README · diff
git:20260523.b49e376 to git:20260711.4f85074
24 added, 3 removed. Audit A to A.
What this update changes in the checks
No check changed: every fixed check gives the same answer before and after this update.
A tripped check is a label, not a verdict: read the matched text and the lines below. This week's updates that newly trip a check.
# Agent catalog
Subagents run in their own context window and return only a summary, so deep analysis never
- bloats the main conversation. They are read-only (no edits). Claude dispatches them
- automatically when the trigger matches (each is described "Use PROACTIVELY"); you can also force
- one — e.g. "run the security-reviewer on these changes."
+ bloats the main conversation.
+ ## Proactive agents
+
+ Claude dispatches these automatically when the trigger matches (each is described "Use
+ PROACTIVELY"); you can also force one — e.g. "run the security-reviewer on these changes."
+ They are read-only (no edits).
+
| Agent | When to use | Model · tools |
| :---- | :---------- | :------------ |
| `convention-checker` | Before a commit, or after editing ≥3 files across `apps/`/`services/`/`packages/`. Fast audit against `docs/conventions/*`. | Haiku · Read/Glob/Grep |
| `migration-reviewer` | After editing `packages/acme-db/src/schema/**` or running `pnpm db:generate`. Checks schema + migration safety and backwards-compatibility. | Sonnet · Read/Glob/Grep |
| `security-reviewer` | After editing auth, API routes, WebSocket handlers, forms, or secret handling. Checks injection, broken access control, secrets exposure, OWASP Top 10. | Opus · Read/Glob/Grep/Bash |
| `architecture-explainer` | A why/how question about service boundaries, data flow, or cross-service interactions. Answers grounded in `docs/`, never invented. | Sonnet · Read/Glob/Grep |
+
+ ## Dispatched agents
+
+ These are workers for a specific skill or hook, not proactive triggers. The `review-*` family
+ is spawned (relevance-gated, model-tiered) by the [`pr-ci-review` skill](../skills/pr-ci-review/SKILL.md);
+ `comment-pruner` is dispatched by the `comment-pruner.sh` Stop hook.
+
+ | Agent | What it does | Model · tools |
+ | :---- | :----------- | :------------ |
+ | `comment-pruner` | Prunes low-value comments added in the current session, delete-when-uncertain, with a hard floor for tooling directives (`@ts-expect-error`, `eslint-disable`, …). The one agent here that edits. Also usable manually for a repo-wide sweep. | Sonnet · Read/Edit/Grep/Glob/Bash |
+ | `review-context` | Spec/protocol contradictions and infrastructure anti-patterns, each finding grounded in the contract it breaks. | Sonnet · Read/Glob/Grep/Bash |
+ | `review-conventions` | Audits changed files against `docs/conventions/*`, accepted ADRs, and `CLAUDE.md`, quoting the exact rule violated. | Sonnet · Read/Glob/Grep/Bash |
+ | `review-correctness` | Logic/behavior defects deterministic tooling can't catch, including silent regressions on the surface tests don't cover. | Opus · Read/Glob/Grep/Bash |
+ | `review-docs` | Accuracy and usefulness of prose a change introduces; strict on noise comments and stale docs. | Sonnet · Read/Glob/Grep/Bash |
+ | `review-maintainability` | Structural regressions — wrong layer, duplication, needless indirection — each with a citable mechanism. | Sonnet · Read/Glob/Grep/Bash |
+ | `review-security` | Real, reachable security issues at trust boundaries, cross-referenced against the threat model. | Opus · Read/Glob/Grep/Bash |
+ | `review-validator` | Adversarial gate: tries to refute each finding against the actual code before it is posted or auto-fixed. | Opus · Read/Glob/Grep/Bash |
**Tuning:** `tools` is the minimum each needs; `model` is matched to the work (Haiku = fast/cheap,
Sonnet = balanced, Opus = deep reasoning). To add an agent, see [`.claude/README.md`](../README.md)
("New agent").