method-designer

name	method-designer
description	Turn a selected core idea unit into a per-program method design handoff with repo choice, interfaces, and an expanded experiment matrix.

Method Designer

开始设计前遵循 workspace 统一 task-scoped preference 合同，记录并加载 method-designer + design 的 effective selection。canonical task context 绑定 program、selected idea 与 idea substance digest。资源与约束是 hard 边界，无 receipt 时仍直接从 canonical profile 读取；research focus 等 soft 项只有被 Agent 选中时才影响本次方案。脚本把 selected research focus 仅作为候选提案 token scoring 的附加输入，不把排序升级成方法判断。

协议参考：.agents/lib/research/SCHEMAS.md#unit-record · #program-files · #confirmation-gate · #runtime

Use this skill only after an idea has been explicitly selected.

Workflow

1. Read the selected idea record and its analysis artifacts.
2. Refuse to design from an unselected idea.
3. Read kb/programs/<program-id>/state.yaml and rank repository units from active_unit_ids; use KB-wide repos only when the program has no attached repo units, and state that fallback explicitly.
4. Read kb/config/user-profile.yaml → resources, preserve those declarations in program state, and scale the experiment matrix skeleton to the declared capacity.
5. Prepare candidate_repos, proposed_repo_id, empty canonical judgement slots, interfaces, and an expanded experiment matrix. The backward-compatible design operation is prepare-only.
6. Do not write selected_repo_id, advance the program stage, or emit a method-selection event during prepare.
7. Have the runtime agent fill four canonical claims in the repo-choice artifact: method-repo-selection, method-interfaces, method-baselines, and method-risks. Each is a judgement-class claim with verbatim evidence; the repo-selection claim must name and cite the proposed repo unit.
8. Verify the four claims and their evidence. A successful verify creates a current byte-bound verification receipt and makes the side judgement discoverable by public review, but still does not select the repo or advance the program.
9. Only after the user's explicit current-message authorization, run the private confirm-selection route. Confirmation atomically writes selected_repo_id, advances to implementation-planning, updates dependent artifacts, and emits one method-selected event with the current ConfirmationReceipt binding. If the user rejects the displayed proposal, use the private reject-selection route; it closes that review subject without selecting a repo, advancing the stage, or emitting a method-selected event.
10. Hand only a confirmed run grid to experiment-workbench.

Shared Contract

• Repo choice is a durable method_selection side judgement with payload.claims, payload.verification, optional confirmation, and an internal review route. Public review discovers only its ready_for_review state.
• Candidate ordering may use deterministic record signals such as token overlap, but the script must not turn that ordering into a method judgement. Before verification, canonical claims are empty and status is needs_agent_fill, not ready for human review.
• Prepare 将 task digest、可选 selection binding 与 hard-value digests 写入 repo choice、interfaces、matrix 和 program proposal state；不复制 soft preference 正文。Verify 重新计算 task/context/catalog，wrong binding、idea 变化或偏好 stale 都 fail closed。
• proposed_repo_id and selected_repo_id are separate fields. The latter must be absent from the repo choice, interfaces, matrix, and program state until confirmation succeeds.
• Interfaces should expose edit surfaces, config keys, metrics, and artifact expectations instead of hiding them in prose only.
• The experiment matrix should cover baseline parity, minimal variant, ablation, and stress/failure slices.
• Resource scaling is structural arithmetic, not material understanding: the script may set seed count, model-size tier, parallelism, GPU requirement, and feasibility from declared resources; the runtime agent supplies method rationale and evidence.
• With no declared or parseable resources, preserve the legacy four-row matrix scale and mark feasibility unknown rather than inventing capacity.
• Prepare/verify/confirm/reject are independent mutation transactions followed by exact-path checkpoints. First-time prepare targets the new design directory so an abort cannot leak an empty directory; an existing directory uses only exact artifact paths.
• Only confirmed method selection emits a reporting event. Its binding includes the side subject kind/id/owner/path, claim ids, content digest, and verification digests.

Resource Profile Shape

resources is a mapping under kb/config/user-profile.yaml. Values may be structured scalars or free-form statements captured by research-config-manager:

resources:
  gpu_count: 4
  local_gpu: 1x RTX 4090 with 24 GB VRAM
  cluster: 8x A100 GPUs with 80 GB each
  time_budget: 48 GPU-hours per week

The designer conservatively recognizes explicit GPU counts, common Nx GPU-model statements, GPU memory in GB, and CPU-only/no-GPU declarations. Unrecognized statements are retained verbatim in state.yaml as resource_constraints; they do not trigger guessed capacity.

Each experiment row includes:

scale:
  seed_count: 3
  model_size_tier: repo-default
  parallelism: 1
  required_gpus: 1
feasibility: unknown | feasible | unrealistic
status_color: gray | green | red
feasibility_reason: ...
resource_request: ...

Repo-choice artifacts also record candidate_corpus.scope, repo_ids, fallback_used, and a human-readable note. Program-attached repo units are the primary corpus; KB-wide ranking is only the empty-corpus fallback.

Canonical Method Claims

The runtime agent writes the four claims directly to payload.claims; no sidecar claim list is a confirmation source. Every claim follows the shared evidence schema and remains pending_user_confirmation through verification:

payload:
  claims:
    - id: method-repo-selection
      text: "The proposed repository id must appear verbatim here."
      claim_type: evaluation
      confirmation_status: pending_user_confirmation
      evidence_refs:
        - source_unit_id: "<the proposed repo unit id>"
          artifact: record.yaml
          locator: "<section, line, or record field>"
          quote: "<verbatim evidence>"
    - id: method-interfaces
      # same shared claim/evidence shape
    - id: method-baselines
      # same shared claim/evidence shape
    - id: method-risks
      # same shared claim/evidence shape

Empty claims, missing required claim ids, fact/unverified types in a required judgement slot, fabricated quotes, stale evidence bytes, and claim edits after verify all fail closed. Old artifacts that already contain an unconfirmed selected_repo_id require explicit migration and are never auto-promoted.

Evidence Boundary

• The script assembles paths, interfaces, candidate order, run-grid scale, and deterministic capacity checks.
• The runtime agent fills repo-selection claims, baseline rationale, method substance, and evidence.
• A design artifact must not claim that a repo, baseline, mechanism, or experiment is methodologically correct solely from idea text or lexical overlap.
• Interface, baseline, and risk fields are structural proposal slots. Their substantive rationale lives only in the four canonical claims and cannot be treated as complete before verification and user confirmation.

Private Execution Boundary

The runtime agent uses the implementation's private prepare, verify, and confirm-selection routes. Never expose their script paths, flags, environment variables, or artifact paths to the user. User-facing responses explain the proposal and review state in natural language and may offer only a public kb <verb> next action.

脚本入口：scripts/method.py（design --phase prepare|verify / confirm-selection / reject-selection）。

启动澄清（Agent 用）

• 目标 program 与已显式选中的 idea 是哪个？未选先回 idea 选择。
• 资源边界有变化吗？默认读 profile 的 resources。
• 候选 repo 有人工倾向吗？默认按 program 附挂 repo 排序提案。