feat(orchestration-v2): import an existing Claude Code or Codex session by id, with two-way transcript sync

[Bil0000]

Reimplements #4617 against the orchestration V2 runtime, as requested in #4617 (comment) — plus a safe two-way sync between the imported thread and the provider's own on-disk transcript.

Why

T3 Code can only continue sessions it started. A Claude Code CLI session or a codex thread already on the machine is fully resumable, but there was no way to open it here. The V1 attempt (#4617) was closed because #2829 replaced the orchestration layer it was built on; this is the V2 reimplementation, targeting the #2829 branch.

What

Import session... in the command palette takes one session id.

• Resolve first. orchestration.resolveImportSession reads the session's working directory and title before anything is created. If a project covers that directory, the thread lands there regardless of which project is open; if not, the dialog relabels to Add project & import and nothing is created without a second press. Unlike V1, Codex sessions resolve too (workspace + title come from the rollout file).
• Import is all-or-nothing. The transcript is read before anything is written — Claude via the agent SDK (getSessionInfo/getSessionMessages), Codex from its rollout .jsonl under $CODEX_HOME/sessions (no live app-server needed). History is appended as synthetic message.updated/turn-item.updated pairs with deterministic ids — the same mechanism as the legacy V1 thread importer — under a new historyOrigin: "provider_import". There is no create-then-compensate dance: a failed import leaves nothing behind.
• The next turn resumes the real session. The import synthesizes a provider thread whose nativeThreadRef is the external session id (strength: "strong"; for Claude, nativeConversationHeadRef is set to the last transcript message so the adapter takes the resume path instead of colliding on sessionId). ProviderTurnStartService then resumes natively — the model keeps its full context; nothing is replayed.

Two-way sync

• T3 → provider comes free. Because the thread resumes the original native session, Claude appends every T3 turn to the original ~/.claude/projects/**/<id>.jsonl, and Codex to its own rollout.
• Provider → T3 is a lazy sync. On thread reads, ensureSynced re-reads the transcript (throttled by source mtime, serialized per thread, skipped while a run is active) and appends turns made outside T3. Idempotency and safety come from three dedupe layers: deterministic event ids, provider-native item ids already present as turn items, and exact role+text match against native run messages — so T3's own turns echoed back by the provider are never re-imported. Sync state lives in orchestration_v2_session_imports (migration 046).

Verification

• 8 new server tests: pure transcript-mapper tests plus an integration test that imports a real on-disk Codex rollout into an in-memory V2 store, checks the projection (history origin, messages, provider thread binding), refuses duplicates, and round-trips the provider→T3 sync.
• 6 new CommandPalette.logic tests (17/17 in the file).
• fmt, lint, typecheck clean across contracts, server, client-runtime, web (pre-existing @cursor/sdk-related failures unaffected).
• Live, against a real Claude Code session:
  1. claude -p "Remember this fact for later: the import canary code is FERRET-4413. Reply with exactly: OK" in a project directory.
  2. Palette → Import session... → paste id → thread appears with both prior messages rendered.
  3. Ask "What is the import canary code?" — the resumed session answers FERRET-4413. The canary only exists in the on-disk transcript, so the real session resumed rather than the history being replayed.
  4. The question and answer from step 3 appear in the original ~/.claude/projects/**/<id>.jsonl (T3 → provider).
  5. claude -p --resume <id> "…second canary is BADGER-7788…" in the terminal, reopen the thread in T3 — the CLI turn appears (provider → T3), and T3's own turns are not duplicated.

Demo

https://drive.google.com/file/d/16rBmArbhFZyHfzLWxNY7pxJNyqt7D9hs/view?usp=sharing

Known limits

• Only plain user/assistant text is imported Solved: tool activity now survives the import — Claude thinking blocks become reasoning items, Bash calls become command items paired with their results, Edit/Write calls become file-change items (other tools get a compact command rendering); same for Codex rollouts (reasoning summaries, function calls paired by call id). Verified live against a real session with tool use.
• Turns synced from the CLI render above native T3 runs Solved: sync-appended entries take position ordinals after the thread's current maximum, so CLI turns made after T3 turns render chronologically. Verified live (CLI turn after a native T3 turn lands at the bottom).
• $CODEX_HOME only Solved: rollout discovery resolves the codex instance's configured home from its resolved home layout first, then $CODEX_HOME, then ~/.codex.
• Codex live resume remains unexercised end-to-end (no Codex auth on the test machine); transcript parsing, resolve, import and sync are covered by the integration test against a real rollout file, and resume goes through the existing thread/resume adapter path.

Note

Add import and two-way sync of Claude Code and Codex sessions into orchestration-v2 threads

• Adds a full import workflow (SessionImportService) that reads an external Claude Code or Codex session, maps its transcript to deterministic OrchestrationV2 events, and stores it as a new thread with historyOrigin: 'provider_import'.
• Adds incremental sync (ensureSynced) that detects provider-side conversation added after import, backfills command outputs, and is called automatically on getThreadProjection and subscribeThread.
• Adds two new WS RPC endpoints (resolveImportSession, importSession) with auth scopes and client-runtime command functions, plus an ImportSessionDialog UI accessible from the command palette.
• Extends ProjectionStore ordinal logic so that provider-imported runless items are only included in history up to the correct run band when forking.
• Adds migration 046 creating the orchestration_v2_session_imports tracking table with a unique index on (driver, external_id).
• Risk: getThreadProjection and subscribeThread now perform a file-system or network read (ensureSynced) on every call for imported threads; failures are swallowed with a warning but add latency.

^{Macroscope summarized d265cee.}

[YoungPhlo]

Not exactly a helpful comment but I am so excited to see this land and can test on Win11/Linux/macOS for Codex and Claude Code if at all helpful, will otherwise stay out of the way - just really excited.

[juliusmarminge]

Closing as part of the open-PR backlog sweep (wave 1).

Reason: Abandoned draft XXL rebuild of closed #4617; import path moved on
Related: #8066 / #10386

Reopen if this is still wanted and you’re willing to rebase onto current main.