hierarchical-codex

hierarchical-codex is a deterministic MCP control plane for native Codex subagents. Codex still creates UI-visible Sol → Terra → Luna threads with spawn_agent; this project supplies the durable state, policy gates, budgets, artifacts, evidence workflow, and recovery protocol around those threads.

Status: engineering MVP. The control plane, Codex project integration, hooks, tests, and operating documentation are implemented. Validate model availability and the exact Codex core version on every target App/IDE deployment.

Design choice

The project follows one explicit route:

Model-driven spawning; code-enforced constraints.

Codex native execution plane
  spawn_agent / wait_agent / send_message / Subagents UI
                  |
                  | task_id in TaskEnvelope
                  v
TypeScript MCP control plane
  SQLite task ledger / leases / budgets / artifacts / evidence / audit
                  ^
                  |
Codex Skills + Profiles + Hooks
  orchestration policy / model-effort routing / spawn veto / stop checks

MCP and hooks do not create native threads. The Sol or Terra model calls native spawn_agent; code validates and records the surrounding workflow.

Related MCP server: dingdawg-governance

Implemented capabilities

• Mission and task ledgers with optimistic versions.

• Direct-parent role policy: Sol director → Terra coordinator → Luna leaf.

• Model/effort matrix:

  • Sol: high, xhigh, max

  • Terra: xhigh, max

  • Luna: high, xhigh, max

• Dependency-aware readiness and bounded child allocation.

• Expiring worker leases, heartbeats, release, and safe reclamation.

• Hierarchical token, cost, wall-time, tool-call, and child-count budgets.

• Content-addressed artifact storage with bounded reads.

• Candidate → checked → verified → committed evidence gates.

• Producer/reviewer separation.

• Request-hashed idempotent mutations and append-only audit events.

• Recovery snapshots after restart or context compaction.

• Project-scoped Codex Skill, Agent profiles, MCP configuration, and hooks.

Requirements

• Node.js 22.5 or newer. Node 26 is used in development.

• Python 3.10 or newer for Codex lifecycle hooks.

• Codex 0.148.0 or newer is the recommended production baseline, with native multi-agent tools, custom agents, MCP, and hooks.

• A trusted Codex project so .codex/config.toml and project hooks are loaded.

Quick start

cd "/mnt/tools/others/codes/web project/hierarchical-codex"
npm install
npm run check
npm run doctor

Then:

1. Open this repository root in Codex App, Codex CLI, or the Codex VS Code extension. npm run doctor does not install the skill into ChatGPT; the App must use this folder as its workspace.

2. Trust the project when prompted. Untrusted projects hide .codex skills.

3. Start a new root conversation with gpt-5.6-sol (skills load at startup).

4. Type $ and select prism, or invoke $prism <mission>. If the picker is empty, AGENTS.md still instructs the root Sol.

5. Inspect native child activity in the Subagents UI and durable state through the MCP tools.

To use the full stack from other folders in the VS Code Codex extension or CLI, run npm run install:user after npm run build. See docs/USER_INSTALL.md. Do not copy this repo's .codex/config.toml into ~/.codex; that would pin Sol as the default model and block ordinary subagents.

The project MCP configuration launches node dist/cli.js with the repository root as its working directory. Run npm run build after source changes.

Development commands

npm run dev          # Run the stdio MCP server from TypeScript
npm run doctor       # Validate runtime and Codex integration files
npm run test         # Unit and integration tests
npm run typecheck    # Strict TypeScript checks
npm run lint         # ESLint
npm run format       # Prettier
npm run build        # Compile dist/
npm run check          # Full local quality gate
npm run install:user   # Install skill, agents, hooks, and MCP into ~/.codex
npm run doctor:user    # Verify the user-global install
npm run uninstall:user # Remove the managed user-global files

The MCP server writes protocol messages to stdout. Application logging must use stderr; stdout logging corrupts stdio MCP transport.

Runtime state

By default, state is project-local and ignored by Git:

.hierarchical-codex/
├── control-plane.sqlite
└── artifacts/
    └── <sha-prefix>/<sha256>

User-global install (npm run install:user) stores the ledger at ~/.local/share/hierarchical-codex/ so Codex MCP sandboxes can write it. If that directory is not writable, the server falls back to a temp path and logs the chosen home on stderr.

Configuration environment variables:

• HIERARCHICAL_CODEX_HOME

• HIERARCHICAL_CODEX_DB

• HIERARCHICAL_CODEX_ARTIFACTS

• HIERARCHICAL_CODEX_MAX_ARTIFACT_BYTES

• HIERARCHICAL_CODEX_DEFAULT_LEASE_SECONDS

• HIERARCHICAL_CODEX_MAX_LEASE_SECONDS

• HIERARCHICAL_CODEX_EVENT_PAGE_SIZE

MCP tools

Mission:

• mission_create

• mission_get

• mission_close

Task and lease:

• task_allocate

• task_get

• task_claim

• task_start

• task_heartbeat

• task_release

• task_block

• task_fail

• task_cancel

• task_supersede

• task_set_effort

• task_commit

Artifacts and evidence:

• artifact_put

• artifact_get

• result_submit_candidate

• result_check

• result_verify

Accounting and recovery:

• budget_report

• recovery_snapshot

See docs/API.md for contracts and docs/PROTOCOL.md for the orchestration sequence.

Repository structure

.agents/skills/             Codex orchestration Skill
.codex/agents/              Sol/Terra/Luna custom profiles
.codex/hooks/               Python policy gates
.codex/config.toml          MCP and native agent configuration
src/domain/                 State and policy definitions
src/infra/                  SQLite repository and artifact storage
src/mcp/                    MCP tool registration
tests/                      Control-plane, policy, and hook tests
docs/                       Architecture, protocol, operations, ADRs, records

Important boundaries

• External MCP code cannot call the internal native spawn_agent registry.

• Hooks can deny, rewrite, or add context, but SubagentStart cannot prevent a subagent after creation.

• Native UI state is observational; SQLite is the durable workflow source.

• This MVP is single-host. SQLite serializes mutations but is not a distributed consensus system.

• Token usage is reported by agents/hosts; the MCP server cannot independently meter model tokens.

• The repository is currently UNLICENSED; add an explicit license before external distribution.

Documentation

• Architecture

• Protocol

• MCP API

• Codex integration

• User-global VS Code / CLI install

• Operations

• Security model

• Development guide

• Test strategy

• Roadmap

• Changelog

• Implementation record

• Architecture decisions