cogames-agents 0.0.0.7__tar.gz → 0.0.0.10__tar.gz

cogames_agents-0.0.0.10/.nimby-version

@@ -0,0 +1 @@
+0.1.18

{cogames_agents-0.0.0.7 → cogames_agents-0.0.0.10}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: cogames-agents
-Version: 0.0.0.7
+Version: 0.0.0.10
 Summary: Optional agent policies for CoGames
 Author: Metta AI
 License-Expression: MIT
@@ -12,8 +12,8 @@ Classifier: Operating System :: POSIX :: Linux
 Classifier: Operating System :: MacOS
 Requires-Python: <3.13,>=3.12
 Description-Content-Type: text/markdown
-Requires-Dist: cogames==0.3.64
-Requires-Dist: mettagrid==0.2.0.74
+Requires-Dist: cogames==0.3.68
+Requires-Dist: mettagrid==0.2.0.82
 Requires-Dist: numpy>=2.0.0
 Provides-Extra: test
 Requires-Dist: pytest; extra == "test"
@@ -44,7 +44,6 @@ Common scripted policy names include:
 - CogsGuard variants: `alignall`, `cogsguard_control`, `cogsguard_targeted`, `cogsguard_v2`
 - CogsGuard roles: `miner`, `scout`, `aligner`, `scrambler`
 - Teacher: `teacher`
-- Pinky: `pinky`
 
 For the full registry snapshot, see `docs/scripted-agent-registry.md`.
 
@@ -64,18 +63,6 @@ metta://policy/role_py?role_cycle=aligner,miner,scrambler,scout
 metta://policy/role_py?role_order=aligner,miner,aligner,miner,scout
 ```
 
-Pinky role counts are applied in a different order than CogsGuard:
-
-- Pinky order: miner -> scout -> aligner -> scrambler, and any remaining agents stay default/noop.
-- CogsGuard order: scrambler -> aligner -> miner -> scout, then fills remaining agents with gear.
-
-Examples:
-
-```
-metta://policy/pinky?miner=4&aligner=2&scrambler=4
-metta://policy/pinky?miner=2&scout=2&aligner=1&scrambler=1&debug=1
-```
-
 ## Recipe usage
 
 The `recipes.experiment.scripted_agents` recipe accepts the same scripted policy names:

{cogames_agents-0.0.0.7 → cogames_agents-0.0.0.10}/README.md

@@ -22,7 +22,6 @@ Common scripted policy names include:
 - CogsGuard variants: `alignall`, `cogsguard_control`, `cogsguard_targeted`, `cogsguard_v2`
 - CogsGuard roles: `miner`, `scout`, `aligner`, `scrambler`
 - Teacher: `teacher`
-- Pinky: `pinky`
 
 For the full registry snapshot, see `docs/scripted-agent-registry.md`.
 
@@ -42,18 +41,6 @@ metta://policy/role_py?role_cycle=aligner,miner,scrambler,scout
 metta://policy/role_py?role_order=aligner,miner,aligner,miner,scout
 ```
 
-Pinky role counts are applied in a different order than CogsGuard:
-
-- Pinky order: miner -> scout -> aligner -> scrambler, and any remaining agents stay default/noop.
-- CogsGuard order: scrambler -> aligner -> miner -> scout, then fills remaining agents with gear.
-
-Examples:
-
-```
-metta://policy/pinky?miner=4&aligner=2&scrambler=4
-metta://policy/pinky?miner=2&scout=2&aligner=1&scrambler=1&debug=1
-```
-
 ## Recipe usage
 
 The `recipes.experiment.scripted_agents` recipe accepts the same scripted policy names:

cogames_agents-0.0.0.10/docs/beta-cvc-policy-validation.md

@@ -0,0 +1,126 @@
+# Beta-CVC Policy Validation (Local)
+
+Date: 2026-01-29
+
+This document summarizes the current state of local validation for the beta-cvc policies referenced in the
+beta-cogsguard leaderboard list. The goal was to re-download the exact uploaded bundles and confirm they start correctly
+on the beta-cvc mission (`cogsguard_machina_1.basic`).
+
+## Scope
+
+- Policies tested (19 total) match the earlier beta-cogsguard leaderboard list:
+  - daveey.pinky:v5/v6/v7/v8/v9
+  - daveey.planky:v6/v7/v8
+  - relh.cogas:v6/v7/v8/v10
+  - relh.wombo-mix:v1/v3/v4
+  - noah::coggernaut:v4/v9
+  - manvi_metcon:v3
+  - cogsguard-roster-mix:v1
+
+- All bundles were re-downloaded from the policy `s3_path` recorded in the backend, not from the older
+  `policy-versions/` location.
+
+## Bundle Download (Exact Uploaded Artifacts)
+
+The backend stores a policy version’s `s3_path` at: `/stats/policies/versions/{policy_version_id}`.
+
+We used the policy version lookup to get `s3_path`, then downloaded via S3:
+
+```
+aws s3 cp s3://observatory-private/cogames/submissions/<user>/<upload_id>.zip \
+  outputs/beta-cvc-policy-bundles-redownload/<policy>_v<version>.zip
+```
+
+All 19 bundles are now in: `outputs/beta-cvc-policy-bundles-redownload/`
+
+## Local Validation Commands
+
+We validated against the beta-cvc mission (CogsGuard Machina1):
+
+```
+uv run cogames scrimmage -m cogsguard_machina_1.basic -c 5 -e 1 -s 300 --format json \
+  -p ./outputs/beta-cvc-policy-bundles-redownload/<policy>.zip
+
+uv run cogames eval -m cogsguard_machina_1.basic -c 5 -e 1 -s 300 --format json \
+  -p ./outputs/beta-cvc-policy-bundles-redownload/<policy>.zip
+```
+
+Notes:
+
+- Use a relative path starting with `./` for local bundles; otherwise the CLI interprets the argument as a class name.
+
+## Results: Scrimmage + Eval (CVC Map)
+
+Summary:
+
+- **18/19** bundles run successfully for both `scrimmage` and `eval`.
+- **1/19** fails due to a missing class in the bundle.
+
+Failures:
+
+- `relh.wombo-mix:v3` fails to import: `cogames_agents.policy.scripted_agent.cogsguard.policy.CogsguardWomboMixPolicy`
+
+Logs and summary:
+
+- `outputs/beta-cvc-policy-bundles-redownload/smoke_logs/summary.txt`
+- Per-policy logs: `outputs/beta-cvc-policy-bundles-redownload/smoke_logs/*__scrimmage.log` and `*__eval.log`
+
+## Diagnose (Diagnostic Evals) Caveats
+
+`cogames diagnose` runs **diagnostic evals**, which are not CogsGuard missions. These maps have:
+
+- Max cogs = 4.
+- A different action set.
+- Variants incompatible with CogsGuard missions.
+
+### Full diagnostic suite
+
+Running `diagnose` with the full `diagnostic_evals` set fails broadly because the evals are incompatible with CogsGuard
+mission variants or cogs > 4.
+
+### Single diagnostic experiment
+
+We ran a single diagnostic experiment as a smoke test:
+
+```
+uv run cogames diagnose -S diagnostic_evals \
+  --experiments diagnostic_chest_deposit_near \
+  -c 4 -e 1 -s 300 \
+  ./outputs/beta-cvc-policy-bundles-redownload/<policy>.zip
+```
+
+Results:
+
+- OK:
+  - cogsguard-roster-mix:v1
+  - daveey.pinky:v5/v6/v7/v8/v9
+  - noah::coggernaut:v4/v9
+  - relh.wombo-mix:v1/v4
+- FAIL:
+  - daveey.planky:v6/v7/v8
+  - manvi_metcon:v3
+  - relh.cogas:v6/v7/v8/v10
+  - relh.wombo-mix:v3
+
+Typical failure pattern for the diagnostic experiment:
+
+- Action mismatch in diagnostics env (example: `KeyError: 'change_vibe_miner'`).
+- `relh.wombo-mix:v3` still fails due to missing class.
+
+Conclusion: `diagnose` is **not a reliable signal** for CogsGuard policies. The best local signal for beta-cvc is
+`scrimmage`/`eval` on `cogsguard_machina_1.basic`.
+
+## Beta-CVC Tournament State (As of 2026-01-29)
+
+Leaderboard is empty because no competition matches have completed yet. Policies currently in beta-cvc:
+
+- Qualifying completed (2 matches): daveey.pinky:v8, daveey.planky:v6, cogsguard-roster-mix:v1, noah::coggernaut:v9,
+  noah::coggernaut:v10.
+- Competition active but no matches completed: manvi_metcon:v3.
+
+## Known Issues / Follow-ups
+
+- `relh.wombo-mix:v3` bundle references a class that does not exist in the repository (`CogsguardWomboMixPolicy`), so it
+  cannot run locally or in the tournament runner.
+- If we need a single “diagnose”-style smoke check for CogsGuard, we should add a dedicated CogsGuard diagnostic mission
+  set (or avoid `diagnose` entirely for this season).

{cogames_agents-0.0.0.7 → cogames_agents-0.0.0.10}/docs/cogames-eval-internals.md

@@ -98,7 +98,7 @@ The core pipeline is in `packages/cogames/src/cogames/evaluate.py`:
 2. **For each mission:** a. `allocate_counts()` distributes agents among policies by weight (largest-remainder method).
    b. `np.repeat(np.arange(len(counts)), counts)` creates the initial assignment array.
 
-3. **Episode loop** (`run_multi_episode_rollout()` in `metta_alo/rollout.py:271`):
+3. **Episode loop** (`run_multi_episode_rollout()` in `mettagrid/runner/rollout.py`):
    - For each episode:
      - Shuffle assignments via `rng.shuffle(assignments)` (randomizes which agent slots get which policy).
      - `run_single_episode_rollout()` creates the environment, loads policies, runs the simulation, collects
@@ -167,16 +167,15 @@ cogames replay <replay_path>
 
 ## 5. Wandb Integration Points
 
-**There is no wandb integration in the eval command.**
+**There is no wandb integration in the eval/diagnose commands.**
 
-The `cogames eval`/`scrimmage` pipeline does not import or call wandb. No metrics are logged to wandb during evaluation.
+The `cogames run`/`scrimmage`/`diagnose` pipelines do not import or call wandb. No metrics are logged to wandb during
+evaluation.
 
-Wandb integration exists in:
+Wandb integration exists only in:
 
 - The **training pipeline** (`cogames train`) -- logs training curves, losses, and periodic eval metrics during
   training.
-- The standalone **`scripts/run_evaluation.py`** -- has optional wandb logging for batch evaluation runs (this is a
-  separate script, not the CLI command).
 
 This is a significant observability gap. Eval results are ephemeral unless captured via `--format json` to a file.
 

{cogames_agents-0.0.0.7 → cogames_agents-0.0.0.10}/docs/cogas-agent-design.md

@@ -7,9 +7,9 @@ Blueprint for a leaderboard-winning agent targeting `cogs.aligned.junction.held
 Cogas uses a **phased goal-tree** architecture, combining Planky's declarative goal decomposition with CogsGuard's
 phase-based state machine and evolution-driven role selection.
 
-**Why not pure behavior-tree (Pinky)?** Pinky's vibe-driven approach lacks explicit precondition reasoning. Agents
-thrash between behaviors when preconditions aren't met (e.g., attempting to align without hearts). Goal-trees naturally
-decompose "align junction" into "have gear AND have hearts AND be adjacent" without custom priority logic.
+**Why not pure behavior-tree?** A vibe-driven approach lacks explicit precondition reasoning. Agents thrash between
+behaviors when preconditions aren't met (e.g., attempting to align without hearts). Goal-trees naturally decompose
+"align junction" into "have gear AND have hearts AND be adjacent" without custom priority logic.
 
 **Why not pure goal-tree (Planky)?** Planky has no phase awareness. It re-evaluates the full goal list every tick, which
 is wasteful when the agent is mid-navigation. Adding phases (BOOTSTRAP, CONTROL, SUSTAIN) gives temporal structure that
@@ -28,7 +28,7 @@ are imperative spaghetti. Goal-tree decomposition inside each phase keeps role l
 ├─────────────────────────────────────┤
 │  GoalEvaluator (per-phase goals)    │  Precondition decomposition
 ├─────────────────────────────────────┤
-│  Navigator + EntityMap + SafetyMgr  │  Shared services (from Planky/Pinky)
+│  Navigator + EntityMap + SafetyMgr  │  Shared services (from Planky)
 └─────────────────────────────────────┘
 ```
 
@@ -347,7 +347,7 @@ accumulation. The catalog tracks:
 3. Implement `PhaseController` with BOOTSTRAP/CONTROL/SUSTAIN transitions
 4. Port Planky's goal-tree evaluator with phase-aware goal lists
 5. Implement role-specific goals (miner, aligner, scrambler, scout)
-6. Reuse Planky's `Navigator`, `EntityMap` and Pinky's `SafetyManager`
+6. Reuse Planky's `Navigator` and `EntityMap`
 
 ### Stage 2: Innovations
 

{cogames_agents-0.0.0.7 → cogames_agents-0.0.0.10}/docs/cogsguard-mechanics-deep-dive.md

@@ -206,8 +206,8 @@ Four resources mined from extractors in map corners:
 
 **Source:** `stations.py:122-206`
 
-**Miner bonus:** With miner gear, extractors yield 10x resources (large_amount vs small_amount in
-`SimpleExtractorConfig`, `stations.py:275-300`).
+**Miner bonus:** With miner gear, extractors yield 10x resources (large_amount vs small_amount in `CvCExtractorConfig`,
+`stations.py:275-300`).
 
 ### Hearts
 

{cogames_agents-0.0.0.7 → cogames_agents-0.0.0.10}/docs/creating-scripted-agents.md

@@ -23,11 +23,6 @@ cogames-agents/src/cogames_agents/policy/
 │   │   ├── roles.py                   # Per-role policies (miner, scout, etc.)
 │   │   ├── types.py                   # CogsGuard-specific state types
 │   │   └── ...
-│   ├── pinky/                         # Behavior-tree style agent
-│   │   ├── policy.py                  # PinkyPolicy (short_name: "pinky")
-│   │   ├── behaviors/                 # Per-role behavior modules
-│   │   ├── services/                  # Map, navigation, safety services
-│   │   └── state.py                   # Agent state
 │   └── planky/                        # Goal-tree agent
 │       ├── policy.py                  # PlankyPolicy (short_name: "planky")
 │       ├── goals/                     # Goal definitions per role
@@ -315,8 +310,7 @@ class MyPolicyImpl(BaselineAgentPolicyImpl):
 
 ## 6. Roles and Vibes
 
-The vibe system is used by team-play agents (CogsGuard, Pinky, Planky) to control agent behavior through in-game visual
-state.
+The vibe system is used by team-play agents (CogsGuard, Planky) to control agent behavior through in-game visual state.
 
 ### What are vibes?
 
@@ -407,7 +401,7 @@ uv run cogames play --mission recipes.experiment.cogsguard.play \
 
 ```bash
 # Full evaluation across all diagnostic missions
-uv run python packages/cogames/scripts/run_evaluation.py --policy my_agent
+uv run cogames diagnose my_agent -S diagnostic_evals
 ```
 
 ## 8. Code Examples from Existing Agents

{cogames_agents-0.0.0.7 → cogames_agents-0.0.0.10}/docs/machina1-agent-analysis.md

@@ -87,17 +87,6 @@ loop.
 | Weaknesses   | Inherits all baseline weaknesses. Unclipping is only relevant when enemy scramblers have clipped extractors -- in pure resource gathering it adds overhead. |
 | Machina1 fit | **Poor to moderate**. Useful only in adversarial scenarios. The baseline movement inefficiency dominates performance.                                       |
 
-### Pinky Agent
-
-**Architecture**: Vibe-based role system with per-agent Brain. Roles: miner, scout, aligner, scrambler. URI-configurable
-role distribution.
-
-| Aspect       | Assessment                                                                                                                                                                                        |
-| ------------ | ------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- |
-| Strengths    | Multi-role coordination. Scout discovers map efficiently. Aligner/scrambler handle junction control. Dynamic role switching based on communal resource levels. Gear acquisition from stations.    |
-| Weaknesses   | Role switching logic may thrash on machina1 where resource availability shifts as biome zones are discovered. Navigator service quality depends on map complexity -- maze biomes can confuse BFS. |
-| Machina1 fit | **Moderate**. Role specialization helps, but the brain/service abstraction adds overhead. Works best when tuned via URI params for specific role ratios.                                          |
-
 ### Planky Agent
 
 **Architecture**: Goal-tree hierarchical policy. Each role has a priority-ordered goal list. Goals evaluate
@@ -344,7 +333,6 @@ This would produce generation-over-generation shifts in the role population.
 | tiny_baseline         | Non-viable           | Testing only                      |
 | baseline              | Poor                 | Ablation baseline                 |
 | ladybug_py            | Poor-Moderate        | Adversarial unclipping            |
-| pinky                 | Moderate             | Tunable role experiments          |
 | planky                | Moderate-Good        | Goal-priority research            |
 | role_py (CoGsGuard)   | Good                 | General competitive play          |
 | cogsguard_v2          | Good-Strong          | Balanced default teams            |

{cogames_agents-0.0.0.7 → cogames_agents-0.0.0.10}/docs/metta-cogames-overview.md

@@ -130,7 +130,7 @@ class=baseline,proportion=0.5
 
 # URI format (role-based)
 metta://policy/role_py?role_cycle=aligner,miner,scrambler,scout
-metta://policy/pinky?miner=4&aligner=2&scrambler=4&scout=0
+metta://policy/role_py?miner=4&aligner=2&scrambler=4&scout=0
 ```
 
 ## 4. CogsGuard Game Mechanics
@@ -237,7 +237,7 @@ Recipe execution:
 - `thinky`, `race_car`, `ladybug` -- Nim-compiled (faster)
 - `role`, `role_py`, `wombo` -- Role-rotation strategies
 - `miner`, `scout`, `aligner`, `scrambler` -- Role specialists
-- `teacher`, `pinky` -- Role assignment meta-policies
+- `teacher` -- Role assignment meta-policy
 
 **Templates** (from cogames):
 

{cogames_agents-0.0.0.7 → cogames_agents-0.0.0.10}/docs/role-distribution-analysis.md

@@ -11,7 +11,7 @@ game-state-driven role switching.
 
 **Key Finding:** The highest-performing agents (CoGsGuard Control, Wombo) share a common pattern: they start with an
 explore-heavy distribution and dynamically shift toward aligner/scrambler-heavy distributions as junctions are
-discovered. Pure static distributions (planky defaults, pinky defaults) leave performance on the table.
+discovered. Pure static distributions (planky defaults) leave performance on the table.
 
 ---
 
@@ -59,50 +59,7 @@ hub recipes.
 
 ---
 
-### 2.2 Pinky (`metta://policy/pinky`)
-
-**Architecture:** Vibe-based role system with per-agent Brain. URI-configurable static distribution.
-
-**Default distribution:** All zeroes -- no agents active unless URI params specify counts.
-
-**Gear-vibe distribution (when `?gear=10`):** The "gear" vibe triggers a hardcoded role selection cycle:
-
-| Agent ID | Role      | Pattern |
-| -------- | --------- | ------- |
-| 0        | aligner   |         |
-| 1        | aligner   |         |
-| 2        | scrambler |         |
-| 3        | miner     |         |
-| 4        | aligner   |         |
-| 5        | scrambler |         |
-| 6        | aligner   |         |
-| 7        | scrambler |         |
-| 8        | aligner   |         |
-| 9        | miner     |         |
-
-**Effective gear-vibe distribution (10 agents):**
-
-| Role      | Count | Percentage |
-| --------- | ----- | ---------- |
-| Miner     | 2     | 20%        |
-| Scout     | 0     | 0%         |
-| Aligner   | 5     | 50%        |
-| Scrambler | 3     | 30%        |
-
-**Key observations:**
-
-- The gear-vibe cycle is **heavily biased toward aligners** (50%).
-- **Zero scouts** -- same blind-discovery problem as planky defaults.
-- Only 2 miners -- may struggle with resource economy on larger maps.
-- Supports dynamic `change_role` parameter: miners become aligners/scramblers when all communal resources exceed 25, and
-  aligners/scramblers revert to miners when any resource drops below 3.
-- Agents beyond the specified count stay `default` (noop) -- pinky does NOT fill remaining slots automatically.
-
-**Source:** `pinky/policy.py:133-148` (gear cycle), `pinky/policy.py:457-477` (constructor).
-
----
-
-### 2.3 Role_py / CoGsGuard (`metta://policy/role_py`)
+### 2.2 Role_py / CoGsGuard (`metta://policy/role_py`)
 
 **Architecture:** Multi-role vibe system with SmartRoleCoordinator. Dynamic role switching.
 
@@ -141,7 +98,7 @@ assigns roles based on team state:
 
 ---
 
-### 2.4 CoGsGuard V2 (`metta://policy/cogsguard_v2`)
+### 2.3 CoGsGuard V2 (`metta://policy/cogsguard_v2`)
 
 **Architecture:** CoGsGuard base with tuned static default allocation formula.
 
@@ -175,7 +132,7 @@ assigns roles based on team state:
 
 ---
 
-### 2.5 CoGsGuard Control (`metta://policy/cogsguard_control`)
+### 2.4 CoGsGuard Control (`metta://policy/cogsguard_control`)
 
 **Architecture:** Phased commander coordinator with active role reassignment.
 
@@ -202,7 +159,7 @@ assigns roles based on team state:
 
 ---
 
-### 2.6 Wombo (`metta://policy/wombo`)
+### 2.5 Wombo (`metta://policy/wombo`)
 
 **Architecture:** CoGsGuard generalist variant prioritizing multi-junction alignment.
 
@@ -244,7 +201,7 @@ algorithm considering:
 
 ---
 
-### 2.7 Teacher (`metta://policy/teacher`)
+### 2.6 Teacher (`metta://policy/teacher`)
 
 **Architecture:** Wrapper that delegates to Nim CoGsGuard agents with forced initial vibes.
 
@@ -286,12 +243,6 @@ metta://policy/role_py?role_order=aligner,miner,aligner,miner    # Exact sequenc
 metta://policy/role_py?evolution=1                                 # Evolutionary roles
 ```
 
-### Pinky additional params:
-
-```
-metta://policy/pinky?miner=4&aligner=3&scrambler=3&debug=1&change_role=100
-```
-
 ### Planky additional params:
 
 ```
@@ -316,7 +267,6 @@ For a standard 10-agent team:
 | Agent                 | Miner        | Scout   | Aligner | Scrambler | Dynamic? | Notes                                           |
 | --------------------- | ------------ | ------- | ------- | --------- | -------- | ----------------------------------------------- |
 | **planky** (default)  | 4 (40%)      | 0 (0%)  | 2 (20%) | 4 (40%)   | No       | No scouts; denial-heavy                         |
-| **pinky** (gear=10)   | 2 (20%)      | 0 (0%)  | 5 (50%) | 3 (30%)   | Partial  | Aligner-heavy; resource-triggered switching     |
 | **role_py** (default) | 4 (40%)      | 0-2\*   | 0-2\*   | 1 (10%)   | Yes      | 5 dynamic `gear` agents fill gaps               |
 | **cogsguard_v2**      | 5 (50%)      | 1 (10%) | 2 (20%) | 2 (20%)   | Limited  | Balanced static formula                         |
 | **cogsguard_control** | 5-7          | 1-2     | 0-2     | 1-2       | Yes      | Phase-aware; commander reassigns every 40 steps |
@@ -405,8 +355,8 @@ Aggressive junction control with double scouts and generalist role switching.
 
 ## 6. Conclusions
 
-1. **No agent currently uses a theoretically optimal distribution out of the box.** Planky and pinky lack scouts
-   entirely in their defaults. Role_py's defaults are miner-heavy with most agents dynamically assigned.
+1. **No agent currently uses a theoretically optimal distribution out of the box.** Planky lacks scouts entirely in its
+   defaults. Role_py's defaults are miner-heavy with most agents dynamically assigned.
 
 2. **Dynamic role switching outperforms static allocation.** CoGsGuard Control and Wombo (the highest-rated agents in
    machina1 analysis) both use adaptive distributions.

{cogames_agents-0.0.0.7 → cogames_agents-0.0.0.10}/docs/scripted-agent-registry.md

@@ -36,7 +36,3 @@ python -c "from cogames_agents.policy.scripted_registry import list_scripted_age
 - `cogsguard_v2` - V2 variant
 - `miner`, `scout`, `aligner`, `scrambler` - Role-specific policies
 - `teacher` - Teacher wrapper over Nim multi-role
-
-## Pinky (Python, CogsGuard)
-
-- `pinky` - Role-count based policy (`?miner=...&scout=...&aligner=...&scrambler=...`)

{cogames_agents-0.0.0.7 → cogames_agents-0.0.0.10}/pyproject.toml

@@ -17,7 +17,7 @@ classifiers = [
   "Operating System :: MacOS",
 ]
 urls = { Homepage = "https://github.com/Metta-AI/metta/tree/main/packages/cogames-agents", Repository = "https://github.com/Metta-AI/metta" }
-dependencies = ["cogames==0.3.64", "mettagrid==0.2.0.74", "numpy>=2.0.0"]
+dependencies = ["cogames==0.3.68", "mettagrid==0.2.0.82", "numpy>=2.0.0"]
 
 [project.optional-dependencies]
 test = ["pytest", "pytest-xdist", "ruff"]

{cogames_agents-0.0.0.7 → cogames_agents-0.0.0.10}/scripts/README.md

@@ -16,7 +16,7 @@ Runs every registered scripted agent through `cogames scrimmage` and saves per-a
 ./scripts/benchmark_agents.sh -e 20 -s 2000 -m cogsguard_arena.basic -o ./my_results
 
 # Subset of agents
-./scripts/benchmark_agents.sh -a role,pinky,baseline,wombo -e 50
+./scripts/benchmark_agents.sh -a role,baseline,wombo -e 50
 ```
 
 **Options:**
@@ -67,7 +67,7 @@ Fast single-agent eval for development iteration (3 episodes, 500 steps by defau
 ./scripts/quick_eval.sh role
 
 # JSON output
-./scripts/quick_eval.sh pinky --json
+./scripts/quick_eval.sh planky --json
 
 # Open in MettaScope GUI
 ./scripts/quick_eval.sh baseline --gui
@@ -84,7 +84,6 @@ Registered scripted agents (from `cogames-agents` package):
 | -------------------- | ------------------------------- |
 | `role`               | Multi-role Nim CogsGuard policy |
 | `role_py`            | Python multi-role CogsGuard     |
-| `pinky`              | Alternative role ordering       |
 | `planky`             | Plank-focused strategy          |
 | `wombo`              | Alternative multi-role          |
 | `baseline`           | Standard baseline               |

{cogames_agents-0.0.0.7 → cogames_agents-0.0.0.10}/scripts/benchmark_agents.sh

@@ -27,7 +27,6 @@ AGENTS=""
 ALL_AGENTS=(
   role
   role_py
-  pinky
   planky
   wombo
   baseline

{cogames_agents-0.0.0.7 → cogames_agents-0.0.0.10}/scripts/quick_eval.sh

@@ -6,7 +6,7 @@
 #
 # Examples:
 #   ./scripts/quick_eval.sh role
-#   ./scripts/quick_eval.sh pinky -e 5 -s 500
+#   ./scripts/quick_eval.sh planky -e 5 -s 500
 #   ./scripts/quick_eval.sh baseline --json
 #   ./scripts/quick_eval.sh role -m cogsguard_arena.basic --seed 99
 
@@ -16,7 +16,7 @@ if [[ $# -lt 1 ]] || [[ "$1" == "-h" ]] || [[ "$1" == "--help" ]]; then
   echo "Usage: $0 AGENT [OPTIONS]"
   echo ""
   echo "Arguments:"
-  echo "  AGENT              Scripted agent name (e.g. role, pinky, baseline)"
+  echo "  AGENT              Scripted agent name (e.g. role, planky, baseline)"
   echo ""
   echo "Options:"
   echo "  -e EPISODES        Number of episodes (default: 3)"

{cogames_agents-0.0.0.7 → cogames_agents-0.0.0.10}/src/cogames_agents/evals/planky_evals.py

@@ -14,7 +14,9 @@ from typing import Dict
 from pydantic import Field
 
 from cogames.cogs_vs_clips.cog import CogConfig
-from cogames.cogs_vs_clips.mission import Mission, Site
+from cogames.cogs_vs_clips.mission import CoGameSite as Site
+from cogames.cogs_vs_clips.mission import CvCMission as Mission
+from cogames.cogs_vs_clips.team import CogTeam
 from mettagrid.config.mettagrid_config import MettaGridConfig
 from mettagrid.map_builder.map_builder import MapBuilderConfig
 from mettagrid.mapgen.mapgen import MapGen, MapGenConfig
@@ -90,7 +92,6 @@ class _PlankyDiagnosticBase(Mission):
         default_factory=lambda: CogConfig(
             energy_limit=255,
             initial_energy=255,
-            energy_regen=255,
             initial_hp=100,
             hp_regen=0,
             influence_regen=0,
@@ -108,6 +109,14 @@ class _PlankyDiagnosticBase(Mission):
         custom_map = _get_planky_map(self.map_name)
         original_map_builder = self.site.map_builder
         self.site.map_builder = custom_map
+
+        # Apply wealth to teams (base class wealth attribute wasn't being used)
+        original_teams = self.teams
+        self.teams = {
+            name: CogTeam(name=team.name, short_name=team.short_name, num_agents=team.num_agents, wealth=self.wealth)
+            for name, team in self.teams.items()
+        }
+
         try:
             cfg = super().make_env()
             cfg.game.map_builder = custom_map
@@ -124,6 +133,7 @@ class _PlankyDiagnosticBase(Mission):
             return cfg
         finally:
             self.site.map_builder = original_map_builder
+            self.teams = original_teams
 
 
 # ==============================================================================
@@ -264,7 +274,6 @@ class PlankySurviveRetreat(_PlankyDiagnosticBase):
         default_factory=lambda: CogConfig(
             energy_limit=255,
             initial_energy=255,
-            energy_regen=255,
             initial_hp=20,
             hp_regen=0,
             influence_regen=0,

{cogames_agents-0.0.0.7 → cogames_agents-0.0.0.10}/src/cogames_agents/policy/nim_agents/__init__.py

@@ -6,14 +6,12 @@ __all__ = [
     "RandomAgentsMultiPolicy",
     "ThinkyAgentsMultiPolicy",
     "RaceCarAgentsMultiPolicy",
-    "LadyBugAgentsMultiPolicy",
     "CogsguardAlignAllAgentsMultiPolicy",
 ]
 
 # Re-export the policy classes for convenience
 from cogames_agents.policy.nim_agents.agents import (  # noqa: F401
     CogsguardAlignAllAgentsMultiPolicy,
-    LadyBugAgentsMultiPolicy,
     RaceCarAgentsMultiPolicy,
     RandomAgentsMultiPolicy,
     ThinkyAgentsMultiPolicy,

{cogames_agents-0.0.0.7 → cogames_agents-0.0.0.10}/src/cogames_agents/policy/nim_agents/agents.py

@@ -55,17 +55,6 @@ class RaceCarAgentsMultiPolicy(NimMultiAgentPolicy):
         )
 
 
-class LadyBugAgentsMultiPolicy(NimMultiAgentPolicy):
-    short_names = ["ladybug"]
-
-    def __init__(self, policy_env_info: PolicyEnvInterface, agent_ids: Sequence[int] | None = None):
-        super().__init__(
-            policy_env_info,
-            nim_policy_factory=na.LadybugPolicy,
-            agent_ids=agent_ids,
-        )
-
-
 class CogsguardAgentsMultiPolicy(NimMultiAgentPolicy):
     short_names = ["role"]
 

{cogames_agents-0.0.0.7 → cogames_agents-0.0.0.10}/src/cogames_agents/policy/nim_agents/nim_agents.nim

@@ -1,6 +1,6 @@
 import
   genny, fidget2/measure,
-  random_agents, thinky_agents, racecar_agents, ladybug_agent, cogsguard_agents,
+  random_agents, thinky_agents, racecar_agents, cogsguard_agents,
   cogsguard_align_all_agents
 
 
@@ -45,12 +45,6 @@ exportRefObject RaceCarPolicy:
   procs:
     stepBatch(RaceCarPolicy, pointer, int, int, int, int, pointer, int, pointer)
 
-exportRefObject LadybugPolicy:
-  constructor:
-    newLadybugPolicy(string)
-  procs:
-    stepBatch(LadybugPolicy, pointer, int, int, int, int, pointer, int, pointer)
-
 exportRefObject CogsguardPolicy:
   constructor:
     newCogsguardPolicy(string)