* fix: escalate to model fallback after rate-limit profile rotation cap
Per-model rate limits (e.g. Anthropic Sonnet-only quotas) are not
relieved by rotating auth profiles — if all profiles share the same
model quota, cycling between them loops forever without falling back
to the next model in the configured fallbacks chain.
Apply the same rotation-cap pattern introduced for overloaded_error
(#58348) to rate_limit errors:
- Add `rateLimitedProfileRotations` to auth.cooldowns config (default: 1)
- After N profile rotations on a rate_limit error, throw FailoverError
to trigger cross-provider model fallback
- Add `resolveRateLimitProfileRotationLimit` helper following the same
pattern as `resolveOverloadProfileRotationLimit`
Fixes#58572
* fix: cap prompt-side rate-limit failover (#58707) (thanks @Forgely3D)
* fix: restore latest-main gates for #58707
---------
Co-authored-by: Ember (Forgely3D) <ember@forgely.co>
Co-authored-by: Peter Steinberger <steipete@gmail.com>
* fix: prevent infinite retry loop when live session model switch blocks failover (#58466)
* fix: remove unused resolveOllamaBaseUrlForRun import after rebase
* fix(media): resolve relative MEDIA: paths against agent workspace dir
* fix(agents): remove stale ollama compat import
* fix(media): preserve workspace dir in outbound access
Instead of exponential backoff guesses on Codex 429, probe the WHAM
usage API to determine real availability and write accurate cooldowns.
- Burst/concurrency contention: 15s circuit-break
- Genuine rate limit: proportional to real reset time (capped 2-4h)
- Expired token: 12h cooldown
- Dead account: 24h cooldown
- Probe failure: 30s fail-open
This prevents cascade lockouts when multiple agents share a pool of
Codex profiles, and avoids wasted retries on genuinely exhausted
profiles.
Closes#26329, relates to #1815, #1522, #23996, #54060
Co-authored-by: ryanngit <ryanngit@users.noreply.github.com>
Add toolsAllow field to cron agent-turn payloads, enabling users to
restrict which tool schemas are sent to the model for a given cron job.
When --tools is set:
- Only listed tools are included in the provider request
- promptMode is forced to 'minimal' (strips skills catalog, reply tags,
heartbeat, messaging, docs, memory, model aliases, silent replies)
- Dramatically reduces input tokens for small local models (~16K to ~800)
CLI surface:
- openclaw cron add --tools exec,read,write
- openclaw cron edit <id> --tools exec
- openclaw cron edit <id> --clear-tools (remove allow-list)
Closes#58435
Co-authored-by: andyk-ms <andyk-ms@users.noreply.github.com>
Allow setting provider params (e.g. cacheRetention) once at the
agents.defaults level instead of repeating them per-model. The merge
order is now: defaults.params -> defaults.models[key].params ->
list[agentId].params.
Co-authored-by: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
* Fix: Add ollama to NON_PI_NATIVE_MODEL_PROVIDERS to ensure correct model selection
* Add test coverage for ollama provider to ensure models are merged correctly
* Fix test case for ollama provider by adding required model properties
* Changelog: add Ollama model picker fix
* Changelog: move Ollama fix entry to section tail
---------
Co-authored-by: Bruce MacDonald <brucewmacdonald@gmail.com>
Strip `reasoning.effort: "none"` from OpenAI-compatible payloads so GPT-5 models do not receive the unsupported value when thinking is off. Cover both the shared payload wrapper path and the websocket `response.create` builder with regression tests.
Regeneration-Prompt: |
Fix the OpenAI request-building bug where thinking set to off still forwards
a reasoning payload with effort "none". GPT-5 mini rejects that value with a
400, so disabled reasoning must be represented by omitting the reasoning
parameter entirely. Keep the change additive: sanitize OpenAI-compatible
payloads in OpenClaw’s wrapper layer, make the websocket request builder stop
emitting a reasoning block for "none", and add focused regression tests for
both code paths.
Tools with no parameters produce { type: "object" } schemas without a
properties field. The OpenAI Responses API rejects these, silently
crashing entire sessions.
Add properties: {} injection in normalizeToolParameters() and
convertTools() to ensure all object-type schemas include a properties
field.
Closes#58246
Vincent Koc
nanakotsai
Forgely3D
sandpile
Peter Steinberger
yi-bot
Neerav Makwana
joshavant
Morrow
Reed
狼哥
Han Yang
Kenny Xie
ryanngit
Andy
Lee Pender
Mingxuan
Gustavo Madeira Santana
oliviareid-svg
Josh Lehman
Christof Salis
yelog