Skip to content

AgentTeams — Feature Inventory

Version baseline: 1.1.0.dev0 · counts last reconciled by hand: 2026-07-29

Features below are grouped by capability area. This inventory may include capabilities currently present on main that are not yet part of a tagged release. Release-specific availability should be verified against CHANGELOG.md and repository tags.

The per-category counts are hand-maintained; the baseline is not

The counts are not derived from the code and will drift between reconciliations. Treat them as indicative and the category descriptions as authoritative; verify any specific capability against CHANGELOG.md or the CLI's own --help.

Deriving the counts requires a machine-readable feature set, which the module does not yet carry. Until it does, this note is the honest alternative to a number that looks computed and is not.

Three things here are checked, because they never required judgment — they were only ever hand-copied:

  • the version baseline above equals agentteams.__version__ (tests/test_api_doc_signatures.py::test_inventory_version_baseline_matches_the_installed_version). It had read 0.1.0 (2026-04-15) while the module was at 1.0.0rc6 — four months and a major version stale, so a reader checking whether a capability had shipped was reading a document describing a different release;
  • the Governance Agents row's members equal analyze.GOVERNANCE_AGENTS;
  • the Total equals the sum of its own column.

Summary

Category Features Description
Core Pipeline 39 Project ingestion, team generation, framework adapters, CLI, templates, schemas
Enrichment 11 Auto-enrichment, notebook scanning, tool catalog, live PyPI metadata
Section Fencing & Safe Merges 5 AGENTTEAMS:BEGIN/END markers, non-destructive merge mode
Security Intelligence 6 Live CVE/CISA-KEV/EPSS data, credential scanning, security reference files
Migration & Update 8 --migrate, --revert-migration, --enrich, --auto-correct, --scan-security, security flags, multi-workspace fleet update (--fleet)
Safety & Backups 6 Automatic backups, restore capability, --no-backup, --list-backups, public backup API
Governance Agents 11 navigator, security, code-hygiene, adversarial, conflict-auditor, conflict-resolution, cleanup, agent-updater, agent-refactor, repo-liaison, git-operations
Workflows 14 Workflows 1–12 with constitutional rules, guards, and final-check termination
Governance Infrastructure 12 @agent-updater auto-triggers, @adversarial guards, truth check, drift-as-trigger, deployment protocol
Interoperability 8 convert, interop, bridge modules; handoff manifest; CI flags
Bridge Automation 7 Daily maintenance script, CI workflows, staleness watchdog, deduplication
Cross-Repository Support 7 @repo-liaison agent, adjacent-repo tracking, impact/update/coordination protocols
Drift Trust & Delivery Gating 9 Structural drift diff, delivery receipts, behavioural-drift detection, trust gating
Retrieval & Review-Time Utilities 3 Memory index, code & API index, session scan
Feature Audit 5 Machine-readable feature registry, per-tier proof enforcement, reachability classification, end-to-end CLI probes

Total: 151 documented features across 15 capability areas — generated from references/feature-registry.csv, which is itself derived from the body of feature-inventory.md. Checked by tests/test_feature_registry.py::test_the_summary_table_is_generated_from_the_registry. It read 125 across 12 until 2026-08-07, having silently omitted two whole capability areas and undercounted the rest by nine: the only check compared the total to its own column, so a table that agreed with itself and with nothing else passed.

Artifact Selection Matrix

Use this matrix when deciding which module/artifact to rely on for a specific operational goal.

Goal Primary Module Primary Artifact Typical Trigger Notes
Detect template/structure drift drift references/build-log.json --check, --update Structural + template fingerprint analysis
Update many workspaces at once fleet <DIR>/.agentteams-fleet/<run-id>/report.json --fleet DIR --update --merge Git-commit snapshot per workspace + git diff content audit; non-destructive (merge-only); dry-run by default, --yes to apply
Verify behavioral conformance eval_suite + behavioral_drift references/eval-suite.json + trajectory payload Post-run behavioral validation Neutral suite + runtime trajectory comparison
Fast historical retrieval memory_index references/memory-index.json --refresh-index, --query-index Use lexical first, vector when thematic recall is needed
Cost/capability routing policy model_routing references/model-routing.json --cost-routing Tier-role contract, concrete models resolved downstream
Cross-framework eval execution eval_adapters Generated adapter output (Inspect module / OpenAI Evals JSON) Adapter export step Converts neutral eval-suite into framework-specific eval assets
Security intelligence refresh security_refs references/security-vulnerability-watch.reference.md + .json Normal generate/update path (unless offline) Live-data artifacts; intentionally non-deterministic
Rollback-ready update safety emit backup APIs .agentteams-backups/<timestamp>/ + _manifest.json --update, --merge, --overwrite Supports restoration and backup integrity checks

Core Pipeline

Team Generation

  1. Agent Teams Module — Generate a complete, coordinated AI agent team from a single project description file
  2. Four-Tier Agent Architecture — Orchestrator → Governance Agents → Domain Agents → Workstream Experts
  3. Project Description Ingestion — Load project briefs from .json or .md format
  4. Project Directory Scanning — Scan an existing project directory to supplement missing brief fields
  5. Project Type Classification — Infer the deliverable type from the project goal
  6. Agent Archetype Selection — Select the appropriate archetype set from the schema-supported archetype catalog, including contextual post-production auto-selection
  7. Team Manifest Generation — Build complete team composition metadata
  8. Template Placeholder Resolution — Resolve {AUTO:token} and {MANUAL:token} placeholders

Core Pipeline Modules

  1. ingest Module — Load and normalize project descriptions; parse JSON/Markdown; scan directories
  2. analyze Module — Classify project type; select archetypes (including contextual post-production auto-selection for any task domain and manual selected_archetypes override support); detect operational tool docs; build manifest
  3. render Module — Resolve placeholders in templates; compute template hashes for drift detection
  4. emit Module — Write rendered files to disk with dry-run and overwrite-protection
  5. drift Module — Detect content drift (template hash comparison) and structural drift (team composition)
  6. scan Module — Proactive security scan for PII paths, credentials, unresolved placeholders
  7. audit Module — Static post-generation audit; optional AI-powered review via Copilot CLI
  8. remediate Module — Auto-correction support via Copilot CLI after audit findings

Visualization & Graph

  1. graph Module — Directed graph inference for team topology
  2. Mermaid Output — Team topology diagram in Mermaid format
  3. DOT Output — GraphViz format for advanced visualization
  4. JSON Output — Machine-readable team topology
  5. Markdown Output — Human-readable topology; auto-generated in references/pipeline-graph.md

Framework Adapters

  1. copilot-vscode Adapter — .agent.md files with YAML front matter; outputs to .github/agents/
  2. copilot-cli Adapter — Plain .md system prompts; outputs to .github/copilot/
  3. claude Adapter — Claude Code front matter .md + CLAUDE.md; outputs to .claude/agents/
  4. Framework-Agnostic Interface — Extensible adapter base class for adding new frameworks

Two additional first-class adapters extend this set:

  • goose Adapter (beta) — Block / AAIF Goose recipe YAML (.goose/recipes/*.yaml, schema 1.0.0); orchestrator delegation encoded natively as sub_recipes (deeper edges become summon load(...)); team brief written to the repo-root AGENTS.md plus a .goosehints integrator. Also a valid --convert-from and --bridge-from target (not an interop target). Beta — generate/convert/bridge supported, interop planned; adapter API not yet under the stability contract. See frameworks.
  • agents-md Adapter — Cross-tool AGENTS.md standard (AAIF / Linux Foundation); emits a single framework-neutral repo-root AGENTS.md (read by ~10 AI coding tools) plus per-specialist detail under .agents/. Generate-only (not a convert/interop/bridge target).

A thinner prep-scoped adapter also ships:

  • codex Adapter — OpenAI Codex target (registry framework codex): emits each agent as a Codex custom agent .codex/agents/<slug>.toml (native generation and interop source/target), writes the repo-root AGENTS.md only when absent or Codex-generated, and splices operator-specified MCP servers into .codex/config.toml (see codex_mcp_emit). Generate-only — a valid --framework target; not a convert/interop/bridge target (codex is absent from the --bridge-source-framework choice list, so it is rejected at argument-parse time). See frameworks.

CLI

  1. agentteams / build_team.py CLI — Command-line interface wiring all pipeline stages
  2. --description — Path to project brief (.json or .md)
  3. --project — Target project directory
  4. --framework — Output framework: copilot-vscode, copilot-cli, claude, goose, agents-md, or codex
  5. --dry-run — Preview output without writing files
  6. --overwrite — Allow overwriting existing agent files
  7. --self — Self-maintenance: regenerate the module's own agent team
  8. --post-audit — Run static + AI-powered conflict and presupposition review

Templates & Schemas

  1. 11 Universal Governance Templates — One tier-2 template per governance agent archetype
  2. 20 Domain Templates — 14 project-type archetype templates (incl. content-enricher and work-summarizer) plus 6 tool templates
  3. 3 Builder Templates — Team Builder agent variants
  4. 6 Workstream Expert Templates — Component-scoped expertise patterns
  5. project-description.schema.json — JSON Schema for brief validation
  6. team-manifest.schema.json — JSON Schema for team manifest validation

The universal/ template library includes the tier-1 Orchestrator template and all 11 always-included tier-2 governance templates.

Enrichment

  1. enrich Package — Context-aware auto-enrichment pipeline
  2. Rule-Based Fill Generation — Pattern-matched placeholder completion
  3. Jupyter Notebook Scanning — Extract evidence and examples from .ipynb sources
  4. Tool Catalog Integration — Reference available tools and frameworks for auto-fill
  5. LLM Model Detection — Identify referenced models and optimize prompts
  6. Enrichment Coverage Audit — Enumerate unresolved defaults; export references/defaults-audit.csv
  7. Dynamic Tool Detection — Automatic framework/tool identification from project structure
  8. Auto-PyPI Metadata Fetch — Fetch package metadata for tool catalog enrichment
  9. Agent Integration Audit — Verify all agents have required integration details
  10. Agent Knowledge Updates — Incorporate discovered tool knowledge into agent content
  11. --enrich Flag — Run enrichment pass after generation

Section Fencing & Safe Merges

  1. AGENTTEAMS:BEGIN/END Fence Markers — Delineate managed regions in agent files
  2. FENCE-CONVENTIONS.md — Specification for fence marker usage and authoring rules
  3. Non-Destructive Merge Mode (--merge) — Update fenced regions only; preserve all user content outside fences
  4. --update --merge Safety Fix — --update now honors --merge; no longer silently overwrites user content
  5. 11 Instrumented Templates — All governance and domain templates ship with fence markers

Security Intelligence

  1. security_refs Module — Render live threat intelligence into generated security reference files
  2. Live CVE/CISA-KEV/EPSS Data — Pull current threat data on every pipeline run
  3. --security-offline Flag — Disable live intelligence fetching for air-gapped or offline runs
  4. --security-max-items Flag — Cap the number of live threat items rendered
  5. --security-no-nvd Flag — Skip NVD fetching while retaining CISA-KEV and EPSS data
  6. Non-Deterministic Snapshot Exclusion — test_snapshot_comparison skips live-data files

Migration & Update

  1. --migrate Flag — One-step fencing migration: tags HEAD as pre-fencing-snapshot, then regenerates all files with fenced templates
  2. --revert-migration Flag — Undo a migration by resetting to pre-fencing-snapshot git tag
  3. --update --merge Canonical Mode — Incremental re-render that preserves manual placeholder values, preserves user-authored content outside fences, and restores expected standard outputs missing on disk
  4. --prune Mode — Remove stale agent components from an existing team
  5. --auto-correct Flag — Invoke Copilot CLI to repair post-audit findings; reruns audit to confirm
  6. --scan-security Flag — Proactive scan for PII paths, credential patterns, unresolved placeholders
  7. available_workflows Fenced Section — Workflow documentation fenced so --update --merge propagates future changes while preserving project-specific rules
  8. Update Deployment Protocol & Fleet Update (--fleet DIR) — A documented multi-step update procedure (dry-run, backup verify, git diff capture, outside-fence analysis (OK/WARN/ERROR), WARN review gate), now productized as the --fleet command that runs it across every workspace under a directory: discovers .github/agents/ and .claude/ targets, snapshots each git workspace via a commit, applies --update --merge in-process, then classifies the git diff by content signals (shrink Notices, USER-EDITABLE deletions). Non-destructive (merge-only); dry-run by default, --yes to apply; report under <DIR>/.agentteams-fleet/<run-id>/. See fleet.

Safety & Backups

  1. Automatic Backup Before Writes — Timestamped backup of all agent files before any --overwrite, --merge, or --update run
  2. Backup Storage — .agentteams-backups/YYYYMMDD-HHMMSS/ (excluded from git)
  3. --no-backup Flag — Suppress backup creation for CI environments
  4. --list-backups Flag — List available backup timestamps
  5. --restore-backup [TIMESTAMP|latest] Flag — Restore a specific backup
  6. emit.backup_output_dir() / emit.list_backups() / emit.restore_backup() — Public backup API

Governance Agents

The following eleven agents are included in every generated team, regardless of project type.

  1. @navigator — Project structure mapping, file location queries, dependency tracking (read-only)
  2. @security — Highest-priority sentinel; clears destructive operations, credential exposure, and external writes (read-only)
  3. @code-hygiene — Architecture auditor; file hygiene, duplication, script lifecycle, agent doc quality — invariant CH-01 through CH-20 plus project extension rules CH-21–CH-24 (Testing, Type Safety, Defensive Programming; CH-24: try/except/finally is a last resort — encode expected conditions in dictionaries/guards and fail hard, reinforcing CH-23 Fail Fast) (read-only)
  4. @adversarial — Presupposition critic; challenges plans and hidden assumptions before execution (read-only)
  5. @conflict-auditor — Detects logical inconsistencies across all output files; logs findings (read + log)
  6. @conflict-resolution — Makes ACCEPT/REJECT/REVISE decisions on flagged conflicts (read + edit)
  7. @cleanup — Removes stale artifacts; requires @security clearance; applies four safety checks before deletion (edit)
  8. @agent-updater — Synchronizes agent documentation when project structure or content changes (edit)
  9. @agent-refactor — Extracts shared data to reference files; enforces spec compliance (edit)
  10. @repo-liaison — Manages cross-repository impact assessment, adjacent-repo updates, and orchestrator coordination (edit)
  11. @git-operations — Handles commit/push, pull/merge/rebase, recovery, and conflict-resolution git workflows (edit)

Workflows

Workflows are step sequences embedded in the generated Orchestrator agent. Every built-in workflow terminates with an unconditional call to Workflow 11 (Final Check). Optional user-added workflow extensions must explicitly include the same terminal handoff.

  1. Workflow 1 — Produce Deliverable — Source review → Component Brief → production → quality audit → style audit → consistency check
  2. Workflow 2 — Revise Deliverable — Revision → adversarial review → quality audit → consistency check → @agent-updater sync
  3. Workflow 3 — Technical Accuracy Audit — Validate claims against source files; conditionally correct and re-audit
  4. Workflow 4 — Compile Final Output — Format conversion → reference verification → final assembly → cleanup
  5. Workflow 5 — Consistency Review — @adversarial challenge → cross-file consistency audit → technical validation → @agent-updater
  6. Workflow 6 — Documentation Maintenance — Sync agent docs → refactor check → consistency audit
  7. Workflow 7 — Cleanup — Identify stale artifacts → @adversarial review → @security clearance → delete → update docs
  8. Workflow 8 — Code Hygiene Audit — CH-01–CH-24 audit (CH-01–CH-20 invariant + CH-21–CH-24 project extensions) → @adversarial guard before deletion plan → conditional cleanup/refactor
  9. Workflow 9 — Cross-Repository Coordination — Impact assessment → approved updates → adjacent-repo writes (security-cleared) → consistency audit
  10. Workflow 10 — Plan Documentation & Review — Plan status scan → pre-execution truth check via @technical-validator → surface blocked steps
  11. Workflow 10B — Work Summary Reporting — Generate daily/weekly/monthly summaries from plan artifacts and git history, then audit them
  12. Workflow 10C — Post-Production Audit Verification (Optional) — User-editable extension workflow for outcome verification after implementation claims; runs @post-production-auditor + adversarial/conflict checks; blocks closeout on FAIL/INCONCLUSIVE
  13. Workflow 11 — Final Check (Part A) — Scan current plan's steps.csv for pending/blocked rows; create audited sub-plans for each
  14. Workflow 11 — Final Check (Part B) — Scan CHANGELOG.md Known Issues, tmp/by-week/YYYY-Www/ plan CSVs (legacy: tmp/ root), and git status for at-large open issues; subject summaries to @adversarial + @conflict-auditor

Governance Infrastructure

  1. Constitutional Rules — Immutable rules encoded in the Orchestrator template: security clearance gates, code-hygiene before merge, conflict audit after multi-file sessions, adversarial review before irreversible changes
  2. Authority Hierarchy — Explicit precedence ordering encoded in every generated Orchestrator (template library → schemas → pipeline source → placeholder conventions → implementation plan)
  3. Automatic @agent-updater Routing in Documentation-Impact Workflows — Routed in Workflows 1, 2, 3, 5, 6, 7, 8, and 9 when knowledge-mutating or coordination changes occur
  4. @adversarial Guards in Audit Workflows — @adversarial runs as step 1 of Workflow 5 and step 2 of Workflow 8 to prevent stale-assumption conclusions
  5. Pre-Execution Truth Check — @technical-validator must verify factual claims in each plan step before it is marked in_progress
  6. Drift-as-Trigger — An explicit trigger in @agent-updater trigger tables: drift detected by --check requires re-render and re-verify before the next workflow
  7. Initialization-as-Trigger — First successful team generation is an explicit lifecycle trigger: it establishes the baseline inventory and trigger corpus used by future update and drift logic
  8. Update Lifecycle Trigger Contract — Canonical --update --merge runs must reconcile drift, emit newly required files, preserve manual values, and preserve user-authored content outside fenced regions
  9. Missing Expected Output Recovery Trigger — During update, absent expected standard outputs are treated as drift and must be restored even if template hashes are unchanged
  10. Cross-Repository Security Rule — Any write outside this project's configured primary output directory must be assessed by @repo-liaison and cleared by @security
  11. User-Editable Gap Safety Pattern — Optional routing rows and optional workflow extensions (such as 10C) must be added outside FENCED sections so --update --merge does not force-propagate them to teams that do not include the required agent
  12. Post-Production Closure Gate Artifacts — Optional post-production audits can emit closure_gate_status.json, capability_check.json, and decision_replay_packet.json to support fail-closed closeout decisions and replay-auditability

Interoperability

  1. convert Module — Direct format migration between framework outputs (copilot-vscode ↔ copilot-cli ↔ claude), and to goose (emits recipes + repo-root AGENTS.md + .goosehints; delegation wired from copilot-vscode sources, flat from claude/copilot-cli)
  2. --convert-from Flag — Convert an existing agent team to a different framework format (targets: copilot-vscode, copilot-cli, claude, goose)
  3. File Classification — Auto-detect file role (agent, instruction, reference) for correct translation
  4. interop Module — Canonical Agent Interface (CAI) normalization and compatibility pipeline (copilot-vscode / copilot-cli / claude targets; goose is refused — the CAI does not carry the handoff graph Goose needs, so use --convert-from instead)
  5. --interop-from Flag — Run the CAI interop pipeline against an existing team
  6. bridge Module — Lightweight runtime compatibility bridge; generate bridge artifacts without regenerating sources (targets include goose — writes fenced AGENTS.md / .goosehints / .goose/README.md pointers)
  7. --bridge-from Flag — Generate bridge artifacts from a source canonical team
  8. Runtime Handoff Manifest — references/runtime-handoffs.json; emitted when handoffs are extracted from non-VS Code adapters, consumed by bridge layers and external tooling

Bridge Automation

  1. run_daily_bridge_maintenance.sh — Daily bridge refresh/check script that halts the run when the security-maintenance preflight fails and continues with warnings only for non-critical bridge refresh/check subtasks
  2. bridge-maintenance.yml CI Workflow — GitHub Actions workflow for daily automated bridge maintenance
  3. bridge-watchdog.yml CI Workflow — Staleness monitoring; creates deduplicated GitHub issues when drift is detected
  4. Bridge Staleness Detection — Compare bridge artifacts against source team; surface divergence
  5. Deduplicated Issue Creation — Watchdog suppresses duplicate GitHub issues for the same staleness event
  6. CI Fallback Mechanism — Bridge refresh/check subtasks continue with warnings, but security-maintenance preflight failures abort the workflow
  7. Snapshot Archive — Pre-update snapshots stored in references/plans/snapshots-*/ for reversible rollback

Cross-Repository Support

  1. @repo-liaison Agent — Manages cross-repository impact assessment, updates, and orchestrator coordination
  2. references/adjacent-repos.md — Living record of adjacent repositories with changelog entries
  3. Protocol 1 — Impact Assessment — Evaluate which adjacent repos are affected by a proposed change; produce an Impact Report
  4. Protocol 2 — Adjacent Updates — Apply approved updates to neighboring project repos (requires @security clearance on each write)
  5. Protocol 3 — Orchestrator Coordination — Formal coordination request between independent orchestrators managing different repositories
  6. @repo-liaison in Constitutional Rules — Any write outside src/ must be assessed by @repo-liaison before @security clearance can be granted
  7. orchestrator-workflows.reference.md — Reference guide documenting the baseline workflow set plus optional extension workflows (for example 10C when configured); kept in sync by @agent-updater

Drift Trust & Delivery Gating

  1. FINGERPRINT_ALGO_VERSION Constant — Module-level integer in agentteams.drift recording the version of the manifest-fingerprint algorithm; recorded into build-log.json on every write and consulted on --check / --update to migrate existing teams when fingerprint semantics change
  2. Algo-Bump Migration Mechanism — compute_structural_diff detects fingerprint_algo_version mismatch between the stored build-log and the current module value, sets manifest_changed, and promotes affected files with _reason = "fingerprint algo version bumped" for a one-shot re-evaluation; legacy logs missing the field do not trigger promotion on their own
  3. Observable Baseline Self-Heal on --update — When an --update run finds no material drift but the recorded fingerprint or algo version is stale, the build-log baseline is refreshed and the run prints ✓ Healed build-log baseline (no material drift; fingerprint refreshed).; heal occurs before delivery-receipt attestation
  4. refine_manifest_promotion Render-Faithful Reconciliation — Pure public function in agentteams.drift that demotes fingerprint-only promotions whose rendered content matches disk byte-for-byte; caller supplies a content_matches(path) closure; preserves manifest_changed as telemetry
  5. --check Option C Render-Faithful Mode — When the structural diff sets manifest_changed with a manifest-promotion reason, --check runs the full render pipeline (render.render_all → framework adapter → handoff/graph append) and applies refine_manifest_promotion against a content-match closure mirroring --update's, eliminating false-positive drift signals
  6. --check / --update --dry-run Parity Contract — --check and --update --dry-run report the same has_changes set for the same inputs; enforced by tests/test_integration.py::test_check_parity_with_update_dry_run
  7. Delivery Receipt Artifact — references/delivery-receipt.json written after every successful --update (non-dry-run) recording artifact_type, receipt_schema_version, delivered_at, project_name, framework, manifest_fingerprint, and fingerprint_algo_version; heal-first-attest-second order ensures the receipt records the post-heal state; write failures warn on stderr but do not fail the run; receipt file is excluded from drift detection
  8. agentteams/schemas/delivery-receipt.schema.json — Draft-07 JSON Schema for the delivery receipt; additionalProperties: false; top-level discriminator is artifact_type (const "delivery-receipt"); receipt_schema_version const "1.0"
  9. docs_src/delivery-procedure.md Operator Guide — Documents the heal/attest order, receipt schema, and operator procedure for verifying a delivered update; registered under Guides in mkdocs.yml

Retrieval & Review-Time Utilities

  1. Confidence-Tiered Retrieval — memory_index.query_index() and code_index.query_partition()/query_partitions() hits carry a confidence field ("reliable" / "candidate" / "weak") computed from score against per-strategy thresholds, printed alongside score= by --query-index/--query-code. Replaces threshold-interpretation prose duplicated across 6 templates (conflict-resolution, conflict-auditor, quality-auditor, research-analyst, retrieval-integrator, tool-doc-researcher). See memory-index, code-index.
  2. agentteams.scan Review-Time Invocability — verdict_for_findings() (HALT / CONDITIONAL_PASS / PASS) and a python -m agentteams.scan <path> entrypoint let a shell-only runtime invoke the deterministic PII/credential scanner at review time, not only at generation time via --scan-security. security.template.md Rules S-1/S-8 cite it as the preferred check. See scan.
  3. session_scan Module — scan_repo_issues() consolidates the orchestrator's Workflow 11 Part B closeout scan (CHANGELOG "Known Issues", plan-steps pending/blocked rows, git status anomalies) into one call, with a python -m agentteams.session_scan entrypoint. See session_scan.

Feature Audit

  1. Machine-Readable Feature Registry — references/feature-registry.csv binds every documented feature to the test that proves it works and a distinct test that fails when it breaks. Derived from this document's own body, so the prose stays the single authored source. See feature-audit.
  2. Per-Tier Proof Enforcement — feature_audit.classify() reports a finding for any requested tier that executed zero proofs, so an empty tier cannot hide behind a populated one. A tier whose rows are all waived or not-provable is exempt.
  3. Live-Tier Reachability Classification — transport failures, anti-bot challenges (HTTP 403/202) and the research CLI's own backend=none exhausted-backends signal classify as UNREACHABLE, which reports but never gates. A missing dependency classifies as harness-broken, not as a failed feature.
  4. End-to-End CLI Probes — tests/test_e2e_probes.py drives the packaged entrypoint against a real tree (generate, --dry-run, --json), covering the gap where in-process tests pass but an installed package does not.
  5. Live-Tier External Probes — tests/test_live_probes.py exercises the agentteams.research egress path (search, scholar) against the real network, asserting positive content rather than absence of error, because the research CLI degrades rather than raising. Env-gated on FEATURE_AUDIT_LIVE so ordinary CI stays deterministic.

Getting Started

Essential Features

Step Feature CLI / Entry Point
1 Create a project brief Write brief.json or brief.md
2 Generate your team agentteams --description brief.json --project /path --framework copilot-vscode
3 Review manual fills Read the generated SETUP-REQUIRED.md
4 Enrich defaults agentteams --enrich
5 Audit output agentteams --post-audit

Advanced Features

Feature When to Use
--migrate Upgrading a pre-fencing team to use AGENTTEAMS:BEGIN/END markers
--update --merge Re-rendering a team after template updates without losing user edits
--convert-from Migrating a team to a different AI framework
--bridge-from Creating lightweight runtime compatibility artifacts without regenerating sources
--check (drift) Detecting whether template or team-composition changes require re-render
--scan-security Auditing generated files for credentials, PII, and unresolved placeholders

Governance Invocations

Need Invoke
Destructive operation (file delete, bulk edit, external write) @security
Code change touching shared utilities or agent docs @code-hygiene
Multi-file session closing @conflict-auditor
Plan with irreversible steps @adversarial
Cross-repository write @repo-liaison → @security