Repository navigation
Tooling batch: nine retro tickets - #433
Merged
Merged
Conversation
The test file defined psql_admin five times, each answering some of the queries and nothing for the rest. One fake now answers from the fixture directory each block fills ($fx/databases, owned, db-state), and a query it doesn't know, or whose fixture file is missing, ends the run with a FAIL line naming it. Co-Authored-By: Claude Fable 5.2 <noreply@anthropic.com>
After each fetch a clean master behind origin/master is fast-forwarded, and the run starts again from scripts/ship-branch.sh when that changed it; a checkout that can't be fast-forwarded and has a different script is refused with the command that updates it. Several branches ship in order, stopping at the first failure and naming those not reached. CLAUDE.md drops the clause that restated the 72-character refusal. Co-Authored-By: Claude Fable 5.2 <noreply@anthropic.com>
…admin fake (#426) Co-Authored-By: Claude <noreply@anthropic.com>
…ip several branches (#431) Co-Authored-By: Claude <noreply@anthropic.com>
One LEFTOVER line with the count of what dev-branch-servers.sh orphans lists, and the command. Quiet when there is nothing or Postgres is down; a stopped Kafka or Redis is not asked. Co-Authored-By: Claude Fable 5.2 <noreply@anthropic.com>
Half of the file's 27 s was the stub backend's shutdown(): it returns at serve_forever's next poll, 0.5 s by default, once per test. The stub now polls every 10 ms. Another 5 s was JUDGE_RUN_SPACING=0.3 in tests that assert nothing about spacing; only the spacing test keeps it, the rest wait 0. What is left (about 9 s) is one Python child per Run. Still 30 tests. Co-Authored-By: Claude Fable 5.2 <noreply@anthropic.com>
…anch databases left behind (#424) Co-Authored-By: Claude <noreply@anthropic.com>
The pointer keeps the two rules whose breach harms other sessions and names the cases that send an agent to docs/agents/parallel-sessions.md. The new file adds a part on the shared compose containers. The Agent Brief convention now says where a brief sends a doc note. Co-Authored-By: Claude Fable 5.2 <noreply@anthropic.com>
…hey take (#425) Co-Authored-By: Claude <noreply@anthropic.com>
…moves behind a pointer (#405) Co-Authored-By: Claude <noreply@anthropic.com>
scripts/python312.sh prints a Python 3.12 or newer (CODEBITE_PYTHON312, python3.12, python3, each asked its version) or exits 2. judge-local.py, replay-local.sh, replay-judge0-python.test.py and check-tracer-helpers.py drop their own lookups; replay-local.sh no longer hands an unasked interpreter or a /opt/homebrew guess to the replay. Co-Authored-By: Claude Fable 5.2 <noreply@anthropic.com>
Every deploy rolled the backend and recreated nginx and the worker, whatever the merge changed. Under Docker 29's containerd image store an image's id is the digest of an index holding the image manifest and BuildKit's provenance attestation, and the attestation is new on every build, fully cached or not. `provenance: false` in the compose file was meant to stop it, but compose 5.1.0 leaves a false out of the bake definition it hands buildx (only true or a mode gets through), so the default attestation was still attached. deploy-prod.sh now exports BUILDX_NO_DEFAULT_ATTESTATIONS=1 for every build it starts, and the three dead `provenance: false` keys go. Loop: two `docker compose build` runs of the same tree under Docker 29.3.0 and compose 5.1.0, compared by image id; and build, up, build, up, watching for a Recreate. Co-Authored-By: Claude Fable 5.2 <noreply@anthropic.com>
…, so unchanged services stay (#403) Co-Authored-By: Claude <noreply@anthropic.com>
#398) Co-Authored-By: Claude <noreply@anthropic.com>
The crash from #392, reproduced: open() wrote a pass's language and theme into localStorage on the home page, where the last pass's signed-in app was still running. When that app's GET /auth/me answered after the write (a backend under load) and before the problem page replaced it (a Vite under load), it wrote the user's language (ko) back. The signed-out English pass then waited for "Description" on a Korean page, and atTab read a tab that wasn't there: TypeError, stack trace. open() now sets the visitor up in a script Chrome runs in the new document before the app's, and goes straight to the problem page. What a pass needs that doesn't show within WALKTHROUGH_WAIT_MS fails one check naming the pass and what it waited for, with a shot of that moment; the pass stops and the next starts. Any other error is a failed check too, never a stack trace. The 15 s bound stays: under a Gradle test run the slowest first load was 4.2 s. Co-Authored-By: Claude Fable 5.2 <noreply@anthropic.com>
…runs out fails a named check (#432) Co-Authored-By: Claude <noreply@anthropic.com>
n-queens and list-bracket-pairings are drawn in full until the idea is shown, then one frame per further result; restore-wifi-ip draws a pruned branch as one frame; merge-cafe-queues draws a heap push or pop with its sift as one frame. Each folding frame's caption says what was skipped and how much. Co-Authored-By: Claude Fable 5.2 <noreply@anthropic.com>
…ong runs (#388) Co-Authored-By: Claude <noreply@anthropic.com>
judge-local.py and check-tracer-helpers.py each stripped 'python312: ' from the finder's stderr; the line now comes bare, for a caller to put after its own name. Co-Authored-By: Claude Fable 5.2 <noreply@anthropic.com>
dev-status counted the indented lines of orphans' human output. dev-branch-servers.sh orphans --count prints the number of databases alone; dev-status stays quiet when it prints none. Co-Authored-By: Claude Fable 5.2 <noreply@anthropic.com>
The blank page and the images of the Diagram-beside shot were waited for with the result dropped; they now stop the section like any other wait. shown takes the passed check's line, so no check(true) follows a wait that throws; the lines printed are the same. Co-Authored-By: Claude Fable 5.2 <noreply@anthropic.com>
restore-wifi-ip's one-call dead ends and merge-cafe-queues' pops and pushes say how much was folded (calls and lengths tried, swaps). restore-wifi-ip takes a dead end's counts from place itself, run with its frames held back, in place of a second copy of the search. The heap-move helper three tracers each carried is codebite.heap_moves; guild-chat-relay and rising-water-path regenerate unchanged. The skill names scripts/python312.sh for finding a Python 3.12 (#398). Co-Authored-By: Claude Fable 5.2 <noreply@anthropic.com>
For what it refuses and how it keeps itself current. Co-Authored-By: Claude Fable 5.2 <noreply@anthropic.com>
…per, named waits Co-Authored-By: Claude <noreply@anthropic.com>
2 tasks
Kyu-seok
marked this pull request as ready for review
November 20, 2026 15:57
This was referenced Nov 22, 2026
Kyu-seok
added a commit
that referenced
this pull request
Nov 22, 2026
* SessionStart hook: fast-forward a stale main checkout A session reads CLAUDE.md, skills and scripts from the main checkout, which nothing kept current: it sat seven merges behind through the #433 build. The hook fast-forwards a clean master and says what it moved, or says how far behind the checkout is and the command that catches it up. Co-Authored-By: Claude Fable 5.2 <noreply@anthropic.com> * Tracer helper check: compare nested functions too A third copy of a nested heap helper reached review in #433: the check compared only top-level functions. The twelve nested copies already there are listed in nested-copies.known, which passes exactly those and fails on a new copy or a stale line; #435 moves them into codebite.py. Co-Authored-By: Claude Fable 5.2 <noreply@anthropic.com> --------- Co-authored-by: Claude Fable 5.2 <noreply@anthropic.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
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.
Nine tickets with no spec, landed on one integration branch.
State directory:
~/.cache/implement-spec/tickets-2026-11-20/(brief:BRIEF.md, merge log:merge-log.md). Integration tip:cea22201.Merged
ticket/426-psql-admin-fake):scripts/dev-branch-servers.test.shhas onepsql_adminfake backed by fixture files (5 definitions to 1, 143 checks to 147); an unknown query ends the run with aFAIL:line naming it.ticket/431-ship-branch):scripts/ship-branch.shfast-forwards a clean main checkout onmasterand restarts from the updated file before any push, refuses when it can't and its own copy differs fromorigin/master's, fast-forwards again after a merge, and takes several branches (--title/--body-filerefused with more than one); the ship-branch: PR title as the squash subject (#400) #428 clause is out of CLAUDE.md (23,016 bytes).ticket/424-dev-status-orphans):scripts/dev-status.shprints oneLEFTOVERline (count, and theorphanscommand) when branch databases or a dropped one's Kafka topics or Redis index are left behind; quiet otherwise, and no Kafka or Redis call when that container is stopped. Exit code unchanged; dev-status takes about 8 s longer with all three containers up.ticket/425-judge-run-test-time):scripts/judge-run.test.pyruns in 9 to 12 s instead of 26 to 35 (the stub backend's shutdown polled every 0.5 s per test; spacing sleeps in tests that assert nothing about spacing), still 30 tests;judge-run.pyuntouched. The script tests as a whole still take about 90 s, spread over six files (timings in the issue comment).ticket/405-parallel-sessions-doc): "Parallel sessions" moves todocs/agents/parallel-sessions.mdin headed parts, with a new part on the shared compose containers; CLAUDE.md keeps a pointer with its trigger cases and the two rules (21,052 bytes);docs/agents/issue-tracker.mdsays where a brief sends a doc note;guard-pr-merge.shcites the new file.ticket/403-deploy-roll-skip):scripts/deploy-prod.shexportsBUILDX_NO_DEFAULT_ATTESTATIONS=1, so a cached build keeps its image id and unchanged backend, nginx and worker containers stay; the three deadprovenance: falsekeys leaveinfra/docker-compose.prod.yml. Cause: compose 5.1.0 dropsprovenance: false, and each build's new attestation changed the image id. Owed: The deploy carrying this still rolls once (ids change form). The next merge that touches nothing the backend image is built from should log--- Backend image and config unchanged: keeping the running replicas ---andContainer codebite-nginx Running; post that run id on Deploy: a no-backend merge still rolls the backend, and nginx recreate returns 502 #403.ticket/398-python312-finder): Newscripts/python312.shprints a Python 3.12 or newer (CODEBITE_PYTHON312,python3.12,python3, each asked its version) or exits 2;judge-local.py,replay-local.sh,replay-judge0-python.test.pyandcheck-tracer-helpers.pyuse it and drop their own lookups; one line in CLAUDE.md.replay-local.shno longer falls back to the Homebrew path unasked.ticket/432-app-check-waits):scripts/walkthrough-app-check.mjs: the crash was a race inopen()(a signed-in pass's late/auth/mewrote the Korean locale back before the English pass), removed by setting localStorage in the new document; a wait that runs out is nowFAIL <pass>: <what> did not show within 15 swith<pass>-stopped.png, and any other error in a pass is aFAILline, never a stack trace.WALKTHROUGH_WAIT_MScuts a wait short; one sentence in the add-walkthrough skill.ticket/388-condense-walkthroughs): Four Walkthroughs regenerated from condensed tracers (longest run:n-queens118 to 33,list-bracket-pairings92 to 30,merge-cafe-queues38 to 27,restore-wifi-ip50 to 44); folding captions say what was skipped in both languages; a note on condensing in the add-walkthrough skill, and two Heap lines in itsviews.md. Owed: A decision onrestore-wifi-ip: folding pruned branches only takes it from 50 to 44.ticket/review-fixes): Code-review fixes: folding captions inrestore-wifi-ipandmerge-cafe-queuesgive a count, andrestore-wifi-iptakes its counts from the instrumented run (thepruned()replica is gone); oneheap_movesincodebite.pyfor three tracers; the two side-by-side waits inwalkthrough-app-check.mjsfail a named check;python312.sh's message has no prefix to strip;dev-status.shcounts withdev-branch-servers.sh orphans --count;parallel-sessions.mdpoints toship-branch.sh's header.#432
Cause.
open()wrote each pass's language and theme into localStorage on the home page, where the previous pass's signed-in app was still running. When that app'sGET /auth/meanswered after the write (slow backend) and before the problem page replaced it (slow Vite), it wrote the user's locale (ko) back. The signed-out English pass then waited for "Description" on a Korean page until its waits ran out, andatTabread a tab that wasn't there. CPU throttling alone (20x, 40x through CDP) does not reproduce it.Reproduction on master (branch servers up; holds
/api/auth/me2 s and the problem page 5 s in Chrome):On master:
exit 1, about 58 s. On this branch the same command exits 0, because the fix removes the race:
open()now sets localStorage in the new document before the app runs.A wait cut short, on this branch:
WALKTHROUGH_WAIT_MS=20 node scripts/walkthrough-app-check.mjs <vite> <api> <out> <slug>printsFAIL out-390-light-en: tabs did not show within 0.02 swithout-390-light-en-stopped.pngfor each pass, exit 1, no stack trace. The named line was not captured under the reproduction itself, which no longer fails.Under load: three runs in a row during
scripts/gradle-test.sh :backend:test --tests 'com.codebite.problem.*'all exit 0 (41 s, 27 s, 33 s); the slowest first load was 4.2 s, so the 15 s bound stays.#405
Every fact of the old "Parallel sessions" paragraph, against its place in
docs/agents/parallel-sessions.md(wording carried over unchanged):master; all branch work in your own worktreegit switchor commit in the shared checkout moves every other session's filesgit worktree add <scratchpad>/wt -b <branch> origin/masternpm ciin the worktree'sfrontend/before any frontend check: nonode_modules, the main checkout's may be staleorigin/master; so can a skill loaded at session startorigin/master, first bulletgit worktree add, rungit -C <worktree> diff $(git rev-parse HEAD) -- CLAUDE.md .claude/skillsfrom the main checkout and read what landed (a newer script, a skill step)origin/master, first bulletorigin/master, second bulletscripts/gradle-test.sh <gradle args>, which waits for another worktree's run; hooks and scripts use it too--tests 'com.codebite.<package>.*'); the full suite is the merge gate'sgrepis the claude binary's ugrep, killed (exit 137) when memory runs low with whatever pipes into it (migrate | grepleft a migration unapplied); write long output to a file, use/usr/bin/grepin pipelinesmaster; the commit-msg hook a subject over 72 charactersscripts/ship-branch.sh <branch>…from the main checkout: stops on unticked- [ ]unless--allow-unticked, waits for CI, squash-merges, deletes the branch (remote, local, worktree, database)gh pr merge --delete-branch, which deletes live worktreesscripts/merge-ticket.sh <branch> <title> [...]from the integration worktree: merges in order, runs the checks a clean merge skips, pushes only when green (header has details); an orchestrator runs it directly, a subagent only for a conflictdockeron PATH orDOCKER_HOST#403
#403: why every deploy rolled the backend and recreated nginx
Root cause. The server runs Docker 29.3.0 with the containerd image store and compose 5.1.0. There an image's id is the digest of an index ("manifest list") holding the image manifest and BuildKit's provenance attestation, and the attestation is new on every build, fully cached or not.
provenance: falsein the compose file (#347) was meant to stop it, but compose 5.1.0 leaves afalseout of the bake definition it hands buildx (docker compose build --printshows anattestentry fortrueormode=max, none forfalse), so buildx attached its default attestation anyway.The deploy logs say the same: in all 12 deploys read (44687133950 to 44734872864) every build step is
CACHEDand theexporting manifestandexporting configdigests repeat from run to run, whileexporting attestation manifestandexporting manifest listare new each time. All 12 rolled the backend and recreatedcodebite-nginxandcodebite-worker-1. The two guesses on record (a missed build cache, a jar that isn't reproducible) are both ruled out by those lines.Loop (an isolated
docker:29.3.0-dinddaemon, which ships compose 5.1.0: the server's versions; this machine's Docker 20.10 has neither attestations nor the containerd store, so it cannot show the bug):docker compose -f infra/docker-compose.prod.yml build backendtwice from the same tree, comparingdocker image inspect -f '{{.Id}}' codebite-backend.sha256:273d60a7…thensha256:ad1e6119…, red. The second build exported the same config and a new attestation manifest.BUILDX_NO_DEFAULT_ATTESTATIONS=1:sha256:07b464f2…both times, green.up -d, build,up -don a service shaped like nginx (built image, fixed container name).upprintsRecreate, red. After:Running, green.Fix.
scripts/deploy-prod.shexportsBUILDX_NO_DEFAULT_ATTESTATIONS=1for every build it starts (the rollback'sup --buildtoo); the three deadprovenance: falsekeys are removed.scripts/deploy-prod.test.shchecks a compose call from the script carries the variable.nginx. The recreate had the same cause: its image id changed on every build, so
up -dreplaced it (about 2.1 s without nginx in run 44687133950:Recreate06:22:16.07,Started06:22:18.15). With stable ids a deploy that doesn't changefrontend/leaves nginx and the worker running. A deploy that does change the frontend still recreates nginx and fails requests for about 2 s.Owed after merge.
backend/,common/,gradle/or the root Gradle files should log--- Backend image and config unchanged: keeping the running replicas ---, withContainer codebite-nginx Runninginstead ofRecreate. Post that run id on Deploy: a no-backend merge still rolls the backend, and nginx recreate returns 502 #403.Not tested. Whether the jar is byte-identical when Gradle really reruns (BuildKit layer cache pruned): every build seen was fully cached.
#388
Walkthroughs condensed (#388)
n-queenslist-bracket-pairingsrestore-wifi-ipmerge-cafe-queuesEvery run's last frame still carries the sample's expected output (the recorder refuses a run that doesn't, and
WalkthroughsMatchSamplesTestpasses unchanged). No other Walkthrough's JSON changed.restore-wifi-iponly drops from 50 to 44: the lengths a call passes over were already folded into captions, so its length is the nine addresses' calls and returns, not pruned branches. Shortening it further needs another decision (for example the first two addresses in full, then one frame per address, as the other two backtracking ones).Folding-frame screenshots (branch app, headless Chrome, dark)
n-queens, 1440 and 390 wide, English:Korean: 1440, 390
list-bracket-pairings, 1440 and 390 wide, English:Korean: 1440, 390
restore-wifi-ip, 1440 and 390 wide, English:Korean: 1440, 390
merge-cafe-queues, 1440 and 390 wide, English:Korean: 1440, 390
Every folding frame's caption, English and Korean
n-queens: folding frames of the longest run (n = 5)list-bracket-pairings: folding frames of the longest run (pairCount = 4)restore-wifi-ip: folding frames of the longest run (digits = "19216811")merge-cafe-queues: folding frames of the longest run (queues = [[2,5,8],[1,3,7],[4,6]])Closes #432
Closes #431
Closes #426
Closes #425
Closes #424
Closes #405
Closes #403
Closes #398
Closes #388
🤖 Generated with Claude Code