* feat: add ElevenLabs Scribe STT transcriber and Telegram SendVoice support
Add ElevenLabsTranscriber as an alternative speech-to-text provider using
the ElevenLabs Scribe API (scribe_v1). This enables voice message
transcription for users who already have an ElevenLabs API key, without
requiring a separate Groq account.
Changes:
- Add ElevenLabsTranscriber implementing the Transcriber interface
- Update DetectTranscriber to check providers.elevenlabs.api_key first,
falling back to Groq for backward compatibility
- Add ElevenLabs to ProvidersConfig
- Add "voice" media type for OGG files with "voice" in filename
- Add SendVoice support in Telegram channel for voice bubble messages
- Add comprehensive tests for ElevenLabs transcriber
Configuration:
"providers": {
"elevenlabs": {
"api_key": "sk_your_key_here"
}
}
Closes#1503 (partial)
* fix: move voice-bubble detection into Telegram channel to avoid regression in other channels
Address review feedback: keep inferMediaType returning "audio" for all
OGG files. Voice-bubble detection (SendVoice vs SendAudio) is now done
inside the Telegram channel based on filename, so other channels that
map "audio" explicitly are unaffected.
* fix: align VoiceConfig struct tags to pass golines formatter
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
* fix(agent): use ModelName in loop test added by upstream
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
---------
Co-authored-by: Claude Sonnet 4.6 <noreply@anthropic.com>
LLM
Prevent LLM from seeing its own credentials (API keys, tokens, secrets)
by filtering sensitive values from tool call results before sending to
the
model. Values are collected from .security.yml and replaced with
[FILTERED] using an efficient strings.Replacer (O(n+m)).
- Add FilterSensitiveData and FilterMinLength to ToolsConfig
- Implement SensitiveDataReplacer() with sync.Once caching in
SecurityConfig
- Use reflection to collect all sensitive values (Model API keys,
channel
tokens, web tool API keys, skills tokens)
- Apply filtering in agent loop at 4 tool result locations
- Add comprehensive tests covering all token types
Anthropic API returns 400 when multiple tool_result blocks share the same
tool_use_id, or when consecutive tool results are sent as separate user
messages. This fix:
1. Adds ToolCallID deduplication in sanitizeHistoryForProvider (context.go)
to drop duplicate tool results before sending to any provider.
2. Merges consecutive tool result messages into a single user message with
multiple tool_result content blocks in Anthropic's buildRequestBody,
for both "user" (with ToolCallID) and "tool" role messages.
3. Adds tests for both behaviors.
Co-authored-by: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
- Add BaiduSearchConfig struct and register in WebToolsConfig/defaults
- Insert Baidu Search in priority chain: DuckDuckGo > Baidu > GLM Search
- Use perplexityTimeout (30s) — Qianfan is LLM-based
- Fix response parsing: use references[] field per API spec
- Add baidu_search block to config.example.json
docs: sync configuration.md and README Documentation table across all languages
- Complete truncated configuration.md for fr/ja/pt-br/vi/zh: add Spawn
async flow diagram, Providers table, Model Configuration (all vendors,
examples, load balancing, migration), Provider Architecture, Scheduled
Tasks, and Advanced Topics links
- Add Hooks/Steering/SubTurn entries to Documentation table in all 8
READMEs (en/zh/fr/id/it/ja/pt-br/vi), ordered before Troubleshooting
- Add Baidu Search row to web search table in all 8 READMEs and
tools_configuration.md (en + 5 i18n); zh README reorders search
engines with China-friendly options first
- Add Matrix channel docs translations (fr/ja/pt-br/vi)
- Add Weixin channel to chat-apps.md and all README Channels tables
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
- spawnSubTurn: set result=nil on panic instead of constructing a non-nil ToolResult
- HardAbort: roll back session history to initialHistoryLength after Finish()
- drainBusToSteering: switch to non-blocking reads after first message so function
returns promptly when the inbound channel is empty
- remove obsolete documentation files
* feat(telegram): stream LLM responses in real-time via sendMessageDraft
Implements real-time token streaming to Telegram using the sendMessageDraft
API (telego v1.6.0). Instead of showing only a "Thinking..." placeholder
until the full response arrives, users now see partial LLM output appear
in the chat as it's generated.
The streaming pipeline threads through all layers:
- StreamingProvider interface (providers/types.go): opt-in ChatStream()
method that receives an onChunk callback with accumulated text
- OpenAI-compatible SSE streaming (openai_compat/provider.go): parses
SSE events with stream:true, handles text deltas and tool call assembly
- Anthropic native streaming (anthropic/provider.go): uses SDK's
NewStreaming() for direct Anthropic API connections
- HTTPProvider delegation (http_provider.go): delegates ChatStream to
the underlying openai_compat provider
- StreamingCapable + Streamer interfaces (channels/interfaces.go):
opt-in channel capability like TypingCapable/PlaceholderCapable
- Telegram streamer (telegram/telegram.go): BeginStream returns a
telegramStreamer that throttles sendMessageDraft calls (3s/200 chars)
with graceful degradation on API errors
- StreamDelegate bridge (bus/bus.go): decouples agent loop from channel
manager without tight imports
- Manager integration (manager.go): implements StreamDelegate, tracks
streamActive state, coordinates with placeholder editing
- Agent loop (loop.go): uses ChatStream when both provider and channel
support streaming, cancels stream on tool calls, skips PublishOutbound
when Finalize already delivered the message
Graceful degradation:
- Bots without forum/topics mode: first sendMessageDraft error sets
failed=true, subsequent Updates become no-ops, Finalize still delivers
via SendMessage. User sees normal non-streaming behavior.
- Non-streaming providers: type assertion fails, falls back to Chat()
- Config opt-out: streaming.enabled (default true) in telegram config
Closes#1098
* fix(telegram): delete placeholder message when streaming delivers response
When streaming was active, the "Thinking..." placeholder message stayed
in the chat because preSend only deleted the tracking entry without
removing the actual Telegram message. Now preSend deletes the placeholder
via the new MessageDeleter interface when streamActive is set.
* refactor(streaming): remove dead code and simplify streaming wiring
- Delete unused Anthropic ChatStream/parseStream (-131 lines) — factory
creates HTTPProvider for all OpenAI-compat providers including OpenRouter
- Simplify runLLMIteration from 4 to 3 return values (remove unused
streamed bool)
- Replace managerStreamer struct with finalizeHookStreamer using embedding
(Update/Cancel promoted, only Finalize overridden)
* fix(streaming): skip streamer acquisition when SendResponse is false
Heartbeat messages set SendResponse=false but the streaming path
was unconditionally acquiring a streamer, causing HEARTBEAT_OK to
leak to Telegram via streamer.Finalize().
* fix(streaming): guard streamer for non-sendable messages, add streaming config
Skip streamer acquisition for heartbeat (NoHistory=true), preventing
HEARTBEAT_OK from leaking to Telegram via streamer.Finalize().
Add streaming.enabled to Telegram defaults and example config.
* feat(telegram): stream LLM responses in real-time via sendMessageDraft
Implements real-time token streaming to Telegram using the sendMessageDraft
API (telego v1.6.0). Instead of showing only a "Thinking..." placeholder
until the full response arrives, users now see partial LLM output appear
in the chat as it's generated.
The streaming pipeline threads through all layers:
- StreamingProvider interface (providers/types.go): opt-in ChatStream()
method that receives an onChunk callback with accumulated text
- OpenAI-compatible SSE streaming (openai_compat/provider.go): parses
SSE events with stream:true, handles text deltas and tool call assembly
- Anthropic native streaming (anthropic/provider.go): uses SDK's
NewStreaming() for direct Anthropic API connections
- HTTPProvider delegation (http_provider.go): delegates ChatStream to
the underlying openai_compat provider
- StreamingCapable + Streamer interfaces (channels/interfaces.go):
opt-in channel capability like TypingCapable/PlaceholderCapable
- Telegram streamer (telegram/telegram.go): BeginStream returns a
telegramStreamer that throttles sendMessageDraft calls (3s/200 chars)
with graceful degradation on API errors
- StreamDelegate bridge (bus/bus.go): decouples agent loop from channel
manager without tight imports
- Manager integration (manager.go): implements StreamDelegate, tracks
streamActive state, coordinates with placeholder editing
- Agent loop (loop.go): uses ChatStream when both provider and channel
support streaming, cancels stream on tool calls, skips PublishOutbound
when Finalize already delivered the message
Graceful degradation:
- Bots without forum/topics mode: first sendMessageDraft error sets
failed=true, subsequent Updates become no-ops, Finalize still delivers
via SendMessage. User sees normal non-streaming behavior.
- Non-streaming providers: type assertion fails, falls back to Chat()
- Config opt-out: streaming.enabled (default true) in telegram config
Closes#1098
* fix(telegram): delete placeholder message when streaming delivers response
When streaming was active, the "Thinking..." placeholder message stayed
in the chat because preSend only deleted the tracking entry without
removing the actual Telegram message. Now preSend deletes the placeholder
via the new MessageDeleter interface when streamActive is set.
* refactor(streaming): remove dead code and simplify streaming wiring
- Delete unused Anthropic ChatStream/parseStream (-131 lines) — factory
creates HTTPProvider for all OpenAI-compat providers including OpenRouter
- Simplify runLLMIteration from 4 to 3 return values (remove unused
streamed bool)
- Replace managerStreamer struct with finalizeHookStreamer using embedding
(Update/Cancel promoted, only Finalize overridden)
* fix(streaming): skip streamer acquisition when SendResponse is false
Heartbeat messages set SendResponse=false but the streaming path
was unconditionally acquiring a streamer, causing HEARTBEAT_OK to
leak to Telegram via streamer.Finalize().
* fix(streaming): guard streamer for non-sendable messages, add streaming config
Skip streamer acquisition for heartbeat (NoHistory=true), preventing
HEARTBEAT_OK from leaking to Telegram via streamer.Finalize().
Add streaming.enabled to Telegram defaults and example config.
* fix(picoclaw): add missing closing brace for StreamingProvider interface
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
* fix: resolve golangci-lint formatting issues
Fix gci import ordering in telegram and anthropic provider, and break
long function signature in openai_compat provider to satisfy golines.
* fix: address code review feedback on streaming PR
- Deduplicate Streamer interface: alias channels.Streamer to bus.Streamer
to prevent type drift across packages
- Increase SSE scanner buffer to 10MB max to handle large single-line
responses that exceed bufio.Scanner's 64KB default
- Switch draftID generation from math/rand to crypto/rand for
collision-resistant random IDs
- Add context cancellation check in SSE parsing loop so cancelled
streams stop processing immediately
- Log Finalize failures with chat_id and content length for debugging
silent message delivery failures
* feat: make streaming throttle interval and min growth configurable
Move hardcoded streamThrottleInterval (3s) and streamMinGrowth (200)
into StreamingConfig so they can be tuned per deployment via config
or environment variables.
* fix(telegram): use parseTelegramChatID in DeleteMessage and BeginStream
These two functions called undefined parseChatID. Use
parseTelegramChatID with _ for the unused threadID instead of adding
a wrapper function. Fixes all three CI checks.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
* fix(streaming): set streamActive only after successful Finalize
Move onFinalize hook to run after Streamer.Finalize succeeds, so that
if Finalize fails the streamActive flag stays false and the regular
placeholder fallback path remains available.
Addresses review feedback from @alexhoshina.
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
---------
Co-authored-by: Claude Opus 4.6 <noreply@anthropic.com>
Refactor agent loop execution around runTurn, add explicit turn state and interrupt semantics, and automatically continue queued steering that misses the current turn boundary.
Changed newTestAgentLoop calls from using 3 blank identifiers to 2 by
assigning the unused provider parameter and explicitly marking it as
unused with `_ = provider`. This fixes the dogsled linter violations
that were causing CI failures.
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
Add nil checks in NewSpawnTool and NewSubagentTool constructors to
handle nil manager gracefully. Fix spelling errors (cancelled->canceled)
and remove unused test code. Update tests to use mock spawner.
Replace hardcoded constants with config-driven parameters in agents.defaults:
- MaxDepth, MaxConcurrent, DefaultTimeout, DefaultTokenBudget, ConcurrencyTimeout
- Support JSON config and env vars (PICOCLAW_AGENTS_DEFAULTS_SUBTURN_*)
- Add getSubTurnConfig() for runtime config resolution with defaults
- Apply defaultTokenBudget when no explicit budget is provided
Rationale: SubTurn is agent execution infrastructure, not a tool, so it belongs
in agents.defaults rather than tools config.
Example:
{
"agents": {
"defaults": {
"subturn": {
"max_depth": 5,
"max_concurrent": 10,
"default_timeout_minutes": 10
}
}
}
}
Add ActualSystemPrompt and InitialMessages fields to SubTurnConfig to enable
stateful worker context passing across multiple evaluation iterations.
Changes:
- Add ActualSystemPrompt field to separate system role from user task description
- Add InitialMessages field to preload ephemeral session history before agent loop starts
- Add Messages field to ToolResult for carrying session history (internal use, not serialized)
- Update runTurn to inject system prompt and preload history from InitialMessages
- Update AgentLoopSpawner to map new fields from tools.SubTurnConfig to agent.SubTurnConfig
This enables the evaluator-optimizer execution strategy in team tool to maintain
worker context across iterations while keeping SubTurn isolation intact.
* fix(tools): propagate tool registry to subagents via Clone
SubagentManager was created with an empty ToolRegistry and SetTools()
was never called, causing all subagent tool invocations to fail with
"tool not found". This was a regression from the multi-agent refactor.
Fix: clone the parent agent's tool registry into the subagent manager
after creation but before spawn/spawn_status registration — giving
subagents access to file, exec, web, and other tools while preventing
recursive subagent spawning.
- Add ToolRegistry.Clone() for independent shallow copies
- Call subagentManager.SetTools(agent.Tools.Clone()) in registerSharedTools
- Add tests for Clone isolation, empty clone, and hidden tool state
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
* fix(tools): fix cron_test build error and add TTL clone test
- Fix cron_test.go:229 — replace non-existent SubscribeOutbound(ctx)
with select on OutboundChan(), matching the MessageBus channel API
- Add TestToolRegistry_Clone_PreservesTTLValue per reviewer feedback
- Add version reset note to Clone() doc comment
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
---------
Co-authored-by: Claude Opus 4.6 <noreply@anthropic.com>
* feat(agent): add structured agent definition loader
Parse AGENT.md frontmatter into a runtime definition and pair it with SOUL.md while keeping a legacy AGENTS.md fallback for transition.
Refs #1218
* refactor(agent): build context from structured agent files
Use AGENT.md and SOUL.md as the structured bootstrap source, ignore IDENTITY.md for structured agents, remove USER.md from the new context flow, and update pkg/agent tests accordingly.
Refs #1218
* refactor(onboard): switch workspace templates to AGENT.md
Replace the legacy AGENTS.md, IDENTITY.md, and USER.md templates with a structured AGENT.md plus SOUL.md, and update the onboard template test to assert the new generated files.
Refs #1218
* docs(readme): update workspace layout for AGENT.md
Refresh the documented workspace tree across the README translations so onboarding now points to AGENT.md and SOUL.md instead of the retired AGENTS.md, IDENTITY.md, and USER.md files.
Refs #1218
* feat(agent): restore workspace USER.md context
* docs(readme): document workspace USER.md layout
* fix: sort agent definition imports for gci
- Critical flag was declared but never acted on; non-critical SubTurns
now break out of the iteration loop when IsParentEnded() returns true
- tools.SubTurnConfig was missing Critical/Timeout/MaxContextRunes,
making those fields unreachable from the tools layer; added fields and
wired them through AgentLoopSpawner.SpawnSubTurn
- Removed subTurnResults sync.Map from AgentLoop — it was a redundant
alias for the same channel already stored in turnState.pendingResults;
dequeuePendingSubTurnResults now reads directly via activeTurnStates
- Replace hardcoded concurrencySem size 5 with maxConcurrentSubTurns constant
- Update affected tests to match new dequeuePendingSubTurnResults API
- Added `/subagents` platform command to visualize the active task tree.
- Implemented GetAllActiveTurns and FormatTree in AgentLoop to support cross-session observability.
- Fixed a bug where sub-turns spawned via tools were not registered in the global `activeTurnStates` map, making them invisible to system queries.
- Enhanced tree rendering logic to identify and display "orphaned" subagents (children that outlive their parent turns).
- Registered the new command in `builtin.go` and injected the turn state provider into the commands runtime.
Modified Files:
- pkg/agent/turn_state.go: Added TurnInfo snapshotting and recursive tree formatting.
- pkg/agent/loop.go: Injected GetActiveTurn hook and implemented multi-root forest rendering.
- pkg/agent/subturn.go: Added child turn registration into activeTurnStates.
- pkg/commands/cmd_subagents.go: New command implementation.
- pkg/commands/builtin.go: Command registration.
This commit addresses several critical concurrency and state management bugs within the SubTurn execution and delivery logic.
1. Fix Goroutine Leak & Deadlock in deliverSubTurnResult:
- Replaced non-blocking select with a safe blocking select that listens to `resultChan` and a new `<-parentTS.Finished()` channel.
- This ensures results are not arbitrarily dropped when the channel is full (preventing orphaned valid results), while also guaranteeing the child goroutine safely unblocks and exits if the parent finishes execution early.
2. Prevent "Send on Closed Channel" Fatal Panics:
- Removed `close(pendingResults)` and `drainPendingResults` from `turnState.Finish()`.
- The pendingResults channel is now naturally garbage collected, completely eliminating the race condition panic when a child attempts delivery at the exact moment the parent finishes.
- Added a `defer recover()` failsafe inside deliverSubTurnResult to gracefully emit Orphan events in extreme edge cases.
3. Fix Truncation Recovery Prompt Drop:
- Fixed the runTurn truncation retry logic by introducing an explicit `promptAlreadyAdded` boolean.
- Ensures that the dynamically generated `recoveryPrompt` is correctly injected into the LLM history sequence on subsequent iterations, adhering to API roles without duplicating arrays.
4. Test Suite Stabilization:
- Fixed TestDeliverSubTurnResultNoDeadlock to accurately wait for deterministic deliveries instead of racing timeouts.
- Replaced defunct closed-channel tests with TestFinishedChannelClosedState matching the new Finished() mechanism.
- Fixed the Finish(true) parameter in TestGrandchildAbort_CascadingCancellation to correctly validate Context cascade behavior.
- All tests now pass cleanly without hanging or emitting false positives.
Problem:
During subturn context limit or truncation recoveries, the recovery loops repeatedly
called `runAgentLoop` with the same or modified `UserMessage`. Because `runAgentLoop`
unconditionally adds the `UserMessage` to the session history, this resulted in:
1. Duplicate User Messages polluting the history upon `context_length_exceeded` retries.
2. The possibility of injecting empty User Messages if `opts.UserMessage` was artificially blanked out to work around the duplication.
3. Messy or duplicate entries during `finish_reason="truncated"` recovery injections.
Solution:
- Introduce `SkipAddUserMessage` boolean to `processOptions` to explicitly control whether the agent loop should write the user prompt to history.
- Add an explicit `opts.UserMessage != ""` check in `runAgentLoop` to prevent polluting history with empty message content.
- In `subturn.go`'s recovery loop, set `SkipAddUserMessage: contextRetryCount > 0` to skip writing the user message on context
* config: add prefer_native and NativeSearchCapable for model-native search
* providers: implement native web search for OpenAI and Codex
* agent: use provider-native search when prefer_native and supported
* tests: add coverage for model-native search
* fix: Golang lint errors
* fix: update the code based on the review
* fix: update codex_provider_test
* fix: Fixed the bug where the bus was closed and consumers had unfinished messages.
* fix: remove unnecessary blank line in Close method
* fix: refactor message bus and channel handling for improved performance and reliability
* fix: improve message handling and bus closure logic for better reliability
* fix: reduce sleep duration in agent loop for improved responsiveness
* fix the test case
Problem:
When parent turn finishes early, all child SubTurns receive "context canceled"
error,because child context was derived from parent context.
Solution:
Implement a lifecycle management system that distinguishes between:
- Graceful finish (Finish(false)): signals parentEnded, children continue
- Hard abort (Finish(true)): immediately cancels all children
Changes:
- turn_state.go:
- Add parentEnded atomic.Bool to signal parent completion
- Add parentTurnState reference for IsParentEnded() checks
- Modify Finish(isHardAbort bool) to distinguish abort types
- subturn.go:
- Add Critical bool to SubTurnConfig (Critical SubTurns continue after parent ends)
- Add Timeout time.Duration for SubTurn self-protection
- Use independent context (context.Background()) instead of derived context
- SubTurns check IsParentEnded() to decide whether to continue or exit
- loop.go:
- Call Finish(false) for normal completion (graceful)
- Add IsParentEnded() check in LLM iteration loop
- steering.go:
- HardAbort calls Finish(true) to immediately cancel children
Behavior:
- Normal finish: parentEnded=true, children continue, orphan results delivered
- Hard abort: all children cancelled immediately via context
- Critical SubTurns: continue running after parent finishes gracefully
- Non-Critical SubTurns: can exit gracefully when IsParentEnded() returns true
Includes JSONL session persistence (#1170), spawn_status tool, Azure provider,
credential encryption, and various fixes. SubTurn features preserved and
integrated with new spawn_status functionality.
* feat(tools): add SpawnStatusTool for reporting subagent statuses
* feat(tools): enhance SpawnStatusTool to restrict task visibility by conversation context
* feat(tests): add Unicode result truncation and channel filtering tests for SpawnStatusTool
* Potential fix for pull request finding
Co-authored-by: Copilot Autofix powered by AI <175728472+Copilot@users.noreply.github.com>
* feat(tools): enhance SpawnStatusTool with task ID validation and sorting by creation timestamp
* feat(tools): update SpawnStatusTool description and parameter documentation for clarity
* refactor(tests): improve comments for clarity in ChannelFiltering test case
* fix(tools): update no subagents message for clarity and remove unnecessary locking in runTask
* fix(tools): improve description clarity for SpawnStatusTool regarding task context
* feat(tools): add spawn_status tool configuration and registration
* Potential fix for pull request finding
Co-authored-by: Copilot Autofix powered by AI <175728472+Copilot@users.noreply.github.com>
* fix(agent): improve subagent management for spawn and spawn_status tools
* Potential fix for pull request finding
Co-authored-by: Copilot Autofix powered by AI <175728472+Copilot@users.noreply.github.com>
* Potential fix for pull request finding
Co-authored-by: Copilot Autofix powered by AI <175728472+Copilot@users.noreply.github.com>
* fix(tests): update ResultTruncation_Unicode test to use valid CJK character
---------
Co-authored-by: Copilot Autofix powered by AI <175728472+Copilot@users.noreply.github.com>
Co-authored-by: lxowalle <83055338+lxowalle@users.noreply.github.com>
When the entire session history is a single Turn (e.g. one user message
followed by a massive tool response), findSafeBoundary returns 0 and
forceCompression previously did nothing — leaving the agent stuck in
a context-exceeded retry loop.
Now falls back to keeping only the most recent user message when no
safe Turn boundary exists. This breaks Turn atomicity as a last resort
but guarantees the agent can recover.
Also updates docs/agent-refactor/context.md to document this behavior.
Ref #1490
- Fix synchronous SubTurn calls placing results in pendingResults channel,
causing double delivery. Now only async calls (Async=true) use the channel.
- Move deliverSubTurnResult into defer to ensure result delivery even when
runTurn panics. Add TestSpawnSubTurn_PanicRecovery to verify.
- Fix ContextWindow incorrectly set to MaxTokens; now inherits from
parentAgent.ContextWindow.
- Add TestSpawnSubTurn_ResultDeliverySync to verify sync behavior.
Critical fixes (5):
- Fix turnState hierarchy corruption in nested SubTurns by checking context
before creating new root turnState in runAgentLoop
- Fix deadlock risk in deliverSubTurnResult by separating lock and channel ops
- Fix session rollback race in HardAbort by calling Finish() before rollback
- Fix resource leak by closing pendingResults channel in Finish() with recovery
- Add thread-safety docs for childTurnIDs and isFinished fields
Medium priority fixes (5):
- Move globalTurnCounter to AgentLoop.subTurnCounter to prevent ID conflicts
- Improve semaphore acquisition to ensure release even on early validation failures
- Document design choice: ephemeral sessions start empty for complete isolation
- Add final poll before Finish() to capture late-arriving SubTurn results
- Remove duplicate channel registration in spawnSubTurn to fix timing issues
Testing:
- Add 6 new tests covering hierarchy, deadlock, ordering, channel lifecycle,
final poll, and semaphore behavior
- All 12 SubTurn tests passing with race detector
This resolves 10 critical and medium issues (5 race conditions, 2 resource leaks,
3 timing issues) identified in code review, bringing SubTurn to production-ready state.
- Fix turnState hierarchy corruption when SubTurns recursively call runAgentLoop
by checking context for existing turnState before creating new root
- Fix deadlock risk in deliverSubTurnResult by separating lock and channel operations
- Fix session rollback race in HardAbort by calling Finish() before rollback
- Fix resource leak by closing pendingResults channel in Finish() with panic recovery
- Add thread-safety documentation for childTurnIDs and isFinished fields
- Move globalTurnCounter to AgentLoop.subTurnCounter to prevent ID conflicts
- Improve semaphore acquisition to ensure release even on early validation failures
- Document design choice: ephemeral sessions start empty for complete isolation
- Add 5 new tests: hierarchy, deadlock, order, channel close, and semaphore
- Add initialHistoryLength field to turnState to snapshot session state at turn start
- Save initial history length in runAgentLoop when creating root turnState
- Implement session rollback in HardAbort via SetHistory, truncating to initial length
- Add TestHardAbortSessionRollback to verify history rollback after abort
- Import providers package in subturn_test.go for Message type
This ensures that when a user triggers hard abort, all messages added during
the aborted turn are discarded, restoring the session to its pre-turn state.
- Add maxConcurrentSubTurns constant (5) and concurrencySem channel to turnState
- Acquire/release semaphore in spawnSubTurn to limit concurrent child turns per parent
- Add activeTurnStates sync.Map to AgentLoop for tracking root turn states by session
- Implement HardAbort(sessionKey) method to trigger cascading cancellation via turnState.Finish()
- Register/unregister root turnState in runAgentLoop for hard abort lookup
- Add TestSubTurnConcurrencySemaphore to verify semaphore capacity enforcement
- Add TestHardAbortCascading to verify context cancellation propagates to child turns
- Add subTurnResults sync.Map to AgentLoop for per-session channel tracking
- Add register/unregister/dequeue methods in steering.go
- Poll SubTurn results in runLLMIteration at loop start and after each tool,
injecting results as [SubTurn Result] messages into parent conversation
- Initialize root turnState in runAgentLoop, propagate via context
(withTurnState/turnStateFromContext), call rootTS.Finish() on completion
- Wire Spawn Tool to spawnSubTurn via SetSpawner in registerSharedTools,
recovering parentTS from context for proper turn hierarchy
- Refactor subagent.go to use SetSpawner pattern
- Add TestSubTurnResultChannelRegistration and TestDequeuePendingSubTurnResults
Session history only stores user/assistant/tool messages — the system
prompt is built dynamically by BuildMessages. Remove the incorrect
system message from TestAgentLoop_ContextExhaustionRetry test data
to match the real data model that forceCompression operates on.
When the entire history is a single Turn (one user message followed by
tool calls and responses, no subsequent user message), the only Turn
boundary is at index 0. Previously the fallback returned targetIndex,
which could land on a tool or assistant message — splitting the Turn.
Return 0 instead, so callers (forceCompression, summarizeSession) see
mid <= 0 and skip compression rather than cutting inside the Turn.
Two estimation bugs fixed:
1. Media tokens were added to the chars accumulator before the chars*2/5
conversion, resulting in 256*2/5=102 tokens per item instead of 256.
Fix: add media tokens directly to the final token count, bypassing
the character-based heuristic.
2. estimateMessageTokens counted both tc.Name and tc.Function.Name for
tool calls, but providers only send one (OpenAI-compat uses
function.name, Anthropic uses tc.Name). Fix: count tc.Function.Name
when Function is present, fall back to tc.Name only otherwise.
Also fix i18n hint text: "auto-detect" was misleading — the backend
uses a 4x max_tokens heuristic, not actual model detection.
Introduce parseTurnBoundaries() which identifies each Turn start index
in the session history. A Turn is a complete "user input → LLM iterations
→ final response" cycle (as defined in the agent refactor design #1316).
findSafeBoundary now uses Turn boundaries instead of raw role-scanning,
making the intent explicit: "find the nearest Turn boundary."
forceCompression drops the oldest half of Turns (not arbitrary messages),
which is simpler and more intuitive. The Turn-based approach naturally
prevents splitting tool-call sequences since each Turn is atomic.
Add tests that reflect actual session data shape: history starts with
user messages (no system prompt), includes chained tool-call sequences,
reasoning content, and media items. Exercises the proactive budget check
path with BuildMessages-style assembled messages.
estimateMessageTokens now counts ReasoningContent (extended thinking /
chain-of-thought) which can be substantial and is persisted in session
history. Media items get a fixed per-item overhead (256 tokens) since
actual cost depends on provider-specific image tokenization.
Session history (GetHistory) contains only user/assistant/tool messages.
The system prompt is built dynamically by BuildMessages and is never
stored in session. The previous code incorrectly treated history[0] as
a system prompt, skipping the first user message and appending a
compression note to it.
Fix: operate on the full history slice, and record the compression
note in the session summary (which BuildMessages already injects into
the system prompt) rather than modifying any history message.
Separate context_window from max_tokens — they serve different purposes
(input capacity vs output generation limit). The previous conflation caused
premature summarization or missed compression triggers.
Changes:
- Add context_window field to AgentDefaults config (default: 4x max_tokens)
- Extract boundary-safe truncation helpers (isSafeBoundary, findSafeBoundary)
into context_budget.go — pure functions with no AgentLoop dependency
- forceCompression: align split to safe boundary so tool-call sequences
(assistant+ToolCalls → tool results) are never torn apart
- summarizeSession: use findSafeBoundary instead of hardcoded keep-last-4
- estimateTokens: count ToolCalls arguments and ToolCallID metadata,
not just Content — fixes systematic undercounting in tool-heavy sessions
- Add proactive context budget check before LLM call in runAgentLoop,
preventing 400 context-length errors instead of reacting to them
- Add estimateToolDefsTokens for tool definition token cost
Closes#556, closes#665
Ref #1439
- Replace duplicate types (ToolResult/Session/Message) with real project types
- Implement ephemeralSessionStore satisfying session.SessionStore interface
- Connect runTurn to real AgentLoop via runAgentLoop + AgentInstance
- Fix subturn_test.go to match updated signatures and types
Co-Authored-By: Claude Sonnet 4 <noreply@anthropic.com>
* make gateway aware of config.json change
* fix according to code review
* fix lint
* fix review comment
* fix for review
* refactor to fix review
* fix for review
* fix for review
* feat(web_search): add load balance and failover for api keys
* feat(web_search): add load balance and failover for api keys
* lint
* new iter to get api key
* deleted conflicts
* feat(session): add SessionStore interface and JSONL backend adapter
Extract a SessionStore interface from the methods the agent loop uses
(AddMessage, GetHistory, SetSummary, TruncateHistory, Save, etc.).
Both SessionManager and the new JSONLBackend satisfy this interface,
allowing the persistence layer to be swapped transparently.
JSONLBackend wraps memory.Store and maps its error-returning API to
the fire-and-forget contract that the agent loop expects — write
errors are logged, reads return empty defaults on failure. Save()
triggers compaction to reclaim space after logical truncation.
Part of #1169
* test(session): add JSONLBackend integration tests
8 tests covering the full SessionStore contract through the JSONL
backend: message roundtrip, tool calls, summary, truncation with
compaction, history replacement, empty sessions, session isolation,
and the complete summarization flow (SetSummary → TruncateHistory →
Save).
Includes compile-time interface satisfaction checks for both
SessionManager and JSONLBackend.
Part of #1169
* feat(agent): wire JSONL session store into agent loop
Replace the concrete *SessionManager field with the SessionStore
interface and initialize the JSONL backend by default. Legacy .json
session files are auto-migrated on first startup. Falls back to
SessionManager if the JSONL store cannot be initialized.
The agent loop code (loop.go) requires zero changes — all method
calls work identically through the interface.
Closes#1169
* fix(session): propagate compact error from Save
Save() was swallowing the error returned by Compact and always
returning nil. Callers checking Save's return value would never
see a compaction failure. Return the error directly so the agent
loop can log or handle it as needed.
* feat(session): add Close to SessionStore interface
Add Close() error to SessionStore so callers can release resources
through the interface. JSONLBackend already had Close; this adds
a no-op implementation to SessionManager for compatibility.
* fix(session): close session stores on shutdown and harden migration
- Add Close() to AgentInstance, AgentRegistry, and AgentLoop so JSONL
file handles are released during gateway shutdown and CLI exit.
- Fall back to SessionManager when migration fails, preventing a split
state where some sessions live in JSONL and others remain in JSON.
- Add defer agentLoop.Close() in the CLI agent command path.
- Document SessionStore interface methods (fire-and-forget contract).
* feat(feishu): implement SendMedia and add send_file tool
Add outbound media support for the Feishu channel so the agent can send
images and files to users via the MediaStore pipeline.
Feishu channel:
- SendMedia dispatches media parts as image or file uploads
- sendImage uploads via Image.Create then sends image message
- sendFile uploads via File.Create then sends file message
- feishuFileType maps extensions to Feishu file_type values
send_file tool:
- New tool lets the LLM send a local file to the current chat
- Validates path, registers file in MediaStore, returns media ref
- Agent loop wires tool registration, MediaStore propagation, and
context updates
Tested on Radxa Cubie A7A (arm64) with Feishu websocket channel.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
* fix(agent): publish outbound media regardless of SendResponse flag
The SendResponse flag controls whether the agent loop publishes the
final text response (callers that publish it themselves set this to
false). However, the media publish path was also gated behind this
flag, which meant tool-produced media was silently dropped for normal
channel messages.
Media should be published immediately when a tool returns media refs,
independent of how the text response is delivered.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
* fix(tools): use magic-bytes MIME detection and add file size limit to send_file
- Replace hardcoded extension-to-MIME map with h2non/filetype (magic
bytes) + mime.TypeByExtension fallback, consistent with the vision
pipeline in resolveMediaRefs
- Add configurable max file size check (defaults to config.DefaultMaxMediaSize,
20 MB) to prevent oversized uploads
- Add tests for magic-bytes detection, extension fallback, size limit,
and default max size
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
* refactor(agent): add ForEachTool to AgentRegistry for cross-agent tool lookup
Extract the pattern of iterating agents to find a named tool into
AgentRegistry.ForEachTool, simplifying SetMediaStore propagation.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
* fix(agent,tools): adapt send_file to ctx-based channel injection after upstream refactor
Replace ContextualTool interface (removed upstream) with direct ctx
reading in SendFileTool.Execute, using ToolChannel/ToolChatID helpers.
Remove updateToolContexts which is no longer needed since ExecuteWithContext
already injects channel/chatID into ctx for all tools.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
* feat(tools): support toggling send_file tool via config
Add SendFileConfig with Enabled field to ToolsConfig, defaulting to
true. Wrap send_file tool registration in loop.go with the config
check, consistent with the pattern used by other tools.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
---------
Co-authored-by: Claude Opus 4.6 <noreply@anthropic.com>
* feat(commands): Session management [Phase 1/2] command centralization and registration
* docs: add design for command registry post-review fixes
Documents the architecture decisions for fixing 5 Important issues
from code review: SubCommand pattern, Deps struct, command-group files,
Executor caching, and Telegram registration dedup.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
* feat(commands): add SubCommand type and EffectiveUsage method
Introduce SubCommand struct for declaring sub-commands structurally
within a parent command Definition. The EffectiveUsage() method
auto-generates usage strings from sub-command names and args,
preventing drift between help text and actual handler behavior.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
* feat(commands): add Deps struct and secondToken helper, remove dead contains()
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
* feat(commands): add sub-command routing to Executor
Uses Registry.Lookup for O(1) command dispatch instead of iterating
all definitions. Definitions with SubCommands are routed to matching
sub-command handlers. Missing or unknown sub-commands reply with
auto-generated usage.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
* refactor(commands): split into command-group files with Deps injection
Extract show/list/start/help into individual cmd_*.go files.
Replace config.Config parameter with Deps struct for runtime data.
Restore /show agents and /list agents sub-commands.
Use EffectiveUsage for auto-generated help text.
Bridge external callers (agent/loop.go, telegram.go) with Deps wrapper
until Task 5 fully wires the Deps fields.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
* perf(commands): cache Executor in AgentLoop, wire Deps with runtime callbacks
Create Executor once in NewAgentLoop instead of per-message. Deps
closures capture AgentLoop pointer for late-bound access to
channelManager and runtime agent model.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
* fix(telegram): remove duplicate initBotCommands, keep async startCommandRegistration only
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
* chore(commands): restore Outcome comments and annotate Deps.Config
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
* refactor(commands): consolidate /switch into commands package, fix ! prefix
Move /switch model and /switch channel handling from inline loop.go
logic into cmd_switch.go using the SubCommand + Deps pattern. This
removes the OutcomePassthrough branch in handleCommand entirely.
Also replace the hardcoded "/" prefix check with commands.HasCommandPrefix
so that "!" prefixed commands are correctly routed to the Executor.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
* chore: add docs/plans to .gitignore and untrack existing files
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
* refactor(commands): address code review findings
- Remove dead ExecuteResult.Reply field and unused branch in loop.go
- Extract shared agentsHandler for /show agents and /list agents
- Remove redundant firstToken/secondToken (use nthToken instead)
- Simplify Telegram startup: pass BuiltinDefinitions directly
- Centralize req.Reply nil guard in executeDefinition
- Extract unavailableMsg constant (was duplicated 5 times)
- Remove unused MessageID from Request
- Remove stale "reserved for Phase 2" comment on Deps.Config
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
* refactor(commands): replace Deps with per-request Runtime
Separate stateless Registry (cached on AgentLoop) from per-request
Runtime (passed to handlers at execution time). This enables future
session management features to inject per-request context without
modifying the command registry.
- Rename Deps → Runtime, move to runtime.go
- Change Handler signature: func(ctx, req) error → func(ctx, req, rt *Runtime) error
- NewExecutor now takes (registry, runtime) — executor is created per-request
- BuiltinDefinitions() no longer takes parameters (stateless)
- AgentLoop caches cmdRegistry, builds Runtime via buildRuntime()
- Update all cmd_*.go handlers and tests
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
* style: fix gci import grouping and godoc formatting
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
* fix(onboard): skip legacy AGENT.md when copying embedded workspace templates
The workspace/ directory contains both AGENT.md (legacy) and AGENTS.md
(current). copyEmbeddedToTarget was copying both, causing the test
TestCopyEmbeddedToTargetUsesAgentsMarkdown to fail. Skip AGENT.md
during the walk to match the expected behavior.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
* refactor(agent): address self-review comments on loop.go
- Move cmdRegistry init into struct literal (review comment #11)
- Rename buildRuntime → buildCommandsRuntime for clarity (review comment #12)
- Add comment to default switch case explaining passthrough (review comment #13)
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
* refactor(commands): address code review findings on naming and correctness
- Rename dispatcher.go → request.go (no Dispatcher type remains)
- Rename cmd_agents.go → handler_agents.go (shared handler, not a top-level command)
- Add modelMu to protect AgentInstance.Model writes in SwitchModel
- Add ListDefinitions to Runtime so /help uses registry instead of BuiltinDefinitions()
- Fix SwitchChannel message: validation-only callback should not say "Switched"
- Propagate Reply errors in executor instead of discarding with _ =
- Add HasCommandPrefix unit test
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
* refactor(onboard): extract legacy filename to constant
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
* fix(agent): handle commands before route error check
Move handleCommand() before the routeErr gate so global commands
(/help, /show, /switch) remain available even when routing fails.
Context-dependent commands that need a routed agent will report
"unavailable" through their nil-Runtime guards.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
* revert: remove unnecessary AGENT.md skip in onboard
Reverts 02d0c04 and 74deae1. The test failure was caused by a local
leftover workspace/AGENT.md file (gitignored but embedded by go:embed).
Deleting the local file fixes the root cause; the code-level skip was
never needed.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
* fix: executeDefinition Unknown option
* fix(agent): use routed agent for model commands, restore Telegram command diff
- Remove modelMu: message processing is serial, no concurrent writes
- Pass routed agent to handleCommand/buildCommandsRuntime instead of
always using default agent
- GetModelInfo/SwitchModel are nil when agent is nil (route failed),
handlers reply "unavailable"
- Restore GetMyCommands + slices.Equal check before SetMyCommands to
avoid unnecessary Telegram API calls on restart
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
* fix(commands): remove unintended config mutation in SwitchModel
SwitchModel should only update the routed agent's runtime Model field.
Writing to cfg.Agents.Defaults.ModelName was a behavioral change that
corrupts the default agent config when switching a non-default agent.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
* refactor(commands): move /switch channel to /check channel
/switch channel only validates availability, not actually switching.
Rename to /check channel to match actual behavior. /switch channel
now shows a redirect message pointing users to the new command.
Addresses review feedback from yinwm on PR #959.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
---------
Co-authored-by: Claude Opus 4.6 <noreply@anthropic.com>
1. CJK token estimation: replace flat rune_count/3 with script-aware
counting — CJK runes (U+2E80–U+9FFF, U+F900–U+FAFF, U+AC00–U+D7AF)
count as 1 token each, non-CJK runes at /4. This fixes a 3x
underestimate for Chinese/Japanese/Korean text that could incorrectly
route complex CJK messages to the light model.
2. Routing observability: SelectModel now returns the computed score as
a third value. selectCandidates logs the score on both paths — Info
level for light model selection, Debug level for primary model
selection.
3. Added tests: TestExtractFeatures_TokenEstimate_Mixed (CJK+ASCII mix),
TestRouter_SelectModel_ReturnsScore.
Addresses review feedback from @mingmxren.
Upstream added ThinkingLevel, SummarizeMessageThreshold,
SummarizeTokenPercent, MaxMediaSize, and maybeSummarize.
Our branch added Router, LightCandidates, and selectCandidates.
Both sets of changes are kept. Dead updateToolContexts removed
(upstream deleted it; no callers exist).
* fix: eliminate data races on shared tool instances
Signed-off-by: Boris Bliznioukov <blib@mail.com>
* fix: remove unused indirect dependency on github.com/gdamore/tcell/v2
Signed-off-by: Boris Bliznioukov <blib@mail.com>
* fix: reviewer comments improve context handling for tool execution and ensure defaults for non-conversation callers
Signed-off-by: Boris Bliznioukov <blib@mail.com>
---------
Signed-off-by: Boris Bliznioukov <blib@mail.com>
* feat: add extended thinking support for Anthropic models
Support configurable thinking levels (off/low/medium/high/xhigh/adaptive)
via `agents.defaults.thinking_level` config field.
- "adaptive": uses Anthropic's adaptive thinking API (Claude 4.6+)
- "low/medium/high/xhigh": uses budget_tokens (all thinking-capable models)
- "off": disables thinking (default)
API constraints handled:
- Temperature cleared when thinking is enabled
- budget_tokens clamped to max_tokens-1
- Thinking response blocks parsed into Reasoning field
Relates to #645, #966
* fix: address PR review feedback for thinking support
- Add ThinkingCapable interface for provider capability detection
- Warn when thinking_level is set but provider doesn't support it
- Warn when temperature is cleared due to thinking enabled
- Adjust budget values per Anthropic best practices (medium=16K, xhigh=64K)
- Add budget clamp warning and 80% threshold warning
- Add parseResponse thinking block tests
- Add thinking_level field to config.example.json
* refactor: move ThinkingLevel from AgentDefaults to ModelConfig
Thinking is a model-level capability, not a global agent property.
Per-model config avoids silent ignoring on non-Anthropic providers
and eliminates spurious warning logs in multi-provider setups.
Addresses PR #1076 review feedback from @yinwm.
Resolve merge conflicts to keep both SearXNG and GLM Search
providers. Updated search priority order to:
Perplexity > Brave > SearXNG > Tavily > DuckDuckGo > GLM Search
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
When the LLM returns multiple tool calls, they are now executed
concurrently using goroutines + sync.WaitGroup instead of sequentially.
Results are collected in an indexed slice and processed in original order
to preserve message ordering. MessageTool.sentInRound is changed to
atomic.Bool for thread safety.
Co-authored-by: Claude Opus 4.6 <noreply@anthropic.com>
* feat(config): add GLMSearchConfig for GLM Search provider
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
* test(tools): add failing tests for GLM Search provider
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
* feat(tools): add GLMSearchProvider for web search
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
* feat(agent): wire GLM Search config into web search tool registration
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
---------
Co-authored-by: Claude Opus 4.6 <noreply@anthropic.com>
- Separate third-party imports from local module imports (gci)
- Fix byte slice literal formatting (gofumpt)
- Rename shadowed err variable to ftErr (govet)
- Remove trailing blank lines in test files
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
- serializeMessages: preserve ToolCallID/ToolCalls when Media is present
- resolveMediaRefs: add 20MB file size limit to prevent OOM
- mimeFromExtension: return empty string for unknown extensions
- Add 11 unit tests for serializeMessages, resolveMediaRefs, mimeFromExtension
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
- Use atomic.Bool for closed flag to prevent TOCTOU race between
CallTool and Close operations
- Add double-check pattern in CallTool for thread-safe closed state
- Use atomic Swap in Close to ensure no new calls can start after
closed flag is set
- Move MCP manager cleanup defer before initialization to handle
partial initialization failures
- Update tests to use atomic.Bool operations
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
instance.go:
- Add Router *routing.Router and LightCandidates []FallbackCandidate
to AgentInstance.
- At agent creation, when routing.enabled and light_model resolves
successfully in model_list, pre-build the Router and resolve the
light model candidates once. If the light model isn't in model_list,
log a warning and disable routing for that agent gracefully.
loop.go:
- Add selectCandidates(agent, userMsg, history) helper.
It calls Router.SelectModel and returns either agent.Candidates /
agent.Model (primary tier) or agent.LightCandidates / light_model
(light tier). Returns primary unchanged when routing is disabled.
- In runLLMIteration, resolve (activeCandidates, activeModel) once
before entering the tool-iteration loop. The model tier is sticky
for the entire turn so a multi-step tool chain doesn't switch
models mid-way.
- Replace hard-coded agent.Candidates / agent.Model references in
callLLM and the debug log with the resolved active values.
The fallback chain and retry logic are untouched. When light_model
returns an error the fallback chain handles escalation normally.
Without this function, media:// refs stored by MediaStore are passed
directly to the LLM API, which rejects them as invalid URLs.
resolveMediaRefs() runs after BuildMessages() and before the LLM call,
converting each media:// ref to a data:image/...;base64,... URL that
vision-capable models can process.
Also adds mimeFromExtension() helper for MIME type inference from
file extensions when ContentType metadata is not available.
Add Media field to processOptions, pass msg.Media from inbound
messages through to BuildMessages and serializeMessages so
vision-capable LLMs receive image_url content parts.
Based on work by @as3k in sipeed/picoclaw#555.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
* fix(tools): allow /dev/null redirection and add read/write sandbox split
- Remove deny pattern that incorrectly blocked redirects to /dev/null
- Expand block device write pattern to cover nvme, mmcblk, vd, xvd,
hd, loop, dm-, md, sr and nbd in addition to sd
- Add safe path whitelist for kernel pseudo-devices so workspace path
check does not reject /dev/null, /dev/zero, /dev/random, /dev/urandom,
/dev/stdin, /dev/stdout and /dev/stderr
- Add allow_read_outside_workspace config option (default true) so file
read and list tools are unrestricted while write tools stay sandboxed
Closes https://github.com/sipeed/picoclaw/issues/964
Closes https://github.com/sipeed/picoclaw/issues/965
Signed-off-by: Huang Rui <vowstar@gmail.com>
* feat(tools): add configurable allow patterns and path whitelists
- Add custom_allow_patterns to exec config so users can exempt specific
commands from deny pattern checks
- Add allow_read_paths and allow_write_paths regex lists to tools config
for whitelisting specific paths outside the workspace
- Introduce whitelistFs that wraps sandboxFs and falls through to hostFs
for paths matching whitelist patterns
- Use variadic constructor signatures to keep backward compatibility
Suggested-by: lxowalle
Signed-off-by: Huang Rui <vowstar@gmail.com>
---------
Signed-off-by: Huang Rui <vowstar@gmail.com>
* fix(tools): close resp.Body on retry cancel and cache http.Client instances
Fix resp.Body leak in DoRequestWithRetry where req.Body (request) was
incorrectly closed instead of resp.Body (response) on context cancel.
Cache http.Client on web search/fetch provider structs and channel
adapters (WeCom, LINE) to avoid per-call allocation overhead.
* fix(channels): preserve original http client timeouts for LINE and WeCom
Split LINE single 60s client into infoClient (10s) for bot info lookups
and apiClient (30s) for messaging API calls. Lower WeCom cached client
base timeout from 60s to 30s (matching uploadMedia), and ensure it is
always >= the configured ReplyTimeout so the per-request context
deadline remains the effective limit.
* refactor(tools): extract timeout consts and deduplicate WebFetchTool constructors
Address PR review feedback from xiaket:
- Define searchTimeout, perplexityTimeout, fetchTimeout, defaultMaxChars,
and maxRedirects as package-level consts instead of magic numbers.
- Remove misleading "No proxy" comment in NewWebFetchTool.
- Deduplicate NewWebFetchTool by delegating to NewWebFetchToolWithProxy.
* test(utils): add context cancellation test for DoRequestWithRetry
Verify that resp.Body is properly closed when the context is canceled
during retry sleep, covering the C8 resp.Body leak fix.
* fix(utils): close resp in test to satisfy bodyclose linter
* fix(utils): eliminate flakiness in context cancellation retry test
Synchronize cancellation using an onRoundTrip callback from the
transport wrapper instead of a timing-based context timeout. This
ensures the first client.Do completes before cancel fires, so
cancellation always hits during sleepWithCtx.
- WhatsApp Start(): use deferred cleanup to nil out c.client/c.container
and disconnect/close resources on any error after struct fields are
assigned, preventing stale references and double-close in Stop()
- handleReasoning: treat bus.ErrBusClosed as an expected condition
(DEBUG level) alongside context timeout/cancel, avoiding WARN noise
during normal shutdown
- WhatsApp Send(): detect unpaired state (Store.ID == nil) and return
ErrTemporary instead of attempting to send while QR login is pending
- handleReasoning: check the returned error type (DeadlineExceeded /
Canceled) instead of ctx.Err() to decide log level, so pubCtx
timeouts on a full bus are correctly classified as expected
- Test: fill bus with a short-timeout loop instead of hardcoding the
buffer size (64), making the test resilient to buffer size changes
- Use c.runCtx for GetQRChannel so the QR producer is canceled on Stop()
- Add atomic stopping guard to prevent wg.Add/wg.Wait race in eventHandler
- Make Stop() context-aware: disconnect client before waiting, respect ctx deadline
- Reduce reasoning publish log noise: use debug level for expected ctx errors
- Add test for handleReasoning when outbound bus is full (timeout path)
Add a 5-second timeout to handleReasoning's PublishOutbound call so
fire-and-forget goroutines do not block indefinitely when the outbound
bus channel is full. Reasoning output is best-effort; on timeout the
publish is abandoned with a warning log instead of holding the
goroutine alive.
Fixes goroutine leak introduced in #802.
* fix: remove redundant tools definitions from system prompt
Tools are already provided to the LLM via JSON schema through
ToProviderDefs(), so the text-based tools section in the system
prompt is redundant.
This removes the buildToolsSection() logic and the tools field
from ContextBuilder, reducing system prompt length while maintaining
the "ALWAYS use tools" rule reminder.
Fixes#731
Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>
* fix: correct spelling 'initialized' (was 'initialised')
---------
Co-authored-by: Claude Opus 4.5 <noreply@anthropic.com>
Avoid rebuilding the entire system prompt on every BuildMessages() call
by caching the static portion (identity, bootstrap, skills summary,
memory) and only recomputing it when workspace source files change.
Key changes:
- ContextBuilder caches the static prompt behind an RWMutex with
double-checked locking. Source file changes are detected via cheap
os.Stat mtime checks so no explicit invalidation is needed.
- Track file existence at cache time (existedAtCache map) so that
newly created or deleted bootstrap/memory files also trigger a
rebuild — the old modifiedSince() silently returned false on
os.IsNotExist.
- Walk the skills directory recursively with filepath.WalkDir to
catch content-only edits at any nesting depth; directory mtime
alone misses in-place file modifications on most filesystems.
- ToolRegistry.sortedToolNames() sorts tool names before iteration,
ensuring deterministic tool definition order across calls — a
prerequisite for LLM-side prefix/KV cache reuse.
- Merge all context (static + dynamic + summary) into a single
system message for provider compatibility: the Anthropic adapter
extracts messages[0] as the top-level system parameter, and Codex
reads only the first system message as instructions.
- Fix a data race in BuildMessages() where cachedSystemPrompt was
read without holding the lock in a debug log statement.
- Add tests: single system message invariant, mtime auto-invalidation,
new-file creation detection, skill file content change, explicit
InvalidateCache, cache stability, concurrent access (20 goroutines
x 50 iterations, passes go test -race), and a benchmark.
HTTP timeouts (context deadline exceeded, Client.Timeout) were
incorrectly classified as context window errors, triggering useless
history compression. Replace broad substring checks ("context",
"token", "length") with specific patterns for real context limit
errors and explicitly exclude timeout errors from that path.
Additionally, timeout errors were not retried at all — the retry
loop only handled context window errors. Now timeouts are retried
up to 2 times with exponential backoff (5s, 10s).
* cleanup dead code.
Signed-off-by: Kai Xia <kaix+github@fastmail.com>
* add these two back with flag.
Signed-off-by: Kai Xia <kaix+github@fastmail.com>
* fix ci
Signed-off-by: Kai Xia <kaix+github@fastmail.com>
* remove this confusing line
Signed-off-by: Kai Xia <kaix+github@fastmail.com>
* make fmt
Signed-off-by: Kai Xia <kaix+github@fastmail.com>
* remove unused method.
picked up by golangci-lint run
Signed-off-by: Kai Xia <kaix+github@fastmail.com>
---------
Signed-off-by: Kai Xia <kaix+github@fastmail.com>
Walk backwards over preceding tool messages to find the nearest assistant
with ToolCalls, instead of only checking the immediate predecessor. Add
unit tests for sanitizeHistoryForProvider covering key edge cases.
- Drain buffered messages in MessageBus.Close() so they aren't silently lost
- Replace all context.TODO() with context.WithTimeout(5s) across 7 call sites
- Fix OneBot pending channel leak: send nil sentinel in Stop() and handle
nil response in sendAPIRequest() to unblock waiting goroutines
The configuration field for specifying the model has been renamed from
"model" to "model_name" for better clarity and consistency with the
model_list configuration.
A GetModelName() accessor method has been added to maintain backward
compatibility. Existing configurations using the old "model" field will
continue to work correctly.
This change affects:
- Configuration structure (AgentDefaults struct)
- All references across the codebase
- Documentation in all language variants
- Example configuration files
- MediaStore: use full UUID to prevent ref collisions, preserve and
expose metadata via ResolveWithMeta, include underlying OS errors
- Agent loop: populate MediaPart Type/Filename/ContentType from
MediaStore metadata so channels can dispatch media correctly
- SplitMessage: fix byte-vs-rune index mixup in code block header
parsing, remove dead candidateStr variable
- Pico auth: restrict query-param token behind AllowTokenQuery config
flag (default false) to prevent token leakage via logs/referer
- HandleMessage: replace context.TODO with caller-propagated ctx,
log PublishInbound failures instead of silently discarding
- Gateway shutdown: use fresh 15s timeout context for StopAll so
graceful shutdown is not short-circuited by the cancelled parent ctx
Add outbound media sending capability so the agent can publish media
attachments (images, files, audio, video) through channels via the bus.
- Add MediaPart and OutboundMediaMessage types to bus
- Add PublishOutboundMedia/SubscribeOutboundMedia bus methods
- Add MediaSender interface discovered via type assertion by Manager
- Add media dispatch/worker in Manager with shared retry logic
- Extend ToolResult with Media field and MediaResult constructor
- Publish outbound media from agent loop on tool results
- Implement SendMedia for Telegram, Discord, Slack, LINE, OneBot, WeCom
PublishInbound/PublishOutbound held RLock during blocking channel sends,
deadlocking against Close() which needs a write lock when the buffer is
full. ConsumeInbound/SubscribeOutbound used bare receives instead of
comma-ok, causing zero-value processing or busy loops after close.
Replace sync.RWMutex+bool with atomic.Bool+done channel so Publish
methods use a lock-free 3-way select (send / done / ctx.Done). Add
context.Context parameter to both Publish methods so callers can cancel
or timeout blocked sends. Close() now only sets the atomic flag and
closes the done channel—never closes the data channels—eliminating
send-on-closed-channel panics.
- Remove dead code: RegisterHandler, GetHandler, handlers map,
MessageHandler type (zero callers across the whole repo)
- Add ErrBusClosed sentinel error
- Update all 10 caller sites to pass context
- Add msgBus.Close() to gateway and agent shutdown flows
- Add pkg/bus/bus_test.go with 11 test cases covering basic round-trip,
context cancellation, closed-bus behavior, concurrent publish+close,
full-buffer timeout, and idempotent Close
Channels previously deleted downloaded media files via defer os.Remove,
racing with the async Agent consumer. Introduce MediaStore to decouple
file ownership: channels register files on download, Agent releases them
after processing via ReleaseAll(scope).
- New pkg/media with MediaStore interface + FileMediaStore implementation
- InboundMessage gains MediaScope field for lifecycle tracking
- BaseChannel gains SetMediaStore/GetMediaStore + BuildMediaScope helper
- Manager injects MediaStore into channels; AgentLoop releases on completion
- Telegram, Discord, Slack, OneBot, LINE channels migrated from defer
os.Remove to store.Store() with media:// refs
Add bus.Peer struct and explicit Peer/MessageID fields to InboundMessage,
replacing the implicit peer_kind/peer_id/message_id metadata convention.
- Add Peer{Kind, ID} type to pkg/bus/types.go
- Extend InboundMessage with Peer and MessageID fields
- Change BaseChannel.HandleMessage signature to accept peer and messageID
- Adapt all 12 channel implementations to pass structured peer/messageID
- Simplify agent extractPeer() to read msg.Peer directly
- extractParentPeer unchanged (parent_peer still via metadata)
Models like Moonshot kimi-k2.5 and DeepSeek-R1 return a
reasoning_content field in assistant messages. When thinking is enabled,
the API requires this field to be echoed back in subsequent requests.
PicoClaw was silently dropping it, causing 400 errors on tool-call
round-trips.
- Add ReasoningContent to Message and LLMResponse types
- Parse reasoning_content in openai_compat parseResponse()
- Carry reasoning_content through assistant tool-call messages
- Add unit test for reasoning_content parsing
Fixes#588
ListAgentIDs() was called on every iteration of the inner tool loop,
causing repeated allocations. Capture the slice once and reuse it for
both agentCount and the registration loop.
Implements SearXNG as a third web search provider to address Oracle Cloud
datacenter IP blocking issues and provide a cost-free, self-hosted alternative
to commercial search APIs.
Changes:
- Add SearXNGConfig struct with Enabled, BaseURL, and MaxResults fields
- Implement SearXNGSearchProvider with JSON API integration
- Update provider priority: Perplexity > Brave > SearXNG > DuckDuckGo
- Wire SearXNG configuration through agent tool registration
- Add default config values (disabled by default, empty BaseURL)
Benefits:
- Solves DuckDuckGo datacenter IP blocking (138 bytes redirect responses)
- Zero-cost alternative to Brave Search API ($5/1000 queries)
- Self-hosted solution with 70+ aggregated search engines
- Privacy-focused with no rate limits or API keys required
- Ideal for Oracle Cloud, GCP, AWS, and Azure VM deployments
The implementation follows the existing provider interface pattern and
maintains backward compatibility with all existing search providers.
Co-Authored-By: Claude Sonnet 4.5 <noreply@anthropic.com>
