feat(usage): add OpenCode as a usage provider - #8260
Conversation
|
Important Review skippedAuto reviews are disabled on this repository. Please check the settings in the CodeRabbit UI or the ⚙️ Run configurationConfiguration used: Repository UI Review profile: CHILL Plan: Pro Plus Run ID: You can disable this status message by setting the Use the checkbox below for a quick retry:
Thanks for using CodeRabbit! It's free for OSS, and your support helps us grow. If you like it, consider giving us a shout-out. Comment |
github-actions
Bot
added
vouch:unvouched
![macroscopeapp[bot]](https://avatars.githubusercontent.com/in/900172?s=60&v=4){kind=small}
![cursor[bot]](https://avatars.githubusercontent.com/in/1210556?s=60&v=4){kind=small}
There was a problem hiding this comment.
Cursor Bugbot has reviewed your changes using high effort and found 1 potential issue.
❌ Bugbot Autofix is OFF. To automatically fix reported issues with cloud agents, enable autofix in the Cursor dashboard.
Want fixes drafted automatically? Bugbot Autofix can create code changes for findings. A team admin can enable Autofix in the Cursor dashboard.
Reviewed by Cursor Bugbot for commit c6fea1f. Configure here.
The usage contract so far covered only Claude Code and Codex transcripts. Widen UsageProviderKind with opencode and bump USAGE_CONTRACT_VERSION so mixed-version environments render as partial coverage instead of failing schema decode.
The usage scan covered Claude Code and Codex transcripts but ignored OpenCode entirely, so its tokens were invisible on the Usage page. OpenCode keeps every message in one WAL-mode database at $XDG_DATA_HOME/opencode/opencode.db instead of JSONL transcripts. Treat the database as a single pseudo-file in the scan: stat db+wal+shm into one (size, mtime) identity for the scan cache, open it read-only with node:sqlite, and map each assistant row to a usage record. Rows are one billed request each, so no dedupe or fork suppression is needed. OpenCode treats reasoning as disjoint from output, so output absorbs it to reconcile with its stored total while keeping the contract invariant.
Add the OpenCode presentation entry reusing the existing OpenCodeIcon so the provider appears in charts, tables, and summaries.
Match the web usage page's new OpenCode provider: series order, label, and a violet band color that holds in both themes.
ApprovabilityVerdict: Not approved Macroscope's review found this PR not approvable — This PR adds a substantial production OpenCode integration, including a new SQLite/WAL reader, usage parser, shared contract provider, and web/mobile presentation. Existing usage requests now inspect and aggregate another local store, creating cross-component runtime behavior beyond a small isolated change. You can add or adjust custom eligibility rules. Learn more. |
c6fea1f to
f38a537
Compare
Three real defects the review bots caught: - decodeScanCache accepted only claude and codex providers, so persisted opencode entries were dropped on every restart and the store was fully rescanned. Accept opencode and cover it with a round-trip test. - json_extract in the WHERE clause throws on a malformed message row before the per-row try/catch, voiding the entire scan. Guard with json_valid so a bad row costs itself only. - The sqlite iteration is fully synchronous, so a cold scan of a large store stalled the event loop where the readline-based JSONL walk yields between chunks. Yield to the event loop every 2048 rows.
|
Note 🤖 GPT-5.6 Sol responding on behalf of Theo We're closing this PR as we clean up the T3 Code backlog. Thank you for taking the time to put this together. #8456 is the active path for OpenCode usage reporting. We recorded the useful differences from this branch, but they do not require a second implementation to remain open. If you believe we closed this in error, please reopen the PR and leave a comment explaining what we missed. |
2 participants
Problem
The Usage page only covers Claude Code and Codex. If you drive OpenCode, every token it spends is invisible there, which makes the page's totals quietly wrong for anyone who uses more than one CLI.
What changed
OpenCode is now a third usage provider. It does not write JSONL transcripts like the other two CLIs; it stores every message in one WAL-mode sqlite database at
${XDG_DATA_HOME:-~/.local/share}/opencode/opencode.db. The usage scan treats that database as one more source:UsageProviderKindgains"opencode"andUSAGE_CONTRACT_VERSIONbumps 4 → 5, so an older client paired with a newer environment renders partial coverage instead of failing schema decode.apps/server/src/usage/opencodeTranscriptReader.tsopens the store read-only throughnode:sqlite(already in the server's dependency set) and streams assistant rows out. The db,-wal, and-shmsiblings stat into one(size, mtime)identity so the existing scan cache invalidates on live writes.costis trusted as provider-reported (a literal0is a real answer from free providers), so most rows skip the LiteLLM fallback entirely.totalsums them. Output absorbs reasoning here to reconcile with that total while keeping the contract's "reasoning is a subset of output" invariant.OpenCodeIcon.Before / after
Before:
After:
Verification
vp test run src/usage/): 46 passed, covering cost trust, reasoning absorption, junk-row rejection, mtime gating, and read-only DB access against a real sqlite file.opencode.db(~19.5k billable rows, zero malformed).Note
Medium Risk
Touches usage aggregation and a contract version bump, but changes are read-only local SQLite access with isolated parsing and no auth or billing paths.
Overview
Adds OpenCode as a third usage source alongside Claude Code and Codex, so Usage charts and totals include its local session spend.
The shared contract gains
"opencode"inUsageProviderKindandUSAGE_CONTRACT_VERSIONbumps 4 → 5 so older clients can still render partial coverage. On the server, OpenCode is read fromopencode.dbunder${XDG_DATA_HOME:-~/.local/share}/opencode(not JSONL): a new reader treats the DB plus WAL/SHM as one cacheable pseudo-file, queries assistant rows with tokens via SQLite JSON, and maps them throughparseOpenCodeMessage(provider-reported cost, no dedupe keys, reasoning folded into output for contract invariants).UsageServiceroutes listing and parsing through that path while reusing the existing scan cache (decode now acceptsopencodeentries). Web and mobile usage UI pick up label, purple styling, andOpenCodeIcon; user docs mention OpenCode.Reviewed by Cursor Bugbot for commit 61975b2. Bugbot is set up for automated code reviews on this repo. Configure here.
Note
Add
opencodeas a usage provider across server, mobile, and webstatOpenCodeDatabase,listOpenCodeDatabaseFiles, andreadOpenCodeRecordsopencodeprovider to call the new reader instead of the transcript reader; cache keying and dedup remain unchangedUsageRecord, mapping reasoning tokens into output and trusting stored non-negative costsopencodewith labelOpenCodeand color#8b5cf6; web adds anOpenCodeIconmarkUSAGE_CONTRACT_VERSIONfrom 4 to 5 and extendsUsageProviderKindto include'opencode'in usage.tsopencodeentries; decodeScanCache now accepts'opencode'so persisted caches round-trip correctlyMacroscope summarized 61975b2.