fix: Matrix transport parity — dedup, message queue, drain-on-stop#103
Open
junlov wants to merge 12 commits intoPleasePrompto:mainfrom
Open
fix: Matrix transport parity — dedup, message queue, drain-on-stop#103junlov wants to merge 12 commits intoPleasePrompto:mainfrom
junlov wants to merge 12 commits intoPleasePrompto:mainfrom
Conversation
Matrix bot was missing several features that Telegram has: 1. **Message dedup**: Events replayed during sync were processed multiple times. Now uses the same DedupeCache as Telegram keyed by event_id. 2. **Pending task tracking**: Messages arriving while a CLI session is active were spawned as independent asyncio tasks with no visibility. Now tracked per chat_id in MatrixMessageQueue. 3. **Drain on /stop**: When /stop killed the active CLI, queued message tasks immediately acquired the lock and started new sessions, making the bot appear to "keep going." /stop and /stop_all now drain all pending message tasks before aborting. 4. **Transport merge bug**: Sub-agents with transport=matrix were started as Telegram bots because merge_sub_agent_config didn't sync the `transports` list, and a Pydantic validator reset `transport` back to the first entry in the stale list. Also adds: - Matrix support in `ductor agents add` (interactive + CLI flags) - Provider-aware model selection (shows codex/gemini models, not just claude) - Transport column in `ductor agents list` - 14 unit tests for MatrixMessageQueue Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Add pytest-xdist to the test extras and default to `-n auto` so the test suite runs across all CPU cores: 145s → 49s (~3x faster). Add a justfile with `fix`, `check`, and `test` recipes using just's [parallel] attribute to run linters, type checks, and tests concurrently. This brings another roughly 20% speedup (62s → 52s). Fix failing tests. Update README with uv installation and just-based dev workflow.
- Add opencode_provider.py: OpenCodeCLI class implementing BaseCLI - Add opencode_events.py: NDJSON stream parser for OpenCode output - Add opencode to factory.py: provider selection branch - Add opencode_cli_parameters to CLIParametersConfig and CLIServiceConfig - Add check_opencode_auth to auth.py and _CHECKERS dict - Add opencode to init_wizard.py: CLI detection and i18n - Add opencode to orchestrator/core.py: both CLIServiceConfig instances - Add opencode entry to en/wizard.toml i18n Allows /model command to switch between claude, codex, gemini, and opencode CLIs.
Gemini CLI installed via Homebrew (macOS) stores gemini-cli-core in Cellar/<version>/libexec/lib/node_modules/ instead of the usual npm layout. Add _gemini_cellar_candidates() that walks up the directory tree from the package root to find the Homebrew Cellar prefix and discovers models.js there. Discovery now finds all 7 models instead of 0: - gemini-2.5-flash, gemini-2.5-flash-lite, gemini-2.5-pro - gemini-3-flash-preview, gemini-3-pro-preview - gemini-3.1-pro-preview, gemini-3.1-pro-preview-customtools
- config.py: add opencode model prefixes (minimax/, kimi-for-coding/, opencode/) to provider_for() so model IDs route to opencode provider - providers.py: add opencode to default_model_for_provider() -> minimax/MiniMax-M2.7 - providers.py: add opencode to resolve_session_directive() for @OpenCode directive - model_selector.py: add opencode branch in _build_model_step() that immediately switches to default MiniMax model without showing a picker - model_selector.py: add opencode to _handle_model_selected() switch so opencode models switch immediately (no reasoning effort picker) - _build_model_step() now takes orch+key args to support async switch
- Remove redundant double-check in _OpenCodeServer.ensure_running() (mypy incorrectly flagged second check as unreachable) - Add explicit None guards for process.stdout/stderr pipes - Change send_streaming to yield events directly via async for delegation instead of returning a coroutine
The cherry-picked OpenCode PR added OpenCode to the provider auth and model selection logic, but the root /model selector button list was in the commit we intentionally skipped (which also removed Claude). Add the OpenCode button without removing Claude. Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Matrix startup now mirrors Telegram's full startup sequence: - Restart sentinel handling (explicit /restart notifications) - Startup kind detection (first_start, system_reboot, service_restart) with user-facing notifications for first start and reboot - Recovery of interrupted work (in-flight turns, named sessions) - Restart marker still works as before Previously Matrix only notified on manual restart markers, missing first-start, reboot, and recovery notifications entirely. Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
…ault The OpenCode PR hardcoded `minimax/MiniMax-M2.7` as the default model, but that model doesn't exist (ProviderModelNotFoundError). Now uses the configured model when provider is already opencode, falling back to `opencode/kimi-k2.5` otherwise — matching the pattern used by Claude. Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
4 participants
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Summary
Brings the Matrix transport closer to parity with Telegram by fixing several missing features:
DedupeCache(same as Telegram) keyed byevent_id.chat_idin a newMatrixMessageQueueclass./stop: Previously,/stopkilled the active CLI but queued message tasks immediately acquired the lock and started new sessions (bot appeared to "keep going")./stopand/stop_allnow drain all pending tasks first.transport=matrixwere started as Telegram bots becausemerge_sub_agent_configdidn't sync thetransportslist — a Pydantic validator resettransportback to"telegram"from the main agent's stale list.Also adds:
ductor agents add(interactive wizard + CLI flags)ductor agents add(shows actual codex/gemini models from cache, not just claude defaults)ductor agents listoutputMatrixMessageQueueFiles changed
messenger/matrix/message_queue.pymessenger/matrix/bot.py/stopmultiagent/models.pytransportslist in sub-agent mergecli_commands/agents.pyi18n/en/cli.tomltests/test_matrix_message_queue.pyTest plan
MatrixMessageQueueunit tests (14 passing)ductor agents add <name>interactive with Matrix transportductor agents add <name> --transport matrix --homeserver ...CLI mode/stop→ verify no queued re-trigger🤖 Generated with Claude Code