fix(agent): Surface real session-init errors and kill leaked CLIs

[charlesvien]

Problem

Connecting to an agent could fail with a generic "Internal error" that named no cause, retried every ~30s, and left orphaned claude subprocesses behind. The real failure — a 30s session-init timeout in ClaudeAcpAgent.createSession — was thrown as a plain Error, which the ACP layer collapses into a JSON-RPC -32603 "Internal error" (real text buried in data.details) before it reaches the logs or UI.

Changes

1. Throw a descriptive RequestError (not a plain Error) at the three session-init timeout sites so the real message survives the ACP/tRPC boundary instead of becoming "Internal error"
2. Abort the timed-out query (abortController.abort() + query.close()) so the claude subprocess can't leak and pile up under the retry loop
3. Surface data.details in the host-side "Failed to create session" log so the exported log names the actual cause for any other masked error
4. Bound the gateway /v1/models fetches with AbortSignal.timeout(10s) so a stalled gateway can't stall session init

How did you test this?

• pnpm --filter @posthog/agent test — 709 passing, including a new refresh-timeout test (asserts a RequestError is thrown and the timed-out query is closed) and a fetchGatewayModels timeout test
• @posthog/agent and @posthog/workspace-server typecheck — clean
• biome lint on the 5 changed files — clean

Automatic notifications

• Publish to changelog?
• Alert Sales and Marketing teams?

[charlesvien]

• fix(analytics): Resolve feature flags before first render #2656
• fix(auth): Auto-recover project-less session after re-auth #2655
• fix(agent): Surface real session-init errors and kill leaked CLIs #2654 👈 (View in Graphite)
• main

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

[github-actions]

React Doctor found no issues in the changed files. 🎉

_{Reviewed by React Doctor for commit 17644d6.}

[greptile-apps]

Prompt To Fix All With AI

Fix the following 1 code review issue. Work through them one at a time, proposing concise fixes.

---

### Issue 1 of 1
packages/agent/src/gateway-models.test.ts:112-138
**`fetchModelsList` timeout not covered by tests**

`fetchModelsList` received the same `AbortSignal.timeout` fix as `fetchGatewayModels`, but only `fetchGatewayModels` has a test verifying the timeout degrades to `[]`. A stalled gateway would leave `fetchModelsList` hanging silently in production and the fix would go undetected if it regressed. The two functions are structurally identical, so a single parameterised test covering both would be cleaner and complete coverage.

_{Reviews (1): Last reviewed commit: "surface real agent init errors and kill ..." | Re-trigger Greptile}

[github-actions]

Clean bug-fix: terminates leaked CLI processes on session-init timeout, surfaces real errors through the ACP layer, and bounds gateway fetch calls. No showstoppers; the bot's missing-test observation is a minor P2 coverage gap, not a blocker.

New commits pushed (delta classified non_trivial_delta) — stamphog approval dismissed; re-review running automatically.

[github-actions]

Clean bug-fix with test coverage for all changed paths; the process-leak and error-masking fixes are correct, the gateway fetch timeout is properly bounded, and the resolved bot comment about fetchModelsList coverage is directly addressed by the new parameterized test.