This is a well-built chart. It reuses the libcoder library the way the coder and provisioner charts do, keeps the imperative surface to a few template helpers, ships render-time validation with precise field-level error messages, and backs it with a deterministic golden-file suite (63% test density) that the panel ran green under -race and a GOMAXPROCS sweep. Bisky's take: "this test suite is lovely... These are real gems." No scope drift: every one of the 40 files belongs to the chart or its mechanical build/release wiring, which mirrors the existing chart entries line-for-line.

Severity count: 1 P2, 10 P3, 4 Nit, 6 Note. No P0/P1, so this is a COMMENT, not a block. The one finding to fix before shipping is the liveness/startup race (CRF-4): three reviewers independently traced that the process does not open its /healthz listener until it has connected to coderd and loaded providers, so on a fresh co-install or during a coderd outage the gateway crashloops. The rest cluster into contract-narrowing (CRF-5, CRF-6, CRF-7), operator-signal correctness (CRF-8, CRF-9, CRF-23), and convention drift (CRF-12, CRF-13).

Process notes, not inline findings:

• The commit subject / PR title misspells "helm" as "heml" (CRF-16). It squash-merges into main permanently. Suggest feat(helm/ai-gateway): add Helm chart for standalone AI Gateway.
• The PR description claims the suite covers "all validation error cases," but 6 of the 13 fail branches have no test (this is CRF-2). Mafu-san flagged this as fake completeness: a reviewer trusting the description approves ~46% of the validation logic as tested when it is not. Either add the cases or scope the wording.
• Netero's round-1 tc := tc nit was dropped: Komugi empirically proved the copy is load-bearing for the parallel multi-namespace subtests (deleting it fails default_values deterministically). Do not remove it.

No findings against the build/release plumbing: the charts list drives the build/ai-gateway_helm_$(VERSION).tgz target before signing, and the attestation ./build/*_helm_*.tgz glob already covers the artifact.

🤖 This review was automatically generated with Coder Agents.

Strong round-2 turnaround. Every one of the 19 round-1 findings the author marked fixed was independently verified against the tree, and the important ones were fixed at the root rather than patched at the symptom: the chart-owned env guard now derives its protected list from the rendered ownedEnv (no more hand-maintained duplicate), automountServiceAccountToken: false moved to the pod spec so it holds on the disableCreate path, extraTemplates delegates to libcoder's coder.renderTemplate again, and the HTTPRoute CRD guard is now actually testable via per-case apiVersions. Bisky: "this test suite is a genuine gem, and I do not say that often" - all 13 validation branches now have a verbatim-message test. TestRenderChart passes.

Contested CRF-9 (listenerTLS behind ingress/httproute): the panel agrees unanimously with the author that a hard validate reject would be wrong, because backend-TLS to the HTTPS listener is legitimate for controllers that re-originate TLS and the chart cannot know the controller's annotation dialect. Kurapika confirmed it fails closed, so there is no security exposure. Five reviewers accept the new README section as a sufficient ceiling. Two (Pariston, Leorio) re-raised a cheap residual: the docs teach the fix but never name the symptom (a 502 / TLS handshake error), and there is no install-time signal even though NOTES.txt already computes the exact condition. Kept open at P3 as a non-blocking doc/UX ask, not a request for validation. See the reply on that thread.

Severity count this round: 2 P3, 1 P4, 4 Nit (all new; no regressions of prior fixes). Two are follow-on consequences of round-1 fixes rather than the original issues re-opening (CRF-27 is the sibling of CRF-7; CRF-28 is guidance that can undo CRF-4).

Acknowledged as intentional by the author and accepted: CRF-18 (metrics listener always-on, matches sibling charts) and CRF-22 (readiness coupled to the coderd DRPC connection).

Process notes:

• The CI title check is red (✗ title). The current PR title parses clean against the conventional-commit validator, so this may be a stale run after the "heml" -> "Helm" rename, but reviewers could not read the failing run (no GitHub credentials in the worktree). Please confirm it is green before merge rather than assume the rename cleared it; a failing check is a bug until proven stale.
• The branch commits still read feat: add heml chart for standalone AI Gateway and agentic review 1. The PR title is corrected, so a squash merge lands a clean subject and this is cosmetic, but if the branch is rebase- or merge-committed the "heml" typo reaches main history. Squash, or reword both subjects.

🤖 This review was automatically generated with Coder Agents.

Round 3 review is blocked on one unaddressed finding, so the panel did not run this round.

The rest of the round looks good: seven findings from round 2 received targeted fixes and per-thread replies (CRF-27 env/envFrom guard wording, CRF-28 probe-comment rewrite, CRF-29 livenessProbe golden coverage, CRF-30 secret-ref naming, CRF-31 grace-period comment, CRF-32 README --namespace, CRF-34 probe-template dedup). Those are queued for panel verification and will be checked on the next round.

Blocking item, needs a response before the panel re-engages:

CRF-9 (P3, listenerTLS behind ingress/httproute). In round 2 the panel agreed with you that a hard validate reject is wrong and accepted the README section as the right layer. Two reviewers left a small, concrete residual open: the README teaches the correct backend-TLS config but never names the failure symptom, so an operator debugging a 502 has nothing to search for. The ask was one sentence, for example: "If the entry point is not configured for backend TLS, it speaks plaintext HTTP to the HTTPS listener and every request fails with a TLS handshake error, typically surfaced as an HTTP 502." An install-time NOTES.txt hint was suggested as optional. This round's commit touches README.md but neither README.md nor NOTES.txt contains "502" or "handshake," and there is no reply on the thread; it was resolved without a code change or response. A resolved thread with no fix and no explanation is not a resolution.

To unblock, do one of: add the symptom sentence to the README TLS section; file a tracking issue and link it here; or reply explaining why it should not be done. Any of the three clears the block.

Transparency on our side: the round-2 re-raise was posted as a reply that tried to reopen the thread, but our bot token lacks the permission to unresolve threads (the API rejected it), so the thread stayed marked resolved and the residual was easy to miss. That is on us. The finding itself still stands.

🤖 This review was automatically generated with Coder Agents.

Round 4 is in good shape. All eight fixes from rounds 2-3 (the seven that round 3's blocked panel never got to verify, plus the round-4 CRF-9 doc sentence) were independently checked against the current tree and confirmed real, structural fixes rather than symptom patches: the env guard now derives its owned-set from the rendered ownedEnv and documents the envFrom gap honestly, the probe comments no longer misdirect, the three probe blocks are collapsed into one coder-ai-gateway.probe template with golden coverage, the secret-ref naming is uniform, and the grace-period coupling is documented on both sides. TestRenderChart passes and Komugi re-forced the parallel-namespace schedule with no flake. Hisoka: "The seams I expected to split held under load. A worthy opponent."

Four new findings this round, severity count 1 P3, 2 Nit, 1 Note. The P3 is a pre-existing correctness gap the panel reached this round, not a regression. The two Nits are the tail of the round-2 namespace-flag fix (the sweep fixed two commands and missed two) and a comment-wording risk. None are blocking; this stays a COMMENT.

Komugi withdrew the round-2 CRF-35 note (goldens hardcoding 0.1.0 matches Chart.yaml and the sibling charts; not a defect). CRF-18 and CRF-22 remain accepted as intentional by the author.

🤖 This review was automatically generated with Coder Agents.