cta-research-loop

name	cta-research-loop
description	Use for CTA research inside Aura when the task is to inspect archived runs, review evaluation artifacts, generate the next structured proposal, or advance approval-gated archive state. Do not use it for generic coding or to invent unsupported execution capability.
allowed-tools	["cta__artifact_find","cta__artifact_load","cta__archive_candidate","cta__archive_lineage","cta__proposal_generate","cta__approval_list","cta__approval_resolve","cta__apply_approved_transition"]

CTA Research Loop

Use this skill for persisted CTA research state, proposal generation, and archive governance inside Aura.

Quick Routing

• Need exact prior state or payloads: read references/tool_contract.md, then use artifact and archive tools.
• Need the next experiment or strategy iteration: load the latest StrategySpec, ExperimentSpec, and optional EvalReport, then call cta__proposal_generate.
• Need promotion, freeze movement, or approval handling: use approval tools first, then materialize the archive transition.
• Need to archive a promising completed run as a new candidate: use cta__archive_candidate.
• Need a fresh backtest run: read references/current_limits.md first. If no deterministic harness exists, stop and propose or implement the missing assembler instead of fabricating execution.
• Need a concrete example of how to execute one of those tasks: read references/research_playbook.md.

Workflow

1. Anchor on persisted CTA state, not chat memory.
2. Load only the specific artifacts or archive records required for the current decision.
3. Emit structured outputs: proposals, approvals, and archive transitions stay machine-readable.
4. Respect the current execution boundary before suggesting or claiming a new run.
5. Do not recursively scan .archive/, .aura/, or large data directories with generic file listing when restoring CTA state; use CTA tools and targeted refs first.

Response Contract

• Default to a compact structured answer with four parts: state_summary, evidence, recommended_action, blocked_by.
• In evidence, cite concrete artifact refs, archive ids, approval ids, or report ids actually loaded in this turn.
• In recommended_action, distinguish between proposal, approval_decision, archive_transition, and implementation_gap.
• If the task is blocked by missing execution infrastructure, say so explicitly and name the missing deterministic component.

Rules

• Treat CTA artifacts, archive records, and approvals as the source of truth.
• Keep outputs structured; do not collapse proposals or governance state into free-form prose.
• Do not bypass approvals when archive state changes are requested.
• Do not claim there is a generic StrategySpec -> executable strategy module path unless the repo has added one.

References

• Read references/tool_contract.md for task-to-tool mapping and payload expectations.
• Read references/current_limits.md when the task touches fresh execution, generic strategy compilation, or runtime boundaries.
• Read references/research_playbook.md for concrete end-to-end task patterns.
• Read references/response_contract.md when you need exact output shapes for state review, proposal, governance, or blocked execution tasks.