linzhe001/iterate

Iterate

References

Read these first:

• ../../../.agents/references/workflow-guide.md
• ../../../.agents/references/language-policy.md
• ./references/iteration-log-schema.json
• ./references/iteration-context.md
• ./references/iteration-constraints.md
• ../../../iteration_log.json
• ../../../PROJECT_STATE.json
• ../../../CLAUDE.md

When To Use

Use this skill for WF8 structured experimentation.

Interpret natural-language requests as one of these canonical intents:

• plan
• code
• run
• eval
• ablate
• status
• log

State Ownership

• iteration_log.json is the only experiment source of truth.
• Do not write stage transitions into PROJECT_STATE.json; $orchestrator owns those.
• Stable interface changes still require project_map.json sync through $code-debug.

Controller Coexistence

• When the auto-iterate controller is active, $iterate phases are invoked via runtime adapter prompt, not directly by the user.
• iteration_log.json ownership is unchanged — $iterate still owns it exclusively.
• The controller only reads iteration_log.json via postcondition validation; it does not write to it.
• The controller does not write to .agents/state/**.
• .auto_iterate/ is controller-owned; $iterate must not write to it.

Canonical Workflow

Startup Cleanup

• Before any sub-command, inspect active iteration context files.
• Prefer .agents/state/current_iteration.json.
• If interrupted state is found, apply the canonical cleanup behavior from the Claude prompt.
• Follow ./references/iteration-context.md for deterministic cleanup.

plan

1. Ensure there is no blocking unfinished iteration.
2. Allocate the next iteration ID.
3. Check prior lessons to avoid repeating known failed ideas blindly.
4. Record hypothesis, changes summary, config diff, and screening recommendation (screening.recommended as a structured boolean field).
5. Preserve the canonical codex_review field behavior in iteration_log.json.

code

1. Select the latest planned iteration.
2. Write canonical iteration context to .agents/state/iterations/<iter-id>/context.json.
3. Mirror active context to .agents/state/current_iteration.json.
4. Follow ./references/iteration-context.md when creating both the persistent and active context files.
5. Invoke $code-debug.
6. Require a semantic commit before moving from coding to training.

run

1. Select the latest training iteration.
2. Build the training command from CLAUDE.md entry scripts and the config.
3. Resolve the tracked metrics from the baseline or evaluation protocol established in WF5 instead of hard-coding PSNR/SSIM-style keys.
4. Record run metadata in run_manifest.
5. Preserve the canonical error handling and manual-mode fallback behavior.

eval

1. Select the latest runnable iteration.
2. Write or refresh active iteration context.
3. Follow ./references/iteration-context.md when reusing an existing context file.
4. Invoke $evaluate when full analysis is needed.
5. Compare against baseline, previous iteration, and best iteration using the tracked metric set inherited from WF5.
6. Record:
   • metrics
   • decision
   • lessons
   • completion state

ablate, status, log

• Preserve the canonical behavior and schema for ablations, summaries, and full history views.

Codex Adaptation

• Treat natural-language requests as the canonical $iterate {plan|code|run|eval|ablate|status|log} interface.
• When the canonical workflow calls for code or analysis substeps, use $code-debug and $evaluate.
• Use .agents/state/ as the local active-context directory; create context files inside it only when needed.
• Use ../../../.agents/references/language-policy.md for reply language and for localizing natural-language iteration summaries; keep schema keys, commands, and decision/status tokens unchanged.

Execution Rule

Follow the local iteration logic and schema rather than replacing it with a generic experiment log loop. ./references/iteration-constraints.md is mandatory behavior, not optional guidance.