Short step names, a wordmark that threads once, Ask closes, deletes in the background, one session count per backend (#546 #543 #547 #542 #535 #536) #550

Merged
cmoriarty merged 11 commits from claude/console-fixes-and-session-cap into main 2026-11-15 18:04:09 -05:00
Owner

Batch A from the 2026-11-15 issue review: six issues in one change, console-fixes-and-session-cap (proposal, specs, design and tasks under openspec/changes/).

What changed

  • #546 Step names. Every step purpose and condition is about three words, from the table approved on #546, and the box's condition drops its only. A gate names the decision ("Approve the merge"), so auto-approve needs no rewrite. A task group is named by its plan's heading in the transcript and the Stop dialog too, not "Implement group N", and the tasks rules ask for each heading in about three words. A run with no plan keeps the number. At 1440 px no name or condition wraps, half-width boxes included (measured on every box of a seeded thorough run).
  • #543 Wordmark. It threads once per page load: not on a remount, a reconnect or a screen change. The count is kept on the page, so a remounted bar cannot restart it. When all is well, the dot's hover says how long the server has been up, from started_at, which is new in /api/status.
  • #547 Ask. osf.repair.opened names its agent. An Ask period closes when its run is no longer live, and Ask going on as Repair closes Ask's period first. The spine's entry and the helper tab read "Ask" for a conversation that was only Ask's. Records from before this change read "Repair".
  • #542 Deleting runs. POST /api/runs/delete answers 202 and deletes in a background job, one job at a time. GET /api/runs/delete/{id} gives the progress and, once the job ends, the results. The dialog sends the whole selection and follows the job. A timed-out request names what was asked and no longer blames other tabs.
  • #535 Session cap. The ceiling is read per run, from that run's own backend, and counts agent steps together with the places granted to their helpers. A helper started in the background asks osfd for a place (POST /api/runs/{run}/helper-place) and gives it back when it ends. A helper its step waits for works in the step's place, one at a time per step. If osfd does not answer within 2 s, the guard falls back to its old gate.
  • #536, its smaller items.
    • Fix rounds and quick fixes fail when git ignores code they wrote.
    • On vLLM, a step is not redone as stalled while the server's prompt counter rises or its cache fills as it runs a request. A waiting request already spared a step.
    • stop takes an optional id and stops only that helper. Each stop call is handled once, so a second stop test works.

Checked

  • ./tools/test.sh passes in 51 s on the laptop.
  • ./tools/sim.sh --smoke passes all 7 scenarios, and helpers-wait-for-a-place, helpers-queue-in-the-background and apply-writes-ignored-source pass.
  • Each backend fix has a test that fails on the code before it: the Ask period, the delete job, the per-backend ceiling, helper places, the ignored-file check on fix steps, vLLM reading, and stop by id.
  • Tried on a seeded local server before this pull request.

The graph at 1440 px

From a seeded test server. The apply groups' titles come from the run's plan; the other step names are the new ones.

The run's graph at the Apply phase

The fix loop's half-width boxes, each name and condition on one line

Not done

  • Task 5.5's dedicated sim scenario, with two review lenses each starting a background reader, is not written. The scheduler test and the guard's node test cover the same case.
  • The vLLM signal is built from how vLLM fills its cache and counts prompts. Run 193's own samples could not be read back, because the run was deleted from production.

🤖 Generated with Claude Code

Batch A from the 2026-11-15 issue review: six issues in one change, `console-fixes-and-session-cap` (proposal, specs, design and tasks under `openspec/changes/`). ## What changed - **#546 Step names.** Every step purpose and condition is about three words, from the table approved on #546, and the box's condition drops its `only`. A gate names the decision ("Approve the merge"), so auto-approve needs no rewrite. A task group is named by its plan's heading in the transcript and the Stop dialog too, not "Implement group N", and the tasks rules ask for each heading in about three words. A run with no plan keeps the number. At 1440 px no name or condition wraps, half-width boxes included (measured on every box of a seeded thorough run). - **#543 Wordmark.** It threads once per page load: not on a remount, a reconnect or a screen change. The count is kept on the page, so a remounted bar cannot restart it. When all is well, the dot's hover says how long the server has been up, from `started_at`, which is new in `/api/status`. - **#547 Ask.** `osf.repair.opened` names its agent. An Ask period closes when its run is no longer live, and Ask going on as Repair closes Ask's period first. The spine's entry and the helper tab read "Ask" for a conversation that was only Ask's. Records from before this change read "Repair". - **#542 Deleting runs.** `POST /api/runs/delete` answers 202 and deletes in a background job, one job at a time. `GET /api/runs/delete/{id}` gives the progress and, once the job ends, the results. The dialog sends the whole selection and follows the job. A timed-out request names what was asked and no longer blames other tabs. - **#535 Session cap.** The ceiling is read per run, from that run's own backend, and counts agent steps together with the places granted to their helpers. A helper started in the background asks osfd for a place (`POST /api/runs/{run}/helper-place`) and gives it back when it ends. A helper its step waits for works in the step's place, one at a time per step. If osfd does not answer within 2 s, the guard falls back to its old gate. - **#536, its smaller items.** - Fix rounds and quick fixes fail when git ignores code they wrote. - On vLLM, a step is not redone as stalled while the server's prompt counter rises or its cache fills as it runs a request. A waiting request already spared a step. - `stop` takes an optional `id` and stops only that helper. Each `stop` call is handled once, so a second `stop test` works. ## Checked - `./tools/test.sh` passes in 51 s on the laptop. - `./tools/sim.sh --smoke` passes all 7 scenarios, and `helpers-wait-for-a-place`, `helpers-queue-in-the-background` and `apply-writes-ignored-source` pass. - Each backend fix has a test that fails on the code before it: the Ask period, the delete job, the per-backend ceiling, helper places, the ignored-file check on fix steps, vLLM reading, and `stop` by id. - Tried on a seeded local server before this pull request. ## The graph at 1440 px From a seeded test server. The apply groups' titles come from the run's plan; the other step names are the new ones. ![The run's graph at the Apply phase](/attachments/e7c3cc98-2c02-4cd5-99bc-349050330247) ![The fix loop's half-width boxes, each name and condition on one line](/attachments/5a34b7eb-6d01-4c86-bb9c-ef53978223a2) ## Not done - Task 5.5's dedicated sim scenario, with two review lenses each starting a background reader, is not written. The scheduler test and the guard's node test cover the same case. - The vLLM signal is built from how vLLM fills its cache and counts prompts. Run 193's own samples could not be read back, because the run was deleted from production. 🤖 Generated with [Claude Code](https://claude.com/claude-code)
The graph's step purposes and conditions follow the table approved on braid#546, the
list and transcript included; a gate names the decision, so auto-approve needs no rewrite.
The wordmark plays once per page load, counted on the page so a remounted bar cannot
restart it, and the dot's hover gives the server's uptime from `/api/status`'s new
`started_at` when all is well.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
`osf.repair.opened` names its agent. A supervisor pass closes an Ask period whose run is no
longer live, and Ask going on as Repair closes Ask's period first. The spine's entry and the
helper's tab read Ask for a conversation that was only Ask's; a record from before reads Repair.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
`POST /api/runs/delete` answers 202 at once and deletes its runs in a job, one job at a time;
`GET /api/runs/delete/{id}` says how far it has got and, once ended, each run's result. The
dialog sends the whole selection and follows the job. A request osfd does not answer in time now
names what was asked and no longer blames other tabs.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
The ceiling is read for each run from the backend it was started on, and counts the agent steps
on that backend with the places granted to their helpers. A helper started in the background asks
osfd for a place (`POST /api/runs/{run}/helper-place`) and gives it back when its session ends; one
its step waits for works in the step's place, one at a time per step. When osfd does not answer,
the guard falls back to its own gate.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
vLLM gives no count of a prompt read so far; its prompt counter rising, or its cache filling while
it runs a request, now marks the server as reading, under the same three-window bound as
llama.cpp's slots (braid#295). A waiting request already spared a step.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
`stop` takes an optional `id`, the sessionID its subagent call returned, and stops only that child
of the step. Stops are kept by the call's own id, so a second `stop test` stops the `test` started
since. A helper the guard had queued is not started once its step stopped helpers of its name.

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): tick the local try of console-fixes-and-session-cap
All checks were successful
test / suite (pull_request) Successful in 1m36s
345d0ae779
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
feat(console): a task group goes by its plan's name in the transcript and the Stop dialog (#546)
All checks were successful
test / suite (pull_request) Successful in 1m38s
2952c9cf82
The graph named a group by its plan's heading; the transcript's step rows and the Stop dialog said
"Implement group N". They now read the same titles, by step id, and only a run with no plan keeps
the number. The tasks rules ask for each group's heading in about three words, as their example
`Draft autosave` already was.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
cmoriarty deleted branch claude/console-fixes-and-session-cap 2026-11-15 18:04:10 -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!550
No description provided.