Fixes from the pinball test: dispatch, merges, CI, reviews, planned projects, judging (#232–#259) #261

Merged
cmoriarty merged 10 commits from claude/batch-232-260 into main 2026-10-27 17:56:39 -04:00
Owner

Four OpenSpec changes for braid#232 to braid#259, from the Braid vs Claude Code pinball test. Each change is archived in this branch.

Reliable dispatch, merges and CI evidence (#237, #245, #247, #249, #258)

  • An attempt opens only in a write that finds its step ready and its run live: no second gate for one step (run 117), nothing after a stop.
  • Open gates are the ones on their step's current attempt, everywhere they are counted.
  • A merge Forgejo defers is read again until it settles when main has not moved; only a base that moved is a conflict.
  • A failed CI job with no log has its steps run in the clone, so the fixer sees the error.
  • Script output reaches the transcript while a step is quiet ("waiting on CI" no longer arrives after the wait).
  • bash is the shell tool's second name, with the shell's guards and rules.

Auto-approve, reviews and gate words (#244, #256, #257)

  • A fast review lens with findings is a review: Sonnet's reviews are no longer held for a person.
  • A gate auto-approve leaves for a person records why (osf.gate.held, schema 18), and the console names the step human. with the reason.
  • Should-fix findings on the change's own documents get one revision, then the gate.
  • Words alone answer a gate where its question names what they stand for; the fix loop's stop passes them to the next fix round.

Iterative planned projects (#246, #250, #251, #252, #259)

  • Every run of a project re-plans and starts the next item, whoever ticked the plan.
  • Plans hold 1 to 3 bigger items; the first is the whole project built to the brief.
  • Every run of a project gets the project's original brief.
  • Unstarted plan stubs pass openspec validate --strict.
  • Home shows a planned project as one row, and its page shows the live plan, runs, state and time (GET /api/projects/{id}).

Checks that match the brief (#232, #234, #248, #253, #254, #255)

  • "Use CI" in Settings, off by default.
  • The judge holds the look to the brief, plays an interactive app through window.__app, and checks the README says how to use it.
  • Briefs that write tests keep the suite small.
  • A Repair turn has no limit unless one is set.

Checks

  • ./tools/test.sh: green.
  • ./tools/sim.sh: 68 of 69. e2e-lane-inferred fails the same way on main.
  • CI: green on a re-run. The first run failed seven test_a_reconnect_backfill… cases at 1.1–2.4 s against the 1 s limit, during a slow job setup (41 s); they take 0.01 s locally and this branch does not touch their path.
  • Browser, against the e2e server: the held label and reason (gate card and review panel), a words-only answer, the project row and page, and the Settings page.
  • openspec validate --all --strict: all but the older repair-agent change, which fails the same way on main.

Schema 18 adds gate.held; take a backup before the deploy.

🤖 Generated with Claude Code

Four OpenSpec changes for braid#232 to braid#259, from the Braid vs Claude Code pinball test. Each change is archived in this branch. **Reliable dispatch, merges and CI evidence** (#237, #245, #247, #249, #258) - An attempt opens only in a write that finds its step ready and its run live: no second gate for one step (run 117), nothing after a stop. - Open gates are the ones on their step's current attempt, everywhere they are counted. - A merge Forgejo defers is read again until it settles when main has not moved; only a base that moved is a conflict. - A failed CI job with no log has its steps run in the clone, so the fixer sees the error. - Script output reaches the transcript while a step is quiet ("waiting on CI" no longer arrives after the wait). - `bash` is the shell tool's second name, with the shell's guards and rules. **Auto-approve, reviews and gate words** (#244, #256, #257) - A fast review lens with findings is a review: Sonnet's reviews are no longer held for a person. - A gate auto-approve leaves for a person records why (`osf.gate.held`, schema 18), and the console names the step `human.` with the reason. - Should-fix findings on the change's own documents get one revision, then the gate. - Words alone answer a gate where its question names what they stand for; the fix loop's stop passes them to the next fix round. **Iterative planned projects** (#246, #250, #251, #252, #259) - Every run of a project re-plans and starts the next item, whoever ticked the plan. - Plans hold 1 to 3 bigger items; the first is the whole project built to the brief. - Every run of a project gets the project's original brief. - Unstarted plan stubs pass `openspec validate --strict`. - Home shows a planned project as one row, and its page shows the live plan, runs, state and time (`GET /api/projects/{id}`). **Checks that match the brief** (#232, #234, #248, #253, #254, #255) - "Use CI" in Settings, off by default. - The judge holds the look to the brief, plays an interactive app through `window.__app`, and checks the README says how to use it. - Briefs that write tests keep the suite small. - A Repair turn has no limit unless one is set. **Checks** - `./tools/test.sh`: green. - `./tools/sim.sh`: 68 of 69. `e2e-lane-inferred` fails the same way on `main`. - CI: green on a re-run. The first run failed seven `test_a_reconnect_backfill…` cases at 1.1–2.4 s against the 1 s limit, during a slow job setup (41 s); they take 0.01 s locally and this branch does not touch their path. - Browser, against the e2e server: the held label and reason (gate card and review panel), a words-only answer, the project row and page, and the Settings page. - `openspec validate --all --strict`: all but the older `repair-agent` change, which fails the same way on `main`. Schema 18 adds `gate.held`; take a backup before the deploy. 🤖 Generated with [Claude Code](https://claude.com/claude-code)
reliable-dispatch-merge-and-ci-evidence (#237, #245, #247, #249, #258),
auto-approve-reviews-and-gate-words (#244, #256, #257),
iterative-planned-projects (#246, #250, #251, #252, #259),
checks-that-match-the-brief (#232, #234, #248, #253, #254, #255).

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
- An attempt opens only in a write that finds its step ready on the attempt before and its run
  live, so a stale ready list opens no second gate (braid#247) and nothing after a stop (#237).
- Open gates are the ones on their step's current attempt, everywhere they are counted.
- A merge Forgejo refuses as not mergeable is read again until it settles when the base is
  already in the branch; only a base that moved is a conflict (braid#245).
- A failed CI job with no log has its run: steps run in the clone until one fails (#249).
- Script output is passed on within the chunk interval even when nothing follows it.
- bash is the shell tool's second name, with the shell's rules (braid#258).
- A fast review lens with findings or reasoning no longer holds auto-approve (braid#244, the
  first task of auto-approve-reviews-and-gate-words).

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
- Auto-approve records osf.gate.held once, with its reason, when it leaves a gate for a person;
  schema 18 keeps it as gate.held, and the console names that step human. and says why
  (braid#244).
- A clean first round whose should-fix findings are on the change's own documents revises
  them once and goes to the gate without another round of lenses (braid#256).
- A gate's question may name the option words alone stand for: Request changes at the
  proposal gates, Try 3 more rounds at a stopped fix loop; the card sends words alone as it,
  and the stop's words reach the next fix round's brief (braid#257).

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
- An item the run's own agent ticked is the run's tick, and a run the plan started carries
  the project on whatever it ticked, so a project no longer stops silently (braid#246).
- A plan's Now milestone holds 1 to 3 items; the first is a working version of the whole
  project built to the brief's quality bar, with an art direction when it asks for a look
  (braid#251).
- A run the plan starts records project.part_of and ends its brief with the project's own
  brief, so every agent sees what the project is for (braid#252).
- The plan's stubs skip their specs until a run writes them, so a delivered repository passes
  openspec validate --all --strict (braid#259).
- Home shows a planned project as one row with its state, items done and time; its page shows
  the live plan, the run behind each item, its runs, why it stopped, and its time split into
  Braid's own and waiting on you (GET /api/projects/{id}, braid#250).

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
- Settings gets Use CI, off by default and kept per run: off, the fix loop never gives a round
  to the project's CI and the briefs say not to add a CI workflow; on, a job's push and pull
  request runs are waited on once (braid#248).
- The change's judge is given the run's brief and holds the look to it, plays an interactive
  app through window.__app, and checks the README says how to use it; apply briefs ask for the
  hook and for the README's use (braid#253, braid#254, braid#232).
- Every brief that writes tests says to keep the suite small: a test per real risk, no sleeps
  or network, several scenarios a test, look and wording to the judge, one browser smoke test
  (braid#255).
- A Repair turn has no limit unless one is set (braid#234).
- The guide follows all four changes, and agent-shell's spec gets a real Purpose.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Tick the browser checks of the four changes.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
chore(openspec): archive the checks-that-match-the-brief change (#232, #234, #248, #253, #254, #255)
All checks were successful
test / suite (pull_request) Successful in 1m11s
5554d64cc0
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
cmoriarty deleted branch claude/batch-232-260 2026-10-27 17:56:40 -04:00
Sign in to join this conversation.
No reviewers
No labels
No milestone
No project
No assignees
1 participant
Notifications
Due date
The due date is invalid or out of range. Please use the format "yyyy-mm-dd".

No due date set.

Dependencies

No dependencies set

Reference
cmoriarty/braid!261
No description provided.