director-cli 0.9.0

Model-agnostic decomposition coding harness that orchestrates agent providers (OpenCode, Claude Code) behind deterministic gates.

director

A model-agnostic decomposition coding harness — a thin orchestrator over agent providers (OpenCode, Claude Code, Codex, …).

director tests one hypothesis:

A strong planner model decomposes a coding task into small, atomic, well-specified units with acceptance tests written first. A cheaper executor model — local or low-cost cloud — implements each unit in an isolated, fresh context. Deterministic gates (tests, lint, typecheck — exit codes, never an LLM's opinion) decide what merges. This cuts token cost dramatically with minimal quality loss.

It is model-agnostic by construction: roles (planner, executor, reviewer, …) bind to provider/model-ref strings in config. The provider is the tool director drives (opencode, claude-code, or codex); the rest of the tier string is whatever that tool expects. Switching the executor from a local 27B to a frontier model — or anything in between — is a one-line config edit, never a code change.

Status: beta. Validated end-to-end (plan → run → bench) under local, cheap-cloud, and all-frontier executor tiers.

---

Install

director is a pure-standard-library Python CLI (no dependencies), so it installs anywhere Python 3.11+ runs:

uv tool install director-cli      # recommended
# or
pipx install director-cli
# or
pip install director-cli

Prerequisites (providers)

director orchestrates other tools rather than replacing them, so it needs:

• Python ≥ 3.11
• git on PATH (isolation is real git worktrees + branches)
• OpenCode on PATH, if you use an opencode/* tier
• Claude Code (claude) on PATH, if you use a claude-code/* tier
• Codex (codex) on PATH, if you use a codex/* tier (e.g. codex/gpt-5-codex; runs via the OpenAI Codex CLI)
• Provider auth: if you use OpenCode tiers, configure auth in OpenCode via opencode auth; Claude Code tiers use the Claude Code CLI's own auth; Codex tiers use the Codex CLI's own auth

director never manages model-provider keys itself — each tool handles its own credentials.

---

Quickstart

cd your-repo

# 1. Create .director/config.toml interactively — director init asks which model to
#    use per role and what your gate commands are, then writes the config for you.
director init
#    (See director/config.example.toml for the full/advanced schema if you want to
#     hand-tune beyond what init prompts for.)
$EDITOR .director/config.toml

# 2. Install director's role agents into .opencode/ (+ gitignore, starter opencode.json)
#    — written only when the opencode provider is configured
director sync-agents

# 3. Plan: brainstorm → spec → test-gated task DAG (two approval gates)
director plan "Add a --json flag to the export command"

#    …review .director/spec.md, then continue; review the plan, then continue:
director plan --continue            # after approving the spec
director plan --continue            # after approving the plan + failing tests

# 4. Run the DAG: isolated worktree per node, deterministic gates, auto-merge
director run

# 5. Inspect
director status

Unattended? Let the planner self-critique at each gate instead of pausing for you:

director plan "…" --auto              # planner self-critiques at each gate
director plan "…" --auto --no-critique  # gates auto-pass, fully hands-off
director run

---

Commands

Command	What it does
director plan "<task>" [--auto] [--no-critique] [--continue]	Brainstorm → spec → test-gated task DAG, with two artifact-based approval gates.
director run [--parallel N] [--max-attempts K]	Execute the DAG: each node in an isolated git worktree, gated by tests/lint/typecheck, auto-merged on pass; escalates a stuck node one tier up.
director status	Per-node progress, attempts, cost, and the executor-tier completion rate.
director bench "<task>" --profiles a,b,c	Run the same task (same frozen acceptance tests) across profile variants and diff cost / quality / wall-time.
director init [--repo .]	Interactively create .director/config.toml — asks which model to use per role and your gate commands.
director sync-agents	(Re)install the role agents into <repo>/.opencode/ (plus gitignore, starter opencode.json) — only when an opencode/* tier is configured.

All state lives under .director/ (resumable, debuggable): plan.json, state.json, costs.jsonl, metrics.jsonl, per-call logs/, and bench/.

Configuration

director init interactively creates .director/config.toml — it asks which model to use for each role and what your deterministic gate commands are, then writes the config for you. A config is just roles → provider/model-ref strings, the deterministic gate commands, per-model pricing, and run limits. For the full/advanced schema, see the complete, commented director/config.example.toml: it shows how to bind the executor tier to a local model (≈ $0 implementation), a low-cost cloud model (zero local infra), or a frontier model (the expensive baseline). See director/README.md for the full architecture (gates, two-stage review, red-green hardening, metrics).

Two-level configuration

director supports two config levels that are combined at runtime:

• User-level defaults live at ~/.director/config.toml. These apply across all repositories on your machine.
• Repo-local overrides live at <repo>/.director/config.toml. These apply only to the current repository and take precedence over user-level settings.

The repo file deep-merges over the user file: nested tables merge recursively, while scalars and arrays replace wholesale. For example, a user file might supply several roles under [tiers]:

# ~/.director/config.toml
[tiers]
planner  = "claude-code/opus"
executor = "opencode/openrouter/deepseek/deepseek-v4-pro"
reviewer = "claude-code/sonnet"

And a repo file overrides just one role:

# <repo>/.director/config.toml
[tiers]
executor = "opencode/lmstudio/qwen3.6-27b-mtp"

The resulting effective config uses the repo-local executor but falls back to the user values for planner and reviewer (each role binds to a single "provider/model-ref" string — the repo [tiers] overrides only the roles it names).

director init target selection

director init auto-detects its write target: inside a git repo it writes <repo>/.director/config.toml; outside any git repo it writes ~/.director/config.toml. You can override this behavior with flags (mutually exclusive):

• --user forces the user-level target (~/.director/config.toml).
• --local forces the repo-level target (<repo>/.director/config.toml).
• --repo PATH still sets the target directory explicitly.

Comparing setups with bench

director bench plans a task once, then runs the same frozen acceptance tests under several config variants to compare cost/quality/wall-time. Create the variants as .director/profiles/<name>.toml (copy your config.toml and change the executor tier in each), then:

director bench "<task>" --profiles all-frontier,cheap-cloud,local-first

For agents & scripting

director is built to be driven by humans or by another agent:

• Deterministic, non-interactive: --auto --no-critique runs plan→run with no prompts; every merge decision is an exit code, never a chat.
• Machine-readable output: .director/metrics.jsonl (per-node + per-run records) and .director/bench/summary.json are stable JSON for downstream tooling.
• Resumable: re-running plan --continue / run picks up from .director/ state.

---

Development

uv sync                       # create the dev environment
uv run python -m unittest discover -s tests -q   # tests
uvx ruff check . && uvx ruff format --check .     # lint + format
uv build                      # build the wheel/sdist

Releases are automated with python-semantic-release on merge to main (conventional-commit messages drive the version bump, changelog, and PyPI publish via Trusted Publishing). See CONTRIBUTING.md.

License

MIT © Christopher Manzi. The ported TDD/review discipline is adapted from obra/superpowers (MIT).