fix: parse nested JSON in LLM responses + runtime bug fixes

[lbartoszcze]

Summary

Fixes 4 runtime bugs that affect every agent running on Singularity, with 57 tests proving correctness.

1. _parse_action silently drops params on nested JSON (CRITICAL)

The bug: The regex r'\{[^{}]*"tool"[^{}]*\}' in cognition.py:_parse_action() uses [^{}]* which rejects any JSON containing nested braces. When the LLM returns the common case:

{"tool": "shell:bash", "params": {"command": "git status"}, "reasoning": "check repo"}

...the regex fails to match because {"command": "git status"} contains braces. The fallback strips ALL params, so the agent knows what to do but loses how to do it.

The fix: Replaced regex with balanced-brace extraction (_extract_tool_json) that scans for {, counts depth to find the matching }, and attempts json.loads. Correctly handles nested objects, arrays, and braces in strings.

19 tests cover flat JSON, nested params, deeply nested objects, markdown code blocks, arrays, missing keys, invalid JSON, and fallback paths.

2. Vertex AI uses sync client in async think() (HIGH)

The bug: cognition.py imports AnthropicVertex (sync) instead of AsyncAnthropicVertex, and calls self.llm.messages.create() without await in the async think() method. This blocks the entire event loop during inference.

The fix: Import AsyncAnthropicVertex, add await to the Vertex call path.

3. Orchestrator _message() / _broadcast() KeyError (HIGH)

The bug: Both methods access _message_boxes[agent_id] without checking if the key exists. If an agent's message box wasn't pre-initialized (e.g., loaded from a previous session), this crashes with KeyError.

The fix: Guard both methods to create the queue on demand: if agent_id not in _message_boxes: _message_boxes[agent_id] = asyncio.Queue().

7 tests cover missing boxes, existing boxes, non-existent agents, self-skip, and dead agent filtering.

4. request.py Linear API call has no timeout (MEDIUM)

The bug: aiohttp.ClientSession() in _create_linear_ticket() has no timeout, risking indefinite hangs if the Linear API is unreachable.

The fix: Added aiohttp.ClientTimeout(total=30).

Files changed

File	Change
singularity/cognition.py	New _extract_tool_json() method, AsyncAnthropicVertex import, await on Vertex call
singularity/skills/orchestrator.py	Guard _message_boxes access in _message() and _broadcast()
singularity/skills/request.py	Add 30s timeout to aiohttp session
pyproject.toml	Add pytest-timeout to dev deps, pytest config
.github/workflows/ci.yml	CI workflow (lint + test on 3.10/3.11/3.12 + build)
tests/test_parse_action.py	19 tests for JSON parsing
tests/test_orchestrator_messaging.py	7 tests for messaging
tests/test_cognition_engine.py	14 tests for engine init, cost calc, prompts
tests/test_smoke.py	17 smoke tests for imports and skill instantiation

Test plan

• 57 tests all pass locally (pytest tests/ -v --timeout=30)
• Lint passes on modified files (ruff check --select F821,F811,F601)
• Package builds and installs cleanly

🤖 Generated with Claude Code