Announcement messages from subagent completion flows were persisted as
regular user messages, polluting conversation history and leaking
orchestration instructions to frontend/CLI.
- Add `internal?: boolean` to SessionEntry message variant
- SessionManager.saveMessage accepts { internal: true } option
- SessionManager.loadMessages filters internal by default
- Agent.runInternal() tags messages as internal, rolls back from memory
- Agent.withRunMutex() prevents concurrent run/runInternal mis-tagging
- AsyncAgent.writeInternal() suppresses event forwarding during internal runs
- Announce flows use writeInternal() instead of write()
- Desktop IPC getHistory reads from session storage (filtered)
- CLI session show parses SessionEntry, supports --show-internal flag
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
- Keep needsCompaction() method from compaction-feedback PR
- Keep tool result pruning from this branch
- Add "pruning" to CompactionEndEvent.reason type in events.ts
Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>
Adding explicit return type to help TypeScript resolve the type
correctly across different build configurations.
Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>
The CompactionResult type includes "pruning" as a reason but SessionEntry
did not, causing type errors when writing compaction entries.
Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>
Two-phase pruning approach based on OpenClaw's microcompact-style:
- Soft Trim (30% utilization): Keep head 1500 + tail 1500 chars of large tool results
- Hard Clear (50% utilization): Replace old tool results with placeholder
Protections:
- Never prunes before first real user message (bootstrap protection)
- Protects last 3 assistant messages and their corresponding tool results
- Skips image-containing tool results
- Respects tool allow/deny lists
Enabled by default in tokens/summary compaction modes.
Introduce MulticaEvent type system parallel to pi-agent-core's AgentEvent,
with compaction_start and compaction_end events. Agent.subscribeAll() merges
both event streams. maybeCompact() now emits events bracketing the compaction
work, gated by a new SessionManager.needsCompaction() pre-check.
Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>
- Log storage write errors in SessionManager.enqueue() instead of
silently swallowing them
- Remove existsSync check in ensureSessionDir to avoid TOCTOU race;
mkdirSync with recursive is idempotent when dir already exists
Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>
The subagent announce flow reads the child's session file to extract
results, but saveMessage() is fire-and-forget (void enqueue). When
agent.run() completes, pending appendEntry writes may not have flushed
to disk yet, causing readLatestAssistantReply() to find no entries.
Add SessionManager.flush() and call it in AsyncAgent.write() after
agent.run() returns. This ensures waitForIdle() won't resolve until
all session data is on disk, fixing the "no return content" issue.
Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>
The mkdir call with { recursive: true } can throw ENOENT on macOS APFS
due to filesystem race conditions. This caused SpawnSession child agents
to silently fail — the session data was never written, so the parent
read back empty results ("子任务完成了，但没有返回内容").
Two fixes:
- Add retry logic to mkdir calls in acquireSessionWriteLock and
ensureSessionDir to handle transient ENOENT
- Add .catch() to SessionManager.enqueue() to prevent unhandled
promise rejections from fire-and-forget storage operations
Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>
Add ReasoningMode type (off/on/stream) to AgentOptions and related
config types. Add extractThinking() for extracting thinking content
blocks from LLM responses, mirroring the existing extractText() pattern.
Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>
Port from OpenClaw:
- session-write-lock: file-level write lock with atomic creation, reference
counting, stale lock detection, and process cleanup handlers
- session-file-repair: auto-detect and repair malformed JSONL lines with
backup and atomic rename
- session-transcript-repair: fix tool call/result pairing issues including
displaced results, duplicates, orphans, and missing inputs
Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>
Add tests for resolveContextWindowInfo, evaluateContextWindowGuard,
and checkContextWindow functions with various threshold scenarios.
Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>
Replace hardcoded ~/.super-multica paths across agent and hub modules
with a single DATA_DIR constant exported from shared.
Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>
* feat(agent): add auto-backgrounding to exec tool
- Add yieldMs parameter to exec tool (default 5s) - commands that don't
complete within this time automatically run in background
- Create shared process-registry.ts for unified process management
- Refactor process.ts to use shared registry
- Add --debug CLI flag for session message logging
- Signal isolation: backgrounded processes ignore abort signals
Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>
* fix(session): preserve tool_use/tool_result pairs during compaction
Previously, session compaction simply kept the last N messages, which
could break tool_use/tool_result pairs if the cut point fell between
them. This caused "tool_call_id is not found" errors from the API.
Now compaction finds a safe cut point that starts from either:
- A user message without tool_result
- An assistant message whose tool_use is needed by the next tool_result
Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>
* feat(session): use Kimi as default model for summary compaction
- Auto-detect MOONSHOT_API_KEY from environment
- Use moonshot-v1-128k (cheaper than k2-thinking)
- Fall back to tokens mode if API key not available
Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>
* docs: add rule to never use git commit --amend
Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>
* docs: clarify git amend rule for immediate fixes
Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>
---------
Co-authored-by: Claude Opus 4.5 <noreply@anthropic.com>
Translate all comments, JSDoc strings, and documentation text from Chinese to English across 19 files for improved accessibility and consistency with the broader developer community.
Co-authored-by: Claude Haiku 4.5 <noreply@anthropic.com>
* feat(agent): add context window guard to prevent token overflow
Implement token-aware context management that validates context window
size on agent initialization and provides intelligent message compaction
based on actual token usage rather than simple message count.
Key changes:
- Add context-window module with guard, token estimation, and types
- Support both "count" (legacy) and "tokens" (new default) compaction modes
- Warn when context window < 32K tokens, block when < 16K tokens
- Trigger compaction at 80% utilization, target 50% after compaction
Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>
* feat(agent): add summary-based compaction using LLM
Implement intelligent compaction that uses LLM to generate summaries
of older messages instead of simply truncating them. This preserves
important context like key decisions, TODOs, and technical details.
Key changes:
- Add summarization.ts with compactMessagesWithSummary functions
- Support chunked summarization for very large histories
- Add "summary" compaction mode alongside "count" and "tokens"
- Auto-resolve API key from environment based on provider
- Graceful fallback to "tokens" mode if model/apiKey unavailable
Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>
---------
Co-authored-by: Claude Opus 4.5 <noreply@anthropic.com>
- Add Agent Profile module for managing agent identity, soul, tools,
memory, and bootstrap configuration
- Add profile CLI (pnpm agent:profile) for creating/listing/showing profiles
- Default sessionId to UUIDv7 instead of "default"
- Expose Agent.sessionId as public readonly property
- Improve exec/process tools error handling (no more crashes on spawn errors)
- Add 'output' action to process tool for reading stdout/stderr
- Better tool descriptions to guide agent in choosing exec vs process
Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>
