Update Luci/PKA Kimi model references for Kimi-K2.7-Code release. Scope: - Claude Code provider-switch harness (`PKA/scripts/claude-provider-env.sh`) default Kimi model. - Dire...
StateDoneNext ActionClosedOwnerLuciRuntimeClosedAge19h ago
Ticket is done; runtime is closed.·profile claude_opus_1m_high
Description
MC-5038
Update Luci/PKA Kimi model references for Kimi-K2.7-Code release.
Scope:
- Claude Code provider-switch harness (`PKA/scripts/claude-provider-env.sh`) default Kimi model.
- Direct Kimi coding API helper (`PKA/scripts/kimi-coding-prompt.py`) default model.
- Scheduler profile metadata (`scheduler.py`) for kimi_cli / kimi_api / claude_kimi where useful.
- Legacy dormant dispatcher env strings only if still used as library/import path.
- Provider smoke/eval harness (`scripts/provider_smoke_test.py`, `scripts/provider_behavior_eval.py`, `tasks/provider-smoke.md`).
- PKA/runtime docs and ai-backends skill/docs if they encode old Kimi model defaults.
Acceptance:
- Validate the exact model id accepted by the Kimi Coding endpoint before committing. If vendor accepts an alias but reports canonical `kimi-for-coding`, document both.
- Smoke-test direct Kimi helper and provider-switched Claude-Kimi path when credentials are present.
- Do not touch historical reports/memory backups except living docs or active harnesses.
- Commit/push source/config/docs changes only; leave unrelated dashboard/generated dirty files untouched.
Activity
done
INTERACTIVE
Luci is working...
Details —
Done
· Critical
· Luci
▼
SState
Done
Closed
PPeople
TTiming / Details▼
telegram (human)
Mission Control
19h ago
18h ago
Advanced / Operator evidence
RRouting owner
ROperator console
Ticket is done; runtime is closed.MC is visibility-only. Hermes Luci launches and gates work outside MC, then mirrors evidence/status here.
WWorkflow
Start Dev Review + QA ▾
Select phases to include:
Agents
Review Gates
Decision
WAT routing: choose an agent, review gate, or decision. Buttons use the live runtime when one is attached.
Started from Telegram: Elmar asked to update all Kimi model aliases/harness/API calls to Kimi-K2.7-Code after release screenshot.
luci18h ago
Kimi K2.7 Code update completed and verified.
Findings:
- Kimi Coding /models endpoint lists stable id `kimi-for-coding` with display_name `K2.7 Code`, 262144 context, reasoning support.
- OpenRouter currently exposes `moonshotai/kimi-latest` and `moonshotai/kimi-k2.6`; no K2.7 slug yet, so OpenRouter aliases were not changed.
Code/source commits:
- luci-workspace branch `kb/MC-5038-kimi-27`, commit `0d5583d9`.
- PKA branch `kb/MC-5038-kimi-27`, commit `2d649bc0`.
Changed:
- Provider smoke/eval harnesses now request `kimi-for-coding` and label it as K2.7 Code.
- Scheduler/runtime metadata documents `display_model: K2.7 Code` for Kimi API/CLI profiles.
- PKA Claude provider switch/helper scripts keep stable id `kimi-for-coding`, document display_name K2.7 Code, and `upgrade-claude-models.sh --dry-run` now resolves the local script and Kimi models endpoint correctly.
- Hermes default + Iris/Miki/builder profile Kimi aliases and auxiliary Kimi models updated from `kimi-k2.6` to `kimi-for-coding`.
- Iris Kimi cron job model values updated in-place.
- Gateway and Iris gateway restarted and active.
Verification:
- Kimi `/models` live check returned `kimi-for-coding` / `K2.7 Code`.
- Direct helper: `KIMI27_OK`, response model `kimi-for-coding`.
- Claude provider-switch path: `KIMI27_PROVIDER_OK` with `ANTHROPIC_MODEL=kimi-for-coding`.
- Hermes default profile smoke: `OK-KIMI27`.
- Iris profile smoke: `OK-IRIS-KIMI27`.
- Workspace py_compile + provider_behavior_eval unittest + simulated and live Kimi eval passed.
- PKA bash -n / py_compile / upgrade-claude-models --dry-run passed.
Notes:
- First direct Kimi helper call briefly hit 429 engine_overloaded, retry succeeded.
- Graphify update was run in both worktrees; report churn was restored to avoid committing noisy generated diffs.
luci18h ago
Final landing proof: luci-workspace `3c34558b` is on origin/master; PKA `adb10a64` is on origin/master. Gateway services active after restart. Marking done.
luci18h ago
Reran Kimi benchmark after K2.7 Code update.
Run: reports/provider_eval/MC-5006-kimi-k27-rerun-20260612T112417Z.md
JSON: reports/provider_eval/MC-5006-kimi-k27-rerun-20260612T112417Z.json
Model identity: requested/resolved `kimi-for-coding (K2.7 Code)`.
Result: 4/5 scenarios passed, parse 5/5, hard score 66/68, average latency 16.1s.
- PASS Life Manager triage: 14/14
- PASS Life Manager digest: 16/16
- PASS Memory extraction: 13/13
- FAIL Ops/watchdog summary: 11/13; swapped severities (`run-2` expected high got medium; `run-3` expected medium got high)
- PASS Code-review-lite: 12/12
Conclusion unchanged for alerting: Kimi K2.7 is fine for triage/digest/memory/code-review-lite style static scheduled tasks, but still not safe as the deterministic ops/watchdog alert severity gate without prompt/rubric tightening.
luci18h ago
Stable copy saved under /home/lucienne/workspace/reports/provider_eval/MC-5006-kimi-k27-rerun-20260612T112417Z.md with raw outputs under /home/lucienne/workspace/reports/provider_eval/runs/MC-5006-kimi-k27-20260612T112417Z/.
luci-board-manager19h ago
[visibility-only] Routed the Kimi K2.7 Code model-default update outside MC to internal Kanban card t_63a9b452 (codexbuilder), using the existing isolated worktree `/home/lucienne/workspace/_mc_internal_worktrees/MC-5038-kimi-27` and instructing it to reconcile the sibling PKA worktree if needed. Verified worker pid=3090364 cwd matches the isolated worktree. MC is parked as waiting/kanban_active while the external worker finishes validation/commit evidence. No MC runtime/pickup/send/harvest endpoint was used.