[lucasheriques]

Why

The pre-PR gate should catch what review bots keep catching. On PostHog/posthog-python#1226 it didn't: after our gate passed, reviewers found six more bugs across five rounds. Those were a broken older supported dependency version, secrets in captured stack-frame source lines, a spec's JSON-member redaction rule not implemented, a background task keeping a traceback alive, an exponential regex, and a quadratic scan on \" runs. #22 and #23's lenses were installed and in the brief the whole time. Two things went wrong:

• Review fixes went out ungated. The gate ran once before the PR opened. Every later fix went straight to reviewers, and two of them caused the next finding.
• The lenses had gaps. No lens covers hostile-input cost. Exposure let an adversary pass a leak because an older path leaked the same way. Completeness took its versions from what was installed, not from what the project declares it supports. Nothing sent adversaries to the spec that governs the behavior.

What changed

• blind-spots.md gains a Cost lens: hostile-input time, grown from small sizes with a per-run timeout, near-match and escape-sequence probes for backtracking regex engines, and memory a long-lived holder keeps alive.
• Parity also triggers on behavior a spec defines, and reads a spec's blanket requirements ("every captured string").
• Completeness takes environments from declared support ranges and the CI matrix, and runs the oldest and newest version of each peer.
• Exposure puts canaries in everything a value copies (source lines, locals, environment), in every shape the contract names, and judges against the contract. A leak on an older path the change doesn't touch becomes a non-blocking follow-up.
• interrogate's Frame also reads the specs the repo's agent instructions name. Its brief, Verify step and Output gain Follow-ups, and Opening a PR step 5 reports them.
• bugbot-triage.md's fix rule runs interrogate on every behavior-changing review fix before it's pushed. A one-character regex or limit change counts. It allows one more pass on the fixes for its findings and never a third, and a finding you won't fix holds that push and goes to the user. Babysit step 8 and the multi-phase fix-round rule point every review fix at it, a person's comment included.
• translation.md records the fork-only departures.

Verification

Replay. Sonnet adversaries got the same brief on the posthog-python commits where reviewers later found bugs. The only thing that changed between runs was the lens file.

Commit	Bug reviewers found later	main lenses	This PR's lenses
8bbd1fb (first gate head)	fastmcp 2.x silently loses all analytics	missed	found
8bbd1fb	secrets in frame source lines	missed	found
8bbd1fb	JSON-member credentials unredacted	missed	found (once Frame named the spec)
8bbd1fb	background task keeps the traceback alive	missed	found
d8a7a55	exponential regex	found (my brief primed it)	found
9c1cb4f	quadratic scan on \" runs	missed (at d8a7a55, where it already existed)	found (escape-sequence probe)

The new-lens runs also found an unreported gap: object keys skip the JSON-member pass. They each reproduced the bug against the real code, with timings, before/after event dumps, or failing calls.

Gate. Two rounds of Opus and Sonnet adversaries.

• Round 1 found that the review-fix rule broke Babysit's "In that case" sentence, had no scope and no round cap, and exempted rewrites of the cited lines. It also found the Cost timing unbounded, older-path leaks blocking unrelated PRs, and translation.md contradicting itself. All of these are fixed in 1331ac5.
• Round 2 found the "one-line value change" exemption, the imported step-2 blocker semantics, human comments skipping the gate, no output slot for follow-ups, and a multi-phase plan conflict. All are fixed in 5cbc168.
• There's no third round, per Opening a PR step 2. I checked each round-2 finding against the final text instead.
• bun scripts/check-port.ts reports 0 findings.

Blast radius

• Gate runs on code with caller-controlled parsing or regex now carry the Cost lens. Over the last 40 non-test commits on posthog-python's main, that's 7.
• Every behavior-changing review fix now costs one interrogate fan-out, or two at most, before it's pushed. That replaces review rounds like #1226's five.