feature-recon

Feature Recon

A reconnaissance sweep, not an authoritative audit: it reports what it actually read, cites path:line for every claim, and declares its blind spots. Output is one JSON state file per feature, a rollup, and a self-contained HTML dashboard built from them.

Bundled files live beside this SKILL.md (${CLAUDE_PLUGIN_ROOT}/skills/feature-recon/): reference/report-spec.md, reference/lens-security.md, reference/lens-ux.md, build_report.sh, template.html. The review agents live at ${CLAUDE_PLUGIN_ROOT}/agents/. Always pass absolute paths.

build_report.sh is the only entry point you call: it runs whichever of python3 / node the machine has. Never call build_report.py or build_report.js directly, and never assume a runtime — if the wrapper reports that neither is installed, say so and stop at step 6, where the JSON state files are already complete and useful on their own.

Procedure

1. Resolve arguments

• <recon-dir> — default docs/recon, or --dir <path>.
• Explicit feature list, if the user gave one → skip step 3.
• --sequential → skip the fan-out in step 4 and sweep features one at a time in this context.
• --lens <list> — which review lenses to run: product (the default, and what a plain run does), security, ux, or all. Comma-separated. Never add a lens the user did not ask for: each one multiplies the agent count by the feature count, and the specialists exist to be opted into.

2. Orient once

Gather this once and reuse it in every subagent prompt, so N agents don't each re-derive it:

git rev-parse --short HEAD && git log -1 --date=short --format=%ad

Plus, from a quick look at the repo root: project name, stack/framework, and where these live — entrypoints/routes, domain modules, models/migrations, UI pages, tests. Read the root README and CLAUDE.md/AGENTS.md if present.

Two more, because the sweep cannot find its best findings without them:

• Shared abstractions — the base controller/job/service classes, the middleware stack, how auth and tenancy are enforced, the shared HTTP/AI/queue client. These are what make a sibling comparison possible ("this job skips the guard its four siblings apply") and what turn eight identical findings into one cross-cutting entry.
• Test conventions — how tests are named, where they live, how they are run.