Skip to content

Repository files navigation

Codex Conductor

A dependency-free Python runtime and Codex skill for bounded coding work. The default path makes one Codex call with no planner or daemon. State stays outside the repository; write-capable work happens in an external staged copy and needs independent verification before application.

Install

Requires Python 3.12 or newer, a signed-in Codex CLI, and macOS or Linux.

Extract dist/codex-conductor-bundle.zip, inspect the plan, and install:

python3 install.py --dry-run
python3 install.py --allow-writes --approve conductor-install

The installer verifies every installed byte and writes only under the configured Codex and Conductor homes, which both default under ~/.codex. Replacing different bytes also requires --replace --approve conductor-update.

The bundle also installs five manifest-bound custom agent profiles without replacing unrelated personal profiles, plus all eight manifest-bound agent-performance skill trees. It leaves $CODEX_HOME/AGENTS.md untouched unless you explicitly opt into the canonical ambient routing block:

python3 install.py --manage-routing-policy --dry-run
python3 install.py --manage-routing-policy \
  --allow-writes --approve conductor-install

Start a new Codex session after installation. See Custom Codex agents for the profile roles, routing policy, update gates, and discovery check.

The core bundle deliberately remains single-writer. If the optional extras isolated-team runtime is not installed, ambient routing serializes write packets; it never falls back to concurrent shared-workspace writers.

Use

For normal repository work, state the outcome. When enabled, the ambient policy defaults to direct execution and silently composes the relevant specialist behavior. Use the CLI as an advanced escape hatch when you need durable receipts, exact resume, reusable workflow JSON, or direct control over staging, budgets, and permissions.

Run one read-only model call:

~/.codex/conductor/bin/conductor-runtime.pyz auto --task "Inspect the parser" \
  --workspace . --allow-agent

Run a staged edit with a read-only verifier:

~/.codex/conductor/bin/conductor-runtime.pyz auto --task "Fix the parser" \
  --workspace . --allow-agent --allow-writes

The source remains unchanged until verified evidence is applied:

~/.codex/conductor/bin/conductor-runtime.pyz apply /path/to/stages/work.json \
  --workspace . --allow-writes --approve verified-stage-apply

From a source checkout, replace the installed path with python3 -m conductor_runtime. Goal mode supports bounded verifier-driven repair; workflow mode is for independent read-only map-to-synthesis work. Captured output can be inspected with results list|get|search|outline.

See the runtime reference for commands and permissions, or the documentation index for optional facilities and compatibility notes.

Agent performance plugin

The release bundle already installs just-do-it plus focused skills for change contracts, debugging, test design, review, simplification, completion proof, and matched comparisons. The separately installable agent-performance plugin is an alternative distribution channel for users who do not install the bundle. From the repository root:

codex plugin marketplace add .
codex plugin add agent-performance@codex-ultracode-local

Do not enable both channels in one Codex installation because they expose the same skill names; differing direct trees are also update-gated. Start a new Codex thread after installation, then invoke it with $just-do-it when an explicit trigger is useful. See the plugin README for behavior, validation, and the optional launch-time identity.

Safety

Agent, write, destructive, network, parallel, risk, and shell capabilities are explicit. Runs bind their workflow, workspace, policy, context, artifacts, and receipts. Staged application verifies source fingerprints and rolls back on failure.

Self-hashes are not signatures; authenticate releases with dist/SHA256SUMS from a trusted channel. Approved shell commands run from an isolated working copy, not an operating-system sandbox against deliberate writes to external absolute paths.

Development

Run the focused gate while editing:

python3 tools/verify.py --quick --skip-artifacts

The release report describes the pinned full gate, evidence refresh, and artifact checks. Historical and specialist commands remain in the optional conductor_extras archive and are in compatibility maintenance.

This is an unofficial project and is not affiliated with or endorsed by OpenAI or Anthropic.

License

MIT. See LICENSE.

About

A dependency-free runtime and skill for bounded coding work, staged writes, and independent verification.

Topics

Resources

Stars

0 stars

Watchers

0 watching

Forks

Releases

Contributors

Languages