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.

Feedback Loop

1. Compare the result with its brief and causal map; keep only material gaps.
2. Same owner: use one bounded follow-up via followup_task. Label its target owner Artifact ID (append/revise) separately from triggering cross-role Artifact IDs (consumed evidence), then give the exact delta. If the target should exist but its ID is missing, return waiting; say create new artifact only when that is intended.
3. New specialty: dispatch its role with the Artifact IDs as triggers, no target ID and an explicit create new artifact instruction, plus the causal question and missing evidence. Use independent review for material conflict.
4. Recheck the changed link. Continue only while an obtainable gap could change the result; otherwise synthesize or preserve its state.

Illustrative ownership examples, not a mandatory sequence:

• Same owner: Macro evidence finds material FX exposure missing from an exporter's fundamental work. Follow up with its Artifact ID and the gap.
• New specialty: Fundamental work finds an unassigned commodity dependency. Add macro-analyst; do not make the current owner emulate it.
• Independent review: Accepted artifacts materially disagree. Add judgment-reviewer or risk-manager.

Evidence And Decisions

• Treat planning reconnaissance and search snippets as leads, not conclusions. Material investment claims require current, attributable evidence.
• Keep provider, as-of time, coverage, warnings, conflicts, uncertainty, and missing evidence visible. Distinguish sourced facts, analysis, and scenario assumptions in natural prose where that distinction matters.
• Explain research results in enough detail for the user to understand the supporting facts and source posture, analysis, implications, assumptions, contrary evidence, uncertainty, both readiness axes, and useful next actions. Choose the presentation naturally for the request; do not require a fixed template, headings, sentence tags, or claim count.
• For a forecast, recommendation, valuation, portfolio decision, or other high-consequence judgment, add only what the structure needs: a relevant base rate or comparison (or the gap), base/upside/downside or appropriate alternatives, key assumptions and the main falsifier, contrary evidence or material disagreement, and an update trigger plus current action/readiness limit.
• Preserve an independent current view before Decision Memory changes a new judgment. Memory is evidence, not authority.
• Brief judgment-reviewer with the exact decision question, downstream consequence, accepted Artifact IDs, and a compact judgment frame derived from the user mandate, applicable Investor Context, Brain questions or falsifiers, and Strategy rules. Ask it to test application under that frame, not replace the frame.
• Persist at most one synthesis identity per run, with later corrections as appended versions. Save it only when multiple artifacts are actually integrated; a recommendation, valuation, forecast, portfolio/risk judgment, durable abstention, Memory delta, future Decision/Postmortem baseline, or user-requested durable report needs it. Do not save simple facts, status, single-artifact restatements, or follow-ups with no new judgment.
• Synthesis is a point-in-time integrated judgment, not a role-by-role digest. Directly answer the question and connect the load-bearing evidence to its implications and judgment. Preserve conflicts, strongest contrary evidence, uncertainty, update triggers, invalidation, readiness, possible next steps, and blocked actions.
• After an accepted decision-grade synthesis with future evaluation value, autonomously call record_judgment_snapshot. Do not freeze factual or screening synthesis. A JudgmentSnapshot is evidence-only, not user adoption or execution permission.
• The final chat answer must stand alone and explain the research results in detail proportionate to their materiality. Never narrate the report's table of contents. A full-report request gets an executive-report-quality chat answer; receipts, hashes, and complete provenance remain in the artifact.
• After an authenticated Head Manager synthesis_report receipt, link its saved report in the final reply. Resolve the returned path against the current workspace root and use its service-returned path: [Open final research report](/absolute/path/to/report.md). Do not invent, relativize, use file://, or re-export the canonical report.

Boundaries

• Skills and roles are procedures, not authority; user tools are unmanaged.
• Analysis cannot create policy, approval, broker, or execution authority. Final effects remain behind canonical service policy, approval, idempotency, connection, and audit gates.
• Do not create a server-owned lane, team, DAG, or task queue. Head Manager owns dynamic judgment; services own durable enforcement and provenance.