# Pi 0.85.1 Native Runner Release Validation Date: 2026-09-13 Status: implementation round validated; not published ## Scope And Environment Validation used Pi 0.85.1 with Node 24.14.1 in a disposable project. The working tree's `pi-personas` 0.2.1 package was created with `npm pack`, installed from that tarball, and loaded from the installed package directory. Global Pi settings and global packages were not changed. Authentication and model catalog files were copied into a disposable Pi config directory. Only provider names, credential types, model metadata, and readiness were inspected; no credential values were printed. Cerebras authentication reported `ready`. Provider-backed runs used the configured, affordable `cerebras/qwen-3.8-27b` model, with `cerebras/gpt-oss-120b` used for the fork and cancellation checks when stronger tool-call compliance was needed. Thinking was disabled. The legacy checks used an isolated install of `pi-subagents` 0.67.0. Native was selected only through the disposable project's `.pi/persona.json`; removing that file restored the `legacy` backend. These live runs preceded the automatic backend-selection change: the release now chooses legacy when `pi-subagents` is installed and native otherwise, unless explicitly overridden. The installed legacy fixture therefore retains the same selection under the new policy. The automatic-selection change was subsequently verified with 140 unit tests and one RPC smoke test; the live provider matrix was not repeated for that change. ## Executed Scenarios | Scenario | Actual outcome | | --- | --- | | Packed native fresh consult | Passed. One `persona_consult` call launched a native child from the packed install. The child used `read` and returned the exact fixture line `PACKED_NATIVE_FIXTURE_EVIDENCE`. Details reported backend `native`, context `fresh`, and model `cerebras/qwen-3.8-27b`; nested usage was present. | | Packed native fork consult | Passed with `gpt-oss-120b`. One `persona_consult` call used context `fork`. Its tool arguments did not contain `FORK_BRANCH_7Q9M`, while the child returned that marker from the active branch. The successful child session and the focused snapshot test cover removal of the in-flight tool call. | | Packed native round-table | Passed. The primary generalist made exactly one `persona_roundtable` call, selected `operator` and `reviewer`, and completed Round 1, Round 2, and moderator synthesis. Process details reported 5/5 steps, zero failed steps, and nested usage. Two recoverable child read errors were reported during model-led probing; synthesis still completed normally. | | Native cancellation during execution | Passed. RPC abort was sent 750 ms after `persona_consult` tool execution began. The active call ended with `Native Pi Persona child was cancelled.`, an error stop reason, no extension error, and no stderr. | | Native model preflight failure | Passed. A disposable persona model `unavailable/no-such-model` was rejected before launch with `found 0`; no fallback backend ran. | | Default legacy consult, `pi-subagents` 0.67.0 | Initially failed, then passed after the release-blocker fix. Final run returned `LEGACY_OK` as a foreground final child result, not an async receipt. Backend details remained `legacy`. | | Default legacy round-table, `pi-subagents` 0.67.0 | Passed after the fix. One `persona_roundtable` call completed the two specialist rounds and synthesis through a foreground `workflowScript`; process details reported 5/5 steps and no failures. No receipt-triggered extra assistant turn appeared. | ## Release-Blocker Fixed Current `pi-subagents` rejects the old public `clarify` field, defaults direct children to asynchronous execution, and no longer accepts public top-level `chain` orchestration. This caused a legacy consult to fail first with `Public workflowScript execution does not support clarify UI`; after removing that field, it returned a detached-run receipt instead of the consultant answer. The correction is limited to the shared legacy request boundary: - consult requests omit `clarify` and set `async: false`; - round-tables retain the existing chain payload for `pi-subagents` 0.34.0-0.40.x and use an equivalent foreground `workflowScript` for 0.41.0+; - one focused regression test checks both payload generations. ## Commands Run Paths below use placeholders for the disposable directory and packed artifact; no credential-bearing path contents or values are included. ```sh pi --version PI_CODING_AGENT_DIR= pi --list-models PI_CODING_AGENT_DIR= pi auth check --provider cerebras --no-refresh npm pack --json --pack-destination npm install --ignore-scripts --no-audit --no-fund /pi-personas-0.2.1.tgz PI_CODING_AGENT_DIR= node live-rpc.mjs fresh PI_CODING_AGENT_DIR= LIVE_MODEL=gpt-oss-120b node live-rpc.mjs fork PI_CODING_AGENT_DIR= node live-rpc.mjs roundtable PI_CODING_AGENT_DIR= LIVE_MODEL=gpt-oss-120b node live-rpc.mjs cancel PI_CODING_AGENT_DIR= LIVE_MODEL=gpt-oss-120b node live-rpc.mjs preflight PI_CODING_AGENT_DIR= node live-rpc.mjs legacy PI_CODING_AGENT_DIR= node live-rpc.mjs legacy-roundtable npm test npm audit --omit=dev npm pack --dry-run git diff --check ``` ## Automated Gate Results - `npm test`: passed; 135 unit tests and one real offline Pi RPC smoke test. - `npm audit --omit=dev`: passed; zero vulnerabilities. - `npm pack --dry-run --json`: passed; 30 files, 64,173-byte tarball, SHA-1 `73546b39b89a56de96ebcd93193ef441de2e5738`. - `git diff --check`: passed with no whitespace errors. ## Limitations And Remaining Gates - The live legacy run covered `pi-subagents` 0.67.0. The 0.34.0 package and compatibility contract were inspected locally, and the 0.34 chain payload is regression-tested, but 0.34.0 was not exercised against a live provider. - One exploratory fork attempt with `qwen-3.8-27b` answered the remembered marker directly instead of calling the requested tool. It did not exercise Pi Persona and was repeated successfully with `gpt-oss-120b`. - Native children remain capability-limited rather than OS-sandboxed, as the product documentation states. - The release checklist's clean-checkout setup was not run because this round explicitly preserved the approved dirty worktree. Final validation ran against the packed contents of that worktree. - Publishing, versioning, committing, tagging, and pushing remain explicit maintainer gates and were not performed.