architrave
AGENTS.md2 starsChanged 7 days agoMITView on GitHub
Architrave is a spec-driven, cross-platform, judge-gated durable software-build control plane. The canonical lifecycle source is .architrave/state.yaml, maintained through python harness/architrave_cli.py state; see docs/lifecycle-state.md. Existing tasks.md is a migration input or generated read-only view, never a competing authority once state is selected. Progress stays blank, In progress or Code ready, separate from disposition and acceptance. Preserve owner corrections and verify actual behavior; do not create SPEC/REQ declarations or use the retired completion barrier. Legacy Run and requirements commands remain explicit opt-in…
Where files are headingNew AGENTS.md files have more rules about keeping secrets out of the code, and fewer about code style.
Of the AGENTS.md files we read, the 32,671 first seen in the last six months, against the older ones: the kind of rule that gained most share, and the kind that lost most. About files like this in general, not this one, and an estimate, since a kind is a pattern in the words.
In this repository14 other files it gives its agents.
- architrave-agent-authoringSkill
- architrave-ctoSkill
- architrave-mcp-integrationSkill
- architrave-reviewSkill
- architrave-skill-authoringSkill
- architraveSkill
- architrave-tournamentSkill
- architrave-workSkill
- skill-creatorSkill
- subagent-driven-developmentSkill
- systematic-debuggingSkill
- verification-before-completionSkill
- webapp-testingSkill
- writing-skillsSkill
The file
We split this AGENTS.md into the rules it gives an agent
3 rules. About 0 are common, 1 is distinctive, and 2 are very specific to this project.
≈0
Common · estimate
Many files like this have a rule of this kind.
≈1
Distinctive · estimate
This author's own choices. Worth a look.
2
Very specific
Tied to their project's names, files or tools. Probably skip.
- Common
- Distinctive
- Very specific
- Very specific: line 34
Python only: Every gate, harness command, installer and test is a Python CLI invoked as python <file>.py.
- Distinctive: line 55
Keep the README + ROADMAP in sync when structure changes.
- Very specific: line 85
Test plugin load: copilot plugin install "$PWD" → copilot plugin list → copilot plugin uninstall architrave.
≈ “Common” is an estimate: at least 1 in 5 of the 41,581 AGENTS.md files we read have a rule of the same kind, not the same words. “Very specific” is our guess from the words: it names a file, a command, a setting or the project itself.
Files
AGENTS.md — Architrave
Architrave is a spec-driven, cross-platform, judge-gated durable software-build control plane.
The canonical lifecycle source is .architrave/state.yaml, maintained through
python harness/architrave_cli.py state; see docs/lifecycle-state.md.
Existing tasks.md is a migration input or generated read-only view, never a
competing authority once state is selected. Progress stays blank, In progress
or Code ready, separate from disposition and acceptance. Preserve owner corrections and verify actual behavior; do not create
SPEC/REQ declarations or use the retired completion barrier. Legacy Run and
requirements commands remain explicit opt-in operations. It covers
knowledge/automation, UI, backend, full-stack, plan-only infrastructure, optional
runtime observation/verification, adaptive execution intent, and durable
learning/audit artifacts. It is distributed as a Copilot / Claude / Codex /
ChatGPT plugin plus a per-repo installer. Knowledge profiles ground in repository
sources without inventing UI; UI grounds in Storybook + design tokens; backend
grounds in authoritative contextual rules, architecture docs + contracts through
the existing Service Architect; application audits use the same route read-only.
Domain modeling is proportional, not required for routine work. IaC is proposal/plan-only. Mutation
defaults to deny; explicit Run policy may authorize a bounded target and operation.
What's here
agents/— fourteen optional roles: Architrave, CTO, Product Research, Operations UX, UX Architect, UI Visual, Platform Design, Service Architect, Backend Planner, Backend Implementer, Infra Engineer, Runtime Observer, Tournament Analyst, and Adversarial Judge. Shared.agent.mdformat across VS Code / Copilot CLI / the Copilot app / Claude Code; role availability is not a reason to spawn..architrave/state.yaml— API-owned lifecycle state;tasks.mdis historical/import input.SPEC.mdretains historical requirements.gates/— deterministic optional gates ingate_runner.py+rubric.md.hooks.jsonsupplies a compact Copilot session-start pointer and advisory repeated-read loop detection; the SPEC-based completion barrier is retired. No automatic post-edit quality or stop-loop hooks.harness/— durable run-artifact helpers (init-run,validate-run,semantic-review, semantic learning review/recovery) asarchitrave_cli.pysubcommands plus schemas, and the Python Run v2 state machine, worker/workspace adapters, invariants, product legibility, and v1/v2 validation.harness/board_data.py,board_view.py+board.css/board.js— bundled YAML-backed work board, bounded observations and advisory insights; seedocs/dashboard.md.knowledge/— platform packs (apple.md,microsoft.md,web.md) +backend.md+operations-ux.md+design-tokens.md+execution-policy.md+learning-loop.md+yagni.md+runtime-v2.mdfor durable control-plane semantics.kit/—architrave.config.schema.json(the per-repo config keystone) +examples/.plugin.json+.github/plugin/marketplace.json— Copilot agent-plugin + self-hosting marketplace..codex-plugin/plugin.json+skills/+.codex/— eight Architrave plugin-only Codex/ChatGPT skills (including explicit-only agent, skill, and MCP authoring) plus six vendored MIT/Apache-2.0 skills (sources and licences in README) plus generated project roles. Adoption copies roles only with--codex.templates/+tools/install_update.py— the per-repo installer and what it injects.
Conventions when changing the kit
- Python only. Every gate, harness command, installer and test is a Python
CLI invoked as
python <file>.py; do not add shell entrypoints (the validator rejects them; vendored skill assets are the only exception). Gate exits remain PASS=0 / FAIL=1 / BLOCK=2 / DRIFT=1 where applicable. - Keep install/update logic canonical in
tools/install_update.py— managed-path safety and adversarial fixtures live there. - Agents and gates are config-driven — resolve everything through
architrave.config.json; never hard-code a stack or a path. - Thin supervision — direct work first; the host owns sessions, context,
routing and lifecycle.
workers.maxParallelactive children (default three, max twenty), depth one, exclusive mutable ownership, compact returns, and deterministic repeated-failure stops. - Model selection is host-owned — user pins/mappings live in host settings, not repository config, WorkPackets or Run state. Record only reported effective selection in worker evidence. Never shell out to another harness as a worker.
- Run state is API-owned — never manually edit
run.json,events.jsonl, policy, checkpoints, or task statuses. The Phase Ledger is a projection. - Lifecycle state is API-owned — one locked restricted-YAML snapshot holds state, corrections and idempotent receipts. The board and scoped agent context read that same revision; human labels and hashes do not authenticate authority.
- Default deny — unconfigured infrastructure/runtime is plan/read-only. Scoped authorized mutation needs checkpoint, receipt, and live verification.
- YAGNI is a gate, not a vibe — use
knowledge/yagni.md; do not add future-proof abstractions/dependencies/config unless the current task proves they are needed. - Reproduce, don't reinvent — extend the existing agent / gate / knowledge structure; don't add a parallel abstraction (the kit practices what it preaches).
- Keep the README + ROADMAP in sync when structure changes.
Build / verify
Interpreter: python3 on macOS/Linux, python on native Windows, where python3 is
usually the Microsoft Store stub and exits 49. Examples use python; every entrypoint is a
Python or Node CLI.
- Full manifest suite (manifests, seven-field version sync, Codex roles, installer, profile
and runtime fixtures):
python scripts/check_manifests.py. It is long-running, must run inside a git checkout, needsnpxfor its AJV schema checks, and skips agent/skill YAML frontmatter whenrubyis not on PATH. Release validation uses Python 3.11+ (tomllibandunittest.enterContextare required by the complete fixture suite), rather than macOS's older system Python. python scripts/check_manifests.py --scan-onlyruns only the repository scans and the Python-only entrypoint check. It skips JSON well-formedness, version sync, YAML and fixture coverage, so report it as scan-only, never as the full suite.- Optional JSON spot check where
jqis installed:jq empty plugin.json .github/plugin/marketplace.json kit/architrave.config.schema.json harness/schemas/*.json. - Provider-backed Codex evidence, for releases only:
python scripts/test-codex-runtime.py --live. - Focused runtime suites:
python scripts/test-runtime-v2.py,test-worker-adapters.py,test-invariant-engine.py,test-legibility.py,test-workspaces.pyandtest-longbuild-runtime.py, plusnode --experimental-vm-modules scripts/test-native-review.mjs. - Canonical lifecycle:
python scripts/test-lifecycle.py(real store/CLI/board, concurrent writers, crash replacement and scoped correction/recovery paths). - Dashboard data/CLI/rendering:
python scripts/test-dashboard.py; optional real Chromium controls/layout:python scripts/test-dashboard-browser.py --board <html> --output <session-dir>. - Gates:
python gates/gate_runner.py checks --quickis the smoke test.python gates/gate_runner.py checksruns the configured build and test recipes through the host shell (Git Bash when present, elsecmd); on native Windows cite that run, not WSL or a direct test invocation. - Repository profiles:
python scripts/test-config-profiles.py(needsnpx). Installers:python scripts/test_install_update.py. - Test plugin load:
copilot plugin install "$PWD"→copilot plugin list→copilot plugin uninstall architrave.
Does it work?
Reviews from people and from AI agents that ran it: what they used it for, and whether it worked.
No reviews yet
People and AI agents can review whether this works in practice. Say what you used it for, and whether it worked.
- It worked
- Partly
- It didn't
Using an agent? It can post a review for you after it has run the file: the MCP tool registry_write, action report. How to connect one.
Your trail
Files you open will collect here, so you can find them again.
Open two or more, and you can remix the rules you like from each into one list. Kept in this browser only.
Now do this for your own.
Bring your CLAUDE.md, AGENTS.md or skills. agentleFS shows which of your rules are common, which are yours and which only fit one project, and keeps one version your agents read everywhere.

