Profiles remember their settings; one OpenSpec pipeline (#274, #275) #277

Merged
cmoriarty merged 7 commits from claude/one-pipeline into main 2026-10-28 13:57:27 -04:00
Owner

Closes #274 and #275 once verified on production.

Profiles remember their settings (#274)

  • A profile holds every setting except Display: its backend, plus the Agents, Runs and Approvals sections. Switching profiles brings its values back; editing saves to the active profile. A profile saved before this reads the global values until it is next saved.
  • New per-profile Parallel sessions: how many sessions the backend serves at once, empty for no limit. Connect fills it from a llama.cpp or tabbyAPI server's slots. Setting it to 1 sets subagents at once to 0, concurrent runs to 1 and Apply to One step. A run keeps the value it started with.
  • Planning and Repair borrow only another profile's backend.

One OpenSpec pipeline (#275)

  • thorough carries one build session through explore, the one propose step, revisions, apply, spec tests, fix rounds and the summary. Reviewers, code review and the judge keep fresh sessions. Apply groups run side by side start their own sessions: opencode keeps the directory a session began in for its system prompt, so a fork would name the wrong worktree, and the fix round could not continue it.
  • Round one of the review is one combined reviewer when parallel sessions is 1. Braid writes each lens's record from that one reading; a lens it did not look at is not reviewed.
  • An apply step does not finish while its own tests have blocking lint findings (#271); the brief asks for tagged tests (#268).
  • minimalist/solo resolve to thorough, the git-flow forms to git-flow. New Run lists thorough, quick-fix and docs, plus release and hotfix in a git-flow repository, and sends git_flow. The solo-only subagents step is gone.

Checks

  • ./tools/test.sh green (158 backend, 48 UI unit, 14 browser).
  • ./tools/sim.sh: 67 of 68 pass; e2e-lane-inferred fails the same way on main.
  • Browser: Settings profile switching and the one-session defaults; New Run choices with and without git flow, and a run recorded on git-flow.
  • No schema migration.

🤖 Generated with Claude Code

Closes #274 and #275 once verified on production. ## Profiles remember their settings (#274) - A profile holds every setting except Display: its backend, plus the Agents, Runs and Approvals sections. Switching profiles brings its values back; editing saves to the active profile. A profile saved before this reads the global values until it is next saved. - New per-profile **Parallel sessions**: how many sessions the backend serves at once, empty for no limit. Connect fills it from a llama.cpp or tabbyAPI server's slots. Setting it to 1 sets subagents at once to 0, concurrent runs to 1 and Apply to One step. A run keeps the value it started with. - Planning and Repair borrow only another profile's backend. ## One OpenSpec pipeline (#275) - `thorough` carries one build session through explore, the one propose step, revisions, apply, spec tests, fix rounds and the summary. Reviewers, code review and the judge keep fresh sessions. Apply groups run side by side start their own sessions: opencode keeps the directory a session began in for its system prompt, so a fork would name the wrong worktree, and the fix round could not continue it. - Round one of the review is one combined reviewer when parallel sessions is 1. Braid writes each lens's record from that one reading; a lens it did not look at is not reviewed. - An apply step does not finish while its own tests have blocking lint findings (#271); the brief asks for tagged tests (#268). - `minimalist`/`solo` resolve to `thorough`, the git-flow forms to `git-flow`. New Run lists thorough, quick-fix and docs, plus release and hotfix in a git-flow repository, and sends `git_flow`. The solo-only subagents step is gone. ## Checks - `./tools/test.sh` green (158 backend, 48 UI unit, 14 browser). - `./tools/sim.sh`: 67 of 68 pass; `e2e-lane-inferred` fails the same way on main. - Browser: Settings profile switching and the one-session defaults; New Run choices with and without git flow, and a run recorded on `git-flow`. - No schema migration. 🤖 Generated with [Claude Code](https://claude.com/claude-code)
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
A profile now keeps the Agents, Runs and Approvals settings beside its backend, plus a new
parallel sessions value Connect reads from llama.cpp's slots. Switching profiles brings them
back; a profile saved before reads the global values until it is next saved.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
thorough carries one build session through explore, the one propose step,
the revisions, apply, the spec tests, the fix rounds and the summary; the
reviewers, the code review and the judge keep sessions of their own. Apply
groups run side by side start sessions of their own in their own worktrees,
since opencode keeps the directory a session began in for its system prompt.

- Round one of the review is one combined reviewer when the run's parallel
  sessions is 1 (agent.review.combined, when: one_session); Braid writes each
  lens's record from its one reading, and a lens it did not look at is not
  reviewed. The `when:` conditions now read the run's parallel sessions.
- An apply step does not finish while its own tests have blocking lint
  findings (tests-lint-clean); the brief asks for tagged tests and ban-check.
- minimalist and solo resolve to thorough, git-flow-minimalist and
  git-flow-solo to git-flow; the New Run page lists thorough, quick-fix and
  docs, plus release and hotfix in a git-flow repository, and sends git_flow.
  The solo-only subagents step is gone.
- Simulated runs move to the one propose step; solo is a one-session thorough
  run showing the combined reviewer. The fake model counts a forked session's
  turns after compaction from its own replies.

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 one-openspec-pipeline change (#275)
Some checks failed
test / suite (pull_request) Failing after 23s
0ee4426417
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
perf(pipelines): derive each built-in's text once per process
All checks were successful
test / suite (pull_request) Successful in 32s
67e70eb452
The default pipeline is now derived from the node library (#275), and the
settings document reads it on every request: 53 ms a call against main's
24 ms, which took a new settings test past its small-test limit on CI. The
text depends on module constants alone, so it is cached: 18 ms a call.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
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!277
No description provided.