Iterate

This vNext run compatibility manual owns experiment semantics in iteration_log.json; .harness_runtime/ owns orchestration and recovery.

Use this only for WF10 experiment-loop state. It owns iteration_log.json; it never writes stage transitions into PROJECT_STATE.json or .auto_iterate/**.

Read First

• Workflow, context, contract, lesson, code-style, review, language, docs, and commit checkpoint rules under ../../../.agents/references/
• Research supervision assets: research-supervision-patterns.md, research-supervision/experiment-and-build-canvas.md, and research-supervision/ai-assisted-research-workflow.md
• ../../../.agents/references/run-artifact-contract.md
• ./references/iteration-log-schema.json, ./references/iteration-context.md, ./references/iteration-constraints.md
• iteration_log.json, PROJECT_STATE.json, CLAUDE.md, evaluation/baseline contracts, glossary, and lessons when present

State Rules

• iteration_log.json is the experiment source of truth.
• New or migrated logs use schema_version: "2" with per-iteration action_state and implementation.
• Stable file/interface changes sync project_map.json when present.
• Auto-iterate may invoke phases through a runtime adapter, but it only reads iteration_log.json postconditions.
• Active context belongs in .agents/state/iterations/<iter-id>/context.json plus .agents/state/current_iteration.json only when needed.

Preflight

Load active iteration plus 5 recent summaries; reference full history and Gate ledger by path. Resolve latest id, status, decision, config_diff.planned_command, target script, git_commit, run_manifest, and metrics/report availability.

Scan auto_paper_output/*/run_request_register.{json,md} and fold the highest priority unclosed request into planning unless a higher-priority WF10 blocker exists.

Recommend exactly one next command unless asked for a playbook. Prefer action_state.next_action; for legacy v2 logs, map status: planned|coding -> code, ready_to_run -> run_screening|run_full, running|ready_to_eval -> eval, needs_debug -> debug, needs_more_evidence -> compare, candidate_for_promotion|promoting -> promote, completed/NEXT_ROUND -> plan, terminal -> stop.

If run-local code becomes reusable, recommend $iterate promote or $change classify before merging into stable src/, scripts, tests, or project_map.json.

Commands

• next: execute one action_state.next_action, update last_action, next_action, reason, and blocked_by, then stop.
• plan: refuse ordinary new work while an iteration is coding or running; allocate ID; check lessons; record hypothesis, files, config_diff, expected effect, screening recommendation, dominant improvement axis, falsifier, assurance axis, minimum artifact, claim/figure implication, canonical codex_review, and implementation scope: config_only, run_local_code, stable_candidate, or delegated_build. Read docs/context/experiments.md when present and consume one queued item or explain the priority choice. Use legacy docs/40_iterations/Experiment_Queue.md only when the project has not migrated to dynamic-context-v2.
• code: select latest planned iteration, write context, apply code-style checklist, preserve slice/glossary boundaries, route implementation through $build, and run a slice or experiment commit checkpoint before status ready_to_run. For run_local_code or stable_candidate, write runs/wf10/<iter>/code_manifest.json.
• run_screening / run_full: select latest ready_to_run; resolve Train/Eval scripts from CLAUDE.md; run config_diff.planned_command exactly when present; verify run-local config paths; create or verify a Semantic Execution Commit covering stable code, configs, eval logic, and run-local code/configs; record run_manifest, pre_train_commit, artifacts, screening result when relevant, and protocol metrics. Set ready_to_eval, needs_debug, or needs_more_evidence; never invent metrics.
• register: record externally executed/manual run paths and expected outputs without inventing metrics; set ready_to_eval only when the bundle is complete.
• debug, compare, ablate: operate inside the same iteration unless an explicit ablation command creates sub-iterations; update action_state.
• eval: refresh context; invoke $evaluate when useful; compare baseline, previous, and best metrics; create or verify pre_eval_commit, or record pre_eval_commit_NOT_CHANGED; record metrics or documented failure, decision, lessons, assurance axis, claim delta evidence, slice/drift/complexity notes, reports, and completion state. Write mutable observations to docs/context/experiments.md, including concrete next experiments, mutable observations, and stable searchable findings, or report NOT_RUN. Legacy docs/45_discoveries/Discovery_Ledger.md, docs/40_iterations/Experiment_Queue.md, and docs/45_discoveries/Research_Wiki.md may be updated only for not-yet-migrated projects. Per-Iteration Light Evidence is recorded through the Campaign controller (harness campaign evidence), not by hand.
• promote: read implementation.code_manifest_path, write/verify implementation.promotion.plan_path, run acceptance commands or report NOT_RUN, merge only approved candidate code into stable surfaces, and update project_map.json / Codebase_Map.md when stable structure changes.
• discard, status, log: preserve canonical schema and history behavior.

Lesson And Gate Rules

• Append MEMORY.md only for accepted lessons satisfying lesson-quality-rule.md; raw observations stay in iteration reports or docs/context/experiments.md; candidates may be promoted into docs/context/memory.md.
• When iteration_log.json, lesson files, accepted memory, reports, queue/wiki docs, claim delta evidence, or evidence indexes change, report a Gate ledger.
• Run check_workflow_state.py near WF10 handoff points; for routine in-loop updates, report whether it was run or deferred.
• After Grill establishes an Automation Policy, WF10 auto-proceeds within that policy. Do not ask for human approval during run/eval loops unless the action would leave the policy, run an explicit approval tool, or perform an irreversible external submission.
• Use commit hashes, Gate ledgers, Claim Delta Evidence, Experiment Queue, and Research Wiki updates for traceability.

Durable Docs Render

temporary drafts.