Claude Code + Codex collaboration

LazyTools connects writing, independent review, durable projects and quota decisions. Start with the worked toy-repo bridge flow: Claude writes the weekday deployment guard, Codex reviews, code executes the boundary tests, and the operator inspects the evidence. Reverse the engines by choosing Codex as writer; tier-based --review-of then requires Claude.

Work	Integration	Evidence
Long write / review	Bridge run --detach, --session, --tier, --review-of	Jobs, routing, native session, result, failure workspace snapshot
Approval	Bridge pending, approve, reject	Durable tickets, actor, reason, expiry
Short turn	Engine-backed MCP tools	Reply header, findings or error, session alias
Shared project	Projects on a configured Store	Plan, claims, contracts, checks, review, decision, notes
Schedule	python -m lazytools.projects.gantt	English HTML/SVG and text computed from recorded dates
Quota	Bridge route; Python route() / recommend()	Rung, exclusions, windows, reservations, admission reason

Tier-based bridge --review-of excludes the producing engine and refuses with human_review_required if no opposite reviewer is eligible. MCP callers choose their reviewer themselves. The bridge runs a write-capable job even for review prompts; MCP reviewers provide a read-only surface when required.

Checks and acceptance are separate steps. Reviewer prose is not a test result. Timeout does not mean no files changed: inspect workspace evidence, Git diff, and checks before continuing. Project verification has evidence gates and an injectable caller authorization policy. Neither a successful job nor the last output of an exhausted LazyBridge verify= loop implies acceptance.

Native conversations do not migrate. Use separate writer/reviewer aliases; share the task, diff and project records. The bridge locks one job per Git repository: wait for the writer to finish before reviewing in that checkout.

Existing CLI collaboration builder

build_cli_collaboration() remains available for the older read/plan workflow: Claude analyses, Codex critiques, a provider-model synthesizer plans and an executor produces the response. It uses CLI subprocess wrappers and separate provider engines, so it is not a detached bridge or an enforced acceptance gate. See its full signature in the reference.

Benchmark in progress. Functional tests establish mechanisms; no comparative quality, time, quota-efficiency or subscription-savings result is published here. Tier ladders are deterministic policy, not a measured ranking.