Move Session Here

Move a session here

Turn one stored Claude Code or Codex transcript into working context for this chat. Resolve a T3 Code thread to its provider transcript when needed. This is an ingestion workflow. Preserve the stored database, WAL, and transcript content.

/move-session-here 9e362a1e-b606-4d60-b3fa-4b6a03d2fcf1
/move-session-here codex://threads/019a1111-2222-7333-8444-555555555555
/move-session-here t3://threads/11111111-1111-4111-8111-111111111111

Require exactly one Claude session ID, Codex thread deeplink, or T3 Code thread deeplink. A bare ID means Claude. A codex://threads/<id> deeplink means Codex. A t3://threads/<id> deeplink means T3 Code. Translate a natural-language request that names a T3 Code thread ID to the T3 Code form.

Extract the session

Resolve <skill-directory> to this skill's installed directory. Create two temporary files, then run:

handoff_path="$(mktemp "${TMPDIR:-/tmp}/move-session-here.XXXXXX")"
handoff_error_path="$(mktemp "${TMPDIR:-/tmp}/move-session-here-error.XXXXXX")"
node <skill-directory>/scripts/session-handoff.mjs \
  '<session-reference>' >"$handoff_path" 2>"$handoff_error_path"

The helper enumerates directory-guarded config homes with find, including CLAUDE_CONFIG_DIR, CODEX_HOME, ~/.claude*, and ~/.codex*. It normalizes the owning transcript and any Claude subagent sidecars into one JSON document. It scopes Codex lookup to sessions/**/rollout-*.jsonl.

For a T3 Code reference, use a current upstream SQLite CLI (3.33.0 or later, with -json support) as sqlite3 on PATH. The helper queries ~/.t3/userdata/state.sqlite in one read-only transaction and resolves its exact provider binding. Set T3_STATE_DB to an absolute regular file to use one custom state database instead. The helper does not scan other T3 stores. It keeps the T3 thread ID and database path under t3; format and sessionId still identify the provider transcript. Some bundled macOS SQLite builds cannot read a WAL-mode database when its sidecars are absent. On macOS, install upstream SQLite with brew install sqlite and prepend $(brew --prefix sqlite)/bin to PATH for the handoff command. A version number alone does not establish that a vendor build supports this case.

Query output is limited to 1 MiB; oversized mappings return invalid_t3_store. SQLite may create an empty -wal file and create or update the database's -shm WAL index and locking state. The helper leaves that shared bookkeeping to SQLite and never checkpoints, migrates, or writes stored database, WAL, or transcript content.

On a nonzero exit, read the JSON error and stop:

• not_found means report the ID and every searchedHomes entry.
• ambiguous_session means report every match and ask for a unique source.
• t3_store_not_found or t3_mapping_not_found means report the T3 thread and database path.
• t3_sqlite_unavailable means install the SQLite CLI and check PATH.
• t3_stale_binding means the mapped provider transcript is missing.
• t3_mapping_ambiguous or t3_ambiguous_binding means report the identifying details and ask for one unambiguous source.
• Any other error means report its code and identifying details.

The helper's result owns session identity. Do not replace it with substring search or pick a transcript that merely mentions the ID. Treat a Claude bridgeSessionId as provenance only. Native bridge and resume mechanisms do not provide a cross-agent handoff.

Build the continuation brief

Read the normalized document through the final event of every track. Process a large document in chronological chunks and carry one cumulative brief forward; skip no chunk. Tool calls show intent, while tool outputs and later messages show what happened. Prefer the later evidence when they disagree.

Build one private continuation brief with all of these fields:

• Source format, provider session ID, transcript paths, cwd, branch, and model.
• T3 Code thread ID and database path when the result contains t3.
• The user's goal and explicit constraints.
• Decisions already made and their recorded rationale.
• Files and other artifacts read, changed, created, or proposed, with their last known state.
• Commands, tests, and external operations already run, with outcomes.
• Failed approaches, unresolved risks, blockers, and open questions.
• Active skills, plugins, MCP tools, and subagent findings that affect the work.
• The last completed step and the single next action.

Separate transcript facts from your inferences. Resolve contradictions in favor of the latest direct result. Replace credentials and other secrets with [redacted] in anything you show or persist.

Reconcile live state and continue

Inspect the current cwd, branch, worktree status, and every artifact needed for the next action. The transcript is historical evidence, not current state. Name any mismatch in the brief and follow the repository's workflow before changing branches or files.

Tell the user the recovered goal, current state, and next action in a short capsule. Then continue from that action without asking the user to restate the session. Stop only for a real permission, product-decision, or safety boundary.

Remove the two temporary files after the brief is established. Leave the T3 state database, its sidecars, and every source transcript in place. Do not clean up SQLite sidecars. Continue from the derived artifact without a native resume command.