Sample report
No client. Our own factory.We audited our own pipeline first.
Before we run the six-area audit on anyone else's factory, we ran it on ours — Ship's own delivery loop, the one shipping this site. Six findings below, two of them still open. This is the register format every audit produces; nothing here is dressed up for the page.
Findings register
Six areas. Six findings. Two still open.
E · Security & Abuse
One workspace's dogfood environment shared a tracker project with another workspace.
Evidence — A PR-lookup path trusted a cached record instead of re-resolving the open PR from source. In a narrow race, that cache could answer for the wrong tenant.
Fix — Fixed: the gate now re-resolves the open PR from source, and tracker bindings became strictly 1:1 — a second workspace claiming the same tracker project gets a hard 409, not a shared lock. A broader tenant-fence hardening pass is still rolling out.
Status: Fixed core bug; hardening in progress
D · Failure Modes & Reliability
A crashed agent could leave a workspace's project_lock held for up to 24 hours.
Evidence — Two independent paths leaked the same lock: the picker returning null mid-acquire, and an agent process dying after acquiring but before releasing. Neither path had a ceiling.
Fix — Root-caused both leak paths; a reconciler now force-releases stale locks past a hard age ceiling instead of waiting on graceful release.
Status: Fixed
C · Cost & Latency
Fifteen identical fix commits landed in a row on one afternoon.
Evidence — A single specialist role was carrying three jobs — checking work against the contract, repairing obvious damage, and triaging regressions. When a fix broke a different test, the role kept re-attempting the same repair instead of escalating.
Fix — Split into three narrow specialists — validation, self-heal, and regression-triage — each with a job a human could describe in one sentence and permission to do only that job.
Status: Fixed
A · Context Engineering
Decomposition agents couldn't read the product brief or write structured output back.
Evidence — The planning step summarized from memory instead of the source document, and had nowhere durable to put a work-breakdown structure once it produced one.
Fix — Wired the brief as a direct read and added a real section for decomposition output instead of routing it through free text.
Status: Fixed
B · Evaluation & Regression Safety
A local import shadowed a module-level name and only failed on the very first run of a fresh install.
Evidence — The bug was invisible in CI because test ordering happened to warm the module first every time. Linters flagged the shape (ruff F823) but nothing gated on it.
Fix — Open. Known, reproduced, not yet shipped — included here because a self-audit that finds nothing outstanding isn't a credible one.
Status: Open
F · Architecture & Operability
A designed harvesting step for operator feedback has never actually run in production.
Evidence — The code path exists and is exercised in tests; the production trigger that should invoke it on a schedule was never wired up.
Fix — Open. Low severity because nothing depends on it yet, but it's exactly the kind of gap that's invisible until someone goes looking for the data it was supposed to produce.
Status: Open
What this means for yours
None of these findings are exotic. A shared cache resolving to the wrong tenant, a lock nobody remembered to time out, a role doing three jobs and quietly retrying instead of escalating — these are the ordinary failure modes of any team running agents in production, including us. The two still open stayed in this report because the point isn't to look finished. It's to show what the register actually looks like when we run it on ourselves — and then run it on you.
Next
Same register. Your pipeline.
Five days, read-only access, a findings register with a dollar figure on every line. $3,500 — half back if it doesn't pay for itself.