Trainer Train

Required inputs

• one selected target file
• the repository root or enough path context to derive the local trainer workspace
• the validation command for the repository
• the concrete stage capability map supplied by the caller
• the currently available specialist-agent roster
• any existing workspace artifacts that should be reused

Core workflow

Follow this order:

1. Resolve the selected target and derive its local trainer workspace.
2. Confirm the prerequisite engineering review artifact exists before training starts.
3. Create or refresh workspace state, source snapshots, required subdirectories, and checkpoint metadata.
4. Reuse existing datasets and eval assets when they fit; otherwise run the missing-data path before optimization.
5. Inspect representative dataset rows before optimization and infer the scoring mode the optimization stage should use.
6. Run at least one optimization pass for the selected target.
7. If optimization produces multiple defensible candidates, run the separate selection stage instead of assuming the optimizer chose a winner.
8. If optimization returns manual follow-up mode, answer the returned model prompt yourself, save the candidate artifact, and continue the loop.
9. Publish steering, candidate bundles, validation logs, and decision summaries under the active iteration.
10. Write the winning candidate back to the source file only when the candidate is defensible and validation passes.

Loop rules

Workspace ownership

• Keep every trainer artifact inside the selected target's local workspace.
• Treat workflow state and required-artifact pointers as checkpoint data, not optional notes.
• Preserve prior iteration artifacts unless the caller explicitly asks to discard them.

Missing-data path

• If train, validation, or authored eval assets are missing, stop optimization and gather them first through the caller-supplied stage capabilities.
• Prefer reuse over regeneration when existing assets already fit the target and scoring shape.
• Keep validation rows as a genuine holdout rather than a paraphrase-only copy of training rows.

Judge-mode inference

• Infer the scoring mode from the row shape before optimization begins.
• Use exact-match scoring only when the rows truly support exact-match grading.
• Use richer scoring whenever the rows depend on structured outputs, normalization, or open-ended judgment criteria.

Manual follow-up recovery

• Treat manual follow-up mode as a supported branch, not as an optimization failure.
• Save the returned report, answer the model-facing prompt, persist the candidate artifact, and continue with steering, validation, and decision-making.

Validation and write-back

• Validate after every meaningful candidate revision.
• Prefer a justified no-op or blocker report over speculative rewrites when prerequisites are missing or the current contract is already fit for purpose.
• Never describe a candidate as accepted until validation passes and the decision summary is written.

Collaboration rules

• The orchestration layer owns sequencing, workspace coordination, and candidate application decisions.
• Research specialists own grounded-source discovery and blocker reporting when source truth is missing.
• Review specialists should critique supplied artifacts and recommend the next steering turn without taking over orchestration.
• Revision specialists should produce bounded candidate changes and explain the reasoning behind them.
• Evaluation specialists should score candidates or trajectories from the supplied artifacts instead of inventing new workflow steps.
• Stress-test specialists should try distinct exploit candidates before finalization when robustness matters.

Steering and candidate rules

• Publish one steering artifact per specialist turn plus rolling summaries for each specialist in the active iteration.
• Stage original, revised, and adversarial candidates as explicit candidate bundles before review or selection.
• Keep cross-run rollups at the workspace root only when they summarize the active best result.

Output contract

Return:

1. workspace status
2. stage decision and any blockers
3. optimization or manual-follow-up status
4. validation status
5. write-back decision
6. next-step guidance when the loop should continue