lumin-repo-lens-lab-write-gate

name	lumin-repo-lens-lab-write-gate
description	Use before/after TS/JS code changes: add, edit, move, rename, refactor, make a helper/type/file/function, or ask if something already exists. Infer intent from plain language, run pre-write reuse screening, then post-write delta checks.

Lumin Repo Lens Write Gate

This is the code-change transaction surface for lumin-repo-lens-lab. It owns pre-write and post-write together because post-write reads the pre-write-advisory.<invocationId>.json produced by the same change.

Keep the human-facing answer short and kind. The assistant reads the cold artifacts; the human gets the next concrete step.

Core Contract

NO STRUCTURAL CLAIM WITHOUT MACHINE EVIDENCE
NO ABSENCE CLAIM WITHOUT STATED SCAN RANGE

Use hedging like "looks like" only when the internal label is degraded or unknown.

Shared Engine

The shared engine lives in the sibling audit skill:

node ${CLAUDE_PLUGIN_ROOT}/skills/lumin-repo-lens-lab/scripts/audit-repo.mjs

In a maintainer checkout, the equivalent command is:

node audit-repo.mjs

Slash commands still read <SKILL_ROOT>/references/command-routing.md from the shared audit skill for exact flag routing. Below, <SKILL_ROOT> means ${CLAUDE_PLUGIN_ROOT}/skills/lumin-repo-lens-lab in plugin installs, or the repo root in a maintainer checkout.

Below, <audit-repo> means whichever of the two commands above applies to the current context.

This surface owns /lumin-repo-lens-lab:pre-write and /lumin-repo-lens-lab:post-write.

References To Load

• Read <SKILL_ROOT>/references/command-routing.md first for slash-command routing.
• Read <SKILL_ROOT>/references/pre-write-intent-shape.md before constructing or repairing intent JSON.
• Read <SKILL_ROOT>/canonical/pre-write-gate.md for the pre-write protocol, domain-cluster hints, exact shape lookup, and advisory language.
• Read <SKILL_ROOT>/canonical/any-contamination.md for planned type escapes, contaminated reuse demotion, and post-write escape deltas.
• Read <SKILL_ROOT>/references/lifecycle-modes.md when exact flags, exit codes, or artifact names matter.
• Read <SKILL_ROOT>/references/write-gate-runtime.md before running or diagnosing pre-write/post-write. It owns current Rust evidence verification, cache use, timing interpretation, and stale-runtime handling.
• Read <SKILL_ROOT>/references/glossary.md and <SKILL_ROOT>/references/false-positive-index.md when advisory output surfaces FP/tier terms that need explanation. The long FP case ledger is maintainer-only; do not load it for ordinary pre-write/post-write.

Pre-Write

Run before adding, implementing, refactoring, moving, renaming, or extending code when a compact intent can be inferred.

Do not ask normal chat users to hand-write JSON. Infer the smallest intent you can from the request and stream it through stdin with --intent -. Generated intent JSON is ephemeral request transport: do not write it under the repository root, the audit output directory, or .audit. Then run:

<audit-repo> --pre-write --root <repo> --output <dir> --intent -

An intent path explicitly supplied by the caller is caller-owned input. Pass it through and never delete or rewrite it. If a host adapter genuinely cannot stream stdin, place only the adapter-owned copy under the OS temporary directory and remove it in finally; do not turn it into repository evidence. Post-write consumes the invocation-specific advisory, not the original intent.

Read the invocation-specific advisory path printed by pre-write before coding during the same uninterrupted change transaction. It is also recorded as artifactPaths.invocationSpecific in the advisory JSON and as manifest.preWrite.advisoryPath when run through the orchestrator. pre-write-advisory.latest.json is only a convenience pointer; use the explicit pre-write-advisory.<invocationId>.json path for post-write across session or task boundaries.

For a normal fresh JS/TS run, verify preWrite.rustEvidencePath and preWrite.rustEvidence.summary before coding. Missing compact Rust evidence is not a successful fresh run; follow references/write-gate-runtime.md instead of repeating a legacy full scan or silently using stale artifacts.

In parallel work, the main controller runs one pair per same-language evidence owner. It preserves the complete intent transport, including language and refactorSources, and runs post-write after workers finish but before broad tests, builds, or generators. Subagents must not launch overlapping scans.

Planned file paths may be grounded NEW_FILE and still carry DOMAIN_CLUSTER_DETECTED when the same directory already has a matching prefix, suffix, or domain-token family. Treat that as a reuse/review hint before creating a new owner file.

P4 shape lookup is exact. Prefer shape.typeLiteral or shape.hash in the intent when asking for shape reuse. Field-name-only shapes are reported as unknown, not fuzzy matches.

Report existing owners/helpers/types/dependencies in plain language, with compact proof only when needed.

Post-Write

Run after the same change when a pre-write advisory exists:

<audit-repo> --post-write --root <repo> --output <dir> --pre-write-advisory <advisory>

Check:

• silent-new type escapes
• planned-but-not-observed escapes
• scan-range parity against the pre-write run
• unexpected new files in the scan range outside intent.files
• advisory failures plus baseline, capability, and scan-range confidence

If post-write cannot find the advisory, say that the post-write check is unknown and name the missing file. Do not invent a clean result. Do not use pre-write-advisory.latest.json after another pre-write run has happened; rerun pre-write or pass the invocation-specific advisory.

Output

Use four short blocks when reporting to a human:

1. What I checked
2. What already exists / changed
3. What needs attention
4. Next command or coding-agent prompt

Keep raw JSON paths, FP ids, and tier names behind the answer unless the user asks for proof or maintainer detail.

Hand Off

If the user shifts to repo-wide structure, dead code, cycles, or a refactor-plan, hand off to lumin-repo-lens-lab. If the user shifts to canonical draft or drift validation, hand off to lumin-repo-lens-lab-canon.