[BinaryMuse]

The write side for Kiro, as #4580 was for Copilot: a thread can now be restored as a Kiro session on another machine, caught up or switched when two copies diverge, forked, and continued into Kiro from another agent, and sync appends to and replaces Kiro's logs. writes_transcripts goes away again, as it did there.

The writer (kiro/rehydrate.rs) puts a session where Kiro looks for it: sessions/<hash of the directory>/sess_<id>/ with the header, the log in Kiro's row shapes (each turn framed by turn_start/turn_end, shell calls as execute_bash with Output:…Exit Code: N), and the session-index line; a fork names its original in the header (parentSessionId, forkedAtMessageId) and under _meta.kiro.atuin. What Kiro needs of a hand-written session was found by probing 2.36.0 and is in the module docs: a tool_call must carry a status or the call and its result drop out of the model's context; foreign ids, foreign tool names, failed results and a missing session_start are all fine; an agent_note reaches the model, so the continuation marker goes there. The reader now links each row to the one before it, so sync sees a Kiro session as a tree, and reads the header's parentSessionId as a fork's parent.

Liveness for an idle copy asks the process table (a kiro-cli-chat or its engine working in the session's directory), since Kiro keeps no per-session lock: a just-restored copy refused its own catch-up otherwise.

Shaken down against the real CLI: a Claude Code thread continued into Kiro resumed there quoting the marker and the carried facts; a captured Kiro session restored under a second KIRO_HOME resumed with its conversation and tool call. The end-to-end suite runs Kiro through the restore, catch-up, switch, fork and continuation cases.

A cross-harness stress test with the real agents (102 hops, 33 of them Kiro, every Kiro pair in both directions) turned up one bug, fixed here: Kiro ends each turn with a redacted reasoning row after the reply, which made every plain catch-up of a native copy look like a divergence, since the restored copy lacked that row and the classifier's fast-forward rule did not know Kiro. The writer now writes the marker back and Kiro is in the rule. The same test showed a written session opening under Kiro's Default profile whatever --agent says, so the header now names the user's chat.defaultAgent.

[greptile-apps]

Confidence Score: 5/5

[Medium impact] The PR appears safe to merge; the previous branch-mixing issue is fixed.

Summary

Adds writing, restoring, catching up, switching, forking, and continuing into Kiro sessions.

• Atuin can restore, fork, and continue threads as Kiro sessions.
• Kiro copies can catch up or switch between conversation branches.

_{Reviews (6) · Last reviewed commit: "fix(threads): fast-forward a Kiro copy o..." · Reviewed by Greptile}

[atuin-bot]

Landed as d784213. GitHub shows this merged once it reaches main.