fix(mistral): repair max-token defaults and doctor migration by vincentkoc · Pull Request #53054 · openclaw/openclaw

vincentkoc · 2026-03-23T17:42:01Z

Summary

Problem: built-in Mistral defaults seeded maxTokens equal to full context for several models, which can deterministically trigger Mistral 422 rejects on normal prompts.
Why it matters: the bad values were duplicated in both the bundled Mistral plugin catalog and the core provider-onboarding defaults, and persisted once written into user config.
What changed: lowered bundled/core Mistral output defaults, taught config normalization and openclaw doctor --fix to repair stale persisted Mistral provider model entries, and added regression coverage to keep the core and bundled defaults aligned.
What did NOT change (scope boundary): this does not change user-overridden Mistral model settings that are already below the safe thresholds, and it does not add a new command path beyond existing doctor/config normalization.

Change Type (select all)

Scope (select all touched areas)

Linked Issue/PR

Closes [Bug]: Mistral AI Error 422 #52599
Related fix(mistral): lower default maxTokens to prevent 422 rejection (#52599) #53006
This PR fixes a bug or regression

Root Cause / Regression History (if applicable)

Root cause: OpenClaw shipped context-sized maxTokens defaults for Mistral models even though Mistral treats max_tokens as output budget, not total context.
Missing detection / guardrail: the bundled plugin catalog and the core provider-onboarding seed path drifted independently, and we had no repair pass for already-persisted provider model values.
Prior context (git blame, prior PR, issue, or refactor if known): #53006 identified the extension-side catalog fix but missed the duplicate core seed path and the persisted-config repair path.
Why this regressed now: any onboarding or provider reconfiguration that wrote the old Mistral defaults preserved them indefinitely in config.
If unknown, what was ruled out: N/A.

Regression Test Plan (if applicable)

Coverage level that should have caught this:
- Unit test
- Seam / integration test
- End-to-end test
- Existing coverage already sufficient
Target test or file:
- extensions/mistral/model-definitions.test.ts
- src/commands/onboard-auth.test.ts
- src/config/model-alias-defaults.test.ts
- src/commands/doctor-legacy-config.migrations.test.ts
Scenario the test should lock in: bundled catalog defaults, core provider seed defaults, config normalization, and doctor migration all converge on the same safe Mistral max-token values.
Why this is the smallest reliable guardrail: the bug existed in both fresh-onboarding and already-persisted config paths; these tests cover both without requiring live Mistral credentials.
Existing test that already covers this (if any): onboarding/provider tests existed, but they only covered the stale 262144 defaults.
If no new test is added, why not: N/A.

User-visible / Behavior Changes

Fresh Mistral onboarding now seeds safer default maxTokens values.
Existing configs with stale context-sized Mistral maxTokens are repaired by config normalization and openclaw doctor --fix.

Security Impact (required)

New permissions/capabilities? (No)
Secrets/tokens handling changed? (No)
New/changed network calls? (No)
Command/tool execution surface changed? (No)
Data access scope changed? (No)
If any Yes, explain risk + mitigation: N/A

Repro + Verification

Environment

OS: macOS
Runtime/container: Node 24.13.0 via NODENV_VERSION=24.13.0
Model/provider: Mistral provider defaults/config repair
Integration/channel (if any): N/A
Relevant config (redacted): N/A

Steps

Seed or onboard Mistral with the old default provider model definitions.
Observe maxTokens equal to the full model context window.
Run provider default application or openclaw doctor --fix.

Expected

Mistral models use safe output-token defaults instead of context-sized output budgets.
Persisted stale Mistral provider configs get normalized.

Actual

Before this patch, fresh and persisted Mistral defaults could retain 262144 / 128000 output budgets and trigger deterministic 422 rejects.

Evidence

Failing test/log before + passing after
Trace/log snippets
Screenshot/recording
Perf numbers (if relevant)

Human Verification (required)

Verified scenarios:
- scoped test suite passed for Mistral catalog, onboarding, config normalization, and doctor migration
- pnpm build passed
Edge cases checked:
- non-Mistral provider max-token clamping still behaves normally
- core and bundled Mistral defaults stay aligned
- persisted Mistral configs below the repaired thresholds are left alone
What you did not verify:
- live Mistral API request/response on a real key in this branch

Review Conversations

I replied to or resolved every bot review conversation I addressed in this PR.
I left unresolved only the conversations that still need reviewer or maintainer judgment.

Compatibility / Migration

Backward compatible? (Yes)
Config/env changes? (No)
Migration needed? (Yes)
If yes, exact upgrade steps:
- upgrade to this build
- run openclaw doctor --fix to rewrite stale persisted Mistral provider model defaults

Failure Recovery (if this breaks)

How to disable/revert this change quickly: revert this commit or manually override Mistral provider model maxTokens in config.
Files/config to restore: extensions/mistral/model-definitions.ts, src/plugins/provider-model-definitions.ts, src/config/defaults.ts, src/commands/doctor-legacy-config.ts
Known bad symptoms reviewers should watch for: Mistral onboarding still writing 262144 / 128000, or doctor not updating persisted Mistral provider model entries.

Risks and Mitigations

Risk: exact Mistral per-model output ceilings are inconsistently documented publicly.
- Mitigation: used conservative safe defaults, verified against public Mistral docs plus secondary public registries, and added repair logic only for stale context-sized values.

greptile-apps · 2026-03-23T17:52:18Z

Greptile Summary

This PR fixes a long-standing bug where Mistral provider models were seeded with maxTokens equal to the full context window (e.g. 262144), which Mistral treats as an output budget and rejects with a 422 on typical prompts. The fix operates across three layers: the bundled Mistral plugin catalog (extensions/mistral/model-definitions.ts), the core provider-onboarding seed path (src/plugins/provider-model-definitions.ts), and a new normalizeLegacyMistralModelMaxTokens migration in doctor-legacy-config.ts that repairs already-persisted configs. A shared resolveNormalizedProviderModelMaxTokens helper in src/config/defaults.ts centralises the decision logic and is reused by both the live applyModelDefaults normalisation path and the migration, ensuring both converge on the same values.

Bundled catalog maxTokens values corrected: mistral-large-latest (262144 → 16384), devstral-medium-latest (262144 → 32768), magistral-small (128000 → 40000), mistral-medium-2508 (262144 → 8192), pixtral-large-latest (128000 → 32768).
New MISTRAL_SAFE_MAX_TOKENS_BY_MODEL lookup table drives the repair; the fallback for unlisted models is DEFAULT_MODEL_MAX_TOKENS = 8192. Note that mistral-small-latest (catalog value 16384) is not in this map — a persisted context-window-sized entry for that model would be conservatively repaired to 8192 rather than 16384. Neither value triggers a 422, and mistral-small-latest was never set to the context-window size in any known shipped catalog, so this is a low-risk gap.
The repair condition (rawMaxTokens >= contextWindow) correctly leaves user-set values that are already below the context window untouched, matching the stated scope boundary.
Test coverage is strong: unit tests lock in the catalog values, applyModelDefaults normalisation, doctor migration messages, and an alignment test between the core and bundled builders to prevent future drift.

Confidence Score: 4/5

Safe to merge; fixes a deterministic 422 regression with no user-data or security risk, and all affected paths have test coverage.
The fix is correctly scoped, the shared helper is logically sound, and three independent test suites verify fresh-onboarding, runtime normalisation, and persisted-config repair. A single minor gap (mistral-small-latest absent from MISTRAL_SAFE_MAX_TOKENS_BY_MODEL) would produce a conservative-but-safe 8192 fallback instead of the catalog-correct 16384 in an edge case that has no known real-world exposure.
src/config/defaults.ts — the MISTRAL_SAFE_MAX_TOKENS_BY_MODEL map could be completed with mistral-small-latest: 16_384 to be fully exhaustive.

Prompt To Fix All With AI

This is a comment left during a code review.
Path: src/config/defaults.ts
Line: 88-89

Comment:
**`MISTRAL_SAFE_MAX_TOKENS_BY_MODEL` missing `mistral-small-latest`**

`mistral-small-latest` (contextWindow: 128000, catalog `maxTokens`: 16384) is absent from `MISTRAL_SAFE_MAX_TOKENS_BY_MODEL`. For a persisted config with `maxTokens: 128000` on this model the fallback would apply `DEFAULT_MODEL_MAX_TOKENS = 8192` instead of the catalog-appropriate `16384`.

Both values are safe (no 422), but a user whose `mistral-small-latest` entry was context-window-sized would end up with a more conservative budget than necessary after a doctor run. Adding the entry would make the repair map exhaustive for all models currently in the bundled catalog:

```suggestion
const MISTRAL_SAFE_MAX_TOKENS_BY_MODEL = {
  "devstral-medium-latest": 32_768,
  "magistral-small": 40_000,
  "mistral-large-latest": 16_384,
  "mistral-medium-2508": 8_192,
  "mistral-small-latest": 16_384,
  "pixtral-large-latest": 32_768,
} as const;
```

`codestral-latest` (catalog `maxTokens`: 4096, contextWindow: 256000) was never set context-window-sized in the old catalog, so it's fine to omit.

How can I resolve this? If you propose a fix, please make it concise.

_{Reviews (1): Last reviewed commit: "Merge branch 'main' into fix/53006-mistr..." | Re-trigger Greptile}

src/config/defaults.ts

aisle-research-bot · 2026-03-23T17:56:48Z

🔒 Aisle Security Analysis

We found 1 potential security issue(s) in this PR:

#	Severity	Title
1	🔵 Low	Terminal/log injection via unsanitized providerId in doctor migration change messages

1. 🔵 Terminal/log injection via unsanitized providerId in doctor migration change messages

Property	Value
Severity	Low
CWE	CWE-117
Location	`src/commands/doctor-legacy-config.ts:860-862`

Description

The doctor legacy-config migration normalizeLegacyMistralModelMaxTokens() constructs human-readable change messages by interpolating the raw config key providerId directly into a string that is later printed to the terminal via note(normalized.changes.join("\n"), ...).

providerId originates from Object.entries(next.models.providers), i.e., user-controlled configuration keys.
note() wraps text but does not escape/strip ANSI escape sequences or control characters (it preserves \n and other control chars).
An attacker who can influence the config file (shared workstation, supply-chain config snippet, malicious PR, etc.) could include newlines and/or ANSI escape sequences in providerId to spoof CLI output, hide warnings, or corrupt logs/CI output.

Vulnerable code:

changes.push(
  `Normalized models.providers.${providerId}.models[${index}].maxTokens ...`,
);

Recommendation

Neutralize untrusted values before embedding them in terminal/log output.

Options:

Sanitize providerId for terminal output (strip control chars / ANSI):

import { sanitizeForLog } from "../terminal/ansi.js";

const safeProviderId = sanitizeForLog(providerId);
changes.push(
  `Normalized models.providers.${safeProviderId}.models[${index}].maxTokens (${maxTokens} → ${normalizedMaxTokens}) ...`,
);

Render untrusted tokens with JSON string escaping to make control characters visible:

const safeProviderId = JSON.stringify(providerId).slice(1, -1);

Also consider applying similar sanitization inside note() (defense-in-depth) so all notes are protected by default.

Analyzed PR: #53054 at commit 1e9bfc8

_{Last updated on: 2026-03-23T18:31:32Z}

chatgpt-codex-connector

💡 Codex Review

Here are some automated review suggestions for this pull request.

Reviewed commit: acf577ccc3

ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

Open a pull request for review
Mark a draft as ready
Comment "@codex review".

If Codex has suggestions, it will comment; otherwise it will react with 👍.

Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".

chatgpt-codex-connector · 2026-03-23T17:58:36Z

src/config/defaults.ts

+    MISTRAL_SAFE_MAX_TOKENS_BY_MODEL[
+      params.modelId as keyof typeof MISTRAL_SAFE_MAX_TOKENS_BY_MODEL
+    ] ?? DEFAULT_MODEL_MAX_TOKENS;


Preserve unknown Mistral max-token overrides

resolveNormalizedProviderModelMaxTokens now rewrites any Mistral model where maxTokens equals contextWindow, but for model IDs not listed in MISTRAL_SAFE_MAX_TOKENS_BY_MODEL it falls back to DEFAULT_MODEL_MAX_TOKENS (8192). That causes custom or newly added Mistral models to be silently reduced from user-configured values during applyModelDefaults (and therefore openclaw doctor --fix), which is a regression from prior clamp-only behavior and can unexpectedly truncate outputs. Consider only rewriting known stale IDs, or returning clamped when the model ID is not in the safe map.

Useful? React with 👍 / 👎.

…w#53054) * fix(mistral): repair max-token defaults and doctor migration * fix(mistral): add missing small-model repair cap

@frankekn

…claw#105) * fix(web-search): mark DuckDuckGo experimental * docs(tools): update DuckDuckGo Search for landed plugin code - Mark as experimental (not just unofficial) - Add region and safeSearch tool parameters (from DDG schema) - Add plugin config example for region/safeSearch defaults - Document auto-detection order (100 = last) - Note SafeSearch defaults to moderate - Verified against extensions/duckduckgo/src/ * fix(agents): deny local MEDIA paths for MCP results * Usage: include reset and deleted session archives (openclaw#43215) Merged via squash. Prepared head SHA: 49ed6c2 Co-authored-by: rcrick <[email protected]> Co-authored-by: frankekn <[email protected]> Reviewed-by: @frankekn * docs(tools): soften DDG wording (scrapes -> pulls/gathers) * fix(build): add stable memory-cli dist entry (openclaw#51759) Co-authored-by: oliviareid-svg <[email protected]> Co-authored-by: Frank <[email protected]> * refactor!: drop legacy CLAWDBOT env compatibility * refactor!: remove moltbot state-dir migration fallback * fix(gateway): preserve async hook ingress provenance * fix(ci): write dist build stamp after builds * perf: trim vitest hot imports and refresh manifests * fix(security): unwrap time dispatch wrappers * fix(plugin-sdk): fall back to src root alias files * fix(ci): skip docs-only preflight pnpm audit * docs(changelog): note time exec approval fix * docs: refresh plugin-sdk api baseline * fix(runtime): make dist-runtime staging idempotent * fix(media): bound remote error-body snippet reads * fix(gateway): gate internal command persistence mutations * fix: restrict remote marketplace plugin sources * fix(runtime): skip peer resolution for bundled plugin deps * docs(agents): prefer current test model examples * fix(exec): escape invisible approval filler chars * test(models): refresh example model fixtures * fix(security): unify dispatch wrapper approval hardening * fix(security): harden explicit-proxy SSRF pinning * fix: gate synology chat reply name matching * docs: clarify sessions_spawn ACP vs subagent policies * refactor(exec): split wrapper resolution modules * refactor(exec): make dispatch wrapper semantics spec-driven * refactor(exec): share wrapper trust planning * refactor(exec): rename wrapper plans for trust semantics * fix: include .npmrc in onboard docker build * test: trim docker live auth mounts * Docs: refresh config baseline for Synology Chat * refactor: clarify synology delivery identity names * refactor: centralize synology dangerous name matching * refactor: narrow synology legacy name lookup * refactor: audit synology dangerous name matching * refactor: dedupe synology config schema * fix: normalize scoped vitest filter paths * fix(voice-call): harden webhook pre-auth guards * fix(synology-chat): fail closed shared webhook paths * docs: credit nexrin in synology changelog * test: fix base vitest thread regressions * test: finish base vitest thread fixture fixes * test(voice-call): accept oversize webhook socket resets * test: honor env auth in gateway live probes * fix: harden plugin docker e2e * Docs: align MiniMax examples with M2.7 * fix(ci): restore stale guardrails and baselines * Test: isolate qr dashboard integration suite * Gateway: resolve fallback plugin context lazily * fix: bind bootstrap setup codes to node profile * fix(tlon): unify settings reconciliation semantics * refactor(synology-chat): type startup webhook path policy * docs(synology-chat): clarify multi-account webhook paths * refactor: unify minimax model and failover live policies * docs: sync minimax m2.7 references * fix: harden Windows Parallels smoke installs * docs: reorder unreleased changelog by user impact * refactor: remove embedded runner cwd mutation * Infra: support shell carrier allow-always approvals * refactor: centralize bootstrap profile handling * refactor: reuse canonical setup bootstrap profile * fix(plugin-sdk): resolve hashed diagnostic events chunks * fix(plugin-sdk): normalize hashed diagnostic event exports * test: fix ci env-sensitive assertions * fix(gateway): fail closed on unresolved discovery endpoints * feat: add slash plugin installs * fix(media): block remote-host file URLs in loaders * fix(media): harden secondary local path seams * test: harden no-isolate reply teardown * docs(changelog): add Windows media security fix * refactor(gateway): centralize discovery target handling * test: narrow live transcript scaffolding strip * test: fix ci docs drift and bun qr exit handling * fix(browser): enforce node browser proxy allowProfiles * refactor(media): share local file access guards * test: stabilize ci test harnesses * test: harden no-isolate test module resets * fix(plugins): preserve live hook registry during gateway runs * test: fix channel summary registry setup * test: harden isolated test mocks * chore(plugins): remove opik investigation checkpoints * ACPX: align pinned runtime version (openclaw#52730) * ACPX: align pinned runtime version * ACPX: drop version example from help text * test: stop leaking image workspace temp dirs * fix(android): gate canvas bridge to trusted pages (openclaw#52722) * fix(android): gate canvas bridge to trusted pages * fix(changelog): note android canvas bridge gating * Update apps/android/app/src/main/java/ai/openclaw/app/node/CanvasActionTrust.kt Co-authored-by: greptile-apps[bot] <165735046+greptile-apps[bot]@users.noreply.github.com> * fix(android): snapshot canvas URL on UI thread --------- Co-authored-by: greptile-apps[bot] <165735046+greptile-apps[bot]@users.noreply.github.com> * test: isolate base vitest thread blockers * fix: sync agent and autoreply e2e updates * test: harden no-isolate mocked module resets * docs: reorder unreleased changelog * fix(changelog): note windows media path guardrails (openclaw#52738) * fix: alphabetize web search provider listings * docs: clarify unreleased breaking changes * test: harden ci isolated mocks * fix: align websocket stream fallback types * test: finish no-isolate suite hardening * style: format image-generation runtime tests * fix(memory-core): register memory tools independently to prevent coupled failure (openclaw#52668) Merged via admin squash because current required CI failures are inherited from base and match latest `main` failures outside this PR's `memory-core` surface. Prepared head SHA: df7f968 Co-authored-by: artwalker <[email protected]> Reviewed-by: @frankekn * fix(status): recompute fallback context window (openclaw#51795) * fix(status): recompute fallback context window * fix(status): keep live context token caps on fallback * fix(status): preserve fallback runtime context windows * fix(status): preserve configured fallback context caps * fix(status): keep provider-aware transcript context lookups * fix(status): preserve explicit fallback context caps * fix(status): clamp fallback configured context caps * fix(status): keep raw runtime slash ids * fix(status): refresh plugin-sdk api baseline * fix(status): preserve fallback context lookup * test(status): refresh plugin-sdk api baseline * fix(status): keep runtime slash-id context lookup --------- Co-authored-by: create <[email protected]> Co-authored-by: Frank Yang <[email protected]> Co-authored-by: RichardCao <[email protected]> * fix(telegram): make buttons schema optional in message tool The Telegram plugin injects a `buttons` property into the message tool schema via `createMessageToolButtonsSchema()`, but without wrapping it in `Type.Optional()`. This causes TypeBox to include `buttons` in the JSON Schema `required` array. In isolated sessions (e.g. cron jobs) where no `currentChannel` is set, all plugin schemas are merged into the message tool. When the LLM calls the message tool without a `buttons` parameter, AJV validation fails with: `buttons: must have required property 'buttons'`. Wrap the buttons schema in `Type.Optional()` so it is not required. * fix: keep message-tool buttons optional for Telegram and Mattermost (openclaw#52589) (thanks @tylerliu612) * test: update codex test fixtures to gpt-5.4 * fix: repair runtime seams after rebase * fix: restore Telegram topic announce delivery (openclaw#51688) (thanks @mvanhorn) When `replyLike.text` or `replyLike.caption` is an unexpected non-string value (edge case from some Telegram API responses), the reply body was coerced to "[object Object]" via string concatenation. Add a `typeof === "string"` guard to gracefully fall back to empty string, matching the existing pattern used for `quoteText` in the same function. Co-authored-by: Penchan <[email protected]> * docs: sync generated release baselines * test: isolate pi embedded model thread fixtures * fix: restore provider runtime lazy boundary * fix: preserve Telegram reply context text (openclaw#50500) (thanks @p3nchan) * fix: guard Telegram reply context text (openclaw#50500) (thanks @p3nchan) * fix: preserve Telegram reply caption fallback (openclaw#50500) (thanks @p3nchan) --------- Co-authored-by: Ayaan Zaidi <[email protected]> * fix: harden gateway SIGTERM shutdown (openclaw#51242) (thanks @juliabush) * fix: increase shutdown timeout to avoid SIGTERM hang * fix(telegram): abort polling fetch on shutdown to prevent SIGTERM hang * fix(gateway): enforce hard exit on shutdown timeout for SIGTERM * fix: tighten gateway shutdown watchdog * fix: harden gateway SIGTERM shutdown (openclaw#51242) (thanks @juliabush) --------- Co-authored-by: Ayaan Zaidi <[email protected]> * build: prepare 2026.3.22-beta.1 * fix: restore provider runtime lazy boundary * test: add parallels npm update smoke * test: split pi embedded model thread fixtures * fix: stop browser server tests from launching real chrome * test: stabilize live provider docker probes * fix: restart windows gateway after npm update * test: isolate server-context browser harness imports * test: inject model runtime hooks for thread-safe tests * test: snapshot ci timeout investigation * test: target gemini 3.1 flash alias * test: stabilize trigger handling and hook e2e tests * build: prepare 2026.3.22 * test: harden channel suite isolation * test: inject thread-safe deps for agent tools * test: raise timeout for slow provider auth normalization * ci: stabilize windows and bun unit lanes * test: inject thread-safe gateway and ACP seams * test: isolate pi model and reset-model thread fixtures * build: prepare 2026.3.23 * test: inject image-tool provider deps for raw threads * test: stabilize e2e module isolation * test: decouple vitest config checks from ambient env * fix: harden parallels smoke agent invocation * test: avoid repo-root perf profile artifacts * test: inject thread-safe base seams * fix: document Telegram asDocument alias (openclaw#52461) (thanks @bakhtiersizhaev) * feat(telegram): add asDocument param to message tool Adds `asDocument` as a user-facing alias for the existing `forceDocument` parameter in the message tool. When set to `true`, media files (images, videos, GIFs) are sent via `sendDocument` instead of `sendPhoto`/ `sendVideo`/`sendAnimation`, preserving the original file quality without Telegram compression. This is useful when agents need to deliver high-resolution images or uncompressed files to users via Telegram. `asDocument` is intentionally an alias rather than a replacement — the existing `forceDocument` continues to work unchanged. Changes: - src/agents/tools/message-tool.ts: add asDocument to send schema - src/agents/tools/telegram-actions.ts: OR asDocument into forceDocument - src/infra/outbound/message-action-runner.ts: same OR logic for outbound path - extensions/telegram/src/channel-actions.ts: read and forward asDocument - src/channels/plugins/actions/actions.test.ts: add test case * fix: restore channel-actions.ts to main version (rebase conflict fix) * fix(test): match asDocument test payload to actual params structure * fix(telegram): preserve forceDocument alias semantics * fix: document Telegram asDocument alias (openclaw#52461) (thanks @bakhtiersizhaev) --------- Co-authored-by: Бахтиер Сижаев <[email protected]> Co-authored-by: Ayaan Zaidi <[email protected]> * fix: refactor deepseek bundled plugin (openclaw#48762) (thanks @07akioni) * fix: declare typebox runtime dep for mattermost plugin * test: reset line webhook mocks between cases * test: split attempt spawn-workspace thread fixtures * test: remove replaced spawn-workspace monolith * refactor: isolate attempt context engine thread helpers * CI: remove npm release preview workflow (openclaw#52825) * CI: remove npm release preview workflow * Docs: align release maintainer skill with manual publish * Docs: expand release maintainer skill flow * test: stabilize gateway thread harness * test: fix status plugin pagination expectation * test: harden channel suite isolation * build: sync lockfile for mattermost plugin * fix: ensure env proxy dispatcher before MiniMax and OpenAI Codex OAuth flows (openclaw#52228) Verified: - pnpm install --frozen-lockfile - NPM_CONFIG_CACHE=/tmp/openclaw-npm-cache-52228 pnpm build - pnpm check - pnpm test:macmini (failed on inherited pre-existing plugin contract test: src/plugins/contracts/registry.contract.test.ts missing deepseek in bundled provider contract registry outside this PR surface) Co-authored-by: openperf <[email protected]> Co-authored-by: Tak Hoffman <[email protected]> * fix: restore ci gates * test: stabilize channel ci gate * docs: refresh generated config baseline * release: verify control-ui assets are included in npm tarball * release-check: include stderr/stdout when npm pack fails * release: add changelog for control UI tarball check * fix: keep session transcript pointers fresh after compaction (openclaw#50688) Co-authored-by: Frank Yang <[email protected]> * fix(msteams): isolate probe test env credentials * release: automate macOS publishing (openclaw#52853) * release: automate macOS publishing * release: keep mac appcast in openclaw repo * release: add preflight-only release workflow runs * release: keep appcast updates manual * release: generate signed appcast as workflow artifact * release: require preflight before publish * release: require mac app for every release * docs: clarify every release ships mac app * release: document Sparkle feed and SHA rules * release: keep publish flow tag-based * release: stabilize mac appcast flow * release: document local mac fallback * Update CHANGELOG.md * Improve PR template regression prompts * fix(agents): preserve anthropic thinking block order (openclaw#52961) * fix(release): ship bundled plugins in pack artifacts * fix(config): keep built-in channels out of plugin allowlists (openclaw#52964) * fix(config): keep built-in channels out of plugin allowlists * docs(changelog): note doctor whatsapp allowlist fix * docs(changelog): move doctor whatsapp fix to top * Update CHANGELOG.md * fix(config): keep built-in auto-enable idempotent * fix(release): preserve shipped channel surfaces in npm tar (openclaw#52913) * fix(channels): ship official channel catalog (openclaw#52838) * fix(release): keep shipped bundles in npm tar (openclaw#52838) * build(release): fix rebased release-check helpers (openclaw#52838) * fix(gateway): harden supervised lock and browser attach readiness * fix(matrix): avoid duplicate runtime api exports * fix(gateway): avoid probe false negatives after connect * docs(changelog): note release and matrix fixes * fix(plugins): unblock Discord/Slack message tool sends and Feishu media (openclaw#52991) * fix(plugins): unblock Discord and Slack message tool payloads * docs(changelog): note Discord Slack and Feishu message fixes * fix(channels): preserve external catalog overrides (openclaw#52988) * fix(channels): preserve external catalog overrides * fix(channels): clarify catalog precedence * fix(channels): respect overridden install specs * fix(gateway): require admin for agent session reset * fix(voice-call): stabilize plivo v2 replay keys * fix(gateway): require auth for canvas routes * fix(clawhub): resolve auth token for skill browsing (openclaw#53017) * fix(clawhub): resolve auth token for skill browsing * docs(changelog): note clawhub skill auth fix * fix(release): raise npm pack size budget * Tests: fix fresh-main regressions (openclaw#53011) * Tests: fix fresh-main regressions * Tests: avoid chat notice cache priming --------- Co-authored-by: Vincent Koc <[email protected]> * fix(config): ignore stale plugin allow entries * fix(browser): reuse running loopback browser after probe miss * fix(clawhub): honor macOS auth config path (openclaw#53034) * docs: fix nav ordering, missing pages, and stale model references - Sort providers alphabetically in docs.json nav - Sort channels alphabetically in docs.json nav (slack before synology-chat) - Add install/migrating-matrix to Maintenance nav section (was orphaned) - Remove zh-CN/plugins/architecture from nav (file does not exist) - Add Voice Call to channels index page - Add missing providers to providers index (DeepSeek, GitHub Copilot, OpenCode Go, Synthetic) - Sort providers index alphabetically - Update stale claude-3-5-sonnet model reference to claude-sonnet-4-6 in webhook docs * fix(clawhub): preserve XDG auth path on macOS * Agents: fix runtime web_search provider selection (openclaw#53020) Co-authored-by: Vincent Koc <[email protected]> * docs: fix CLI command tree, SDK import path, and tool group listing - Remove non-existent 'secrets migrate' from CLI command tree - Add actual secrets subcommands: audit, configure, apply - Add missing plugin subcommands: inspect, uninstall, update, marketplace list - Fix plugins info -> inspect (actual command name) - Add message send and broadcast subcommands to command tree - Remove misleading deprecated import from sdk-overview - Add sessions_yield and subagents to group:sessions tool group docs - Fix formatting * fix(gateway): guard openrouter auto pricing recursion (openclaw#53055) * test: refresh thread-safe agent fixtures * Release: fix npm release preflight under pnpm (openclaw#52985) Co-authored-by: Vincent Koc <[email protected]> * docs(changelog): add channel catalog override note (openclaw#52988) (openclaw#53059) * fix: harden update dev switch and refresh changelog * fix(mistral): repair max-token defaults and doctor migration (openclaw#53054) * fix(mistral): repair max-token defaults and doctor migration * fix(mistral): add missing small-model repair cap * fix(plugins): enable bundled Brave web search plugin by default (openclaw#52072) Brave is a bundled web search plugin but was missing from BUNDLED_ENABLED_BY_DEFAULT, causing it to be filtered out during provider resolution. This made web_search unavailable even when plugins.entries.brave.enabled was configured. Fixes openclaw#51937 Co-authored-by: Ubuntu <[email protected]> Co-authored-by: Vincent Koc <[email protected]> * fix(release): fail empty control ui tarballs * Revert "fix(plugins): enable bundled Brave web search plugin by default (openclaw#52072)" This reverts commit 0ea3c4d. * Telegram: preserve inbound debounce order * Telegram: fix fire-and-forget debounce order * fix(reply): refresh followup drain callbacks * Update CHANGELOG.md * fix(reply): preserve no-debounce inbound concurrency * fix(reply): clear idle followup callbacks * fix(inbound): bound tracked debounce keys * fix: preserve debounce and followup ordering (openclaw#52998) (thanks @osolmaz) * fix(discord): reply on native command auth failures (openclaw#53072) * docs(changelog): add missing recent fixes * fix: bound tracked debounce key accounting * fix packaged control ui asset lookup (openclaw#53081) * fix(cli): preserve posix default git dir * build: prepare 2026.3.23-beta.1 * test: harden canvas host undici isolation * docs(changelog): credit web search runtime fix * fix(openai-codex): bootstrap proxy on oauth refresh (openclaw#53078) Verified: - pnpm install --frozen-lockfile - pnpm exec vitest run extensions/openai/openai-codex-provider.runtime.test.ts extensions/openai/openai-provider.test.ts * release: harden preflight workflows (openclaw#53087) * release: harden preflight-only workflows * release: require main for publish runs * release: select xcode for macos workflow * release: retry flaky macos preflight steps * ci: shard bun test lane * Fix Control UI operator.read scope handling (openclaw#53110) Preserve Control UI scopes through the device-auth bypass path, normalize implied operator device-auth scopes, ignore cached under-scoped operator tokens, and degrade read-backed main pages gracefully when a connection truly lacks operator.read. Co-authored-by: Val Alexander <[email protected]> * build: prepare 2026.3.23 * fix(agents): prefer runtime snapshot for skill secrets * docs(changelog): note skill secretref runtime fix * fix(memory): bootstrap lancedb runtime on demand (openclaw#53111) Bootstrap LanceDB into plugin runtime state on first use for packaged/global installs, keep @lancedb/lancedb plugin-local, and add regression coverage for bundled, cached, retry, and Nix fail-fast runtime paths. Co-authored-by: Val Alexander <[email protected]> * build: finalize 2026.3.23 release * release: upload macos preflight artifacts (openclaw#53105) * release: upload macos preflight artifacts * release: speed up macos preflight * release: use xlarge macos runner * release: skip dmg path in macos preflight * fix(subagents): recheck timed-out announce waits (openclaw#53127) Recheck timed-out subagent announce waits against the latest runtime snapshot before announcing timeout, and keep that recheck best-effort so transient gateway failures do not suppress the announcement. Co-authored-by: Val Alexander <[email protected]> * docs(feishu): replace botName with name in config examples (openclaw#52753) Merged via squash. Prepared head SHA: 5237726 Co-authored-by: haroldfabla2-hue <[email protected]> Co-authored-by: altaywtf <[email protected]> Reviewed-by: @altaywtf * fix(plugins): accept clawhub uninstall specs * test(auth): align device scope expectations (openclaw#53151) * fix: prevent delivery-mirror re-delivery and raise Slack chunk limit (openclaw#45489) Merged via squash. Prepared head SHA: c7664c7 Co-authored-by: theo674 <[email protected]> Co-authored-by: altaywtf <[email protected]> Reviewed-by: @altaywtf * Infra: tighten shell-wrapper positional-argv allowlist matching (openclaw#53133) * Infra: tighten shell carrier allowlist matching * fix(security): tighten shell carrier allowlist matcher * fix: generalize api_error detection for fallback model triggering (openclaw#49611) Co-authored-by: Ayush Ojha <[email protected]> Co-authored-by: altaywtf <[email protected]> * feat(modelstudio): add standard (pay-as-you-go) DashScope endpoints for Qwen (openclaw#43878) Add Standard API Key auth methods for China (dashscope.aliyuncs.com) and Global/Intl (dashscope-intl.aliyuncs.com) pay-as-you-go endpoints alongside the existing Coding Plan (subscription) endpoints. Also updates group label to 'Qwen (Alibaba Cloud Model Studio)' and fixes glm-4.7 -> glm-5 in Coding Plan note messages. Co-authored-by: wenmeng zhou <[email protected]> * Release: privatize macOS publish flow (openclaw#53166) * fix(diagnostics): redact credentials from cache-trace diagnostic output Refs openclaw#53103 * Release: document manual macOS asset upload (openclaw#53178) * Release: document manual macOS asset upload * Release: document macOS smoke-test mode * docs(changelog): reorder release highlights * test(whatsapp): stabilize login coverage in shared workers * test(whatsapp): preserve session exports in login coverage * test(whatsapp): preserve media test module exports * test(whatsapp): preserve harness session exports * fix(ci): stabilize whatsapp extension checks * test: make update-cli checkout path assertion platform-safe * fix(auth): prevent stale auth store reverts (openclaw#53211) * Doctor: prune stale plugin allowlist and entry refs (openclaw#53187) Signed-off-by: sallyom <[email protected]> * test: stabilize test isolation * test: update command coverage * test: expand gemini live transcript stripping * test: fix update-cli default path assertion * chore(sre:PLA-920): adopt upstream sync changes * fix(sre:PLA-920): align branch with adopted upstream tree * build(sre:PLA-920): refresh dist artifacts * test(sre:PLA-920): align incident-format expectations --------- Signed-off-by: sallyom <[email protected]> Co-authored-by: Vincent Koc <[email protected]> Co-authored-by: Peter Steinberger <[email protected]> Co-authored-by: Rick_Xu <[email protected]> Co-authored-by: rcrick <[email protected]> Co-authored-by: frankekn <[email protected]> Co-authored-by: oliviareid-svg <[email protected]> Co-authored-by: oliviareid-svg <[email protected]> Co-authored-by: Frank <[email protected]> Co-authored-by: scoootscooob <[email protected]> Co-authored-by: ruochen <[email protected]> Co-authored-by: Onur Solmaz <[email protected]> Co-authored-by: greptile-apps[bot] <165735046+greptile-apps[bot]@users.noreply.github.com> Co-authored-by: Frank Yang <[email protected]> Co-authored-by: artwalker <[email protected]> Co-authored-by: RichardCao <[email protected]> Co-authored-by: create <[email protected]> Co-authored-by: RichardCao <[email protected]> Co-authored-by: liuyang <[email protected]> Co-authored-by: Ayaan Zaidi <[email protected]> Co-authored-by: Matt Van Horn <[email protected]> Co-authored-by: Penchan <[email protected]> Co-authored-by: Penchan <[email protected]> Co-authored-by: Julia Bush <[email protected]> Co-authored-by: Bakhtier Sizhaev <[email protected]> Co-authored-by: Бахтиер Сижаев <[email protected]> Co-authored-by: wangchunyue <[email protected]> Co-authored-by: Tak Hoffman <[email protected]> Co-authored-by: evann <[email protected]> Co-authored-by: Robin Waslander <[email protected]> Co-authored-by: Sathvik Veerapaneni <[email protected]> Co-authored-by: Nimrod Gutman <[email protected]> Co-authored-by: Luke <[email protected]> Co-authored-by: scoootscooob <[email protected]> Co-authored-by: Jamil Zakirov <[email protected]> Co-authored-by: TheRipper <[email protected]> Co-authored-by: Quinn H. <[email protected]> Co-authored-by: Ubuntu <[email protected]> Co-authored-by: Val Alexander <[email protected]> Co-authored-by: betoblair <[email protected]> Co-authored-by: haroldfabla2-hue <[email protected]> Co-authored-by: altaywtf <[email protected]> Co-authored-by: Altay <[email protected]> Co-authored-by: theo674 <[email protected]> Co-authored-by: theo674 <[email protected]> Co-authored-by: Ayush Ojha <[email protected]> Co-authored-by: Ayush Ojha <[email protected]> Co-authored-by: George Zhang <[email protected]> Co-authored-by: wenmeng zhou <[email protected]> Co-authored-by: Onur <[email protected]> Co-authored-by: Sally O'Malley <[email protected]>

…w#53054) * fix(mistral): repair max-token defaults and doctor migration * fix(mistral): add missing small-model repair cap (cherry picked from commit dd586d5)

vincentkoc self-assigned this Mar 23, 2026

openclaw-barnacle bot added commands Command implementations size: M maintainer Maintainer-authored PR labels Mar 23, 2026

vincentkoc marked this pull request as ready for review March 23, 2026 17:48

greptile-apps bot reviewed Mar 23, 2026

View reviewed changes

src/config/defaults.ts Show resolved Hide resolved

vincentkoc added 2 commits March 23, 2026 10:57

fix(mistral): repair max-token defaults and doctor migration

1eadc11

fix(mistral): add missing small-model repair cap

1e9bfc8

vincentkoc force-pushed the fix/53006-mistral-max-tokens branch from 968e7ac to 1e9bfc8 Compare March 23, 2026 17:57

vincentkoc merged commit dd586d5 into main Mar 23, 2026
7 checks passed

vincentkoc deleted the fix/53006-mistral-max-tokens branch March 23, 2026 17:57

chatgpt-codex-connector bot reviewed Mar 23, 2026

View reviewed changes

github-actions bot mentioned this pull request Mar 23, 2026

📡 Upstream Digest — 2026-03-23 18:42 UTC curtismercier/openclaw-mods#344

Open

iclem pushed a commit to iclem/openclaw that referenced this pull request Mar 23, 2026

fix(mistral): repair max-token defaults and doctor migration (opencla…

0a180f4

…w#53054) * fix(mistral): repair max-token defaults and doctor migration * fix(mistral): add missing small-model repair cap

steipete mentioned this pull request Mar 24, 2026

[Bug]: Mistral AI STILL fails with error 422 status code: (no body) #53363

Closed

Arry8 pushed a commit to Arry8/openclaw that referenced this pull request Mar 25, 2026

fix(mistral): repair max-token defaults and doctor migration (opencla…

2fc0dc9

…w#53054) * fix(mistral): repair max-token defaults and doctor migration * fix(mistral): add missing small-model repair cap

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

Uh oh!

fix(mistral): repair max-token defaults and doctor migration#53054

fix(mistral): repair max-token defaults and doctor migration#53054
vincentkoc merged 2 commits intomainfrom
fix/53006-mistral-max-tokens

vincentkoc commented Mar 23, 2026

Uh oh!

greptile-apps bot commented Mar 23, 2026

Uh oh!

Uh oh!

aisle-research-bot bot commented Mar 23, 2026 •

edited

Loading

Uh oh!

Uh oh!

chatgpt-codex-connector bot left a comment

Uh oh!

chatgpt-codex-connector bot Mar 23, 2026

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

1 participant

Uh oh!

Conversation

vincentkoc commented Mar 23, 2026

Summary

Change Type (select all)

Scope (select all touched areas)

Linked Issue/PR

Root Cause / Regression History (if applicable)

Regression Test Plan (if applicable)

User-visible / Behavior Changes

Security Impact (required)

Repro + Verification

Environment

Steps

Expected

Actual

Evidence

Human Verification (required)

Review Conversations

Compatibility / Migration

Failure Recovery (if this breaks)

Risks and Mitigations

Uh oh!

greptile-apps bot commented Mar 23, 2026

Greptile Summary

Confidence Score: 4/5

Uh oh!

Uh oh!

aisle-research-bot bot commented Mar 23, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

🔒 Aisle Security Analysis

1. 🔵 Terminal/log injection via unsanitized providerId in doctor migration change messages

Description

Recommendation

Uh oh!

Uh oh!

chatgpt-codex-connector bot left a comment

Choose a reason for hiding this comment

💡 Codex Review

Uh oh!

chatgpt-codex-connector bot Mar 23, 2026

Choose a reason for hiding this comment

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

1 participant

aisle-research-bot bot commented Mar 23, 2026 •

edited

Loading