fix(usage): OpenCode usage fidelity (reasoning, cache key, latency, aborts, local pricing)

[pacphi]

Summary

PR 4 (last) of the usage-fidelity plan: OpenCode host. Every finding was re-verified at 8ebde36 before it was touched.

Finding	Status	What changed
O-1 reasoning missing from tokens and estimated cost	Fixed	A usage row's output now adds reasoning (reasoningOutput keeps it as detail). Guarded so it is never added twice.
O-2 cache key blind to in-place row rewrites	Fixed	Key gains upd = latest message/session time_updated.
O-3 latency measured to the row insert (~15 ms)	Fixed	Ends at assistant time.completed; no completed stamp = no sample. Completed stamp and each row's generation span join engaged time.
O-4 user stop counted as an exception	Fixed	MessageAbortedError goes to aborts. Per-host byHost[h].aborts; tile, session strip, CLI and tooltip corrected.
O-5 local model with no cost priced at the $3/$15 fallback	Fixed	Local providers (lmstudio, ollama, llama.cpp, registry billing: local) are left unpriced and counted as costEvidence.unpricedMessages; no cache saving quoted for them.
O-7 rows keyed (day, model), provider merged/last-wins	Fixed at row level	Rows keyed (day, model, provider). Session-level provider bucket still last-wins (deferred, below). Done before O-5 because O-5 needs the row's provider.
O-8 unreported tokens indistinguishable from zero	Fixed	Completed, error-free, all-zero-token responses are marked tokensUnreported on the row; one usage-not-reported:N warning on OpenCode source health (existing generic warnings, shown in the existing pill tooltip).

Also pinned (verified-correct behaviour): a message with a step-finish part carrying tokens/cost is counted once.

O-1: how additivity was verified

Not from memory and not from data. This machine has no reasoning-bearing message with a priced model (2 zero-token live-test sessions in opencode.db; the 0.9.10 legacy JSON messages are all lmstudio, all reasoning: 0, cost: 0), so the price cross-check is impossible here. Instead I read getUsage out of the installed OpenCode 1.18.31 binary (strings):

tokens = { input: max(0, inputTokens - cacheRead - cacheWrite),
           output: max(0, outputTokens - reasoningTokens),
           reasoning: reasoningTokens, cache: {write, read} }
cost   = input*p_in + output*p_out + cache.read*p_cr + cache.write*p_cw + reasoning*p_out

So output is stored net of reasoning and reasoning is a separate, additive field, priced at the output rate. Recorded cost already includes reasoning, so observed-cost rows are unchanged; only token totals and the estimated cost of missing-cost rows rise.

Risk I could not close: builds that predate that split stored reasoning inside output. The parser guards it with the message's own tokens.total: a total equal to input + cache + output (no reasoning add-on) means output already contained reasoning, so nothing is added. No total, or one that fits neither shape, falls to the verified current convention. Unit-tested both ways.

Before / after on real data (read-only)

Real data is nearly empty, said plainly:

• ~/.local/share/opencode/opencode.db: 2 sessions, each 1 user + 1 assistant message, all zero tokens, cost 0, no time.completed. Before: each carried a fabricated latency sample (~10 ms created-gap) and rows with no provider. After: 0 latency samples (nothing completed), rows carry provider: lmstudio, nothing else changes (there is nothing to re-price). Both are the mid-turn/abandoned shape O-2 describes.
• Legacy 0.9.10 JSON messages (8 sessions, 17 messages; only numeric/id fields copied into a scratch db, no text): latency histogram before [2,0,0,0,0,0], [2,0,1,0,0,0], [2,0,0,0,0,0] (all sub-2 s) and after [0,0,0,2,0,0], [0,0,0,2,1,0], [0,0,2,0,0,0] (10-30 s and 5-10 s, matching the 8-21 s the JSON shows); engaged time +19-20 s per session; 10 completed all-zero lmstudio responses flagged tokensUnreported (0 before, no signal). Tokens stay 0: those servers reported none.

Schema

SCHEMA_VERSION 24 to 25 with a v25: note. Test that a forged v24 cache is discarded. (The claude-window test's === 24 became >= 24.)

Docs / ADR

ADR-0038 dated correction note (no history rewrite), ADR index status, USAGE-SCORECARD-METRICS (cost, latency, reliability/aborts), TRANSCRIPTS, DASHBOARD, TROUBLESHOOTING, in current-state voice. ADR-0042/0052 do not describe OpenCode counting and are untouched. file:line citations remapped; doc-citations green.

Dropped / deferred

• Session-level provider bucket (byProvider) is still the last observed provider for a session; attributing a mixed-provider session across provider buckets is a larger aggregation change. Rows themselves are now correct.
• Custom-named local providers (an id that is not lmstudio/ollama/llama.cpp/local-openai) cannot be recognised: the transcript records the id, not the endpoint, and the base URL lives in a config file the reader must not open. They keep the fallback rate.
• Out of scope, untouched: O-6, O-9, O-10, O-11, O-12, O-13, any Claude/Codex code, dashboard layout.

Verification

Node 22.22.3 / pnpm 11.17.0, local (macOS). Every fix has a test written before the code (a mutation check on O-1 confirmed the new tests bind it).

• pnpm run check: typecheck, lint (0 errors; no warnings in touched files), lint:cc, lint:md, build all green. pnpm test node:test stage: 4092 tests, 4085 pass, 6 skipped, 1 fail. Coverage: 95.26% lines / 85.64% branches / 93.57% functions (floors 70%).
• The 1 failure, on two full runs, is opencode-stock-ruflo-gateway "stock OpenCode keeps Ruflo and Agentic QE connected..." with spawnSync npm ETIMEDOUT (live npm install). It is the known slow-registry flake; rerun alone it passes (1/1).
• Because that failure short-circuits the && chain, the remaining pnpm test scripts were run separately and pass: statusline-segments, statusline-window-ledger, statusline-brain, agentdb, health-history, harvest, dashboard, admin-model, admin.
• pnpm run test:ui (dashboard client text/logic touched): 491 passed, 0 failed, then 8/8.

Windows pitfalls checked in the new tests: expected paths built with path.join, no bare absolute import(), no POSIX mode-bit or sh child commands, the CLI test sets HOME, USERPROFILE, XDG_CONFIG_HOME, APPDATA and XDG_DATA_HOME, every fixture db is closed before its tmp dir is removed, no CRLF-sensitive assertions, and no real ~/.local/share/opencode or ~/.config touched.

🤖 Generated with Claude Code