Free profile: what is left after #164 (a run that finishes, checks not yet recorded) #205

Closed
opened 2026-10-18 11:09:38 -04:00 by cmoriarty · 4 comments
Owner

Follow-up to #164, #186 and #195, archived in 5a953a0. The Free profile is live (production c1228b4) and its refusal handling is proven on runs 76-78. What remains:

Make a Free run finish

  • #202: Kilo's beam refuses any session another model began (unfinished work on wip/202-beam-reasoning).
  • #204: a refused session's move picks a Zen model stalled as busy.
  • A run on free providers alone that finishes: the local solo run of free-profile's 13.2, or a production run after #202.

From #164's tasks, not yet done (see the archived free-profile/tasks.md notes)

  • 1.6: Gemini's free-tier limits from AI Studio (operator only), and their numbers in the design (12.2).
  • 4.2: run the deploy README's commands as written.
  • 11.2: three simulated runs never written: free-hold-between-steps, free-vision-spent-mid-browser, free-no-vision.
  • 11.3: the slow router test drives requests itself; make opencode go through the router.
  • 13.3: on production, osfd doctor, osfd's env and mounts holding no key, and the reach checks.

From #195

  • malformed_function_call (Gemini's finish reason) counted against the model, like invalid_output.
  • See a 10-minute cut handled live under the new 3,600 s router cap.

From #186: 10.4, the timing of a review lens's and the judge's compactions (runs 73/74 measured the apply's).

Follow-up to #164, #186 and #195, archived in 5a953a0. The Free profile is live (production c1228b4) and its refusal handling is proven on runs 76-78. What remains: **Make a Free run finish** - #202: Kilo's beam refuses any session another model began (unfinished work on `wip/202-beam-reasoning`). - #204: a refused session's move picks a Zen model stalled as busy. - A run on free providers alone that finishes: the local solo run of free-profile's 13.2, or a production run after #202. **From #164's tasks, not yet done** (see the archived `free-profile/tasks.md` notes) - 1.6: Gemini's free-tier limits from AI Studio (operator only), and their numbers in the design (12.2). - 4.2: run the deploy README's commands as written. - 11.2: three simulated runs never written: free-hold-between-steps, free-vision-spent-mid-browser, free-no-vision. - 11.3: the slow router test drives requests itself; make opencode go through the router. - 13.3: on production, `osfd doctor`, osfd's env and mounts holding no key, and the reach checks. **From #195** - `malformed_function_call` (Gemini's finish reason) counted against the model, like `invalid_output`. - See a 10-minute cut handled live under the new 3,600 s router cap. **From #186:** 10.4, the timing of a review lens's and the judge's compactions (runs 73/74 measured the apply's).
Author
Owner

Looked at with the Free batch of 2026-10-30 (#325, #326, #327). Still open here:

  • The two missing simulated runs (free-hold-between-steps, free-vision-spent-mid-browser) need the fake router to play quota running out between steps, and a screenshot read. That is harness work.
  • malformed_function_call: no production event records how Gemini's finish reason arrives, so there is nothing yet to count. The nearest case seen is #327, beam-free's spilled paths, now repaired in the guards.
  • Operator-only: 1.6, Gemini's limits from AI Studio.
  • Still to check on production: 13.3, osfd doctor and the reach checks. I'll do it after the batch deploys.
**Looked at with the Free batch of 2026-10-30 (#325, #326, #327).** Still open here: - **The two missing simulated runs** (`free-hold-between-steps`, `free-vision-spent-mid-browser`) need the fake router to play quota running out between steps, and a screenshot read. That is harness work. - **`malformed_function_call`**: no production event records how Gemini's finish reason arrives, so there is nothing yet to count. The nearest case seen is #327, beam-free's spilled paths, now repaired in the guards. - **Operator-only:** 1.6, Gemini's limits from AI Studio. - **Still to check on production:** 13.3, `osfd doctor` and the reach checks. I'll do it after the batch deploys.
Author
Owner

Where it stands, 2026-10-30.

Done:

  • #202 is closed (on the sim, free-reasoning-required).
  • A Free run has finished: run 144.
  • 13.3: osfd doctor on production (c959528) reads 0 failed. Every Free check is ok: the router answers at the pinned 0.1.103, the admin key authenticates, a secret reference is refused, and all five tiers have models, vision being Gemini 3.5 and 3.1 Flash-Lite. One warning: the sandbox is off (#201, #203).
  • Free runs now resume (#331), get Repair (#325), keep test artifacts out of their checks (#326), have beam-free's spilled paths cut (#327), and look with subagents off (#333).

Left:

  • #204's proof: a fake Zen lane in the sims.
  • 11.2: the simulated runs free-hold-between-steps and free-vision-spent-mid-browser, which need the fake router to play quota running out between steps, and a screenshot read.
  • 11.3: the slow router test through opencode.
  • #195: malformed_function_call, never yet seen on production, and a 10-minute cut seen live.
  • #186 10.4: the compaction timings.
  • Operator-only: 1.6, Gemini's limits from AI Studio; 4.2, the deploy README run as written.
**Where it stands, 2026-10-30.** **Done:** - #202 is closed (on the sim, `free-reasoning-required`). - A Free run has finished: run 144. - 13.3: `osfd doctor` on production (c959528) reads 0 failed. Every Free check is ok: the router answers at the pinned 0.1.103, the admin key authenticates, a secret reference is refused, and all five tiers have models, vision being Gemini 3.5 and 3.1 Flash-Lite. One warning: the sandbox is off (#201, #203). - Free runs now resume (#331), get Repair (#325), keep test artifacts out of their checks (#326), have beam-free's spilled paths cut (#327), and look with subagents off (#333). **Left:** - #204's proof: a fake Zen lane in the sims. - 11.2: the simulated runs `free-hold-between-steps` and `free-vision-spent-mid-browser`, which need the fake router to play quota running out between steps, and a screenshot read. - 11.3: the slow router test through opencode. - #195: `malformed_function_call`, never yet seen on production, and a 10-minute cut seen live. - #186 10.4: the compaction timings. - Operator-only: 1.6, Gemini's limits from AI Studio; 4.2, the deploy README run as written.
Author
Owner

Progress from PR 362. Staying open for what is left.

Done or replaced

  • #204: proven by free-zen-busy-passed-over, closed.
  • 11.2 free-vision-spent-mid-browser: replaced by #338's work.
  • 11.3: replaced by tools/replay.sh free, where opencode goes through the real router.

Not written: free-hold-between-steps. The hold a scenario reaches is a day-long demand hold, which it can't sit out. Testing it needs the fake ledger to spend quota on a served request.

Still open

  • 1.6: Gemini's free-tier limits (operator only).
  • malformed_function_call counted against the model.
  • 11.2 free-no-vision, unchecked in this pass.
  • 4.2, 13.3, the 10-minute cut and #186's 10.4 were not touched.
Progress from PR 362. Staying open for what is left. **Done or replaced** - #204: proven by `free-zen-busy-passed-over`, closed. - 11.2 `free-vision-spent-mid-browser`: replaced by #338's work. - 11.3: replaced by `tools/replay.sh free`, where opencode goes through the real router. **Not written:** `free-hold-between-steps`. The hold a scenario reaches is a day-long demand hold, which it can't sit out. Testing it needs the fake ledger to spend quota on a served request. **Still open** - 1.6: Gemini's free-tier limits (operator only). - `malformed_function_call` counted against the model. - 11.2 `free-no-vision`, unchecked in this pass. - 4.2, 13.3, the 10-minute cut and #186's 10.4 were not touched.
Author
Owner

Closing: a Free run finishes and what this tracked has shipped.

  • Run 144 finished on Free alone. braid#202 and braid#204 are closed. 13.3 passed on production on 10-30.
  • The free-no-vision sim exists (sim/free-no-vision-passed.yaml). 11.2 and 11.3 were replaced by braid#338's work and tools/replay.sh free.

Dropped, unless a run shows a need:

  • malformed_function_call counted against the model. It has never been seen on production, and nothing in the code reads it.
  • The free-hold-between-steps sim, the live 10-minute cut, and braid#186's 10.4 timings.
  • Operator only: 1.6, Gemini's free-tier limits, and 4.2, the deploy README run as written.

The archived change free-runs-finish closes with this.

Closing: a Free run finishes and what this tracked has shipped. - Run 144 finished on Free alone. braid#202 and braid#204 are closed. 13.3 passed on production on 10-30. - The `free-no-vision` sim exists (`sim/free-no-vision-passed.yaml`). 11.2 and 11.3 were replaced by braid#338's work and `tools/replay.sh free`. Dropped, unless a run shows a need: - `malformed_function_call` counted against the model. It has never been seen on production, and nothing in the code reads it. - The `free-hold-between-steps` sim, the live 10-minute cut, and braid#186's 10.4 timings. - Operator only: 1.6, Gemini's free-tier limits, and 4.2, the deploy README run as written. The archived change `free-runs-finish` closes with this.
Sign in to join this conversation.
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#205
No description provided.