| name | evaluation-batch-runner |
| description | Use when: orchestrating an end-to-end J-stock batch evaluation workflow, or when the user explicitly names evaluation-batch-runner. Schedule evaluation-batch-generator, evaluation-batch-executor, and evaluation-result-summary in order, preserving the manifest approval boundary and allowing subagent delegation only during approved execution. |
Evaluation Batch Runner
Purpose
Coordinate the three split J-stock evaluation skills as one higher-level workflow.
Use this skill as the scheduler, not as a replacement for the lower-level contracts:
evaluation-batch-generator: generate the worker-ready manifest and stop for approval.
evaluation-batch-executor: execute an already-approved manifest, monitor boundedly, and report completion.
evaluation-result-summary: aggregate completed output directories into markdown under G:/My Drive/AI-Stock-Sync/summary.
Scheduling Flow
- Start from the earliest missing phase. If no worker-ready manifest exists or the requested parameters changed, enter Generation.
- If the exact manifest has not been approved, do not enter Execution.
- If an approved manifest is running or needs to be launched, enter Execution.
- If execution completed and the user asked for results, comparison, reporting, or end-to-end handling, enter Summary.
- If the user asks for only one phase, route directly to the relevant split skill while preserving the same invariants.
Phase Contracts
1. Generation
- Use
evaluation-batch-generator.
- Let the generator resolve defaults, worker partitioning, and strict
tmp/evaluationtmp.json contents.
- Do not execute any evaluation, replay, or runner command during this phase.
- Present the manifest content and wait for explicit user approval before continuing.
2. Execution
- Use
evaluation-batch-executor.
- Keep the approved manifest unchanged: no repartitioning, no reordering, no silently updated flags.
- Launch one parent runner per approved manifest and keep bounded monitoring until completion unless the user requests leave-running behavior.
- Execution-phase subagents are allowed only after manifest approval:
- Use them for approved launch or monitoring when the run is long, multiple approved manifests need parallel ownership, or the user asks for subagents.
- Delegate only to
evaluation-batch-executor; pass the exact approved manifest path or content and the executor reporting contract.
- Keep the main agent as scheduler of record; reconcile the subagent's
summary.json, output directories, worker status, and warning/error report before final response.
- Do not let a subagent regenerate, edit, reorder, or reinterpret the approved manifest.
- If subagents are unavailable, continue in the current session with
evaluation-batch-executor.
3. Summary
- Use
evaluation-result-summary only after completed output directories or a runner summary.json exist.
- Prefer the executor's local
summary.json or exact completed output directories as the source of truth.
- For multi-output batches, require the summary phase to run
tools/score_evaluation_outputs.py through evaluation-result-summary before interpreting cross-worker rankings.
- Treat the generated global
mdd_win_score as the default cross-worker ranking source; do not rank one-candidate workers by their local prs_train_score.
- Generate the markdown summary artifact under
G:/My Drive/AI-Stock-Sync/summary.
- Do not rerun evaluations, generate manifests, or change source code during this phase.
Approval Boundary
- Preserve the approval boundary even for "end-to-end" requests.
- Stop after Generation unless the user has already approved the exact manifest content to execute.
- Treat approval of a high-level plan as insufficient for execution; execution requires approval of the concrete manifest.
- If the user approves an existing manifest by path, verify that the path and content are the intended approved artifact before Execution.
Subagent Prompt Shape
When delegating approved execution, keep the prompt narrow:
Use evaluation-batch-executor for this approved J-stock manifest.
Manifest path: <path>
Approval evidence: <short reference>
Required behavior: launch, monitor, and report completion for this manifest only. Keep the manifest unchanged. Return parent command, terminal id, local run dir, summary.json path, output dirs, worker status, and warnings/errors.
Do not include generator instructions, summary instructions, or expected performance conclusions in the execution subagent prompt.
Reporting Discipline
- State the active phase and why it is active.
- After Generation, return the manifest approval request from
evaluation-batch-generator.
- After Execution, return the executor's run status and output locations.
- After Summary, return the markdown summary path, any global score CSV/Markdown paths, plus concise findings.
evaluation-progress-monitor is legacy-only. New flows keep monitoring inside evaluation-batch-executor.