Research

research

Purpose

Gathers external market/news context — the one skill in this project with no access to the user's private data at all, by design (see Isolation, below). Output feeds Action Drafting's rationale and supporting_research_refs.

This is the named candidate skill for execution via Managed Agents API, implemented using ADK's native ManagedAgent class (see ADR-0009, which refines ADR-0005 and ADR-0007) — it was chosen specifically because it never touches sensitive data, which is what makes it safe to route through a Pre-GA sandbox in the first place. The orchestrator constructs a fresh ManagedAgent node each planning cycle, with this file's content resolved from the Agent Registry as its description — not statically configured, and not a long-lived instance reused across cycles.

Isolation

This skill has no Firestore or BigQuery access. It receives its research question as a direct input from the orchestrator — it doesn't look up what to research on its own. This is a deliberate boundary, not an oversight: it means this skill's tool surface can never grow to leak private financial data into an external call, regardless of how its prompt or reasoning evolves.

When this skill runs

• Root planner invokes it before Action Drafting, when a candidate action would benefit from external context (market conditions, recent news for a specific ticker)
• User asks a market-conditions question directly

Inputs

Output: Research Brief

Not a persisted contract — transient, like Portfolio Analysis's Drift Report. Cited by ID from ProposedAction.supporting_research_refs (see proposed-action.schema.json), so it needs a stable identifier even though the brief itself isn't stored long-term.

Failure mode: no reliable data found

If search/market-data tools don't return anything conclusive, the brief says so explicitly (confidence: low, summary states the gap) rather than presenting speculation as fact. Action Drafting is expected to reflect that uncertainty in its own rationale, not paper over it.

Tools / permissions required

• google_search (ADK/ManagedAgent built-in, server-side) — no custom MCP server. ManagedAgent doesn't support MCP tools at all; see ADR-0009 for why this is sufficient for what this skill actually needs, not just a limitation worked around.
• No Firestore, no BigQuery, no trade-execution tools of any kind. This is the strictest tool surface of any skill in this project.

Search budget

• At most 3 google_search calls per invocation. If 3 searches do not produce enough evidence, return the brief with confidence: low and an explicit gap statement in summary. Do not keep querying.
• Prefer one broad search followed by at most two follow-ups scoped to named tickers, sectors, or dates from the first result. Do not run synonym-variant queries of the same question.
• One search per distinct sub-question. If two sub-questions can be answered by the same search, run it once.

Registry metadata

• Registered as: projects/{project}/locations/{location}/skills/private-research
• Skill revision: 0.1.0 (draft — not yet registered)
• Approval scope: read:external_market_data

Acceptance criteria

1. This skill's tool surface never includes any internal (Firestore/ BigQuery) read or write scope — enforced at the tool-list composition level, not just documented intent
2. Inconclusive research produces confidence: low with an explicit gap statement, not a fabricated conclusion
3. Every ResearchBrief has a research_run_id that Action Drafting can cite verbatim in supporting_research_refs
4. The ManagedAgent node is constructed fresh each planning cycle, with description resolved from the Agent Registry at that moment — not a long-lived instance built once and reused. A revoked skill should be absent from the next cycle's construction, verified the same way as the project's other live-revocation behavior.