inspect-swe: lower inspect_ai floor 0.3.249 → 0.3.233

[gas2own]

What

Lower the solvers/inspect-swe inspect_ai floor from >=0.3.249 to >=0.3.233 in both dependencies and [tool.uv].override-dependencies, regenerate the frozen uv.lock, and correct the misattributing comment.

Why

0.3.233 is the tightest floor that carries both fixes this arm needs:

Fix	First landed
Bridged-tool list[ContentText] result serialization (#3879)	0.3.220
Accept inline role: "system" from bridged Anthropic clients (commit 8829d740 / #4095) — folds into ChatMessageSystem instead of raising RuntimeError: Unexpected message role: system	0.3.233

The previous >=0.3.249 floor over-specified: the old comment attributed the system-role fix to 0.3.249, but it actually landed at 0.3.233. The 0.3.249 changelog (Moonshot provider, inspect ctl control-channel changes, an OpenAI with_raw_response bridge fix) contains nothing this arm requires. inspect_swe==0.2.66 itself declares inspect-ai with no floor, so nothing transitive forces 0.3.249.

Beyond being unnecessary, >=0.3.249 is actively harmful: when astabench is overlaid via --with and pins inspect_ai==0.3.233 (allenai/asta-bench#156), a >=0.3.249 override conflicts and blocks the resolver from landing the intended 0.3.233. Lowering the floor removes the conflict.

Lock

uv lock regenerated — the base project still resolves inspect_ai 0.3.249 (newest satisfying the floor, no astabench group active); the runtime downgrade to 0.3.233 happens when the astabench overlay re-resolves. Only the override/specifier metadata lines changed.

Coordination

Paired with allenai/asta-bench#156 (astabench pin ==0.3.203 → ==0.3.233). Neither alone lands 0.3.233 at runtime — the astabench pin sets the exact version, this floor stops over-constraining it. Tracking + full analysis: allenai/gas2own#346.

🤖 Generated with Claude Code

[gas2own]

⚠️ Draft — paired with allenai/asta-bench#156, which is blocked: astabench also depends on agent-eval[leaderboard]==0.1.51 → inspect-ai==0.3.203, so the solve overlay can't resolve to 0.3.233 until agent-eval relaxes its pin. This floor change is still correct and necessary, but not sufficient on its own. See allenai/gas2own#346.