Recursive Outcome Delivery for Codex

A Codex plugin for Outcome-driven recursive decomposition: turn a mission into necessary outcomes and unknowns, run decisive experiments, deliver evidence, integrate, measure, and recursively replan.

The unit of progress is a verified outcome or a consequential uncertainty resolved. For example, a migration is done only after destination readback proves the requested data exists; successful export and import commands alone do not establish success.

Install

Requires a Codex host with plugin support. The authoring/runtime target is Codex CLI 0.159.2 and the local Codex Desktop host. Python 3.10+ is needed only for the optional state tool.

codex plugin marketplace add trukhinyuri/outcome-driven-recursive-decomposition --ref main
codex plugin add outcome-driven-recursive-decomposition@recursive-outcome-delivery
codex plugin list --json

For local development:

codex plugin marketplace add /absolute/path/to/outcome-driven-recursive-decomposition
codex plugin add outcome-driven-recursive-decomposition@recursive-outcome-delivery

Refresh skills or start the next chat after installation; some Desktop versions require restarting the app to see changes. Do not interrupt active work just to restart. Follow your host's plugin installation UI if it does not expose CLI installation. Portable plugin.json and the Codex compatibility manifest carry the same identity; the repository includes its own marketplace catalog.

Use

Explicit:

Use $outcome-decomposition to migrate this system and prove the result in the destination.

The host may display the namespaced name outcome-driven-recursive-decomposition:outcome-decomposition. The skill is eligible for automatic selection when planning complex tasks and when implementing or continuing their approved plans. It covers dependent outcomes, uncertain feasibility, alternative designs, integration risk and replanning. Clear one-step edits, translations and simple lookups should stay lightweight. Automatic eligibility is enabled with allow_implicit_invocation: true; selection remains a host/model decision.

If you ask only for a plan, the agent presents outcomes, dependencies and acceptance checks and stops before implementation. A subsequent human confirmation such as “plan approved, implement it” or “план подтверждаю” continues that task directly through implementation, integration and verification without requiring the skill name or a second approval. Partial approval authorizes only its accepted scope. A direct request to build or fix already authorizes work in its stated scope; the method does not impose an extra planning gate.

For a standing personal preference, a short rule in your global Codex AGENTS.md can explicitly select this skill for those complex-task stages; see Codex custom instructions. This supplements normal discovery and does not install a background process or override task-specific instructions. The official skill documentation explains description-based matching and implicit invocation.

The workflow is:

Original outcome and acceptance
 → necessary outcomes / unknowns / dependencies
 → hypothesis + discriminating test + decision threshold
 → execute and observe
 → retain, reject, or replace the branch
 → integrate and verify the parent
 → update the model and repeat where needed

The agent preserves user intent, authorization and resource limits. It checks current state on resume and does not equate children completed with the root done. See the skill and the optional local state tool.

Version 2 task journals

Version 2 fixes a stale-evidence defect after replacing an outcome branch. Changes to the replacement now invalidate the dependent outcomes, their children and downstream consumers. Existing version 1 journals remain readable, but require an explicit upgrade and fresh verification before they can report a completed mission. The upgrade preserves the historical events and hashes; it does not approve or execute external work. See the migration guide and verification record.

Quality and limitations

The package combines a concise skill, progressive references, a deterministic local evidence/state engine, adversarial cases and tests. Its design basis includes NASA, DARPA, ADaPT and Reflexion. Those sources support design choices; they do not prove this composition improves every task.

The agent still judges source quality, performs real work and decides whether the goal matters. The state tool verifies recorded graph and evidence invariants, not the truth of model-entered observations. It has no autonomous executor or hidden background activity. No comparative benchmark establishes a world-best claim. The validation record distinguishes unit tests, host readback and observed model behavior. The frontier research and candidate screening record reports two tested candidates, the reasons for retaining 1.0.1 and the remaining comparative limits. The public frontier comparator contracts identify historical and current benchmark targets, exact known settings, and the infrastructure and access gaps preventing a full run.

python3 -m unittest discover -s tests -v
python3 tools/host_check.py --cwd /path/to/project

Live model evaluation is opt-in: prompts and decision criteria are in evals/cases.json. Host readback alone launches no model turn. A passed smoke case proves that case on the tested host, not universal automatic activation.

Privacy and removal

No MCP connection, hooks, telemetry, remote service, credential reader or model default change. The helper writes only explicitly selected local task state. Keep private observations outside the public plugin checkout.

codex plugin remove outcome-driven-recursive-decomposition@recursive-outcome-delivery
codex plugin marketplace remove recursive-outcome-delivery

Uninstallation does not delete task state you created. MIT licensed.