Problem
Stop retrying blind into the same crash — the Avenger workflow's Claude Code CLI process segfaults in Bun on every single attempt, so all 4 built-in retries burn ~90s per run for zero benefit. This is a 100% failure rate for the workflow in this window (2/2 runs, 8/8 attempts).
Affected workflows and runs
- Workflow:
Avenger (.github/workflows/avenger.lock.yml)
- Failed runs (both 100% attempt failure): §31465993823 (2026-08-11 06:41 UTC), §31462571906 (2026-08-11 05:42 UTC)
- Last known success (comparator): §31459752206 (2026-08-11 04:50 UTC)
Probable root cause
Bun 1.4.0 crashes with Segmentation fault (SIGSEGV) inside the Claude Code CLI process, at a different memory address on every attempt (0x12, 0x21, 0x0, 0xFFFE000000000015). That variability, combined with Bun's own panic message ("This indicates a bug in Bun, not your code"), points to heap/GC corruption in the Bun runtime rather than a deterministic input-triggered bug in the workflow or prompt.
audit-diff against the last successful run rules out a sandbox/network cause: the failed run completed 43 real api.anthropic.com calls and 19 GitHub API calls before crashing, with no anomalous firewall/domain deltas versus the successful run. The crash happens mid-session, after genuine model traffic — this is an in-process Bun fault, not a policy or connectivity issue.
A second, compounding bug lives in the harness itself: in §31462571906, attempt 2's --continue retry failed with "no deferred tool marker," and the harness responded by permanently disabling --continue for the rest of the run (failure_reason=harness_retry_path_invalid), forcing attempts 3 and 4 into full fresh-context restarts — which then also segfaulted, discarding all prior conversation progress on every retry.
Proposed remediation
- Pin the Bun version used by the Claude Code CLI harness to a known-stable release, or bisect against recent Bun updates to find the regression — do not float latest.
- Add segfault-aware fast-fail: if attempt 1 exits with
signal=SIGSEGV, cap retries at 1 instead of exhausting all 4 — a repeat segfault on retry 1 is extremely unlikely to succeed on retries 2-4 and just wastes runtime.
- File the captured
bun.report crash links upstream to Bun's team (three distinct crash reports are already in the run logs).
- Fix the
harness_retry_path_invalid fallback so a failed --continue doesn't silently degrade into full fresh-run retries — surface it as its own diagnostic signal so it doesn't get conflated with normal retry behavior.
Success criteria / verification
- Avenger completes successfully, or fails with a distinct non-segfault error, on 3 consecutive scheduled runs after the fix.
- Zero
signal=SIGSEGV entries in Claude Code CLI harness logs for Avenger over the following 7 days.
- If the Bun bug isn't resolved upstream in that window, confirm the harness now fails fast (≤1 retry) on a confirmed segfault instead of exhausting all 4 attempts.
Parent report: #51983
Related to #51983
Generated by 🔍 [aw] Failure Investigator (6h) · agent · 123.6 AIC · ⌖ 41.5 AIC · ⊞ 5.3K · ◷
Problem
Stop retrying blind into the same crash — the Avenger workflow's Claude Code CLI process segfaults in Bun on every single attempt, so all 4 built-in retries burn ~90s per run for zero benefit. This is a 100% failure rate for the workflow in this window (2/2 runs, 8/8 attempts).
Affected workflows and runs
- Workflow:
Avenger (.github/workflows/avenger.lock.yml)
- Failed runs (both 100% attempt failure): §31465993823 (2026-08-11 06:41 UTC), §31462571906 (2026-08-11 05:42 UTC)
- Last known success (comparator): §31459752206 (2026-08-11 04:50 UTC)
- New (2026-08-13):
Claude Code User Documentation Review (.github/workflows/claude-code-user-docs-review.lock.yml) — §31700716464 (2026-08-13 12:34 UTC)
Probable root cause
Bun 1.4.0 crashes with Segmentation fault (SIGSEGV) inside the Claude Code CLI process, at a different memory address on every attempt (0x12, 0x21, 0x0, 0xFFFE000000000015, and now 0xFFFE000000000068). That variability, combined with Bun's own panic message ("This indicates a bug in Bun, not your code"), points to heap/GC corruption in the Bun runtime rather than a deterministic input-triggered bug in the workflow or prompt.
audit-diff against the last successful run rules out a sandbox/network cause: the failed run completed 43 real api.anthropic.com calls and 19 GitHub API calls before crashing, with no anomalous firewall/domain deltas versus the successful run. The crash happens mid-session, after genuine model traffic — this is an in-process Bun fault, not a policy or connectivity issue.
A second, compounding bug lives in the harness itself: in §31462571906, attempt 2's --continue retry failed with "no deferred tool marker," and the harness responded by permanently disabling --continue for the rest of the run (failure_reason=harness_retry_path_invalid), forcing attempts 3 and 4 into full fresh-context restarts — which then also segfaulted, discarding all prior conversation progress on every retry.
Proposed remediation
- Pin the Bun version used by the Claude Code CLI harness to a known-stable release, or bisect against recent Bun updates to find the regression — do not float latest.
- Add segfault-aware fast-fail: if attempt 1 exits with
signal=SIGSEGV, cap retries at 1 instead of exhausting all 4 — a repeat segfault on retry 1 is extremely unlikely to succeed on retries 2-4 and just wastes runtime.
- File the captured
bun.report crash links upstream to Bun's team (three distinct crash reports are already in the run logs).
- Fix the
harness_retry_path_invalid fallback so a failed --continue doesn't silently degrade into full fresh-run retries — surface it as its own diagnostic signal so it doesn't get conflated with normal retry behavior.
Success criteria / verification
- Avenger completes successfully, or fails with a distinct non-segfault error, on 3 consecutive scheduled runs after the fix.
- Zero
signal=SIGSEGV entries in Claude Code CLI harness logs for Avenger over the following 7 days.
- If the Bun bug isn't resolved upstream in that window, confirm the harness now fails fast (≤1 retry) on a confirmed segfault instead of exhausting all 4 attempts.
Update — 2026-08-13 12:38 UTC
Still not fixed, and now spreading to a second workflow. Claude Code User Documentation Review hit the identical signature in run §31700716464, job 94449125253: panic(main thread): Segmentation fault at address 0xFFFE000000000068, oh no: Bun has crashed, signal=SIGSEGV, all 3 harness retries exhausted, total duration 1m 19s.
audit-diff vs. the last successful run of this workflow (§31597056301) again shows no firewall/domain anomalies — but call volume collapsed: api.anthropic.com dropped 170% (37 vs. 100 allowed calls) and Sentry/OTLP ingest dropped 329% (7 vs. 30). That's consistent with the process dying ~15s into a normal session, before it could complete its usual call sequence — further confirming this is an in-process Bun runtime fault, not a network/firewall regression. Raise this to P0-critical: the crash is no longer isolated to one workflow.
Parent report: #51983
Related to #51983
Generated by 🔍 [aw] Failure Investigator (6h) · agent · 123.6 AIC · ⌖ 41.5 AIC · ⊞ 5.3K · ◷
Update — 2026-08-15 (6h window)
Widen the fix scope — this is now a 3rd workflow, not 2. Fix the Bun runtime crash before it spreads further; each new hit costs a full retry budget for zero output.
New occurrence: §31855680444 — [aw] Failure Investigator (6h) — identical Bun 1.4.0 Illegal instruction panic (bun.report/1.4.0/...) under the Claude engine, matching this issue's signature exactly. Auto-filed as #52807, now closed as duplicate of this issue.
No new root-cause information beyond what's already documented above — just confirming the blast radius is growing (Avenger x2 + Failure Investigator x1 = 3 workflows affected).> Generated by 🔍 [aw] Failure Investigator (6h) · agent · 210.5 AIC · ⌖ 9.45 AIC · ⊞ 5.6K · ◷
Update — 2026-08-16 (6h window)
Still not fixed — 2 more Avenger occurrences today. Same signature, exit code 139 (SIGSEGV), no other error detail in this window's logs:
- §31931891565 — Avenger, 06:38 UTC, step "Execute Claude Code CLI"
- §31929386213 — Avenger, 05:36 UTC, step "Execute Claude Code CLI"
No new root-cause detail beyond what's already documented — confirming Avenger is still 100% blocked by this crash. Fix the Bun pin / fast-fail-on-SIGSEGV remediation above before the next scheduled Avenger run to stop burning the full retry budget for zero output.> Generated by 🔍 [aw] Failure Investigator (6h) · agent · 121.4 AIC · ⌖ 6.57 AIC · ⊞ 5.6K · ◷
Update — 2026-08-16 22:37 UTC (new occurrence)
Still not fixed — a new Avenger occurrence hit in this 6h window, exit code 139 again.
- New occurrence: §31976674745 (2026-08-16 22:33 UTC), step
Execute Claude Code CLI, ##[error]Process completed with exit code 139. — identical signature to every prior hit in this issue.
audit confirms genuine mid-session traffic before the crash (67 api.anthropic.com calls, 0 blocked firewall requests) — same in-process Bun fault pattern, not a network/policy regression.
No new root-cause detail — just confirming Avenger is still fully blocked. The Bun-pin / fast-fail-on-SIGSEGV remediation already proposed above remains the fix.> Generated by 🔍 [aw] Failure Investigator (6h) · agent · 237.9 AIC · ⌖ 8.49 AIC · ⊞ 5.6K · ◷
Problem
Stop retrying blind into the same crash — the Avenger workflow's Claude Code CLI process segfaults in Bun on every single attempt, so all 4 built-in retries burn ~90s per run for zero benefit. This is a 100% failure rate for the workflow in this window (2/2 runs, 8/8 attempts).
Affected workflows and runs
Avenger(.github/workflows/avenger.lock.yml)Probable root cause
Bun 1.4.0 crashes with
Segmentation fault(SIGSEGV) inside the Claude Code CLI process, at a different memory address on every attempt (0x12,0x21,0x0,0xFFFE000000000015). That variability, combined with Bun's own panic message ("This indicates a bug in Bun, not your code"), points to heap/GC corruption in the Bun runtime rather than a deterministic input-triggered bug in the workflow or prompt.audit-diffagainst the last successful run rules out a sandbox/network cause: the failed run completed 43 realapi.anthropic.comcalls and 19 GitHub API calls before crashing, with no anomalous firewall/domain deltas versus the successful run. The crash happens mid-session, after genuine model traffic — this is an in-process Bun fault, not a policy or connectivity issue.A second, compounding bug lives in the harness itself: in §31462571906, attempt 2's
--continueretry failed with "no deferred tool marker," and the harness responded by permanently disabling--continuefor the rest of the run (failure_reason=harness_retry_path_invalid), forcing attempts 3 and 4 into full fresh-context restarts — which then also segfaulted, discarding all prior conversation progress on every retry.Proposed remediation
signal=SIGSEGV, cap retries at 1 instead of exhausting all 4 — a repeat segfault on retry 1 is extremely unlikely to succeed on retries 2-4 and just wastes runtime.bun.reportcrash links upstream to Bun's team (three distinct crash reports are already in the run logs).harness_retry_path_invalidfallback so a failed--continuedoesn't silently degrade into full fresh-run retries — surface it as its own diagnostic signal so it doesn't get conflated with normal retry behavior.Success criteria / verification
signal=SIGSEGVentries in Claude Code CLI harness logs for Avenger over the following 7 days.Parent report: #51983
Related to #51983
Problem
Stop retrying blind into the same crash — the Avenger workflow's Claude Code CLI process segfaults in Bun on every single attempt, so all 4 built-in retries burn ~90s per run for zero benefit. This is a 100% failure rate for the workflow in this window (2/2 runs, 8/8 attempts).
Affected workflows and runs
Avenger(.github/workflows/avenger.lock.yml)Claude Code User Documentation Review(.github/workflows/claude-code-user-docs-review.lock.yml) — §31700716464 (2026-08-13 12:34 UTC)Probable root cause
Bun 1.4.0 crashes with
Segmentation fault(SIGSEGV) inside the Claude Code CLI process, at a different memory address on every attempt (0x12,0x21,0x0,0xFFFE000000000015, and now0xFFFE000000000068). That variability, combined with Bun's own panic message ("This indicates a bug in Bun, not your code"), points to heap/GC corruption in the Bun runtime rather than a deterministic input-triggered bug in the workflow or prompt.audit-diffagainst the last successful run rules out a sandbox/network cause: the failed run completed 43 realapi.anthropic.comcalls and 19 GitHub API calls before crashing, with no anomalous firewall/domain deltas versus the successful run. The crash happens mid-session, after genuine model traffic — this is an in-process Bun fault, not a policy or connectivity issue.A second, compounding bug lives in the harness itself: in §31462571906, attempt 2's
--continueretry failed with "no deferred tool marker," and the harness responded by permanently disabling--continuefor the rest of the run (failure_reason=harness_retry_path_invalid), forcing attempts 3 and 4 into full fresh-context restarts — which then also segfaulted, discarding all prior conversation progress on every retry.Proposed remediation
signal=SIGSEGV, cap retries at 1 instead of exhausting all 4 — a repeat segfault on retry 1 is extremely unlikely to succeed on retries 2-4 and just wastes runtime.bun.reportcrash links upstream to Bun's team (three distinct crash reports are already in the run logs).harness_retry_path_invalidfallback so a failed--continuedoesn't silently degrade into full fresh-run retries — surface it as its own diagnostic signal so it doesn't get conflated with normal retry behavior.Success criteria / verification
signal=SIGSEGVentries in Claude Code CLI harness logs for Avenger over the following 7 days.Update — 2026-08-13 12:38 UTC
Still not fixed, and now spreading to a second workflow.
Claude Code User Documentation Reviewhit the identical signature in run §31700716464, job 94449125253:panic(main thread): Segmentation fault at address 0xFFFE000000000068,oh no: Bun has crashed,signal=SIGSEGV, all 3 harness retries exhausted, total duration 1m 19s.audit-diffvs. the last successful run of this workflow (§31597056301) again shows no firewall/domain anomalies — but call volume collapsed:api.anthropic.comdropped 170% (37 vs. 100 allowed calls) and Sentry/OTLP ingest dropped 329% (7 vs. 30). That's consistent with the process dying ~15s into a normal session, before it could complete its usual call sequence — further confirming this is an in-process Bun runtime fault, not a network/firewall regression. Raise this to P0-critical: the crash is no longer isolated to one workflow.Parent report: #51983
Related to #51983
Update — 2026-08-15 (6h window)
Widen the fix scope — this is now a 3rd workflow, not 2. Fix the Bun runtime crash before it spreads further; each new hit costs a full retry budget for zero output.
New occurrence: §31855680444 —
[aw] Failure Investigator (6h)— identical Bun 1.4.0Illegal instructionpanic (bun.report/1.4.0/...) under the Claude engine, matching this issue's signature exactly. Auto-filed as #52807, now closed as duplicate of this issue.No new root-cause information beyond what's already documented above — just confirming the blast radius is growing (Avenger x2 + Failure Investigator x1 = 3 workflows affected).> Generated by 🔍 [aw] Failure Investigator (6h) · agent · 210.5 AIC · ⌖ 9.45 AIC · ⊞ 5.6K · ◷
Update — 2026-08-16 (6h window)
Still not fixed — 2 more Avenger occurrences today. Same signature,
exit code 139(SIGSEGV), no other error detail in this window's logs:No new root-cause detail beyond what's already documented — confirming Avenger is still 100% blocked by this crash. Fix the Bun pin / fast-fail-on-SIGSEGV remediation above before the next scheduled Avenger run to stop burning the full retry budget for zero output.> Generated by 🔍 [aw] Failure Investigator (6h) · agent · 121.4 AIC · ⌖ 6.57 AIC · ⊞ 5.6K · ◷
Update — 2026-08-16 22:37 UTC (new occurrence)
Still not fixed — a new Avenger occurrence hit in this 6h window, exit code 139 again.
Execute Claude Code CLI,##[error]Process completed with exit code 139.— identical signature to every prior hit in this issue.auditconfirms genuine mid-session traffic before the crash (67api.anthropic.comcalls, 0 blocked firewall requests) — same in-process Bun fault pattern, not a network/policy regression.No new root-cause detail — just confirming Avenger is still fully blocked. The Bun-pin / fast-fail-on-SIGSEGV remediation already proposed above remains the fix.> Generated by 🔍 [aw] Failure Investigator (6h) · agent · 237.9 AIC · ⌖ 8.49 AIC · ⊞ 5.6K · ◷