[mgratzer]

Whole-codebase reviews keep finding the same classes of problem, and their reports are lost with the session that wrote them. forge-survey records each finding in docs/surveys/ with its class, evidence, remedy and the rule or check that let it through, and turns a class that recurs into a guardrail.

Notes

• A tenth skill, 911 words. It is not part of the pipeline, so the lens is unchanged.
• The class table repeats terms that feat(skills): ask where a decision already lives, in the plan, the review and the risk verdict #154 defines in deep-modules.md (test seam, leftover of a replacement). The two PRs merge in either order; README is the only shared file.
• No unattended mode: triage asks the user, and a headless run stops with every finding open.
• Judgment calls (run unattended from another project's session): the name forge-survey, the eight classes, the record's home docs/surveys/.

Exercised

Snapshot forge-new at 3fcbd19, run headless in a fresh clone of the sandbox. There is no forge-old run: the skill is new.

• /forge-new:forge-survey: four fresh-context readers (three areas, one cross-module). It recorded six findings in docs/surveys/2026-11-10.md, each with class, file:line evidence and a Let through by cell, for example Quality gate in AGENTS.md:9 ("syntax check"): missing. It listed three possible bugs separately, dropped three weak candidates, proposed no guardrail (first survey), and opened sandbox PR feat(ship): local-first shipping and a separate publish step #56 (closed).
• Failure case, no one to ask: it printed the triage question, filed no Issue, and left every Status open.
• wrong: the run left the checkout on the docs/survey-2026-11-10 branch. Fixed in 73185cb (record from a temporary worktree); that fix was not re-run.
• Not exercised: a second survey that marks recurring findings and proposes a guardrail; a scoped run.
• Checks: scripts/check-skills.sh prints nothing. claude plugin validate and claude plugin test not run.

Try it: in any repository, /forge-survey, then open docs/surveys/<today>.md.

Plan

Plan as first made (no approver: the owner asked for the skill, item B6 of the design guardrails proposal, and did not see this plan)

• Design: Interface: /forge-survey [paths], output docs/surveys/<date>.md. Hides: how areas are split and verified. Already lives: nowhere in forge; forge-reflect reviews a diff, and its review flow is not reused because a survey has no diff.
• Phase 1: the skill, with its class table and record format.
• Phase 2: README table and count, docs/architecture.md, a Related Skills line in forge-setup-project, scenarios in docs/testing.md.
• Not changing: every other skill's process, the lens, the shared modules.

Deviations (as of 73185cb): none