[mashkovd]

Why

The mctl-agents cheap profile moved to claude-haiku-5-5 in mctlhq/mctl-agents#666, and claude-remote already runs on it. claude-firstparty.json had no claude-haiku-5-5 entry, so Catalog.Calculate returned ErrNoPricing and every Haiku 5.5 usage record was stored unpriced.

What

• claude-firstparty.json: a new append-only entry, version 2026-10-17-firstparty-1, effective_from: 2026-10-17T00:00:00Z. Its rates come from the pricing page's "Model pricing" table (https://platform.claude.com/docs/en/about-claude/pricing, retrieved 2026-10-17), the row for prompts up to 100,000 tokens, in USD per MTok: input 0.10, output 0.50, 5m cache write 0.125, cache hit 0.01. Web search costs $0.01 per call. No existing entry is edited.
• README.md: adds the table row and a section stating the known limitation.
• mctl-api.yaml: bumps ROLLOUT_MARKER, because mctl-api reads the catalog only at startup.

Known limitation (owner decision: option 1)

Haiku 5.5 is priced per request by prompt length. Over 100,000 prompt tokens, every dimension costs 5x. mctl-api's catalog holds one flat rate set per entry, and a usage record aggregates a whole query, so the tier cannot be resolved from the record. This card uses the up-to-100k rates, which fit the short mechanical prompts cheap is meant for. Any request with a longer prompt is under-priced by up to 5x, so Haiku 5.5 costs should be read as a lower bound. Tier-aware pricing is tracked in mctlhq/mctl-api#642. When it lands, a new entry supersedes this one.

Records already ingested without a price before this card existed are not re-priced by this change.

Verification after merge

• The mctl-api pod restarts on the new ROLLOUT_MARKER and loads the catalog without error.
• A new claude-haiku-5-5 usage record carries pricing_version = 2026-10-17-firstparty-1 and a non-null calculated_cost.

🤖 Generated with Claude Code

https://claude.ai/code/session_01BKnkncXtavYRiryZ2k33nM

[claude]

Claude finished @mashkovd's task in 2m 20s —— View job

---

Review complete ✅

• Gather context (CLAUDE.md, pinned diff, existing catalog/README)
• Verify new claude-haiku-5-5 pricing entry against PR's stated rates and append-only invariant
• Check ROLLOUT_MARKER bump and its comment
• Post inline findings
• Submit review verdict

Verdict: APPROVE (0 P1, 0 P2, 1 P3)

Checked git diff 63cd99835...dcd0653d1:

• claude-firstparty.json: new entry's input_per_mtok/output_per_mtok/cache_write_per_mtok/cache_read_per_mtok/web_search_per_call (0.10 / 0.50 / 0.125 / 0.01 / 0.01) match the rates stated in the PR description. version/effective_from follow the existing naming convention, and (canonical_model, provider, effective_from) doesn't collide with any existing entry — append-only invariant (ADR-012 deploy: libertex/nginx → plain-text #7, per the README) holds.
• README.md: new table row and "known limitation" section accurately describe the tiered-pricing gap and point at the tracking issue (mctl-api#642). Good-faith, well-documented under-pricing tradeoff.
• mctl-api.yaml: ROLLOUT_MARKER bump is required per this file's own "Restart requirement" doc, since the catalog is read only at startup — correctly done in the same commit.

P3 (nit, doesn't block): the Bump only after zitadel-iac has applied generated identity configuration... sentence above ROLLOUT_MARKER is leftover guidance from the previous rollout (mctl-gitops#1923, loading ZITADEL_PORTAL_CLIENT_ID) and doesn't apply to this one. Left as-is it could make a future contributor think they need to wait on zitadel-iac for an unrelated bump. See inline comment for a suggested trim. Fix this → on branch feat/usage-pricing-haiku-5-5.

One thing outside the diff itself, flagged for awareness only (not a code issue): the PR description cites mctlhq/mctl-agents#666 for the cheap profile move, while the new README text (and the prior commit's subject, chore(agents): shepherd issue-663-...) cites #663. Worth double-checking which issue number is correct before merge, though it has no effect on the catalog/pricing logic.

[claude]

P3: Bump only after zitadel-iac has applied generated identity configuration: envFrom is read at container start. is leftover guidance from the previous rollout (loading ZITADEL_PORTAL_CLIENT_ID, mctl-gitops#1923) and has no bearing on this one (loading the usage-pricing catalog). Left as-is, it reads as a precondition for this bump too, which could make a future contributor wait on zitadel-iac unnecessarily. Consider trimming it to the parts that are actually rollout-reason-agnostic (the 503-until-restart rationale) and keeping reason-specific preconditions only in the line directly above ROLLOUT_MARKER.

[mashkovd]

Valid, deferring as P3: that sentence predates this PR (it came with the #1923 rollout) and this PR only replaces the reason line directly above ROLLOUT_MARKER. I will trim it to the reason-agnostic rationale the next time the marker is bumped, rather than spending another review round on a comment-only change here.

[claude]

No P1/P2 findings (1 P3). Good to merge.

[github-actions]

Antigravity (agy) review — informational

Model: gemini-3.8-flash-high · commit dcd0653.

No significant issues found.