[yaacovcorcos]

Problem

A queue entry runs the full CI matrix (about 25–45 minutes, and one of the queue's two build slots) even when its PR's own CI just passed on exactly the same content. Separately, #582's main-push reuse refused whenever several queue entries landed in one push: #597 and #598 landed together, so #598's push ran full CI with "Qualification receipt differs".

Resulting behavior

• Queue reuse. A merge_group run reuses the PR's passing run only when that run tested the same content (tree) against the same base, its workflow ran from that tested merge, and the PR head already holds that content. In practice: the PR contained the latest main when its CI ran, nothing landed since, and nothing is queued ahead. The probe checks, via the API, that the queue commit merges exactly that PR head onto the group base and that the PR run's tested merge has the same tree and parents. It then republishes the run's seven reports.
• Why the PR head must hold the content. Then the merge changes nothing, so Compute recipe sealing's path-limited git log follows the PR head and never a merge commit, whose timestamp differs between the PR's test merge and the queue's.
• Required shard names. A matrix job skipped as a whole reports one unexpanded name, never Test Web 1–3 or Test Server 1–3. A reusing queue run therefore adds small jobs that report those required names.
• PR receipts. Same-repository PR runs record a qualification receipt (schema 2: event, tested merge, base, tree). Fork PRs never do.
• Multi-entry pushes. A main push reuses when each queue entry it lands qualified that exact commit from its first parent, back to the previous main, up to the queue's four-entry merge limit.
• Mobile gate. It now diffs the event's base against the tested commit with Git on every event, like the compute and packaging gates, instead of a PR's live file list. That list could change during a run, and the API capped it at 300 or 3,000 files. It matches raw bytes, so newline or non-UTF-8 filenames can't hide a native source.
• Fallback. Any missing, expired or mismatched evidence still runs normal CI.

Expected effect: of the last 27 merged PRs with a passing run on their final head, 6 met the queue condition (13 had an unchanged main, but only 6 of those also contained it). More PRs qualify when the latest main is merged into the branch before the final push.

Verification

• node --test .github/scripts/ci-queue-reuse.test.cjs: 84 pass. The tests cover:
  • the queue path, two-entry and four-entry pushes, and one past the merge limit;
  • every mismatch and failure path refusing reuse without partial reports;
  • PR and queue receipts;
  • the workflow structure, including the stand-in shard names;
  • a behavioral test that runs the mobile gate's own script against commits built with Git plumbing.
• Mutation checks: removing each of nine guards fails at least one test. Reverting the gate's newline mapping, byte matching or --no-renames each fails the behavioral test.
• Real objects: PR test(server): keep server tests out of the source checkout #601's test merge has parents [main at test time, PR head]. Queue commit c145ae778f has parents [queue base, fix(citations): build each citation style's engine once #598 head].
• Gate script on real commits: a ci.yml change and a Swift change report changed=true; a docs-only commit reports changed=false; a native rename reports true with both paths; a bad base fails open.
• vp fmt --check, vp lint on the scripts, YAML parse, and git diff --check pass.
• Not yet observed on hosted CI: a reusing queue run. This PR's own run will be the first to write a PR receipt, which checks that GITHUB_WORKFLOW_SHA equals the tested merge on pull_request. If GitHub reports another value, reuse simply never applies.

Independent review: Codex gpt-6.1-sol, five rounds. Round 1 found the mobile-gate and recipe-timestamp gaps, round 2 the live PR file list and the workflow binding, round 3 newline filenames, and round 4 non-UTF-8 filenames. Each was fixed in this branch. Round 5 approved, with no remaining P0–P2.

Documentation impact

Updated — docs/internals/ci.md (reuse rules, required shard names, mobile gate comparison).

Model and harness: Claude Opus 5.5 in Claude Code (Scient), with Codex gpt-6.1-sol as independent reviewer.

🤖 Generated with Claude Code

[github-actions]

Thread transfer impact

✅ Thread transfer remains within every enforced ceiling.

Provider	Metric	Main baseline	This PR	Impact	PR ceiling
Codex	Total thread wire	5.6 KiB	5.6 KiB	0 B (0.0%)	6.8 KiB	✅
Codex	Thread snapshot wire	4.3 KiB	4.3 KiB	0 B (0.0%)	4.9 KiB	✅
Codex	Live turn WebSocket wire	1.3 KiB	1.3 KiB	0 B (0.0%)	2.0 KiB	✅
Codex	Live turn WebSocket decoded	26.2 KiB	26.2 KiB	0 B (0.0%)	29.3 KiB	✅
Codex	Live turn messages	2	2	0 (0.0%)	8	✅
Claude	Total thread wire	5.6 KiB	5.6 KiB	−36 B (−0.6%)	6.8 KiB	✅
Claude	Thread snapshot wire	4.3 KiB	4.3 KiB	0 B (0.0%)	4.9 KiB	✅
Claude	Live turn WebSocket wire	1.3 KiB	1.3 KiB	−36 B (−2.6%)	2.0 KiB	✅
Claude	Live turn WebSocket decoded	26.6 KiB	26.5 KiB	−41 B (−0.2%)	29.3 KiB	✅
Claude	Live turn messages	2	1	−1 (−50.0%)	8	✅

Baseline: 129f4af · PR result: 6e16191 · Source CI: success

Scenario and decoded snapshot size

10 historical turns, 5 command tools per turn, 878.9 KiB retained MCP result per historical turn, and a 1.05 MiB retained result in the measured turn.
Full compact HTTP snapshot with live cursor.

• Codex decoded thread snapshot: 133.1 KiB
• Claude decoded thread snapshot: 133.4 KiB

Updated in place by a trusted workflow. PR artifacts are strictly validated and never executed.