7504b8803f363aa6322edcce94b; full delta from M is 43655 B / c7bcfd981608757c01715c5ad2c30127a08e72cef0b69b27c16d1e709097c505. The 7697b46 source train completed with 62 success, one intended skip, zero statuses. Its results remain valid. It did NOT execute the new four-method diagnostic test class because the PHPUnit Feature command selects Unit classes explicitly. General missed that wiring when adding the class; no queue action was taken. This successor adds that class to the existing PHPUnit invocation. The Composer inventory then correctly refused the changed full shell line locally. Re-review confirms the only command change is a PHPUnit filename after Composer finishes: the same two Composer occurrences, source-free checked tool image, server-verify guard and distance 86 remain exact. The one manifest line is updated; the classifier, trust boundary, permissions, sink count, CA, image, installs and retry policy do not change. No waiver. Local hard-isolated Inventory + RunComposer controls: 19 run, 17 passed, two real-Docker checks skipped; the initial unreviewed-line refusal is preserved. Existing PHP functional/lint checks and all preceding source/queue attributions stay recorded. Recheck clean exact head/tree/delta; main and merge-base M; PR/remote old head; all prior runs terminal; no queue or auto-merge. Push once with the exact old-head lease and let normal CI execute the four new tests. No rerun/cancel, ready/queue, release, runner-ref or #1099 effect is selected. |
|
|
General — one wiring successor push completed at 14:47:52 UTC under selection 6722918514 (raw body 7997e3932865c0e32074e910bfb6774769f6dbed2453cf925ce5d387cbc221c9). Exact old-head lease 7697b46 → d8d40e5, plain fast-forward. Remote and PR agree on d8d40e5 / tree fa4163bc; draft and out of the queue. Main remains M. Exactly two files since 7697b46: the explicit PHPUnit class list and its reviewed Composer inventory line. Successor 9,212 B / 664c9e833eebd5609354d71658b0663d181ad7504b8803f363aa6322edcce94b; full 43,655 B / c7bcfd981608757c01715c5ad2c30127a08e72cef0b69b27c16d1e709097c505. Composer confinement count/guard/distance and every runtime assertion are unchanged. Normal runs: Shared fixtures 42491566759; PHPUnit Feature 42491566827; SonarQube 42491566807; CodeQL 42491562962; Boundary 42491566898 and push 42491563208. The four new PHPUnit methods must be verified on this head before integration. The earlier 7697b46 train stays green with its disclosed new-test execution gap; prior NONPASSs and the failed queue invocation remain attributed. No cancellation, rerun, queue, release or runner update occurred. |
|
General accepts the exact #1104 source head d8d40e5, tree fa4163bca3c244450a55ad2ff44b45f10fba307d, on base/merge-base M 2a15179. Full delta 43,655 B / c7bcfd981608757c01715c5ad2c30127a08e72cef0b69b27c16d1e709097c505. Worktree clean; no new source change is selected. Exact-head terminal result: 62 success, one intended path-gated skip, zero statuses. All three required checks are green. Shared fixtures 42491566759 passed; PHPUnit Feature 42491566827 passed, including the explicitly selected ServerParityLargeCommandBatchesProbeTest.php in job 128764935529 (main invocation: 3,633 tests / 90,112 assertions, 51 skips). Its four new methods have no skip or group condition. Sonar and CodeQL passed. The earlier 7697b46 train stays green with its disclosed missing-test execution gap. Source NONPASSs and the 14:07 failed merge group remain attributed. Select one fresh ordinary queue entry after re-reading this body and rechecking main/base/merge-base M; exact head/tree/delta; open/draft/mergeable/clean; no auto-merge; these terminal checks/statuses; and empty queue. Mark ready once, then enqueue once with expectedHeadOid=d8d40e5e. No forced/bypass merge, cancellation or workflow rerun. Set a 45-minute ceiling from the actual enqueue timestamp; if this exact entry remains live then, perform one exact dequeue and stop that entry. Do not silently re-enter. After merge require valid signature, ordered parents [M,d8d40e5e], tree equality to fa4163bc, actual-main terminal check/status custody and ancestry. Delete only the owned ci/parity-cli-once topic, with exact-head/ancestry/main-terminal guards; preserve all WIP refs. Return custody. The 3.3.0 cut/tag/image stay bound to M; frozen runner #1101 and optimization #1099 are excluded. No tag, publication, acquisition, qualification dispatch or alias effect is selected here. |
|
General — fresh ordinary queue entry created once under selection 6723285934 (2cbb48d8fbcde15e49c9043568e0c42bd9a723896193991cb8dc184c0286a075). Final guards matched: main/base/merge-base M, head d8d40e5, tree fa4163bc, full delta c7bcfd98, clean worktree, 62 success + intended skip, zero statuses, no auto-merge and empty queue. Marked ready once and enqueued with expectedHeadOid=d8d40e5e. Entry MQE_lQDOSATFBs8AAAABNpAH084ABPZYzgPQ_Is, enqueuedAt 2026-11-08T15:22:59Z, initial state QUEUED, position 1. Ceiling 2026-11-08T16:07:59+00:00. Watch only; one exact dequeue if this same entry remains live then. No silent re-entry. The prior b14296f merge-group failure remains recorded at 6722556394. The d8d40e5 source train is terminal green, including the now-wired four diagnostic tests. Main custody is still owed; no release, runner refresh/dispatch, image or #1099 effect was performed by this queue action. |
|
General: #1104 second queue entry merged at 15:46:30Z, inside its 16:07:59Z ceiling, with no dequeue/re-entry/rerun/cancel. Actual main C = f77b093. GitHub signature is valid; ordered parents are [M 2a15179, accepted head d8d40e5]; tree fa4163bca3c244450a55ad2ff44b45f10fba307d equals the accepted head. The actual M→C binary diff is byte-equal to the accepted delta: 43,655 B / c7bcfd981608757c01715c5ad2c30127a08e72cef0b69b27c16d1e709097c505. Main readback is C. The merge-group runs passed. Actual-main push checks are running, so custody is not yet terminal. The owned topic stays until main is terminal and successful, then exact-head/ancestry cleanup. All WIP refs remain untouched. Earlier NONPASSs and the initially missing PHPUnit wiring remain attributed. Next: terminal actual-main custody, then review/select the unchanged #1101 runner merge on top of C. Native 3.3.0 tag/image remain bound to M; #1099 stays deferred. No tag, publication, runner dispatch or alias effect is selected by this merge receipt. |
|
General returns terminal #1104 actual-main custody at 16:01:45Z. C = f77b093 remains current main. Valid signature; ordered parents [M 2a15179, d8d40e5]; tree fa4163bc equals the accepted head. The 43,655 B / c7bcfd98 delta is byte-equal to the accepted full diff. All 79 check runs are terminal: 71 success, 8 expected skips, 0 failures/cancelled/timed-out; 0 statuses. After main-success, exact-head and ancestry guards, the owned ci/parity-cli-once topic was deleted once with the d8d40e5 lease. Git readback is empty and API is 404. No WIP/preparation ref was touched. Delivered: bounded pinned parity-CLI acquisition shared to consumers; Composer 2.10 transient transport matching; deterministic #1105 timer hold/control; bounded missing-poll diagnostics and their explicit PHPUnit wiring. API fallback availability remains unproven; no broader outage or zero-flake claim. All earlier NONPASSs and the initially unexecuted test-class gap stay attributed. Next: unchanged #1101 runner merge on top of C, as a new ordinary queue entry. Native 3.3.0 release M and image D stay unchanged; #1099 deferred. No dispatch, tag, alias or publication by this custody receipt. Both repo agents remain Off. |
The parity CI train repeatedly lost GitHub release-asset and Composer downloads, and a timer test could race its scheduler. Fetch the pinned CLI once per run with bounded HTTPS retries and a digest-checked official API fallback; share its artifact with every corpus consumer. Match Composer 2.10 transient errors within the existing locked-install retry policy, and hold the specific debug test timer until its deliberately due sweep has been observed.
A subsequent merge-group invocation failed in the frozen PHP reference recorder before Native comparison: the large-command-batches poll returned no task, but its final response was not retained. Keep that assertion and capture bounded HTTP/refusal facts and poll-specific retries on failure, without payloads or headers. This diagnostic is not a claimed fix or acceptance for the earlier reference failure.
Validation and limits:
Fixes #1103 and #1105 after successful integration. General owns the remaining work while the Server agent is Off.