Agents

The plugin ships twelve subagents in its agents/ directory: two plan-task executors (used by model routing) and, since 3.10.0, ten quality agents that power the profile-gated quality loop. All are dispatched by the skills — you never invoke them directly.

Roster

Agent	Model / effort	Tools	Role
plan-task-executor	sonnet / medium	read + write + bash	Executes one dispatched trivial/standard unit from an execution plan
plan-task-executor-opus	opus / high	read + write + bash	Executes one dispatched complex/sensitive unit
phase-reviewer	opus / high	read-only + bash	Reviews one phase diff through the base + add-on lenses; RV findings
spec-test-author	fable / high	read + write + bash	Writes spec tests implementation-blind from the packet; confirms the red phase
security-auditor	fable / high	read-only + bash	One dispatch per phase with the union of active security checklists; SA findings
preflight-auditor	opus / high	read-only + bash	Audits one artifact against one preflight dimension cluster; PA findings
design-challenger	fable / high	read-only, no bash	Independent second opinion on a decision, blind to the dispatcher's pick
perf-auditor	opus / high	read-only + bash	Hot paths, allocations, complexity, N+1s, blocking I/O; PE findings
concurrency-auditor	inherit / high	read-only + bash	Races, deadlocks, lost updates, unsafe retry; each finding carries an interleaving; CA findings
financial-integrity-auditor	inherit / high	read-only + bash	Idempotency, precision, atomicity, audit trails on money movement; FA findings
semantics-reviewer	inherit / high	read-only + bash	Formal semantics across parsing, typing, lowering, diagnostics, compatibility; SR findings
codebase-scout	sonnet / low	read-only	Facts with file:line only; honest "not found"; capped at 3 per skill run

The three specialists inherit the session model rather than pinning a tier, so they run at whatever you are already running at. Raise or lower them per repo through agent_models like any other agent.

Supersession

A dedicated auditor supersedes the matching dimension in a shared reviewer for that phase, so the same ground is never covered twice at two different depths:

When this dispatches	It supersedes	In
security-auditor	the security lens	phase-reviewer
perf-auditor	the perf lens	phase-reviewer
concurrency-auditor	the concurrency lens	phase-reviewer
financial-integrity-auditor	the financial-integrity checklist	security-auditor, which still runs for its other checklists
semantics-reviewer	(nothing — no shared reviewer covers formal semantics)	—

Supersession is written into both prompts, not just the profile: the specialist claims the ground and the shared reviewer is told to stand down. Each dispatch packet names the dimensions withdrawn for that phase, and a reviewer that skipped one says so in its report — a dropped dimension has to be visible, never assumed.

Activation

Agent	Activates on
security-auditor	security_profile is non-empty
perf-auditor	perf_critical: true and the diff touches code
concurrency-auditor	lenses contains concurrency and the diff touches code
financial-integrity-auditor	security_profile contains financial-integrity
semantics-reviewer	compiler-and-language is among the selected domains and the diff touches code

The semantics reviewer is the one agent keyed on a domain rather than a profile field, because compiler-and-language already means "this system has formal transformation semantics". Domains are classified for every repo, but this dispatch still requires a quality profile — as every agent does. See Domains.

Packet contracts

Every quality dispatch carries a machine-readable header on line 1 — [codeops-dispatch agent=<name> feature=<slug> phase=<id>] — plus a self-contained packet: the reviewers/auditors get the phase worktree snapshot, task lines, active lenses, and verify context; the spec-test-author gets spec excerpts, planned interfaces, and a FORBIDDEN implementation-file list it must never open; the challenger gets the problem and options without the dispatcher's preference. The canonical packet definitions live in the plugin's _shared/quality-profile.md.

For telemetry, the agent is identified from the dispatch tool's own subagent_type, not from that header — so a run is attributed even on the dispatch paths that do not carry one. A dispatch counts as CodeOps when its subagent_type is codeops:<name> or a bare <name> matching an agent the plugin ships, which also attributes project-local overrides in .claude/agents/. Ordinary agent use (Explore, general-purpose, your own agents) is recorded without an agent name, so it never appears in per-agent statistics. The header remains the only source of feature and phase, and is still read as a fallback when subagent_type is absent.

All finding-producing agents are read-only — they never edit, fix, or commit — and report "no findings" explicitly rather than returning empty output.

Model fallback

A pinned model that is unavailable on your account (for example, absent from an organization's model allowlist) silently falls back to the session model. There is no error and no warning — if review quality seems off on a restricted account, check this first. The same is true of an effort level the model does not offer.

Customizing an agent for one repo

Per-repo model and effort overrides go through the quality profile's agent_models map (see Quality profile) — for example {phase-reviewer: {effort: xhigh}}.

Do not copy an agent file into .claude/agents/ to change its model or effort. A copy shadows the plugin's agent permanently and freezes its prompt at the release you copied from, so every later improvement silently passes the repo by. Overrides carrying an effort are generated and version-stamped for you by "${CLAUDE_PLUGIN_ROOT}/scripts/codeops-agents-sync.sh", which /setup_routing runs; --check afterwards tells you whether anything has gone stale.

Copying an agent file is still the right move when you want a genuinely different prompt. Such a file carries no CODEOPS-GENERATED marker, and the sync engine leaves it alone permanently.