Add corpus-side trace distiller + transcript exporter

[jack-champagne]

Lets us view anyone's agent run from the cloud corpus, without asking them for their SQLite database.

The problem

otlp_export.py produces the clean AGENT→LLM→TOOL traces we actually want to look at — but it builds them by reading opencode's SQLite at ~/.local/share/opencode. That file is on the user's machine, so it can only ever export your own sessions. When Raghav's first telemetry landed today, the only way to get a readable trace of it would have been to ask him to copy his database over.

Why that isn't necessary

The raw wire capture already contains everything: ai.prompt, ai.response.text/reasoning, ai.toolCall.args/result, ai.usage.*, ai.model.id, and the real opencode session id. It's just diluted ~300:1 by opencode's own internal instrumentation — sql.execute alone was 3,653 spans of 18,892 on Raghav's session, plus Session.getPart, SessionProcessor.*, http.server.

So the clean view is a filter over data we already own, not a different data source.

What's added

• corpus_distill.py — corpus *.otlp.gz → OpenInference trace, POSTed as protobuf. Measured 18,918 → 67 spans (282:1) on Raghav's real session.
• corpus_export.py — corpus *.otlp.gz → <session>.md transcript + <session>.spans.json, same format as the SQLite path.

Verified structurally equivalent to the existing pipeline:

	corpus-distilled (Raghav)	otlp_export.py known-good
observations	67	67
types	38 TOOL / 28 GENERATION / 1 AGENT	37 TOOL / 29 GENERATION / 1 AGENT
root	run a x gate locally I want to test…	Optimize X gate on transmon

Identity: the submitter, not the client-reported id

Three user identities are in this data and only one is trustworthy:

source	trustworthy	what it is
submitter (S3 object metadata)	yes	derived by the ingest lambda from sha256 of the bearer token; never client-supplied
ai.telemetry.metadata.userId	no	opencode config value (raghavchari) — spoofable
amicode.user	no	VS Code machineId hash

The scripts prefer --submitter, fall back to the client id, and tag the trace identity:verified or identity:client-reported so provenance is never implicit. All three are recorded as amicode.submitter / amicode.opencode_user / amicode.machine_id.

Note S3 object metadata is not preserved by aws s3 sync, so the submitter has to be carried out-of-band — hence the flag and the _submitter sidecar. The README documents the head-object one-liner.

Langfuse's OTLP mapping was verified empirically against a live instance rather than assumed: langfuse.user.id → userId (a plain user.id is silently ignored), langfuse.session.id → sessionId, langfuse.tags (JSON array string) → tags.

Known fidelity limit, measured

OTEL_ATTRIBUTE_VALUE_LENGTH_LIMIT=16384 clips long values. On a real session:

attribute	intact
ai.toolCall.args	38/38
ai.toolCall.result	35/37
ai.response.text	56/56
ai.response.reasoning	52/52
ai.prompt	1/28 — often invalid JSON

Neither script relies on ai.prompt parsing. ai.prompt is the whole conversation re-sent on every call, so it's quadratically redundant — turn N restates turns 1..N-1, which we already hold untruncated. Both scrape it with a regex fallback and rebuild from the pieces.

Root titles come from a quirk worth documenting: opencode's first LLM call in every session is its own title generator ([system, "Generate a title for this conversation:", <real request>]) and is small enough to escape the cap. Skipping that boilerplate is required, or every root gets named Generate a title for this conversation: — which is exactly what my first pass did.

Genuinely lost content is limited to tool results over 16 KB, marked [TRUNCATED at 16 KB cap] rather than silently dropped. The uncapped fix is the transcript channel on feat/telemetry-transcript-channel, still unshipped.

Verification

Smoke-tested from the committed copies against real prod corpus data: transcript reconstruction (27 steps / 38 tool calls / 2 clipped) and a Langfuse push returning HTTP 200 with userId='raghav', sessionId='ses_052d6d2e…', tags ['distilled','identity:verified','repo:amico','run-corpus'].

.gitignore extended for *.spans.json and _submitter — these outputs carry raw prompts and chain-of-thought, same as the existing exports.

Not in scope

• Deleting the superseded traces I pushed to the local Langfuse while iterating (no clean per-trace delete via the public API).
• Raising the attribute cap, or shipping the transcript channel. Both are the real fix for the 16 KB loss and belong in their own PRs.