[roodboi]

Why

The HACK-1211 hosted experiment (PR #297) ran on Bun 1.3.9 and 1.4.2 with a 30-minute budget. Each run produced only 2 iterations: one focused pass (~16.5 s) and one full bun test that ran into the deadline (overrun, ~1783 s). The full suite was started on a 600 s estimate, but it takes about 30-35 minutes on hosted macOS.

What changes

Budget (scripts/bun-crash-loop.ts)

• An iteration starts only when its estimate ends at or before the deadline. Before a command completes a run, its estimate is a conservative prior: 120 s for focused, and now 2400 s for full (up from 600 s). After that, the estimate is the longest completed run times 1.2.
• A command that does not fit is recorded once in summary.json → skippedForBudget (with the estimate and the remaining time), and it leaves the rotation. Without a run its estimate cannot change, and the remaining budget only shrinks, so it could never fit again.
• The deadline stop stays as a safety net. It is still recorded as overrun and never counts as a crash unless the crash banner is in the log.

Modes

• --mode=focused|full|alternate, default focused. Focused mode runs only tests/native-https-owner.test.ts, the file where both original segfaults happened.
• --role=baseline|candidate (optional). It is used only in the verdict.

Workflow (.github/workflows/bun-crash-experiment.yml)

• New mode choice input for dispatch. The plan job validates it and rejects full with a budget_minutes under 40, because no run could start.
• The label-triggered matched pair uses focused, with roles baseline (1.3.9) and candidate (1.4.2). The matrix is now include entries of {bun_version, role}.
• The job timeout is still budget + 15.

Summary

• summary.json gains verdict, mode, role, counts per command (iterations, passes, crashes, failures (other failures), overruns, skippedForBudget) and skippedForBudget records. The step summary shows the same counts and the verdict.
• Example verdicts: baseline 0/97 focused crashes, Bun 1.4.2 0/40 focused crashes, 0/1 full crashes, candidate 0/0 full crashes; no iteration fit the budget.

Classifier (scripts/lib/bun-crash-classify.ts)

• Exit statuses 134 (SIGABRT) and 138 (SIGBUS) now count as crash exits, along with 133 and 139. Signal deaths were already crashes. The census shares this classifier, so a hosted job that exits with 134 or 138 now counts as a crash there too.

Test semantics, assertions and socket behavior of the tested suite are unchanged. The harness still runs plain bun test / bun test <file>.

Validation

• bun test tests/bun-crash-loop.test.ts tests/ci-bun-crash-census.test.ts tests/agent-instruction-source.test.ts: 52 pass. New tests use an injected clock and an injected executor. The executor stops at the deadline, so starting a run that cannot fit would show up as an overrun. The tests cover:
  • the hosted shape (30-minute budget, alternate mode, 16.5 s focused, 35-minute full): 108 focused runs, the full suite never starts, both skips recorded, no overrun;
  • estimates moving from the prior to measured durations;
  • focused mode running only the owner file;
  • full mode below the prior starting nothing;
  • injected crashes (banner, SIGSEGV, SIGBUS, exit 134) counted as crashes, and exits 1 and 2 counted as other failures;
  • counts and verdict text;
  • JSON and Markdown summaries staying value-free.
• Mutation check: when the start-fits-budget condition is removed, 8 tests fail.
• No test sends a crash signal or runs a real loop. The existing executeWithLog test still checks that no bun*.ips report appears.
• Pinned actionlint@1.7.12 on the CI workflow set: clean.
• I ran the plan step's shell locally for the label, dispatch focused, dispatch full at 30 minutes (rejected), dispatch full at 45 minutes and an invalid mode. I also checked how the loop step expands its arguments: --role is passed only when a role is set.
• bun run --cwd packages/cli typecheck, bun run --cwd packages/cli check, Biome on the changed files and bun run privacy:check: all pass.

Not done / gates

• I have not dispatched the experiment or labelled any PR. The new behavior is proven by unit tests and a local run of the plan shell, not by a hosted run. A real matched pair still needs an explicit go-ahead, because it spends about 45 hosted macOS minutes per Bun version.
• The workflow's full-mode budget floor (40 minutes) copies DEFAULT_FULL_ESTIMATE_SECONDS. If one changes, the other must change too (both places have a comment saying so).

🤖 Generated with Claude Code

---

^{Need help on this PR? Tag @codesmith-bot with what you need. Autofix is disabled.}