fix(e2e): honor E2E_TIMEOUT for cursor-cli and raise the default to 4m

[Soph]

This runner hardcoded a 90s per-prompt ceiling and never read E2E_TIMEOUT, which CLAUDE.md documents as the per-prompt timeout for every agent. Only the hardcoded default and a per-test override applied, so the Cursor leg couldn't be widened from the environment at all. opencode.go:98 already reads it.

90s is independently too tight for current Cursor. On a healthy account with no quota error, one TestSingleSessionManualCommit turn measured 188.7s and passed only once the ceiling was lifted. At 90s the harness tears down tmux mid-turn, so the stop hook never fires and the checkpoint is silently absent — and the assertion names the timeout, not the missing hook, which makes a harness limit read as agent flakiness.

Changes

• Read E2E_TIMEOUT. Precedence now matches opencode.go: default < E2E_TIMEOUT < per-test.
• Raise the default 90s → 4m (~2x the measured turn). No workflow sets E2E_TIMEOUT, so the env knob alone would leave the default in force. A too-high default only costs wall clock when an agent genuinely hangs.

This does NOT fix the currently red cursor-cli leg

Worth being explicit, because I initially misdiagnosed it. The red job's failures are quota, not the timeout. The captured pane carries:

Error: Increase limits for faster responses
You're out of usage. Switch to Auto or Composer 2.5, or ask your admin to
increase your limit to continue.

66 occurrences alongside 68 timeouts in the same run. The agent stops responding and the harness then times out, so the timeout is the downstream symptom. Raising the ceiling will not help a rate-limited account — that needs the Cursor account limit raised.

The change stands on its own merits: the documented env var should work, and 90s is below a healthy turn. But it should not be merged expecting green CI.

Deliberately not addressed

No --model. The runner builds agent --force --workspace <dir> with no model pinned, so Auto routing picks per run and turn duration is non-deterministic by construction. An E2E_CURSOR_MODEL knob mirroring E2E_CLAUDE_MODEL would fix that properly.

Per-test overrides below the measured turn duration — edge_cases_test.go:138 (120s), split_commits_test.go:69,82 (2m), multi_session_test.go (3m). Agent-agnostic, so widening them changes every agent's run time; wanted a maintainer call rather than a silent edit.

Verification

E2E_TIMEOUT=7m mise run test:e2e --agent cursor-cli TestSingleSessionManualCommit → passes in 188.7s, full lifecycle (SessionStart → TurnStart → TurnEnd → checkpoint saved → condensed).

Not re-run across the full Cursor leg — with the account currently out of usage, that run would fail on quota regardless of this change.

mise run check green.

🤖 Generated with Claude Code

[Copilot]

Pull request overview

Updates the Cursor CLI E2E agent harness to honor the documented E2E_TIMEOUT environment variable and increases the default per-prompt timeout so Cursor’s slower auto-routed turns don’t get cut off mid-turn (which prevents the stop hook from firing and leads to missing checkpoints).

Changes:

• Raise Cursor CLI runner default prompt timeout from 90s to 4m.
• Add E2E_TIMEOUT parsing with the same precedence as OpenCode: default < E2E_TIMEOUT < per-test PromptTimeout.

---

💡 Add a code-review agent skill or configure MCP servers for context-aware, tailored reviews. Learn more in the docs.