agent-sync API server

Lightweight HTTP API to run or resume coding agents (Codex + Claude) on a workspace machine and persist normalized transcripts to InstantDB. Provider adapters live in src/providers/; persistence uses @repo/agent-trajectory-persist.

Build / run

- Build: bun run build (bundles to dist/bundle/server.js).
- Run dev (hot): bun run dev:server
- Run built bundle: bun run --bun dist/bundle/server.js (or via the bin agent-sync).

Endpoints

- GET /health
- POST /runs – start a run. Body matches RunRequestSchema (provider, prompt?, model?, reasoningEffort?, etc.).
- POST /runs/:id/resume – resume a trajectory.
- POST /runs/:id/stop – stop a specific trajectory.
- GET /runs/:id – status snapshot.

Schema alignment

- Request/response schemas live in @repo/agent-trajectory-core/api-schema (Zod). The server, TRPC, and frontend share these for validation and options.
- Reasoning support is model-aware (e.g., xhigh only on gpt-5.1-codex-max; Claude reasoning only on Opus).

Stop semantics

- Stop is per-trajectory via POST /runs/:id/stop; the server aborts the specific run and marks it stopped. No DB polling.

Notes

- Designed to run inside VM images under pm2 as agent-sync-api (default port 43117, path /operator/agent-sync).
- Providers: Codex via @openai/codex-sdk; Claude via @anthropic-ai/claude-agent-sdk.

Build & publish

- bun run build produces dist/bundle/server.js.
- Package name @dubdubdublabs/agent-sync (public).