Skip to content

Stop /ship counting a reviewer's "nothing blocks" summary as a blocking finding (GST-2495) - #4301

Closed
garrytan wants to merge 3 commits into
mainfrom
ship-ledger-nothing-that-blocks
Closed

garrytan wants to merge 3 commits into
mainfrom
ship-ledger-nothing-that-blocks

Conversation

@garrytan

Copy link
Copy Markdown
Owner

/ship no longer turns an adversarial reviewer's "nothing blocks" summary into a blocking finding. This serves the shipping goal directly: a false blocker stops a crew's /ship receipt from going green when nothing in the review blocks.

What went wrong. On founderstar#926 the native adversarial reader opened its review with "I found nothing that blocks. There are three Low findings, all FIXABLE, and none can change outcomes, draws or save replay." The review ledger had a special case for this kind of verdict line, but it only recognised a few spellings of "nothing blocks", and only when the line named no severity besides FIXABLE. This line missed on both counts ("nothing that blocks", and "Low"). The general closing-summary reader refuses any line with FIXABLE in it, so the line fell through to the default rule, where the native reader's FIXABLE means CRITICAL. Lead judgment may not lower a CRITICAL row, so the receipt said blocking_found=1 with nothing blocking.

What changed (lib/review-ledger.ts, the verdict-line rule only):

  • The no-blocker clause now covers the common phrasings: "nothing that blocks", "nothing blocking", "none of which would block", "none of them is a blocker", "no finding that would block", "no blockers", "nothing here blocks", alongside the old ones.
  • A verdict line may now count its findings at a severity ("three Low findings", "one Medium and two Low findings"). It collects at the most severe count it states (Low here, High if it says High, CRITICAL if it says CRITICAL), never lower and never raised to CRITICAL by its FIXABLE tag.
  • On the relaxed review class the same line used to be raised to CRITICAL after collection. It now stays at its collected label there too; the existing v1.94.0.0 ship: CI is the one full test run (CI carry, tree cache, /ship beside CI) #3901 verdict line ("…all FIXABLE, and none blocks.") changes from CRITICAL to FIXABLE on that class for the same reason.

What stays strict. A line is still collected as CRITICAL when it also says something blocks ("I found one blocker", "One finding blocks the merge", a Blocks: naming a finding), adds an exception ("nothing that blocks except the token leak", "apart from…", a "but" after the denial), names a file and line, carries its own Severity or Blocks field, or uses FIXABLE outside a count or an "all/both/are FIXABLE" tag ("The cache key drops the tree, FIXABLE."). That last one was quietly accepted before when the line said "nothing blocks"; it's now a blocker, which is the fail-closed direction. Real findings under the summary parse exactly as before.

Evidence. The new tests in test/review-ledger-verdict-line.test.ts use the captured founderstar verdict text and fail on main (4 of 8 tests). Every captured reader fixture and pinned input the ledger corpus holds (test/fixtures/review-ledger/main-labels.json with its generated shapes, plus every reader output under test/fixtures/review-ledger, outside-review-severity and codex-review, each under four reader roles on both review classes: 3,874 parses, 4,818 rows, 1,666 CRITICAL) gives identical labels on main and on this branch.

Agent notes
  • Lane GST-2495 for GStack Minder (GSTA-36), routed by FOUN-32 from FOUN-298 (founderstar#926, gstack 1.99.1.0, head 78171ea37132). Freeze exception: blocks a crew's /ship.
  • Judge behaviour, severity policy and reviewer prompts are untouched.
  • Local checks at d4614e5: the 11 ledger and dispatch test files pass except one review-dispatch case that hits Bun's 5 s default timeout on this 4-core machine on main as well; it passes alone in 7.6 s with a longer timeout. bun run typecheck clean; typecheck:test reports no new diagnostics.

Open in Capy

…ng finding (GST-2495)

The adversarial reader's verdict line 'I found nothing that blocks. There are three Low findings, all FIXABLE' collected as CRITICAL because the no-blocker matcher missed 'nothing that blocks' and the line named a severity beside FIXABLE. The ledger now reads the common no-blocker phrasings and collects such a line at its counted severity on every review class, while exceptions, other blocking claims, loci, fields and stray FIXABLE tags still collect as CRITICAL.

Co-authored-by: capy-ai[bot] <230910855+capy-ai[bot]@users.noreply.github.com>
@trunk-io

trunk-io Bot commented Nov 11, 2026

Copy link
Copy Markdown

Merging to main in this repository is managed by Trunk.

  • To merge this pull request, check the box to the left or comment /trunk merge below.

After your PR is submitted to the merge queue, this comment will be automatically updated with its status. If the PR fails, failure details will also be posted here

garrytan and others added 2 commits November 11, 2026 22:45
…ad counts that start a sentence

A reader stuck repeating counts could produce a line long enough to stall the review ledger, so lines over 2,000 characters are no longer read as a verdict line and collect as before. Counts written with a capital at the start of a sentence ("Three Low findings, all FIXABLE, and nothing blocks.") are now read like mid-sentence ones.

Review-Finding: p1/red-team#1 reader=red-team
Review-Finding: p1/red-team#2 reader=red-team
Co-authored-by: capy-ai[bot] <230910855+capy-ai[bot]@users.noreply.github.com>
…on it is the verdict

Second review round on #4301: a 'nothing blocks' line that also named a defect, an exception, a hedge or a wrapped line above was collected at Low. The rule now requires every sentence to be the no-blocker clause, a count, a tag or a trailing denial, skips lines that continue the one above, and flags the summary at collection so the relaxed class no longer re-reads it with its joined continuation.

Co-authored-by: capy-ai[bot] <230910855+capy-ai[bot]@users.noreply.github.com>
@capy-ai

capy-ai Bot commented Nov 11, 2026

Copy link
Copy Markdown
Contributor

Parking this one. The fix taught the review ledger to recognise a reviewer's "nothing blocks" summary from its wording, and two rounds of review each found wordings where that rule let a reviewer's stated blocker through, which is the wrong failure for a merge check. Rather than patch the same wording heuristic again during the tooling freeze, we'll replace it after the freeze with a structured verdict: reviewers will state "no blockers" in a machine-readable field and the ledger will stop inferring it from prose. The failing cases and test lines are kept for that work.

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant