fix(agents): L1 — crew/agent runtime reliability (parallel gather, per-tool timeout, dead code)

[AVADSA25]

Summary

Four findings from a focused review of codec_agents.py (the self-contained multi-agent runtime). All reliability/correctness — no behavior change for the happy path.

#	Severity	Line	Fix
1	High	:872	Parallel crews no longer sink on one agent's failure — gather now return_exceptions=True
2	High	_execute_tool_with_hooks	Per-tool 120s wall-clock budget — a hanging skill can't hang the agent forever
3	Medium	:838	Dead tautology removed (agent.tools != agent.tools)
4	Low	:72	Dead eager SERPER_API_KEY global removed (Keychain shellout at import, never read)

Detail

1. Parallel crew isolation. asyncio.gather(*coros) had no return_exceptions=True. If one agent raised (e.g. a stuck-detection abort), every other agent's result was discarded and the whole crew failed — while sequential mode isolates each agent with try/except. Parallel now degrades gracefully: a failed agent contributes [Name failed: Type: msg], the rest still return and join, and each failure gets its own agent_finish (outcome=error) audit.

2. Per-tool timeout. Built-in tools have internal timeouts (web_fetch 30s, shell 30s) but arbitrary skill tools invoked via _make_lazy_fn had none — input(), a no-timeout network call, or a deadlock would hang the agent and its default-thread-pool worker indefinitely. The executor call is now asyncio.wait_for(..., 120s); on timeout the agent gets a recoverable error string + a tool_result/timeout audit and its ReAct loop continues. Documented residual: the abandoned worker thread can't be killed (no thread.kill in Python) and stays parked until the blocking call returns — the generous 120s budget keeps that rare and tools doing real work finish well inside it.

3 & 4. Pure cleanup — the tautology collapsed to the existing if stripped: (allowlist scoping unchanged, pinned by test); the SERPER_API_KEY global was never read (web_search fetches its own key via codec_search.search). The _serper_api_key() getter stays (live reads + satisfies the test_remaining_secrets source scan).

Test plan

• tests/test_agents_reliability_l1.py — 7 tests: parallel degrade vs all-succeed, hanging-tool-bounded (measures the coroutine's return at the budget, not the orphaned thread's 2s finish), fast-tool-still-works, allowlist-strips + no-allowlist-keeps-all, dead-global-gone
• Existing 236 agent/crew tests still pass (no regression)
• python3.13 -m pytest --ignore=tests/test_skills.py -q → 2,062 passed, 77 skipped
• ruff check: 0 issues

Branches off main; touches only codec_agents.py — independent of the other queued PRs.

🤖 Generated with Claude Code