tcx-workflow

TCX Workflow

Use the smallest safe path. Coordinate rather than impersonate specialists.

When a named tool is deferred, use one names-only exact-name lookup, then inspect the selected exact name's schema once. Never print full tool records, scan descriptions, or repeat a schema lookup.

Fast Path

Answer a narrow fact, definition, or status from a trusted read-only source. Do not start a run, artifact, or child to restate it. State unsupported gaps.

Start a workflow only for research, valuation, forecasts, recommendations, portfolio or risk judgment, order preparation, approval review, or a question that needs fresh evidence or more than one distinct expertise.

Workflow

1. Preserve the user's outcome, constraints, and exclusions. Ask only when a missing choice changes the result or a sensitive action.
2. For a workflow, call begin_analysis_run once. It seals provenance but creates no child. Never wait for it: either dispatch a chosen role through spawn_agent or stop.
3. Apply one explicitly selected Investment Brain or Strategy only as sealed context. Do not infer, blend, inspect, or change one during the run. When an Investment Brain, Strategy, or Decision Memory source applies, read Context Overlays before applying it or dispatching children.
4. For fresh research, valuation, forecasts, recommendations, or material portfolio/risk judgment, read the Research Framing playbook, find the smallest decision-relevant unknowns, and choose the smallest useful role set. Skip the playbook for a narrow fact or recorded order, approval, or execution status. Dispatch only for distinct expertise or independent challenge. Use risk-manager where its separate risk authority materially improves the result. Use judgment-reviewer when a recommendation, portfolio/risk decision, material conflict, or high-consequence uncertainty benefits from adversarial adjudication; do not force it onto narrow facts or requests without a meaningful competing case.
5. Use an exact fixed role when one is available. Only an unavailable evidence-producing role may use a generic child for the same bounded brief and no-order boundary; it cannot approve, execute, access secrets, or act as a broker. Do not replace an independent risk-manager or judgment-reviewer review with a generic child: return an explicit profile gap as blocked or waiting. Call spawn_agent for an exact profile with its agent_type, a compact message, task_name, and fork_turns="none"; let its role TOML supply the fixed model settings. Success requires a live target; correct a rejected spawn only from its error.
6. Before refetching a reusable data family, check only current-workflow Snapshot/Dataset candidates. Give it one owner and brief every child with that ownership, exact reusable IDs, and the missing slice. Require reuse of supplied IDs; no child recollects another owner's complete family. Tell the owner to load $tcx-source-gate and return IDs, not raw output.
7. Reuse a live child's session with followup_task for work it still owns. Start another only for a new specialty, unavailable session, or independent review; parallelize independent questions. Never claim activity absent from native tool and child-lifecycle results in this run.
8. Call wait only after spawn_agent or followup_task has returned a live target in this run and only while that child has useful work. Otherwise do not call it. Update only on observable material change; a timeout is not progress.
9. Save an authenticated role artifact when evidence changes a conclusion, supports a consequential judgment or handoff, records a material conflict or gap, or has future reuse or audit value. Do not save a narrow fact, status, temporary exploration, or restatement of one existing artifact.
10. After each accepted result, apply the Feedback Loop below. Synthesize only accepted, authenticated, run-local artifacts.