# fix\(pi\): fork completed turns while the next turn is running · PR \#23052 · pingdotgg/t3code [View on GitCafe](https://git.cafe/pingdotgg/t3code/pull/23052) Repository: [pingdotgg/t3code](https://git.cafe/pingdotgg/t3code) Visibility: public State: merged Source repository: pingdotgg/t3code Source branch: fix/issue-23050 Target branch: main Head commit: 1372a7af05a99986ff154c41ec3d6c83b21f51c3 [Discussion and reviews](https://git.cafe/pingdotgg/t3code/pull/23052) [Commits](https://git.cafe/pingdotgg/t3code/pull/23052/commits?format=markdown) [Files changed](https://git.cafe/pingdotgg/t3code/pull/23052/diffs?format=markdown) ## Description --- ## Problem Forking a completed Pi turn while the next turn is still running fails with “Pi turn 2 has no captured session-tree entry and none was found in the session tree.” The running turn has a synthetic weak reference and no completion timestamp, although Pi has already persisted its prompt. ## Change Allow the existing session-tree recovery window to remain open at its upper end for a running boundary turn. Keep the retained turn’s lower bound, ancestry validation, and earliest-user-entry selection, so the fork cuts before the running turn’s prompt rather than a later steer. Completed boundary turns keep their finite upper bound. Add one adapter regression for the running weak-reference boundary. This is a server-only Pi adapter change; no client, wire contract, or other provider changes. ## Scope and approval Fixes #23050 Triaged bug via the automated issue flow. Focused repair of the existing fork behavior, following [#20521](https://git.cafe/pingdotgg/t3code/pulls/20521); no authentication, credential, permission, approval, mode, sandbox, or MCP changes. ## Verification - Test-first: the running-boundary regression fails on main with the exact reported protocol error; the existing three native-fork tests pass. Independent proof audit reruns red and green. - `./node_modules/.bin/vp test run packages/provider-pi/src/server/adapter.test.ts`: 84/84 pass on the final head, including the existing completed-uncaptured-boundary cases. - Scoped Pi typecheck, changed-file lint, formatting, whitespace checks, and `knip --workspace packages/provider-pi --no-config-hints` pass. Lint reports one existing unused `layer` warning outside this change. - Temporary fake-Pi runtime driver: before fails with the exact protocol error; after returns `/fake/forked.jsonl` and sends native fork at `u2`, excluding turn 2 and its later steer. Evidence is from the adapter’s fake Pi protocol harness, not native session persistence. - Codex GPT-6.1 Sol and Claude Opus 5.5 independent reviews: MERGE, no confirmed bugs. Removed their nonblocking fixture-only ancestry assertion. - UI evidence matrix is empty. No visual changes; no screenshots or recordings required. [23050-runtime-proof.txt](https://gh-file-drop-api-prod-mi5fy3sowv63ufte.pinglabs.workers.dev/f/be53774a16dfccf0/23050-runtime-proof.txt) [23050-proof-matrix.txt](https://gh-file-drop-api-prod-mi5fy3sowv63ufte.pinglabs.workers.dev/f/f806ac5eab7bc328/23050-proof-matrix.txt) Pi 1.3.0 installed without model credentials. Attempted the isolated worktree web flow, but repeated dev watcher restarts on source/test-cache changes and reconnect failures prevented completing the live two-turn sequence. Live Pi, authenticated model behavior, Windows, Electron desktop, and mobile remain unverified. No live T3 state was written. Main was fetched and merged locally; it was already up to date. Implementation and regression test: Claude Opus 5.5. Triage, integrated verification, proof audit, and review: Codex GPT-6.1 Sol and Claude Opus 5.5 through T3 Code (Codex and Claude Agent harnesses).