[silvervest]

Merge order: any time: independent of the other PRs of this set (#178-#185).

New feature.

Two rules ask agents of two models to agree, so one model's blind spot isn't taken for proof: shared knowledge, and the verdict that no C makes a function. Agents of one model never get there, and nothing told the user it was theirs to settle.

• Knowledge: on an x86 game worked by one model only, 55 proposals waited for a second model that never came, and every session found the same things out again. The knowledge panel now offers Give it to agents and Reject on proposals with one backer too. decide() already took any proposal; its docstring, the reply to a proposal and the prompt's heading ("or checked by the user") now say so.
• No C makes it: on the Xbox demo, whose last open function is its hand-written _ftol2, two agents of one model claimed it a dozen times between them, each time writing the same note, then failed session after session until the five-failure pause. The claim order now skips a function the agent itself said no C makes (another model's agent can still take it), and nothing_to_claim says which ones wait on that verdict. When every open function does, the agent is told to stop, the user gets one notification per function (set it aside in its panel, or let another model look), and a helper with nothing else stops before starting a session, saying so.

Tests: test_the_user_can_be_the_second_opinion (tests/test_knowledge.py), test_an_agent_isnt_handed_back_what_it_said_no_c_makes_and_the_user_hears_when_only_that_is_left (tests/test_claims.py), test_a_helper_stops_when_all_that_is_open_is_what_it_said_no_c_makes (tests/test_agents.py). ruff is clean.

🤖 Generated with Claude Code

[silvervest]

Rebased onto main, with one new commit from my own review: "Only an agent whose own verdict is on every open function is told to stop". "Stop working on this project" went to any agent that found nothing to claim while every open function had a no-C verdict from anyone. That included an agent of another model that had only released them for now, so the agent whose verdict could set them aside quit. It's now said only when every open function has the agent's own verdict. The test adds that case, and it fails on the previous head.

[RetroGameDeveloper]

I reviewed head 58763e5dcf0ab7523e65d0fc2658cd1a7ce09e64. I'd request one fix before merging:

[P2] The new verdict handling doesn't reach room guests. The host's /collab/claim handler calls claim_row() directly. When the new filter excludes a function the guest already ruled out, it returns None without calling nothing_to_claim(). This bypasses both the verdict explanation and the user notification.

New handling being bypassed: https://github.com/silvervest/decompdecomp/blob/58763e5dcf0ab7523e65d0fc2658cd1a7ce09e64/src/decompnet/network.py#L638

I reproduced this with separate host/guest databases and a stand-in room link that delegates to the host's claim_row(), matching the current room handler:

• The last function remains open with the guest's no-C verdict.
• The guest receives “Nothing left to claim in this room — every open function is matched or someone is on it. 🎉”
• The host records no verdict notification.

The room protocol should carry the host's verdict-aware response, with a regression test. Both the ROM room path (engine/judging.py) and project room path (forkengine.py) currently turn the empty claim into the generic response. This falls under the repository's requirement that features reach rooms.

Otherwise, the knowledge controls and local claim handling look sound. All six CI checks pass; locally, 83 focused tests passed, and lint passed. The PR merges cleanly.