Kilo’s Migration API Lets Agents Resume Without Shipping Their Transcripts

Kilo Code’s public migration endpoint lets a configured server rediscover Claude Code or Codex history it can read, create a Kilo session per source, and make a normal retry a skip instead of a duplicate.

A rough-paper flow shows a client sending formats and IDs to a configured Kilo server, which discovers a session, records migration metadata, and skips it on a normal rerun.

Diagram Punkidempotent migration starts with a server that can see the source.

repo Kilo-Org/kilocode evidence

4 source signals 1 repo commit 6cc21af

Evidence: commit 6cc21af / September 1, 2026 / Daily Edition

API discovery: Kilo Code’s public OpenAPI contract defines POST /kilocode/migrate/sessions. Its fresh server-side migration change removes the awkward older pattern of sending transcript bytes and a target session from the client. Instead, the configured Kilo server rediscovers Claude Code or Codex sources it can read, creates one Kilo session per source, and records migration provenance so an ordinary rerun can return skipped: true.

Try it

Build a small migration concierge for a team changing agent clients. Let a developer select a prior Codex session, then have the configured Kilo service migrate only that source and return the new Kilo session ID for a review screen. The useful part is not bulk import; it is a retry-safe handoff that never asks a browser or remote client to upload transcript bytes merely because the server needs to do the import.

POST /kilocode/migrate/sessions
Content-Type: application/json

{
  "cwd": "/workspace/atlas",
  "formats": ["codex"],
  "ids": ["selected-session-uuid"],
  "agent": "build"
}

The public contract also allows a model and an explicit force flag. Leave force off for a normal retry: the integration tests show that a second identical migration points back to the existing session and is counted as skipped.

Copy-paste agent instruction

In a disposable, authenticated Kilo Code server that is authorized to read the selected machine's Claude Code or Codex session directories, inspect the public OpenAPI contract for POST /kilocode/migrate/sessions. Ask the user to select one session ID and one format, then call the configured server with filters only; do not upload transcript bytes. Surface each returned session ID, skipped flag, dropped-content notice, and per-session error. Treat a normal rerun as a no-op, and do not use force unless the user explicitly wants a new migration.

Test caveat

This reporting run did not call a live Kilo server. The inspected public sources describe the endpoint and its tests but do not publish a base URL, authentication recipe, rate limit, or permission model. The server must actually have authorized access to the relevant source directories; it cannot discover a developer’s files on another machine. Kilo’s public CLI documentation also says imports can skip content it cannot represent. Test against disposable transcripts, review the returned dropped entries, and do not treat server-side discovery as a privacy guarantee.

Primary Evidence

Public API source: https://github.com/Kilo-Org/kilocode/blob/6cc21af6d0e19887d5be887a9771865ccf2963af/packages/sdk/openapi.json

Shows: Kilo Code publishes the POST /kilocode/migrate/sessions API surface, including formats, selected IDs, agent, model, force, response schema, and the provider's description of an idempotent external-session migration service.

Supporting definition or documentation: https://github.com/Kilo-Org/kilocode/blob/6cc21af6d0e19887d5be887a9771865ccf2963af/packages/opencode/test/kilocode/session-resume-integration.test.ts

Shows: the public integration test checks request behavior: server-side discovery, one created Kilo session per transcript, persisted source metadata, skipped reruns, force remigration, selected IDs, and per-entry errors for unreadable transcripts.

Migration-design receipt: https://github.com/Kilo-Org/kilocode/commit/6cc21af6d0e19887d5be887a9771865ccf2963af

Shows: the main-branch change removes client transcript bytes and a client target session from the request, re-discovers Claude Code and Codex sources server-side, and removes an empty session when a post-create write fails.

Public CLI documentation: https://github.com/Kilo-Org/kilocode/blob/6cc21af6d0e19887d5be887a9771865ccf2963af/packages/kilo-docs/pages/code-with-ai/platforms/cli.md

Shows: the documented interactive import flow, source directories, empty-session requirement, and the limit that unrepresentable content can be skipped.

Evidence Limits

The public sources establish a source-level API contract and test behavior, not a live endpoint, deployment URL, authentication setup, rate limit, access control, or release availability.

A server must have authorized filesystem access to the source session roots. Do not send production transcript contents to an unverified service or assume this endpoint can discover history that exists only on another machine.

Kilo’s Migration API Lets Agents Resume Without Shipping Their Transcripts

Try it

Copy-paste agent instruction

Test caveat

Receipts below the story

Primary Evidence

Evidence Limits

Send a note to the desk

Kilo’s Migration API Lets Agents Resume Without Shipping Their Transcripts

Try it

Copy-paste agent instruction

Test caveat

Receipts below the story

Primary Evidence

Evidence Limits

Atlas Context

Memory Is a Typed Storage Contract

Send a note to the desk

Same Edition

A Completed Agent Run Is Not a Successful One

Builder Work: Mastra Tests Memory Prompts Against Their Own Noise