After the comparison: lint, briefs, spine, review findings, inference per backend (#279) #297

Merged
cmoriarty merged 10 commits from claude/batch-294 into main 2026-10-28 21:58:58 -04:00
Owner

The batch after the backend comparison (#279): three OpenSpec changes, one merge.

Fixes after the comparison (fixes-after-the-comparison)

  • #294: lint rule B7 is gone. ban-check prints only the rules that block (--all for every rule), and the briefs list only those. The test-only refusal and the scenario-test brief say a helper goes under a test directory.
  • #286: the propose brief says Braid writes the .gitignore right after it; the file's header names Braid.
  • #285: a simulated run whose osfd hangs up says so, and on which call. --logs prints the end of its osfd log, and the image self-check passes it.
  • #293: script.scenarios and git-flow's release steps have a phase. A drift test checks every built-in node.
  • #290: leaving a run for another page forgets its landing, so its tab opens on its active step.
  • #291, #292: the running row's hover is steady (full name, activity, no clock). The line drops the per-action clocks and keeps the silence warning; the step's time (or time left) sits under its ring.

Review findings panel (review-findings-panel)

  • #287: a review's chip counts as the digest merges (6 to fix · 5 notes) in the tone of the worst finding: red, amber, grey or green. The code review's major findings count as things to fix.
  • #288: lenses give each finding a title, and the revision writes outcomes.json. GET /api/runs/{id}/review/{round} serves a round, and a panel under the review and revision steps lists it by weight; a click on a review chip opens it.

Inference per backend (inference-per-backend)

  • #289: the page draws only what its backend reports. llama.cpp's /metrics and /slots give its requests, each slot's context, prompt reuse and draft acceptance; tabbyAPI shows Braid's own rows; a hosted API shows Braid's tokens and their cost at the profile's prices. Every chart has a value axis.
  • #295: while llama.cpp's slots keep reading a prompt, the watchdog counts the server as working, so a long prefill is no longer redone as a stall.

Checked

  • ./tools/test.sh: 198 backend tests, 56 unit tests, 18 browser specs.
  • ./tools/sim.sh: 67 of 68. e2e-lane-inferred fails on main as well, filed as #296.
  • In a browser: the findings panel on a seeded review; the Inference page against trogdor's llama.cpp (context slot 0 · 1.3k of 262k, reuse 99%, drafts 91%) and on a priced Claude profile (no slots, tokens and cost).

🤖 Generated with Claude Code

The batch after the backend comparison (#279): three OpenSpec changes, one merge. **Fixes after the comparison** (`fixes-after-the-comparison`) - #294: lint rule B7 is gone. `ban-check` prints only the rules that block (`--all` for every rule), and the briefs list only those. The test-only refusal and the scenario-test brief say a helper goes under a test directory. - #286: the propose brief says Braid writes the `.gitignore` right after it; the file's header names Braid. - #285: a simulated run whose osfd hangs up says so, and on which call. `--logs` prints the end of its osfd log, and the image self-check passes it. - #293: `script.scenarios` and git-flow's release steps have a phase. A drift test checks every built-in node. - #290: leaving a run for another page forgets its landing, so its tab opens on its active step. - #291, #292: the running row's hover is steady (full name, activity, no clock). The line drops the per-action clocks and keeps the silence warning; the step's time (or time left) sits under its ring. **Review findings panel** (`review-findings-panel`) - #287: a review's chip counts as the digest merges (`6 to fix · 5 notes`) in the tone of the worst finding: red, amber, grey or green. The code review's major findings count as things to fix. - #288: lenses give each finding a title, and the revision writes `outcomes.json`. `GET /api/runs/{id}/review/{round}` serves a round, and a panel under the review and revision steps lists it by weight; a click on a review chip opens it. **Inference per backend** (`inference-per-backend`) - #289: the page draws only what its backend reports. llama.cpp's `/metrics` and `/slots` give its requests, each slot's context, prompt reuse and draft acceptance; tabbyAPI shows Braid's own rows; a hosted API shows Braid's tokens and their cost at the profile's prices. Every chart has a value axis. - #295: while llama.cpp's slots keep reading a prompt, the watchdog counts the server as working, so a long prefill is no longer redone as a stall. **Checked** - `./tools/test.sh`: 198 backend tests, 56 unit tests, 18 browser specs. - `./tools/sim.sh`: 67 of 68. `e2e-lane-inferred` fails on main as well, filed as #296. - In a browser: the findings panel on a seeded review; the Inference page against trogdor's llama.cpp (context `slot 0 · 1.3k of 262k`, reuse 99%, drafts 91%) and on a priced Claude profile (no slots, tokens and cost). 🤖 Generated with [Claude Code](https://claude.com/claude-code)
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
B7 asked that a test import the module it tests. It went by names and import
shapes, and runs 133, 134 and 136 renamed code, rewrote a script's loading and
failed a step to please it. ban-check now prints only the rules that block
(--all for every rule), and the briefs list only those. The test-only refusal
and the scenario-test brief say a helper goes under a test directory.

The propose brief says Braid writes the .gitignore right after it; the file's
header names Braid. A simulated run whose osfd hangs up says so, on which call,
and --logs prints the end of its osfd log; the image self-check passes it.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
script.scenarios and git-flow's release steps had no phase, so the spine drew a
band of their first segment; a drift test now checks every built-in node.
Leaving a run for another page forgets its landing, so its tab opens on its
active step again. The running row's hover holds the full name and the
activity in words with no clock, so the browser shows it; the line drops the
per-action clocks and keeps the silence warning, and the step's time (or its
time left) sits under the ring.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
A review's chip counts its findings as the digest merges them, in words a
reader knows (2 blocking · 6 to fix · 5 notes, nothing found), and takes the
tone of the worst: red, amber, grey or green. The code review counts major as
to fix and minor as notes. The verdict carries the tone; the step stores it.

Lenses give each finding an optional title; the revision records what it did
with each finding by its id in .osf/review/r<n>/outcomes.json.
GET /api/runs/{id}/review/{round} serves a round, and ReviewFindings shows it
under the transcript of the round's review and revision steps, grouped by
weight; a click on a review chip opens it.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
A metrics kind per backend replaces the vLLM-only flag. llama.cpp's /metrics
and /slots give its requests, prompt reuse, draft acceptance and each slot's
context in use; a server without --metrics reports none rather than reading as
down. The snapshot declares the rows its backend feeds. Samples keep the draft
counters, and the history works out draft acceptance and Braid's own tokens
per bucket; a profile may have prices, its own and never another's.

The sampler follows the prompt tokens llama.cpp's slots have read: while that
count keeps rising, the watchdog counts the server as working, so a long
prefill is no longer redone as a stall.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Seen on the page against trogdor's llama.cpp and a priced Claude profile.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
chore(openspec): tick the batch's checks done so far
Some checks failed
test / suite (pull_request) Failing after 28s
218badae7b
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
ci: run the suite again (a small test over its time limit on a loaded runner)
All checks were successful
test / suite (pull_request) Successful in 35s
b939e554c9
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!297
No description provided.