메뉴
Unified planning entry point: classify tier, run tier-sized research, then route to the smallest safe plan→task-artifacts→review workflow. Use after scope or for "plan this/how build this" requests. Do NOT trigger for scope, specific bug diagnosis, or direct task breakdown.
Codex 또는 Claude로 설치 이 Prompt를 복사해 Codex, Claude 또는 다른 어시스턴트에 붙여 넣으면 Skill 페이지를 검토하고 설치를 진행할 수 있습니다.
SOC 직업 분류 기준
Use when modifying the Codex CASPAR install flow, project skill syncing, learn/recall registry files, or Codex-specific runtime files. TRIGGER when: codex, caspar, codex install, project skills, registry, caspar-learn, caspar-recall, config.toml, doctor
Run an independent adversarial code review of completed work using a pinned high-effort opposing runtime, with a same-contract native fallback. Finds bugs, correctness and security failures, regressions, performance landmines, overengineering, and missing behavioral tests. Review only; never edits code.
Build the planned feature from a compact execute.md index and sliceable tasks.json detail, dispatching parallel task waves through deterministic checks, lightweight sentinel review, and final adversarial review/validation. Trigger after execute.md/tasks.json exist or to resume a partially-built wave plan. Do NOT trigger for scoping/planning, unplanned bug-fixing, or dead-code cleanup (caspar-prune).
Use when rendering the Next Steps footer at the end of any caspar command, suggesting which caspar command to run next, or answering "what command should I run?" / "how does the workflow work?". Do NOT use mid-task or for executing a phase — this only renders guidance and the footer.
Independent adversarial review of plan.md before task generation. Use a pinned high-effort opposite-runtime reviewer; fall back to one native reviewer with the same contract when unavailable. Trigger after create_plan and before create_tasks. Do NOT review generated tasks, finished code, or apply scope changes.
Adversarial review of generated execute.md + tasks.json against a reviewed plan using a pinned high-effort opposite runtime, with a same-contract native fallback. Checks translation, criteria, RED pairing, dependencies, waves, and alignment. Do NOT review plan quality, finished code, or change scope.
| name | caspar-plan |
| description | Unified planning entry point: classify tier, run tier-sized research, then route to the smallest safe plan→task-artifacts→review workflow. Use after scope or for "plan this/how build this" requests. Do NOT trigger for scope, specific bug diagnosis, or direct task breakdown. |
| user-invocable | true |
Planning router: classify tier, spend a tier-sized research budget, then route to the smallest safe workflow. Clear on WHAT, silent on HOW. Plan review stabilizes intent before task generation; comprehensive task review checks generated execution artifacts.
$ARGUMENTS — feature/problem reference + optional user-specified output dir or tier.{OUT_DIR}/ if present: concepts/scope.md, specs/prd.md/specs/ux.md, task_context.md, specs/plan.md, specs/execute.md, specs/tasks.json, research/*.md. Read what exists; reuse research instead of repeating it.branch = git rev-parse --abbrev-ref HEADOUT_DIR = user-specified || docs/tasks/{branch}; mkdir -p it.rg/glob/manifests. No subagents until tier is set.1. Classify (CheapLocalScan → CheckHardStops → DetermineTier). Any hard-stop true ⇒ automatic COMPREHENSIVE:
db_schema_destructive (drop/rename/non-additive column) · data_migration_required (backfill/transform) · new_service_or_component · auth_or_pii_change · secrets_or_credentials_handling · payment_billing_logic · public_api_change · concurrent_writes_or_locking · caching_consistency · cross_service_or_cross_workspace_change · slo_sla_risk.
| Tier | Use when |
|---|---|
| MICRO | Obvious ≤1-file change, no artifact needed, no hard-stop. |
| LIGHT | Clear internal change, known pattern, ~≤5 files / 1–2 components, no hard-stop. |
| STANDARD | Multi-file or moderate ambiguity, no hard-stop; needs adversarial review. |
| COMPREHENSIVE | Any hard-stop, new architecture/service, high ambiguity, or broad blast radius. |
Between adjacent tiers: prefer lower if execution can check the gap; prefer higher for scope, security/privacy, data correctness, public API, or rollback risk. Log tier inline; do not ask for confirmation.
2. Research by tier. MICRO/LIGHT: cheap scan only (LIGHT may dispatch @caspar:finder if files unclear). STANDARD: ≤2 of @caspar:finder/@caspar:analyst/@caspar:patterns; no @caspar:web-research unless a new external dep/API/framework is likely. COMPREHENSIVE: all needed dimensions incl. @caspar:web-research only when external choices matter. For non-MICRO, write findings to {OUT_DIR}/task_context.md ## Technical Research so downstream skills consume instead of re-research.
3. Initial Design Proposal Gate — STANDARD/COMPREHENSIVE only (skip MICRO/LIGHT). Present inline a single high-level proposed approach — components + key decisions/rationale — with: Verification spine (major change → test|observable|state signal), Filled assumptions (default + source), Open questions (each with recommended default). A fork = an open question, not a menu.
HOLD FOR USER CONFIRMATION HERE. End with exactly: "High-level design proposed. Reply Approved to generate the full reviewed plan/tasks, or give design feedback." Do not call caspar-create_plan, caspar-create_tasks, or caspar-plan_review until the user approves this gate or gives feedback that has been incorporated. Once approved, append ## Selected Design to {OUT_DIR}/task_context.md, including the approved design and any resolved assumptions/questions.
4. Route — invoke the required skills only after required gates are satisfied (do not describe, summarize, or hand-write plan/tasks):
Skill(caspar-create_plan) {OUT_DIR}/task_context.md --depth light --no-review → Skill(caspar-create_tasks) … --depth light. Verify execute.md sections + tasks.json parse; summarize artifacts; state LIGHT skipped review + human gate.task_context.md must contain ## Selected Design from the approved Initial Design Proposal Gate before routing. If missing, return to step 3 and stop.Skill(caspar-create_plan) {OUT_DIR}/task_context.md --depth standard --no-review → Skill(caspar-plan_review) {OUT_DIR} --mode adversarial --auto-apply scope-safe → Skill(caspar-create_tasks) {OUT_DIR} --depth standard.Skill(caspar-create_plan) {OUT_DIR}/task_context.md --depth comprehensive --no-review → Skill(caspar-plan_review) {OUT_DIR} --mode full --auto-apply scope-safe → Skill(caspar-create_tasks) {OUT_DIR} --depth comprehensive → Skill(caspar-task_review) {OUT_DIR} --mode adversarial --auto-apply scope-safe.After any user conversation, re-orient: if STANDARD/COMPREHENSIVE and the Initial Design Proposal Gate is not approved, stay in step 3; if approved and ## Selected Design exists in task_context.md, continue to step 4; only then execute the next required Skill invocation. YOU MUST call these via the Skill tool — never write plan.md, execute.md, or tasks.json content yourself, never self-author plan_review, never end the turn saying "I'll now…".
concepts/scope.md (then specs/prd.md/specs/ux.md) as canonical. Planning, task artifacts, review, and feedback integration may change approach, sequencing, verification, references, and YAGNI fences — but MUST NOT cut, narrow, expand, or reinterpret scope without an explicit user scope-change gate. Never overwrite an existing plan.md, execute.md, or tasks.json; use scoped names.caspar-plan_review must produce a saved plan-only review through an opposite-runtime CLI reviewer when available, or through explicit same-runtime Caspar subagent fallback when not. The planner MUST NOT substitute an inline/self-authored review. Read the report; ensure every scope-safe Blocker/High is reflected in plan.md before task generation. Do not apply Scope-Change-Required findings — surface them.caspar-task_review checks only generated execute.md + tasks.json against the reviewed plan. Apply scope-safe Blocker/High task fixes to tasks.json and affected execute-index rows only; never edit plan.md or scope docs from task review.{OUT_DIR}/specs/plan.md + specs/execute.md + specs/tasks.json.plan_review report under {OUT_DIR}/reviews/, with scope-safe plan findings integrated before task generation.task_review report under {OUT_DIR}/reviews/, with scope-safe task-artifact findings integrated.Post-tasks tier re-check: count parent/sub-tasks, unique files, Phase-0 deps. Recommend, do not silently re-run: escalate if LIGHT touches >3 files or >2 parents; STANDARD reveals a missed hard-stop; or any Out-of-Bounds violation. Downgrade only if COMPREHENSIVE collapses to one parent with no migration/new-component/API change. Say: "I planned this as {tier}, but tasks revealed {signal}. Recommend re-running as {new}. Reply 'rerun' or 'keep'." and wait.
DONE when: tier was classified and logged; the tier's required skills ran (not described); required artifacts exist for the tier; execute.md has Document Manifest / Task Detail Source / Execution Summary / Wave Plan / Parent Task Index / Slicing Rules; referenced tasks.json parses; for STANDARD/COMPREHENSIVE, the Initial Design Proposal Gate was approved and recorded as ## Selected Design, independent plan review produced a saved report before task generation, scope-safe plan findings are integrated, the Final Gate was presented, and the post-tasks re-check passed/was surfaced; for COMPREHENSIVE, task review produced a saved report after task generation and scope-safe task findings are integrated; canonical scope is unchanged (or a scope gate was raised).
Present artifact paths (plan.md, execute.md, tasks.json, reviews/…), applied review changes, skipped items, and any blocked scope-change recommendation. If scope change is required, say "This requires a scope change; I did not apply it" and ask whether to reopen scope or keep the scope-preserving plan. Else prompt: "Final reviewed plan/tasks are ready. Reply Approved to proceed to execution, or give final feedback." Wait.
plan.md edit, re-run plan_review … --auto-apply scope-safe, then regenerate task artifacts; if it only corrects task translation, edit tasks.json plus affected execute.md rows and re-run task_review for COMPREHENSIVE./caspar:scope; do not edit plan/tasks against stale scope.Then render the inline Next-Steps line: suggest /caspar:execute.
/caspar:scope.