[shafty023]

A vibe session ends one of two ways now: a pull request the person takes through review and merge, or passing it on (operator ruling, 2026-10-15). In a real product-manager session, handoff asked "Who should pick this up next?" twice. She said "i will get the pr merged", and the session answered with "the ticket and the branch to merge". It never started the requested pull request path, wrote no tests, never pushed and never assigned the ticket. This fixes each of those.

• Plain-language PR intent. "raise a PR", "open a pull request", "I'll get the PR merged", "I'll get it merged", "I'll ship it myself" and "I want to take it through review" are an explicit request and take handoff's "A requested pull request" path. The canonical statement is in guardrails.md ("Tests"); the vibe and handoff skills, quality-loop.md and the change worker point to it. Tests are still written only on that request. Nothing offers the person "the branch to merge".
• The PR actually opens. Once requested, the same writer writes the tests in its pull-request turn, and step 7 commits, pushes and publishes. As handoff's last step, a new pull-request mode in vibe-environment-worker opens the PR on the pushed branch with gh pr create. It fills symphony-alpha's current template from the session record, with an honest Breaking changes answer and a UI flag attestation only from the flag record. Anything the record can't support is noted for the reviewer. It never merges, enqueues, requests reviewers, triggers CI or adds warning-acknowledged. The ticket stays with the person, In Progress, and they get the PR link. Every brief after the request carries the pull-request phase, and no brief may say both "PR" and "no tests". The old "handoff opens none" wording is gone.
• Two-way question. "Who should pick this up next?" is gone. When the person's words don't settle it, handoff asks: "Do you want me to open a pull request so you can take it through review and merge, or pass it to someone, usually design and then engineering, to finish?" It asks this at the start (new step 0, "Settle how it ends") because a PR changes what runs. Only passing it on asks for a name or email. A confused reply gets both endings explained, never a name prompt. The vibe skill's section 8 asks the same question and keeps "keep it" and "throw it away".
• Early notice. A new session's first reply carries "When you're done, I can either open a pull request for you, or pass it to design or engineering to finish." It's a non-blocking notice and never appears on resume.
• No unexplained "handoff" in text the person sees. $handoff and its triggers are unchanged. The skill and guardrails.md no longer assume an executive who always passes work on.
• Both harnesses. The prose is shared. The new pull-request record action is declared once in tools/vibe-workers/src/record-context.ts and read by both the Claude launcher (claude-worker.mjs) and the Codex native grant (writer-state.mjs acquire-record). The dist bundles are rebuilt.

Tests:

• New test_vibe_session_endings_contract.py: 10 cases, all 10 fail on 1.0.38 and pass here.
• test_handoff_skill_contract.py and test_vibe_handoff_docs_tests_control_contract.py pinned the old question and "handoff opens none"; they now pin the new wording.
• New vitest cases: the Claude launcher admits the environment worker's pull-request action with no record writes; the Codex native record CLI grants it to that worker and refuses it to the ticket worker. Both fail on 1.0.38.
• New plugins/vibe/evals/ suite (15 cases for claude plugin eval, executive and product-manager personas, including the real exchange with a neutral name). The README says how to run it. It is not wired into CI. Scores are in a comment below.

Bumps vibe to 1.0.39 (both manifests) with a CHANGELOG entry.

[shafty023]

Eval results (claude plugin eval, 15 cases, 1 run each, sonnet agent, sonnet judge, -j 2):

Arm	Cases passed	Mean score
This branch (vibe 1.0.39)	14/15, then 15/15 after the grader fix below	0.99
No-plugin baseline arm (same run)	6/15	0.70
vibe 1.0.38 (main, --ablation none)	5/15	0.70

• 1.0.38 fails the PR-intent cases ("I'll get the PR merged", "I'll ship it myself", "I want to take it through review": no requested pull request route), both "I'm done" cases (it asks "Hand this off now, keep it to come back to, or throw it away?"), the confused-reply case, both new-session notice cases and "someone else can finish". On the real exchange it answered "I'll look you up in ClosedLoop now. I don't open the pull request myself, so the ticket and the branch will be yours", the same failure as the real session. It scored 0.4 on 3 of 3 runs.
• The one branch miss was the judge failing a correct reply ("I'll open a pull request for you... The ticket stays with you. When the pull request is open, I'll send you its link, along with the ticket and app links"). The grader now allows sending the PR link alongside the ticket link. On a rerun of that case with 3 runs, the branch passed 3/3 and 1.0.38 scored 0.4 on 3/3.
• "can you raise a PR", "give it to ", the resume case and the PR-link closing case also pass on 1.0.38. The first two are plain enough that the old skill gets them right; the last two guard against regressions.
• Cost: about $2.30 for the branch with its baseline arm, about $1.90 for 1.0.38.

claude plugin eval exercises Claude Code only. Codex coverage is the shared contract tests plus the Codex native record test for the new pull-request action.