fix(opencode): price Z.AI usage with provider-qualified rates#21646

Approvability

Verdict: Needs Human Approval

Macroscope's review found this PR requires human approval — This focused fix correctly routes selected OpenCode Z.AI usage through provider-qualified rates and adds coverage for both storage formats and pricing precedence. Because it changes production usage cost/metering results, human review is required despite the limited scope.

^{You can add or adjust custom eligibility rules. Learn more.}

Review in Change Stack →

No actionable comments were generated in the recent review.

• Autofix · Keep fixing CodeRabbit findings and required CI, and resolving merge conflicts

_{Comment @coderabbitai help to get the list of available commands.}

thanks @JesstLe. your fix uses the existing rateModel lookup and preserves display names, tokens, reported costs, and custom overrides. i kept the implementation and both regressions intact and merged current main (cc93cc564b05) into your branch at 514df5ea1e419e538b20be097546bbf611250e04, resolving the test conflict without dropping main's instance-history coverage. the final diff is still seven production lines and two tests in existing files.

verification passed through blacksmith-testbox:

• both regressions fail with main's reader: missing zai/glm-5.3, and $0.25 instead of $0.250215.
• all 70 focused reader/service/pricing/aggregation tests pass, without platform filters.
• vp run check:changed passes formatting, lint, both knip checks, graphql validation, five package typechecks, and all 47 nearby tests.
• real server and web client, matching isolated OpenCode history and deterministic conflicting rates: glm-5.3 changes from unpriced to $2.15, retaining 1.50m tokens. these are fixture rates, not a claim about current published prices.

current main no longer caches OpenCode rows in the transcript scan cache: openCodeUsageReader scans on each summary, and cache decoding rejects providers without transcript formats. the older cache-bump review finding therefore no longer applies; no global cache invalidation was added.

unverified: a billed live Z.AI request, Electron, native mobile, and remote/relay/tunnel execution. no client or wire code changed; every client consumes the shared server summary. all 24 hosted checks are now green, github reports this head mergeable, and coderabbit generated no actionable comments in its latest review. two independent gpt-6.1-sol reviews also returned merge-ready for this exact sha. the old cache thread is resolved with the reason above. human approval of the metering change remains Maria's decision.

the uploaded captures were opened and inspected at full size. gitcafe currently renders external image markdown as links in this comment; inline rendering here is unverified.

before: glm-5.3 is unpriced with 1.50m tokens

after: the same glm-5.3 record costs $2.15 with 1.50m tokens

Verified in a real dev environment

On this Linux Boat VM I installed dependencies and ran vp run dev first from a separate github/main worktree at 4e4d4ba3959cae17db7bba1915e6a25885844484, then from the exact PR head 514df5ea1e419e538b20be097546bbf611250e04. Main was stopped by its captured process group before starting the PR server. Each used isolated T3 state with automatic data seeding disabled.

Using Playwright from the repo dependencies and /usr/bin/google-chrome, I added the real /home/user/t3code git project through the app, selected Codex gpt-6.1-sol, and submitted the same read-only prompt on each revision. Both turns returned usage verification ok. through the actual web composer/server/provider path. The real usage cost view loaded on both; switching Cost to Tokens on the PR retained matching model totals. These are neighboring Codex checks, not proof of the OpenCode Z.AI fix. Aggregate usage includes this verification agent and grows between captures; these are not matching-data Z.AI before/after evidence.

Runtime transcript:

Both full-size captures were inspected, uploaded, and reopened successfully as 1440x1000 images. I reopened this comment in the browser; GitCafe renders the external image Markdown as links rather than inline images.

main: real Codex usage cost view, neighboring behavior only

PR: real Codex usage cost view, neighboring behavior only

Not verified

The changed OpenCode Z.AI pricing behavior remains unverified. This VM has no Z.AI credentials or existing OpenCode Z.AI history. Codex turns do not produce zai/zai-coding-plan OpenCode records. I could not reproduce the original bug on main or prove qualified-rate pricing, positive reported-cost precedence, custom overrides, or restart persistence with real Z.AI usage. I did not manufacture history or rates, use fixtures, or count the prior fixture-based acceptance as live-provider proof. Completing this check requires a configured Z.AI account and a real OpenCode turn, then comparing that same real usage history on main and this head.

Codex Limits fails on both revisions with Codex could not read usage (JSON-RPC -32600) through the local proxy. This is a baseline limitation, not evidence of a PR regression. Electron, native iOS/Android, and remote/relay/tunnel paths were not exercised. No interaction/timing change is claimed; evidence is the server-path transcript and static neighboring-usage captures.

The forge reports 24 successful hosted checks, zero failing and zero pending. I did not rerun those suites or check:changed, and CI does not establish the missing real-provider proof. No source changes, commits, pushes, labels, merge, rebase, force-push, or closure were made. Verdict for the requested changed behavior: unverifiable on this VM with its current provider access.