# Native fork usage evidence

Both sides: Linux web development app, Chrome 157, 1440×1000, light theme, UTC,
30-day Usage window (2026-10-15 through 2026-11-13), OpenCode filter only.
Each app used its own fresh T3 database and isolated OpenCode 2.1.2 data directory.
The workspace contained only a synthetic README and initial git commit. No real
user database was copied. The original reply was generated through T3's CLI and
real OpenCode using anthropic/claude-haiku-5-5. The prompts and native API sequence
were identical on both sides; live-provider cache hits naturally differed.

Replay cannot create OpenCode's native session_v2 fork lineage. T3's app-history
fork is lazy: the native provider fork is materialized when a subsequent turn
starts. To exercise the reported no-new-request native history copies, the driver
called the real OpenCode session fork API twice, then moved the second native
fork to another actual git worktree. It refreshed the real T3 Usage UI after each
operation. It then submitted a genuinely new prompt to that second native session.
No usage rows were synthesized or inserted, and no application source was edited.
The recordings show the Usage UI during the API operations and refreshes; they
are not recordings of clicking T3's fork menu.

The screenshots were accepted only after two frames 500 ms apart were byte-for-byte
identical. All uploaded images and recording contact sheets were visually inspected.
The driver reads only its synthetic provider database, read-only, to retain the
actual lineage and request usage. Head transcripts additionally contain the actual
app Usage query buckets, with source paths removed. Tiny costs round to cents in
screenshots; head's precise app total stayed 0.001347045 through both forks and
rose to 0.00365817 after the new reply.

Base f313d0a6cb843551fb60e57a454a71e71a3a8c04:
15.5K tokens / 1 session → 31K / 2 → 46.5K / 3 without any new request.
The UI cost went from $0.00 to $0.01. A new reply then raised tokens to 65.7K.

Head 4160650b22bb0031efb2fa752bf29951eded14e4:
15.5K tokens / 1 session → 15.5K / 1 → 15.5K / 1 without any new request.
A new reply raised tokens to 34.7K and sessions to 2. Exact app cost increased by
0.002311125, the new native request's reported cost.

The base had discarded setup attempts before recording: an unavailable model,
a successful Luna request that was deleted through OpenCode's API, and an
unsupported old Haiku model that recorded no tokens or cost. The recorded base
baseline contained exactly one assistant record with usage. The head started
with that single successful Haiku request. The native driver and transcripts
make the recorded starting state and every counted record explicit.

Run native-proof.py with `forks` and then `reply`; supply --browser, --session,
--checkout, --run (a fresh timestamped output folder inside that side's DATA),
--worktree, --provider-url and --provider-pid for the provider owned by that app.
It reads that provider's ephemeral password without printing or persisting it.
Use a fresh isolated data directory for each commit; never use a live user home.