Fixes from the backend comparison: builders look, questions, sessions, judges, lint, guard, loop, spine (#279) #284

Merged
cmoriarty merged 5 commits from claude/batch-283 into main 2026-10-28 18:36:25 -04:00
Owner

Nine findings of the backend comparison (#279), ranked with the operator, in one change so they merge and deploy once. The issues stay open until the llama.cpp rerun on production verifies them.

What changes

  1. The builders see what they build (for #283, #266). Implement, apply and fix steps look once at a change a person sees before finishing: through browser, or with the Playwright tools the build agent now holds in a run without browser. The judge still checks the whole result.
  2. Questions Braid answers itself (for #280). A question asked while the asker's own subagent works gets "end your turn; the report wakes you". On an auto-approved run, a one-option question is answered with that option. Both use #147's recorded automatic reply.
  3. A continued session is not compacted (for #276). At its alarm, a step in a fork goes on in a fresh session from Braid's note, which names the files it changed and the tasks it ticked.
  4. A typed answer keeps its spaces (for #281).
  5. Judge rounds continue (for #282, #266). Each later judge forks the last one (fork_optional), starts from a fresh browser, re-checks what failed and sweeps what passed.
  6. B7 asks for the project's own code (for #278). The apply brief names the blocking rules and calls the rest advice.
  7. The shell guard reads /** in a non-shell heredoc as text (for #264).
  8. The fix loop stops after two rounds in a row without progress (for #267).
  9. The spine (for #265). The glyph shows status only, a verdict chip shows what a check found, and the band shows round N · pass|fail → fixing.

Checks

  • ./tools/test.sh green: 170 backend, 51 UI unit, 16 browser. New: build looks, the guard, questions answered, B7, the judge chain, the loop band, a typed answer with spaces, verdict chips.
  • ./tools/sim.sh: 67 of 68. e2e-lane-inferred fails as on main. The compaction scenarios now show a fresh session at a fork's alarm.
  • Browser: a fix-loop spine with chips and the band.
  • No schema migration.

🤖 Generated with Claude Code

Nine findings of the backend comparison (#279), ranked with the operator, in one change so they merge and deploy once. The issues stay open until the llama.cpp rerun on production verifies them. ## What changes 1. **The builders see what they build** (for #283, #266). Implement, apply and fix steps look once at a change a person sees before finishing: through `browser`, or with the Playwright tools the `build` agent now holds in a run without `browser`. The judge still checks the whole result. 2. **Questions Braid answers itself** (for #280). A question asked while the asker's own subagent works gets "end your turn; the report wakes you". On an auto-approved run, a one-option question is answered with that option. Both use #147's recorded automatic reply. 3. **A continued session is not compacted** (for #276). At its alarm, a step in a fork goes on in a fresh session from Braid's note, which names the files it changed and the tasks it ticked. 4. **A typed answer keeps its spaces** (for #281). 5. **Judge rounds continue** (for #282, #266). Each later judge forks the last one (`fork_optional`), starts from a fresh browser, re-checks what failed and sweeps what passed. 6. **B7 asks for the project's own code** (for #278). The apply brief names the blocking rules and calls the rest advice. 7. **The shell guard** reads `/**` in a non-shell heredoc as text (for #264). 8. **The fix loop** stops after two rounds in a row without progress (for #267). 9. **The spine** (for #265). The glyph shows status only, a verdict chip shows what a check found, and the band shows `round N · pass|fail → fixing`. ## Checks - `./tools/test.sh` green: 170 backend, 51 UI unit, 16 browser. New: build looks, the guard, questions answered, B7, the judge chain, the loop band, a typed answer with spaces, verdict chips. - `./tools/sim.sh`: 67 of 68. `e2e-lane-inferred` fails as on main. The compaction scenarios now show a fresh session at a fork's alarm. - Browser: a fix-loop spine with chips and the band. - No schema migration. 🤖 Generated with [Claude Code](https://claude.com/claude-code)
Work in progress: code, tests and the OpenSpec change; simulated runs not yet updated.

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>
ci: run the suite again (the runner was loaded: seven unrelated tests over their limits)
All checks were successful
test / suite (pull_request) Successful in 35s
5baf68e616
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!284
No description provided.