cold-start: lazy-load agent runner heavy dependencies

[x0ba]

Stack Context

This stack reduces Vercel cold-start time by shrinking the shared SvelteKit server boot path and isolating heavy agent/GitHub dependencies.

Why?

The agent runner statically imported Playwright, Browserbase, and tool-agent provider SDKs at module load. Dashboard routes that only read run snapshots were paying that cost. Heavy modules now load only when the matching run mode starts.

Test plan

• Start a tool-agent run and confirm it completes
• Start a browser run and confirm Browserbase session creation still works

[vercel]

The latest updates on your projects. Learn more about Vercel for GitHub.

Project	Deployment	Actions	Updated (UTC)
flightlog	Ready	Preview, Comment	May 28, 2026 4:53pm

[x0ba]

• cold-start: tune Sentry sampling and Vercel function limits #45
• cold-start: split Neon and pg database drivers #44
• cold-start: lazy-load agent runner heavy dependencies #43 👈 (View in Graphite)
• cold-start: split regression execution into executor module #42
• cold-start: decouple webhook registration from server hook #41
• staging

This stack of pull requests is managed by Graphite. Learn more about stacking.

[greptile-apps]

Greptile Summary

This PR lazy-loads four heavy module dependencies (./tool-agent, ./browser, ./openai, $lib/server/openai-transport) using dynamic import() so that dashboard routes and other SvelteKit server paths that never start an agent run no longer pay the cold-start cost of loading Playwright, Browserbase, and provider SDKs at module initialisation.

• The runToolAgent import is moved inside the existing tool_agent try/catch (line 155), so any import failure is caught and surfaced via failRun.
• The three browser-path imports (lines 179–181) are placed outside the browser try block (line 234), meaning a module-load failure there bypasses failRun and leaves the run stuck in the database without cleanup.

Confidence Score: 4/5

The change is safe for the tool-agent path, but the browser path has an unguarded window where a dynamic import failure leaves a run permanently stuck in the database.

The runToolAgent dynamic import is correctly placed inside its try/catch so errors are handled gracefully. The three browser-path imports (lines 179–181) are outside the browser try block: if any throw, failRun is never called, the run record stays in a bad state, and the rejection propagates to the void-discarded promise with no cleanup. This is a real gap introduced by the refactor, not a pre-existing condition.

src/lib/server/agent-runner/service.ts — specifically the three browser dynamic imports at lines 179–181 that fall outside the browser try/catch block.

Important Files Changed

Filename	Overview
src/lib/server/agent-runner/service.ts	Converts four static imports to dynamic imports for cold-start savings; tool-agent import is correctly guarded inside try/catch, but the three browser-path imports land outside the browser try block, leaving runs in a broken state if any import fails.

Flowchart

%%{init: {'theme': 'neutral'}}%%
flowchart TD
    A[runAgent called] --> B{runMode === tool_agent?}
    B -- Yes --> C["try { ... }"]
    C --> D["await import('./tool-agent') ✅ inside try/catch"]
    D --> E[runToolAgent]
    E --> F[return]
    C -- catch --> G[failRun]
    B -- No / browser --> H["await import('openai-transport') ⚠️ outside try"]
    H --> I["await import('./browser') ⚠️ outside try"]
    I --> J["await import('./openai') ⚠️ outside try"]
    J --> K[credential checks]
    K --> L["try { ... }"]
    L --> M[createBrowserSession → agent loop]
    L -- catch --> N[failRun]
    H -- throws --> O["❌ no failRun, run stuck in DB"]

Loading

_{Reviews (2): Last reviewed commit: "Lazy-load browser and tool-agent modules..." | Re-trigger Greptile}

[x0ba]

Merge activity

• May 28, 4:49 PM UTC: A user started a stack merge that includes this pull request via Graphite.
• May 28, 4:53 PM UTC: Graphite rebased this pull request as part of a merge.
• May 28, 4:53 PM UTC: @x0ba merged this pull request with Graphite.