No LLM spans emitted on opencode 1.14.x — event stream renamed (message.updated and message.part.updated never fire)

[evan-burrell]

Summary

Running @braintrust/trace-opencode@0.0.8 against opencode CLI 1.14.48 (the version bundled in cloudflare/sandbox:0.10.1-opencode), the plugin loads correctly and the root OpenCode: … span lands in Braintrust with a duration — but zero LLM spans are emitted for any turn. Braintrust dashboards show LLM calls: 0, Input: —, Output: — for every session.

Root cause

opencode 1.14.x emits a different event stream than the plugin handles. I captured a full event-receive histogram from the plugin's own event hook over a real ~25-minute session (5 opencode sub-sessions, lots of tool activity):

event_type                        count   plugin handler?
─────────────────────────────────  ─────  ─────────────────────────────────────
message.part.delta                  548   NO  (catchall "unhandled event")
session.status                      122   NO  (catchall)
session.diff                         61   NO  (catchall)
file.watcher.updated                 16   NO  (catchall)
file.edited                          16   NO  (catchall)
todo.updated                         10   NO  (catchall)
session.updated                       5   NO  (catchall)
session.idle                          5   YES (closes turn span ✓)
tui.toast.show                        1   NO  (catchall)

message.updated                       0   needed by handleMessageUpdated to emit LLM span
message.part.updated                  0   needed to accumulate llmOutputParts (text + tool calls + reasoning)
session.created                       0   plugin works around this via chat.message lazy-init ✓

Cross-check from the plugin's own I/O ndjson (LOG_TO_FILE=auto):

    784 \"hook\":\"event\"
    222 \"hook\":\"tool.execute.before\"
    220 \"hook\":\"tool.execute.after\"
    112 \"hook\":\"experimental.chat.system.transform\"
     10 \"hook\":\"chat.message\"               # = 5 calls × (input+output)

So chat.message fires correctly (5 opencode sessions × 1 hook call each — the lazy-init root-span path in the chat.message handler runs). What's broken is everything after the turn span gets created:

• dist/index.js's handleMessageUpdated is gated on message.updated, which opencode 1.14.x never sends. The LLM span is emitted from this handler with the message's time.completed, tokens, providerID/modelID, and accumulated text — none of which the plugin can collect.
• handleMessagePartUpdated is gated on message.part.updated, which is also never sent. The plugin's state.llmOutputParts / state.llmToolCalls / state.llmReasoningParts maps stay empty for every turn.

opencode now streams partials via message.part.delta (token-by-token) and never publishes a single "message complete" or "part complete" event the plugin recognises.

Reproduction

1. npm install @braintrust/trace-opencode@0.0.8.
2. Run inside cloudflare/sandbox:0.10.1-opencode (opencode 1.14.48), with the plugin registered in ~/.config/opencode/opencode.json:

   {
     \"plugin\": [\"@braintrust/trace-opencode\"]
   }

3. Set BRAINTRUST_API_KEY, TRACE_TO_BRAINTRUST=true, BRAINTRUST_PROJECT=…, BRAINTRUST_DEBUG=true, LOG_TO_FILE=auto.
4. Start a session, fire any prompt, let it finish.
5. Observe in Braintrust: root span lands, has a duration, but LLM calls: 0.
6. grep 'service=braintrust-trace unhandled event' $XDG_DATA_HOME/opencode/log/*.log | awk '{print $NF}' | sort | uniq -c → reproduces the histogram above.

Suggested fix direction

The minimal path (without changing the public schema in Braintrust):

1. Accumulate output via message.part.delta. Treat the delta.field=\"text\" payload as an append into state.llmOutputParts.get(messageID). Tool-call deltas land via partial JSON in delta too — buffer and parse on completion.
2. Use session.idle (or a synthesized terminal marker per-message) as the LLM-completion event. When session.idle fires for a session that has a currentTurnSpanId, emit one LLM span with the buffered output, using currentTurnStartTime for start and wallClock.now() for end (token counts may be unavailable — set tokens: undefined instead of 0 so the dashboard doesn't show "0 tokens").
3. Stop logging the catchall unhandled event … at INFO — opencode 1.14+ sends thousands of these per session and they're swamping the log file.

I'd be happy to draft a PR if helpful. Reference data and full plugin-decision dump available; happy to share if it'd help triage.

Environment

• @braintrust/trace-opencode: 0.0.8 (latest as of writing)
• opencode CLI: 1.14.48 (from cloudflare/sandbox:0.10.1-opencode base image)
• @opencode-ai/sdk: 1.16.0 (host)
• Runtime: Cloudflare Workers + Sandbox containers, hosting opencode in server mode and driving it via client.session.promptAsync