StripConfigLeak was blocking legitimate responses when predefined agents
mentioned SOUL.md/IDENTITY.md/AGENTS.md in architecture explanations.
Improved detection logic to exclude code blocks but disabled the gate
entirely for now until a more robust approach is designed.
Spans are now emitted BEFORE execution starts with status "running" and
input visible, then updated when the step completes. This lets users see
what the system is doing during long-running LLM calls and tool executions.
- Add EmitSpanUpdate() to collector with separate update buffer
- Flush ordering: batch INSERT new spans first, then process updates
- Split LLM/tool/agent spans into start/end pairs (agent loop + subagent)
- Emit tool span end inside goroutines for parallel calls (no orphans)
- EmitSpanUpdate is a channel send — works after ctx cancellation
- Update go.mod and Dockerfile to Go 1.26
- Apply `go fix ./...` stdlib modernizations across 170+ files
- Add `go fix` to post-implementation checklist in CLAUDE.md
- Fix go fix misapplied rewrite in loop_history.go
- Add tool status display on channels during tool execution (streaming preview + reactions)
- Emit agent.activity events at phase transitions (thinking, tool_exec, compacting)
- Enrich delegation progress with per-member activity and tool info
- Add LLM-based intent classifier for DM status queries when agent is busy
- Keyword fast-path for cancel/status patterns (no LLM cost)
- Falls back to LLM classification with 5s timeout
- Supports status_query (immediate reply) and cancel (abort run) intents
- Register/unregister runs in makeSchedulerRunFunc for channel inbound tracking
- Add sessionRuns secondary index in Router for O(1) IsSessionBusy lookups
- Add intent_classify config toggle (global default + per-agent override)
- Add tool_status config toggle for channel tool status display
- Add i18n keys and translations (en/vi/zh) for status messages
- Add web UI config toggles for intent_classify and tool_status
* fix(mcp-bridge): add per-session agent context and HMAC verification
- Add per-session MCP config with X-Agent-ID/X-User-ID headers instead
of shared global config file
- Sign bridge context headers with HMAC-SHA256 to prevent forgery
- Add bridgeContextMiddleware to verify signatures on MCP bridge requests
- Store MCP configs in ~/.goclaw/mcp-configs/ outside agent workDir
- Use atomic writes (tmp + rename) for MCP config files
- Fix provider rename leaving ghost registry entries
- Remove provider_type from mutable fields on update
- Tighten temp dir permissions from 0755 to 0700
* feat(mcp-bridge): propagate channel routing context through MCP bridge
- Pass channel, chat_id, and peer_kind from agent loop to CLI provider options
- Inject X-Channel, X-Chat-ID, X-Peer-Kind headers in bridge context middleware
- Add BridgeContext struct to bundle per-call context for MCP config generation
- Include channel routing headers in per-session MCP config files
- Expose "message" tool via MCP bridge for cross-channel messaging
- Add extract helpers for new option keys in claude_cli_session.go
* feat(mcp-bridge): forward media attachments to outbound message bus
- Wire MessageBus into gateway server and MCP bridge handler
- Publish tool result media files to outbound bus for channel delivery
- Extract channel/chatID/peerKind from tool context for proper routing
- Add mimeFromExt helper for content-type detection on attachments
* feat(mcp-bridge): inject per-agent DB-backed MCP servers into Claude CLI config
- Add MCPServerLookup type to resolve agent-specific MCP servers from DB
- Wire MCPServerStore through provider registration and HTTP handler
- Extract mcpServerEntryToConfig helper to deduplicate transport config logic
- Add JSON-to-Go helpers (jsonToStringSlice, jsonToStringMap) for DB fields
- Merge per-agent MCP servers at config write time without overriding static entries
* fix(mcp-bridge): use Media struct fields and prefer explicit MimeType
- Map Media.Path to attachment URL instead of treating Media as string
- Use Media.MimeType when available, fall back to extension-based detection
* refactor(providers): deduplicate option extractors and extract bridge media forwarding
- Replace per-field extractors (extractSessionKey, extractAgentID, etc.) with generic extractStringOpt/extractBoolOpt
- Add bridgeContextFromOpts helper to build BridgeContext in one call
- Extract forwardMediaToOutbound from inline block in makeToolHandler
- Change NewBridgeServer msgBus param from variadic to explicit pointer
* fix(providers): validate provider_type on update instead of silently dropping it
- Add explicit validation against ValidProviderTypes with 400 response
- Remove silent delete(updates, "provider_type") that hid invalid values
- Caller now receives clear error when submitting unsupported provider_type
* fix(providers): add header injection validation to MCP bridge headers
- Extend CRLF/null-byte checks to agentID, channel, chatID, and peerKind
- Previously only userID had header injection prevention
- Prevents HTTP header injection via crafted values in MCP config
* fix(mcp-bridge): sign all context fields in HMAC and remove legacy code
- Sign all 5 bridge context fields (agentID|userID|channel|chatID|peerKind)
in HMAC instead of only agentID|userID to prevent channel routing forgery
- Propagate context.Context into MCPServerLookup to respect request
cancellation instead of using context.Background()
- Remove legacy BuildCLIMCPConfig, WithClaudeCLIMCPConfig, mcpConfigPath,
and mcpCleanup (dead code since system is PG-only)
- Use mime.TypeByExtension before custom fallback in mimeFromExt
- Add debug log when media forwarding is skipped due to missing context
- Add thread-safety comment to SetMCPServerLookup
---------
Co-authored-by: Nam Nguyen Ngoc <namnn.0911@gmail.com>
Co-authored-by: viettranx <viettranx@gmail.com>
* fix(agent): use ChannelType in system prompt for proper channel context
The system prompt was using the channel instance name (e.g. "zep-lao") instead
of the platform type (e.g. "zalo_personal"), causing the LLM to not understand
which messaging platform it's running on. This led to context confusion where
the bot would ask users which channel to send to instead of using the current one.
Changes:
- Add ChannelType field to RunRequest and SystemPromptConfig
- Thread channel type from consumer/cron → agent loop → system prompt
- Add WithToolChannelType/ToolChannelTypeFromCtx for tool context
- Register channel types for both config-based and DB-loaded instances
- Fix Zalo group thread type detection with approvedGroups cache
- Update cron handler to resolve channel type for cron-triggered runs
* refactor(channels): add Type() to Channel interface, remove channelTypes map
Move channel type from a separate map in Manager to the Channel interface
itself. BaseChannel.Type() falls back to Name() for config-based channels
where name == type. Extracts resolveChannelType helper to DRY up 6
repeated resolution blocks across consumer and cron handlers.
* feat(zalo): add pending group history for conversation context
Zalo personal groups now record non-@mentioned messages in a ring buffer
(default 50, configurable via history_limit). When the bot IS mentioned,
pending history is flushed as context — matching Telegram/Discord/Feishu.
Separated mention gating from policy gating in checkGroupPolicy for
cleaner control flow.
- Fix create_video: use predictLongRunning API instead of generateContent
(async polling flow: POST → poll every 10s → download video from URI)
- Fix durationSeconds as int (not string) per actual Gemini API requirement
- Fix MediaRef collection order: historical first, current last, so
refs[len-1] always returns the most recent file (fixes read_audio
picking up old file instead of current voice message)
- Remove misleading "video not yet supported" text from Telegram handler
that prevented LLM from calling read_video tool
- Add isNonChatModel() to skip chat-based verify for generation models
(veo-*, dall-e-*, imagen-*, gemini-*-image)
- Add read_audio tool with Gemini File API, OpenAI input_audio, and fallback support
- Add read_video tool with Gemini File API and base64 fallback for video analysis
- Add create_video tool with Gemini Veo and OpenRouter chat completions support
- Add shared gemini_file_api.go for upload → poll → generateContent pipeline
- Add shared openai_compat_call.go for custom JSON chat completions
- Fix system prompt showing denied tools: use filteredToolNames() instead of tools.List()
- Wire audio/video MediaRef context propagation in agent loop
- Register new tools in seed data, policy groups, and web UI settings
- Enforce duration (max 30s) and aspect_ratio limits on create_video
Self-Evolution: predefined agents can now optionally evolve their SOUL.md
(communication style/tone only) when self_evolve is enabled in other_config.
Identity, name, and operating instructions remain locked. Context propagation
flows through LoopConfig → Loop → context.WithValue → interceptor carve-out.
System prompt guides the agent on what it can/cannot evolve.
Instances Tab: new HTTP endpoints and UI tab for viewing/editing per-user
USER.md files on predefined agents. Includes owner-only access checks,
fileName validation (USER.md only), and cache invalidation.
UI: self-evolve toggle in General tab, create dialog, and setup wizard.
Agent type and evolve/static badges with tooltip explanations on cards
and detail header. TooltipProvider added to agents list and detail pages.
- Add persistent media storage (internal/media/) replacing temp file deletion
- Add MediaRef type for lightweight media references in session messages
- Refactor media pipeline to use bus.MediaFile{Path, MimeType} across all channels
- Add read_document builtin tool for PDF/DOCX/XLSX analysis via Gemini native API
- Move image sanitization from Telegram to shared agent/media layer
- Add media reload for multi-turn conversations (images from last 5 messages)
- Add reply-to-message media resolution for Telegram (re-download on reply)
- Add media inventory to compaction summary to preserve awareness after truncation
- Fix coreToolSummaries for read_image, read_document, create_image tools
- Add real-time trace update events via WebSocket broadcast
- Improve trace detail UI with media refs and tool result display
- FilesHandler now serves files by absolute path (auth-token protected)
instead of requiring a workspace root, supporting multiple agent workspaces
- mediaToMarkdown generates relative URLs (/v1/files/...) instead of
absolute http://host:port URLs so images work from any client origin
- Deduplicate media collection in agent loop: prefer result.Media over
MEDIA: prefix parsing to prevent duplicate images
Previous approach failed because:
1. PublishOutbound with channel "ws" is silently dropped (no handler)
2. LLM strips image URLs despite instruction text
3. Post-run AddMessage races with frontend loadHistory()
New approach: ContentSuffix field on RunRequest is appended to the
assistant response inside the agent loop BEFORE saving to session
and BEFORE emitting run.completed. This guarantees image markdown
is in the session when frontend calls loadHistory().
Only affects WS channel — other channels still use ForwardMedia
and their respective outbound handlers.
- Add GET /v1/files/{path} endpoint with Bearer token + query param auth
- Pre-convert media to markdown HTTP URLs in announce messages for WS channel
- Add HideInput flag to skip persisting system messages in session history
- Add fallback to append image URLs if LLM strips them from announce response
- Fix stream-to-history DOM flash by promoting streamed text locally
- Fix tool call card word-wrap and expandable arguments panel
- Fix snake_case mismatch in Message types (toolCalls → tool_calls)
- Add clickable image rendering in markdown renderer
- Scan web_fetch/web_search tool results for prompt injection patterns via inputGuard
- Strip hidden HTML elements (display:none, aria-hidden, sr-only classes) during conversion
- Scope web tool caches per channel to prevent cross-channel cache poisoning
- Enforce domain blocklist and allowlist checks on HTTP redirect targets
- Add untrusted content reminder to external content wrapper
- Log redirect source URL in fetch results for transparency
Co-authored-by: Nam Nguyen Ngoc <namnn.0911@gmail.com>
Add Claude CLI as an LLM provider (subscription-based, no API key needed).
The CLI manages session history, tool execution, and context while GoClaw
forwards messages and streams responses.
Key features:
- Claude CLI provider with session persistence (--resume)
- MCP bridge server exposing GoClaw tools to CLI via streamable-http
- Security hooks (shell deny patterns, workspace path restrictions)
- Per-session mutex preventing concurrent CLI calls
- Onboard wizard for Claude CLI setup and auth verification
- Web UI for adding/managing Claude CLI provider with auth status
- Provider registry Close() for proper shutdown cleanup
Security:
- CLI path validation (only "claude" or absolute paths from DB)
- Token auth middleware for MCP bridge endpoint
- Shell injection prevention in hook scripts (single-quoted paths)
- Relative path resolution before workspace boundary checks
- Resource leak prevention on provider replace/unregister
Co-authored-by: nhokboo <nhokboo@users.noreply.github.com>
Add block.reply event that delivers intermediate assistant text to non-streaming
channels during multi-tool iterations. Includes 2-tier config toggle:
gateway-level default (disabled) + per-channel override (inherit/on/off).
Backend:
- Emit block.reply events from agent loop between tool iterations
- Add BlockReply *bool to GatewayConfig and all 6 channel config structs
- Add BlockReplyChannel interface with ResolveBlockReply() resolution
- Guard delivery in HandleAgentEvent by RunContext.BlockReplyEnabled
- Resolve config at RegisterRun time, pass to consumer goroutine
- Conditional dedup: skip final message if identical to last block reply
UI:
- Gateway settings: Switch toggle for global default
- Per-channel: tri-state select (Inherit from gateway / Enabled / Disabled)
- Protocol: BLOCK_REPLY constant in AgentEventTypes
- Form: coerceBoolSelects for proper JSON boolean serialization
Remove hardcoded "personal assistant running inside GoClaw" identity.
Replace with channel+peer-aware context: "running in Telegram (a group chat)".
Remove model name from Runtime section to prevent leaking provider info.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
Backend: add message to run.started, content to run.completed,
arguments to tool.call events. Frontend: fix arrow direction
(agent → tool), compact card layout (badge+content+timestamp in
single row), show subtype in outer badge for agent events, fix
JSON dialog newline consumption in regex.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
Add 16 new event constants and typed payload structs for full WS
visibility into team agent operations. Enrich AgentEvent with delegation
and routing context (delegationId, teamId, parentAgentId, userId,
channel, chatId). Emit thinking/chunk events for non-streaming runs
so delegate member agents also produce WS events.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
- Copy resp.Thinking to assistant message in agent loop so reasoning_content
is actually passed back to thinking models (Kimi, DeepSeek) on multi-turn
- Fix stripGarbledToolXML returning empty string when valid content remains
after stripping XML artifacts
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
Root cause: TEAM.md instructed LLM to call team_tasks create + spawn
in the "SAME turn", but parallel tool execution caused LLM to
hallucinate task IDs. pendingTasksHint then leaked real task IDs
into error messages, causing cascade failures.
Changes:
- Auto-create team task when spawn is called without team_task_id
(subject from label param, fallback to task description)
- Remove pendingTasksHint() — stop leaking task IDs in error messages
- Update TEAM.md workflow to one-step delegation (just spawn)
- Orphan detection now queries DB for actual pending tasks
- Include team_task_id in spawn tool result for tracing
- Improve in-progress reminder wording
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
Part A — Channel quota limiter (managed mode):
- DB-backed per-user/group request quotas with in-memory 60s TTL cache
- Config merge priority: Groups > Channels > Providers > Default
- Per-group quota override via channels.telegram.groups[chatID].quota
- Migration 000009: index on channel_requests for quota queries
- Hot-reload quota config via pub/sub (TopicConfigChanged)
Part B — Per-run tool call budget:
- Soft stop at configurable limit (default 25, per-agent override)
- MaxToolCalls field on AgentDefaults + AgentSpec + LoopConfig
- LLM gets one final call to summarize when budget exceeded
Part C — Web UI + config page refactor:
- QuotaSection with provider/channel dropdowns (useProviders, useChannelInstances)
- Config page refactored to vertical sidebar tabs layout
- Categories: General, Quota, Agents, Tools, Connections, Advanced, Raw Editor
- Fixed config.patch RPC to serialize raw JSON + baseHash correctly
- Config change pub/sub broadcast from handleApply/handlePatch
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
Send a "still working" message to users after 45s if async delegation is
still running, using PublishOutbound directly. Also expand the team task
reminder to include in-progress tasks so lead agents wait for delegates.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
- Add sendMessageDraft transport (disabled pending Telegram client fix for
"reply to deleted message" artifact — tdesktop#10315, bugs.telegram.org/c/561)
- Split stream_mode into dm_stream/group_stream boolean flags (both default false)
- DM messages no longer set reply_to_message_id (cleaner UX, matching TS)
- Progressive placeholder editing for DMs: "Thinking..." → stream chunks → final
- Update web UI with separate DM/Group streaming toggles
fix(agent): prevent false MEDIA: detection in tool output
parseMediaResult() used strings.Index to find "MEDIA:" anywhere in tool output,
causing false positives when external content (e.g. GitHub releases page)
contained commit messages like "return MEDIA: path from screenshot".
Changed to strings.HasPrefix to only match at start of output.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
Tasks from one Telegram group were being injected into another group's
session because the pending-task hint and /tasks command queried all
tasks team-wide without filtering by the group-scoped userID.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
LLM models often hallucinate UUIDs when delegating, passing a wrong
team_task_id that doesn't exist. Previously the error was bare
("task not found") with no guidance, causing the model to get stuck.
Now the error includes a list of pending tasks so the model can
self-correct. Also refactored prepareDelegation to resolve team once
instead of 3 separate GetTeamForAgent calls, and extracted
pendingTasksHint() to deduplicate hint-building logic.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
Add mid-loop compaction to prevent context overflow during long-running
delegated agent runs (e.g. 225K+ tokens causing DashScope timeouts).
Uses same threshold as maybeSummarize (contextWindow * historyShare)
with actual PromptTokens from LLM response. Only compacts the in-memory
messages slice; pendingMsgs preserves full history for session flush.
Add user_id/channel columns to team_tasks so end users only see their
own tasks. Delegate/system channels bypass the filter to see all tasks.
Group chats use the group-scoped UserID (group:channel:chatID) so all
members share visibility.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
When LLMs call team_tasks create + spawn in parallel, the spawn
tool receives a hallucinated task_id that fails uuid.Parse, causing
a misleading error and bypassing orphan detection.
- Include pending task IDs in spawn error message so LLM can retry
with the correct UUID
- Move spawn counting to post-execution so failed spawns don't
increment teamTaskSpawns, allowing orphan detection to fire
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
Port 4 missing Telegram forum/topic features from TypeScript OpenClaw:
1. Thread-not-found fallback: retry sends without message_thread_id when
a topic is deleted (sendHTML, sendPhoto, sendVideo, sendAudio,
sendDocument, stream flush).
2. Per-topic config: hierarchical config resolution (global → wildcard
group "*" → specific group → specific topic) for groupPolicy,
requireMention, allowFrom, enabled, skills, systemPrompt.
New TelegramGroupConfig/TelegramTopicConfig structs, resolveTopicConfig()
with 10 unit tests.
3. DM topic support: preserve message_thread_id in private chats for
session isolation. New BuildDMThreadSessionKey, parseRawChatID handles
🧵 suffix.
4. createForumTopic agent tool: ForumTopicCreator interface decoupled
from telego, lazy bot resolution via channel manager.
5. Web UI: structured group/topic config form with tri-state booleans
(Inherit/Yes/No), nested collapsible group and topic entries.
Also fix: forum group pairing reply and approval notification now
correctly set MessageThreadID so messages land in the right topic.
Send() extracts threadID from localKey suffix as fallback for cases
where metadata is absent (e.g. pairing approval via SendToChannel).
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
- Fix resolvePath for nested non-existent dirs (use resolveThroughExistingAncestors)
- Channel-isolated workspace: user_agent_profiles.workspace stores channel prefix,
used as source of truth with backward compat for existing users
- Loop caches workspace per-user with CacheKindUserWorkspace invalidation via pubsub
- ContractHome/ExpandHome for portable ~-based paths in DB
- create_image saves to workspace/generated/YYYY-MM-DD/ instead of OS temp dir
- SOUL.md template: add ## Expertise section for domain knowledge
- Summoner buildEditPrompt: section guide, complete file output, frontmatter update
- Bus: Topic* constants for Subscribe/Broadcast keys, CacheKind* for payload kinds
- Teams, delegates, sessions, agent links: various enhancements
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
Team delegation:
- Unify spawn/subagent/delegate into single spawn tool
- Sibling-aware announce suppression with artifact accumulation
- Fix auto-complete race (isLastDelegation guard)
- Add team tasks list limit (20) with search guidance
- Multi-round orchestration patterns in TEAM.md
- Communication guidance for initial vs follow-up delegations
Telegram resilience:
- Add retrySend wrapper (3 attempts, escalating delay) for network errors
- Fix HTML fallback: strip tags + unescape entities instead of showing raw HTML
- Pre-process HTML tags in LLM output to markdown before conversion pipeline
- Skip caption truncation entirely when > 1024 bytes, send text separately
- Auto-send large images (>5MB) as documents to avoid compression
Artifact forwarding:
- Fix missing ContentType on forwarded media (mimeFromExt for result.Media/ForwardMedia)
- Add deliver parameter to write_file for file attachment delivery
- Extend mimeFromExt with document MIME types
UI: fix regenerate dialog overflow, improve task list layout, delegation detail view
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
Implement full thinking mode system: per-agent thinking_level config (off/low/medium/high)
stored in other_config JSONB, per-provider param injection (Anthropic budget_tokens, OpenAI
reasoning_effort, DashScope enable_thinking+thinking_budget), Anthropic extended thinking with
streaming parse and tool use block preservation via RawAssistantContent, thinking token tracking
in trace spans, and Web UI with agent config selector, chat thinking block rendering, and trace
thinking token display. No DB migration needed.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
- Add DashScope (Qwen) native provider with tools+streaming fallback
- Add Bailian Coding provider with hardcoded model list (no /v1/models API)
- Parse reasoning_content in OpenAI-compat streaming/non-streaming responses
- Emit ChatEventThinking events in agent loop for thinking models
- Add vision support for DashScope (qwen3-vl)
- Fix provider form dialog not updating API base URL when switching types
- Update README provider count from 11+ to 13+
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
- Add tool loop detection (toolloop.go): tracks repeated no-progress tool calls
using SHA256 hashing of args+results. Warning at 5 identical calls, force stop
at 10. Prevents Gemini models from burning tokens in infinite loops.
- Inject AVAILABILITY.md negative context when agent has no team/delegation
targets, so models don't waste iterations probing unavailable capabilities.
- Fix Dockerfile: create /app/.goclaw directory so Docker volume initializes
with correct goclaw:goclaw ownership instead of root:root.
- Update collapseToolCallsWithoutSig comments for clarity.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
In managed mode, multiple agents share a single tool registry. The spawn
and subagent tools were constructed with a hardcoded parentID of "default",
which does not exist as an agent in managed mode deployments. This caused
subagent announces to fail with "agent default not found" when any
non-default agent (e.g. a secondary agent) tried to spawn a subagent.
Fix: add ctxAgentKey to tool context keys (same pattern as channel/chatID),
inject the calling agent's key from the loop, and read it in spawn/subagent
tools with a fallback to the construction-time default for backward
compatibility with standalone mode.
viettranx