* docs: add CLAUDE.md for agent onboarding Inherits platform conventions from molecule-core: - Cron discipline and triage rules - Build/test/release commands (pytest, PyPI tag workflow) - Platform integration notes (Phase 30 endpoints table) - SDK-specific conventions (async patterns, token security, tar safety) - Known issues policy (file GH issue before patching silently) Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com> * docs: add known-issues.md and .claude/settings.json - known-issues.md: 5 entries (KI-001 inbound A2A server, KI-002 delegation idempotency, KI-003 symlink silent skip, KI-004 token file races, KI-005 plugin manifest secret scanning) - .claude/settings.json: permissions for pip/npm/go tools, PreToolUse Bash hook, cleanupPeriodDays 30 Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com> --------- Co-authored-by: Molecule AI SDK-Dev <sdk-dev@agents.moleculesai.app> Co-authored-by: Claude Sonnet 4.6 <noreply@anthropic.com>
8.6 KiB
CLAUDE.md — molecule-sdk-python
Project overview
Python SDK for the Molecule AI agent platform. Exposes two user-facing packages:
-
molecule_agent— Phase 30.8 remote-agent client. Write an agent that runs outside the platform's Docker network; it registers with the platform, pulls secrets, sends heartbeats, and detects pause/delete. Wraps the Phase 30.1–30.7 HTTP contract (register, secrets, heartbeat, state-poll, A2A peer discovery, delegation, plugin install). -
molecule_plugin— Plugin-authoring SDK. Build installable plugin directories that ship rules, skills (agentskills.io format), and per-runtime adaptors to any Molecule AI workspace. Ships validators for plugin.yaml, SKILL.md (agentskills.io spec), workspace/org/channel templates, and apython -m molecule_pluginCLI.
Both packages are published together as molecule-ai-sdk on PyPI (setuptools,
pyproject.toml, requires-python = ">=3.11").
Build and test
# Install in dev mode
pip install -e .
# Run the full suite
pytest
# Run only molecule_agent tests (remote-agent client)
pytest tests/test_remote_agent.py
# Run only molecule_plugin tests (SDK + validators)
pytest tests/test_sdk.py tests/test_validators.py
# CLI smoke
python -m molecule_plugin validate --help
python -m molecule_plugin validate plugin /path/to/my-plugin/
python -m molecule_plugin validate workspace /path/to/workspace-template/
Tests use standard pytest fixtures with in-memory mocks — no live platform
required. The molecule_agent tests mock requests.Session directly via
unittest.mock.MagicMock.
Package conventions
molecule_agent/ # Remote-agent client (blocking requests, Phase 30)
client.py # RemoteAgentClient, WorkspaceState, PeerInfo,
# make_idempotency_key, _safe_extract_tar
molecule_plugin/ # Plugin-authoring SDK
protocol.py # PluginAdaptor (runtime_checkable Protocol),
# InstallContext, InstallResult
builtins.py # AgentskillsAdaptor (default),
# SKIP_ROOT_MD, _install_claude_layer
manifest.py # PLUGIN_YAML_SCHEMA, validate_manifest,
# parse_skill_md, validate_skill, validate_plugin
workspace.py # validate_workspace_template, SUPPORTED_RUNTIMES
org.py # validate_org_template
channel.py # validate_channel_config, validate_channel_file,
# SUPPORTED_CHANNEL_TYPES
__main__.py # CLI: python -m molecule_plugin validate [plugin|workspace|org|channel]
template/ # Reference plugin layout (NOT pip-installable)
adapters/
claude_code.py # AgentskillsAdaptor — one-liner per runtime
deepagents.py # AgentskillsAdaptor — one-liner per runtime
examples/remote-agent/ # Runnable Phase 30.1–30.5 demo
run.py
Adding a new tool or endpoint to molecule_agent
- Pick the Phase 30 sub-phase that matches the contract (e.g. 30.6 = peer discovery).
- Add the method to
RemoteAgentClientinclient.py. Follow the existing pattern:_auth_headers()for bearer token,raise_for_status()on the response,logger.warning()instead of re-raising for transient errors in loops. - Add a corresponding test fixture + test cases in
tests/test_remote_agent.py. Mockclient._session.get/.postwithFakeResponseor aside_effect. - Export from
__init__.pyand add to__all__.
Adding a new validator to molecule_plugin
- Add the validation function to the appropriate module (
manifest.pyfor SKILL.md,workspace.pyfor workspace templates, etc.). - Return a list of error strings (manifest layer) or a list of
ValidationErrorobjects (workspace/org/channel layer — see existing patterns inworkspace.py). - Re-export from
molecule_plugin/__init__.py. - Add
python -m molecule_plugin validate <kind> /pathCLI cases or hook into the existing dispatch in__main__.pyif the kind is new. - Add tests in
tests/test_sdk.pyortests/test_validators.py.
Release process
PyPI publication is automated via GitHub Actions and triggered by git tags with
a v prefix matching the version in pyproject.toml (e.g. tag v0.2.1 publishes
molecule-ai-sdk==0.2.1):
# 1. Update version in pyproject.toml
# 2. Tag and push
git tag v0.2.1
git push origin v0.2.1
The GitHub Actions workflow handles sdist + wheel build and upload to PyPI. No manual steps required. Ensure you have PyPI token permissions in the repo secrets before the first release.
Platform integration notes
molecule_agent wraps these Phase 30 HTTP endpoints (all require bearer token
unless noted):
| Method | Endpoint | Phase | Auth |
|---|---|---|---|
POST |
/registry/register |
30.1 | none (issues token) |
GET |
/workspaces/:id/secrets/values |
30.2 | bearer |
POST |
/registry/heartbeat |
30.1 | bearer |
GET |
/workspaces/:id/state |
30.4 | bearer |
GET |
/registry/:id/peers |
30.6 | bearer + X-Workspace-ID |
GET |
/registry/discover/:id |
30.6 | bearer + X-Workspace-ID |
POST |
peer direct URL (A2A) | 30.6 | bearer + X-Workspace-ID |
POST |
/workspaces/:id/a2a (proxy) |
30.6 | bearer + X-Workspace-ID |
POST |
/workspaces/:id/delegate |
30.6 | bearer + X-Workspace-ID, 300s timeout |
GET |
/workspaces/:id/plugins/:name/download |
30.3 | bearer |
POST |
/workspaces/:id/plugins |
30.3 | bearer |
Token is cached at ~/.molecule/<workspace_id>/.auth_token with 0600
permissions. On restart the client reuses the cached token — the platform
refuses to issue a second token when one is on file.
Idempotency (KI-002): delegate() auto-generates an idempotency key as
SHA256(task + current_minute) (rounded to the minute). Two container restarts
within the same minute that send the same task string share the key, preventing
duplicate processing.
Plugin install: Tars are extracted with _safe_extract_tar() — rejects
.. path components and absolute paths; silently skips symlinks/hardlinks.
Atomic rename via staging dir + rename prevents partial installs.
SDK-specific conventions
-
Python:
>=3.11, no external async dependencies inmolecule_agent(uses blockingrequestsso it embeds in any event loop).molecule_pluginadaptor methods areasync(install/uninstallsatisfyPluginAdaptor). -
Async:
molecule_pluginusesasync def/awaitforPluginAdaptor. Callasyncio.run(adaptor.install(ctx))to run inline in a sync context. -
Error handling: Network errors in loops are logged and swallowed so a transient platform hiccup does not take a remote agent offline. API-level errors (4xx) propagate via
raise_for_status(). -
Token security: Token file created with
0o600— other local users must not be able to read it._safe_extract_targuards against tar-slip attacks in plugin install. -
Validation:
validate_manifest/validate_skill/validate_pluginare pure and have no external dependencies (nojsonschema). They return lists of error strings. The workspace/org/channel validators returnlist[ValidationError]objects with.fileand.messagefields. -
First-party plugins:
test_first_party_plugins_are_spec_compliant()intests/test_sdk.pyvalidates every plugin in the repo's top-levelplugins/directory against full agentskills.io spec. Keep that test passing.
Known issues
-
Before patching a silent failure or quirky behavior, file a GitHub issue first. Do not patch silently — the SDK is consumed across multiple runtime environments and silent patches can cause subtle breakage elsewhere.
-
molecule_agentdoes not yet bundle an inbound A2A server helper. Platform-initiated calls to a remote agent without a publicly reachable endpoint will not succeed. See Phase 30.8b in the platform'sPLAN.md.
Relevant platform docs
- Platform conventions:
docs/development/constraints-and-rules.md— no auth for MVP, Postgres as source of truth, no secrets in bundles, generic workspace-template. - Secrets runbook:
docs/runbooks/saas-secrets.md— read before rotating any secrets. - Cron learnings:
cron-learnings.md(platform root) — read before reviewing PRs; write a 1-line reflection to.claude/per-tick-reflections.mdafter triage. - CLAUDE.md/PLAN.md sync PRs: treat these as always noteworthy.
- molecule-core docs: Full platform
PLAN.mdand architecture docs athttps://github.com/hongmingw/molecule-monorepo