agent-orchestrator/packages/core
prateek 91dd7cc15f
fix: keep Claude Code interactive after initial prompt (#145)
* fix: keep Claude Code interactive after initial prompt

Claude Code's -p flag runs in one-shot mode (exits after responding),
which prevents follow-up messages via `ao send`. Instead, launch Claude
interactively and deliver the initial prompt post-launch via
runtime.sendMessage().

Adds `promptDelivery` property to the Agent interface so each agent
plugin can declare whether prompts should be inlined in the launch
command or sent after the agent starts.

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>

* fix: make post-launch prompt delivery non-fatal and add test coverage

- Move sendMessage call outside the try/catch that destroys the session.
  A prompt delivery failure should not kill a running agent — user can
  retry with `ao send`.
- Add tests: no-prompt + post-launch agent, sendMessage failure resilience,
  5s delay verification, systemPrompt/systemPromptFile alongside omitted -p.

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>

* test: add integration test for prompt delivery (proves bug and fix)

Two real-Claude integration tests that contrast:
1. `-p` mode: Claude exits after responding (the bug)
2. Interactive + sendMessage: Claude stays alive, follow-up works (the fix)

Runs in CI with ANTHROPIC_API_KEY. Skips when prerequisites missing.

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>

* fix(test): skip interactive test without auth, improve TUI readiness detection

The integration test failed in CI because interactive Claude requires
full login auth (not just ANTHROPIC_API_KEY). Skip the interactive suite
when `claude auth status` reports not logged in.

Also fix local flakiness: replace blind 5s sleep with polling for
Claude's TUI prompt character (❯) before sending the first message,
and increase scrollback capture from 200 to 500 lines.

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>

* fix(test): skip interactive test in CI, fix TUI readiness detection

The interactive test ran in CI despite hasInteractiveAuth() — Claude
reports logged in when ANTHROPIC_API_KEY is set, but interactive mode
requires OAuth. Use `!process.env.CI` as the skip condition instead.

Also fix waitForTuiReady false positive: the OAuth screen's
"Paste code here if prompted >" matched the `>` regex. Now checks
the last non-empty line for Claude's specific ❯ prompt character,
and bails early if the OAuth/login screen is detected.

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>

---------

Co-authored-by: Claude Opus 4.6 <noreply@anthropic.com>
2026-02-27 18:11:06 +05:30
..
__tests__ fix: dashboard config discovery + CLI service layer refactoring (#70) 2026-02-18 17:08:48 +05:30
src fix: keep Claude Code interactive after initial prompt (#145) 2026-02-27 18:11:06 +05:30
README.md docs: update port references to reflect configurability (#122) 2026-02-19 07:37:00 +05:30
package.json feat: publish to npm under @composio scope (#32) 2026-02-15 04:28:57 +05:30
tsconfig.build.json feat: implement SCM and tracker plugins (github, linear) (#4) 2026-02-14 15:45:51 +05:30
tsconfig.json feat: implement SCM and tracker plugins (github, linear) (#4) 2026-02-14 15:45:51 +05:30
vitest.config.ts feat: seamless onboarding with enhanced documentation (#66) 2026-02-16 22:22:13 +05:30

README.md

@agent-orchestrator/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 schemas
  • src/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 manager
  • PluginModule — 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 sessions
  • get(sessionId) — get session details
  • kill(sessionId) — terminate session
  • cleanup(projectId?) — kill completed/merged sessions
  • send(sessionId, message) — send message to agent

Data flow in spawn():

  1. Load project config
  2. Validate issue exists via Tracker.getIssue() (if issueId provided, fails-fast if not found)
  3. Reserve session ID
  4. Determine branch name
  5. Create workspace via Workspace.create()
  6. Generate prompt via Tracker.generatePrompt()
  7. Build launch command via Agent.getLaunchCommand()
  8. Create runtime session via Runtime.create()
  9. Run Agent.postLaunchSetup() (optional)
  10. Write metadata file
  11. 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 agent
  • changes-requested → send review comments to agent
  • approved-and-green → notify human (or auto-merge)
  • agent-stuck → notify human

Polling loop:

  1. For each session: check agent activity state (Agent.getActivityState())
  2. If PR exists: check CI status (SCM.getCISummary()), review state (SCM.getReviewDecision())
  3. Update session status based on state
  4. Trigger reactions if state changed
  5. Emit events

src/services/plugin-registry.ts — Plugin Discovery + Loading

Loads plugins and provides access to them:

  • register(plugin, config?) — register a plugin instance
  • get<T>(slot, name) — get plugin by slot + name
  • list(slot) — list all plugins for a slot
  • loadBuiltins(config?) — load built-in plugins (runtime-tmux, agent-claude-code, etc.)
  • loadFromConfig(config) — load plugins from config (npm packages, local paths)

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
  • scm-github
  • notifier-desktop, notifier-slack, notifier-composio, notifier-webhook
  • terminal-iterm2, terminal-web

src/config.ts — Configuration Loading

Loads and validates agent-orchestrator.yaml:

Main config sections:

  • dataDir — where session metadata lives (~/.agent-orchestrator)
  • worktreeDir — where workspaces are created (~/.worktrees)
  • 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)
  • projects — per-project config (repo, path, branch, symlinks, reactions, agentRules)
  • notifiers — notification channel config (Slack webhooks, etc.)
  • notificationRouting — which notifiers get which priority events
  • reactions — 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

  1. Edit src/types.tsSession interface
  2. Edit src/services/session-manager.ts → initialize field in spawn()
  3. Rebuild: pnpm --filter @agent-orchestrator/core build

Adding an Event Type

  1. Edit src/types.tsEventType union
  2. Emit the event: eventEmitter.emit() in relevant service
  3. Add reaction handler (optional): src/services/lifecycle-manager.ts

Adding a Reaction

  1. Edit src/services/lifecycle-manager.ts → add handler function
  2. Wire it up in the polling loop
  3. Add config schema in src/config.ts if new reaction type

Testing

# Run all core tests
pnpm --filter @agent-orchestrator/core test

# Run in watch mode
pnpm --filter @agent-orchestrator/core test -- --watch

# Run specific test
pnpm --filter @agent-orchestrator/core test -- session-manager.test.ts

Tests are in src/__tests__/:

  • session-manager.test.ts — session CRUD, spawn, cleanup
  • lifecycle-manager.test.ts — state machine, reactions
  • plugin-registry.test.ts — plugin loading, resolution
  • tmux.test.ts — tmux utility functions (not a plugin test)
  • prompt-builder.test.ts — prompt generation utilities

Building

# Build core
pnpm --filter @agent-orchestrator/core build

# Typecheck
pnpm --filter @agent-orchestrator/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/my-app-3 shows 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)