No response grows with the size of the change

compute_diff returned the whole diff beside the run_id every later call needs. On a 205,432-character change the response was 234,337 characters — past the client’s tool-result cap, so it was spilled to a file whole and the run_id went with it. The reviewer has no file access by design, so it could neither record a round nor submit. The bigger the change, the more certain the failure.

Now compute_diff returns the handle, the SHAs and a numbered hunk index; the bytes come from read_diff, paged on hunk boundaries — the unit the anchors already speak in. Envelope on the change that prompted it: 234,337 → 15,622 characters, twelve pages, every walked hunk byte-identical on reassembly.

Breaking: compute_diff no longer returns diff. The document format is unchanged.

Also

• •A run handle, not a working directory. Rounds were filed under the server’s cwd while submits passed the repo — different keys in a multi-repo workspace, so the submit found no interview and the reviewer re-ran everything. A run is now sha256(repo|base_sha|head_sha), required by both tools.
• •Anchors by hunk id. A stop says H3, H4 instead of restating line numbers from a diff header. Ids are expanded before validation, so nothing downstream changes.
• •A transcript parser per agent. Claude Code and Codex transcripts share nothing; guessing at both half-worked in silence — get_spine reduced 831 Codex entries to one gap item.
• •A tour grouped by the plan, not the diff. Grouping by file structure produces stops that are really just filenames.
• •The reviewer lost the orchestrator’s tools. It could reset consent for every repository on the machine, and in one run reached for it after two failed submits.