fix(run): set non-zero exit code on session error

[cokekitten]

Issue for this PR

Closes #26509

Type of change

• Bug fix
• New feature
• Refactor / code improvement
• Documentation

What does this PR do?

opencode run -m <invalid-model> "<prompt>" printed the error but exited 0,
breaking shell scripts that key off $?. Without a prompt the same command
exits 1, so the bug only hits the prompt path.

In run.ts, the non-interactive path runs loop() as fire-and-forget. When a
session.error event arrives, the handler stores the message locally and prints
it — but never propagates it to the process exit code. After the session goes
idle the loop breaks, the event subscription closes, the event loop drains, and
Node exits 0.

Fix: set process.exitCode = 1 in the session.error branch. Same pattern used
elsewhere in the CLI for soft-failure exits (thread.ts:126,222,
attach.ts:54,81). Using exitCode rather than process.exit(1) lets the loop
finish, the subscription tear down, and stdout flush before Node exits — relevant
for --format json consumers.

Refs older duplicates that were never actually closed by a fix: #15558, #17854,
#2489. PR #15787 took a different approach (await loopDone then process.exit(1))
but was auto-closed by the template-compliance bot before any human review.

How did you verify your code works?

Manual repro on dev tree (also reproduces on installed 1.14.41):

```
$ bun --conditions=browser packages/opencode/src/index.ts run -m invalid-model "test"
Error: Model not found: invalid-model/
$ echo $?
1 # was 0 before the fix
```

No-prompt path still exits `1` (no regression):

```
$ bun --conditions=browser packages/opencode/src/index.ts run -m invalid-model </dev/null
Error: You must provide a message or a command
$ echo $?
1
```

`bun run typecheck` passes; `bunx oxlint` reports no new warnings.

Screenshots / recordings

N/A — CLI exit code change.

Checklist

• I have tested my changes locally
• I have not included unrelated changes in this PR

[github-actions]

The following comment was made by an LLM, it may be inaccurate:

Based on my search results, I found one related PR that addresses a similar issue:

Potential Related PR:

• PR fix(cli): set non-zero exit code when run command encounters errors #14625: "fix(cli): set non-zero exit code when run command encounters errors"
  • fix(cli): set non-zero exit code when run command encounters errors #14625
  • This PR also addresses setting non-zero exit codes for the run command, specifically when errors are encountered. It may be handling a similar or related issue to the current PR fix(run): set non-zero exit code on session error #26588.

The current PR (26588) appears to be addressing a specific edge case where session.error events weren't propagating to the process exit code in the non-interactive/prompt path, whereas PR #14625 may have handled a broader set of error scenarios. These could be complementary fixes or addressing different aspects of the same underlying issue.

[cokekitten]

For visibility: #14625 (open since 2026-02-22) takes a different approach to the same bug — it converts loop() into an awaitable, awaits it after the prompt is submitted, and then sets process.exitCode based on the outer error variable. Both fixes produce the same end behavior; this one is smaller (+1 line) and doesn't change the control flow. Happy to defer if maintainers prefer the other approach.