SERVICE_AGENT_TYPE was always 'claude-code' in the unified service,
making it a misleading constant. Each group already has its own
agentType field.
- Removed SERVICE_AGENT_TYPE from config.ts
- Session functions now accept agentType parameter (default: 'claude-code')
- Task scheduler uses task-level agent_type for token rotation decisions
- All fallback defaults changed to 'claude-code' literal
- Logging uses 'unified' instead of the misleading type
In paired rooms, cron output was posted by the owner bot. Since the
owner can't respond to its own messages, the cron report just sat
there with no follow-up action.
Now cron output in paired rooms is posted via the reviewer bot. The
owner picks it up as a peer request, analyzes the report, and acts
on it (e.g., fixing Sentry errors). The normal paired review loop
then kicks in to verify the fix.
Replace LLM-per-tick polling with direct `gh api` calls for GitHub
Actions watchers. New `ci_provider` discriminator column routes tasks
to either the host-driven path (zero LLM tokens) or the existing
generic LLM path. GitHub watchers poll at 15s intervals (min 10s)
via `checkGitHubActionsRun()` in the scheduler, with completion
messages sent directly to chat on terminal state.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
Replace scattered reason string literals with a centralized type hierarchy:
- AgentTriggerReason: all possible trigger reasons
- ClaudeRotationReason, CodexRotationReason, NoFallbackCooldownReason:
derived subtypes for specific contexts
- AgentErrorClassification, FallbackTriggerResult, CodexRotationTriggerResult:
discriminated unions for compile-time narrowing
Remove dead export isUsageExhausted() (superseded by
isPrimaryNoFallbackCooldownActive). Replace NO_FALLBACK_COOLDOWN_REASONS
Set with isNoFallbackCooldownReason() type guard.
Add tests for agent-error-detection and provider-retry (423 total passing).
Known limitation: 6 `as CodexRotationReason` casts in Codex rotation paths
where streamed trigger reason is AgentTriggerReason but runtime value is
always CodexRotationReason-compatible.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
When a Claude account is suspended, the API returns "Your organization does
not have access to Claude" on the success path or "Failed to authenticate.
API Error: 403 terminated" on the error path. Both are now classified as
'org-access-denied', suppressed from Discord output, and trigger automatic
token rotation. If all tokens are exhausted, enters cooldown without Kimi
fallback (same as usage-exhausted and auth-expired).
Changes:
- agent-error-detection: add isClaudeOrgAccessDeniedMessage(), expand
classifyClaudeAuthError() and shouldRotateClaudeToken()
- streamed-output-evaluator: detect org-access-denied in success-path chain
- provider-fallback: add NO_FALLBACK_COOLDOWN_REASONS set and
isPrimaryNoFallbackCooldownActive() helper
- provider-retry: handle org-access-denied in rotation loop
- message-agent-executor / task-scheduler: use generalized no-fallback
cooldown check
- Tests: +8 test cases across 5 test files (415 total passing)
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
The streamed-output-evaluator used 'claude' while the rest of the
codebase used 'claude-code'. Align the evaluator type and all call
sites (message-agent-executor, task-scheduler) to use the canonical
'claude-code' | 'codex' domain consistently.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
- Fix token rotation not taking effect: getCurrentToken() was shadowed
by static .env CLAUDE_CODE_OAUTH_TOKEN value in agent environment
- Don't fall back to Kimi on usage-exhausted (only on transient 429/network)
- Add disk cache for Claude usage API data (survives restarts, 429s)
- Rate-limit usage API calls to 1 per token per 5 minutes
- Add ignoreRateLimits option to rotateToken() for exhausted recovery
- Rotate to next token in getActiveProvider() when current is exhausted
- Add isUsageExhausted() helper to provider-fallback
- Parse "try again at" time from rate-limit errors instead of
fixed 1-hour cooldown, with 3-min buffer after reset
- Pass error message to rotateToken/rotateCodexToken for parsing
- Cache usage % and reset time per Codex account for dashboard
- Persist cached usage in rotation state file across restarts
- Show rate-limited Codex accounts with cached 100% + reset time
- Fix dashboard text indicators (*!/space) for monospace alignment
- New codex-token-rotation module: rotates between multiple
~/.codex-accounts/{n}/auth.json on rate-limit
- Copies active account auth to session dir before each spawn
- Fix detectFallbackTrigger to match "usage limit" / "hit your limit"
- Fix streamed error detection: remove claude-only guard so codex
rate-limit errors are caught even when output.status is "success"
- Add rotation in both task-scheduler and message-agent-executor
- Auto-suspend tasks after 3 consecutive quota/auth errors with
retry-after date parsing and Discord notification
- Add tool_progress and tool_use_summary streaming from Claude Agent SDK
- Prefix progress messages with invisible marker so bots ignore them
- Fix misleading provider label in codex service logs
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>
- 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
Scheduled tasks that send messages via send_message (IPC) instead of
returning text as result left the container idle for ~30 minutes until
the hard timeout killed it (exit 137). This blocked new messages for
the group during that window.
Root cause: scheduleClose() was only called inside the
`if (streamedOutput.result)` branch. Tasks that communicate solely
through IPC (e.g. heartbeat check-ins) complete with result=null,
so the 10s close timer was never set.
Fix: also call scheduleClose() on status==='success', covering both
result-based and IPC-only task completions.
Co-authored-by: Claude Opus 4.6 <noreply@anthropic.com>
The send_message tool description incorrectly stated that a scheduled
task's final output is not delivered to the user, instructing agents to
use the MCP tool for any communication. In reality, task-scheduler.ts
unconditionally forwards the agent's result to the user via a streaming
output callback (deps.sendMessage), which is a direct call to the
channel layer — entirely separate from the MCP tool path.
This caused agents following the description to call send_message
explicitly, resulting in duplicate messages: once via MCP and once via
the native streaming callback.
- Remove the incorrect note from the send_message tool description
- Fix the misleading comment at task-scheduler.ts which attributed
result delivery to the MCP tool rather than the streaming callback
