fix: improve plugin system robustness — agent/command resolution, async errors, hook timing, two-phase init

[ryanskidmore]

Issue for this PR

Closes #18310

Type of change

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

What does this PR do?

Agent.get() and Command.get() return undefined for unknown names, but five call sites in prompt.ts dereference the result without null checks — producing a raw TypeError instead of a useful error. The prompt_async route drops errors silently because the detached SessionPrompt.prompt() call has no .catch(). Plugin config hooks aren't error-isolated, so one throwing hook kills the rest. The command.execute.before hook fires after template parts are merged with caller parts, making clean template replacement fragile. And plugins added by a config hook aren't loaded in the same startup pass because Plugin.state() is already cached.

Fixes:

1. Agent/command null guards — Added null checks at all five Agent.get()/Command.get() call sites (createUserMessage, loop normal + subtask, shell, command). On missing agent/command, throws NamedError.Unknown with the list of available names and publishes Session.Event.Error. Follows the existing guard pattern already at line 1858 of prompt.ts. Empty-string agent falls back to Agent.defaultAgent() instead of looking up "".

2. prompt_async error handling — Added .catch() on the detached SessionPrompt.prompt() call in the prompt_async route that logs the error and publishes Session.Event.Error. The route still returns 204 immediately — the fire-and-forget semantics are preserved, errors are just no longer invisible.

3. Config hook error isolation — Wrapped each individual hook.config?.() call in Plugin.init() with try/catch so a failing hook is logged but doesn't prevent others from running.

4. Command hook timing — Moved Plugin.trigger("command.execute.before") to fire before template parts are merged with input.parts. Plugins now receive template-only parts in output.parts and can replace template content cleanly. input.parts are merged after the hook returns. Subtask path unchanged. No hook type signature change.

5. Two-phase plugin init — After config hooks fire in Plugin.init(), the code now diffs config.plugin against what was originally loaded, imports any newly-added plugins using the same install/import path as state(), runs their config hooks once, and pushes them into the shared hooks array. One extra pass only — no recursive loading.

How did you verify your code works?

7 new test cases added to existing test files, all passing:

• test/session/prompt.test.ts — unknown agent throws NamedError.Unknown (not TypeError), error includes available names, empty agent falls back to default, unknown command throws with names; command hook receives template-only parts before input merge
• test/server/session-messages.test.ts — structural test verifying .catch() + Bus.publish(Session.Event.Error) on prompt_async route
• test/plugin/auth-override.test.ts — two-phase plugin loading structure verified, config hooks individually error-isolated with try/catch

bun typecheck clean across all 13 packages. bun test passes.

Screenshots / recordings

N/A — no UI changes.

Checklist

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

[github-actions]

Thanks for updating your PR! It now meets our contributing guidelines. 👍

[rekram1-node]

lgtm, just need to resolve merge conflicts