Agents Catalog

Agents are named specialists invoked explicitly for a bounded kind of work. Each agent defines its explicit scope, hard constraints, step-by-step approach, and falsifiable completion criteria (Done When).

Overview Table

Agent Name	File	Role & Mandate	Domain Guide	Key Constraints	Paired Skill / Instruction
Thinker	think.agent.md	Explores options and blast radius, producing the Think record (think.md).	Think & Plan	No production or test code changes; 2 to 4 options; blast radius classified; decision records drafted, never signed; no code blocks.	think instruction, think skill
Planner	plan.agent.md	Turns a validated Think record into the Plan record (plan.md) with file scope, criteria, and atomic steps.	Think & Plan	Derived strictly from think.md; no source code blocks; preserve bootability; blast radius verified.	plan instruction, plan skill
Code Agent	code.agent.md	Implements the validated Plan steps in dependency order, runs the checks and controls until green, and opens the pull request with its record.	Code	Edits only a step's Files; never weakens or edits a control; no production credentials; sets no human validation; stops at the fix-loop limit.	code skill, code instruction
Prove Validator	prove.agent.md	Executes the full Prove verification suite and performs semantic review of the PR diff against think.md and plan.md, producing specs/issue-<n>/prove.md.	Prove	Never edits application or test code; never weakens a control; never claims line-by-line review; validates prove.md with validate-prove.mjs.	prove skill, prove instruction
Gate Fixer	gate-fix.agent.md	Diagnoses and fixes a failing deterministic gate within the step and file scope (agent auto-remediation).	Prove	Never weakens, skips, or edits a control; never touches protected paths; stops at the fix-loop limit.	prove skill, prove instruction
Learn Analyst	learn.agent.md	Watches what a release does in production: detects and explains anomalies, proposes hypothesis verdicts, drafts incident records, runs the context review, and emits Triage issues.	Learn	Never decides, mitigates in production, holds production credentials, or writes a verdict.	telemetry-scan, hypothesis-ledger, incident-record, learn-signals, context-health skills
Release Planner	release.agent.md	Prepares and validates release.md for automated authorization.	Release	Never deploys, holds production credentials, loosens thresholds, or fabricates human approval.	release-plan skill
Decision Scribe	decision-scribe.agent.md	Chooses and records attributed autonomous decisions, or escalations.	Governance & Decisions	No source changes, fabricated human signatures, or waived controls.	decision-record skill
Meeting Analyst	meeting-pack.agent.md	Turns a client meeting transcript into an immutable Meeting Pack with candidate business needs.	Meet	Zero code changes; verbatim client quotes only; transcript never edited; never confirms.	meeting-pack skill
Triage Analyst	triage.agent.md	Turns a candidate business need into an evidence-linked triage.md with a scorecard and a proposed decision.	Triage	Zero code changes; verbatim evidence only; never sets the final decision or a blast radius tier.	triage skill
Dead Code Auditor	dead-code.agent.md	Audits a repository for unused exports, orphan files, dead branches, and redundant dependencies.	Codebase Audits	Read-only; verify dynamic references and public package.json entrypoints before flagging.	Structured audit-findings.json & audit-report.md
Test Gap Auditor	test-gap.agent.md	Audits a repository for untested public APIs, missing edge-case coverage, and weak assertions.	Codebase Audits	Read-only; prioritize high-risk public contracts and error paths over line coverage.	Structured audit-findings.json & audit-report.md
Docs Sync Auditor	docs-sync.agent.md	Audits documentation (README.md, docs/, JSDoc) against actual exported APIs and CLI flags.	Codebase Audits	Read-only; report concrete factual drift only, never stylistic preferences.	Structured audit-findings.json & audit-report.md
Dependency Leverage Auditor	dependency-leverage.agent.md	Audits how effectively a repository leverages its installed dependencies and upstream ecosystem libraries.	Codebase Audits	Read-only; distinguishes local fixes from upstream library escalations.	Structured audit-findings.json & audit-report.md
Spec from Code	spec-from-code.agent.md	Reverse-engineers technical specifications from existing code.	Specs & Docs	Grounded strictly in source code; no fabricated features; comprehensive interfaces.	spec-from-code skill
Unit Tester	unit-test.agent.md	Writes or updates isolated unit tests with full branch and edge-case coverage.	Testing & Verification	One module scope; edit test files only; never change production code without explicit consent.	Installed domain test instruction
API Tester	api-test.agent.md	Authors or updates Supertest HTTP API route integration tests.	Testing & Verification	Route scope only; Supertest assertions; edit test files only; zero production code altered.	node-api-tests instruction
E2E Tester	e2e-test.agent.md	Writes or updates Playwright browser end-to-end tests for user journeys.	Testing & Verification	e2e/ scope only; POM pattern; edit test files only; zero production code altered.	angular-e2e-tests instruction
VitePress Docs	vitepress-docs.agent.md	Scaffolds or updates a VitePress documentation site in a user-named folder.	Specs & Docs	Confine all files to docs root; follow vitepress-docs skill; default theme only.	vitepress-docs skill
Demo Videos	demo-videos.agent.md	Films user journeys with Playwright and encodes them as MP4 + poster for a website.	Specs & Docs	Edits demos/** and the output folder only; local dev stack only; repeatable runs; UI defects reported, not fixed.	playwright-demo-videos skill
Performance Tester	performance-tests.agent.md	Designs, scaffolds, or extends k6 load, stress, and spike test suites.	Testing & Verification	Follow k6 Docker execution pattern; assert realistic thresholds; isolate scenarios.	k6-performance-tests skill
Fuzz Tester	fuzz-tests.agent.md	Configures and runs RESTler API fuzzing pipelines.	Testing & Verification	OpenAPI-driven grammar compilation; handle auth refresh; fail on 500 status.	restler-fuzzing-tests skill
Migration Documenter	migration-docs.agent.md	Documents an infrastructure migration (for example on-premise to cloud): assessment, migration plan, and cutover runbook.	Migration	Documents only; infrastructure absorbs differences before any application change; every change is registered; no secret values.	migration-docs skill and instruction

---

Agent Deep Dives

1. Dead Code Auditor (dead-code.agent.md)

• Role: Identifies unused exports, orphan modules, unreachable code paths, and redundant dependencies.
• Hard Boundaries: Read-only; never flags public entrypoints declared in package.json (main, module, exports, bin).
• Completion Gate: Produces valid audit-findings.json and audit-report.md with file paths and line-level evidence.
• Full Workflow: See Codebase Audits.

2. Test Gap Auditor (test-gap.agent.md)

• Role: Audits public APIs, classes, CLI commands, and error/boundary paths for missing or tautological tests.
• Hard Boundaries: Read-only; zero source or test edits; prioritizes behavioral risk over line-coverage percentages.
• Completion Gate: Produces valid audit-findings.json and audit-report.md with concrete test scenarios and acceptance criteria.
• Full Workflow: See Codebase Audits.

3. Spec from Code (spec-from-code.agent.md)

• Role: Reverse-engineers technical specifications and data models directly from source code.
• Hard Boundaries: Grounded strictly in observable code; no undocumented or assumed capabilities.
• Completion Gate: Area specification markdown files pass skills/spec-from-code/scripts/spec-docs.mjs validation.
• Full Workflow: See Technical Specification Pipeline.

4. Unit Tester (unit-test.agent.md)

• Role: Authors or improves isolated unit tests with complete branch coverage.
• Hard Boundaries: Strictly one module in scope. Edits test files only; never alters production code without explicit approval.
• Completion Gate: All execution paths asserted, narrowest test command passes with exit code 0.
• Full Workflow: See Unit & Component Testing.

5. API Tester (api-test.agent.md)

• Role: Authors and maintains Supertest HTTP API route integration tests.
• Hard Boundaries: Strictly route module and tests/routes/ scope. Edits test files only; drives full Express middleware stack; mocks at database and external HTTP boundaries.
• Completion Gate: All route verbs, status codes, and response envelopes asserted; narrowest route test command passes with exit code 0.
• Full Workflow: See HTTP API Testing.

6. E2E Tester (e2e-test.agent.md)

• Role: Authors and updates Playwright browser end-to-end tests.
• Hard Boundaries: Confined to e2e/. Enforces Page Object Model (POM). Production code is read-only.
• Completion Gate: All journey checkpoints pass headlessly against the containerized dev stack.
• Full Workflow: See End-to-End Testing.

7. VitePress Docs (vitepress-docs.agent.md)

• Role: Scaffolds and maintains product documentation websites using VitePress.
• Hard Boundaries: Requires an explicit user-named root (e.g. website/ or docs/). Confined strictly to docs directory.
• Completion Gate: npm run build succeeds with exit code 0 and all guide pages are indexed in the sidebar navigation.
• Full Workflow: See Product Documentation Website.

8. Performance Tester (performance-tests.agent.md)

• Role: Designs and executes automated k6 API performance and load benchmarking.
• Hard Boundaries: Parameterized scenarios with explicit SLA thresholds; runs via Docker runner.
• Completion Gate: k6 scenarios execute cleanly and SLA thresholds pass.
• Full Workflow: See Performance & Load Testing.

9. Fuzz Tester (fuzz-tests.agent.md)

• Role: Configures and runs Microsoft RESTler stateful API fuzzing.
• Hard Boundaries: Grammar compilation from OpenAPI contract; handles auth refresh; zero 500 error tolerance.
• Completion Gate: Fuzz-lean sequence completes with zero unhandled 500 exceptions.
• Full Workflow: See API Fuzz Testing.

10. Thinker (think.agent.md) & Planner (plan.agent.md)

• Thinker Role: Explores 2 to 4 options from a prioritized backlog issue, classifies blast radius, and writes the think.md record without code.
• Planner Role: Takes a validated think.md and breaks it down into explicit file scopes, observable acceptance criteria, ordered atomic steps (Files, Action, Verification, Bootable Check, Serves), and test plan in plan.md without writing source code.
• Hard Boundaries: No production or test code changes and no source code blocks in think.md or plan.md. Completed decisions and approvals use explicit automation identity and date, never fabricated human signatures.
• Completion Gate: think.md passes skills/think/scripts/validate-think.mjs and plan.md passes skills/plan/scripts/validate-plan.mjs validation with exit code 0.
• Full Workflow: See Business Needs to Specifications.

11. Triage Analyst (triage.agent.md)

• Role: Enriches a candidate business need with evidence and system mapping, fills the scorecard, and proposes a decision in triage.md, without writing code. In review mode it reports the ranked backlog for the weekly Triage review.
• Hard Boundaries: Zero code changes; codebase is read-only; one need per run; quotes verbatim from real sources; leaves Final (PO) as Pending and assigns no blast radius tier; never edits the ledger or weights.
• Completion Gate: triage.md passes skills/triage/scripts/validate-triage.mjs --stage proposed with exit code 0.
• Full Workflow: See Candidate Need to Backlog Issue.

12. Meeting Analyst (meeting-pack.agent.md)

• Role: Turns a plain-text client meeting transcript into a normalized transcript.md and a Meeting Pack (pack.md) with an executive summary, domain terminology, and candidate business needs backed by verbatim client quotes. In read-back mode it prepares the Product Owner's list and writes nothing.
• Hard Boundaries: Zero code changes; never edits a transcript line, the glossary, another meeting, or a confirmed pack; every quote verbatim and attributed as Name (Role); leaves all four confirmations Pending; stops without recording consent or on secrets; one meeting per run.
• Completion Gate: transcript.md is unchanged from the normalizer and pack.md passes skills/meeting-pack/scripts/validate-meeting.mjs --stage draft --glossary docs/glossary.md with exit code 0.
• Full Workflow: See Transcript to Meeting Pack.

13. Code Agent (code.agent.md)

• Role: Implements a validated plan.md in dependency order: implements each step's Action, runs the checks and controls (scope-guard, secret-guard, dependency-guard, control-guard, test-probe, preview-smoke, criteria-verify), repairs failures within the fix-loop limit, commits on a short-lived branch (branch-guard.mjs), and opens the pull request with the code skill's PR record body.
• Hard Boundaries: Edits only a step's Files; never weakens, skips, or edits a control; never holds production credentials; never sets the Human Validation fields; records a context-correction suggestion instead of hand-fixing a misunderstood pattern; stops for the Developer at the fix-loop limit.
• Completion Gate: Every step's Verification and Bootable Check pass, criteria-verify confirms every AC-<n> is covered by a passing test, and validate-pr.mjs --stage code passes.
• Full Workflow: See Code.

14. Gate Fixer (gate-fix.agent.md)

• Role: Reads a failing deterministic gate's output, diagnoses the exact fault, and fixes the code within the step and file scope (agent auto-remediation).
• Hard Boundaries: Can never weaken, skip, edit, or waive a control; never touches protected paths; re-runs the failed control and control-guard after every fix; records each attempt and stops at the fix-loop limit.
• Completion Gate: The failed gate passes, control-guard reports no finding, and the attempt is recorded.
• Full Workflow: See Prove.

15. Release Planner (release.agent.md)

• Role: Drafts release.md for one merged change: tier and rollback from Think, the outcome metric from the Triage hypothesis, rollout stages, and thresholds proposed from the tier defaults.
• Hard Boundaries: Never deploys, never holds production credentials, never disables a threshold; DevOps approves.
• Completion Gate: release.md passes validate-release.mjs with --think.
• Full Workflow: See Release.

16. Decision Scribe (decision-scribe.agent.md)

• Role: Drafts decision records and escalations from a Think record or a described decision.
• Hard Boundaries: Completed decisions may be Accepted with explicit Automation attribution. Incomplete decisions stay Drafted and unsigned; never fabricate human sign-off or weaken controls.
• Completion Gate: The record passes validate-decision.mjs.
• Full Workflow: See Governance & Decisions.

17. Learn Analyst (learn.agent.md)

• Role: Owned by the QA and maintainer. In its modes it detects and explains (telemetry scan plus correlated deploys), presents proposed hypothesis verdicts, drafts and validates incident records, answers "what should a control have caught?", runs the Persistent Context health checks, and emits Triage issues so the loop closes.
• Hard Boundaries: Never decides, never mitigates in production, never holds production credentials, never writes a ledger verdict.
• Completion Gate: Every artifact it produces passes its skill's validator, and every Learn output is a Triage issue.
• Full Workflow: See Learn.

18. Migration Documenter (migration-docs.agent.md)

• Role: Inventories the current infrastructure from the repository and manifests, maps each component to the target, registers the unavoidable application changes, and writes the migration plan and cutover runbook.
• Hard Boundaries: Writes only under docs/migration/; never modifies application code, infrastructure code, or pipelines; resolves gaps infrastructure-side first; no secret values; a Refactor decision record stays drafted and unsigned.
• Completion Gate: The records pass skills/migration-docs/scripts/validate-migration.mjs --dir with exit code 0, and scope-guard and secret-guard are clean.
• Full Workflow: See Migration.

19. Prove Validator (prove.agent.md)

• Role: Executes the 8 deterministic controls on a pull request branch (control-guard, secret-guard, dependency-guard, scope-guard, task-gate, criteria-verify, test-probe, preview-smoke) and performs semantic review of the PR diff against think.md and plan.md across Client Intent, Behavioral Experience, and Safety & Blast Radius, writing specs/issue-<n>/prove.md.
• Hard Boundaries: Writes only specs/issue-<n>/prove.md; never edits application or test code; never weakens or skips a control; never claims line-by-line review.
• Completion Gate: specs/issue-<n>/prove.md passes skills/prove/scripts/validate-prove.mjs --dir specs/issue-<n> with exit code 0.
• Full Workflow: See Prove.

20. Demo Videos (demo-videos.agent.md)

• Role: Films named user journeys as short demo videos for a product website.
• Hard Boundaries: Edits only demos/**, the app demos script and gitignore, and the encoded files in the user-named output folder; production code is read-only; runs against the local dev stack only.
• Completion Gate: Each scenario passes twice in a row on the same database, and its MP4 + JPG poster are reviewed in the output folder.
• Full Workflow: See Product Demo Videos.

21. Docs Sync Auditor (docs-sync.agent.md)

• Role: Audits README.md, docs/, JSDoc comments, and usage examples against actual exported code, signatures, configuration keys, and CLI scripts.
• Hard Boundaries: Read-only; reports only concrete factual inaccuracies or missing public API documentation, never stylistic preferences.
• Completion Gate: Produces valid audit-findings.json and audit-report.md citing both the documentation file and the source code file.
• Full Workflow: See Codebase Audits.

22. Dependency Leverage Auditor (dependency-leverage.agent.md)

• Role: Audits how effectively a repository leverages its installed dependencies and identifies hand-rolled duplicates or reusable capabilities that belong in an upstream ecosystem library.
• Hard Boundaries: Read-only; never proposes new third-party dependencies; classifies each finding as scope: "local" or scope: "upstream".
• Completion Gate: Produces valid audit-findings.json and audit-report.md.
• Full Workflow: See Codebase Audits.