* fix(lifecycle): reduce GitHub API rate limiting from batch enrichment bypass
Three optimizations to prevent API storms in the lifecycle manager poll cycle:
1. **CRITICAL - maybeDispatchMergeConflicts**: Gate the getMergeability()
fallback to only run when batch enrichment didn't run at all. Previously
it called getMergeability() (3 REST calls) whenever hasConflicts was
undefined, even when the batch had already fetched PR data. Now uses
cachedData.hasConflicts ?? false when the batch ran.
2. **HIGH - maybeDispatchCIFailureDetails**: Use batch enrichment ciChecks
when available instead of calling getCIChecks() (separate REST call)
on every poll. The GraphQL batch query now fetches statusCheckRollup
contexts (individual check names, statuses, URLs) alongside the rollup
state. Falls back to getCIChecks() only when batch didn't run.
3. **MEDIUM - maybeDispatchReviewBacklog**: Throttle getPendingComments +
getAutomatedComments API calls to at most once per 2 minutes per session.
These were called every 30s even when nothing had changed.
Impact: ~8-10 API calls/PR/poll reduced to ~2-4, enabling 3-4x more
concurrent sessions before hitting GitHub's 5,000/hr REST limit.
Also extends PREnrichmentData with ciChecks?: CICheck[] and adds
parseCheckContexts() helper to graphql-batch.ts for parsing CheckRun
and StatusContext nodes from the GraphQL statusCheckRollup.contexts field.
* fix(scm-github): fall back to getCIChecks() when contexts list is truncated
When a PR has >20 CI checks, contexts(first: 20) silently truncates the
list. Setting ciChecks to undefined when pageInfo.hasNextPage is true
ensures maybeDispatchCIFailureDetails falls back to the getCIChecks()
REST call, which returns all checks without truncation.
Also adds pageInfo { hasNextPage } to the contexts GraphQL query so
truncation can be detected.
* fix(lifecycle): prune lastReviewBacklogCheckAt in pollAll cleanup loop
Add the new throttle map to the existing pruning loop that removes stale
entries for sessions no longer in the session list. Previously the map
was only cleared on terminal status transitions, leaving orphaned entries
for sessions removed externally (killed + cleaned up without transition).
* fix(lifecycle): bypass throttle on review transition; fix StatusContext conclusion
Two fixes for automated review findings:
1. Bypass review backlog throttle when a transition reaction just fired for
humanReactionKey or automatedReactionKey. The transitionReaction branch
needs to read the current fingerprint via the API to record
lastPendingReviewDispatchHash. Without bypassing, the throttle prevents
this write and the next unthrottled poll sees a stale (empty) hash,
clears the reaction tracker, and fires a duplicate dispatch.
2. Set conclusion on StatusContext nodes in parseCheckContexts() to match
the REST getCIChecksFromStatusRollup() format (rawState.toUpperCase()).
The CI failure fingerprint includes c.conclusion ?? '', so inconsistent
conclusion values between GraphQL and REST paths caused phantom fingerprint
changes when switching sources, triggering duplicate dispatches.
* fix(scm-github): normalize CheckRun conclusion and map NEUTRAL to skipped
Two consistency fixes in parseCheckContexts() vs the REST path:
1. NEUTRAL conclusion: was mapped to 'passed' (with SUCCESS), but
mapRawCheckStateToStatus() in the REST path maps NEUTRAL to 'skipped'.
Changed to treat NEUTRAL the same as SKIPPED.
2. CheckRun conclusion: was stored as the raw GraphQL string (may be
lowercase). REST getCIChecks/getCIChecksFromStatusRollup always store
conclusion as rawState.toUpperCase(). Now stores rawConclusion which
is already uppercased during the status branching logic.
Both fixes prevent phantom fingerprint changes when maybeDispatchCIFailureDetails
switches between GraphQL batch and REST fallback across poll cycles.
* fix(scm-github): map STALE/NOT_REQUIRED/NONE conclusions to skipped
parseCheckContexts() was mapping these conclusions to 'failed' via the
else fallback, while mapRawCheckStateToStatus() in the REST path
explicitly maps all of them to 'skipped'. Added them to the skipped
branch alongside SKIPPED and NEUTRAL to fully mirror the REST mapping.
* fix(scm-github): map QUEUED/WAITING to pending not running
parseCheckContexts() mapped QUEUED and WAITING CheckRun statuses to
'running', but mapRawCheckStateToStatus() in the REST path maps both
to 'pending'. Only IN_PROGRESS maps to 'running' in the REST path.
Fixes fingerprint inconsistency when switching between GraphQL batch
and REST fallback across poll cycles.
* fix(scm-github): map STARTUP_FAILURE to skipped; guard null pageInfo
- STARTUP_FAILURE conclusion now falls through to the "skipped" branch
(matching mapRawCheckStateToStatus() REST default) instead of the
explicit failure enumeration catch-all
- Null pageInfo guard prevents TypeError from typeof null === "object"
JavaScript quirk when accessing hasNextPage on a null pageInfo field
- Tests added for both cases
* fix(scm-github): map COMPLETED+null conclusion to skipped not passed
When a CheckRun has status COMPLETED and conclusion null, the REST path's
mapRawCheckStateToStatus() converts it to "" which maps to "skipped".
The GraphQL path was incorrectly mapping it to "passed" via !rawConclusion.
Fix: only map rawConclusion === "SUCCESS" to "passed"; null falls through
to the else branch → "skipped", matching the REST path exactly.
|
||
|---|---|---|
| .. | ||
| __tests__ | ||
| src | ||
| CHANGELOG.md | ||
| README.md | ||
| package.json | ||
| tsconfig.build.json | ||
| tsconfig.json | ||
| vitest.config.ts | ||
README.md
@composio/ao-core
Core services, types, and configuration for the Agent Orchestrator system.
What's Here
src/types.ts— All TypeScript interfaces (Runtime, Agent, Workspace, Tracker, SCM, Notifier, Terminal, Session, events)src/services/— Core services (SessionManager, LifecycleManager, PluginRegistry)src/config.ts— Configuration loading + Zod schemassrc/utils/— Shared utilities (shell escaping, metadata parsing, etc.)
Key Files
src/types.ts — The Source of Truth
Every interface the system uses is defined here. If you're working on any part of the orchestrator, start by reading this file.
Main interfaces:
Runtime— where sessions execute (tmux, docker, k8s)Agent— AI coding tool adapter (claude-code, codex, aider)Workspace— code isolation (worktree, clone)Tracker— issue tracking (GitHub Issues, Linear)SCM— PR/CI/reviews (GitHub, GitLab)Notifier— push notifications (desktop, Slack, webhook)Terminal— human interaction UI (iTerm2, web)Session— running agent instance (state, metadata, handles)OrchestratorEvent— events emitted by lifecycle managerPluginModule— what every plugin exports
src/services/session-manager.ts — Session CRUD
Handles session lifecycle:
spawn(config)— create new session (workspace + runtime + agent)list(projectId?)— list all sessionsget(sessionId)— get session detailskill(sessionId)— terminate sessioncleanup(projectId?)— kill completed/merged sessionssend(sessionId, message)— send message to agent
Data flow in spawn():
- Load project config
- Validate issue exists via
Tracker.getIssue()(if issueId provided, fails-fast if not found) - Reserve session ID
- Determine branch name
- Create workspace via
Workspace.create() - Generate prompt via
Tracker.generatePrompt() - Build launch command via
Agent.getLaunchCommand() - Create runtime session via
Runtime.create() - Run
Agent.postLaunchSetup()(optional) - Write metadata file
- Return Session object
Note: If issue validation fails (not found, auth error), spawn fails before creating any resources (no workspace, no runtime, no session ID). This prevents spawning sessions with broken issue references.
src/services/lifecycle-manager.ts — State Machine + Reactions
Polls sessions, detects state changes, triggers reactions:
State machine:
spawning → working → pr_open → ci_failed/review_pending/approved → mergeable → merged
Reactions:
ci-failed→ send fix prompt to agentchanges-requested→ send review comments to agentapproved-and-green→ notify human (or auto-merge)agent-stuck→ notify human
Polling loop:
- For each session: check agent activity state (
Agent.getActivityState()) - If PR exists: check CI status (
SCM.getCISummary()), review state (SCM.getReviewDecision()) - Update session status based on state
- Trigger reactions if state changed
- Emit events
src/services/plugin-registry.ts — Plugin Discovery + Loading
Loads plugins and provides access to them:
register(plugin, config?)— register a plugin instanceget<T>(slot, name)— get plugin by slot + namelist(slot)— list all plugins for a slotloadBuiltins(config?)— load built-in plugins (runtime-tmux, agent-claude-code, etc.)loadFromConfig(config)— load built-ins today; external plugin descriptors are the marketplace extension point
Built-in plugins (loaded by default):
- runtime-tmux, runtime-process
- agent-claude-code, agent-codex, agent-aider, agent-opencode
- workspace-worktree, workspace-clone
- tracker-github, tracker-linear, tracker-gitlab
- scm-github, scm-gitlab
- notifier-desktop, notifier-discord, notifier-slack, notifier-composio, notifier-openclaw, notifier-webhook
- terminal-iterm2, terminal-web
src/config.ts — Configuration Loading
Loads and validates agent-orchestrator.yaml:
Main config sections:
- Runtime data paths are auto-derived from the config location under
~/.agent-orchestrator/{hash}-{projectId}/ port— web dashboard port (default 3000, set different values for multiple projects)terminalPort— terminal WebSocket port (auto-detected if not set)directTerminalPort— direct terminal WebSocket port (auto-detected if not set)defaults— default plugins (runtime, agent, workspace, notifiers)plugins— installer-managed external plugin descriptors (registry, npm, or local)projects— per-project config (repo, path, branch, symlinks, reactions, agentRules)notifiers— notification channel config (Slack webhooks, etc.)notificationRouting— which notifiers get which priority eventsreactions— auto-response config (ci-failed, changes-requested, approved-and-green, etc.)
Zod schemas validate all config at load time.
Common Tasks
Adding a Field to Session
- Edit
src/types.ts→Sessioninterface - Edit
src/services/session-manager.ts→ initialize field inspawn() - Rebuild:
pnpm --filter @composio/ao-core build
Adding an Event Type
- Edit
src/types.ts→EventTypeunion - Emit the event:
eventEmitter.emit()in relevant service - Add reaction handler (optional):
src/services/lifecycle-manager.ts
Adding a Reaction
- Edit
src/services/lifecycle-manager.ts→ add handler function - Wire it up in the polling loop
- Add config schema in
src/config.tsif new reaction type
Feedback Tools (v1)
@composio/ao-core exports two structured feedback tool contracts:
bug_reportimprovement_suggestion
Both share the same required input fields:
titlebodyevidence(array of strings)sessionsourceconfidence(0..1)
Example:
import { FEEDBACK_TOOL_NAMES, FeedbackReportStore, getFeedbackReportsDir } from "@composio/ao-core";
const reportsDir = getFeedbackReportsDir(configPath, projectPath);
const store = new FeedbackReportStore(reportsDir);
const saved = store.persist(FEEDBACK_TOOL_NAMES.BUG_REPORT, {
title: "SSO login loop",
body: "Google SSO redirects back to /login repeatedly.",
evidence: ["trace_id=abc123", "screenshot: login-loop.png"],
session: "ao-22",
source: "agent",
confidence: 0.84,
});
Storage format:
- Reports are persisted under
~/.agent-orchestrator/{hash}-{projectId}/feedback-reports - Each report is a typed key=value file (
report_<timestamp>_<id>.kv) for easy inspection - A deterministic dedupe key (
sha256, 16 hex chars) is generated from normalized tool+content
Migration notes:
- No migration needed for existing AO installs
- The
feedback-reportsdirectory is created lazily on first persisted report
Testing
# Run all core tests
pnpm --filter @composio/ao-core test
# Run in watch mode
pnpm --filter @composio/ao-core test -- --watch
# Run specific test
pnpm --filter @composio/ao-core test -- session-manager.test.ts
Tests are in src/__tests__/:
session-manager.test.ts— session CRUD, spawn, cleanuplifecycle-manager.test.ts— state machine, reactionsplugin-registry.test.ts— plugin loading, resolutiontmux.test.ts— tmux utility functions (not a plugin test)prompt-builder.test.ts— prompt generation utilities
Building
# Build core
pnpm --filter @composio/ao-core build
# Typecheck
pnpm --filter @composio/ao-core typecheck
This package is a dependency of all other packages. Build it first if working on the codebase.
Architecture Notes
Why flat metadata files?
- Debuggability:
cat ~/.agent-orchestrator/<hash>-my-app/sessions/app-3shows full state - No database dependency (survives crashes, easy to inspect)
- Backwards-compatible with bash script orchestrator
Why polling instead of webhooks?
- Simpler (no webhook setup, no ngrok for local dev)
- Works offline (CI/review state is fetched, not pushed)
- Survives orchestrator restarts (no missed events)
Why plugin slots?
- Swappability: use tmux locally, docker in CI, k8s in prod
- Testability: mock plugins for tests
- Extensibility: users can add custom plugins (e.g., company-specific notifier)