Gemini returns finish_reason="tool_calls" (not "length") even when the
thinking budget exhausted max_tokens before tool arguments could be
emitted. Extend the existing truncation retry condition to detect this
case via an allowlist of mutating tools (write_file, edit, exec,
create_image, read_file) that virtually never legitimately call with
empty args. Nullary tools (datetime, heartbeat) pass through unchanged.
Trace: 019d8f33-2de1-7ab2-9a32-9df92cd610dd
Gemini 3 defaults to "high" thinking budget internally when no
reasoning_effort is set, consuming the full max_tokens budget and
leaving no tokens for tool call arguments on small models. Map
OptThinkingLevel to reasoning_effort on Gemini routes (native endpoint
or proxy-by-model), with "medium" downgraded to "high" per Gemini 3
Preview constraint. Non-Gemini OpenAI-compat hosts unchanged.
Trace: 019d8f33-2de1-7ab2-9a32-9df92cd610dd
Google Gemini's OpenAI-compat shim maps role=tool messages to native
FunctionResponse{name, response}; an empty name trips HTTP 400
("Name cannot be empty"). Build raw-ID → tool-name index at serialize
time and populate the name field on role=tool wire messages.
Trace: 019d8f33-2de1-7ab2-9a32-9df92cd610dd
Port goclaw context pruning to match upstream TS design in
openclaw/src/agents/pi-hooks/context-pruning/:
- Opt-in default: prune only when mode="cache-ttl" (was opt-out)
- Remove Pass 0 per-result 30% guard (duplicated Pass 1 with different
suffix, caused wobble)
- Dedupe double prune call per iteration: PruneStage owns the single
entry point; loop_history only runs limitHistoryTurns + sanitizeHistory
- Add cache-TTL gate for Anthropic prompt cache: skip prune while cache
is live, scoped per-session via sync.Map
- Add context.pruned event emission for observability
- Configurable TTL as Go duration string ("5m", "30s")
BREAKING CHANGE: context pruning now opt-in. Add
contextPruning.mode: "cache-ttl" to config.agents.defaults to restore.
Migration 51 / SQLite v19 backfills mode="cache-ttl" for agents with
existing custom context_pruning config missing the mode field, so
previously-configured agents keep pruning after the opt-in flip.
NULL configs stay NULL (new opt-in default applies).
Web UI adds Cache TTL input + toggle wiring mode to cache-ttl/off.
Add tts namespace to desktop i18n (en/vi/zh). Extend tools.json with STT form keys including WhatsApp privacy banner. Mirrors web ui/web locale structure.
Port web voice picker and STT provider form to desktop frontend. Singleton audio preview (hide when preview_url null), whatsapp_enabled toggle with privacy banner. Reuses desktop's custom Combobox primitive. Voice id persists via other_config.tts_voice_id merge in AgentDetailPanel. ToolSettingsDialog routes stt tool to new SttProviderForm.
Add stt-provider-form.tsx React component with form fields for model, language, and API key configuration. Mirror TTS provider form structure for consistency. Integrate into builtin-tool-settings-dialog.tsx. Includes comprehensive unit tests for form validation, submission, and error handling.
Add STT builtin tools seeding:
- PG: migration 000050 (49→50) with stt_scribe and stt_proxy entries
- SQLite: schema version 17→18 with inline seed
- Both: register in gateway_builtin_tools.go with encrypted API key support
Migrations support rollback. Version bumps gated to prevent schema drift.
Implement legacy STT bridge adapter to support existing per-channel STTProxyURL database configuration. Wraps arbitrary proxy endpoints and auto-registers them via Manager at boot. Includes integration tests for URL validation, authentication, and error handling.
Add Transcribe() method to Manager with STT provider dispatch and channel-specific overrides. Implements channelSTTOverrides map, WithChannel context propagation, RegisterChannelSTT registration, and resolveSTTChain fallback logic. Includes manager-level unit tests.
Implement STT proxy provider that wraps arbitrary HTTP-based STT services. Ported 12 test cases from legacy Telegram STT integration to validate multipart/form-data submission, language/model overrides, and API error handling.
Implement native ElevenLabs Scribe STT provider with POST /v1/speech-to-text endpoint. Supports 20MB audio cap, multipart form submission, configurable model and language. Includes comprehensive unit tests for happy path, audio size validation, API errors, and edge cases.
Add STTInput, STTOptions, and TranscriptResult types to support native STT providers (ElevenLabs Scribe). Includes audio_duration_secs JSON tag for transcription metadata.
Update architecture overview with streaming TTS provider layer. Expand tools
system docs with voice/model resolution. Update HTTP and WebSocket RPC
documentation for voice endpoints. Record Phase 02 completion in changelog.
Update CLAUDE.md with voice picker and streaming TTS patterns.
Add VoicePicker with live search and preview button. Integrate into prompt
settings section. Add voice API hooks (useVoices, useRefreshVoices). Localize
UI strings for en/vi/zh. Fix snake_case JSON field mapping (voice_id, preview_url).
Add StreamingTTSProvider interface for audio streaming support. Implement voice
caching with TTL+LRU. Add ElevenLabs model validation and voice list retrieval.
Update module map to list internal/audio/ (unified manager, TTS active,
STT/Music/SFX stubbed/partial) and clarify internal/tts/ as a
24-symbol backward-compat alias layer. Add changelog entry for Phase 1.
Add optional Audio *AudioConfig pointer field on Config with STT and
Music sub-structs. Nil-safe — absent in JSON5 decodes as nil, no
breaking change. cfg.Tts retained unchanged for backward compat.
setupAudioExtras stub wired for Phase 3/4 STT/Music provider
registration.
internal/tts becomes a thin backward-compat alias over internal/audio:
15 type aliases, 6 constants, 5 constructors, 5 compile-time signature
guards. All pre-refactor callers compile unchanged. alias_test.go
enforces symbol coverage and type identity. Old per-provider files
(manager, types, elevenlabs, openai, edge, minimax) are removed in
the same commit to keep history bisectable.
MCP servers with require_user_credentials (e.g. Notion) were defined
but never loaded into the agent's tool registry. Three gaps:
1. getUserMCPTools was defined but never called — add call in
makeBuildFilteredTools before FilterTools runs each iteration.
2. hasMCPTools stayed false when only user-credential servers existed,
so agentToolPolicyWithMCP never injected "group:mcp" into alsoAllow.
Now set true when mcpUserCredSrvs is non-empty.
3. Per-user BridgeTools were registered in the registry but never added
to the "mcp" tool group, so expandSpec("group:mcp") returned empty.
Add MergeToolGroup helper for additive group updates.
Also add debug log when getUserMCPTools skips due to empty userID.
Stale recovery sweeps traces by `start_time < NOW() - threshold`, which
measures trace age rather than inactivity. Any threshold low enough to
be useful (2-10 min) kills legitimate long-running agent runs: research
chains, large code generation, extended shell commands routinely exceed
10 minutes.
Disabled in Start() — function kept in place for easy re-enable once a
`last_span_at` column is added so recovery can gate on "no activity for
N minutes" instead of "started > N min ago".
Trade-off: zombie traces from gateway crashes may remain `running` in
DB. Accepted: primary abort path (router 2-phase + trace.status WS
event) handles the common case; safety-net gap preferred over false
kills of healthy runs.
Integration test RecoverStaleNow() still works (manual trigger, not
loop-dependent) so coverage of the recovery function itself is
preserved for when it's re-enabled.
viettranx