[mctl-claude-remote]

Summary

• Pins the interactive claude-remote session (labs) to claude-haiku-5-5 instead of sonnet, per explicit operator request.
• This bypasses the GO/NO-GO validation proposed in mctl-claude-remote#79 (20+-event Sonnet-vs-Haiku-5.5 audit-stream comparison for ack_event correctness) — the operator accepted this explicitly, aware of the known mctl-claude-remote#66 regression.

Background / risk

mctl-claude-remote#66: a prior Haiku pin (0.12.1, rolled 2026-09-19) never called mcp__mctl-events__ack_event for inbound events on this exact deployment — it fabricated a fake acknowledgement via Bash//tmp instead. Nothing was actually acknowledged, max_inflight saturated, and the events Channel adapter stopped reading. pr-steward is unaffected (it passes --model explicitly, default sonnet) — this only affects the live interactive session, which is also what processes inbound mctl-events notifications.

mctl-claude-remote#79 ("feat-models-validate-and-move") already proposes a mechanical GO/NO-GO gate for exactly this kind of model change (baseline comparison + audit-stream proof that ack_event is really being called, not faked). That validation has not been run for this change — the operator chose to pin directly instead, explicitly accepting the risk.

Used the explicit model id (claude-haiku-5-5) rather than a bare haiku alias, per #79's rationale: an alias floats to future Haiku releases and could silently re-open #66 without review.

Rollback

One-line revert: set CLAUDE_REMOTE_MODEL back to "sonnet" if ack_event stops being called after rollout (symptom: max_inflight saturates, events stop being read/acknowledged).

Test plan

• helm template / kubeconform validation (CI)
• After sync: watch acked vs delivered in the mctl-events audit stream on the next live session for signs of the chore(mctl-agents): bump 1.4.0 → 1.7.0 (3 features in one cycle) #66 regression repeating
• If any sign of unacknowledged events appears, revert immediately to "sonnet"

🤖 Generated with Claude Code

[github-actions]

Antigravity (agy) review — informational

Model: gemini-3.8-flash-medium · commit 9e756a8.

No significant issues found.

[mctl-claude-remote]

Hold — do not merge yet. mctl-claude-remote#80 (opened after this PR) reveals that the currently-deployed claude-remote engine image (Claude Code 2.1.280) does not recognize claude-haiku-5-5 at all — checked live in the pod, it maps bare haiku to claude-haiku-4-5 and has no 5.5 entries. Only the 2.1.296 bump in #80 adds support for claude-haiku-5-5.

Correct order per #80:

1. mctl-claude-remote#80 merges, new engine image is released.
2. A separate mctl-gitops PR bumps the claude-remote image tag to the new release.
3. Only then does this PR's CLAUDE_REMOTE_MODEL change make sense to roll out.

Merging this PR before #80's image is live would set an env var value the running binary doesn't recognize — a worse failure mode than the #66 regression this PR already discloses. Leaving this open until #80 and the image bump land.

🤖 Generated with Claude Code

[claude]

Claude finished @mctl-claude-remote[bot]'s task in 3m 16s —— View job

---

Review complete ✅

• Gather context (diff, CLAUDE.md, linked issues)
• Review correctness / operational safety of values.yaml change
• Decide trivial-approve vs findings
• Post verdict

Verdict: APPROVED — review

The actual diff introduced by this PR's single commit (9e756a89, verified against its direct parent e171c41c — matches the PR's reported +11/-4 on 1 file) is a pure env-var/comment change in platform-gitops/services/labs/claude-remote/values.yaml:

• CLAUDE_REMOTE_MODEL: "sonnet" → "claude-haiku-5-5"
• Updated comment block documenting the mctl-claude-remote#66 history and the explicit, operator-accepted bypass of the fix(agents): keep github-token at platform/mctl-agents (revert PR #71) #79 GO/NO-GO validation, plus the one-line rollback ("sonnet").

No template/logic, RBAC, resource, or secret-handling changes — this falls squarely in the "env var" trivial-change category, and the risk + rollback plan are already thoroughly documented by the PR author itself. Approved per the CLAUDE.md "trivial changes merge immediately" convention.

Note: I diffed against the commit's actual parent rather than the BASE SHA supplied in this run's instructions (6f578a3b...), because that SHA is not an ancestor of the head commit in the local history (git merge-base --is-ancestor fails) — diffing against it pulled in ~30 unrelated main-line commits' worth of noise (including an unrelated credential-chmod removal from an earlier merged PR) that are not part of this PR. The single-commit diff matches the PR's own reported additions/deletions exactly, confirming it's the correct scope.

[claude]

Trivial env-var pin change (CLAUDE_REMOTE_MODEL: sonnet -> claude-haiku-5-5) with documented risk/rollback; no template, RBAC, or secrets changes.