fix(models): preserve stream usage compat opt-ins#45733
fix(models): preserve stream usage compat opt-ins#45733ademczuk merged 1 commit intoopenclaw:mainfrom
Conversation
Greptile SummaryThis PR fixes a Key changes:
Confidence Score: 4/5
Prompt To Fix All With AIThis is a comment left during a code review.
Path: src/agents/model-compat.ts
Line: 69-73
Comment:
**`supportsDeveloperRole: true` opt-in silently discarded**
The old code preserved an explicit `supportsDeveloperRole: true` from user/catalog config (the now-renamed test "respects explicit supportsDeveloperRole true…" asserted `.toBe(true)`). The new `needsDeveloperRoleOverride` condition is `compat?.supportsDeveloperRole !== false`, which is `true` whenever the flag is `true` *or* `undefined`, so an explicit `supportsDeveloperRole: true` in a user's model definition is now silently overridden to `false`.
This is an asymmetric treatment compared to `supportsUsageInStreaming`: streaming usage preserves both explicit `true` *and* explicit `false`, but developer-role only preserves explicit `false`. Any user who previously opted in with `compat: { supportsDeveloperRole: true }` will lose that capability without warning after this change.
If this is intentional (the comment says "force it off for non-native endpoints"), consider adding an explanatory note in the config schema docs or a console warning so users aren't surprised, and make sure the changelog/migration notes call this out explicitly.
How can I resolve this? If you propose a fix, please make it concise.Last reviewed commit: 7b50052 |
![greptile-apps[bot]](https://avatars.githubusercontent.com/in/867647?s=60&v=4){kind=small}
![chatgpt-codex-connector[bot]](https://avatars.githubusercontent.com/in/1144995?s=60&v=4){kind=small}
There was a problem hiding this comment.
💡 Codex Review
Here are some automated review suggestions for this pull request.
Reviewed commit: 7b500520f3
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".
|
Addressed the review feedback in
Validation rerun after the update:
|
|
Addressed the review feedback in 85817c8.\n\n- restored the existing behavior of preserving explicit \ overrides on non-native \ endpoints, so this PR stays focused on streaming-usage compat\n- moved built-in streaming-usage opt-ins out of the static provider catalogs and into a provider-side helper that only enables them for known native Moonshot / Model Studio baseUrls\n- added regression coverage for native-vs-custom baseUrl behavior so custom OpenAI-compatible proxies stay on the conservative default\n\nValidation rerun after the update:\n- ✓ src/agents/models-config.providers.moonshot.test.ts > moonshot implicit provider (#33637) > uses explicit CN baseUrl when provided 25ms Test Files 3 passed (3)
OK apps/macos/Sources/OpenClaw/HostEnvSecurityPolicy.generated.swift
Checking formatting... All matched files use the correct format.
Found 0 warnings and 0 errors.
OK: bundled plugin source files use scoped plugin-sdk subpaths (790 checked).
A2UI bundle up to date; skipping.
[copy-hook-metadata] Copied 4 hook metadata files. |
There was a problem hiding this comment.
💡 Codex Review
Here are some automated review suggestions for this pull request.
Reviewed commit: 85817c864f
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".
|
Reviewed the diff and the approach is solid. The split between preserving explicit compat overrides in model-compat.ts and applying per-provider native URL opt-ins via applyNativeStreamingUsageCompat is clean. One blocker: there's a rebase conflict in model-compat.ts. Upstream added supportsStrictMode as a third compat flag since your last push, which changes the early-return condition and the compat object shape. The conflict touches the same lines your PR modifies (the forcedDeveloperRole/needsDeveloperRoleOverride logic and the compat spread). Could you rebase onto current main and integrate the supportsStrictMode handling? The upstream changes are in the same function (normalizeModelCompat) around lines 65-87. The strictMode flag follows the same pattern - force it off for non-native endpoints unless explicitly set. Happy to re-review once rebased. |
ademczuk
left a comment
ademczuk
left a comment
There was a problem hiding this comment.
Approach looks correct. The model-compat.ts change properly preserves explicit streaming usage overrides (both true and false) without changing behavior for unset cases. The applyNativeStreamingUsageCompat function is scoped to known-good native URLs for moonshot and modelstudio only, applied at the right stage in the plan pipeline.
Approving the approach. Needs rebase to resolve the supportsStrictMode conflict before merge.
|
Hey @pezy - this needs a rebase after #45497 landed (adds The conflict is in
I pushed a rebased version to Happy to help get this landed - the approach is solid and I'd like to close out the competing PRs in this area. |
pezy
force-pushed
the
codex/preserve-stream-usage-compat-upstream
branch
from
1e6bcb9 to
41fd398
Compare
|
Thanks again @ademczuk — your rebase notes and reference branch were very helpful. I wanted to follow up on how I handled this. I did not directly cherry-pick
I treated your branch as a reference implementation for the intended resolution, while reapplying the PR on top of a newer I also reran:
Really appreciate you pushing the rebased branch and spelling out the three conflict points — that made it much easier to land this cleanly. |
pezy
force-pushed
the
codex/preserve-stream-usage-compat-upstream
branch
from
ed225f1 to
0994eab
Compare
Preserves explicit `supportsUsageInStreaming` overrides from built-in provider catalogs and user config instead of unconditionally forcing `false` on non-native openai-completions endpoints. Adds `applyNativeStreamingUsageCompat()` to set `supportsUsageInStreaming: true` on ModelStudio (DashScope) and Moonshot models at config build time so their native streaming usage works out of the box. Closes openclaw#46142 Co-authored-by: pezy <[email protected]>
* feat: make compaction timeout configurable via agents.defaults.compaction.timeoutSeconds (openclaw#46889) * feat: make compaction timeout configurable via agents.defaults.compaction.timeoutSeconds The hardcoded 5-minute (300s) compaction timeout causes large sessions to enter a death spiral where compaction repeatedly fails and the session grows indefinitely. This adds agents.defaults.compaction.timeoutSeconds to allow operators to override the compaction safety timeout. Default raised to 900s (15min) which is sufficient for sessions up to ~400k tokens. The resolved timeout is also used for the session write lock duration so locks don't expire before compaction completes. Fixes openclaw#38233 Co-Authored-By: Claude Opus 4.6 (1M context) <[email protected]> * test: add resolveCompactionTimeoutMs tests Cover config resolution edge cases: undefined config, missing compaction section, valid seconds, fractional values, zero, negative, NaN, and Infinity. Co-Authored-By: Claude Opus 4.6 (1M context) <[email protected]> * fix: add timeoutSeconds to compaction Zod schema The compaction object schema uses .strict(), so setting the new timeoutSeconds config option would fail validation at startup. Co-Authored-By: Claude Opus 4.6 (1M context) <[email protected]> * fix: enforce integer constraint on compaction timeoutSeconds schema Prevents sub-second values like 0.5 which would floor to 0ms and cause immediate compaction timeout. Matches pattern of other integer timeout fields in the schema. Co-Authored-By: Claude Opus 4.6 (1M context) <[email protected]> * fix: clamp compaction timeout to Node timer-safe maximum Values above ~2.1B ms overflow Node's setTimeout to 1ms, causing immediate timeout. Clamp to MAX_SAFE_TIMEOUT_MS matching the pattern in agents/timeout.ts. Co-Authored-By: Claude Opus 4.6 (1M context) <[email protected]> * fix: add FIELD_LABELS entry for compaction timeoutSeconds Maintains label/help parity invariant enforced by schema.help.quality.test.ts. Co-Authored-By: Claude Opus 4.6 (1M context) <[email protected]> * fix: align compaction timeouts with abort handling * fix: land compaction timeout handling (openclaw#46889) (thanks @asyncjason) --------- Co-authored-by: Jason Separovic <[email protected]> Co-authored-by: Claude Opus 4.6 (1M context) <[email protected]> Co-authored-by: Ayaan Zaidi <[email protected]> * fix: harden compaction timeout follow-ups * Docs: fix stale Clawdbot branding in agent workflow file (openclaw#46963) Co-authored-by: webdevpraveen <[email protected]> * docs: replace outdated Clawdbot references with OpenClaw in skill docs (openclaw#41563) Update 5 references to the old "Clawdbot" name in skills/apple-reminders/SKILL.md and skills/imsg/SKILL.md. Co-authored-by: imanisynapse <[email protected]> * Docs: switch README logo to SVG assets (openclaw#47049) * fix: Disable strict mode tools for non-native openai-completions compatible APIs (openclaw#45497) Merged via squash. Prepared head SHA: 20fe05f Co-authored-by: sahancava <[email protected]> Co-authored-by: frankekn <[email protected]> Reviewed-by: @frankekn * fix: forward forceDocument through sendPayload path (follow-up to openclaw#45111) (openclaw#47119) Merged via squash. Prepared head SHA: d791190 Co-authored-by: thepagent <[email protected]> Reviewed-by: @frankekn * fix(android): support android node `calllog.search` (openclaw#44073) * fix(android): support android node `calllog.search` * fix(android): support android node calllog.search * fix(android): wire callLog through shared surfaces * fix: land Android callLog support (openclaw#44073) (thanks @lxk7280) --------- Co-authored-by: lixuankai <[email protected]> Co-authored-by: Ayaan Zaidi <[email protected]> * fix(whatsapp): restore append recency filter lost in extensions refactor, handle Long timestamps (openclaw#42588) Merged via squash. Prepared head SHA: 8ce59bb Co-authored-by: MonkeyLeeT <[email protected]> Co-authored-by: scoootscooob <[email protected]> Reviewed-by: @scoootscooob * fix(web): handle 515 Stream Error during WhatsApp QR pairing (openclaw#27910) * fix(web): handle 515 Stream Error during WhatsApp QR pairing getStatusCode() never unwrapped the lastDisconnect wrapper object, so login.errorStatus was always undefined and the 515 restart path in restartLoginSocket was dead code. - Add err.error?.output?.statusCode fallback to getStatusCode() - Export waitForCredsSaveQueue() so callers can await pending creds - Await creds flush in restartLoginSocket before creating new socket Fixes openclaw#3942 * test: update session mock for getStatusCode unwrap + waitForCredsSaveQueue Mirror the getStatusCode fix (err.error?.output?.statusCode fallback) in the test mock and export waitForCredsSaveQueue so restartLoginSocket tests work correctly. * fix(web): scope creds save queue per-authDir to avoid cross-account blocking The credential save queue was a single global promise chain shared by all WhatsApp accounts. In multi-account setups, a slow save on one account blocked credential writes and 515 restart recovery for unrelated accounts. Replace the global queue with a per-authDir Map so each account's creds serialize independently. waitForCredsSaveQueue() now accepts an optional authDir to wait on a single account's queue, or waits on all when omitted. Co-Authored-By: Claude Opus 4.6 <[email protected]> * test: use real Baileys v7 error shape in 515 restart test The test was using { output: { statusCode: 515 } } which was already handled before the fix. Updated to use the actual Baileys v7 shape { error: { output: { statusCode: 515 } } } to cover the new fallback path in getStatusCode. Co-Authored-By: Claude Code (Opus 4.6) <[email protected]> * fix(web): bound credential-queue wait during 515 restart Prevents restartLoginSocket from blocking indefinitely if a queued saveCreds() promise stalls (e.g. hung filesystem write). Co-Authored-By: Claude <[email protected]> * fix: clear flush timeout handle and assert creds queue in test Co-Authored-By: Claude <[email protected]> * fix: evict settled credsSaveQueues entries to prevent unbounded growth Co-Authored-By: Claude <[email protected]> * fix: share WhatsApp 515 creds flush handling (openclaw#27910) (thanks @asyncjason) --------- Co-authored-by: Jason Separovic <[email protected]> Co-authored-by: Claude Opus 4.6 <[email protected]> Co-authored-by: Ayaan Zaidi <[email protected]> * Deduplicate repeated tool call IDs for OpenAI-compatible APIs (openclaw#40996) Merged via squash. Prepared head SHA: 38d8048 Co-authored-by: xaeon2026 <[email protected]> Co-authored-by: frankekn <[email protected]> Reviewed-by: @frankekn * fix(gateway): skip Control UI pairing when auth.mode=none (closes openclaw#42931) (openclaw#47148) When auth is completely disabled (mode=none), requiring device pairing for Control UI operator sessions adds friction without security value since any client can already connect without credentials. Add authMode parameter to shouldSkipControlUiPairing so the bypass fires only for Control UI + operator role + auth.mode=none. This avoids the openclaw#43478 regression where a top-level OR disabled pairing for ALL websocket clients. * fix: preserve Telegram word boundaries when rechunking HTML (openclaw#47274) * fix: preserve Telegram chunk word boundaries * fix: address Telegram chunking review feedback * fix: preserve Telegram retry separators * fix: preserve Telegram chunking boundaries (openclaw#47274) * test(whatsapp): fix stale append inbox expectation * chore(gateway): ignore `.test.ts` changes in `gateway:watch` (openclaw#36211) * fix: harden remote cdp probes * feat(feishu): add ACP and subagent session binding (openclaw#46819) * feat(feishu): add ACP session support * fix(feishu): preserve sender-scoped ACP rebinding * fix(feishu): recover sender scope from bound ACP sessions * fix(feishu): support DM ACP binding placement * feat(feishu): add current-conversation session binding * fix(feishu): avoid DM parent binding fallback * fix(feishu): require canonical topic sender ids * fix(feishu): honor sender-scoped ACP bindings * fix(feishu): allow user-id ACP DM bindings * fix(feishu): recover user-id ACP DM bindings * ACP: fail closed on conflicting tool identity hints (openclaw#46817) * ACP: fail closed on conflicting tool identity hints * ACP: restore rawInput fallback for safe tool resolution * ACP tests: cover rawInput-only safe tool approval * fix: harden mention pattern regex compilation * Nodes: recheck queued actions before delivery (openclaw#46815) * Nodes: recheck queued actions before delivery * Nodes tests: cover pull-time policy recheck * Nodes tests: type node policy mocks explicitly * refactor: drop deprecated whatsapp mention pattern sdk helper * added a fix for memory leak on 2gb ram (openclaw#46522) * Nodes tests: prove pull-time policy revalidation * fix: harden device token rotation denial paths * style: format imported model helpers * Plugins: preserve scoped ids and reserve bundled duplicates (openclaw#47413) * Plugins: preserve scoped ids and reserve bundled duplicates * Changelog: add plugin scoped id note * Plugins: harden scoped install ids * Plugins: reserve scoped install dirs * Plugins: migrate legacy scoped update ids * CLI: reduce channels add startup memory (openclaw#46784) * CLI: lazy-load channel subcommand handlers * Channels: defer add command dependencies * CLI: skip status JSON plugin preload * CLI: cover status JSON route preload * Status: trim JSON security audit path * Status: update JSON fast-path tests * CLI: cover root help fast path * CLI: fast-path root help * Status: keep JSON security parity * Status: restore JSON security tests * CLI: document status plugin preload * Channels: reuse Telegram account import * Integrations: tighten inbound callback and allowlist checks (openclaw#46787) * Integrations: harden inbound callback and allowlist handling * Integrations: address review follow-ups * Update CHANGELOG.md * Mattermost: avoid command-gating open button callbacks * ACP: require admin scope for mutating internal actions (openclaw#46789) * ACP: require admin scope for mutating internal actions * ACP: cover operator admin mutating actions * ACP: gate internal status behind admin scope * Changelog: add missing PR credits * Changelog: add more unreleased PR numbers * Subagents: restrict follow-up messaging scope (openclaw#46801) * Subagents: restrict follow-up messaging scope * Subagents: cover foreign-session follow-up sends * Update CHANGELOG.md * Webhooks: tighten pre-auth body handling (openclaw#46802) * Webhooks: tighten pre-auth body handling * Webhooks: clean up request body guards * Tools: revalidate workspace-only patch targets (openclaw#46803) * Tools: revalidate workspace-only patch targets * Tests: narrow apply-patch delete-path assertion * CLI: trim onboarding provider startup imports (openclaw#47467) * Scope Control UI sessions per gateway (openclaw#47453) * Scope Control UI sessions per gateway Signed-off-by: sallyom <[email protected]> * Add changelog for Control UI session scoping Signed-off-by: sallyom <[email protected]> --------- Signed-off-by: sallyom <[email protected]> * Gateway: scrub credentials from endpoint snapshots (openclaw#46799) * Gateway: scrub credentials from endpoint snapshots * Gateway: scrub raw endpoint credentials in snapshots * Gateway: preserve config redaction round-trips * Gateway: restore redacted endpoint URLs on apply * fix(config): avoid failing startup on implicit memory slot (openclaw#47494) * fix(config): avoid failing on implicit memory slot * fix(config): satisfy build for memory slot guard * docs(changelog): note implicit memory slot startup fix (openclaw#47494) * CLI: lazy-load auth choice provider fallback (openclaw#47495) * CLI: lazy-load auth choice provider fallback * CLI: cover lazy auth choice provider fallback * fix(ci): config drift found and documented * Gateway: tighten forwarded client and pairing guards (openclaw#46800) * Gateway: tighten forwarded client and pairing guards * Gateway: make device approval scope checks atomic * Gateway: preserve device approval baseDir compatibility * Changelog: note CLI OOM startup fixes (openclaw#47525) * Commands: lazy-load model picker provider runtime (openclaw#47536) * Commands: lazy-load model picker provider runtime * Tests: cover model picker runtime boundary * docs: fork rebase spec + per-patch diffs for upstream v2026.3.13 merge Generated after failed merge attempt (2026-03-15). Contains: - FORK-PATCHES-SPEC.md: implementation instructions per patch group (249 lines) - FORK-REBASE-SPEC.md: technical context, errors, SSE protocol (292 lines) - fork-patches/by-patch/: 31 per-patch git diffs (consultable on demand) - fork-patches/fork-vs-upstream-src-only.patch: full squashed diff (5813 lines) Co-authored-by: Bob * docs: add merge plan from feat/upstream-merge-3.13 branch Co-authored-by: Bob * docs: remove old merge plan — superseded by FORK-PATCHES-SPEC + FORK-REBASE-SPEC Co-authored-by: Bob * chore(fmt): format changes and broken types * Commands: split static onboard auth choice help (openclaw#47545) * Commands: split static onboard auth choice help * Tests: cover static onboard auth choice help * Changelog: note static onboard auth choice help * CLI/completion: fix generator OOM and harden plugin registries (openclaw#45537) * fix: avoid OOM during completion script generation * CLI/completion: fix PowerShell nested command paths * CLI/completion: cover generated shell scripts * Changelog: note completion generator follow-up * Plugins: reserve shared registry names --------- Co-authored-by: Xiaoyi <[email protected]> Co-authored-by: Vincent Koc <[email protected]> * fix(plugins): load bundled extensions from dist (openclaw#47560) * fix(models): preserve stream usage compat opt-ins (openclaw#45733) Preserves explicit `supportsUsageInStreaming` overrides from built-in provider catalogs and user config instead of unconditionally forcing `false` on non-native openai-completions endpoints. Adds `applyNativeStreamingUsageCompat()` to set `supportsUsageInStreaming: true` on ModelStudio (DashScope) and Moonshot models at config build time so their native streaming usage works out of the box. Closes openclaw#46142 Co-authored-by: pezy <[email protected]> * Plugins: reserve context engine ownership * docs(zalo): document current Marketplace bot behavior (openclaw#47552) Verified: - pnpm check:docs Co-authored-by: Tomáš Dinh <[email protected]> Co-authored-by: Tak Hoffman <[email protected]> * Docs: move release runbook to maintainer repo (openclaw#47532) * Docs: redact private release setup * Docs: tighten release order * Docs: move release runbook to maintainer repo * Docs: delete public mac release page * Docs: remove zh-CN mac release page * Docs: turn release checklist into release policy * Docs: point release policy to private docs * Docs: regenerate zh-CN release policy pages * Docs: preserve Doctor in zh-CN hubs * Docs: fix zh-CN polls label * Docs: tighten docs i18n term guardrails * Docs: enforce zh-CN glossary coverage * Scripts: rebuild on extension and tsdown config changes (openclaw#47571) Merged via squash. Prepared head SHA: edd8ed8 Co-authored-by: gumadeiras <[email protected]> Co-authored-by: gumadeiras <[email protected]> Reviewed-by: @gumadeiras * fix: reset chat buffer on tool-start to prevent intermediary text accumulation The Pi SDK resets lastStreamedAssistantCleaned between tool calls, but the gateway chatRunState.buffers was not reset — causing mergedText to accumulate text from ALL prior turns. The SSE subscriber (which resets lastTextLen=0 on tool-start) then re-emitted the entire conversation. Co-authored-by: Bob * fix(release): block oversized npm packs that regress low-memory startup (openclaw#46850) * fix(release): guard npm pack size regressions * fix(release): fail closed when npm omits pack size * Plugins: reserve context engine ownership (openclaw#47595) * Plugins: reserve context engine ownership * Update src/context-engine/registry.ts Co-authored-by: greptile-apps[bot] <165735046+greptile-apps[bot]@users.noreply.github.com> --------- Co-authored-by: greptile-apps[bot] <165735046+greptile-apps[bot]@users.noreply.github.com> * fix: restore Patch #4 (chat mirror) and Patch #5 (inbound push) lost in Codex merge Patch #4: webchat-originated replies on WA-scoped sessions now mirror to WhatsApp via sendMessageWhatsApp(). The Codex merge kept the runContext.mirror registration but lost the delivery block. Patch #5: inbound messages (WA/Slack/etc.) now broadcast to WS/SSE clients via message.inbound event, restoring real-time cross-channel message display in the webchat dashboard. Co-authored-by: Bob * fix: mirror delivery in emitChatFinal (embedded runner path) The previous commit placed mirror only in the !agentRunStarted branch of server-methods/chat.ts, but the embedded runner sets agentRunStarted=true and delivers via emitChatFinal in server-chat.ts instead. This restores the mirror block in the correct location — matching the alpha. Co-authored-by: Bob * Gateway: sync runtime post-build artifacts * Plugins: harden context engine ownership * fix: complete Patch #5 inbound push + fix mirror static import - Add emitInboundMessageEvent() call in dispatch-from-config.ts (was only defined but never called — WA messages never reached SSE/webchat) - Switch mirror from dynamic import() to static import (dynamic import failed silently in bundled build) Co-authored-by: Bob * fix: globalThis singleton for WA listeners to survive chunk duplication The bundler splits active-listener.ts into a different chunk than server-chat.ts (mirror) and auto-reply/monitor.ts (listener registration). Static/dynamic imports resolve to different module instances, so mirror always sees an empty listeners Map. Using globalThis ensures all chunks share the same Map instance. Co-authored-by: Bob * fix(plugins): fix bundled plugin roots and skill assets (openclaw#47601) * fix(acpx): resolve bundled plugin root correctly * fix(plugins): copy bundled plugin skill assets * fix(plugins): tolerate missing bundled skill paths * chore: remove temporary mirror debug logs Co-authored-by: Bob * fix: globalThis singleton for inbound event listeners (chunk duplication) Same root cause as the WA listener fix: dispatch-from-config.ts emits inbound events in one chunk, server.impl.ts subscribes in another. Module-level Set gets duplicated across chunks. Co-authored-by: Bob * fix: emit inbound events from WA process-message path (not dispatch-from-config) dispatch-from-config.ts is NOT in the WA message processing chain. WA messages go through process-message.ts → provider-dispatcher.ts. Moved emitInboundMessageEvent to process-message.ts where WA messages are actually processed. Co-authored-by: Bob * fix(ci): restore config baseline release-check output (openclaw#47629) * Docs: regenerate config baseline * Chore: ignore generated config baseline * Update .prettierignore Co-authored-by: greptile-apps[bot] <165735046+greptile-apps[bot]@users.noreply.github.com> --------- Co-authored-by: greptile-apps[bot] <165735046+greptile-apps[bot]@users.noreply.github.com> * CLI: support package-manager installs from GitHub main (openclaw#47630) * CLI: resolve package-manager main install specs * CLI: skip registry resolution for raw package specs * CLI: support main package target updates * CLI: document package update specs in help * Tests: cover package install spec resolution * Tests: cover npm main-package updates * Tests: cover update --tag main * Installer: support main package targets * Installer: support main package targets on Windows * Docs: document package-manager main updates * Docs: document installer main targets * Docs: document npm and pnpm main installs * Docs: document update --tag main * Changelog: note package-manager main installs * Update src/infra/update-global.test.ts Co-authored-by: greptile-apps[bot] <165735046+greptile-apps[bot]@users.noreply.github.com> --------- Co-authored-by: greptile-apps[bot] <165735046+greptile-apps[bot]@users.noreply.github.com> * fix: emit message.inbound directly on gatewayEventBus Bypass inbound-events.ts entirely — its module-level Set suffers from chunk duplication even with globalThis (timing/ordering issues). gatewayEventBus already uses globalThis singleton and is proven to work for chat/agent events. SSE listens on gatewayEventBus for message.inbound. Co-authored-by: Bob * fix(dev): align gateway watch with tsdown wrapper (openclaw#47636) * Commands: lazy-load non-interactive plugin provider runtime (openclaw#47593) * Commands: lazy-load non-interactive plugin provider runtime * Tests: cover non-interactive plugin provider ordering * Update src/commands/onboard-non-interactive/local/auth-choice.plugin-providers.runtime.ts Co-authored-by: greptile-apps[bot] <165735046+greptile-apps[bot]@users.noreply.github.com> --------- Co-authored-by: greptile-apps[bot] <165735046+greptile-apps[bot]@users.noreply.github.com> * Plugins: relocate bundled skill assets * Plugins: skip nested node_modules in bundled skills * Plugins: clean stale bundled skill outputs * feat(plugins): move provider runtimes into bundled plugins * build(plugins): add bundled provider plugin manifests * Channels: move onboarding adapters into extensions * Channels: use owned helper imports * Plugins: broaden plugin surface for Codex App Server (openclaw#45318) * Plugins: add inbound claim and Telegram interaction seams * Plugins: add Discord interaction surface * Chore: fix formatting after plugin rebase * fix(hooks): preserve observers after inbound claim * test(hooks): cover claimed inbound observer delivery * fix(plugins): harden typing lease refreshes * fix(discord): pass real auth to plugin interactions * fix(plugins): remove raw session binding runtime exposure * fix(plugins): tighten interactive callback handling * Plugins: gate conversation binding with approvals * Plugins: migrate legacy plugin binding records * Plugins/phone-control: update test command context * Plugins: migrate legacy binding ids * Plugins: migrate legacy codex session bindings * Discord: fix plugin interaction handling * Discord: support direct plugin conversation binds * Plugins: preserve Discord command bind targets * Tests: fix plugin binding and interactive fallout * Discord: stabilize directory lookup tests * Discord: route bound DMs to plugins * Discord: restore plugin bindings after restart * Telegram: persist detached plugin bindings * Plugins: limit binding APIs to Telegram and Discord * Plugins: harden bound conversation routing * Plugins: fix extension target imports * Plugins: fix Telegram runtime extension imports * Plugins: format rebased binding handlers * Discord: bind group DM interactions by channel --------- Co-authored-by: Vincent Koc <[email protected]> * feat(plugins): add compatible bundle support * feat(plugins): move provider runtimes into bundled plugins * build(plugins): add bundled provider plugin packages * fix(plugins): restore provider compatibility fallbacks * Changelog: note plugin agent integrations * chore: remove inbound-push debug logs Co-authored-by: Bob * refactor: decouple channel setup discovery * refactor: move telegram onboarding to setup wizard * docs: describe channel setup wizard surface * fix: tighten setup wizard typing * fix: deduplicate inbound events + use raw body instead of envelope - Remove emitInboundMessageEvent from dispatch-from-config.ts (WA uses process-message.ts path, causing double emit) - Use params.msg.body (clean) instead of combinedBody (with envelope prefix) to avoid showing [WhatsApp ...] metadata in chat UI Co-authored-by: Bob * Commands: lazy-load auth choice plugin provider runtime (openclaw#47692) * Commands: lazy-load auth choice plugin provider runtime * Tests: cover auth choice plugin provider runtime * refactor: expand setup wizard flow * refactor: move discord and slack to setup wizard * refactor: drop onboarding adapter sdk exports * docs: update setup wizard capabilities * feat(plugins): test bundle MCP end to end * fix(onboarding): use scoped plugin snapshots to prevent OOM on low-memory hosts (openclaw#46763) * fix(onboarding): use scoped plugin snapshots to prevent OOM on low-memory hosts Onboarding and channel-add flows previously loaded the full plugin registry, which caused OOM crashes on memory-constrained hosts. This patch introduces scoped, non-activating plugin registry snapshots that load only the selected channel plugin without replacing the running gateway's global state. Key changes: - Add onlyPluginIds and activate options to loadOpenClawPlugins for scoped loads - Add suppressGlobalCommands to plugin registry to avoid leaking commands - Replace full registry reloads in onboarding with per-channel scoped snapshots - Validate command definitions in snapshot loads without writing global registry - Preload configured external plugins via scoped discovery during onboarding Co-Authored-By: Claude Opus 4.6 <[email protected]> * fix(test): add return type annotation to hoisted mock to resolve TS2322 * fix(plugins): enforce cache:false invariant for non-activating snapshot loads * Channels: preserve lazy scoped snapshot import after rebase * Onboarding: scope channel snapshots by plugin id * Catalog: trust manifest ids for channel plugin mapping * Onboarding: preserve scoped setup channel loading * Onboarding: restore built-in adapter fallback --------- Co-authored-by: Claude Opus 4.6 <[email protected]> Co-authored-by: Vincent Koc <[email protected]> * feat(plugins): add provider usage runtime hooks * feat(plugins): move bundled providers behind plugin hooks * docs(plugins): document provider runtime usage hooks * docs(plugins): unify bundle format explainer * fix: repair onboarding adapter registry imports * refactor: expand setup wizard input flow * refactor: move signal imessage mattermost to setup wizard * docs: document richer setup wizard prompts * feat(plugins): move anthropic and openai vendors to plugins * fix: repair onboarding setup-wizard imports * test(discord): cover startup phase logging * fix: reduce plugin and discord warning noise * chore: raise plugin registry cache cap * build: suppress protobufjs eval warning in tsdown * refactor: tighten setup wizard onboarding bridge * refactor: move bluebubbles to setup wizard * refactor: move nextcloud talk to setup wizard * CLI: restore lightweight root help and scoped status plugin preload * Matrix: lazy-load runtime-heavy channel paths * CI: add CLI startup memory regression check * MSTeams: lazy-load runtime-heavy channel paths * refactor: expand setup wizard flow * refactor: move whatsapp to setup wizard * refactor: move irc to setup wizard * refactor: move tlon to setup wizard * refactor: move googlechat to setup wizard * refactor: expose setup wizard sdk surfaces * Feishu: lazy-load runtime-heavy channel paths * Google Chat: lazy-load runtime-heavy channel paths * fix: gate setup-only plugin side effects * feat(web-search): add plugin-backed search providers * fix(web-search): restore build after plugin rebase * refactor(web-search): move providers into company plugins * WhatsApp: lazy-load setup wizard surface * fix: align channel adapters with plugin sdk * fix: repair node24 ci type drift * refactor(google): merge gemini auth into google plugin * feat(plugins): merge openai vendor seams into one plugin * refactor(plugins): lazy load provider runtime shims * perf(cli): trim help startup imports * perf(status): defer heavy startup loading * fix(matrix): assert outbound runtime hooks * refactor: extend setup wizard account resolution * refactor: move feishu zalo zalouser to setup wizard * refactor: move matrix msteams twitch to setup wizard * refactor: drop channel onboarding fallback * fix: quiet discord startup logs * Slack: lazy-load setup wizard surface * Feishu: drop stale runtime onboarding export * Discord: lazy-load setup wizard surface * Signal: lazy-load setup wizard surface * perf(plugins): lazy-load setup surfaces * fix(cli): repair preaction merge typo * Signal: restore setup surface helper exports * iMessage: lazy-load setup wizard surface * Nextcloud Talk: split setup adapter helpers * fix: remove stale dist plugin dirs * BlueBubbles: split setup adapter helpers * test(plugins): cover retired google auth compatibility * refactor(tests): share plugin registration helpers * refactor(plugins): share bundled compat transforms * refactor(google): split oauth flow modules * refactor(plugin-sdk): centralize entrypoint manifest * fix(docs): harden i18n prompt failures * docs(i18n): sync zh-CN google plugin references * fix(docs): run i18n through a local rpc client * build(plugin-sdk): enforce export sync in check * docs(google): remove stale plugin references * IRC: split setup adapter helpers * refactor: move line to setup wizard * refactor: trim onboarding sdk exports * Telegram: split setup adapter helpers * fix: allow plugin package id hints * Tlon: split setup adapter helpers * LINE: split setup adapter helpers * fix: restore ci type checks * fix: resolve line setup rebase drift * Mattermost: split setup adapter helpers * refactor: merge minimax bundled plugins * docs: refresh zh-CN model providers * perf(plugins): lazy-load channel setup entrypoints * Google Chat: split setup adapter helpers * Matrix: split setup adapter helpers * MSTeams: split setup adapter helpers * feat(telegram): add topic-edit action * fix(telegram): normalize topic-edit targets * fix: add Telegram topic-edit action (openclaw#47798) * Feishu: split setup adapter helpers * fixed main? * Zalo: split setup adapter helpers * refactor(plugins): split lightweight channel setup modules * Zalouser: split setup adapter helpers * Status: skip unused channel issue scan in JSON mode * fix(plugins): tighten lazy setup typing * fix: tighten outbound channel/plugin resolution * fix(ci): repair security and route test fixtures * secrets: harden read-only SecretRef command paths and diagnostics (openclaw#47794) * secrets: harden read-only SecretRef resolution for status and audit * CLI: add SecretRef degrade-safe regression coverage * Docs: align SecretRef status and daemon probe semantics * Security audit: close SecretRef review gaps * Security audit: preserve source auth SecretRef configuredness * changelog Signed-off-by: joshavant <[email protected]> --------- Signed-off-by: joshavant <[email protected]> * Gateway: add presence-only probe mode for status * refactor: move group access into setup wizard * feat: add nostr setup and unify channel setup discovery * fix: drop duplicate channel setup import * feat: add openshell sandbox backend * feat(system-prompt): replace hardcoded identity with butley-system-prompt.md Cherry-picked from work (9e1137a). Guilherme's PR #4. Co-authored-by: Bob * fix: suppress SSE finalization on retryable rate-limit errors (openclaw#32) Cherry-picked from work (456f091). Retryable provider errors (429, overload) no longer kill the SSE stream — keeps it open during gateway retries/failover so text flows when retry succeeds. Co-authored-by: Bob * Status: scope JSON plugin preload to configured channels * feat: persist previousSessionId chain across session resets (openclaw#34) Cherry-picked from work (b465233). Adapted: upstream already has previousSessionEntry — only added previousSessionIdForChain for fallback chain persistence on reset/new/idle-expiry. Co-authored-by: Bob * Status: lazy-load read-only account inspectors * refactor(core): land plugin auth and startup cleanup * chore: restore butley-api + clickup-api custom plugins from alpha These custom extensions were missing from the rebase branch. Copied from alpha verbatim. Co-authored-by: Bob * CLI: route gateway status before program registration * feat: add remote openshell sandbox mode * docs: expand openshell sandbox docs * feat: add firecrawl onboarding search plugin * Gateway: lazy-load SSH status helpers * refactor: rename channel setup flow seam * refactor: move setup fallback into setup registry * feat: add synology chat setup wizard * build: add setup entrypoints for migrated channel plugins * docs: update channel setup docs * fix: update feishu setup adapter import * Status: lazy-load channel summary helpers * Agents: skip eager context warmup for status commands * Status: route JSON through lean command * refactor(plugins): move auth and model policy to providers * fix: control UI sends correct provider prefix when switching models The model selector was using just the model ID (e.g. "gpt-5.2") as the option value. When sent to sessions.patch, the server would fall back to the session's current provider ("anthropic") yielding "anthropic/gpt-5.2" instead of "openai/gpt-5.2". Now option values use "provider/model" format, and resolveModelOverrideValue and resolveDefaultModelValue also return the full provider-prefixed key so selected state stays consistent. * fix: format default model label as 'model · provider' for consistency The default option showed 'Default (openai/gpt-5.2)' while individual options used the friendlier 'gpt-5.2 · openai' format. * Nostr: break setup-surface import cycle * Tests: stabilize bundle MCP env on Windows * Status: lazy-load channel security and summaries * Docs: refresh generated config baseline * test: silence vitest warning noise * Status: lazy-load text scan helpers * refactor: rename setup helper surfaces * test: fix fetch mock typing * docs: update channel setup wording * Security: lazy-load channel audit provider helpers * fix(ui): centralize control model ref handling * CLI: route gateway status through daemon status * Status: restore lazy scan runtime typing * feat: token usage tracking via llm_output hook * fix: remove duplicate previousSessionEntry declaration * fix: butley-api extension imports — use openclaw/plugin-sdk instead of relative source paths * fix: ensure llm_output hook is included in butley-api build output - Add optional bundled cluster filtering to listBundledPluginBuildEntries() to skip extensions with native dependencies (matrix, whatsapp, etc.) that cannot be bundled by rolldown on all platforms - Filter plugin-sdk entries for optional clusters to prevent native .node binary bundling failures - Matches upstream's shouldBuildBundledCluster() pattern - butley-api dist output now correctly contains both the tool registration AND the llm_output hook in the register() default export * Revert "fix: ensure llm_output hook is included in butley-api build output" This reverts commit 4558c9d. * fix: revert tsdown changes, copy butley-api to dist via Dockerfile * feat: [FORK-PATCH-37] token usage tracking via direct Convex POST * fix: improve FORK-PATCH-37 logging and add cache token tracking --------- Signed-off-by: sallyom <[email protected]> Signed-off-by: joshavant <[email protected]> Co-authored-by: Jason <[email protected]> Co-authored-by: Jason Separovic <[email protected]> Co-authored-by: Claude Opus 4.6 (1M context) <[email protected]> Co-authored-by: Ayaan Zaidi <[email protected]> Co-authored-by: Praveen K Singh <[email protected]> Co-authored-by: webdevpraveen <[email protected]> Co-authored-by: SkunkWorks0x <[email protected]> Co-authored-by: imanisynapse <[email protected]> Co-authored-by: Onur Solmaz <[email protected]> Co-authored-by: Sahan <[email protected]> Co-authored-by: frankekn <[email protected]> Co-authored-by: Frank Yang <[email protected]> Co-authored-by: thepagent <[email protected]> Co-authored-by: Ace Lee <[email protected]> Co-authored-by: lixuankai <[email protected]> Co-authored-by: Ted Li <[email protected]> Co-authored-by: MonkeyLeeT <[email protected]> Co-authored-by: scoootscooob <[email protected]> Co-authored-by: 助爪 <[email protected]> Co-authored-by: xaeon2026 <[email protected]> Co-authored-by: Andrew Demczuk <[email protected]> Co-authored-by: Peter Steinberger <[email protected]> Co-authored-by: Harold Hunt <[email protected]> Co-authored-by: Tak Hoffman <[email protected]> Co-authored-by: Vincent Koc <[email protected]> Co-authored-by: Aditya Chaudhary <[email protected]> Co-authored-by: Sally O'Malley <[email protected]> Co-authored-by: Nimrod Gutman <[email protected]> Co-authored-by: Lucas Machado <[email protected]> Co-authored-by: xiaoyi <[email protected]> Co-authored-by: Xiaoyi <[email protected]> Co-authored-by: peizhe.chen <[email protected]> Co-authored-by: Tomáš Dinh <[email protected]> Co-authored-by: Gustavo Madeira Santana <[email protected]> Co-authored-by: gumadeiras <[email protected]> Co-authored-by: greptile-apps[bot] <165735046+greptile-apps[bot]@users.noreply.github.com> Co-authored-by: Mason <[email protected]> Co-authored-by: Josh Avant <[email protected]> Co-authored-by: Christopher Chamaletsos <[email protected]>
Preserves explicit `supportsUsageInStreaming` overrides from built-in provider catalogs and user config instead of unconditionally forcing `false` on non-native openai-completions endpoints. Adds `applyNativeStreamingUsageCompat()` to set `supportsUsageInStreaming: true` on ModelStudio (DashScope) and Moonshot models at config build time so their native streaming usage works out of the box. Closes openclaw#46142 Co-authored-by: pezy <[email protected]> (cherry picked from commit 42837a0)
Preserves explicit `supportsUsageInStreaming` overrides from built-in provider catalogs and user config instead of unconditionally forcing `false` on non-native openai-completions endpoints. Adds `applyNativeStreamingUsageCompat()` to set `supportsUsageInStreaming: true` on ModelStudio (DashScope) and Moonshot models at config build time so their native streaming usage works out of the box. Closes openclaw#46142 Co-authored-by: pezy <[email protected]>
2 participants
Summary
This fixes a
model-compatpolicy bug for non-nativeopenai-completionsendpoints.OpenClaw currently forces both
supportsDeveloperRoleandsupportsUsageInStreamingtofalsefor any non-native OpenAI-compatible base URL. That is too broad for streaming usage: built-in catalogs and user config can already declare explicit compat opt-ins, but the normalizer overwrites them.Bailian/DashScope is a concrete repro. It can return streaming token usage, but OpenClaw suppresses
stream_options.include_usagebefore the request is sent, sousage/lastCallUsagenever shows up in the final result.Root cause
This is not a Bailian-specific parser bug in this repo.
The issue is that
normalizeModelCompat()treats streaming usage as a blanket "non-native OpenAI" incompatibility, even when a provider model or user config has already opted in explicitly. That makes the compat schema inconsistent: callers can setsupportsUsageInStreaming, but the normalizer silently discards it.There is one important constraint here:
@mariozechner/pi-aienablesstream_options.include_usageunlesscompat.supportsUsageInStreaming === false. So removing the default-off behavior entirely would reopen unknown custom proxies by default, which would be too risky.What this changes
This PR keeps the conservative default for unknown non-native endpoints, but preserves explicit streaming-usage opt-ins.
Specifically:
openai-completionsendpoints still forcesupportsDeveloperRole = falsesupportsUsageInStreaming = falseis only injected when the flag is unspecifiedsupportsUsageInStreaming: trueandfalsenow survive normalizationThis means known compatible providers can expose token usage, while unknown custom proxies stay on the current safe default.
Scope boundary
This PR does not change
pi-aiparsing or retry behavior.It also does not enable streaming usage for arbitrary custom OpenAI-compatible endpoints by default. Those remain opt-in.
Validation
Checks run from this branch:
CI=1 pnpm exec vitest run --maxWorkers=1 --reporter=verbose src/agents/model-compat.test.tsCI=1 pnpm exec vitest run --maxWorkers=1 --reporter=verbose src/agents/models-config.providers.moonshot.test.tsCI=1 pnpm exec vitest run --maxWorkers=1 --reporter=verbose src/agents/models-config.providers.modelstudio.test.ts -t 'should build the static Model Studio provider catalog'pnpm buildpnpm checkcodex review --base openclaw-upstream/mainManual verification:
openclaw agent --local --jsonrequestusageandlastCallUsagewere present in the resultusage.input=16571,usage.output=76,usage.total=16647This keeps the implementation generic while using Bailian as the motivating repro and validation case.