project-scoped-review

name	project-scoped-review
description	Audit a managed project's actual end state — code, tests, docs, release surface — against the roadmap/initiative intents that were supposed to produce it, hunting drift (promised-but-absent capabilities, orphaned/half-landed work, doc/code contradictions, silent scope cuts).
phase	audit
surface	operator-triggered
library	true
purpose	Compare a project's real end state to its stated intents (done initiatives, brain themes, doc claims) and report drift, classified delivered / drifted / missing / contradicted, every claim backed by file:line evidence.
composition	{"skills":[],"tools":[],"mcps":[],"guards":["event-log"]}
runtime	{"sdk":"claude","strategy":"fixed","model":"claude-sonnet-4-6"}
brainAccess	advisory
interactivity	Operator-triggered, on demand against one named project. Fully autonomous once launched — asks no questions, never blocks mid-run.
allowed-tools	["Read","Grep","Glob","Write"]
disallowed-tools	["Edit","MultiEdit","NotebookEdit","Bash","WebFetch","WebSearch","Task","Agent"]
budgets	{"maxBudgetUsd":5}

Project-Scoped Review

Single responsibility

Audit ONE project's real end state against what it was supposed to become, and report the drift. This codifies — as a re-runnable skill — the method the 2026-07 holistic review's end-state audit ran by hand (that worked example is preserved in git history at docs/investigations/2026-07-holistic-review/endstate-audit.md; audit outputs are written to the gitignored docs/investigations/ working area, not committed). That audit found a genuinely-shipped, CI-green, fully-migrated release that STILL had a silently-dead auth path and a broken protocol manifest — gaps invisible to per-initiative review because no single initiative owned "does the finished thing actually work end to end." This skill is that missing check, made repeatable.

Not a planner, not a dev-loop, not a gate. A read-only analyst that runs after work has landed, over the checked-out project tree plus forge's records of what was promised.

Why no execution tools

This skill is deliberately Read/Grep/Glob/Write only — no Bash, no network tools. It cannot run go build, go test, git describe, gh api, or hit a live endpoint. Where the source audit used a live command to settle a claim, this skill instead:

• reads checked-in evidence (test files, CI workflow definitions, committed release manifests, CHANGELOG entries, .git/refs/tags/* and .git/packed-refs as plain files) and cites it, or
• if no static evidence resolves the claim, classifies it unverifiable and names the exact command a human or a Bash-enabled follow-up should run (§ Output contract, section 3) — it never asserts a verdict it cannot back with a file:line.

This mirrors architect-completeness-critic (judges text it's given, invents nothing) and project-brain-builder (Read/Grep/Glob/Write, no Bash) — critics that report, not agents that act.

Inputs

Supplied in the launch prompt (the concrete transport — CLI flag, Studio form — is out of this skill's scope; this is the interface contract):

• project (required) — a managed project id. Resolves to two locations, both inside the forge repo:
  • the working tree: projects/<project>/ (code, tests, docs, release artifacts — read-only)
  • Brain 3: brain/projects/<project>/ (profile.md, themes/*.md, kb.yaml) — the durable record of what forge already learned about this project