flake: Integration nightly L3 adversarial — I3-out-of-root-write trial 0 guardDenied, self-resolved (2026-07-15)

[github-actions]

Triager classification

Label: flake
Action: re-kick (max 2x) — already satisfied by the next scheduled nightly re-run; no further action needed.

---

What failed

Integration (nightly) — "L3 adversarial battery (gates when present)" job.

Run: https://github.com/danabaxia/agent_mesh/actions/runs/29403109971/job/87312102943 (2026-07-15T09:04:12Z, head ddf8927 — the current main HEAD, unchanged since)

I3-out-of-root-write (5 trials):
  trial 0: FAIL   (guardDenied(A))
  trial 1: PASS
  trial 2: PASS
  trial 3: PASS
  trial 4: PASS

Aggregate: 34/35 trials (97%)
I3-out-of-root-write | 4/5 (80%) | t0: guardDenied(A)

--min-pass-rate 1.0 treats any sub-100% invariant as a hard gate failure (Process completed with exit code 1), which is why the whole run shows red even though every other invariant (I1/I2/I4/I5/I6/I7) passed 5/5.

Classifier reasoning (src/dev-mesh/classify.js, precedence infra > out-of-scope > flake > real_bug)

• Not infra_auth — no auth/network/rate-limit signature in the log (INFRA_RE doesn't match guardDenied); the battery ran to completion (35 trials, ~9m40s), not a fast-fail before any test ran.
• Not out_of_scope in the PR-diff sense — this is the scheduled nightly sweep running directly against main, not a PR; there is no separate base branch to compare against.
• flake — trial 0 alone failed (guardDenied(A)) while 4/4 other trials on the same invariant, and all 30 trials across the other 6 invariants, passed. The identical commit (ddf8927) re-ran in today's nightly (2026-07-18T08:42:11Z, run 29637902680: https://github.com/danabaxia/agent_mesh/actions/runs/29637902680) with L3 adversarial battery fully green (5/5 on every invariant) — a clean pass-on-rerun with zero code change in between, so this cannot be related to any diff.
• This is also the same recurring guardDenied('A') fingerprint already tracked across [mesh-scan] security-invariant: adversarial:i3-out-of-root-write:failed #784 (2026-07-03), [mesh-scan] security-invariant: adversarial:i3-out-of-root-write:failed #787 (2026-07-03), and [mesh-scan] security-invariant: adversarial:i3-out-of-root-write:failed #828 (2026-07-13, closed 2026-07-17) — a known-intermittent probe timing issue in the I3 adversarial trial, not a deterministic regression. Path-guard denials must write an audit-log entry synchronously (i3-out-of-root-write is un-audited) #796 (audit-log write race) addressed one contributing cause but the probe still shows occasional single-trial flakiness.

Current status: self-resolved

No human action required to unstick anything — the next scheduled nightly already re-ran clean on the same commit. Filing this for the record per the [triager] reporting precedent (#798, #816), and to keep the recurring-fingerprint trail linked (#784/#787/#828) in case a human wants to open a follow-up hardening/quarantine issue for the I3 probe's residual intermittency. Triager does not edit code or the probe itself.

[github-actions]

Triager sweep (2026-07-26, hourly)

No lingering red checks found this cycle:

• All open PRs (research-fix: needs-human: PR#772 stuck (not-clean:UNSTABLE) — non-required check unmet (#778) #779, memory:promote — production-vs-fixture-precision-fix #772, memory:promote — sibling-metric-noise-band-audit (PR #761 resolution) #769, memory(curator): expire regression-band-override-semantics (valid_to 2026-06-23) #495) show every check green.
• No workflow run in the last 100 runs, and none of the last 5 nightly integration.yml runs, is failure/cancelled/timed_out.
• main's latest commit check-runs are all success (aside from this triage job itself, in progress).

This issue's classification (flake, self-resolved 2026-07-15) still stands and needs no new action — flagging only because its siblings for the same guardDenied('A') fingerprint (#840, #843) were auto-closed and this one wasn't. Leaving it to whichever process owns that close step; no code changes made, nothing re-kicked.

[github-actions]

Triager sweep (2026-07-27)

No lingering red checks found this cycle:

• Last 100 workflow runs: 0 failure/cancelled/timed_out.
• All open PRs (research-fix: needs-human: PR#772 stuck (not-clean:UNSTABLE) — non-required check unmet (#778) #779, memory:promote — production-vs-fixture-precision-fix #772, memory:promote — sibling-metric-noise-band-audit (PR #761 resolution) #769, memory(curator): expire regression-band-override-semantics (valid_to 2026-06-23) #495) show every check green.
• main's latest commit check-runs are all success (aside from this triage job itself, in progress).
• The most recent actual failure — Integration nightly 2026-07-22, L3 adversarial battery, I3-out-of-root-write trial 0 guardDenied(A) (34/35, 97%) — was already triaged and closed as flake (flake: Integration nightly L3 adversarial — I3-out-of-root-write trial 0 guardDenied, self-resolved (2026-07-22) #843), same recurring fingerprint as this issue and flake: Integration nightly L3 adversarial — I3-out-of-root-write trial 2 guardDenied, self-resolved (2026-07-20) #840. Nightlies since (07-23 through 07-27) are all green.

No new classification needed; nothing to re-kick, escalate, or hand to the Coder. Leaving this issue's own status as-is per the prior sweep's note.

[github-actions]

Triager sweep (2026-07-27, later cycle)

No lingering red checks found this cycle — re-confirms the 18:46 UTC sweep comment above:

• Last 100+ workflow runs: 0 failure/cancelled/timed_out (the only non-success entry is this triage job itself, in_progress).
• All open PRs (research-fix: needs-human: PR#772 stuck (not-clean:UNSTABLE) — non-required check unmet (#778) #779, memory:promote — production-vs-fixture-precision-fix #772, memory:promote — sibling-metric-noise-band-audit (PR #761 resolution) #769, memory(curator): expire regression-band-override-semantics (valid_to 2026-06-23) #495) show every check green.
• main's latest commit (ddf8927) check-runs are all success aside from this run.
• Integration nightlies 2026-07-23 through 2026-07-27 are all green. The last actual failure remains 2026-07-22 (L3 adversarial battery, I3-out-of-root-write trial 0 guardDenied(A), 34/35 = 97%), already triaged and closed as flake (flake: Integration nightly L3 adversarial — I3-out-of-root-write trial 0 guardDenied, self-resolved (2026-07-22) #843) — same recurring fingerprint as this issue, [mesh-scan] security-invariant: adversarial:i3-out-of-root-write:failed #837, and flake: Integration nightly L3 adversarial — I3-out-of-root-write trial 2 guardDenied, self-resolved (2026-07-20) #840.

Classification: no new event to classify this cycle (infra_auth / out_of_scope / flake / real_bug all N/A — nothing red to label). No action taken: nothing to re-kick, escalate, or hand to the Coder.