Hongming Wang 467d47e3a3 v0.1.1: exponential backoff on platform errors + error-path test coverage

Production fix:
- wait_for_message exceptions now trigger exponential backoff (1s → 60s
  cap, resets to 0 on first successful poll) instead of a flat 1s retry.
  Under platform outage, N daemons under flat 1Hz retry would hammer the
  endpoint unnecessarily; the cap-and-reset shape keeps the daemon
  responsive while being a good citizen.

Correctness gate:
- Test coverage for the six error branches that operators actually hit:
  the backoff progression itself, backoff reset on first success,
  inbox_pop failure (codex must not re-run the same message),
  peer_agent without peer_id (poison drained, not looped),
  unknown message kind (poison drained, not looped),
  empty codex output (placeholder reply, not silent drop),
  canvas_user falling back to workspace_id when arrival_workspace_id
  absent, and four malformed-payload shapes from wait_for_message
  (parametrised: invalid JSON, non-dict, timeout sentinel, missing
  activity_id).
- Backoff tests verified to FAIL on the old flat-1s code by stashing
  only bridge.py and re-running — pinning a real regression, not a
  tautology.

Cleanup:
- _RealTools imports molecule_runtime.a2a_tools once at construction
  instead of four times per message.
- README documents CODEX_CHANNEL_MOLECULE_STATE_DIR override.

Test: pytest -q → 28 passed (was 17).

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

2026-05-04 20:10:17 -07:00

4.5 KiB

Raw Blame History

codex-channel-molecule

Bridge daemon — gives codex CLI push parity with the Molecule AI platform's other external runtimes.

The Molecule platform's hermes-channel-molecule plugin gives hermes-agent true push delivery — peer agents and canvas-user messages land mid-session as conversation turns. Codex CLI has no plugin API today and its MCP runtime drops inbound notifications, so this daemon is the equivalent push surface — built outside the codex process.

How it works

canvas user / peer agent ──► molecule platform inbox
                                       │
                              wait_for_message (long-poll)
                                       │
                                       ▼
                       codex-channel-molecule daemon
                                       │
                              codex exec --resume <sid> "<msg>"
                                       │
                              capture stdout
                                       │
                              send_message_to_user / delegate_task
                                       │
                              inbox_pop(activity_id)
                                       │
                                       ▼
                         canvas chat / peer workspace

Per chat thread (one canvas-user thread or one peer-workspace thread) gets its own codex session_id, persisted to ~/.codex-channel-molecule/sessions.json so daemon restarts don't lose conversation context. Set CODEX_CHANNEL_MOLECULE_STATE_DIR to override the default location (e.g. when running under systemd with a per-instance state dir).

The codex tab wires the molecule MCP server into ~/.codex/config.toml so codex can call platform tools (list_peers, delegate_task, send_message_to_user, commit_memory, etc.). That's outbound — codex calls out to the platform.

This daemon is the inbound counterpart — the platform pushes to codex. Run both for full bidirectional integration.

Install

npm install -g @openai/codex@^0.57
pip install codex-channel-molecule

Configure + run

The same env-var contract as hermes-channel-molecule's outbound MCP path (WORKSPACE_ID, PLATFORM_URL, MOLECULE_WORKSPACE_TOKEN):

export WORKSPACE_ID=<uuid from External Connect modal>
export PLATFORM_URL=https://<your-tenant>.moleculesai.app
export MOLECULE_WORKSPACE_TOKEN=<bearer token from External Connect modal>

codex-channel-molecule

The daemon runs in the foreground; logs go to stderr. For systemd hosts, register a unit; for one-off use, nohup ... & plus a log file works.

Deprecation path

When openai/codex#17543 lands upstream — a generic path for handling MCP custom notifications in codex and forwarding them into the active session as user submissions — this daemon becomes redundant. Codex itself will accept inbound molecule messages as Op::UserInput directly through the MCP server already wired in ~/.codex/config.toml. Until then, this is the operator-facing answer.

Development

git clone https://github.com/Molecule-AI/codex-channel-molecule
cd codex-channel-molecule
pip install -e ".[test]"
pytest -q

Tests are entirely real-subprocess (no mocking the spawn boundary) so the boot path is covered the same way the daemon runs in production.

Releasing

Tag-on-push triggers publish.yml which builds + publishes to PyPI via OIDC trusted publishing (no API token needed).

# Bump pyproject.toml `version`, commit, then:
git tag v0.1.1 && git push origin v0.1.1

The workflow refuses to publish if the tag doesn't match pyproject.toml's version — keeps PyPI versions and git tags in lockstep.

One-time PyPI setup (before the first release):

Create the project on PyPI by uploading the first wheel manually, OR
Pre-register the project on PyPI under a "Pending publisher" config so the first tagged push creates it.

Either way, on the project's PyPI page → "Manage" → "Publishing" → "Add a new publisher", configure:

Owner: Molecule-AI
Repository: codex-channel-molecule
Workflow filename: publish.yml
Environment name: pypi

After this, every git push origin v*.*.* ships the wheel to PyPI without any further intervention.

License

Apache-2.0

4.5 KiB Raw Blame History