feat(agent-bundles): au-quant-8 multi-agent quant team bundle#691
feat(agent-bundles): au-quant-8 multi-agent quant team bundle#691TatsuKo-Tsukimi wants to merge 1 commit into
Conversation
![chatgpt-codex-connector[bot]](https://avatars.githubusercontent.com/in/1144995?s=60&v=4){kind=small}
There was a problem hiding this comment.
💡 Codex Review
Here are some automated review suggestions for this pull request.
Reviewed commit: 2921e71be3
ℹ️ About Codex in GitHub
Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you
- Open a pull request for review
- Mark a draft as ready
- Comment "@codex review".
If Codex has suggestions, it will comment; otherwise it will react with 👍.
Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".
| # the ``{{agent_name}}`` template placeholders). | ||
| agent_dir = _agent_dir(agent.id) | ||
| soul_path = agent_dir / "soul.md" | ||
| soul_path.write_text(ba.soul_md or "", encoding="utf-8") |
There was a problem hiding this comment.
When STORAGE_BACKEND is S3 or another non-local backend, initialize_agent_files() writes the scaffold into the storage backend and does not create/update this local path. This Path.write_text() therefore either raises FileNotFoundError during bundle hire or only mutates ephemeral local disk, while build_agent_context() and container materialization read soul.md and skills from storage; hired bundle agents can keep the scaffolded soul and miss bundle skills. Use the storage backend under the agent prefix for the bundle soul/skill copies.
Useful? React with 👍 / 👎.
| .where( | ||
| ChatMessage.participant_id == from_participant, | ||
| ChatMessage.created_at > since, | ||
| ChatMessage.role == "assistant", |
There was a problem hiding this comment.
When an agent sends via send_message_to_agent or send_file_to_agent, the A2A ChatMessage is stored with role="user" and the source agent's participant id (see agent_tools.py around the A2A message/file inserts). Adding this assistant-only predicate makes on_message triggers seeded from bundle relationships ignore the actual A2A sends/files they are meant to wake on, so daemon fallback/callback triggers stay silent unless a separate assistant-role row happens to match. Include participant-attributed user rows or remove the role filter.
Useful? React with 👍 / 👎.
8-agent Shanghai-gold (AU) quant team — bull/bear researchers, research manager, trader, three risk analysts, risk-committee chair — with one-shot hire, per-team paper-trading isolation via Bearer api_key, zh + en bundle variants, plus supporting backend (bundle seeder, hire flow, MCP exec authz, alembic migrations) and frontend (hire modal, talent market). MCP endpoints are placeholders (YOUR_HOST / YOUR_DATA_HOST); set them to your own deployment before seeding.
TatsuKo-Tsukimi
force-pushed
the
feat/agent-bundles
branch
from
2921e71 to
110cf17
Compare
There was a problem hiding this comment.
💡 Codex Review
Here are some automated review suggestions for this pull request.
Reviewed commit: 110cf17c79
ℹ️ About Codex in GitHub
Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you
- Open a pull request for review
- Mark a draft as ready
- Comment "@codex review".
If Codex has suggestions, it will comment; otherwise it will react with 👍.
Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".
| agent_id=to_id, | ||
| name=trigger_name, | ||
| type="on_message", | ||
| config={"from_agent_name": from_name}, |
There was a problem hiding this comment.
For bundle graphs where one source agent has edges to multiple recipients, this trigger config only records the sender name. I checked check_new_agent_messages, and the lookup only filters by the sender participant and excludes trigger sessions; it does not constrain the A2A session to this trigger's recipient. As a result, the first message from that sender to any agent can fire every seeded trigger for that sender, and because these bundle triggers are capped at max_fires=1, the intended recipient-specific fallback can be consumed before the actual edge message arrives.
Useful? React with 👍 / 👎.
| f"{self.FEISHU_CONTACT_URL}/{open_id}", | ||
| params={"user_id_type": "user_id"}, |
There was a problem hiding this comment.
When the OAuth user_info response supplies an open_id, this request puts that open_id in the URL but declares user_id_type=user_id. The existing Feishu webhook resolver in this repo uses the same contact endpoint with user_id_type=open_id for an open_id path, so this call will fail to populate contact_user; SSO then misses the stable employee user_id/authoritative contact fields and can fail to link org-synced members keyed by external_id.
Useful? React with 👍 / 👎.
1 participant
What
Adds au-quant-8, an 8-agent Shanghai-gold (AU) futures quant team bundle, plus the supporting agent-bundle system (one-shot hire, per-team paper-trading isolation, bundle i18n).
Agents (zh + en variants)
Backend
agent_bundles/bundle format (bundle.yaml / mcps.yaml / relationships.yaml / per-agent soul.md + skills)agent_bundle_hire.py), bundle i18n / locale-scoped Talent MarketFrontend
Config
MCP endpoints in each bundle's
mcps.yamlare placeholders (YOUR_HOST/YOUR_DATA_HOST) — set them to your own deployment before seeding.Notes
mainand is currently behind upstream; it will need a rebase before a clean merge (a few backend / i18n files conflict with recent upstream changes). Opening now to start the discussion / let CI run.dummy-*bundles used as test fixtures — happy to drop them if you'd prefer them out of the contribution.🤖 Generated with Claude Code