Follow-up to #1260. A model (notably the claude-cli proxy under a
degraded session) sometimes emits a tool call as a bare
<invoke name="...">...</invoke> block with no <function_calls> wrapper.
fullToolCallBlockPattern only matched the wrapped form, so the bare
block slipped through to the tag-only strip, leaking the inner
<parameter> command text into the reply while the tool never ran.
Match and remove complete bare <invoke> blocks too (after the wrapped
form), add "<invoke name=" as a detection indicator, and log the
dropped tool name(s). Partial/unterminated artifacts still fall through
to the existing tag strip.
Co-authored-by: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
stripGarbledToolXML removed only the tags of a model-emitted
<function_calls> block, leaking the inner <parameter> argument values
into the user-facing reply, and the dropped call was completely silent.
Detect a complete tool-call block, remove it whole, and log the
attempted tool name(s) at WARN so the no-op is diagnosable. Partial
artifacts still fall through to the existing tag-level strip.
Co-authored-by: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Models occasionally emit NO_REPLY with trailing underscores, quotes,
markdown emphasis, or short explanations (e.g. `NO_REPLY_`,
`**NO_REPLY**`, `NO_REPLY: offline`). Previous word-boundary check
treated `_` as a word char so `NO_REPLY_` slipped through and was
delivered to end users.
Rewrite IsSilentReply to strip decorative wrappers, then prefix-match
case-insensitively. Token is only rejected when glued to another
word character (e.g. `NO_REPLYING`). Intentional divergence from
upstream #19537 — documented in the doc comment.
Port PR #685 fixes for HTTP 400 on Together AI and strict OpenAI-compat hosts,
with additional improvements:
- Gate reasoning_content on assistant history to allowlisted models only
(OpenAI o-series/GPT-5, DeepSeek, Kimi) — prevents HTTP 400 on Together/Qwen
- Gate reasoning_effort to models that support it (OpenAI reasoning family)
- Skip stream_options for Together endpoints (causes HTTP 400)
- Scope DashScope enable_thinking/thinking_budget to DashScope providers only
- Reorder multimodal parts: text before images (Together/Qwen preferred order)
- Add redacted_thinking tag to sanitization patterns with early-exit guard fix
- Upgrade Together detection from URL-only to URL + providerType + name fallback
(mirrors dashScopePassthroughKeys pattern for reverse-proxy compatibility)
- Add comprehensive tests for all new behavior
- Introduced StripMessageDirectives in the agent loop to remove internal
[[name:value]] routing tags from user-facing content while preserving
them in session store for model context
- Narrowed regex from wildcard `(?s)\[\[.*?\]\]` to structured
`\[\[\w+(?::[^\]\n]+)?\]\]` to avoid false positives on legitimate content
- Preserved [[tts...]] tags for TTS AutoTagged pipeline compatibility
- Added 15 unit tests covering strip, preserve, and edge cases
- Fixed step numbering in agent loop (steps 5-9)
Major refactoring of the team system with multiple improvements:
## Removed legacy delegation tools
- Delete `delegate.go`, `delegate_async.go`, `delegate_sync.go`, `delegate_events.go`,
`delegate_policy.go`, `delegate_prep.go`, `delegate_state.go`, `delegate_search_tool.go`
- Delete `evaluate_loop_tool.go`, `handoff_tool.go`
- Remove all references and registrations from tool manager and policy
- Clean up TEAM_PLAYBOOK_IDEAS.md and TEAM_SYSTEM.md (moved to docs)
## Rename await_reply → ask_user
- Rename action `await_reply` → `ask_user`, `clear_followup` → `clear_ask_user`
- Rename functions `executeAwaitReply` → `executeAskUser`, `executeClearFollowup` → `executeClearAskUser`
- Update system prompt with stronger wording to prevent model misuse
- Model was confusing "await_reply" with general waiting; "ask_user" is unambiguous
## Fix auto-followup false positives
- Add `HasActiveMemberTasks(ctx, teamID, excludeAgentID)` store method
- Guard `autoSetFollowup()` in consumer: skip when lead has active member tasks
- Prevents auto-followup when lead is orchestrating teammates (not waiting for user)
## Task identifier zero-padding
- Change format from `T-1-xxxx` → `T-001-xxxx` (3-digit minimum)
## Refactor workspace WS handlers to filesystem-only
- Rewrite `teams.workspace.list/read/delete` to use pure filesystem (os.ReadDir/ReadFile/Remove)
- Remove DB dependency from workspace WS handlers
- Consistent with storage handler and workspace tools
- Simplify TeamWorkspaceFile type and frontend hook
## Add team events listing API
- New WS method `teams.events.list` with team_id, limit, offset params
- New HTTP endpoint `GET /v1/teams/{id}/events` with bearer auth
- New `ListTeamEvents(ctx, teamID, limit, offset)` store method
- JOIN with team_tasks for team-wide event filtering
## Extract team access policy
- New `team_access_policy.go` — centralized team tool access control
## Migration 000019: team_id columns
- Add team_id foreign key columns to relevant tables
## Other improvements
- Add team_id propagation through agent loop, tracing, sessions
- Update i18n locale files (en/vi/zh) for new tool labels
- Update frontend builtin-tools page and require-setup component
- Bump RequiredSchemaVersion for migration 000019
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.
- 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>
Multi-agent AI gateway with WebSocket RPC, HTTP API, and messaging channel integrations.
Go port of OpenClaw with multi-tenant PostgreSQL, per-user isolation, security hardening,
and production observability.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
