Honest speeds and effort, per-run backends, Home by repository (#344 #351 #345 #349 #324 #321 #322) #355

Merged
cmoriarty merged 7 commits from claude/batch-b-numbers-home-4f128c into main 2026-11-01 03:02:21 -05:00
Owner

Batch B: honest numbers and a tidier Home.

  • #344: a request's decode ends at its last token or its last tool call's start, not at step-finish after the tools ran.
  • #351: while nothing finished in the window, the decode tile shows the live rate (vLLM's own generation counter where it has one), and the prefill tile its last value and age.
  • #345: a subagent is named from opencode 2's info.agent; an unnamed one reads "Subagent".
  • #349: reasoning effort in the model's own levels (Qwen3.8 xhigh/medium/low, GLM-5.3 max/high/low/none), Server default and Auto; Connect drops levels the template refuses or ignores; sent on every self-hosted kind, fastllm and tabbyAPI included. An unsaved profile sends nothing (never below the template's default); Qwen3.8 recommends xhigh.
  • #324: GET /api/backends; the top bar shows Braid's services only; each run's bar and Home row show its own backend; the Inference page lists each backend in use with its runs.
  • #321, #322: Home grouped by repository; a planned project is archived/restored in one step and deleted whole.

OpenSpec: honest-speeds-and-effort, backends-and-home (both validate --strict). ./tools/test.sh green, ./tools/sim.sh --smoke 7/7.

🤖 Generated with Claude Code

Batch B: honest numbers and a tidier Home. - #344: a request's decode ends at its last token or its last tool call's start, not at `step-finish` after the tools ran. - #351: while nothing finished in the window, the decode tile shows the live rate (vLLM's own generation counter where it has one), and the prefill tile its last value and age. - #345: a subagent is named from opencode 2's `info.agent`; an unnamed one reads "Subagent". - #349: reasoning effort in the model's own levels (Qwen3.8 xhigh/medium/low, GLM-5.3 max/high/low/none), Server default and Auto; Connect drops levels the template refuses or ignores; sent on every self-hosted kind, fastllm and tabbyAPI included. An unsaved profile sends nothing (never below the template's default); Qwen3.8 recommends xhigh. - #324: `GET /api/backends`; the top bar shows Braid's services only; each run's bar and Home row show its own backend; the Inference page lists each backend in use with its runs. - #321, #322: Home grouped by repository; a planned project is archived/restored in one step and deleted whole. OpenSpec: `honest-speeds-and-effort`, `backends-and-home` (both `validate --strict`). `./tools/test.sh` green, `./tools/sim.sh --smoke` 7/7. 🤖 Generated with [Claude Code](https://claude.com/claude-code)
A request's time ends at its last streamed token or when its last tool call started running,
not at the step-finish after its tools ran. While nothing finished in the window, the decode
tile shows the live rate (vLLM's own generation counter where it has one), and the prefill
tile the last value and its age.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Each profile offers its model's levels (Qwen3.8 xhigh/medium/low, GLM-5.3 max/high/low/none),
Server default (sends nothing) and Auto, whose tiers map onto those levels. Connect renders each
level through /tokenize or /apply-template and drops those the template refuses or ignores.
fastllm and tabbyAPI now get chat_template_kwargs too. A profile with nothing saved sends
nothing, so it never thinks lower than the template's default; Qwen3.8 is recommended xhigh.
Old saved words keep their meaning (high off Anthropic sends nothing).

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
GET /api/backends lists each backend in use with its runs, slot use and reachability, and
Braid's services (osfd, the forge, each backend, the Free router). The top bar drops the
Inference and Slots cells for a Services cell; the run bar names its run's backend; the
Inference page lists each backend in use with its runs.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Each repository is a group of its projects and standalone runs, ordered by latest activity with
a live one first; a live run's row names its backend. A project's row is reached by Select and
has Archive/Restore; its page has Archive and Delete. POST /api/projects/{id}/archive and
/unarchive hide or restore every run of it in one write, refusing a project with a run going.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
docs(openspec): proposals, designs and tasks for honest-speeds-and-effort and backends-and-home
All checks were successful
test / suite (pull_request) Successful in 34s
b0a30a1ee7
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
fix(effort): the old high and an unsaved profile think at the model's highest level (#349)
All checks were successful
test / suite (pull_request) Successful in 35s
7212af0cca
Sending nothing for high was run 160's bug: fastllm's Qwen3.8 template defaults to medium.
A tier's word now maps to the model's level for that tier, so high is xhigh on Qwen3.8 on
every kind, and the default never thinks lower than the model's highest.

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!355
No description provided.