OpenClinXR OpenClaw-Style Bridge

Purpose

Use this as Codex's bridge into the same repo-native OpenClaw-style / OpenClaw-inspired operating pattern that Grok uses. This is not an external OpenClaw runtime. The canonical sources stay in the repo; do not duplicate their contents into Codex config.

Required Start

1. Read AGENTS.md + PROJECT_STATUS.md (first ~40 lines).
2. Load active slice brief: .openclinxr/slices/<slice-id>/brief.json.
3. Spawn parallel team: pnpm openclaw:team-spawn -- --slice-id <id> --phase scout|execute.
4. Verify completion: pnpm openclaw:slice:verify -- --slice-id <id>.
5. Acquire lease before integrator merges: pnpm openclaw:lease -- acquire --owner integrator --slice "<slice>" --ttl-minutes 60.

New slice from template: pnpm openclaw:slice:init -- --template real-garment-v1 --slice-id <id>.
Autonomous execution: pnpm openclaw:run-next or pnpm grok:automation-prompt for scheduler-based heartbeat loops.

Legacy PROJECT_COORDINATION_INDEX.md / AUTONOMOUS_WORK_PLAN.md are historical audit ledgers — do not append per-slice checkpoints there. Canonical state is in PROJECT_STATUS.md.

Repo-Agent Consultation

• Coordinator first: read agents/coordinator/chief-coordinator/charter.md and memory.md for orchestration/drift-sensitive work.
• Drift concerns: read agents/adversarial/openclaw-drift-police/charter.md and memory.md.
• Planning gaps: read agents/core/implementation-planning-lead/ and agents/adversarial/implementation-plan-gap-attacker/ when selecting slices or repairing drift.
• XR/assets: read XR Systems Architect and Asset Pipeline Lead role files before runtime, humanoid, asset, IWSDK, or visual evidence changes.

Use live subagents only when available and materially useful. Otherwise perform local role consultation and record only durable outcomes.

Shared Guardrails

• Rules live canonically in agents/rules/*.md; Grok mirrors them via .grok/rules.
• Grok hooks live in .grok/hooks; Codex lifecycle hooks live in .codex/hooks.json and route through pnpm codex:hook.
• Codex project hooks must be reviewed/trusted with /hooks when their hash changes.
• No canonical state update for no-op readiness checks. Routine runner output belongs in .openclinxr/openclaw/run-next-report.json.
• Canonical state updates are for product changes, verification evidence, or real blockers.

Model routing (harness-specific)

• Grok: prefer direct DeepSeek (deepseek-v4-flash for scouts, deepseek-v4-pro for implementation/review) and grok-build for frontier synthesis. Do not use Moonbridge as the primary Grok subagent route when direct DeepSeek is available.
• Codex Desktop: cannot select DeepSeek in the native model picker. Use tier-appropriate models from generated .codex/agents/*.toml. Moonbridge (pnpm local:moonbridge:probe) is Codex-only optional assist for bounded first-pass review on scout/expert roles—not for implementation or readiness judgment.
• Codex Luna: an efficient Codex-side coordinator/triage alternative when available, not a claim of superiority over DeepSeek. DeepSeek dispatch remains through BothyBoard/Grok tooling; do not invent a direct Desktop dispatch path.
• Codex harness details: when an external monitor, fresh codex exec wake, or parallel Codex worker is involved, read the conditional BothyBoard Codex harness addendum. It covers sync-based observation, no Desktop-task resume, disjoint worktrees, and board mutation authority without duplicating the core protocol.
• Production pipeline: asset generation and scene optimization may require agentic evaluation behind a swappable ModelAssistProvider (Moonbridge today; approved online models later). Procedural-only pipelines are the goal, but online AI is permitted behind explicit gates.

See docs/agent-factory/model-assignment-policy.md and packages/openclinxr/agent-loop/src/role-harness-policy.ts.

Memory write-back

When a slice yields a durable role lesson:

pnpm agent:memory:append -- --role <role-id> --topic <topic> --lesson "<text>"

Verification

After coordination file edits (PROJECT_STATUS.md, teams/, AGENTS.md) run:

pnpm docs:drift-check
pnpm agent:alignment

Per-slice completion uses machine verify (not MD ritual):

pnpm openclaw:slice:verify -- --slice-id <id>

For product code, also run focused tests for touched packages before broader gates.