[PR #393] [MERGED] fix(cursor-review): move the default Moonshot cell back to kimi-k3-max #395

New issue

Closed

opened 2026-10-26 22:39:20 +00:00 by bsakaguchi · 0 comments

bsakaguchi commented

2026-10-26 22:39:20 +00:00

Owner

📋 Pull Request Information

Original PR: https://github.com/Comfy-Org/github-workflows/pull/393
Author: @mattmillerai
Created: 10/9/2026
Status: ✅ Merged
Merged: 10/9/2026
Merged by: @mattmillerai

Base: main ← Head: matt/default-panel-kimi-max

📝 Commits (3)

1175e24 fix(cursor-review): move the default Moonshot cell back to kimi-k3-max
90e4e38 docs(cursor-review): rejoin the panel_models sentence the Kimi note split
2dadf0e docs(cursor-review): move the remaining kimi-k3-high examples to the -max pin

📊 Changes

4 files changed (+60 additions, -12 deletions)

View changed files

📝 .github/cursor-review/README.md (+16 -1)
📝 .github/cursor-review/prompt-judge.md (+1 -1)
📝 .github/workflows/cursor-review.yml (+40 -7)
📝 docs/callers/cursor-review.md (+3 -3)

📄 Description

kimi-k3-high is capacity-bound, so the built-in panel is currently one lane short on every consumer that has not overridden it — which is most of them.

The cells don't run slow, they die

~25 seconds in, RetriableError: [resource_exhausted] after three internal reconnects, zero findings, leg red, Panel integrity red with it.


measured	56 panels, one consumer
`kimi-k3-high` cells red	24 of 76 (32%)
of those, this error	19
on the last day	12 of 22

Not a property of one repo — the same model threw the same error the same day on two unrelated consumers. And #388's in-step retry does not rescue it, because the condition is sustained rather than transient: both cells on this repo's own panel for #392 exhausted all three attempts (30s, then 90s).

`-max`, not removal — the lane earns its place

The 2026-08-27 attribution that cut Gemini was re-run over 1,158 judge-kept findings (82 PRs), parsed from the judge's own _Raised by N of M reviewers (...)_ lines and normalised per finding raised, so panel size flatters nobody:

model	raised	sole	distinct%	sole/PR	c/h sole
`kimi-k3-max`	229	39	17%	1.05	3/58
`kimi-k3-high`	81	22	27%	1.05	0/11
`gemini-3.1-pro`	99	12	12%	0.33	0/58

Kimi contributes 1.05 unique findings per PR at both tiers and uniquely caught 3 of 58 critical/high at -max. Gemini — cut for exactly this reason — managed 0.33 and 0 of 58. Dropping Kimi would cost roughly three times what dropping Gemini did, so the fix is the tier, not the lane.

What this costs, stated plainly

It gives back the spend the -high notch in #243 was for. This is a deliberate partial reversal of that change, not an oversight.
The reliability it buys is not established at this tier under present conditions. -max was 0 of 36 cells red, but that was measured 2026-10-07, before the capacity trouble, and the two tiers may share a pool.

Both are recorded in the workflow and the README so the next person doesn't rediscover them — along with what the evidence says to do if -max goes the same way: drop to a two-lab Cursor panel rather than substitute Gemini or Grok. Grok 4.7-high was tried on 2026-10-09 and raised 0 unique findings in 17 while timing out on 11 of 20 cells — worse than Gemini on value and worse than Kimi on reliability.

Scope

Default change, so it reaches every consumer on the next SHA bump. A repo that wants the old tier back sets CURSOR_PANEL_MODELS. README panel table and history updated alongside, and every remaining kimi-k3-high example outside the era-pinned tests moved to -max — the CURSOR_PANEL_MODELS copy-paste examples in docs/callers/cursor-review.md (so they again differ from the built-in list by exactly one axis), the judge prompt's attribution example, and the leg-context-name examples.

.github/cursor-review/tests     1257 OK
.github/workflow-pins/tests      551 OK  + check_workflow_pins.py OK
check_agents_md.py --root .     passed

The historical-era fixtures and test_catalog_drift.py's REAL_PANEL pin their own eras by design (per #243) and are deliberately untouched.

Provenance

Verified:
- .github/cursor-review/tests: 1257 OK; .github/workflow-pins/tests: OK (43 skipped)
- check_workflow_pins.py, check_agents_md.py --root ., check-org-repo-literals.sh: pass
- PR CI at head 2dadf0e: pending at handoff; no merge-queue run yet
Deviations: none — the diff matches the description (default Moonshot cell kimi-k3-high → kimi-k3-max, workflow comment, README rationale, and the stale -high doc/prompt examples)

🤖 Generated with Claude Code

_{🔄 This issue represents a GitHub Pull Request. It cannot be merged through Gitea due to API limitations.}

## 📋 Pull Request Information **Original PR:** https://github.com/Comfy-Org/github-workflows/pull/393 **Author:** [@mattmillerai](https://github.com/mattmillerai) **Created:** 10/9/2026 **Status:** ✅ Merged **Merged:** 10/9/2026 **Merged by:** [@mattmillerai](https://github.com/mattmillerai) **Base:** `main` ← **Head:** `matt/default-panel-kimi-max` --- ### 📝 Commits (3) - [`1175e24`](https://github.com/Comfy-Org/github-workflows/commit/1175e244817267da89d4839e4777cb17a7b30f7d) fix(cursor-review): move the default Moonshot cell back to kimi-k3-max - [`90e4e38`](https://github.com/Comfy-Org/github-workflows/commit/90e4e388e5ebdda637491af4235f527cf3871bd4) docs(cursor-review): rejoin the panel_models sentence the Kimi note split - [`2dadf0e`](https://github.com/Comfy-Org/github-workflows/commit/2dadf0eb448ec82f69b06b3c2625b397e5877c53) docs(cursor-review): move the remaining kimi-k3-high examples to the -max pin ### 📊 Changes **4 files changed** (+60 additions, -12 deletions) <details> <summary>View changed files</summary> 📝 `.github/cursor-review/README.md` (+16 -1) 📝 `.github/cursor-review/prompt-judge.md` (+1 -1) 📝 `.github/workflows/cursor-review.yml` (+40 -7) 📝 `docs/callers/cursor-review.md` (+3 -3) </details> ### 📄 Description `kimi-k3-high` is capacity-bound, so the built-in panel is currently **one lane short on every consumer that has not overridden it** — which is most of them. ## The cells don't run slow, they die ~25 seconds in, `RetriableError: [resource_exhausted]` after three internal reconnects, zero findings, leg red, `Panel integrity` red with it. | | | |---|---| | measured | 56 panels, one consumer | | `kimi-k3-high` cells red | **24 of 76 (32%)** | | of those, this error | 19 | | on the last day | **12 of 22** | Not a property of one repo — the same model threw the same error the same day on two unrelated consumers. And **#388's in-step retry does not rescue it**, because the condition is sustained rather than transient: both cells on this repo's own panel for #392 exhausted all three attempts (30s, then 90s). ## `-max`, not removal — the lane earns its place The 2026-08-27 attribution that cut Gemini was re-run over **1,158 judge-kept findings (82 PRs)**, parsed from the judge's own `_Raised by N of M reviewers (...)_` lines and normalised per finding raised, so panel size flatters nobody: | model | raised | sole | distinct% | **sole/PR** | c/h sole | |---|---|---|---|---|---| | `kimi-k3-max` | 229 | 39 | 17% | **1.05** | **3/58** | | `kimi-k3-high` | 81 | 22 | 27% | **1.05** | 0/11 | | `gemini-3.1-pro` | 99 | 12 | 12% | 0.33 | 0/58 | Kimi contributes **1.05 unique findings per PR at both tiers** and uniquely caught 3 of 58 critical/high at `-max`. Gemini — cut for exactly this reason — managed 0.33 and 0 of 58. Dropping Kimi would cost roughly **three times** what dropping Gemini did, so the fix is the tier, not the lane. ## What this costs, stated plainly - It **gives back the spend** the `-high` notch in #243 was for. This is a deliberate partial reversal of that change, not an oversight. - The reliability it buys is **not established at this tier under present conditions**. `-max` was 0 of 36 cells red, but that was measured 2026-10-07, before the capacity trouble, and the two tiers may share a pool. Both are recorded in the workflow and the README so the next person doesn't rediscover them — along with what the evidence says to do if `-max` goes the same way: **drop to a two-lab Cursor panel rather than substitute Gemini or Grok.** Grok `4.7-high` was tried on 2026-10-09 and raised **0 unique findings in 17** while timing out on 11 of 20 cells — worse than Gemini on value *and* worse than Kimi on reliability. ## Scope Default change, so it reaches every consumer on the next SHA bump. A repo that wants the old tier back sets `CURSOR_PANEL_MODELS`. README panel table and history updated alongside, and every remaining `kimi-k3-high` example outside the era-pinned tests moved to `-max` — the `CURSOR_PANEL_MODELS` copy-paste examples in `docs/callers/cursor-review.md` (so they again differ from the built-in list by exactly one axis), the judge prompt's attribution example, and the leg-context-name examples. ``` .github/cursor-review/tests 1257 OK .github/workflow-pins/tests 551 OK + check_workflow_pins.py OK check_agents_md.py --root . passed ``` The historical-era fixtures and `test_catalog_drift.py`'s `REAL_PANEL` pin their own eras by design (per #243) and are deliberately untouched. ## Provenance - **Verified:** - `.github/cursor-review/tests`: 1257 OK; `.github/workflow-pins/tests`: OK (43 skipped) - `check_workflow_pins.py`, `check_agents_md.py --root .`, `check-org-repo-literals.sh`: pass - PR CI at head 2dadf0e: pending at handoff; no merge-queue run yet - **Deviations:** none — the diff matches the description (default Moonshot cell `kimi-k3-high` → `kimi-k3-max`, workflow comment, README rationale, and the stale `-high` doc/prompt examples) 🤖 Generated with [Claude Code](https://claude.com/claude-code) --- <sub>🔄 This issue represents a GitHub Pull Request. It cannot be merged through Gitea due to API limitations.</sub>

bsakaguchi

2026-10-26 22:39:20 +00:00