The run 181 batch: steer carries on, judge look lists, per-step thinking levels, adopt-on-resume and Ask #421

Merged
cmoriarty merged 1 commit from batch-2026-11-08 into main 2026-11-08 01:55:03 -05:00
Owner

What

The batch after run 181 (owner, 2026-11-08: "collect a large batch of issues to implement together"), as five OpenSpec changes on one branch, each validated --strict:

  • briefs-after-run-181 (#412, #413, #416, #418): a Steer or queued message to a step's agent ends with "act on this, then carry on with your tasks"; the whole-change apply's retry with tasks left continues its session, told which tasks remain (retry_continues widened to unfinished); the apply and fix briefs name ../screenshots; the motion probe is named at the judge's rule and the probes block sits by the rules; a read of a file the code review's brief pasted whole is answered with its heading until the file changes (pasted.json + the guard); the sim gains a steer operator action and the scenario apply-steered.
  • judge-round-two-mechanics (#414, #415, #417, #419): Braid computes round 2's look list and carry list by id from the records, writes them to .osf/fix/rN/look.json, and false_carries fails a carry from the look list; the paused clock holds WebSocket and EventSource message events and fetch responses until stepped past or resumed; the text probe reads a gradient's first stop; the browser subagent loses read, grep, glob and list (run 181: 80 file reads in 350 calls, 87% of a check's time in short turns), and the guidance briefs it with the hook's names and selectors; scenario judge-round-two-css-fix.
  • per-step-thinking-level (#420): step_efforts per profile by step family, the model's highest for every step by default, kept by a run, a table under Reasoning effort, the hover lists the levels in force.
  • resume-adopts-profile (#352): a stopped or failed run's detail lists what moved in its profile (backend_moved), the run page offers "Resume with the backend as it is now" (POST /resume {adopt: true} rewrites the copy; the supervisor re-provisions a server still up), and the apply fan-out decides one step or groups on the adopted copy when reached. No re-plan step was needed.
  • ask-mode (#223): the helper conversation is Ask while the run is live (ask agent, read-only by permissions), Repair once it stops, with a note that it turned; POST /repair accepts a live run and records the agent; an Agent | Ask switch in the message box (A), remembered per run, Send only in Ask; a nudge (Ask, Stop and repair) beside a step silent past the watchdog's threshold.

Also: docs/guide.md for each, sim/README.md for steer.

Verification

  • ./tools/test.sh green in 34 s (backend, types, vitest, build, 48 browser journeys).
  • ./tools/sim.sh --smoke, apply-steered and judge-round-two-css-fix pass.
  • New tests: tests/engine/test_operator_messages.py, test_mediation (unfinished tasks), test_fork_room (continue on tasks left), test_browser_delegation (screenshots line, browser guidance), test_see_motion (rule and probes order); tests/pipeline/test_later_rounds (look list, carry on the look list fails); tests/agent/test_path_guard (pasted file), test_review_agent (browser and ask permissions), test_effort (step levels); tests/api/test_settings (step levels kept per profile and per run), test_resume_adopt, test_ask; ui vitest effort.test.ts and ask.test.ts; journeys: clock holds a socket, gradient probe, step level saved, adopting resume, Ask switch and nudge.
  • Not covered here: the production run that measures the judge after these briefs (the next Scratch run), and #419's re-measurement after it.

🤖 Generated with Claude Code

## What The batch after run 181 (owner, 2026-11-08: "collect a large batch of issues to implement together"), as five OpenSpec changes on one branch, each validated `--strict`: - **briefs-after-run-181** (#412, #413, #416, #418): a Steer or queued message to a step's agent ends with "act on this, then carry on with your tasks"; the whole-change apply's retry with tasks left continues its session, told which tasks remain (`retry_continues` widened to `unfinished`); the apply and fix briefs name `../screenshots`; the motion probe is named at the judge's rule and the probes block sits by the rules; a `read` of a file the code review's brief pasted whole is answered with its heading until the file changes (`pasted.json` + the guard); the sim gains a `steer` operator action and the scenario `apply-steered`. - **judge-round-two-mechanics** (#414, #415, #417, #419): Braid computes round 2's look list and carry list by id from the records, writes them to `.osf/fix/rN/look.json`, and `false_carries` fails a carry from the look list; the paused clock holds WebSocket and EventSource `message` events and `fetch` responses until stepped past or resumed; the text probe reads a gradient's first stop; the `browser` subagent loses `read`, `grep`, `glob` and `list` (run 181: 80 file reads in 350 calls, 87% of a check's time in short turns), and the guidance briefs it with the hook's names and selectors; scenario `judge-round-two-css-fix`. - **per-step-thinking-level** (#420): `step_efforts` per profile by step family, the model's highest for every step by default, kept by a run, a table under Reasoning effort, the hover lists the levels in force. - **resume-adopts-profile** (#352): a stopped or failed run's detail lists what moved in its profile (`backend_moved`), the run page offers "Resume with the backend as it is now" (`POST /resume {adopt: true}` rewrites the copy; the supervisor re-provisions a server still up), and the apply fan-out decides one step or groups on the adopted copy when reached. No re-plan step was needed. - **ask-mode** (#223): the helper conversation is Ask while the run is live (`ask` agent, read-only by permissions), Repair once it stops, with a note that it turned; `POST /repair` accepts a live run and records the agent; an Agent | Ask switch in the message box (`A`), remembered per run, Send only in Ask; a nudge (Ask, Stop and repair) beside a step silent past the watchdog's threshold. Also: `docs/guide.md` for each, `sim/README.md` for `steer`. ## Verification - `./tools/test.sh` green in 34 s (backend, types, vitest, build, 48 browser journeys). - `./tools/sim.sh --smoke`, `apply-steered` and `judge-round-two-css-fix` pass. - New tests: tests/engine/test_operator_messages.py, test_mediation (unfinished tasks), test_fork_room (continue on tasks left), test_browser_delegation (screenshots line, browser guidance), test_see_motion (rule and probes order); tests/pipeline/test_later_rounds (look list, carry on the look list fails); tests/agent/test_path_guard (pasted file), test_review_agent (browser and ask permissions), test_effort (step levels); tests/api/test_settings (step levels kept per profile and per run), test_resume_adopt, test_ask; ui vitest effort.test.ts and ask.test.ts; journeys: clock holds a socket, gradient probe, step level saved, adopting resume, Ask switch and nudge. - Not covered here: the production run that measures the judge after these briefs (the next Scratch run), and #419's re-measurement after it. 🤖 Generated with [Claude Code](https://claude.com/claude-code)
Five OpenSpec changes, each validated strict, from the issues run 181 exposed and the
owner's batch of 2026-11-08:

- briefs-after-run-181 (braid#412 #413 #416 #418): a Steer or queued message to a step's
  agent ends with "act on this, then carry on"; the whole-change apply's retry with tasks
  left continues its session, told which tasks remain; the apply and fix briefs name
  `../screenshots`; the motion probe is named at the judge's rule and the probes sit by the
  rules; a `read` of a file the code review's brief pasted whole is answered with its heading
  until the file changes; the sim gains a `steer` operator action and `apply-steered`.
- judge-round-two-mechanics (braid#414 #415 #417 #419): Braid computes round 2's look list
  and carry list by id and a carry from the look list fails the check; the paused clock holds
  WebSocket and EventSource messages and fetch responses; the text probe reads a gradient's
  first stop; the `browser` subagent loses read, grep, glob and list, and the guidance briefs
  it with the hook's names and selectors; sim `judge-round-two-css-fix`.
- per-step-thinking-level (braid#420): `step_efforts` per profile by step family, the
  highest for every step by default, kept by a run, shown in the effort hover.
- resume-adopts-profile (braid#352): a stopped or failed run lists what moved in its profile
  and offers "Resume with the backend as it is now"; the apply fan-out then decides on the
  adopted copy.
- ask-mode (braid#223): the helper conversation is Ask while the run is live, read-only by
  its permissions, Repair once it stops; an Agent | Ask switch in the message box; a nudge on
  a silent step.

./tools/test.sh green (34 s); ./tools/sim.sh --smoke, apply-steered and
judge-round-two-css-fix pass.

Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
cmoriarty deleted branch batch-2026-11-08 2026-11-08 01:55:03 -05:00
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!421
No description provided.