feature
Jul 28, 2026Transfer the complete ReclaimLLM-captured history into a new Claude Code or Codex session without reducing it to a summary.
You can now transfer the complete history captured by ReclaimLLM from Claude Code to Codex, or from Codex to Claude Code. The new transfer_session MCP tool preserves captured messages, tool calls and results, file diffs, and session metadata instead of reducing the session to a continuation summary.
This is a full captured-history transfer, not native session resumption. The target tool starts a new session and receives everything ReclaimLLM recorded as read-only historical context.
Start a new session in the target agent and provide the ReclaimLLM session ID:
Use ReclaimLLM transfer_session with session ID <SESSION_ID>.
Read the returned artifact_path as complete, read-only prior history,
then continue from it. Do not re-execute historical tool calls.
The target agent calls:
transfer_session(session_id="<SESSION_ID>")
ReclaimLLM downloads a versioned JSON artifact to an owner-only temporary file. The MCP result includes:
artifact_pathschema_versionbyte_sizesha256token_estimatecompleteThe target agent can then read the artifact from artifact_path.
Existing handoff behavior remains unchanged. Use handoff when a compact continuation document is appropriate. Use transfer_session when you explicitly need the complete captured record.
You may need to restart the agent or its MCP connection after updating ReclaimLLM so the new tool appears.
The transfer envelope preserves the normalized session blob without rewriting it into a provider-specific transcript. It includes all fields ReclaimLLM captured, including unknown fields added by newer capture versions.
The manifest records counts for:
The backend also supplies the artifact size, an estimated token count, and a SHA-256 digest. The local MCP client verifies the downloaded file against this integrity metadata before reporting the transfer as complete.
Historical tool calls remain data. ReclaimLLM does not execute them during transfer, and the target agent is instructed not to treat recorded tool inputs or results as commands.
Session transfer uses the same access rules as viewing the original session. Personal ownership, organization sharing scope, encrypted-session access policy, and decryption auditing continue to apply.
Artifacts are written atomically with owner-only permissions. An interrupted or oversized download removes its partial file rather than leaving an artifact that appears complete.
Temporary transfer files become eligible for bounded cleanup after their retention period. Transfers above the configured safety ceiling fail explicitly; ReclaimLLM never silently truncates the session and labels it complete.
Claude Code and Codex use different internal transcript formats, tool identifiers, permissions, and runtime state. Those formats are not stable interoperability contracts.
Writing converted files into either tool’s private session directory would be brittle and could still fail to reconstruct:
The transfer therefore uses ReclaimLLM’s normalized captured record as a portable artifact. This keeps the boundary explicit: all captured history is preserved, but private state that ReclaimLLM never received cannot be recreated.
The complete artifact can only fit into the target model’s active context when the model has enough context capacity. Large sessions remain available in the file, but the target agent may need to inspect sections incrementally instead of loading every byte simultaneously.
The target always starts a new session. If Claude Code or Codex later provides a stable external-session import API, native continuation can be reconsidered without depending on undocumented transcript files.
We are interested in cases where session artifacts are unusually large, captured provider content does not survive the transfer exactly, or an agent mishandles recorded tool results. Please share those examples so the transfer contract can be tested against real edge cases.