select（select 搜索）

Select

Intent

Pick a task source, refine it into dependency-aware atomic tasks, schedule parallel waves when safe, and emit an orchestration plan.

This skill is plan-only:

• It does not implement changes (no code edits; no running workers).
• It does not mutate the task source (no SLICES.md writes, no plan-N.md writes); instead it emits explicit manual writeback steps when needed (e.g., "mark as in_progress").
• It is orchestration-agnostic: the output is a neutral plan schema (not tied to a specific executor).

It may also emit a small pipeline for driving planning artifacts into execution (manual steps; optionally loopable).

Swarm-ready planning

$select is optimized for parallel multi-agent execution, so it prefers plans that are explicit, decomposed, and lock-safe.

• Explore the codebase (read-only) when needed to ground tasks in real paths/components and to set tight scope locks.
• Ask clarifying questions when multiple reasonable approaches exist; include a recommended default.
• Atomic tasks: each task should be independently executable by a single worker.
• Workstream shaping: identify major workstreams first, then map atomic tasks into those workstreams.
• Role-aware shaping: use role labels when useful (contract, implementation, integration, checkpoint) to improve sequencing and review quality.
• Explicit dependencies: prefer explicit edges over relying on implicit serialization via overlapping scope.
• Delegation metadata: include scope (required for safe parallelism), plus location and validation whenever possible.
• Unit scope quality: mark scope as invalid when missing/unknown/overly broad; invalid scope units are not parallel-eligible.
• First-ready claim quality: when emitting waves[0], include the full safe first wave only after dependency and scope checks pass.
• Review before yielding: run a separate reviewer-mode pass for missing deps/order/lock overlaps/validation gaps.

Decomposition heuristics (parallelism quality)

When the selected source is coarse or linear, refine task structure before scheduling waves:

• Identify workstreams and keep task descriptions scoped to one workstream where possible.
• Create explicit contract tasks for API/schema/interface/config decisions that unblock multiple downstream tasks.
• Create explicit checkpoint/integration tasks as join points after parallel branches.
• Keep medium granularity: each task should be independently PR-able.
• Prefer explicit dependency edges to document true prerequisites and avoid accidental serialization.

Dependency semantics (hard vs soft)

Use two dependency channels with distinct meaning:

• depends_on: hard prerequisite; contributes to DAG readiness and wave scheduling.
• related_to: soft ordering/context edge; never gates readiness or wave eligibility.

If ordering is uncertain or advisory ("nice first", "reduces rework"), prefer related_to over depends_on.

Invocation directives (optional)

If present, interpret these directives from the invocation text:

• mode: both|triage|new

- both (default): triage in_progress first; then select new work. - triage: only triage in_progress and recommend close/reopen/continue. - new: skip triage and select new work (still warn about in_progress).

• max_tasks: auto|<int>

- If omitted: default auto. - For slices, auto means: keep any carried-forward in_progress task and add every additional dependency-independent task that safely fits in waves[0]; it may still resolve to 1 when scope or dependencies only permit one task. - Applies after triage decisions.

• review: required|auto|off

- required (default): run a reviewer pass and iterate until it passes (or stop+ask if blocked). - auto: run one reviewer pass; fix what you can; proceed with remaining warnings. - off: skip the reviewer pass.

Source precedence

When multiple sources exist, pick exactly one using this precedence:

1. Explicit user-provided task list in the invocation text.
2. SLICES.md (if present and parseable).
3. plan-N.md (highest N, repo root).

If the highest-precedence source exists but has no viable tasks, do not default to an empty plan.

• First try to select an unblocker (a blocked leaf task that would unlock future ready work).
• Only stop+ask about falling back if the source is empty/unparseable, or if *everything* is closed/unschedulable.

Source detection + preflight (how)

Detect sources without mutating them:

1. Invocation task list:

- A numbered/bulleted list that follows an explicit phrase like "Use $select" / "$select:". - If present, treat it as the source and do not look elsewhere.

1. SLICES.md:

- Exists at repo root (SLICES.md). - Parseable (each slice section contains exactly one YAML mapping).

1. plan-N.md:

- Find files matching plan-(\d+).md in the repo root and pick the highest N.

Preflight (best-effort):

• For the chosen source, collect candidate tasks + dependency edges.
• If a dependency refers to an unknown task ID, treat the referencing task as blocked; warn only after auto-remediation fails.
• If the dependency graph is cyclic/unschedulable, schedule any work outside the cycle (if possible) and emit a warning.
• If the source is present but unparseable/empty, stop and ask before falling back.

Warning auto-remediation (read-only)

Before emitting warnings, attempt safe, deterministic fixes that do not mutate sources.

Order (stop when resolved):

1. ID normalization + aliasing:

- Canonicalize ids/depends_on/related_to (trim, lowercase, drop leading #). - If a depends_on is unknown, try to map it to a unique known id via a numeric suffix alias (e.g., 1 -> t-1, sl-1) or an exact canonical match. - If resolved, replace the dep with the canonical id and suppress the unknown-dep warning.

1. Scope normalization (safe):

- Normalize each scope entry: trim whitespace; drop a leading ./ (normalize ./ to .); collapse repeated /; remove a trailing / (except /). - Record as auto_fix: scope_normalize if any scope entry changed.

1. Scope inference (conservative):

- If scope is missing, scan location, verification, id, title, description, and subtasks for explicit path or glob tokens (contain /, **, or a file extension). - Only adopt tokens that are existing paths or obvious globs; prefer file or module globs over directory roots, and keep the narrowest non-overlapping set that still makes ownership explicit. - If inferred, set scope and suppress the missing-scope warning.

1. Orchestrator downgrade:

- If agent: orchestrator lacks subtasks, downgrade to worker. - This is a semantic change; still emit a warning even after the fix.

Record auto-fixes in the Decision Trace (auto_fix) and keep warnings for unresolved issues.

Source adapters (read one; do not improvise)

After you choose the source kind, read and apply exactly one adapter spec:

• list: codex/skills/select/ADAPTER_LIST.md
• slices: codex/skills/select/ADAPTER_SLICES.md
• plan: codex/skills/select/ADAPTER_PLAN.md

Regression fixtures live in codex/skills/select/FIXTURES.md.

Parallelism rules (safety-first)

Parallelism is only scheduled when tasks provide enough metadata to make it defensible.

• The canonical lock-root contract lives in codex/skills/select/references/lock-roots.md.
• A task is eligible for parallel waves only if it has a non-empty scope list.
• Two tasks may share a wave only if their scope sets do not overlap.

- Treat scope entries as exclusive locks. - Recommended lock style: tight file or module globs; use directory roots only when finer locks are unavailable. - Conservative overlap check: compare lock roots, not raw strings. - For each scope entry: normalize it (drop leading ./; collapse /), drop a trailing /** or /**/*, then take the prefix up to the first glob metachar (*, ?, [). - Treat overlap if any lock root is equal OR one lock root is a path-prefix of another.

• Tasks missing scope are treated as overlapping everything and therefore scheduled alone.

- Warn missing_scope only when it affects this plan's wave packing. - Hint: add a narrow scope list (paths/globs) to unlock parallel waves.

• Tasks with overly-broad scope are treated as overlapping everything and therefore scheduled alone.

- Broad examples: "", ., ./, /, *, **, **/*. - Warn broad_scope only when it affects this plan's wave packing.

• If tasks must be serialized due to overlapping lock roots, prefer an explicit depends_on edge to make the order intentional.

- Warn implicit_order only when lock roots are nested (strict prefix), the tasks are otherwise dependency-independent (no DAG path), and the order was chosen by tie-breaks/stable order.

Delegation readiness (recommended)

Parallel waves are only useful if each task is independently executable by a worker.

• Prefer tasks that include:

- location: where to work (paths/globs; navigation only; does not affect scheduling) - validation: how to prove done (commands/checks; does not affect scheduling)

• If emitting any wave with 2+ tasks and that wave contains at least one task with validation, warn missing_validation for tasks in that wave that lack validation.

Claiming selected work (required)

When $select selects work to *start now* (i.e., the tasks in the first scheduled wave waves[0]), it must also emit a claim so other planners/executors do not pick the same work concurrently.

• If the chosen source supports status, the claim is: set the selected tasks to an in-progress status using the source's spelling.

- Canonical token: in_progress - Accept common variants as equivalent when reading/triaging: in progress, in-progress, in_progress - When emitting a claim, prefer the token already used by the source; otherwise default to in_progress.

• If the source is list or plan, emit claim: none.
• If the resulting OrchPlan will feed durable execution, the follow-through path is st import-orchplan then st claim before any spawn_agent or $mesh run.

This is still plan-only: $select does not perform the writeback; it spells out what to change.

Orchestration-of-orchestration

Tasks may be delegated to an agent: orchestrator only when the user provided subtasks. If agent: orchestrator is set but subtasks is empty/missing, downgrade to agent: worker, record an auto-fix, and warn.

Output: OrchPlan v1

Emit one YAML block with the canonical plan.

Schema (YAML; best-effort, omit unknown keys):

schema_version: 1
kind: OrchPlan

created_at: "<rfc3339>"

source:
  kind: list|slices|plan
  locator: "<freeform; e.g. 'invocation', 'SLICES.md', 'plan-3.md'>"

# Optional. If omitted, treat as "auto" (unbounded by cap; waves are dependency/lock driven).
cap: auto

# Optional. Human/worker context only (does not affect scheduling).
prereqs: []
risks: []

tasks:
  - id: t-1
    title: "..."
    description: "..."          # optional
    workstream: "..."           # optional
    role: contract|implementation|integration|checkpoint  # optional
    parallelism_impact: "unlocks <n> tasks"  # optional best-effort
    agent: worker|orchestrator
    scope: ["path/**"]
    location: ["path/file"]     # optional
    validation: ["..."]         # optional
    depends_on: []
    related_to: []              # optional non-gating links
    subtasks: []

waves:
  - id: w1
    tasks: [t-1, t-2]
  - id: w2
    tasks: [t-3]

integration:
  boundary: patch-first
  order: [t-1, t-2, t-3]
  conflict_policy: rebase-author

warnings:
  - "..."

Decision Trace (required)

After the OrchPlan YAML, emit a short plaintext trace (tight and structured):

• source: chosen source kind + locator
• mode: resolved mode + resolved max_tasks
• triage: if any in_progress was seen, state: continue <id> OR recommend close <id> OR recommend reopen <id> OR none
• claim: mark <in_progress token> <id,...> OR already <in_progress token> <id> OR none
• counts: totals for the chosen source (at minimum: leaf, ready, blocked, in_progress)
• fanout_possible: number of dependency-independent tasks that could safely fit in waves[0] after scope checks
• fanout_selected: number of tasks actually placed into waves[0]
• fanout_left_on_table: fanout_possible - fanout_selected
• pick: selected task id + 3-10 word reason
• next2: next two candidates (or none) + 3-10 word reason each
• waves: (recommended when tasks were scheduled) N + a compact wave listing (e.g. w1[t-1,t-2]; w2[t-3])
• review: pass|warn|skipped|blocked + 0-6 word note
• warnings: list count + top 1-3 keys (e.g. unknown_deps, status_drift, cycle, broad_scope, implicit_order, underfilled_wave, missing_validation, linear_graph, missing_role, missing_checkpoint)
• auto_fix: list count + top 1-3 keys (e.g. dep_alias, scope_normalize, scope_infer)

Procedure (high-level)

1. Resolve invocation directives (mode/max_tasks/cap).
2. Source detection (pick exactly one; do not merge sources).
3. Read the corresponding adapter spec (above) and extract tasks.
4. If tasks are too coarse or missing metadata required for safe parallelism, refine them:

- Decompose into atomic tasks with explicit depends_on. - Identify workstreams and annotate workstream where useful. - Insert contract tasks when they can unlock parallel implementation branches. - Insert checkpoint/integration tasks as explicit join points across branches. - Encode soft ordering/context in related_to instead of depends_on. - Populate scope locks (tight paths/globs), plus location and validation where possible. - Keep each task independently PR-able (medium granularity). - Explore the repo (read-only) and consult authoritative docs when needed. - Stop and ask targeted questions if blocked by ambiguity.

1. Normalize tasks: ensure id; apply orchestrator rule; treat unknown deps as blocked (pending auto-remediation).
2. Run warning auto-remediation (above); finalize warnings.
3. Schedule waves using depends_on + scope locks. 7.5. If selecting new work, compute claim from waves[0]; for source-backed flows emit manual status writeback, and for durable execution handoffs keep the first-wave ids ready for st import-orchplan + st claim.
4. Reviewer pass (per review): check deps/order/locks/validation/delegation gaps; revise as needed.

- Detect unnecessary linear chains and downgrade advisory edges from depends_on to related_to when safe. - If waves[0] leaves safe ready work unclaimed and max_tasks was not explicitly capped, treat it as a planning defect and revise until fanout_left_on_table=0 or emit a blocking explanation. - Require explicit roles/workstreams in multi-wave plans when inference is feasible. - Require checkpoint/integration joins when multiple parallel implementation branches converge. - In reviewer mode: do not expand scope; do not redesign; only close gaps and reduce risk. - If review=required: iterate until review: pass OR stop+ask if blocked.

1. Emit OrchPlan v1 YAML (always) + Decision Trace (required). Add pipelines only when useful.

Scheduling algorithm (parallelism-first)

Build waves using dependency readiness and scope locks:

1. Build a DAG from depends_on edges only (related_to is non-gating and excluded from DAG readiness).
2. Maintain ready = unscheduled tasks whose deps are all scheduled.
3. While tasks remain:

- Treat missing scope as overlapping everything (i.e. it can only be scheduled alone). - Treat overly-broad scope locks ("", ., ./, /, *, **, **/*) as overlapping everything. - Pick a maximal subset of ready whose scope lock roots do not overlap (greedy is fine). - When cap and max_tasks are auto, keep the full maximal subset; do not trim to one task merely because the source is slices. - If cap is a number, limit the wave to cap tasks. - Remove scheduled tasks from the pool; proceed to next wave.

When you must choose between conflicting tasks (overlapping scope or cap pressure), use scoring beyond priority:

1. Risk/hardness/blast (if present): prefer lower risk, lower hardness, and smaller blast radius first.
2. Priority (if present): 0/P0 first.
3. Kind order (if present): task > bug > feature > chore > epic > docs > question.
4. Parallelism impact (same risk tier): prefer role=contract|checkpoint, then integration, then implementation.
5. Unlock count (same risk tier): prefer tasks that unblock more other tasks in this plan.
6. Soft-order penalty: if a task has related_to another ready task, prefer the related-first ordering when it likely reduces rework.
7. Delegation readiness: prefer tighter scope and explicit validation.
8. Stable order: preserve the source order.

Emit warnings when unresolved (noise-controlled; warn only when it affects this OrchPlan):

• missing_scope: a task missing scope prevented adding at least one other ready task to the same wave.
• broad_scope: a task with overly-broad scope prevented adding at least one other ready task to the same wave.
• implicit_order: two dependency-independent tasks had nested lock roots and were concurrently ready; order chosen by tie-breaks/stable order.
• missing_validation: a parallel wave mixed tasks with and without validation.
• unknown_deps: a depends_on points at an unknown ID.
• orchestrator_without_subtasks: a task declared agent=orchestrator without subtasks.
• underfilled_wave: waves[0] left safe ready tasks unclaimed without an explicit max_tasks cap.
• linear_graph: the dependency graph is mostly chain-like and a safe split into workstreams/contracts/checkpoints appears possible.
• missing_role: multi-wave plan lacks role annotations where they are needed for reasoning/review.
• missing_checkpoint: multiple parallel implementation branches have no explicit integration/checkpoint join. If a warning is auto-remediated, omit it from warnings and list it under auto_fix (except orchestrator downgrade, which must still warn).

Examples (synthesized)

Example A: clean parallel wave (list source)

schema_version: 1
kind: OrchPlan

created_at: "2026-02-02T00:00:00Z"

source:
  kind: list
  locator: "invocation"

cap: auto

tasks:
  - id: cfg
    title: "Add config loader"
    agent: worker
    scope: ["src/config/**"]
    location: ["src/config/loader.ts", "src/config/index.ts"]
    validation: ["npm test -w config"]
    depends_on: []
    subtasks: []
  - id: ui
    title: "Update settings UI"
    agent: worker
    scope: ["src/ui/**"]
    location: ["src/ui/Settings.tsx"]
    validation: ["npm test -w ui"]
    depends_on: []
    subtasks: []
  - id: wire
    title: "Wire config into UI"
    agent: worker
    scope: ["src/app/**"]
    location: ["src/app/bootstrap.ts"]
    validation: ["npm test"]
    depends_on: [cfg, ui]
    subtasks: []

waves:
  - id: w1
    tasks: [cfg, ui]
  - id: w2
    tasks: [wire]

integration:
  boundary: patch-first
  order: [cfg, ui, wire]
  conflict_policy: rebase-author

warnings: []

Decision Trace:

• source: list (invocation)
• mode: both; max_tasks=auto
• triage: none
• claim: none
• counts: leaf=3 ready=2 blocked=1 in_progress=0
• pick: cfg; unblocks wire; parallel-safe scope
• next2: ui; parallel-ready; disjoint scope
• waves: 2 w1[cfg,ui]; w2[wire]
• review: pass
• warnings: 0
• auto_fix: 0

Example B: parallel wave with mixed validation + broad scope

schema_version: 1
kind: OrchPlan

created_at: "2026-02-02T00:00:00Z"

source:
  kind: list
  locator: "invocation"

cap: auto

tasks:
  - id: api
    title: "Add /health endpoint"
    agent: worker
    scope: ["src/api/**"]
    location: ["src/api/health.ts", "src/api/router.ts"]
    validation: ["npm test -w api"]
    depends_on: []
    subtasks: []
  - id: docs
    title: "Document /health endpoint"
    agent: worker
    scope: ["docs/**"]
    location: ["docs/api.md"]
    depends_on: []
    subtasks: []
  - id: big
    title: "Repo-wide rename OldName -> NewName"
    agent: worker
    scope: ["**"]
    location: ["."]
    validation: ["rg -n \"OldName\" . || true"]
    depends_on: []
    subtasks: []

waves:
  - id: w1
    tasks: [api, docs]
  - id: w2
    tasks: [big]

integration:
  boundary: patch-first
  order: [api, docs, big]
  conflict_policy: rebase-author

warnings:
  - "missing_validation: [docs]"
  - "broad_scope: [big]"

Decision Trace:

• source: list (invocation)
• mode: both; max_tasks=auto
• triage: none
• claim: none
• counts: leaf=3 ready=3 blocked=0 in_progress=0
• pick: api; explicit validation; tight scope
• next2: docs; parallel-ready; missing validation
• waves: 2 w1[api,docs]; w2[big]
• review: warn missing_validation,broad_scope
• warnings: 2 missing_validation,broad_scope
• auto_fix: 0

Source adapters (extraction only)

Adapter specs live in:

• codex/skills/select/ADAPTER_LIST.md
• codex/skills/select/ADAPTER_SLICES.md
• codex/skills/select/ADAPTER_PLAN.md

Pipelines

Pipelines live in codex/skills/select/PIPELINES.md.

Output

• Always emit OrchPlan v1 YAML, then Decision Trace.
• If no actionable source tasks exist, prefer selecting an unblocker; only then emit an empty plan + warnings.
• If the chosen source is empty/non-viable and a fallback source exists, stop and ask before falling back.