Repository navigation
Conversation
chore: factory block for softfac
Red: ./index.ts does not exist yet, so the import fails. Issue: #1
Green: factory started is logged once. Issue: #1
Red: factory/scripts/check-seams.sh does not exist yet, so both cases fail. Issue: #1
Green: both check-seams cases pass. Issue: #1
No red phase: the tree already complies, so the real-tree test passes on arrival. The synthetic cases show each rule catching a violation. Issue: #1
Typecheck could not resolve the bare vitest module; the repo imports it via vite-plus/test. Issue: #1
…order-free Red: "names an upstream file renamed into an allowed path" fails because check-seams.sh lets git rename detection hide the deleted upstream path. The fail case now changes server.ts itself, and the boundary real-tree test asserts server.ts imports ./factory/index.ts; both already pass. Issue: #1
A git mv of an upstream file into an allowed path listed only the new path, so the deleted upstream file passed the seam check. --no-renames lists both sides. Issue: #1
- boundary.test.ts flags factory code outside factory/t3/ that imports the server's workspace packages (@t3tools/*, effect-acp, ...), read from apps/server/package.json. - check-seams.sh captures the diff before the loop, so a failed git diff fails the script instead of passing with nothing checked. - README and the script header say the check covers committed HEAD only. - check-seams tests cover apps/web/src/factory/ and a failing diff. - Node built-ins imported as NodeFS/NodePath/... per namespace-node-imports.
Factory skeleton: an empty FactoryLayer wired in through one seam
… first failure Issue: #2
…ushes to main Issue: #2
… plainly Review nits: the cd to the repo root was never exercised, and the expected calls rebuilt the constants they came from. No red phase: both behaviors already held. Issue: #2
Red on purpose: install.sh, rollback.sh and build.sh do not exist yet. Issue: #3
…64 binary Issue: #3
… tail Red: apps/server/src/factory/process.ts does not exist yet. Issue: #8
…utput tail Issue: #8
…ests Ports the run id, next_attempt, format_ts and parse_ts tests from softfac's ledger.rs, plus the issue key parse and refuse cases from issue_key.rs. Red: runId.test.ts and time.test.ts fail because runId.ts, time.ts and errors.ts (LedgerError, used by testSupport.ts) do not exist yet. Issue: #9
Ports softfac's RunId, IssueKey, next_attempt, format_ts and parse_ts to plain TypeScript, and LedgerError with the Rust Display messages. - Run ids are lowercase owner/name plus a canonical issue (up to u64 max) and attempt (up to u32 max); file names that are not exactly what this code writes are not run ids. - Lowercasing is ASCII-only so a Kelvin sign cannot fold into a legal name. - validateRunId also checks the issue and attempt, since they end up in the file name too. Issue: #9
Ports the run ledger tests from softfac's ledger.rs: one exact line for every one of the 30 event types, create/open semantics (create-new, newline repair of a torn last line, symlinks and directories refused), the read rules (torn last line, blank lines, invalid lines counted, unknown events and odd data kept as written, causes kept as written), finish (total_ms and gate_attempts computed by the writer), the failed-write rules through an injected AppendIo, file and directory modes, and the directory-sync failure that removes the new file. The feedback read file and the list/summary readers are not ported. Red: ledger.test.ts fails because events.ts, appendFile.ts, read.ts and ledger.ts do not exist yet. Issue: #9
Ports softfac's find_project and status cases from src/t3.rs, adds launch and thread cases, and a contract test over the real orchestration layers with the Claude replay transcript. Red: paths.ts and sessions.ts do not exist yet, so none of the three files compile. Issue: #5
Red: the wrapper sh dies on SIGTERM and the group is SIGKILLed at once, so the grandchild's marker never reaches the tail. Issue: #8
A run whose last complete line is a finished event is skipped after one bounded read from the end of its file. A cut-off or invalid last line, or a finished line longer than the tail, falls back to the full read. Issue: #49
…s, isolates failures
Red: the tests call unassignedClaims with a known set and a ClaimMemo and expect a
{ claims, unread } result, none of which exist yet (compile errors count as red).
Issue: #48
…eady explained Fails before the fix: the labeledAt tracker method does not exist yet, and releaseClaim skips the comment on an old cut-short comment and never reads the label history. Issue: #50
… issues The scan took one serial GraphQL history call per unassigned triage/building issue on every pass, including issues the pass then dropped for having a run, and one failed read failed the scan for the whole repo. It now takes the issue numbers the pass has runs for and reads nothing for them, remembers issues judged not the factory's while their updatedAt is unchanged (for the life of the process), and reports a failed read for that issue only, which the pass prints as one error line. Issue: #48
…ady explained Issue: #50
A run file holding only blank lines, or a large file whose last 8192 bytes are all blank, must be listed as unfinished exactly as readRun reads it (state null, or the state of its first line). Red: before the fix the run hung node (the vitest worker printed no result and had to be killed after the 60s alarm did not stop it). Issue: #49
The blank-line step-back called lastIndexOf with a byteOffset of -1 once it reached the first byte, which Buffer reads as counting from the end, so the walk looped forever. With no earlier line it now returns false and the file is read in full. Issue: #49
…short scan The two new tests fail: such a failure is still swallowed into unread and the scan goes on to the next issue. Issue: #48
…scan A rate limit, a timeout or a gh spawn failure would fail every remaining history read, so it now ends the scan like a failing issue list does. Only GitHubFailed, GitHubParse and GitHubOutputTooLarge are kept per issue. Issue: #48
…c failures Four cases a mutation review showed were missing. They pass against the current code (test-only cycle): - index: a judgement that an issue is not the factory's lasts across passes of one world (one history read over two scans) - index: a failed history read reaches the pass as an unread entry with the issue and a non-empty message - github: an unparseable history is that issue's GitHubParse error and the next issue is still read - pass: a run in another repo is left out of the known set handed to the scan Issue: #48
…shed are read in full Both tests pin existing behavior, so there was no red phase. They kill two mutation survivors: a last line naming finished that is not a valid entry, and a cut-off fragment after a complete finished line, must each force a full read. Issue: #49
A pass skips finished run files without parsing them
…se reads Review round 1: f1, f2. - labeledAt and unassignedClaims now pick the claim's label event through one newestLabelEvent helper. - alreadySaid and releaseClaim share newestSays for 'the newest factory comment is this text, in GitHub's time form'. - A release from building is tested against a history with both triage and building label events, so the released state must reach the label lookup; labeledAt is tested for building as well.
…ssue and label The memo now keys by login, repo, issue and state label and stores the updatedAt it judged at, so a re-authenticated gh judges again and a newer judgement replaces the old one instead of piling up. Tests pin each part of the key, a re-read after a failed read, an oversized history as that issue's error, and a gh that cannot start ending the scan.
…on a set read A run whose session reported its pull request is no longer limited until it next enters building: the wait for the gate is not the build's time. A triage run whose launch was sent with no session recorded is left to resume, which names its session, so the limit never finishes a run whose session may be running untracked. The events wake checks an in-memory set of owned threads, read at start and after every pass, instead of reading every run file per status change. The time-limit comment no longer promises that the next attempt continues from the old branch. sanitize is imported from text.ts everywhere, and advance's reminder send and ended-badly statuses each have one home. Tests: an unconfirmed run is never time-limited, watched or advanced; a hostile branch name reaches the comment and ledger only as prepared text; a run back in building from pr-open is limited again.
Until the gate is wired into the pass, a building run whose session reported its pull request stays in building. It has no session at work, so like a pr-open run it no longer takes one of max_concurrent_runs, and new claims are not starved by finished builds. The listing carries prReported from the ledger, cleared when the run enters building again.
…ad again Once the session has reported its pull request, the build is done and the run waits for the gate (#19). The pass no longer watches its text or calls advance for it, so each pass makes no tracker or T3 call for a finished build and only says it is ready for the gate.
Building: follow the session to its result line
Cut-short scan stops re-reading label history for issues it never acts on
# Conflicts: # apps/server/src/factory/index.test.ts # apps/server/src/factory/index.ts # apps/server/src/factory/tracker/github.ts
A cut-short release comments unless this claim was already explained
Fails until the scripts check health: the unhealthy install, closed port, old log line, no-previous, and both rollback-check cases. The healthy and truncated-log cases already pass. Issue: #59
After the restart, install.sh and rollback.sh wait up to T3FAC_HEALTH_WAIT seconds (90) for the port and a fresh factory started or no factory config log line. A failed install runs rollback.sh; a failed rollback exits 1 naming both versions. Issue: #59
…il is 50 lines Review findings: nothing exercised the "no factory config" line, and the tail assertion could not tell 50 lines from the whole log. Both already pass. Issue: #59
github-actions
Bot
added
vouch:unvouched
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.
Closes #59
Contract
restart_job,install.shwaits up toT3FAC_HEALTH_WAITs (90) for a TCP connect on the port and afactory startedorno factory configline appended to the log after the restart (size noted before; a smaller file is read from offset 0). Either missing: last 50 log lines,rollback.sh, exit 1 withinstall of <version> failed its health check; rolled back to <previous>; no previous: exit 1 and says so.rollback.shruns the same check on the restored version and exits 1 naming both versions. README gains a Recovery section.vp test run apps/server/src/factory/installScripts.test.ts→ pass (31 tests)Acceptance criteria
install.shwaits up toT3FAC_HEALTH_WAIT(default 90) for the port and a fresh log line —wait_healthyinfactory/scripts/lib/t3fac.sh; testskeeps a healthy install,rolls back an install whose port does not accept connections,rolls back an install whose server never logs that it started,keeps an install whose server logs that it has no factory configrollback.sh, exit 1 with the message —rolls back an install whose server never logs that it started(asserts line 11 shown, line 10 cut)exits 1 and says so when an unhealthy first install has nothing to roll back torollback.shchecks the restored version and names both versions —rollback.sh exits 1 naming both versions when the restored version is unhealthy too,install.sh reports when the rollback fails its health check as wellfactory startedline does not count; truncated log reads from offset 0 —does not count a factory started line that was there before the restart,counts a line written after the log was truncated to a size below the noted onelaunchctl, a stub listener and a written log —apps/server/src/factory/installScripts.test.ts, describeinstall.sh and rollback.sh health checkfactory/README.mdsection## RecoveryAssumption: the issue says tests in
factory/scripts/; the existing script tests live inapps/server/src/factory/installScripts.test.ts, which is what the gate runs, so the new cases are there.T3FAC_PORTis a test-only override for the health check port; the plist still serves 3876.Gate
factory/scripts/check.sh→ exit 0 (seams, 33 test files / 1138 tests, server and web typecheck)Models
Review
no factory configline)no factory configline; the log-tail assertion now proves the 50-line bound (both mutants re-run and killed by the orchestrator)Lessons
HEAD, not by the name in the launch snapshot. Symptom:git push -u origin <branch>fails with "src refspec does not match any". Cause: the git status snapshot at session start named a different branch (t3/8c8fa6d2) from the worktree's real one (t3/factory-issue-59). Default: readgit branch --show-currentfirst, or pushHEAD, and never pass the snapshot's name.cdinto the skill directory. Symptom: aftercd <skill>/references && cat ...the harness reports the primary working directory moved off the worktree. Cause: the session's cwd follows the shell. Default:cat /abs/path/filefrom the worktree; every command keeps starting withcd <worktree> &&.Commits
520aaab0btest: install and rollback wait for the server to come up— Make slice (red: 6 failed)b33d8dfcdfeat(factory): roll back an install that fails its health check— Make slice (green)626a4971bdocs(factory): recovery section— Make slice (docs, no red phase)51fe118b3test: a server that logs no factory config is healthy, and the log tail is 50 lines— Check cycle 1 (tests for behavior that already worked, no red phase)Files
log,port(T3FAC_PORT),log_beforenoted inrestart_job, newwait_healthyLAUNCHCTL_HEALTHY/LAUNCHCTL_HEALTHY_LINE/LAUNCHCTL_TRUNCATE_LOGin the launchctl stub, nine health-check casesIsolate: the factory's worktree was behind
origin/main(ancestor, clean tree); fast-forwarded withgit merge --ff-only origin/mainbefore provisioning. Repo labels: all seven existed; none created.