Draft — GitHub issue for anomalyco/opencode

Title: SSE /api/event fibers interrupted mid-turn by provider/model config reload triggered from other connected directories (v2.0.26, Windows)

Environment

• opencode: v2.0.26 (@opencode/cli via npm global, Bun runtime) on Windows 11
• Architecture: one shared background server on a fixed port, serving multiple clients via ?directory= workspaces (6 distinct directories connected concurrently)
• Providers: default + one custom OpenAI-compatible provider with a 16-model whitelist (relevant because config reloads are frequent)

Summary

When provider/model configuration is reloaded on the shared server — apparently triggered by clients connecting from other directories — in-flight SSE streams on /api/event for unrelated, active sessions are torn down with InterruptError: All fibers interrupted without error. The HTTP response itself completes with status 200, but the event stream simply stops mid-turn. CLI clients running opencode run --format json then exit silently (no terminal event), which downstream tools interpret as a crash and retry.

Observed behavior (2026-10-10, all times UTC)

Server log (~/.local/share/opencode/log/opencode.log) shows bursts of:

timestamp=2026-10-10T07:00:16.095Z level=INFO run=bf476822 cause="InterruptError: All fibers interrupted without error\n at ServerProcess.start (B:/~BUN/root/chunk-gn3byahv.js:2:7477)..." http.span=15810 http.method=GET http.url=/api/event http.status=200

• 22 such interrupts between 06:4x and 07:0x on one hour alone.
• Each affected GET /api/event has a large http.span (e.g. 15810 ms) and still returns http.status=200 — the interruption happens while the SSE response is streaming.
• The server process does not restart (stable PID across the whole window, 19 active connections) — this is not a watchdog/service-replacement issue.

Correlation on the client side (a desktop tool spawning opencode run --format json --session <id> per task):

cli starting ... args=["run","--format","json","--dangerously-skip-permissions","--model","...","--session","ses_edb74a31dffedlcMIDfbjAJjWZ"]

followed by no further output (silent death). The same session was resumed 16 times between 06:41 and 07:00 — each resume starting a fresh CLI that again hit an interrupted stream — burning tokens each time.

The server log in the same period is dominated by config-churn events (provider:updated etc., on the order of 1300+ events/day across the connected directories), consistent with ScopedCache invalidation racing in-flight request fibers.

Impact

• Any multi-client setup where one workspace's config reload can kill unrelated active streams.
• Silent CLI exit gives callers no usable error to distinguish "stream interrupted" from "model/provider failure", causing retry storms in wrapping tools.

Suspected mechanism

Same class as #32755 (ScopedCache invalidation → instance disposal interrupts all in-flight lookup fibers): a provider/model config reload — here triggered by other ?directory= clients connecting/reconnecting — disposes shared instances and interrupts fibers of in-flight SSE requests, including /api/event streams for sessions in unrelated directories. Differs from #52049 / #51216 in that the server is not restarted/replaced; differs from #48907 in that no skill loading is involved.

Minimal reproduction sketch

1. Start shared server, connect client A with directory DA and an active long-running session streaming /api/event.
2. From client B (directory DB), trigger provider config reload (e.g. connect/reconnect, or touch provider config).
3. Observe A's /api/event stream terminate with InterruptError ... http.status=200 and the CLI exit without a terminal event.

Secondary observation (Windows-specific)

~38 occurrences of unhandled rejections like:

EPERM: operation not permitted, rename '...session.json.<pid>.<uuid>.tmp' -> '...session.json'

when multiple processes write shared session state concurrently — suggests atomic-rename on shared state files needs retry/serialization under Windows.

Related issues

• bug: InterruptError — All fibers interrupted when InstanceStore.dispose races with Agent.list #32755 — same InterruptError mechanism (CLOSED, but clearly still reproducible)
• Skill tool fails with "All fibers interrupted without error" on Windows #48907 — same error, skill-loading trigger
• cli(win): 45s event-stream idle watchdog restarts the managed service, aborting all sessions/subagents #52049, desktop(win): background service SIGKILLed mid-turn when snapshot capture blocks /api/info for >6s -> ~60s restart loop, aborted turns #51216 — service replacement/SIGKILL restart loops (different trigger: server actually restarts there; here it does not)