Reviews with another agent's transcript: prior art (2026-09-25)¶
The brief behind /review (DESIGN.md §26): one agent reviews another's work with the engineering context (what was tried, why, what was rejected), not only the diff.
(2026-09-28: /review became an alias of /catchup, which hands the agent exact session ids and a fixed protocol and lets it use whatever session-history tool it has, instead of an agentsview command; see DESIGN.md §26. The research below is unchanged.)
Verdict
Nobody ships exactly this. The closest is hcom. Its agents read each other's real transcripts across Claude, Codex and Cursor, and its README lists "review what claude did" as an example. But a session started by hand needs hooks installed and a restart first, and it doesn't support Devin. The closest product built for review is entire review, but it launches new reviewer agents for a branch instead of using the sessions that are already running.
Closest prior art
| Tool | What B gets from A | Cross-vendor | Live sessions | Link |
|---|---|---|---|---|
| hcom | Real transcript and tool I/O, no hidden reasoning | Yes, no Devin | Only after hook install and restart | repo |
Entire review |
Checkpoint summaries plus raw transcript on request; uncommitted work gives only the last prompt | Yes | No: reviewers are launched fresh, scoped to a branch | repo |
| tandem navigator | Full translated transcript plus diff, every turn | Claude↔Codex | Only sessions run through tandem | PR #83 |
| VS Code Agent Host | Another session's recent conversation | Copilot/Claude/Codex | VS Code-hosted; terminal sessions must be adopted into VS Code | 1.129 |
| codex-plugin-cc | Review sees only the diff; /codex:transfer copies a Claude transcript into Codex |
Claude→Codex only | One-off copy into a new Codex thread | repo |
| agentsview MCP | Raw messages via get_messages |
All four switchboard harnesses (Devin from v0.36.1) | Yes, with a lag | docs |
What's actually open
- The part switchboard can own: @codex-1 review claude-1 between sessions that are already running, across all four vendors, in a room the human reads. No shipped tool does all of that.
- Already common: indexing transcripts from many agents (agentsview, cass), cross-model review of the diff only (roborev), and live messaging between sessions (Claude Code SendMessage, which is Claude-only and deliberately never shares history).
- Build on agentsview and skip writing transcript parsers. Let agentsview deal with vendor format changes (tandem broke on Codex 0.155.1). agentsview has an MCP server from v0.35.0 (agentsview mcp) and indexes Devin sessions from v0.36.1. Prefer MCP or the CLI to curl: in one real attempt, Codex's sandbox blocked localhost:8080. Give B a summary and a way to pull more when it needs it, as Entire and Amp do. Amp also checks whether later messages reverted what it found.
Risks to design for
- Injection and secrets: transcripts contain raw tool output. B must treat it as data, and secrets should be redacted first (or the human warned that they may be there). Reading A's transcript also sends A's session to B's vendor.
- Cost: whole transcripts are expensive, so set a budget. second-opinion caps the transcript at 10% of the reviewer's input.
- Staleness: agentsview search skips sessions active in the last 10 minutes unless include_active is passed, so fetch by session ID. Never --resume A's live session to read it: the two processes' messages get mixed into A's transcript.
- Review quality: reading A's reasoning can bias B toward A's conclusions, so B should review the diff first and read the transcript second. Which model reviews which also matters: on a benchmark of single-file coding tasks, Codex reviewing Claude lowered the pass rate from 91.4% to 82.8% (arXiv 2607.21656).