core

Review a diff or package for correctness, security, and compliance with the project's architecture and discipline rules. Read-only; produces a severity-ordered report. Use before merging any PR — one pass on the full branch diff, never iterative re-review loops.

Tier-up second opinion on the main loop's own work — designs, plans, diffs, invariant-adjacent decisions, anything where being wrong is expensive. A duplicate of the built-in `advisor` tool with caller-controlled model selection (the `advisorModel` settings key is one fixed value for all sessions; this agent picks per consult). INVOCATION RULE — the caller selects the model one tier above the session model via the Agent tool's `model` param: haiku→sonnet, sonnet→opus, opus→fable; if the session already runs Fable (the top tier), omit the override — `inherit` yields Fable. On Fable sessions the built-in `advisor` tool is unsupported (as of 2026-07), so this agent is the only advisor path there. Advisory only: it judges and recommends, never edits. Pack the full decision context into the prompt — the question, options considered, constraints, and pointers to the relevant files.

Audit the .claude/ steering layer itself — frontmatter validity, model-tier consistency vs the CLAUDE.md tiering rule, dangling skill/agent/command/hook references, and broken docs/ and [[memory]] cross-links. Read-only; produces a severity-ordered report. Use after editing any .claude/ config, or before merging a PR that touches it.

Check that code matches the specs/docs/architecture and that discoveries were written back in the same PR. Flags drift and proposes the doc/skill/ADR edits to fix it. Use only on PRs whose diff touches a docs surface (specs, design docs, skills, ADRs, CI workflow) or at a milestone close — it is the heaviest per-call agent; no docs surface, no invocation.

Read-only code/document search and locate across the repo (and any pinned external-dependency source) — broad fan-out to find files, symbols, naming conventions, and call sites, returning conclusions as `file:line` anchors, not file dumps. Use this instead of the built-in `Explore` so search runs on Sonnet (the model floor) rather than the inherited session model. It locates; it does not review, audit, edit, or decide.

Use when planning or executing any multi-step task to decide which model does what. Routes mechanical, low-judgment work (boilerplate, transcription, config, docs, repetitive edits, and all git/GitHub ops like commit/push/PR/CI-watching) to a cheaper Sonnet 5 subagent, keeps judgment-heavy work on the strong model, and verifies delegated output. Quality is paramount; tokens are not wasted on mechanical workflows.

Use to run a deep, multi-source research report efficiently — the tiered, budget-guarded, checkpointed version of a fan-out research workflow. Trigger whenever you would reach for the bundled `deep-research` harness, or any time you fan out web search → fetch → verify → synthesize across many subagents. Encodes the fix for this repo's ~1.95M-token burn: per-stage model pins, a budget guard, per-phase checkpoint-to-memory, and synthesis that runs even when partial.

Use for the git/GitHub development flow — creating branches, opening and merging PRs, waiting on CI, branch protection. Trunk-based (GitHub Flow): branch off main → PR into main → gates green → rebase-merge → delete. Trigger before any branch/PR/CI/merge operation. Delegate CI-waits and routine gh ops to a Sonnet subagent; keep the merge decision on the main loop. Repo-agnostic; a repo may pin specifics (branch routing, required checks) in its own CLAUDE.md.

Use at the start of any milestone build, large review, or migration that will spawn subagents — the tiered orchestrator-worker model: the strong model orchestrates and does the judgment work, tiered Haiku/Sonnet workers do the legwork, workers consult main, critics consult no one, and deterministic gates (tests, linters, type-checkers, schema/contract validators) outrank every LLM judge. Trigger from any task you'll parallelize across subagents. The full model behind the cost-aware-delegation skill. Validated against Anthropic's published multi-agent guidance (see Grounding).

1 hook across 1 event

Terse principal-engineer voice with deep cross-domain systems expertise; verifies before asserting, never fabricates an API, reasons about failure modes and security by default, self-reviews before declaring done.