plan-review
/plan-review
Przegląda lokalny plik planu, zanim ktokolwiek go wdroży: planning-agent proponuje zmianę, a człowiek ją zatwierdza.
- Autor
- qgadrian ↗
- Repozytorium
- aidrian ↗
- Licencja
- Apache-2.0
- Ostatnia zmiana repozytorium
- 12.11.2026
Ocena i bezpieczeństwo
- Pliki w paczce
- 3
- Skrypty
- 0
- Komentarze HTML
- 0
- Adresy URL w SKILL.md
- 0
Brak sygnałów. Repozytorium zmienione 12.11.2026, bez odwołań do wycofanych modeli i API.
Apache-2.0, otwarta: można używać i zmieniać, z zachowaniem licencji i informacji o zmianach.
Co robi i kiedy
Przegląda lokalny plik planu, zanim ktokolwiek go wdroży: planning-agent proponuje zmianę, a człowiek ją zatwierdza.
Oryginalny opis (EN)
Review a local plan before anyone implements it. Use when asked to review a plan file. planning-agent proposes a diff; the Human approves it.
Jak działa
- Inputs
- 1. Find the plan
- 2. Preflight
- 3. Gates
- 3b. Quorum
- 4. Critique
- 5. Apply
- 6. Cap
- Close
- Do not
Podgląd SKILL.md
--- name: plan-review description: Review a local plan before anyone implements it. Use when asked to review a plan file. planning-agent proposes a diff; the Human approves it. --- # plan-review Run one critique-and-refine pass over a plan before anyone implements it. This session coordinates: it finds the plan, runs the gates, has `planning-agent` critique the plan and propose a diff, and apply it only after the Human approves it. It never critiques or rewrites the plan itself, and never commits it. Before you start, follow the `knowledge` skill. Load your `product_sensitive` docs for the flows the project declares critical, and `security` docs for the surfaces that need a security pass. ## Inputs ```text [<plan path> | <issue key or URL>] [--findings-only] [--apply] [--extended] [--quorum | --no-quorum] ``` - A plan path: absolute or from the repository root. An issue key or URL: the plan for that issue, found as step 1 says. - `--findings-only`: critique without a proposed diff. - `--apply`: the Human approves the diff this session proposed earlier; go to step 5. Valid only as step 5 says. - `--extended`: allow one refinement pass more than step 6 does. - `--quorum`: run step 3b on any plan. `--no-quorum`: skip it. **stop** ends the run with a report; **stop and ask** adds one question only the Human can answer. ## 1. Find the plan Plans live in `<state>/plans/`, where `<state>` is the folder `aidrian state` prints. A design import keeps its plan at `docs/design-imports/<slug>/plan.md` in its worktree, on branch `feat/design-<slug>`. Skip backups, editor and temporary files, and `.json` files. Without a path, take the first of these that finds a plan: 1. The issue: the argument, else the key in the current branch name, in the format the project's `issue_tracker` knowledge gives. Look for `<state>/plans/<key>-*.md`, with the key in lowercase, and for a plan whose `issue_id` is the key. 2. A plan whose `workspace_decision.branch_name` is the current branch. 3. The newest plan whose `status` is not `complete`. Several candidates: list up to five, newest first, and ask. None: stop, and point to the `planning-agent` skill, or the `issue-plan` skill for an issue. Never build a plan from the conversation. ## 2. Preflight - Read the plan. With YAML frontmatter it is a schema plan of its `plan_kind`; without `plan_kind`, review it as `implementation` and add a finding. Without frontmatter it is a free-form plan: it gets the generic lens, never a penalty for the missing schema. - `status: complete`: ask whether to review it again. - An issue-anchored plan whose `ticket_updated_at` is older than the issue's `updatedAt` is stale. Ask: the Human revises it by running the `issue-plan` skill with `--mode revise`, which only the Human starts; review it as it is; or stop. On a critical flow, offer only revise or stop. - Record the plan's hash, `git hash-object <plan>`, for step 5. ## 3. Gates Both run before the critique; their verdicts go into its brief as authoritative. A gate's brief carries the plan path, the trigger (the plan lines and the flow or surface they touch), and the knowledge docs to check against. It carries nothing of your own: no suspected findings, risks, or hints. The gate reviews the plan unanchored. - **Critical flow.** The plan's goal, scope, approach, affected surfaces, or slice files touch a flow the project's `product_sensitive` knowledge declares critical (the whole text, for a free-form plan): `product-manager` first. `blocked` or `needs_evidence` stops the run with its `decision_options`, attributed: no `authority` covers this gate (`knowledge/references/decisions.md`). - **Security.** The plan touches auth, sessions, tokens, tenancy, permissions, payments, webhooks, uploads, redirects, secrets, untrusted content, privileged access, or a surface the project's `security` knowledge lists: `security-engineer` first. Its verdict and required invariants bind the critique; no critique finding drops them. ## 3b. Quorum When a gate in step 3 ran, or the plan names a decision-record path or the `adr-writer` skill, load the `quorum` skill and follow it, with the gate verdicts and the trigger. Its `references/ledger.md` § Under plan-review defines the triggers, its reruns, and the `blocked on quorum` stop that resumes here. Its accepted objections, the Human's answers, and its Report's revision-history line go into step 4's brief as authoritative. Skip it when it is not installed. ## 4. Critique Dispatch `planning-agent` alone, in the foreground. Its brief carries the plan path, the gate verdicts, the flags, the issue's acceptance criteria and non-goals for an issue-anchored plan, and `references/review-lens.md` from this folder, which it applies. It consults `product-manager`, `architect`, or `explorer` itself when the lens needs one: dispatch no consult beside it. When it returns blocked on missing evidence, dispatch `explorer` for that evidence, then critique once more. Then report, in this order, and stop: 1. **Verdict**: `plan ready: no findings`, `findings: diff proposed`, `findings only`, or `blocked on <gate or evidence>`. 2. **Plan**: path, `plan_kind` or free-form, `status`, and freshness against the issue. 3. **Gates**: each verdict, or not applicable; the quorum Report's Ledger item with its trigger, or why it was skipped. 4. **Findings**: blocking, important, then improvements, each with its anchor. 5. **Open questions** from the plan, word for word, then **escalations**: product or architecture decisions, scope growth, disputes. 6. **Proposed diff** with its revision-history line, unless `--findings-only`. 7. **Ask**: `apply`, `apply with edits: <edits>`, or `skip`. ## 5. Apply Only on the Human's explicit `apply` in this session, after step 4 in this same session: it needs step 2's hash and step 4's diff. `--apply` without both: refuse, and say to rerun without `--apply`. Then read [steps/apply.md](steps/apply.md) and follow its steps 5.1 to 5.4. ## 6. Cap One critique, one apply, and one more critique only when step 5.4 finds a new blocking flaw. `--extended` adds one refinement pass, never more. Past the cap, escalate to the Human. A restart from step 5.1 reviews a different plan, so its count starts again; it waits on the Human's next `apply` like the first. A consult that returns a dispute stops the run: escalate both positions and the missing evidence; never settle on a compromise. ## Close The run is done only when each gate is cleared or not applicable, the Human answered step 4, an applied plan passed step 5.4, and no dispute or blocked consult was dropped. Otherwise report `blocked` or `needs_input` with the step to resume from. End with the next step: the `issue-implement` skill for an issue-anchored plan, and the `plan-implement` skill for any other, design imports included. A changed design-import plan is a file change the report names; this skill never commits it. ## Do not - Critique, rewrite, or restyle the plan yourself, in a report or in any brief, or prefer your structure to the author's when both are valid. - Apply a change without the Human's approval in this session, or a scope-growing change without its own approval. - Skip a gate, or loop past the cap. - Hand work an agent can do to the Human as a follow-up.
Pliki w paczce (3)
- SKILL.md 7 KB
- references/review-lens.md 5 KB
- steps/apply.md 1 KB
Wersje
| Repozytorium | Wynik | Licencja |
|---|---|---|
| Ta strona · qgadrian/aidrian | 100 | Apache-2.0 |
| aws-solutions/konductor | 100 | Apache-2.0 |
| mathews-tom/armory | 100 | MIT |
| noes-25h/agent-kit | 100 | MIT |
| unhaeos/hhmode | 100 | MIT |
| ndisisnd/msg | 100 | MIT |
| sd0xdev/sd0x-dev-flow | 100 | MIT |
| sean-sunagaku/claude-code-plugin | 100 | MIT |
| longranger2/claude-gpt-workflow | 100 | MIT |
Migawka bazy z 13.11.2026 · wersja paczki c07db795796e (skrót SHA-256 całej paczki)
Pobrania w ostatnich 30 dniach: 0 · ranking
Zgłoś problem z tym skillem: szkodliwa treść, błędny opis albo prośba autora o usunięcie.
Podobne skille
- 100plan-canvas
Otwiera plany i artefakty HTML w lokalnym płótnie przeglądarki z szablonem planu proponowanym po słowach plan lub roadmap, gdzie człowiek komentuje elementy i zatwierdza zmiany. - 100plan-canvas
Otwiera plany i artefakty HTML w lokalnym podglądzie przeglądarki, gdzie człowiek zaznacza elementy, rozmawia i zatwierdza albo prosi o zmiany. - 100plan-canvas
Otwiera plany i artefakty HTML w lokalnym podglądzie przeglądarki, gdzie człowiek zaznacza elementy, rozmawia i zatwierdza albo prosi o zmiany. - 100plan-canvas
Otwiera plany i artefakty HTML w lokalnym płótnie przeglądarki, gdzie człowiek komentuje elementy, rozmawia i zatwierdza albo zgłasza zmiany bez opuszczania strony. - 100plan-canvas
Otwiera plany i artefakty HTML w lokalnym płótnie przeglądarki, gdzie człowiek komentuje elementy, rozmawia i zatwierdza albo zgłasza zmiany bez opuszczania strony. - 90contextual-plan
Planuje aplikację lub stronę przed budową: zadaje najwyżej pięć pytań, proponuje stos i kierunek designu i zapisuje plik planu.