Agnostic AI

One rules file for every AI coding tool you use, plus a terminal coding agent that runs against local or hosted models under the same safety policy.

• Harness. Write your working agreement once in core/rules/global-rules.md. npm run sync compiles it into the rules file of 18 clients (Claude Code, Codex, Cursor, Windsurf, Copilot, Cline, Aider, OpenHands, Goose, Continue, Zed, Trae, Amazon Q, Cody, Antigravity, OpenClaw, Hermes, generic system prompt), links your skills directory into the 14 that have one, and registers the same guard hook in the 10 that support hooks. Corrections and errors from those sessions are harvested back into candidate rules, which you approve from a local dashboard.
• Agent. agnostic is a Textual TUI coding agent with file / shell / search tools, subagents, an AST symbol index, checkpoints, a test-and-fix loop and an arrow-key /model picker covering LM Studio and Ollama, hosted Gemini / Claude / GPT / DeepSeek presets, and logged-in claude / codex / agy CLIs with no API key (with per-subscription model pinning and session continuity). It has persistent auto-memory, MCP tool servers (.mcp.json), per-model cost/latency tracking, a multi-line composer, double-Esc rewind, a /diff turn browser, !cmd shell escape and a headless agnostic -p mode for scripting it as a subagent.
• Policy. core/safety/guards.json is read by the Python guard, the Node hooks and the dashboard. Secret paths are always blocked; hard-stop commands need a human unless you opt in; a missing or unreachable policy fails closed.

Developed on Windows 11; CI runs on Ubuntu with Python 3.9 and 3.12 and Node 20.

Quick start

Requirements: Python 3.9+, Node 18+. There are no npm dependencies.

git clone https://github.com/ucsandman/agnostic-harness.git
cd agnostic-harness
pip install -e .

# Edit your rules, then compile them to every client
#   core/rules/global-rules.md
npm run sync:check     # what would change (exit 1 if stale)
npm run sync           # write, with a backup of every file it overwrites

# Start the coding agent (LM Studio default endpoint)
agnostic
agnostic --url http://localhost:11434/v1 --model qwen2.5-coder   # Ollama
agnostic -p "explain #CodebaseIndexer in @agent/tools/indexer.py"  # one shot

# Open the command center
npm run dashboard      # http://127.0.0.1:7842

python launch.py runs first-run setup (harvest, skill consolidation, sync), checks parity, checks DashClaw, runs the engine self-tests and then opens the dashboard. It is not interactive.

The agent in two minutes

agnostic > /model                      # pick endpoint, preset and effort
agnostic > @agent/loop.py how does tool dispatch work?
agnostic > /plan add a --dry-run flag to sync
agnostic > /test                       # detect runner, loop fixes until green
agnostic > /review                     # reviewer subagent over the diff
agnostic > /commit

• @path injects a file, #symbol injects one function or class from the AST index. Tab completes both against the workspace index, in either shell.
• Read-only tools (read_file, grep_search, find_files, get_outline, find_symbol) run in parallel when the model asks for several at once. read_file and get_outline truncate past 120 lines to head and tail (an explicit line range is never truncated); grep_search and find_files stop at 40 and 50 results and say so.
• /trust reads|tests|all sets the session tier. Only all lets hard-stop commands run without a prompt; start with --ask-permissions to be asked instead of denied. Secret paths are blocked in every tier.
• /undo, /checkpoint, /session, /compact, /swarm, /diagram, /pr, /learn and the rest are in docs/slash-commands.md.
• agnostic --web (or /web) starts a browser companion on 7843 (next free port if taken) with live telemetry, diffs and a context meter.
• Prefer a classic readline shell? agnostic-legacy runs the same loop with prompt_toolkit.

The harness in two minutes

Command	What
npm run sync / sync:check	Compile rules + traits + policy into every client listed in core/templates/targets.json. Drifted (hand-edited) targets are backed up and skipped unless --force.
npm run merge / merge:global	Pull lessons from project CLAUDE.md / AGENTS.md / GEMINI.md files back into the rules.
npm run harvest	Collect errors and corrections from local agent logs into storage/candidates.jsonl.
npm run distill	Run the promotion ladder (observation → fact → rule) and the pruner; write a digest and proposal. Rules are only written into the SSOT when you approve them.
npm run dashboard	Command center on 7842: candidates, rules, skills matrix, per-project recommendations, routines, governance settings, guard simulator.
npm run skills:consolidate / skills:recommend	Gather skills from every client into one catalog; score them per project.
npm run parity / recall	Per-target sync status (7845), rule and memory search (7844).
npm run setup:default	First-run onboarding (what launch.py calls).
npm run dashclaw:setup / dashclaw:status	Optional governed autonomy via DashClaw.

All local UIs bind 127.0.0.1. Mutating routes need a per-process token and a loopback origin, so a page in another tab cannot trigger a sync or approve a rule.

Full client table (rules file, hook config, skills dir per client), generated from targets.json: docs/targets.md.

Documentation

docs/architecture.md	How sync, hooks, harvest, distill and the agent loop fit together; storage layout.
docs/configuration.md	Files you edit, every env var, CLI flags, trust tiers, DashClaw, scheduled jobs, uninstall.
docs/slash-commands.md	Every slash command, and whether the TUI or legacy shell handles it.
docs/targets.md	The 18 supported clients.
CONTRIBUTING.md	Setup, test commands, where things live, rules for a change.
SECURITY.md	Scope, reporting, what the guard is not.
CHANGELOG.md	Release notes.

Repository layout

agent/          Python coding agent: tui.py (default), cli.py (legacy), loop.py,
                llm/ (client, presets, endpoint detector), tools/ (registry, indexer,
                subagents), governance/ (guard, audit, undo, context, sessions),
                workflows/ (swarm, tester, pr_pilot, diagram, scheduler), web/
core/           Single source of truth: rules/, traits/, safety/guards.json,
                templates/targets.json, examples/
engine/         Zero-dependency Node: sync/, hooks/, harvest/, distill/, ingest/,
                skills/, audit/, setup/, docs/, tests/
tools/          Local web UIs: dashboard/, recall/, sync/ (parity)
jobs/           PowerShell wrappers for scheduled sync and nightly distill
tests/          pytest suite for the agent
storage/        Runtime state (gitignored except .gitkeep); backups of every synced file
launch.py       First-run + dashboard launcher

Development

pip install -e ".[dev]"
ruff check .
python -m pytest tests/ -q
npm test                 # engine + sync + hook regression suites
npm run docs:check

See CONTRIBUTING.md.

Using this repository as a template

Click Use this template on GitHub, clone your copy, replace core/rules/global-rules.md and core/traits/traits.md with your own working agreement, and run npm run sync. Set AGNOSTIC_PROJECTS_DIR if your projects do not live in C:\Projects / ~/Projects.

License

MIT. See LICENSE.