opencode run --format json drops text and step-finish events in containerized environments

[jangel97]

Description

opencode run --format json produces incomplete JSONL output in containerized environments. Only step_start is emitted — text and step_finish events are lost.

Root Cause

The loop() function in run.ts breaks immediately when it receives a session.status = idle event. In containerized environments (higher event delivery latency), this idle event races ahead of text and step-finish part events in the SSE pipeline, causing the loop to exit before all events are delivered.

Steps to Reproduce

Run inside a container:

OPENCODE_EXPERIMENTAL=true opencode run --format json -m "google-vertex-anthropic/claude-sonnet-4@20250514" "Say hello in one sentence"

Actual output (only step_start):

{"type":"step_start","timestamp":1780961221439,"sessionID":"ses_...","part":{"type":"step-start"}}

Expected output (all three events):

{"type":"step_start","timestamp":...,"sessionID":"ses_...","part":{"type":"step-start"}}
{"type":"text","timestamp":...,"sessionID":"ses_...","part":{"type":"text","text":"Hello! ..."}}
{"type":"step_finish","timestamp":...,"sessionID":"ses_...","part":{"type":"step-finish","tokens":{...}}}

Environment

• OpenCode run inside a Linux container (podman/docker)
• Provider: Vertex AI (Google Cloud), but likely affects all providers
• The issue does not reproduce on macOS natively, where event delivery is faster

Proposed Fix

Track active step lifecycle with a counter. When idle arrives while steps are still open, defer the break and drain remaining events with a safety timeout.

See PR #31434.

[github-actions]

This issue might be a duplicate of existing issues. Please check:

• CLI: JSON stream output truncated when process exits quickly (stdout buffer not flushed) #29866: Very similar symptom — opencode run --format json only emits step_start, missing text and step_finish events. That issue attributes the truncation to stdout not being flushed before process exit, while this issue attributes it to an SSE event ordering race in containerized environments. The root causes may differ, but the symptoms and affected code path overlap significantly.

[jangel97]

The race condition is easily reproducible using the official container image. Running the same command multiple times shows non-deterministic event delivery.

Setup

podman run --rm \
  -v "${HOME}/.config/gcloud:/root/.config/gcloud" \
  -e GOOGLE_APPLICATION_CREDENTIALS="/root/.config/gcloud/application_default_credentials.json" \
  -e GOOGLE_VERTEX_PROJECT="my-project" \
  -e GOOGLE_VERTEX_LOCATION="global" \
  -e OPENCODE_EXPERIMENTAL=true \
  ghcr.io/anomalyco/opencode:latest \
  run --format json -m google-vertex-anthropic/claude-sonnet-4-6@default "Say hello"

Upstream image — 5 consecutive runs

=== Run 1 === empty (zero events)
=== Run 2 === empty (zero events)
=== Run 3 === empty (zero events)
=== Run 4 === empty (zero events)
=== Run 5 === empty (zero events)

0/5 runs produced any output. The session.status = idle event fires before any part events reach the client iterator. Exit code is 0 despite producing nothing.

With fix from PR #31446 — 5 consecutive runs

Built from the jangel97/opencode fork (dev branch) and run in a container with the same provider/model:

=== Run 1 === 3 events -> [step_start, text, step_finish]
=== Run 2 === 3 events -> [step_start, text, step_finish]
=== Run 3 === 3 events -> [step_start, text, step_finish]
=== Run 4 === 3 events -> [step_start, text, step_finish]
=== Run 5 === 3 events -> [step_start, text, step_finish]

5/5 runs produced all three events consistently.

Occasional partial output

In some runs, the upstream image emits partial results — step_start and text arrive but step_finish is lost:

{"type":"step_start","timestamp":1780992719574,"part":{"type":"step-start"}}
{"type":"text","timestamp":1780992719700,"part":{"type":"text","text":"Hello!"}}

This confirms the race is timing-dependent: idle fires at different points in the SSE pipeline depending on scheduling and load.