Define Once.
Run Across Every AI Coding Agent.
Write canonical agent roles in pure, portable Markdown. Agent Fabric compiles them deterministically into native configurations for OpenCode, Kilo, Antigravity CLI, Codex, and Claude Code — with atomic SHA-256 state tracking and zero vendor lock-in.
One Canonical Definition. Five Native Targets.
Never duplicate prompt instructions, role boundaries, or tool permissions across different AI harnesses. Agent Fabric parses canonical schema v1 frontmatter and compiles it cleanly into native configs.
---
description: Authors grounded, independently reviewed implementation plans as executable vertical slices
mode: primary
hooks: [load-task, pre-plan, post-plan, decompose]
x-agent-fabric:
schema: 1
profile: planner
effort: high
visibility: public
isolation: sandbox
permissions:
edit: allow
bash: allow
task: allow
---
# Planner
You author and refine grounded implementation plans. You do not execute plan
phases, deploy, or treat a proposal as approval.
## Operating Modes
- Author: create a new canonical plan from a goal or task seed.
- Refine: update an eligible draft plan in place.
- Approval/projection: invoke decompose hook with validated slices.8 Purpose-Built Autonomous Roles
Agent Fabric bundles eight rigorously specified roles with explicit evidence contracts, failure boundaries, and vertical slicing principles.
planner
Planner Profile • High EffortAuthors and refines grounded, multi-phase implementation plans as executable vertical slices. Enforces strict definition of done and test-first verification.
plan-reviewer
Reviewer Profile • High EffortCritiques proposed plans for grounded evidence, completeness, dependency ordering, failure modes, and vertical slicing feasibility before execution begins.
plan-supervisor
Supervisor Profile • High EffortOrchestrates the iterative discovery, interactive grilling, plan refinement loop, and final human approval gate before projection into task systems.
implementor
Worker Profile • High EffortExecutes single vertical slices test-first with minimal, targeted changes. Strictly adheres to contracts, coding standards, and definition of done.
expert-debugger
Worker Profile • High ReasoningSystematically isolates hard bugs, test failures, and race conditions using hypothesis validation and reproducible test cases with minimal diff footprint.
qa-runner
Reviewer Profile • High EffortExecutes test suites, builds reproduction cases for edge scenarios, verifies end-to-end behavior, and guards against regressions.
code-reviewer
Reviewer Profile • High EffortConducts dual-axis reviews: Standards Review (repository conventions, security, linting) and Spec Adherence Review (does code match original requirements).
loop-supervisor
Supervisor Profile • High EffortDrives autonomous implementation loops, delegates work slice-by-slice across workers and reviewers, tracks progress, and prevents infinite loops.
Architectural Pillars
Built from first principles in Go to provide rock-solid predictability and prevent the common failure modes of ad-hoc agent setups.
Atomic Manifest Ownership
Every generated agent file is fingerprinted with a SHA-256 checksum in .agent-fabric-manifest.json. When running agf sync, your local customizations are automatically detected and preserved — never silently overwritten without --force.
~/.config/agent-fabric/.agent-fabric-manifest.json
6 Deterministic Lifecycle Hooks
Canonical agents declare portable lifecycle hook placeholders (load-task, pre-plan, classify, label, decompose, post-plan). Fabric resolves hooks once at install/sync time from host-global ~/.agent-hooks/.
Abstract Profile Tiers
Decouple abstract compute requirements (planner, worker, reviewer, supervisor) from specific provider model IDs. Configure custom overrides in one clean JSON config without touching canonical agent bodies.
Hub Agent Ecosystem
Extend Fabric with external curated agent catalogs via agf hub install. Validates dependency graphs, prevents naming collisions, and enforces strict schema compliance before writing any file to disk.
The Agentic Execution Pipeline
How canonical agent roles collaborate from initial problem intake through verified implementation.
Task Intake & Hook Resolution
The load-task and classify hooks hydrate ticket context, repository state, and constraints into canonical format.
Grounded Plan Authoring
planner drafts vertical implementation slices; plan-reviewer stress-tests assumptions and boundary conditions.
Supervisor & Human Gate
plan-supervisor coordinates plan grilling and pauses for explicit human operator approval before vertical slice decomposition.
Vertical Slice TDD Execution
implementor executes slice-by-slice, expert-debugger isolates blockers, and qa-runner runs test suites.
Dual-Axis Review & Sync
code-reviewer validates Standards & Specs adherence. agf sync updates atomic manifest state safely.
Interactive CLI Explorer
Explore the fast, intuitive Agent Fabric CLI commands. Click any command to preview its execution.
[INFO] Agent Fabric v1.0.0 (linux/amd64)
[INFO] Resolving canonical source definitions from bundled assets...
[INFO] Found 8 canonical agents: planner, plan-reviewer, plan-supervisor, implementor, expert-debugger, qa-runner, code-reviewer, loop-supervisor
[INFO] Resolving target adapters: opencode, kilo, antigravity, codex, claude
[INFO] Resolving hooks from ~/.agent-hooks/ (4 installed, 2 skipped)
[OK] Rendered 8 agents -> ~/.opencode/agents/
[OK] Rendered 8 agents -> ~/.kilo/agents/
[OK] Rendered 8 agents -> ~/.gemini/config/agents/
[OK] Rendered 8 agents -> ~/.codex/agents/
[OK] Rendered 8 agents -> ~/.claude/agents/
[OK] Manifest recorded at ~/.config/agent-fabric/.agent-fabric-manifest.json
[SUCCESS] Installed 40 target configurations successfully. Run 'agf doctor' to verify.Ready to Standardize Your Agent Fleet?
One command is all it takes to bootstrap Agent Fabric on Linux or macOS.
1. Bootstrap Binary
Installs agent-fabric and the agf alias into ~/.local/bin with automated SHA-256 verification.
curl -fsSL https://ericmaster.ninja/agent-fabric/install | bash
2. Install Target Agents
Install globally or scope agents to a specific Git repository with --project.
agf install --all --tools opencode,kilo,agy --yes
3. Verify Integrity
Doctor inspects generated files, SHA-256 hashes, file permissions, and mapping health.
agf doctor && agf validate
Optional: Target Profile Overrides
~/.config/agent-fabric/config.json{
"profiles": {
"opencode": {
"worker": {
"model": "openai/gpt-5.6",
"effort": "high",
"permissions": { "network": "deny" }
}
},
"antigravity": {
"planner": {
"model": "gemini-3.1-pro",
"effort": "high"
}
}
}
}Everything You Need To Know
During install and sync, Fabric records the SHA-256 hash of each generated file in .agent-fabric-manifest.json. When you run agf sync, if a file has been modified by you, Fabric detects the checksum delta and leaves your file untouched while reporting its path. You must explicitly pass --force if you wish to overwrite custom edits.
Modern AI harnesses (OpenCode, Kilo, Antigravity CLI, Codex, Claude Code) each use different file formats (Markdown vs TOML), differing frontmatter schemas, distinct tool permission syntaxes, and provider-specific model strings. Managing them manually across repositories leads to prompt drift and subtle behavioral bugs. Agent Fabric provides a single source of truth.
No. Agent Fabric bundles its canonical source definitions directly beside the compiled executable. Commands like agf install, agf sync, agf validate, and agf doctor execute completely offline with zero telemetry.
Yes. You can use agf hub install <url-or-path> to install supplemental agent catalogs (such as simplification-planner) from GitHub repos, local paths, or HTTPS tarballs with automatic dependency graph validation.