🧭 Agent Compass

One operating system for every coding agent

CI Version Node License Agents

Skills

Ops Packs Specs Memory

Start in 30 seconds · Explore capabilities · Browse skills · Read the contract

Importable operating manual for Claude Code, Codex, Gemini, and GitHub Copilot: one project contract, quality gates, workflows, templates, operational knowledge, and helper scripts.

Use it as a submodule in real projects, or clone it standalone to bootstrap a new project from proven defaults.

Agent Compass exists for three jobs:

  1. Teach agents how to work: context first, spec when needed, plan, implement, validate, report.
  2. Bootstrap projects: stack presets, templates, specs, memory, PR workflow, and tool pointers.
  3. Grow safely: pull reusable knowledge out of real projects without leaking secrets, personal data, or project-specific facts.
  4. Operate systems: opt-in DevOps, security, infrastructure, and compliance packs with authorization and production-safety gates.

Fastest path: open any agent CLI (Claude Code, Codex, Gemini, Copilot, …) in this repo and say what you want — "set up ~/projects/foo", "bootstrap a new API from these guidelines", "add a skill for X". The agent routes itself through MISSIONS.md and executes the matching playbook.

Table Of Contents

Quick Start

30-Second Route

git submodule add https://github.com/<owner>/agent-compass.git docs/agent-compass
node docs/agent-compass/scripts/adopt.mjs .

Then tell your agent what you want. MISSIONS.md routes setup, bootstrap, and Compass extension work to the right playbook.

Add To Existing Project

From the host project root:

git submodule add https://github.com/<owner>/agent-compass.git docs/agent-compass
node docs/agent-compass/scripts/adopt.mjs .

adopt runs the whole chain non-interactively: detection, setup, fit-based skill sync (core + detected stacks only), readiness verification, and next steps. Granular equivalent when you want control:

node docs/agent-compass/scripts/setup-wizard.mjs . --yes
node docs/agent-compass/scripts/setup-host.mjs . --strict
node docs/agent-compass/scripts/apply-recommendations.mjs . --policy solo-dev

The wizard writes agent-compass.answers.json and .agent/setup-plan.md, then can run the full setup. setup-host installs missing pointers/templates, applies safe fixes, runs deep doctor checks, and writes .agent/context.json, .agent/doctor-report.md, and .agent/RUNBOOK.md. apply-recommendations applies safe missing pieces: provider verification, MCP readiness, spec validation map, quality gates, migration plan, dashboard, and optional skills/policy setup.

Manual equivalent:

node docs/agent-compass/scripts/install.mjs
node docs/agent-compass/scripts/install.mjs --doctor
node docs/agent-compass/scripts/install.mjs --doctor --fix
node docs/agent-compass/scripts/install.mjs --doctor --deep

This creates missing agent pointers, hooks, command registry, specs/memory starters, MCP examples, PR templates, and Copilot instructions. It never overwrites existing files.

Then run:

node docs/agent-compass/scripts/context.mjs .
node docs/agent-compass/scripts/doctor-report.mjs . --write
node docs/agent-compass/scripts/runbook.mjs . --write

Result:

  • AGENTS.md tells agents where the shared contract lives.
  • agent-compass.commands.json tells agents which commands exist.
  • .agent/doctor-report.md shows host readiness gaps.
  • .agent/RUNBOOK.md gives agents a compact startup path.

Use docs/workflows/upgrading.md when bumping the submodule later.

Adopt Without Submodule

Use a standalone clone when the host should copy standards but not vendor this repo:

git clone https://github.com/<owner>/agent-compass.git /tmp/agent-compass
node /tmp/agent-compass/scripts/install.mjs --dry /path/to/host
node /tmp/agent-compass/scripts/install.mjs /path/to/host
node /tmp/agent-compass/scripts/install.mjs --doctor --fix /path/to/host
node /tmp/agent-compass/scripts/install.mjs --doctor --deep /path/to/host

Review and commit only host-local files. Copy .mcp/*.example.json into local MCP client config, replace /absolute/path/to/repo, and never commit that local client config.

Bootstrap New Project

git clone https://github.com/<owner>/agent-compass.git
cd agent-compass
node scripts/bootstrap.mjs

The bootstrap writes:

  • BOOTSTRAP_PROMPT.md: a copy-paste build prompt with stack choices, spec-first setup, TDD, docs, validation, and quality gates.
  • agent-compass.answers.json: saved answers for repeatable setup.

Paste BOOTSTRAP_PROMPT.md into your agent. It starts with specs/000-project/ artifacts before scaffolding code.

Agents (and CI) skip the questionnaire with an answers file:

node scripts/bootstrap.mjs --schema                       # answers contract
node scripts/bootstrap.mjs --answers answers.json --out /path/to/new-project

Or let the agent do the whole mission — see skills/compass-bootstrap/SKILL.md.

Browse And Copy

Agent Compass is mostly Markdown and templates:

Capability Packs

Agent Compass keeps broad operational knowledge opt-in. Normal project and global adoption stay small; choose packs only when the host needs them.

Pack Skills Covers
devops-platform 22 CI/CD, containers, Kubernetes, observability, AI pipelines, developer environments, and release operations.
security 35 Defensive scanning, secrets, hardening, network security, incident response, and AI security.
infrastructure 70 Cloud, IaC, servers, networking, databases, storage, platforms, IT, and AI infrastructure operations.
compliance 19 Framework mapping, governance, evidence, continuity, auditing, and incident management.

Focused subpacks

Use smaller packs when a host only needs one cloud, Kubernetes, observability, AI operations, scanning, secrets, hardening, or compliance area.

Pack Skills Covers
aws 12 AWS compute, containers, IAM, networking, data, secrets, auditing, cost, CloudFormation, and Terraform.
azure 9 Azure compute, AKS, networking, SQL, functions, Key Vault, audit monitoring, ARM/Bicep, and Terraform.
gcp 8 GCP compute, GKE, networking, Cloud SQL, functions, secrets, audit logs, and Terraform.
kubernetes 9 Kubernetes operations, packaging, GitOps, managed clusters, GPU workloads, scaling, and hardening.
observability 8 Metrics, traces, logs, alerts, audit telemetry, and cloud audit trails.
ai-ops 14 AI pipelines, model serving, GPU operations, gateways, caching, cost, RAG, vector stores, and inference scaling.
security-scanning 7 Dependency, source, dynamic, container, vulnerability, SBOM, and supply-chain scanning.
secrets 5 Vault, cloud secret managers, and encrypted GitOps secrets.
hardening 6 CIS, Linux, Windows, container, Kubernetes, and agent deployment hardening.
compliance-frameworks 6 FedRAMP, GDPR, HIPAA, ISO 27001, PCI DSS, and SOC 2 framework guidance. 
# Discover packs and skills through the unified CLI
node docs/agent-compass/scripts/cli.mjs skills-sync --list-packs
node docs/agent-compass/scripts/cli.mjs skills --pack devops-platform
node docs/agent-compass/scripts/cli.mjs skills --grep kubernetes --md
node docs/agent-compass/scripts/cli.mjs skills kubernetes-ops

# Install one pack's skills from the pin, safety gate applied
node docs/agent-compass/scripts/cli.mjs external-skills . \
  --source devops-security --skill kubernetes-ops,helm-charts,argocd-gitops

# The whole curated operational baseline
node docs/agent-compass/scripts/cli.mjs external-skills . \
  --source devops-security --recommended

DISCOVER          SELECT             INSTALL              ACTIVATE
host context  →   fitting pack   →   fetch + harden   →   agent loads one skill

All 146 curated operational skills are pinned to an audited upstream commit and carry authorization/dry-run/rollback rules. Agent Compass stores none of them: external-skills fetches them from the pin and applies the safety gate and the argv-secret narrowings on the way in, so the corrected text is what lands in your project. Executable payloads are refused unless you allow them. Full list: skills/README.md. Safety contract: operational-safety.md.

Every external source is tracked, never copied

All nine external skill sources — the operational corpus, Google's Android skills, an Apple-platform corpus, and the style/design/document sources — are pinned in one registry with "strategy": "reference". Agent Compass records the pin, the licence, the full upstream inventory, and which skills it endorses. It copies nothing.

agent-compass external-skills --list                 # every tracked source
agent-compass upstream-skills --verify               # offline: pins, pointers, inventories
agent-compass check-update --remote                  # cached remote-head check
agent-compass upstream-skills --update caveman --dry

Installing is the same command whether a skill is local or tracked, for a project or for the user, across Claude Code, Codex, and Copilot:

# Fit-based adoption resolves both kinds from one list
agent-compass recommend /path/to/host --json          # → assets.skills
agent-compass skills-sync /path/to/host --only <that list>

# Or name a source directly
agent-compass external-skills /path/to/host --source devops-security --recommended
agent-compass external-skills --source ponytail --recommended --global

skills-sync writes .claude/skills/ for Claude Code and .agents/skills/ for Codex and Copilot, adds a .github/instructions/ file so Copilot — which has no skills directory — is told the tree exists, and drops the source's licence notice beside the install. The cached update check is read-only, refresh is explicit, and remote content is never executed. See upstream-sources.md, operational-skills.md, native-mobile-skills.md, style-and-design-skills.md, and ADR 002.

Status

Agent Compass is usable across real projects. Current version: 0.9.1.

Area Current state
Agent contract Tool-agnostic AGENTS.md with Claude, Codex, Gemini, and Copilot pointers.
Installer Non-destructive host setup with --doctor and --doctor --deep.
Bootstrap Interactive prompt generator for new projects and stack presets.
Specs Native specs/ templates and spec workflow; optional upstream Spec Kit path documented.
Memory projectmem workflow, templates, policy, and MCP examples.
Code intelligence Optional codebase-memory-mcp layer: code-intel CLI, MCP template, doctor checks, and navigation rules.
PR workflow GitHub PR create/review helpers and reviewer/label/base rules.
Figma MCP guidance, frontend skill, and design-system extraction worksheet.
Documents Pinned local anydoc extraction for office files and text PDFs, with prompt-injection and hosted-upload guardrails.
Operations 146 opt-in DevOps, security, infrastructure, and compliance skills with safety gates.
CI Node 24, latest tracked action majors, tests, naming/index/docs/action guards.
Release Changelog/version/tag guard and a release helper that tags, pushes, and publishes the forge release on every remote.

Safety Model

  • Installer never overwrites existing host files.
  • Agents must not commit, push, deploy, publish, or open PRs unless explicitly asked.
  • Commands must come from agent-compass.commands.json, package.json, or documented equivalents.
  • Knowledge capture refuses likely secrets, personal data, and known project/domain tokens.
  • Project memory must not store secrets, credentials, tokens, personal data, or temporary brainstorming.
  • Optional tools stay optional: a plain clone never installs a binary, and machine-level installs need an interactive confirm or --yes.
  • A code-graph answer never becomes an exhaustive or negative claim without source corroboration.
  • Specs hold product intent; implementation decisions belong in plans.
  • PR helpers use existing labels only and require at least one reviewer.
  • CI guards reject stale action majors and broken local docs/template links.

Feature Map

Need Start here Helper
Let an agent run a full mission MISSIONS.md skills/compass-*
Browse every asset programmatically MISSIONS.md scripts/catalog.mjs
Install into host repo docs/workflows/upgrading.md scripts/install.mjs
Start a new project docs/workflows/new-project.md scripts/bootstrap.mjs
Get repo context fast docs/architecture/repo-map.md scripts/context.mjs
Check host readiness docs/tooling/prerequisites.md scripts/doctor-report.mjs
Full host setup docs/workflows/upgrading.md scripts/setup-host.mjs
Host recommendations docs/workflows/upgrading.md scripts/recommend.mjs
Provider verification docs/tooling/agent-provider-capabilities.md scripts/provider-verify.mjs
Agent dashboard docs/workflows/upgrading.md scripts/dashboard.mjs
Global user setup docs/agent-setup.md scripts/global-setup.mjs
MCP readiness docs/tooling/mcp.md scripts/mcp-probe.mjs
Navigate code structurally docs/tooling/codebase-memory.md scripts/code-intel.mjs
Policy packs docs/workflows/upgrading.md scripts/policy-pack.mjs
Generate agent runbook AGENTS.md scripts/runbook.mjs
Create specs docs/workflows/spec-driven-development.md templates/specs/
Use durable memory docs/workflows/project-memory.md templates/memory/
Wire MCP tools docs/tooling/mcp.md templates/mcp/
Build from Figma skills/figma-mcp-frontend/SKILL.md templates/design-system/
Create a PR docs/workflows/pull-requests.md scripts/pr.mjs
Review a PR docs/workflows/pr-review.md scripts/pr-review.mjs
Capture reusable knowledge docs/workflows/knowledge-capture.md scripts/pull-knowledge.mjs
Release Agent Compass docs/workflows/releasing.md scripts/release.mjs

What You Get

Area What it provides
Agent contract One canonical AGENTS.md for all tools.
Spec workflow Native specs/<id-slug>/ flow for idea → spec → plan → tasks → implementation.
Project memory projectmem guidance and templates for durable factual memory.
Command registry agent-compass.commands.json so agents stop inventing commands.
Repo context context.mjs, repo map template, and generated runbook/report helpers.
PR workflow PR creation/review docs, templates, and gh helper scripts.
MCP setup MCP docs and examples for projectmem, codebase-memory, and Figma.
Code intelligence Optional structural code graph so agents query before they grep.
Figma frontend flow Design-system extraction template and Figma MCP skill.
Quality gates TDD, lint/typecheck/test reporting, docs sync, security discipline.
Knowledge capture Safe pull/promote workflow for reusable project lessons.

Repository Layout

agent-compass/
├── AGENTS.md                       # canonical agent contract
├── CLAUDE.md                       # Claude pointer + notes
├── CODEX.md                        # Codex pointer + notes
├── GEMINI.md                       # Gemini pointer + notes
├── MISSIONS.md                     # mission router → compass-* playbooks
├── agent-compass.commands.json     # command registry
├── .github/                        # CI workflows (tests, drift check)
├── docs/
│   ├── architecture/
│   ├── decisions/
│   ├── guidelines/
│   ├── tooling/
│   └── workflows/
├── knowledge/                      # instincts, worked examples, incoming
├── migrations/                     # versioned sync migrations for hosts
├── scripts/
├── skills/
├── stacks/
├── templates/
└── test/

Important entry points:

Path Purpose
AGENTS.md Canonical agent contract. Read first.
docs/guidelines/ Coding, testing, security, docs, workflow rules.
docs/architecture/ Generic architecture principles and repo-map guidance.
docs/tooling/ Tool setup: prerequisites, CLI, MCP + recommended servers, headroom, projectmem, pnpm, turbo, Docker, security, prose and terminology linting.
docs/workflows/ Playbooks for project creation, specs, memory, PRs, release, upgrades.
skills/ Portable agent skills for work style, review, security, stacks, Figma.
templates/ Copyable config, MCP examples, PR template, specs, memory, design system, docs-lint.
scripts/ Bootstrap, install, doctor, PR, review, release, upgrade helpers.

Core Contract

The contract in AGENTS.md is tool-agnostic. Host project guidance wins when it conflicts:

host project AGENTS.md > agent-compass AGENTS.md > model defaults

Baseline behavior:

  • Gather context before code.
  • Clarify ambiguity before planning.
  • Use specs for new projects, new features, ambiguous behavior, and high-risk work.
  • Plan before implementation.
  • Reuse existing code, then standard library, then existing dependencies.
  • Validate with real commands from agent-compass.commands.json or package.json.
  • Use provider-native capabilities when they reduce risk: skills/prompts, MCP, hooks, subagents/custom agents, goal/plan/review modes.
  • Teach selectively when users ask for explanations or repeat a costly prompting/tool-use pattern.
  • Report with the Completion Gate.
  • Never commit, push, deploy, publish, or open PRs unless explicitly asked.

Completion Gate:

Goal:
Mode:            implementation | review-only | docs-only | partial
Files changed:
Commands run:
Validation:      one line per command - passed | failed | partial | not run + reason
Risks:
Next step:

Agent Support

Tool Entry file Notes
Claude CLAUDE.md → AGENTS.md Skills can be referenced directly or synced globally.
Codex CODEX.md → AGENTS.md Uses same contract and repo context helpers.
Gemini GEMINI.md → AGENTS.md MCP servers configure via .gemini/settings.json; the installer ships a .gemini/settings.example.json starter.
Copilot .github/copilot-instructions.md → AGENTS.md Extra .github/instructions/*.instructions.md templates included.

Other tools: point a rule file at AGENTS.md.

Workflows

Workflow Use when
new-project.md Starting from zero.
architecture-decision.md Choosing & justifying a new project's architecture (research-first, tech-neutral).
new-module.md Adding a feature/module to existing code.
spec-driven-development.md Turning ideas into specs, plans, tasks, and synced docs.
project-memory.md Reading/writing durable project memory during work.
pull-requests.md Creating GitHub PRs with sane defaults.
pr-review.md Local/GitHub PR reviews and implementing submitted review fixes.
agent-teaching.md Teaching users without coaching every prompt.
agent-improvement-loop.md Turning repeated friction into reusable agent artifacts.
agent-value-expansion.md Further ways to automate and guide developer work.
review-and-ship.md Final self-review, validation, handoff, PR.
knowledge-capture.md Pulling reusable lessons into Agent Compass.
releasing.md Releasing this repo.
upgrading.md Bumping a host project submodule.

CLI

All scripts below are reachable through one entrypoint, agent-compass (scripts/cli.mjs), which dispatches agent-compass <command> [...args] and passes flags through. From a host that vendors the submodule:

alias ac="node docs/agent-compass/scripts/cli.mjs"
ac install && ac doctor . --deep && ac onboard .
ac help            # grouped command list

Full install paths and command reference: docs/tooling/cli.md.

Scripts

All scripts are dependency-free Node scripts (also reachable via the CLI above).

Command Purpose
npm run cli Unified dispatcher (agent-compass <command>).
npm run bootstrap New-project prompt generator (interactive, or --answers for agents).
npm run catalog Machine-readable asset catalog (skills, stacks, templates, docs, commands).
npm run context Print compact repo snapshot for agents.
npm run agent-conformance Check provider customization artifacts and print smoke prompts.
npm run agent-evals Validate teaching/tool-offer eval fixtures.
npm run agent-drift Read-only drift dashboard across all guidance validators.
npm run agent-trace Validate an agent trace/outcome log (no secrets/PII).
npm run agent-onboard One-command readiness check (doctor + drift + sync) for a host.
npm run context-pack Generate .agent/context.json machine-readable repo index.
npm run run-command Run a command from the registry (refuses unknown/destructive).
npm run new Scaffold a skill, ADR, or spec stub.
npm run redact Scan files/staged diff for secret/PII leaks.
npm run trace-to-evals Turn failed trace rows into regression eval scenarios.
npm run gen-depgraph Generate a Mermaid dependency graph from local imports.
npm run check-companions Fail when source changes ship without a test.
npm run doctor-report Print host readiness report.
npm run runbook Print compact agent runbook.
npm run install-into Install pointers/templates into a host.
npm run sync-into Update a host's managed files from the submodule (safe, no clobber).
npm run check-update Cheap cached check whether a host is behind (zero LLM tokens).
npm run pull-knowledge Stage reusable knowledge from another project.
npm run pr Create PR through gh with Agent Compass defaults.
npm run pr-review Build local PR review packet or submit review.
npm run release Prepare version/changelog release metadata.
npm run upgrade-host Update host submodule and run deep doctor.
npm run check Run tests + conformance + evals + naming/index/docs guards.
npm run lint:actions Enforce supported GitHub Action majors.

Examples:

node scripts/doctor-report.mjs /path/to/host --write
node scripts/runbook.mjs /path/to/host --write
node scripts/pr.mjs --reviewer alice --label enhancement --dry
node scripts/pr-review.mjs 123 --out .agent/pr-123-review.md
node scripts/release.mjs 0.7.0 --dry
node scripts/upgrade-host.mjs /path/to/host docs/agent-compass --dry

Templates

Templates are copyable starters, not generated framework magic.

Group Includes
agent/ PR template, Copilot instructions, prompt files, custom agents, agent-ready issue form.
commands/ agent-compass.commands.json starter.
intake/ Agent-ready work intake (Goal/Context/Constraints/Done/Validation).
architecture/ New-project architecture decision scaffolds (intake, decision, matrix, diagrams, backlog, meetings).
context/ Repo map + task routing template.
conformance/ Lightweight agent smoke test.
design-system/ Figma/design-token extraction worksheet.
memory/ projectmem README and policy.
mcp/ MCP examples + per-tool contract for projectmem and Figma.
trace/ Agent trace/outcome log schema.
specs/ Constitution, spec, plan, tasks, checklist templates.
monorepo/ pnpm/turbo/tsconfig/prettier/husky/env starters.
ci/ GitHub Actions examples.
docker/ Dockerfiles and local compose.
eslint/ Stack-specific ESLint configs.
security/ OSV scanner config.
sonar/ SonarQube project configs.

Full index: templates/README.md.

Skills

Skills are portable SKILL.md folders. Agents can load them directly or use them as referenced context.

Useful groups:

  • Working style: caveman, ponytail, caveman-review, ponytail-review.
  • Teaching: agent-teacher.
  • Architecture: architecture-advisor (new-project architecture decisions).
  • Quality: gen-docs, verify-module, verify-quality, verify-change, verify-security.
  • Workflow: spec-workflow, project-memory, pr-workflow.
  • Frontend/design: figma-mcp-frontend, react-admin-dashboard-patterns, expo-react-native-patterns.
  • Backend: nestjs-patterns, drizzle-postgres-patterns, bullmq-patterns, resilience-observability-patterns, external-service-patterns.
  • Operations: devops-platform, security, infrastructure, and compliance packs (146 skills; opt-in through skills-sync --pack).

Full index: skills/README.md.

MCP, Memory, And Figma

Agent Compass supports MCP as optional tooling, not a hard dependency.

Start here:

projectmem flow:

Before work: read summaries and pre-action warnings.
During work: log failed attempts and important findings.
After work: log decisions, fixes, files changed, validation, remaining risks.
Never log secrets, credentials, tokens, personal data, or temporary brainstorming.

Figma flow:

Verify Figma MCP tools → pull selected frame/file context → extract tokens,
components, variants, states, layout rules → map to existing code components →
implement → validate visually.

Use skills/figma-mcp-frontend/SKILL.md and templates/design-system/README.md for design-system work.

PR Workflow

Agent Compass assumes GitHub PRs even if product issues live elsewhere, such as Jira.

Defaults:

  • base branch: develop
  • assign PR to self: --assignee @me
  • require at least one reviewer
  • use only labels that exist in the repo
  • include what changed, why, validation, risks, and reviewer notes

Create PR:

node docs/agent-compass/scripts/pr.mjs --reviewer <login> --label <existing-label>

Local review packet:

node docs/agent-compass/scripts/pr-review.mjs 123

Submit prepared review:

node docs/agent-compass/scripts/pr-review.mjs 123 --submit request-changes --body /tmp/review.md

Read:

Prompt Examples

These assume Agent Compass is installed at docs/agent-compass/.

Bootstrap A Service

Read docs/agent-compass/AGENTS.md and docs/agent-compass/stacks/nestjs-api.md. Scaffold a NestJS + Drizzle + BullMQ API in a pnpm/turbo monorepo. Use TDD. Create specs/000-project/ first. Stop after the plan.

Add A Feature

Following docs/agent-compass/AGENTS.md, add POST /invoices. Update or create the feature spec first. Write the test first. Keep OpenAPI/Scalar, Bruno, and Gherkin in sync. Report against the Completion Gate.

Build From Figma

Use Figma MCP and docs/agent-compass/skills/figma-mcp-frontend/SKILL.md. Pull the selected checkout frame, extract tokens/components/states, map them to existing UI components, implement the screen, and validate visually.

Review A PR Locally

Use docs/agent-compass/docs/workflows/pr-review.md for local PR review #123. Findings first. No GitHub review submission unless I ask.

Implement Submitted Review Fixes

Implement review fixes for PR #123. Verify every review item against current code first, skip outdated comments with reason, patch relevant issues, run validation, and report.

Capture Knowledge Back

Run node docs/agent-compass/scripts/pull-knowledge.mjs ../other-project. Review knowledge/incoming/, promote only generic patterns, and redact any project-specific facts.

Validation And Release

Repo validation:

npm run check
for f in scripts/*.mjs; do node --check "$f"; done
git diff --check

Release prep:

node scripts/release.mjs 0.7.0 --dry
node scripts/release.mjs 0.7.0
npm run check
git add package.json CHANGELOG.md
git commit -m "chore: release v0.7.0"
git tag -a v0.7.0 -m "v0.7.0"
npm run lint:release

Full guide: docs/workflows/releasing.md.

Using In Host Projects

Recommended host files after install:

AGENTS.md
agent-compass.commands.json
docs/architecture/repo-map.md
docs/decisions/000-template.md
specs/README.md
specs/constitution.md
.projectmem/README.md
.projectmem/projectmem-policy.md
.mcp/README.md
.github/PULL_REQUEST_TEMPLATE.md
.github/instructions/*.instructions.md

Minimal host startup prompt:

Read AGENTS.md, agent-compass.commands.json, relevant specs, project memory summaries, and docs/architecture/repo-map.md. Then plan the requested change and list validation commands before editing.

Maintaining Agent Compass

Before changing this repo:

  1. Read CONTRIBUTING.md.
  2. Put new rules in the narrowest useful place: guideline, workflow, tooling doc, skill, template, or script.
  3. Update indexes: docs/workflows/README.md, docs/tooling/README.md, skills/README.md, or templates/README.md.
  4. Add or update focused tests.
  5. Run validation.

Knowledge capture:

node scripts/pull-knowledge.mjs /path/to/project

The pull step refuses likely secrets, personal data, and known project/domain tokens before staging. Promote only generic material. See docs/workflows/knowledge-capture.md.

Provenance

v0.1 was distilled from a production pnpm/turbo monorepo and mature global agen