Add _is_system_injected() filter to conversation.py that drops user
messages starting with known system-injected prefixes. These messages
(hook outputs, system reminders, teammate notifications) appear in
JSONL session logs as type: "user" with string content but are not
human-typed input.
Filtered prefixes:
- <system-reminder> — Claude Code system context injection
- <local-command-caveat> — local command hook output
- <available-deferred-tools> — deferred tool discovery messages
- <teammate-message — team agent message delivery (no closing >
because tag has attributes)
This brings Claude parsing to parity with the Codex parser, which
already filters system-injected content via SKIP_PREFIXES (line 222).
The filter uses str.startswith(tuple) with lstrip() to handle leading
whitespace. Applied at line 75 in the existing content-type guard chain.
Affects both chat display and input history navigation — system noise
is removed at the source so all consumers benefit.
tests/test_conversation.py:
- TestIsSystemInjected: 10 unit tests for the filter function covering
each prefix, leading whitespace, normal messages, mid-string tags,
empty strings, slash commands, and multiline content
- TestClaudeSystemInjectedFiltering: 5 integration tests through the
full parser verifying exclusion, sequential ID preservation after
filtering, and all-system-message edge case
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
Split the monolithic context.py (117 lines) into five purpose-specific
modules following single-responsibility principle:
- config.py: Server-level constants (DATA_DIR, SESSIONS_DIR, PORT,
STALE_EVENT_AGE, _state_lock)
- agents.py: Agent-specific paths and caches (CLAUDE_PROJECTS_DIR,
CODEX_SESSIONS_DIR, discovery caches)
- auth.py: Authentication token generation/validation for spawn endpoint
- spawn_config.py: Spawn feature configuration (PENDING_SPAWNS_DIR,
rate limiting, projects watcher thread)
- zellij.py: Zellij binary resolution and session management constants
This refactoring improves:
- Code navigation: Find relevant constants by domain, not alphabetically
- Testing: Each module can be tested in isolation
- Import clarity: Mixins import only what they need
- Future maintenance: Changes to one domain don't risk breaking others
All mixins updated to import from new module locations. Tests updated
to use new import paths.
Includes PROPOSED_CODE_FILE_REORGANIZATION_PLAN.md documenting the
rationale and mapping from old to new locations.
Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>
Add conversation_mtime_ns field to session state that tracks the actual
modification time of conversation files. This enables more responsive
dashboard updates by detecting changes that occur between hook events
(e.g., during streaming tool execution).
Changes:
- state.py: Add _get_conversation_mtime() to stat conversation files
and include mtime_ns in session payloads when available
- conversation.py: Add stable message IDs (claude-{session}-{n} format)
for React key stability and message deduplication
- control.py: Fix FIFO eviction for dismissed_codex_ids - set.pop()
removes arbitrary element, now uses dict with insertion-order iteration
- context.py: Update dismissed_codex_ids type from set to dict
The mtime approach complements existing last_event_at tracking:
- last_event_at: Changes on hook events (session boundaries)
- conversation_mtime_ns: Changes on every file write (real-time)
Dashboard can now detect mid-session conversation updates without
waiting for the next hook event.
Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>
The Codex conversation parser was only handling "message" payload types,
missing tool calls entirely. Codex uses separate response_items:
- function_call: tool invocations with name, arguments, call_id
- reasoning: thinking summaries (encrypted content, visible summary)
- message: user/assistant text (previously the only type handled)
Changes:
- Parse function_call payloads and accumulate as tool_calls array
- Attach tool_calls to the next assistant message, or flush standalone
- Parse reasoning payloads and extract summary text as thinking
- Add _parse_codex_arguments() helper to handle JSON string arguments
This fixes the dashboard not showing Codex tool calls like exec_command,
read_file, etc.
Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>
Split the 860+ line bin/amc-server into a modular Python package:
amc_server/
__init__.py - Package marker
context.py - Shared constants (DATA_DIR, PORT, CLAUDE_PROJECTS_DIR, etc.)
handler.py - AMCHandler class using mixin composition
logging_utils.py - Structured logging setup with signal handlers
server.py - Main entry point (ThreadingHTTPServer)
mixins/
__init__.py - Mixin package marker
control.py - Session control (dismiss, respond via Zellij)
conversation.py - Conversation history parsing (Claude JSONL format)
discovery.py - Session discovery (Codex pane inspection, Zellij cache)
http.py - HTTP response helpers (CORS, JSON, static files)
parsing.py - Session state parsing and aggregation
state.py - Session state endpoint logic
The monolithic bin/amc-server becomes a thin launcher that just imports
and calls main(). This separation enables:
- Easier testing of individual components
- Better IDE support (proper Python package structure)
- Cleaner separation of concerns (discovery vs parsing vs control)
- ThreadingHTTPServer instead of single-threaded (handles concurrent requests)
Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>
