fix(heartbeat): isolate heartbeat API client from main streaming session

[tfatykhov]

Summary

• Problem: HeartbeatRunner shared the same AnthropicClient (httpx pool of 5 connections) as the main chat streaming session. When heartbeat cognitive triage fired during active tool execution, both competed for connections — causing pool timeout.
• Fix: Give HeartbeatRunner its own dedicated AnthropicClient with an isolated connection pool via a new AgentRunner.fork() method.
• Review: 3-agent architecture review (architect + concurrency specialist + devil's advocate) + implementation review. All issues resolved.

Changes

File	Change
nous/api/runner.py	Add fork(api_client) method — creates sibling runner with shared cognitive/dispatcher but isolated API client
nous/heartbeat/runner.py	Accept api_client param, create dedicated runner via fork(), route triage through it, cleanup with try/finally
nous/main.py	Create second AnthropicClient for heartbeat
tests/test_runner_fork.py	4 tests for fork() correctness
tests/test_heartbeat_isolation.py	11 tests for routing, cleanup, start() integration, backward compat

Key Design Decisions

• fork() sets _api_shared=True so the forked runner doesn't close the client — HeartbeatRunner.stop() owns that lifecycle
• try/finally blocks in stop() ensure api_client is closed even if dedicated runner cleanup fails
• _get_triage_runner() falls back to shared runner with warning log if start() wasn't called
• Backward compatible: existing callers without api_client param work unchanged

Test plan

• 4 tests: AgentRunner.fork() creates isolated runner, shares cognitive/dispatcher, doesn't close client on close()
• 3 tests: triage routing uses dedicated runner when available, falls back with warning when not
• 4 tests: stop() cleanup closes both runner and client, handles errors, backward compat
• 2 tests: start() calls fork() when api_client provided, skips when not
• 65 existing heartbeat tests pass (no regressions)
• 67 heartbeat lifecycle/intelligent/tuner tests pass

🤖 Generated with Claude Code

[chatgpt-codex-connector]

You have reached your Codex usage limits for code reviews. You can see your limits in the Codex usage dashboard.
To continue using code reviews, you can upgrade your account or add credits to your account and enable them for code reviews in your settings.