One OpenSpec pipeline #275

Closed
opened 2026-10-28 12:48:30 -04:00 by cmoriarty · 1 comment
Owner

What happens
Braid ships 13 built-in pipelines. Three build an OpenSpec change (thorough, minimalist, solo), and six are git-flow copies. The #260 re-run showed minimalist saved little and built the worst game, and solo exists only for backends that serve one session at a time.

Fix

  • One OpenSpec pipeline, still named thorough. One build session carries explore, propose, revise, apply and fix; review, code review and the judge stay fresh sessions.
  • Proposal review: four reviewers side by side, or one combined reviewer covering all four lenses when the profile has one parallel session (#274).
  • The build session writes every test, lints them before it finishes, and the scenario check sends gaps back to it (#271, #268).
  • Git-flow is a setting that applies base: develop, work_prefix: feature/; release and hotfix stay their own pipelines.
  • minimalist, solo and the git-flow-* copies become aliases, so old runs and labels still resolve.

Done when
New runs offer thorough, quick-fix, docs, plan and git-flow release and hotfix; a one-session profile runs thorough with one reviewer and no subagents.

Depends on #274.

**What happens** Braid ships 13 built-in pipelines. Three build an OpenSpec change (`thorough`, `minimalist`, `solo`), and six are git-flow copies. The #260 re-run showed `minimalist` saved little and built the worst game, and `solo` exists only for backends that serve one session at a time. **Fix** - One OpenSpec pipeline, still named `thorough`. One build session carries explore, propose, revise, apply and fix; review, code review and the judge stay fresh sessions. - Proposal review: four reviewers side by side, or one combined reviewer covering all four lenses when the profile has one parallel session (#274). - The build session writes every test, lints them before it finishes, and the scenario check sends gaps back to it (#271, #268). - Git-flow is a setting that applies `base: develop, work_prefix: feature/`; release and hotfix stay their own pipelines. - `minimalist`, `solo` and the `git-flow-*` copies become aliases, so old runs and labels still resolve. **Done when** New runs offer `thorough`, `quick-fix`, `docs`, `plan` and git-flow release and hotfix; a one-session profile runs `thorough` with one reviewer and no subagents. Depends on #274.
Author
Owner

Shipped in PR #277 (merge ec8535e), live on production.

  • thorough carries one build session: explore → propose (one step) → revise → apply → spec tests → fix rounds → summary. Reviewers, code review and the judge start fresh.
  • Round one is one combined reviewer when parallel sessions is 1 (#274), otherwise four lenses.
  • An apply step doesn't finish while its own tests have blocking lint findings; the brief asks for tagged tests and ban-check.
  • minimalist/solo → thorough, git-flow-minimalist/git-flow-solo → git-flow. New Run lists thorough, quick-fix, docs (+ release/hotfix in a git-flow repo) and sends git_flow.
  • One change from the plan: apply groups run side by side start their own sessions. opencode keeps the directory a session began in for its system prompt, so a forked group would be told the wrong worktree, and the fix round couldn't continue it. Single and one-after-another apply still carry the session.

Verified on production, run 133 (helloworld, Sonnet 5.5, 2 min 37 s to merge and archive). Each fork's first turn read the parent's whole prompt from cache:

fork from cache read parent peak
propose explore 14,864 14,866
revise propose 23,925 23,927
apply revise 30,352 30,354
summarize apply 41,398 41,400

Four reviewers ran (the combined one skipped: parallel sessions empty). Apply ran ban-check itself and wrote a tagged test for every scenario, so spec tests and reconcile were skipped and the lanes passed in round 1.

Also checked: the suite; 67 of 68 simulated runs (e2e-lane-inferred fails the same way on main); New Run in a browser. The combined reviewer is verified in the simulated solo run only; a production run on the llama.cpp profile with Parallel sessions 1 would show it for real.

Found on the way: #276 (compacting a forked session) and #278 (B7 judges tests by file name).

Shipped in PR #277 (merge ec8535e), live on production. - `thorough` carries one build session: explore → propose (one step) → revise → apply → spec tests → fix rounds → summary. Reviewers, code review and the judge start fresh. - Round one is one combined reviewer when parallel sessions is 1 (#274), otherwise four lenses. - An apply step doesn't finish while its own tests have blocking lint findings; the brief asks for tagged tests and `ban-check`. - `minimalist`/`solo` → `thorough`, `git-flow-minimalist`/`git-flow-solo` → `git-flow`. New Run lists thorough, quick-fix, docs (+ release/hotfix in a git-flow repo) and sends `git_flow`. - One change from the plan: apply groups run **side by side** start their own sessions. opencode keeps the directory a session began in for its system prompt, so a forked group would be told the wrong worktree, and the fix round couldn't continue it. Single and one-after-another apply still carry the session. **Verified on production, run 133** (helloworld, Sonnet 5.5, 2 min 37 s to merge and archive). Each fork's first turn read the parent's whole prompt from cache: | fork | from | cache read | parent peak | |---|---|---|---| | propose | explore | 14,864 | 14,866 | | revise | propose | 23,925 | 23,927 | | apply | revise | 30,352 | 30,354 | | summarize | apply | 41,398 | 41,400 | Four reviewers ran (the combined one skipped: parallel sessions empty). Apply ran `ban-check` itself and wrote a tagged test for every scenario, so spec tests and reconcile were skipped and the lanes passed in round 1. Also checked: the suite; 67 of 68 simulated runs (`e2e-lane-inferred` fails the same way on main); New Run in a browser. The combined reviewer is verified in the simulated `solo` run only; a production run on the llama.cpp profile with Parallel sessions 1 would show it for real. Found on the way: #276 (compacting a forked session) and #278 (B7 judges tests by file name).
Sign in to join this conversation.
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#275
No description provided.