openclaw/extensions/kimi-coding
Peter Steinberger f024e67860
fix: preserve thinking choices across models and runtimes (#150106)
* fix: preserve thinking choices across models and runtimes

Consolidate thinking defaults, runtime capability selection, and request intent.
Keep native none distinct from logical off, honor declared effort restrictions,
and share provider wire policy with standalone completions.

Preserve mandatory OpenRouter reasoning without an effort selector, refresh
incomplete cached capabilities, and retain default-on reasoning replay.
Remove duplicate policy and wrappers while keeping existing configuration.

* fix: preserve prepared thinking state and defaults

Keep standalone wrappers alive across calls and resolve provider thinking before serialization. Preserve trusted runtime alias context, align transport fixtures with verified provider contracts, and consolidate duplicate test setup.

* fix: preserve omitted reasoning and scoped model capabilities
2026-09-16 10:56:40 -07:00
..
assets feat: give every plugin a compact chat activity icon (#147333) 2026-09-13 13:49:51 -07:00
api.ts
implicit-provider.test.ts fix: preserve thinking choices across models and runtimes (#150106) 2026-09-16 10:56:40 -07:00
index.test.ts fix: preserve thinking choices across models and runtimes (#150106) 2026-09-16 10:56:40 -07:00
index.ts fix(kimi): stop treating quota exhaustion as authentication failure (#142656) 2026-09-09 08:16:50 +05:30
kimi-coding.live.test.ts fix(live): repair model E2E selection and session fixtures (#131229) 2026-08-27 20:09:39 -07:00
onboard.test.ts refactor(kimi): remove duplicate onboarding model assertion (#149190) 2026-09-15 23:47:02 +08:00
onboard.ts fix(providers): avoid saving generated setup model rows (#140139) 2026-09-06 19:10:23 +05:30
openclaw.plugin.json fix: preserve thinking choices across models and runtimes (#150106) 2026-09-16 10:56:40 -07:00
package.json chore(release): close out 2026.9.4 on main (#144708) 2026-09-11 18:46:37 +08:00
provider-catalog.test.ts fix: preserve thinking choices across models and runtimes (#150106) 2026-09-16 10:56:40 -07:00
provider-catalog.ts fix(kimi): normalize legacy model IDs from manifest metadata (#146261) 2026-09-12 11:00:50 -07:00
provider-policy-api.test.ts fix(kimi): refresh K3 catalog and add 256K variant (#113918) 2026-07-25 20:57:26 -07:00
provider-policy-api.ts fix(kimi): refresh K3 catalog and add 256K variant (#113918) 2026-07-25 20:57:26 -07:00
README.md fix(kimi): stop treating quota exhaustion as authentication failure (#142656) 2026-09-09 08:16:50 +05:30
replay-policy.test.ts
replay-policy.ts
stream.test.ts fix: preserve thinking choices across models and runtimes (#150106) 2026-09-16 10:56:40 -07:00
stream.ts fix: preserve thinking choices across models and runtimes (#150106) 2026-09-16 10:56:40 -07:00
tsconfig.json refactor(plugins): inherit shared package boundary settings (#117474) 2026-08-01 09:11:09 -07:00

OpenClaw Kimi Coding Provider

Official OpenClaw provider plugin for Kimi Coding.

Install from OpenClaw:

openclaw plugins install @openclaw/kimi-provider
openclaw gateway restart

See https://docs.openclaw.ai/providers/moonshot for setup and configuration.

Quota errors

Kimi can return HTTP 403 when a weekly usage limit is exhausted. OpenClaw treats explicit weekly, seven-day, or quota-reset errors as rate limits. Wait for the provider's quota window to reset or use another configured provider; replacing a valid API key does not restore quota. Invalid keys and access restrictions keep their authentication error handling.

Catalog notes

Model rows live in openclaw.plugin.json under modelCatalog.providers.kimi.

  • k3 serves up to 1M context, tier-gated server-side; k3-256k is the cheaper 256K variant of the same weights. Both point at moonshot/kimi-k3 through upstreamModel, which keeps their compat capability tiers aligned with the moonshot catalog for the same model.
  • Legacy k3[1m] was retired upstream and normalizes to k3 for shipped configurations.
  • KIMI_K3_MODEL_IDS in provider-policy-api.ts must cover exactly the catalog rows that carry a K3 thinkingLevelMap; provider-catalog.test.ts asserts it.