The run 187 batch: apply mode, one test policy, rules past compaction, review diffs (#471) #472

Merged
cmoriarty merged 2 commits from run187-batch into main 2026-11-10 00:32:07 -05:00
Owner

The batch the owner approved from run 187's findings (braid#467), as one OpenSpec change, run187-batch. Closes braid#471.

What changes

  • Apply mode: auto applies in one step only up to six task groups (run 187: eight groups, 10 compactions, 335 min). The tasks rules now hold in either mode.
  • One test policy: one browser smoke test file per project, extended or created; B8 is enforced from the first new file. B10 (blocking) fails a fixed wait or sleep; B11 (recorded) marks a test that only asserts on a source file's text.
  • Past a compaction: the git line, ticking, the test rules, delegation and the running background commands, each sent only to the steps they fit.
  • Background commands: a step waiting on its own isn't judged at the idle or redone as stalled (30-minute cap, within the deadline).
  • Text by kind of step: resume, fork and fresh-start text no longer sends git, task lists or "read only what you change" to read-only and document steps. The note's last words are the agent's own. An attempt that did nothing restarts fresh. A forked planner names the brief by path.
  • Reviews: the digest keeps two findings of one lens on one quote. Round one gets the plan whole and a 30k share of code. Later rounds get only the revision. The code review diffs modified tests and source first, names changed seeds, and leaves openspec/ and lockfiles undiffed.
  • Environment: explore has no file count. Briefs name the installed browsers and an app whose page is built once (the old reload check called run 187's build && tsx watch reloading). The Node .gitignore leaves out Playwright's reports.

Verified

  • ./tools/test.sh passes (690 backend, 119 vitest, 51 browser journeys, 43 s).
  • All 87 sims pass, including a new seven-group auto scenario.
  • No brief grew on the fixture: explore −321, the Visual check −44, apply −16, a fix round −13.
  • Scripted replays of run 187's steps (old size → new):
    • explore: 21,684 → 21,363;
    • round-one review: 174,798 → 146,365;
    • round-two review: 100,444 → 11,479;
    • code review: 243,959 → 235,235, now diffing the modified code and naming the re-seed.
  • A Free replay of run 187's explore didn't run: four Free models refused it. The model-facing check will be the re-run of run 187's brief on vLLM after deploy.

🤖 Generated with Claude Code

The batch the owner approved from run 187's findings (braid#467), as one OpenSpec change, `run187-batch`. Closes braid#471. **What changes** - **Apply mode:** `auto` applies in one step only up to six task groups (run 187: eight groups, 10 compactions, 335 min). The tasks rules now hold in either mode. - **One test policy:** one browser smoke test file per project, extended or created; B8 is enforced from the first new file. B10 (blocking) fails a fixed wait or sleep; B11 (recorded) marks a test that only asserts on a source file's text. - **Past a compaction:** the git line, ticking, the test rules, delegation and the running background commands, each sent only to the steps they fit. - **Background commands:** a step waiting on its own isn't judged at the idle or redone as stalled (30-minute cap, within the deadline). - **Text by kind of step:** resume, fork and fresh-start text no longer sends git, task lists or "read only what you change" to read-only and document steps. The note's last words are the agent's own. An attempt that did nothing restarts fresh. A forked planner names the brief by path. - **Reviews:** the digest keeps two findings of one lens on one quote. Round one gets the plan whole and a 30k share of code. Later rounds get only the revision. The code review diffs modified tests and source first, names changed seeds, and leaves `openspec/` and lockfiles undiffed. - **Environment:** explore has no file count. Briefs name the installed browsers and an app whose page is built once (the old reload check called run 187's `build && tsx watch` reloading). The Node `.gitignore` leaves out Playwright's reports. **Verified** - `./tools/test.sh` passes (690 backend, 119 vitest, 51 browser journeys, 43 s). - All 87 sims pass, including a new seven-group auto scenario. - No brief grew on the fixture: explore −321, the Visual check −44, apply −16, a fix round −13. - Scripted replays of run 187's steps (old size → new): - explore: 21,684 → 21,363; - round-one review: 174,798 → 146,365; - round-two review: 100,444 → 11,479; - code review: 243,959 → 235,235, now diffing the modified code and naming the re-seed. - A Free replay of run 187's explore didn't run: four Free models refused it. The model-facing check will be the re-run of run 187's brief on vLLM after deploy. 🤖 Generated with [Claude Code](https://claude.com/claude-code)
Run 187 (braid#467) took 417 minutes and stopped before its lanes. This is the batch the owner
approved from its findings:

- auto applies in one step only up to six task groups; the tasks rules hold in either mode
- one browser smoke test file per project, B8 enforced from the first; B10 fails a fixed wait,
  B11 records a test that only reads a source file's text
- the post-compaction prompt keeps the git line, the tick, the test rules, delegation and the
  running background commands, each for the steps it fits
- a step waiting on its own background command is not judged at the idle, nor redone as stalled
- resume, fork and fresh-start text by kind of step; the note's last words are the agent's;
  an attempt that did nothing restarts fresh; a forked planner names the brief by path
- the digest keeps two findings of one lens on one quote; round one reads the plan whole and a
  share of code; later rounds the revision only; the code review sees modified tests and source
  first, with changed seeds named
- explore has no file count; the briefs name the installed browsers and an app built once;
  the Node .gitignore leaves out Playwright's reports

No brief grew on the fixture: explore -321, the Visual check -44, apply -16, a fix round -13.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
The doubled diff budget grew the review brief by 42k on run 187's capture; diff and new files now
share 190k, so the modified code takes its room from the new files (235,235 against 243,959). The
one-shot sim checked the scope paragraph in propose's message, which a fork now names by path.
Measured sizes are in the change's design.

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!472
No description provided.