Major reliability improvements to the message processing pipeline:
- Add messages.seq monotonic cursor replacing timestamp-based cursors,
preventing message loss from timestamp collisions with LIMIT queries
- Add work_items table separating agent production from delivery,
enabling delivery retry without re-running the agent
- Propagate Discord send failures instead of silently swallowing them
- Add Claude 429 → Kimi K2.5 automatic provider fallback with cooldown
- Fix follow-up turn state reset in live index.ts path (not just
message-runtime.ts) to prevent final output from being lost
- Add restart context tracking for graceful restart announcements
- Lazy migration from timestamp cursors to seq cursors for existing data
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
watch_ci tasks in perf channel were silently failing because
NANOCLAW_CHAT_JID was set to the group folder name (e.g. "eyejokerdb-4")
instead of the actual Discord JID. The scheduler rejected the task with
"target group not registered".
Also adds a fallback in ipc.ts to resolve folder names to registered
JIDs when a folder name is passed instead of a JID.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
Two bugs fixed:
1. agent-runner error path race condition: when agent process exited
with error, the promise resolved immediately without waiting for
outputChain, allowing a late wrappedOnOutput to re-persist a stale
session ID after clearSession had already run.
2. follow-up turn state not resetting: when a follow-up IPC message
joined an active run, turn-level flags (finalOutputSentToUser,
sawNonProgressOutput) from the first turn prevented the last
progress from being promoted to a final message.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
- Add MEMENTO_MCP_SSE_URL/ACCESS_KEY/REMOTE_PATH to readEnvFile
- Merge .env vars into cleanEnv for runner process inheritance
- Claude Code runner: add memento-mcp via mcp-remote in mcpServers
- Codex runner: inject memento-mcp section into config.toml
- Various improvements: CI watch, task scheduler, DB, IPC auth
No longer using Docker containers — agents run as direct host
processes. The directory name now reflects the actual purpose.
Updated all references across source code, docs, and skills.
Phase 0-2 of data unification:
- Enable WAL mode + busy_timeout for safe concurrent access
- router_state: keys prefixed with SERVICE_ID (lazy migration)
- sessions: composite PK (group_folder, agent_type)
- registered_groups: filtered by SERVICE_AGENT_TYPE on load
- Logger: service name tag for unified log streams
All changes are backward-compatible with the current split setup.
Actual directory merge (Phase 3-4) is a separate step.
Replace raw JSON-RPC app-server protocol with @openai/codex-sdk.
The SDK wraps `codex exec` which ensures complete task execution per
turn, fixing the issue where app-server mode ended turns prematurely
(agent saying "하겠습니다" without doing the work).
Also fix AGENTS.md copy in agent-runner (was copying instructions.md
which Codex CLI doesn't read).
The codex service uses its own DISCORD_BOT_TOKEN via systemd
EnvironmentFile, so the discord-codex secondary channel is only
relevant for the primary service. Removes spurious credential
missing warning on codex service startup.
Delete token-refresh.ts and credentials.json sync. Auth is now
via CLAUDE_CODE_OAUTH_TOKEN in .env (1-year setup-token).
No more auto-refresh loop or credentials.json dependency.
When CLAUDE_CODE_OAUTH_TOKEN or ANTHROPIC_API_KEY is set in .env,
the credentials.json refresh loop is unnecessary. setup-token
generates a 1-year token, making auto-refresh redundant.
Download and inline content of .txt and other text-based file
attachments instead of showing just the filename placeholder.
Handles long copy-pastes that Discord auto-converts to .txt files.
Truncates at 8000 chars with a note.
Groq runs the same whisper-large-v3-turbo model at 200x+ real-time speed.
Falls back to OpenAI Whisper if GROQ_API_KEY is not set.
Logs elapsed time per transcription for monitoring.
Both services (nanoclaw + nanoclaw-codex) now share cache/transcriptions/
directory. When one service transcribes audio, the other reads from cache
instead of making a second Whisper API call. Uses .pending file to
coordinate concurrent transcription attempts.
- Add purgeChannel() to Channel interface and DiscordChannel
- Uses bulkDelete for recent messages, individual delete for old ones
- Called once at startup before creating fresh dashboard messages
- Remove "📊 Usage 보고" title line
- Use os.uptime() instead of process.uptime() for actual server uptime
- Handle null limitName in Codex rate limits (fallback to limitId or 'Codex')
- Fix Codex usage parsing: extract from rateLimitsByLimitId object
- Fix formatResetKST to handle unix timestamps (number type)
- Remove seconds from update timestamp display
- Fetch Claude Code usage via OAuth API (five_hour/seven_day)
- Fetch Codex usage via app-server JSON-RPC (rateLimits/read)
- System resources: CPU%, memory, disk in GB with emoji thresholds
- Show actual Discord channel names (#name) instead of DB-stored names
- Guard against overlapping usage updates (async API calls)
- Add getChannelMeta to Discord channel (batch guild fetch)
- Status dashboard groups by Discord category, sorts by position
- Add live usage dashboard (CPU, memory, disk, uptime) with message edit
- Fix 4 failing discord.test.ts tests (sendMessage format, image fetch mock)
Dedicated Discord channel shows per-group agent status with live
elapsed time. Single message is edited every 10s instead of spamming.
- GroupQueue tracks startedAt timestamps and exposes getStatuses()
- Channel interface gets editMessage/sendAndTrack for message editing
- STATUS_CHANNEL_ID env var to configure the dashboard channel
- Shows processing/idle/waiting/inactive per registered group
Auto-continue was punishing the model for thinking — text-only turns
were treated as failures and re-prompted up to 5 times with aggressive
"execute now" messages. This made Codex overly compliant and suppressed
critical thinking.
Now each turn result is delivered directly. The model decides whether
to think, push back, or execute.
