…-resolution deck counts (#6194) * fix(ai-commander): surface driver stalls instead of exiting as normal game end #6080 diagnosis: an empty run_ai_actions batch while waiting_for != GameOver was silently treated as a normal game end, hiding the whole p0-softlock issue family (#5250, #4345, #5958, #6172, #3886, #3919, #3233). The only prior signal at any of run_ai_actions' three break doors (no AI actor, choose_action None, apply() Err) was a tracing::error/warn that no harness subscriber captures. run_ai_actions now returns AiActionsRun (Deref<Target = Vec<AiActionResult>> for source compatibility with existing callers) carrying an Option<AiActionsBreakReason> naming which door fired. ai_commander prints a machine-readable STALL line (waiting_for variant, turn, active/priority player, pending-cast summary) plus the break reason, and exits with a distinct code (3) instead of falling through to a clean exit — distinguishing a stall from both a completed game (0) and an action-cap abort (2). The existing "Game did NOT reach GameOver" substring pod-lab's runner.py matches on is preserved verbatim. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> * fix(ai-commander): hard-fail unknown --difficulty labels instead of silently downgrading parse_difficulty previously matched only five exact-case strings and fell back to Easy for anything else, including typos and CEDH (which has no arm here yet) — silently poisoning an entire batch of games with the wrong skill tier and no indication anything went wrong. Routes the actual label -> AiDifficulty mapping through the crate's existing single-authority AiDifficulty::from_label (so this binary's understanding of a label can never drift from every other transport that parses one), but layers a hard startup error on top for anything from_label would otherwise silently downgrade to Medium. Accepted values are listed explicitly (VeryEasy, Easy, Medium, Hard, VeryHard, CEDH) so a wrong flag names every valid spelling instead of just failing quietly; adding CEDH here is what unlocks the CEDH tier for pod-lab's gauntlet plan (§3.8). Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> * feat(ai-commander): print per-seat resolved deck counts after name resolution resolve_deck_list silently skips any card name the database doesn't recognize, and the only existing print (pre-resolution, from the raw feed JSON) happens before that resolution runs — so no prior output could prove a deck actually loaded full-size (plan §3.9). A harness driving this binary at scale had no way to detect silent-skip drift short of re-deriving counts from engine internals. Adds a per-seat "Resolved deck sizes (post name-resolution...)" block right after resolve_deck_list, summing each PlayerDeckPayload's main_deck and commander DeckEntry counts. Purely additive — the existing pre-resolution print is untouched, and pod-lab's runner.py doesn't currently parse either line, so this carries no stdout-compatibility risk. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> * fix(ai-commander): capture break_reason from every AI-action batch, not only empty ones `run_ai_actions` can complete one or more actions and still stop on ApplyFailed/ChooseActionNone before the safety cap. The driver previously only inspected `break_reason` when the returned batch was empty, silently discarding the diagnostic whenever a batch was non-empty and looping again until a later, unrelated empty batch produced a misleading NoActor/unknown reason in the stall report. Extract `driver_step` in auto_play.rs as the single place that decides whether a batch's break_reason should stop the driver, and use it in ai_commander's loop. Add regression coverage: a unit test on `driver_step` itself, and an integration test forcing `run_ai_actions` to return a real non-empty batch with a break_reason (DeclareBlockers applies, then priority reaches an AI seat with no registered config). Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> * fix(ai): split MissingAiConfig out of NoActor in AiActionsBreakReason The missing-configuration break door recorded `NoActor`, so `ai_commander` reported a caller wiring gap (an AI seat in `ai_players` with no `ai_configs` entry) as "no actor" — two stalls with different remediation printed as one. Adds `MissingAiConfig { player }` and records it at that door. `NoActor` stays fieldless and now documents the two causes it still covers: `acting_players()` returned empty (`GameOver`, or an empty pending set), or it returned players none of whom map to an AI seat. It carries no `PlayerId` because the first cause has no player at all, and `MulliganDecision`/`OpeningHandBottomCards` can pend several at once, so naming one would be arbitrary. The regression expectation now asserts `MissingAiConfig { player: P1 }` via `matches!`, which pins the missing-config door specifically — `NoActor` has no field and cannot match it. Also addresses two bot notes: corrects the `ApplyFailed` doc comment (only `action` is boxed; `EngineError` is four variants with a `String` at worst, so `large_enum_variant` does not fire on it), and replaces hardcoded `ai_decks[0]`/`[1]` indexing in the deck-count print with seat iteration. Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com> --------- Co-authored-by: Claude Fable 5 <noreply@anthropic.com> Co-authored-by: Matt Evans <1388610+matthewevans@users.noreply.github.com>

Uh oh!

Uh oh!

Stuck decision: ReplacementChoice #3886

Description

Metadata

Metadata

Assignees

Labels

Type

Projects

Milestone

Relationships

Development

Issue actions