feat(usage): mirror token usage where usage reporters read it

[filipeforattini]

Token accounting lived only inside the session store, so nothing outside redcode could report on it — and that store is ours to change. This writes it a second time, in two places, carrying only what a report needs: tokens, cost, model, provider and timestamps, one row per assistant turn, and never a prompt, a tool result or a file path.

The two targets

Our sidecar — ~/.red/code/data/usage/opencode.db, in OpenCode's layout, which ccusage reads when pointed at the directory:

OPENCODE_DATA_DIR=~/.red/code/data/usage ccusage opencode daily

It is ours, so it survives whatever the session store becomes. That was the point of doing this at all: when storage moves to RedDB, this file is the contract that stays, and no reporting tool ever needs a redcode-specific adapter or a dependency on our client.

The fan-out into OpenCode's own database, when one exists, so ccusage opencode finds redcode usage with no configuration. OpenCode's message rows are keyed to session, and session to project, so the fan-out writes those two as well. Every id it introduces is marked — prj_redcode_…, ses_redcode_…, redcode_… — so its rows are recognizable and removable by hand. Session and project go in with INSERT OR IGNORE: this adds rows and never edits what OpenCode owns. The file is opened read-write and never created, so a machine without OpenCode gets no fan-out and no invented database.

Why not a PR to ccusage instead

ccusage already ships an OpenCode adapter, and its record shape is ours (redcode forked OpenCode's message layout). Only path discovery differed. Its CONTRIBUTING.md also gates PRs behind an approved issue, and a redcode adapter there would have to be rewritten the day storage moves off SQLite — while a client dependency on RedDB (BSL) inside an MIT CLI is not something to ask a maintainer for. Writing the format ourselves needs nobody's approval and cannot rot upstream.

Verification

• Against a copy of a real OpenCode database (1,783 existing messages): PRAGMA foreign_key_check clean afterwards, existing rows untouched, and the published ccusage reporting redcode's days beside OpenCode's own.
• End to end with the published ccusage: daily, session, monthly and weekly all read the sidecar correctly.
• packages/core/test/usage.test.ts pins the layout, the record shape, the absence of content, the upsert on re-publication, both opt-outs, and the fan-out's foreign keys against OpenCode's real schema (copied into test/fixtures/opencode-schema.sql).

Costs accepted deliberately

Redcode's sessions become visible in OpenCode's own session list (empty, since no parts are written), and deleting one there takes its usage rows with it through ON DELETE CASCADE — the sidecar keeps its own copy. REDCODE_DISABLE_USAGE_FANOUT=1 turns off the second target; REDCODE_DISABLE_USAGE_SIDECAR=1 turns off both.

Not included: a backfill of the messages already in redcode.db. The mirrors start from the moment this lands.

https://claude.ai/code/session_01MyZkjLGy7k8uQdRWzU3aSB

---

^{Need help on this PR? Tag @codesmith-bot with what you need. Autofix is disabled.}