[calvin-archastro]

Review on ArchCode

Merge order: after ArchAstro/firstlanding#21117 ships in a release. This skill text describes that CLI's behavior: every applicable rule is judged, and not_evaluated shows as one warning per task.

Problem and author intent

archdev tasks graph import and archdev tasks review judge each plan node's brief against the repository's rules (plan.rule-verdicts, Jev). The node's scope drives that check: git pathspecs in scope pick the rules for the touched languages and supply up to 8 KB of repository context. A scope entry that matches no files selects every rule with no context.

The tasks skill:

• showed "scope": ["Export API"] in its example, a label that matches no files;
• never mentioned the repository-rule check, its --allow-lint override, or how to read not_evaluated. The firstlanding task that added the check (tsk_69hAt30rUu5VfmR1uFMiR2) listed this skill update as remaining work.

Intent: agents write scope as paths and clear lint errors before opening the review.

What changed

• Field guidance: scope lists the repository paths the task changes (files, directories, git globs), not a prose label, with the reason.
• Example: both nodes use path scopes.
• Section 3, step 5: run tasks graph import --dry-run. Fix errors in the node, or add an Exemplar: line. Pass --allow-lint / allow_lint only with the human's approval. not_evaluated means the model was unavailable and does not block. Use extract run plan.rule-verdicts --plan … --json to see every verdict.
• Harness copies under plugins/*/skills/tasks/ were regenerated with tools/plugin-gen (npm run build).

Scope: docs/skill only.

Risk

Low. Skill text only. The added commands exist in the released CLI. The new dry-run step adds about 15–25s of model time per plan on a cold run.

User impact

Agents using the tasks skill write path scopes and run the plan lint before handing a plan to the human, so rule violations show up before review instead of at save.

Testing

• tools/plugin-gen: npm run build regenerated 180 files, npm test passes, npm run check passes ("180 generated files match the catalog").
• Ran the skill's example plan through archdev --json tasks graph import --dry-run (firstlanding branch build). It parses and lints with the new path scopes. Inside firstlanding it is blocked by that repo's infrastructure-features-have-local-proof-paths rule, which is expected for a generic example.
• Verified flags against the CLI: tasks graph import --dry-run --allow-lint, extract run plan.rule-verdicts --plan --json; dry-run JSON has a top-level lint list.
• No end-to-end test applies: this is skill text with no executable path.

🤖 Generated with Claude Code