Skip to content

Advisors: Advise, Proposal, the lookahead as the first advisor (carries the actors-seats merge) - #134

Merged
ptescher merged 5 commits into
masterfrom
actors-advisors
Nov 15, 2026
Merged

ptescher merged 5 commits into
masterfrom
actors-advisors

Conversation

@ptescher

Copy link
Copy Markdown
Owner

Step 4 (first half) of plans/actors.md: advisors in the rules crate and the fast arena.

This branch also carries the merge of actors-seats (PR 131) into actors-roster, as its first commit. One conflict: domains/teleoperation/command/src/seat.rs (the queue branch added queue: vec![] there, the seats branch turned it into a re-export of the moved file). Resolved by keeping the re-export and carrying the queue edit into towers_ground/src/seat.rs. plans/actors.md merged with both Results notes. The computer commander in towers_ground::commander is now driven as a Command and sends each directive's queue over the seat (Seat::order_with_queue). World::command_standing became test-only (nothing calls it after the merge).

What changed

  • recreation_towers::advice: trait Advise { fn advise(&mut self, view: &TeamView, hand: &Hand) -> Vec<Proposal>; fn clone_box(&self) -> Box<dyn Advise>; }, Proposal { drone, order, queue, gain, made, until, why }, Hand { orders, free, away } (what the commander has in hand and the map does not show) and Lookahead, the old lookahead moved out of Commander with the same model, cadence and numbers.
  • Commander holds advisors with a delay in game time (with_advisor, add_advisor). Each decision it takes the best proposal that is not stale, gains more than ADVICE_MARGIN (2 points, a constant: CommanderTuning is the league's candidate format) and still fits (drone free, tower not given up, and for a proposal made on an earlier map not a tower the team holds now). Commander::advise puts proposals in hand from outside. A taken proposal's queue goes out with the directive. with_lookahead(false) is a commander with no advisor.
  • Arena: Roster::with_advisor(team, advisor, delay), Command::advised_by, Captain::advised_by. The delay is game time, no threads.
  • Expectations: 18-advice.expect (taken, small gain, stale, held now, no advisor), vocabulary in 09-towers/section.expect, glossary terms Advisor and Proposal, a pointer in 05-commander.expect. cargo run -p expectation -- check expectations: 0 errors. In the recreation_towers acceptance driver the suite's coverage mode reaches the first three; the last two only arranged and visited, so tests/acceptance.rs has mod advice playing all five by hand.
  • Example cognition_towers --example advice.
  • The text Advisor is unchanged; it carries no order or gain, so it is not an Advise.

Not done

The seat side (proposals over the wire, the app's accept), a queue in the lookahead's proposals, the Claude advisor.

Gate: same results

denial, ruleset v2 200 and judge --game captain --champion timid --challenger default --seeds 24 were saved before and after, with SIM001_LOOKAHEAD unset and 0: all six pairs byte-identical. Only judge and the "default against default" line of ruleset use the lookahead (Blue wins 50 % with it, 68 % without); denial and the other policies play with it off.

Gate: what advice is worth

cargo run --release -p cognition_towers --example advice 400: commander + lookahead advisor at the delay, against a commander with no advisor, 400 seeds, both sides (800 games each).

Delay Margin Won Lost
0 s -0.47 ± 0.04 268 532
2 s -0.46 ± 0.04 272 528
5 s -0.44 ± 0.04 276 524
10 s -0.28 ± 0.05 333 467
30 s -0.81 ± 0.04 182 618

By the default rules the lookahead is worth less than nothing in the arena, at every delay. Lateness up to 5 s changes nothing measurable; 30 s is worst. The reason is not measured.

Tests

cargo test --release -p recreation_towers -p cognition_towers -p recreation_towers_ground, -p recreation_towers_server --bins and -p cognition_league --lib pass; cargo build --release -p recreation_towers_server -p teleoperation_command builds; fmt and clippy clean in the repo's code. cognition_league's tests/acceptance.rs was not run (known to time out).

🤖 Generated with Claude Code

ptescher and others added 5 commits November 14, 2026 08:34
The seat client and the station move from the commander app into recreation_towers_ground, with
a headless computer commander (commander::run, binary towers_commander). The server no longer
decides or sends plans: it starts a computer commander as a seat client for a computer team.
A seat changes hands during a game: a person relieves a computer, and a computer takes a seat a
person leaves. Two people can play.

Co-Authored-By: Claude <noreply@anthropic.com>
…d reckoning; crash log lines

Measured on SITL: the first version asked for AGENT in the middle of the take-off transition (two
games lost a drone at its base), held its clock to one game second past the last map (slow plans at
10x) and planned from the position of the last report. The server prints what a crashed drone was
doing and who was near.

Co-Authored-By: Claude <noreply@anthropic.com>
Co-Authored-By: Claude <noreply@anthropic.com>
The computer commander in towers_ground now implements the Command trait's
directive and sends each directive's queue over the seat.

Co-Authored-By: Claude <noreply@anthropic.com>
The commander takes advice (not stale, gain over 2 points, still fits); a
roster can give a team an advisor with a delay in game time. Expectations
18-advice, an advice example, and the results in plans/actors.md.

Co-Authored-By: Claude <noreply@anthropic.com>
@ptescher
ptescher changed the base branch from actors-roster to master November 15, 2026 02:48
@ptescher
ptescher merged commit 2f2dd96 into master Nov 15, 2026
2 checks passed
@ptescher
ptescher deleted the actors-advisors branch November 15, 2026 05:36
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant