Quality profile

The quality profile is a small sentinel-fenced block in your project's CLAUDE.md that opts the repo into CodeOps' quality loop: post-phase reviews by dedicated agents, security/perf audits, and workflow telemetry. No block, no behavior change — repos without a profile run exactly as before.

The block

## Quality profile (CodeOps)
<!-- CODEOPS-QUALITY:START -->
lenses: [security, concurrency]
security_profile: [auth-protocol]
perf_critical: false
review_hook: on
telemetry: on
agent_models: {}
<!-- CODEOPS-QUALITY:END -->

Onboard a repo with /setup_routing — it analyzes the code, proposes the routing block and the quality block together from the evidence it finds, and writes both only after you confirm. Hand-editing the block afterwards is fully supported.

Fields

Key	Values	Default	Effect
lenses	add-on lens names	[]	Extra review lenses beyond the always-on base (correctness, maintainability, standards)
security_profile	owasp-web, auth-protocol, financial-integrity, tenant-isolation, mcp-agent	[]	Activates the security auditor with the union of the named checklists, once per phase
perf_critical	true / false	false	Activates the perf auditor on code-touching phases
review_hook	on / off	on	off switches the whole loop off while keeping the profile on record
telemetry	on / off	on	Per-repo telemetry kill switch
agent_models	map agent → model, or → {model, effort}	{}	Per-repo model and effort overrides for the quality agents — see below

Add-on lenses: security, perf, api-surface, concurrency (which owns data-integrity). Parsing is lenient per key — an unknown key or value is warned about and ignored, never blocking. Both enums are grow-only.

What activates when

• Every executed phase (and non-trivial task mini-plan) ends with a parallel dispatch of the phase reviewer plus any active auditors on the phase's worktree snapshot — the whole change set, in any commit mode. Trivial tasks are never reviewed; docs-only diffs get the reviewer only.
• Findings gate commits: critical and major findings pause execution for your ruling in every commit mode — auto-commit automates the git operation, never the ruling. Minor findings are report-only. Accepted fixes are re-reviewed once, from a fresh snapshot scoped to the fix.
• Supersession: an active security profile replaces the reviewer's security lens; perf_critical replaces its perf lens — the same ground is never reviewed twice.
• Budget caps: one re-review per phase, at most; three codebase-scout dispatches per skill run; preflight fans out as five clustered auditor dispatches (--thorough goes per-dimension).

Per-repo model and effort overrides

Some repos want a harder reviewer than the plugin ships, or a cheaper scout. agent_models takes three value forms:

agent_models: {codebase-scout: opus, phase-reviewer: {effort: xhigh}, security-auditor: {model: fable, effort: max}}

Form	Meaning
name: model	Model only
name: {effort: E}	Effort only; the agent keeps its own model pin
name: {model: M, effort: E}	Both

Efforts are low, medium, high, xhigh, max — availability depends on the model (Sonnet has no xhigh), and an unavailable level degrades silently, exactly like an unavailable model.

Why effort works differently. Claude Code lets a model be redirected when an agent is dispatched, but reads effort only from the agent's own frontmatter. An effort override therefore has to exist as a file in .claude/agents/.

Do not write that file by hand. A hand copy of a plugin agent freezes its prompt at today's release and never receives another improvement — a silent, permanent fork to change one line. /setup_routing instead runs "${CLAUDE_PLUGIN_ROOT}/scripts/codeops-agents-sync.sh", which copies the plugin's body byte-for-byte, rewrites only the frontmatter, and stamps the result:

<!-- CODEOPS-GENERATED agent=phase-reviewer version=3.12.0 source=agents/phase-reviewer.md — … -->

That marker is the ownership boundary. Files carrying it are regenerated when the plugin moves on and pruned when you withdraw the override; files without it are yours, and the engine never touches them — so a genuinely customized prompt stays a deliberate choice. After upgrading the plugin, "${CLAUDE_PLUGIN_ROOT}/scripts/codeops-agents-sync.sh" --check reports (exit 1) if any generated agent is stale.

The full convention — packet contents, dispatch headers, model and effort resolution — lives in the plugin's _shared/quality-profile.md, which every dispatching skill links. See also Agents and Telemetry.