Consult Claude

The consulted Claude owns a sealed laboratory. You own the living checkout, integration, and final decision. A consultation is not a packet-only print call: Claude gets an editable snapshot of the current repository and may research, rewrite code, run tests, commit experiments, and change its mind there. The living checkout is not part of Claude's tool boundary.

Give Claude one outcome, the settled values it must preserve, and source territory worth starting from. Do not give it your working theory, a menu of answers, or a prescribed method. Tell it what would make the research complete. The runner supplies the snapshot ID, laboratory boundary, checkpoint path, and permission profile.

Start the native Claude background session from the repository root and keep the command attached when the user wants the answer before this task continues:

bun .agents/skills/consult-claude/scripts/consult-claude.ts start --wait

Type the compact research brief, then send EOF. start prints a run ID, native Agent View ID, native session ID, and checkpoint path. It creates an independent replica, removes its Git remote, and keeps Bash network egress offline. Claude may search the web, but direct fetching or any external action is a deliberate escalation, not part of the default run. It does not recreate Claude's session manager: inspect, attach, steer, or stop the live worker with claude agents.

Read the latest result without waking the worker:

bun .agents/skills/consult-claude/scripts/consult-claude.ts status <run-id>

When Claude writes state: needs-decision or state: complete, you may continue the same laboratory conversation. The runner archives the prior checkpoint and waits for a fresh one when requested:

bun .agents/skills/consult-claude/scripts/consult-claude.ts follow-up <run-id> --wait

Type the follow-up, then send EOF. Do not use this to interrupt a working laboratory; reply through claude agents when immediate steering is needed. Every checkpoint is tied to one snapshot. Treat a claim that cannot be re-verified against the living checkout as a question, not a fact. When live work has materially moved on, start a new laboratory snapshot; never silently refresh a worker's replica beneath it.

Use each checkpoint to decide whether to edit, ask Claude a sharper question, or stop. Consultation does not transfer authorship: Claude-generated patches are evidence, not changes to apply. Return Claude's recommendation, material objections, and any remaining disagreement to the user.