Start plan review
Skill 540lyle/lens-temper/plugins/lens-temper/skills/start-plan-review
Use first when a user asks LensTemper to review a plan, spec, proposal, or implementation approach, including full reviews, selected-lens reviews, reruns, or archived review runs.From its SKILL.md
npx -y skills add 540lyle/lens-temper --skill start-plan-reviewAssembled from the repository path, not quoted from the project. Check it against their README if it does not work.
One thing to look at
- 1 stars1 stars. Stars are a popularity signal and not a quality one, but at this level it is likely that nobody has read this closely except its author, and you would be relying on your own review.
SKILL.md
6.0 KB, as published. Nobody here has run it
LensTemper Start Plan Review
Use reviews/registry.json from the skill package or repository root as the entry
point. Read reviews/README.md, reviews/AGENT.md, selected lens manifests,
and selected role manifests before running a review.
Inputs
- Target plan or spec path.
- Repository-relative review input JSON containing the feature request, relevant context, constraints, and optional previous adjudications.
- Selected lenses, or enough context to select lenses.
- Pass id, deterministic target revision, and normalized review input revision.
Outputs
- Ledger JSON.
- Trace event log (
events.jsonl) when orchestration artifacts are generated. - Per-lens prompt packets or reviewer instructions.
- Detached orchestrator packet when the host supports an independent orchestrator agent.
- Captured review outputs.
- Synthesis output.
- Rerun decisions and completion summary.
Procedure
- Resolve lens scope before creating the ledger or spawning reviewers:
- If the user explicitly names lenses, validate each id against
reviews/registry.jsonand use exactly that set. Do not add or remove lenses; stop on unknown or duplicate ids. - Otherwise, use the registry's default core profile and run the canonical selector against the normalized review input and current target. The core lenses are always required; any deterministically triggered specialists are added to the run's required set.
- The orchestrator may add registry-valid lenses through a validated
--lens-proposal. Each addition needs a concise reason and concrete evidence from the canonical review input or target. Additions are unioned with the deterministic minimum and may never remove or replace it. - Focused selector-only workflows may still stop for clarification when no domain matches. A normal full review does not: the core profile supplies the required baseline.
- If the user explicitly names lenses, validate each id against
- Create or update a ledger with deterministic target, template, and lens revisions plus the repository-relative review input path and normalized review input revision. Stop if the feature request is missing.
- Use
full_hostedorfull_detachedby default. A request to run LensTemper, review with LensTemper, or run a full LensTemper review means detached-context reviewer subagents, one per selected lens. A detached-context reviewer subagent is fresh and receives none of the host, parent, or orchestrator conversation or history; it reads only the run packet and permitted workspace files. Host equivalents:- Codex: use the current detached-context subagent mechanism once per
selected lens. See
docs/hosts/codex.mdfor current mechanics. - Claude Code: each
Agent(Task) invocation is already a fresh subagent with isolated context, so spawn one Agent call per selected lens -- no flag required. - Claude Desktop / Claude.ai: use only if the host can launch
detached-context reviewer subagents and can provide the shared
reviews/workflow resources. Otherwise stop full-review requests; inline/advisory mode is only valid when the user explicitly asks for a non-lockable advisory pass. - Cursor, plain CLI, and other manual hosts: use LensTemper materials as advisory unless a fresh independent-agent mechanism and artifact validation have been verified for that host. Do not perform an inline/advisory substitute unless the user explicitly asks for inline or advisory mode. If fresh subagents cannot be spawned, stop and report that the full review could not be completed.
- Codex: use the current detached-context subagent mechanism once per
selected lens. See
- Choose the run family:
inline: explicitly requested current-context advisory review, no lockable claims.full_hosted: this agent orchestrates fresh lens reviewers.full_detached: a fresh orchestrator owns ledger, reviewers, synthesis, reruns, archive, and completion claims.
- For
full_detached, make the parent agent a launcher/reporter. Generate the host-neutral packet withreviews/scripts/assemble-orchestrator-prompt.mjsorreviews/scripts/run-plan-review.mjs --execution-mode fresh_spawned_orchestrator, then hand that Markdown packet to the host's independent-agent mechanism. - Assemble reviewer prompts with
reviews/scripts/assemble-review-prompt.mjswhen possible. - Run reviewers as independent fresh agents only when the host supports that and artifact validation can prove the run. Codex and Claude may provide different spawning mechanics; Cursor, plain CLI, and manual hosts remain advisory until their fresh-agent path is verified. Reviewer execution may be concurrent or sequential. Each selected lens still requires its own detached-context reviewer subagent.
- Validate review, synthesis, and ledger JSON with
reviews/scripts/validate-review-fixtures.mjsor the individual validators. For full runs, bind review and completion validation to the run with--ledger <run>/ledger.json; do not validate them against free-standing revision strings. - Assemble synthesis with
reviews/scripts/run-synthesis.mjs --ledger <run>/ledger.json, then archive completed runs withreviews/scripts/archive-review-run.mjs. - After current review and synthesis artifacts are attached, finalize derived
readiness state with
reviews/scripts/update-ledger.mjs --ledger <run>/ledger.json --finalize --write. Do not authorcompleted_lens_ids,completion_validation, ledger completion status, orcore_gate_passeddirectly.
The orchestrator may update ledger state. Lens reviewers may not.
Detached orchestration may not claim completion unless events.jsonl, ledger,
reviewer outputs, synthesis, and archive evidence agree.
Unqualified completion additionally requires run_scope: core_profile, a
registry-valid core_profile_id, every entry in required_lens_ids in
completed_lens_ids, and core_gate_passed: true.