cassandra-alpha/feedback
2
0
Issues 8 Activity

Agent is not opting out of auto return to command zone well #31

New Issue
Closed
opened 2026-07-09 23:27:25 +00:00 by bstein · 1 comment
bstein commented 2026-07-09 23:27:25 +00:00
Owner
Copy Link

What happened

image.png Commander gets TEMPORARILY exiled (would return at next endstep) and instead the agent allows the automatic return to command zone for the commander causing it to gain the +2 tax, when it would have returned automatically on its own without permanently increasing the recast tax for the commander for the rest of the game.

What should have happened

This is a unique situation in which the commander may be allowed to stay exiled rather than return to the commander zone, because the exile effect on it automatically returns the commander to the battlefield (which is a more powerful/better place to be returned to than to the commander zone, which would require recasting it at a higher cost and any additional removal of the commander to be more expensive).

Category

  • kind: misplay
  • severity: debatable

Context

## What happened <img width="1440" alt="image.png" src="attachments/f7f9752c-0da4-4e0a-84e2-32db7bc90551"> Commander gets TEMPORARILY exiled (would return at next endstep) and instead the agent allows the automatic return to command zone for the commander causing it to gain the +2 tax, when it would have returned automatically on its own without permanently increasing the recast tax for the commander for the rest of the game. ## What should have happened This is a unique situation in which the commander may be allowed to stay exiled rather than return to the commander zone, because the exile effect on it automatically returns the commander to the battlefield (which is a more powerful/better place to be returned to than to the commander zone, which would require recasting it at a higher cost and any additional removal of the commander to be more expensive). ## Category - kind: misplay - severity: debatable ## Context - Page URL: https://veles.bstein.dev/table?path=%2Fdata%2Fveles-artifacts%2Flogs%2Fdashboard%2Fstrategy-plan-poc-0538-20260709225328-a-baseline-lazav-v8-phelia-v1%2Fduel_commander_phelia_exuberant_shepherd__1_g1_2026070902%2Fgame.jsonl - Route: table - Path: /table?path=%2Fdata%2Fveles-artifacts%2Flogs%2Fdashboard%2Fstrategy-plan-poc-0538-20260709225328-a-baseline-lazav-v8-phelia-v1%2Fduel_commander_phelia_exuberant_shepherd__1_g1_2026070902%2Fgame.jsonl - Backend version: 0.5.38 - Frontend version: 0.5.38 - Sim version: 0.5.38
image.png
1.0 MiB
bstein added the
kind/misplay
severity/debatable
pilot/proficient
 labels 2026-07-09 23:27:25 +00:00
bstein commented 2026-07-10 04:00:40 +00:00
Author
Owner
Copy Link

Verified and closing in Cassandra 0.5.40.

Fix:

  • 0.5.39 added the commander temporary-exile exception: a commander temporarily exiled with a pending battlefield return can choose/keep commander_zone:no so the pilot does not throw away the return or add unnecessary commander tax.
  • 0.5.40 hardened that exception after the first rerun found an over-broad Phelia-history false positive. The exception now requires the prompted commander to be visibly in exile and the blink source/target evidence to match that same commander; ordinary death/permanent-exile prompts keep the command-zone default.

Proof:

Leaving #29 open: that issue tracks an adjacent Table/UI duplicate-commander artifact, not the runtime commander-zone policy fixed here.

Verified and closing in Cassandra 0.5.40. Fix: - 0.5.39 added the commander temporary-exile exception: a commander temporarily exiled with a pending battlefield return can choose/keep `commander_zone:no` so the pilot does not throw away the return or add unnecessary commander tax. - 0.5.40 hardened that exception after the first rerun found an over-broad Phelia-history false positive. The exception now requires the prompted commander to be visibly in exile and the blink source/target evidence to match that same commander; ordinary death/permanent-exile prompts keep the command-zone default. Proof: - Source: focused `tests.test_strategy_plan_runtime tests.test_playtest_decisions tests.test_codex_decision_worker tests.test_misplay_audit -q` ran 549 tests OK; full Python discovery ran 1951 tests OK, 5 skipped; frontend build passed; `git diff --check` passed. - Live deploy: `/api/v1/app-config` reports backend/frontend/sim-worker `0.5.40` and `/api/v1/ready` is green. - Live Veteran A/B rerun used the same seed `2026071001`. - A baseline: `strategy-plan-poc-0540-20260710025131-a-veteran-baseline-lazav-v8-phelia-v1`, queue job `d4162c3d3881`, completed; Phelia/player_2 won turn 14; 216 prompts; 0 strategy-plan rows. Table: https://veles.bstein.dev/table?path=/data/veles-artifacts/logs/dashboard/strategy-plan-poc-0540-20260710025131-a-veteran-baseline-lazav-v8-phelia-v1/duel_commander_phelia_exuberant_shepherd__1_g1_2026071001/game.jsonl - B strategy-plan: `strategy-plan-poc-0540-20260710025131-b-veteran-strategy-plan-lazav-v9-phelia-v2`, queue job `766450e09b82`, completed; Phelia/player_2 won turn 11; 177 prompts; 170 strategy-plan rows; 673 annotations. Table: https://veles.bstein.dev/table?path=/data/veles-artifacts/logs/dashboard/strategy-plan-poc-0540-20260710025131-b-veteran-strategy-plan-lazav-v9-phelia-v2/duel_commander_phelia_exuberant_shepherd__2_g1_2026071001/game.jsonl - Commander-zone scan: A events idx 104 and 152 both chose `commander_zone:yes`; B event idx 43 chose `commander_zone:yes` with `strategy_plan:hard_guard`; `commander_zone:no` was tagged as `breaks_commander_access` / `requires_deviation` and was not selected. - `audit_stage2_logs` reported 0 findings over the pair. - Kubernetes proof: both 0.5.40 sim jobs completed with 0 pod restarts; Veteran queue payloads used `active_deadline_seconds=10800`. Leaving #29 open: that issue tracks an adjacent Table/UI duplicate-commander artifact, not the runtime commander-zone policy fixed here.
Sign in to join this conversation.
No Branch/Tag Specified
Branches Tags
Labels
Clear labels
battery-0.5.21
Cassandra edge-case evaluation finding
bridge
Cassandra edge-case evaluation finding
commander-timing-stack
Cassandra edge-case evaluation finding
draw-go-vs-storm
Cassandra edge-case evaluation finding
dredge-vs-lantern
Cassandra edge-case evaluation finding
format/commander
Commander-specific issue.
format/duel-commander
Duel Commander-specific issue.
kind/bridge
Forge bridge/runtime ownership gap.
kind/metadata
Card metadata, image, or content import issue.
kind/misplay
Pilot quality issue or bad strategic/tactical choice.
kind/ops
Deployment, auth, capacity, or operational issue.
kind/ui
Replay, Lab, Runs, or visual review issue.
manual-review
Cassandra edge-case evaluation finding
misplay
Cassandra edge-case evaluation finding
pilot/adept
Observed at adept pilot skill.
pilot/proficient
Observed at proficient pilot skill.
pilot/veteran
Observed at veteran pilot skill.
rules
Cassandra edge-case evaluation finding
severity/debatable
Misplay should be absent at adept and above.
severity/defensible
Borderline line; veteran+ should avoid if identifiable.
severity/obvious
Misplay should not happen at proficient.
No Label
kind/misplay
pilot/proficient
severity/debatable
Milestone
No items
No Milestone
Assignees
Clear assignees
bstein veles-dev
No Assignees
1 Participants
Notifications
Due Date
No due date set.
Dependencies

No dependencies set.

Reference: cassandra-alpha/feedback#31
No description provided.
Delete Branch "%!s()"

Deleting a branch is permanent. Although the deleted branch may continue to exist for a short time before it actually gets removed, it CANNOT be undone in most cases. Continue?