[lucasbraum]

UPT-111. On dms-saas, the known MongoMemoryServer flake (private-admission.test.ts) failed the end-of-session check in all 4 SDK sessions, so every edit was discarded. The #237 pilot got 0/51 coverage, 10 type errors, and spent $1.02 on a $1 cap.

What changes

• The session gate reruns failing test files once before it rejects. When a session's tests fail with failures that aren't preexisting, the gate takes the test files the runner named and drops the ones that already failed at the parent or at the session's start. It runs the rest once more, alone (sdk-agent.ts, judge).
  • Every rerun file passes: the files are recorded as llm.tasks[].flaky in report.json and the edits stay.
  • Any file fails again: the session is rejected, as before.
  • A failing run that names no test file (a build step, a timeout), or a rerun that runs nothing: also rejected, as before. No run is no pass.
• How a file runs alone (TestOptions.only in planTests): with its runner directly, vitest run <files> or jest --runTestsByPath <files>, with the same unit-only isolation as the full run. It never runs a test script, a fallback, a pretest (the first run already built this tree) or the existing unrelated-file retry. A file no runner can take alone (a service test without --with-services, or no runner at all) gets a missing plan, so it can't count as a pass.
• Final verification keeps its full gate. It still runs the full suite, so a flaky file that fails there fails verification. The PR body's review list gets a line naming the flaky files. It says whether verification's full run passed them or failed them again ("rerun before merging"). A failing verify never turns into a pass.
• No session the budget can't cover. A verification feedback round is no longer started when what's left of maxCost is below the mean cost of the run's sessions so far. llm.feedback.stopped records the remaining and typical amounts, and the PR's feedback line says so. In feat(fix): fetch the Agent SDK on the first fix, never fall back silently (UPT-108) #237 the fourth session started with $0.09 left after three sessions at about $0.30 each, crossed the cap, and stopped at error_max_budget_usd. It would not start now.
• Docs: docs/commands/fix.md. Changeset: patch.

Tests

• sdk-agent.test.ts, session gate:
  • A file that fails once and passes alone keeps the edits and the commit, and is recorded as flaky. Only the new file is rerun, not the parent's own failure.
  • A file that fails twice is rejected and reverted.
  • A rerun that ran nothing is rejected, and so is a failure that names no file.
• feedback.test.ts, end to end through fix:
  • The flaky session is kept, and report.json and the PR body name the file. If verification's full run fails that file again, verification fails and the body says so; otherwise it passes.
  • Budget stop: at $0.10 a session against a $0.25 cap, the third session is not started, and spend stays at or under the cap. Two existing tests that expected the old overshoot (3 rounds, $0.30 on $0.25) now expect 2 rounds and $0.20.
• services.test.ts: the exact rerun commands for vitest (unit-only config, and with services) and for jest. A test script that builds first is skipped, and an e2e file is reported as missing.

pnpm turbo run typecheck test for uptide (517) and @uptide/core (2305) passes. biome check . shows only the pre-existing warning and info.

dms-saas rerun

Same harness as #237: #169's bench as the pilot ran it, the packed tarball through plain npx -y file:<tgz>, the Uptide arm, SDK executor, claude-sonnet-5-5, and the $1 cap (--max-cost default). One run. The human reference is round 3's.

	executor	local checks (install / typecheck / lint / test / build)	flaky files	sessions	cost	wall	fork CI	coverage vs human
this PR	sdk	ok / ok / ok / ok / ok	none	1, accepted; 0 feedback rounds	$0.27	234 s	1 pass, 0 fail (fork PR)	10/51, 0 unrelated, no regressions
#237 pilot	sdk	ok / 10 errors / ok / 1 fail / fail	(not detected)	4: 3 rejected on tests, 1 stopped at the budget	$1.02	612 s	fail at typecheck	0/51

• The SDK session fixed 10 sites and the rules fixed 2. Uptide's verification passed. Its 3 failing test files (Cannot find package 'vue' from the nested frontend-vue project) are marked preexisting: they fail the same way at the parent.
• The flake did not fire on this run. The session passed its gate on the first test run, so no rerun happened and nothing was recorded as flaky. This run shows the end result on dms-saas; the rerun path itself is covered by the fixtures above, not by this pilot. Proving it on the real flake needs either more runs or a forced timeout in private-admission.test.ts.
• The budget stop didn't apply either: one session at $0.27, with no feedback round needed.

🤖 Generated with Claude Code