Metta-AI

Use when improving a Coworld player through one diagnosed failure and completed comparison evidence.

Use when compiling, validating, and uploading a notsus change as a new Crewrift policy version.

Use when decoding a Crewrift replay for authoritative deaths, survival, movement, roles, tasks, or rewards.

Use when verifying that the notsus Bedrock vote advisor fires before an advisor-sensitive evaluation.

Use when classifying a stuck, failed, empty, or zero-scoring evaluation as infrastructure or policy failure.

Use when finding and statistically confirming a reproducible Crewrift crux before changing the policy.

Use before evaluations or after roster churn to resolve live policy versions and restore the champion flag.

Use when running a hosted Crewrift policy A/B and judging it with Wilson confidence intervals.

Use when designing or repairing a CoGames-style game and the main risk is that the rules, map, roles, or play loop do not yet match the real game or prompt closely enough.

Use when creating a new mettagrid game from scratch and you need the design-first workflow that writes the rules contract, checks engine fit, and hands a clean build plan to `cg.game.build-game`.

Use when a CoGames-style game already works visually and mechanically and now needs to be split into a composable variant registry, dependency tree, and optional curriculum tree without regressing the default game.

Use when implementing or overhauling a CoGames-style game once the target contract is known, including scaffold, registration, mechanics, assets, and variant factoring.

Use when designing or repairing a CoGames-style game and the main risk is that the rules, map, roles, or play loop do not yet match the real game or prompt closely enough.

Use when a CoGames-style game is mechanically working but still has missing, placeholder, wrongly themed, or non-namespaced art, terrain, or atlas entries.

Use when creating a new mettagrid game from scratch and you need the design-first workflow that writes the rules contract, checks engine fit, and hands a clean build plan to `cg.game.build-game`.

Use when a CoGames-style game already works visually and mechanically and now needs to be split into a composable variant registry, dependency tree, and optional curriculum tree without regressing the default game.

Use to identify and download Coworld episode artifacts — replays, results, and per-agent logs — for completed episodes, or STREAMED live from a still-running experience request (--watch: each episode downloads as it finishes, no waiting for the batch). Triggers: 'download the replay/logs for episode X', 'pull the last N episodes <player> played', 'fetch artifacts for this experience request / pool / round', 'stream artifacts while the request runs', 'get the episode data for post-mortem'. Game-agnostic (Crewrift, amongthem, etc.).

Use to create and monitor Coworld experience requests — hosted batches of episodes you define (target, roster, roles, count) for evaluating a policy against a live field. Triggers: 'run crewborg vs the top opponents', 'make an experience request', 'request N hosted games', 'A/B a policy against the league', 'measure the imposter', 'set up an evaluation battery'. DEFAULT FLOW after create: launch the STREAMING pipeline in the background (step 4) — artifacts download and analysis prep overlap the still-running episodes; don't wait for the batch to drain.

Use to run your own built player policy in a LOCAL Coworld episode and watch it — the Gate-1 smoke test ('did my change take, does it connect → play → exit cleanly?'). Triggers: 'smoke test the player locally', 'run a local game', 'run my policy locally and watch it', 'does the build still work', 'open the replay locally'. Game-agnostic; this is correctness only, NOT a competitive matchup (you can't run other users' policies locally).

Use to submit an ALREADY-UPLOADED policy version to a Coworld league and monitor whether it QUALIFIES, then competes (and becomes champion). Triggers: 'submit crewborg to the league', 'did it qualify', 'is it disqualified / champion', 'watch the qualifier', 'monitor standings after submitting'. SUBMIT is the public, effectively-irreversible, champion-making action — explicit human go-ahead only. (Building + uploading a version is the separate `build & upload` skill.)

Use to decide whether a crewborg change ACTUALLY helped — A/B the candidate against the baseline head-to-head, fresh, right now, against the same field. Triggers: 'did my change help', 'compare v71 vs v70', 'A/B test crewborg', 'is the candidate better', 'did fixing the imposter regress the crew'. Runs two MATCHED fresh experience requests, diffs role-split metrics with significance (compare.py), renders a comparison report, and guides a qualitative side-by-side. It's also the 'designed run' that crewrift-experiment hands off to.

Use to build and query the Crewrift event warehouse — a policy-indexed DuckDB/Parquet dataset of per-tick gameplay events (kills, follows, proximity, votes, tasks, chat) over many episodes — for deep, mechanistic, cross-episode questions about a policy's behaviour. Triggers: 'how often is crewborg near a crewmate without killing', 'who trails crewborg', 'where does it abandon tasks', 'does it vote real imposters', 'build the warehouse for this round', 'query the events', 'stream an experience request into a warehouse'. For a FRESH experience request, stream_eval.py builds the warehouse WHILE the episodes run (the default; builds are incremental). This is the deep-dig mainstay after crewrift-survey flags something.

Use to turn a SET of Crewrift episodes (an experience request, a policy's recent league games, a tournament batch) into a fast, polished HTML survey: a per-policy stats table, a policy×policy win heat map, and a short list of interesting episodes (with replay links). Triggers: 'survey these episodes', 'how is crewborg doing in this batch', 'who beats whom', 'give me the report on this XP request', 'which games should I watch'. Fast + lightweight (results.json + episode.json only). Pair with coworld-episode-artifacts (to pull the episodes); the DEEP per-episode dissection is a separate skill.

Use to build the crewborg image and upload it as a new policy version — the routine, inert, every-iteration action that gives you a runnable artifact to smoke-test and evaluate. Triggers: 'build crewborg', 'upload a new version', 'rebuild and re-upload', 'ship a version for testing', 'upload with the LLM on'. Uploading enters NO competition; submitting to a league is the separate coworld-policy-lifecycle (submit & monitor) skill.

Generate an HTML training dashboard from .coach/ state showing experiments, scores, learnings, TODOs. Pulls live tournament status. Opens in browser. Use for "dashboard", "training status", "show progress".

Save current coaching session state. Writes log, updates state.json and todos.md, commits and pushes if git is configured. Use at the end of each improvement cycle.

Run one full coaching step — start session, choose approach (PCO or IntelligentDesign), improve, save session. This is the skill that /loop calls every 30 minutes. Use when asked to "run a coaching step" or "coach step".

IntelligentDesign improvement iteration. Directly analyzes code and prompts, identifies a specific improvement, implements it, tests, and submits. Use when asked to "review the code", "improve the prompts", or "intelligent design".

PCO improvement iteration. Plays a game, collects experience, runs CvCLearner to propose program patches, tests locally, and submits if improved. Use when asked to "run PCO" or "evolve programs".

Run one improvement iteration. Chooses between PCO and IntelligentDesign based on recent effectiveness, runs the chosen approach, falls back to the other if it fails. Use when asked to "improve the agent", "run a coaching session", or "coach improve".

Start the coaching loop. Initializes state, then runs /coach.step every 30 minutes. Use when asked to "start coaching", "coach the agent", or "start the coaching loop".

Initialize a coaching engagement. Asks for a guide document, brainstorms a coaching strategy, generates guide.md, and creates .coach/ state. Use when starting a new coaching engagement or "coach init".

Use to A/B test two Crewrift policy versions head-to-head and decide whether a change actually helped — RIGHT NOW, against the current field. Triggers: 'did my change help', 'compare v2 vs v1', 'A/B test crewborg', 'is the candidate better than the baseline', 'test this hypothesis', 'did fixing the imposter regress the crewmate'. Runs two matched fresh experience requests, compares hard metrics (compare.py) AND guides a context-driven qualitative investigation of the logs/replays. Distinct from crewrift-report, which surveys ONE batch descriptively.

Use to turn a SET of Crewrift episodes (an experience request, a policy's recent league games, a round/tournament batch) into a dense, high-signal report on a player policy's strengths and weaknesses. Triggers: 'analyze/report on these XP-request episodes', 'where is crewborg weak', 'find the interesting/bad/good games', 'what did crewborg do wrong as imposter', 'summarize this batch of replays/logs', post-mortem of an evaluation. Crewrift-specific (roles, votes, kills, expand_replay). Pair with coworld-episode-artifacts (to pull the episodes) and coworld-experience-requests (to create them).

Use to create and monitor Coworld experience requests — hosted batches of episodes you define (target, roster, roles, count) for evaluating agents against a live roster. Triggers: 'run <policy> vs the top opponents', 'make an experience request', 'request N hosted games', 'A/B a policy against the league', 'set up an evaluation battery / matchup'. Game-agnostic; pair with coworld-episode-artifacts to pull the resulting episodes.

Use to run your own built player policy in a LOCAL Coworld episode and watch it — the Gate-1 smoke test ('did my change take, does it connect → play → exit cleanly?'). Triggers: 'smoke test the player locally', 'run a local game', 'run my policy locally and watch it', 'does the build still work', 'open the replay locally'. Game-agnostic; this is correctness only, NOT a competitive matchup (you can't run other users' policies locally).

Use to identify and download Coworld episode artifacts — replays, results, and per-agent logs — for completed episodes. Triggers: 'download the replay/logs for episode X', 'pull the last N episodes <player> played', 'fetch artifacts for this experience request / pool / round', 'get the episode data for post-mortem'. Game-agnostic (Crewrift, amongthem, etc.).

Use to take a built player image through the Coworld policy lifecycle: upload it as a new version, (gated) submit it to a league, and monitor its live standings. Triggers: 'upload the new policy version', 'submit crewborg to the league', 'is my submission live / champion', 'how is my policy ranking', 'ship the player'. Game-agnostic. Upload is routine and inert; SUBMIT is the public, irreversible, champion-making action — only on explicit human go-ahead.

Use to turn a Crewrift report's signals into evidence-grounded, MECHANISTIC improvement hypotheses — by investigating the replays/logs/code for *why* a behavior happens (or fails to), then presenting candidate directions to the human. Triggers: 'why is crewborg weak at X', 'what should we try / where should we go next', 'form hypotheses', 'diagnose this weakness', 'what's the mechanism behind Y', 'turn this report into directions', 'we got lucky in some games — can we engineer that'. Optional AUGMENTATION of the Report step — it adds candidate hypotheses alongside the report's replay-pointers; it does not replace the human's direction call. Pairs with crewrift-report (signals in) and crewrift-ab (to test a hypothesis).

Create or modify a game brain (decision engine) that subclasses BaseBrain. Use when implementing per-tick decision logic, adding roles, tuning thresholds, or writing the decide() method for a new game plugin. Triggers on "brain", "decide", "role", "threshold", "heuristic", "decision logic".

Configure or run the self-improving code evolution pipeline. Use when setting up evolution for a new game, customizing editable files, tuning the evolution prompt, running manual evolution, or managing the weekly fork/reset lifecycle. Triggers on "evolution", "evolve", "self-improving", "str_replace", "fork", "reset", "promote", "weekly cycle".

Create or modify an LLM harness that subclasses BaseHarness. Use when implementing LLM context building, directive parsing, system prompts, or post-game analysis hooks for a game plugin. Triggers on "harness", "LLM context", "system prompt", "directive", "narrator", "advisor".

Create or modify the game policy (control loop) that subclasses BasePolicy. Use when implementing perceive() to parse game engine observations, execute() to convert commands to game actions, or wiring brain + harness together. Triggers on "policy", "perceive", "execute", "control loop", "game engine bridge".

Run or configure post-game analysis on completed game memory dumps. Use when setting up the analysis pipeline, customizing the analysis prompt, reviewing learnings files, or running manual analysis. Triggers on "analyze", "post-game", "learnings", "score", "review game", "what went wrong".

Create or modify game event triggers that subclass BaseEventDetector. Use when implementing event detection for LLM consultation, adding new trigger types, tuning priorities, or configuring debounce. Triggers on "trigger", "event detection", "consult", "LLM scheduling", "when to call LLM".

Create or update wiki knowledge pages for a game plugin or the framework. Use when writing strategy docs, skill guides, mechanics references, or updating the transferable knowledge base. Triggers on "wiki", "strategy doc", "skill doc", "mechanics doc", "game knowledge", "write documentation".

Diagnose the biggest CvC policy weakness from eval metrics and LLM-Python trace data

Run multi-episode CvC evaluation and produce structured JSON metrics

Implement one targeted CvC code change from analysis, verify improvement, submit if better

Run a CvC game and capture LLM-Python communication trace for analysis

Install softmax-cli, authenticate via softmax login, and find or create a cogames player

Crawl any website and save pages as local markdown files. Use when you need to download documentation, knowledge bases, or web content for offline access or analysis. No code required - just provide a URL.

Extract content from specific URLs using Tavily's extraction API. Returns clean markdown/text from web pages. Use when you have specific URLs and need their content without writing code.

Search the web using Tavily's LLM-optimized search API. Returns relevant results with content snippets, scores, and metadata. Use when you need to find web content on any topic without writing code.

Build production-ready Tavily integrations with best practices baked in. Reference documentation for developers using coding assistants (Claude Code, Cursor, etc.) to implement web search, content extraction, crawling, and research in agentic workflows, RAG systems, or autonomous agents.

Use when a game agent (bot, NPC, AI) appears frozen, oscillates between two adjacent tiles, "moves back and forth without making progress", gets stuck near walls, or stops cooperating partway through a session. Symptom is "stopped doing what it should" rather than "doing the wrong thing".

Use when a game "feels off" but you don't know where — bots dominate that shouldn't, prey never gets caught, scores don't match the design, players die invisibly, or one strategy beats every other. Open-ended diagnostic, not a targeted fix.

Use when iterating on stag_hunt — tuning prey rewards, flee probabilities, capture coalition sizes, energy economy, or testing a server-side mechanic change. Specific to the stag_hunt game in this repo; for the general process see investigating-game-balance and testing-game-balance-change.

Use when you (or the user) have a specific proposed change to a game tunable — "lower this prob 40 → 30", "give rabbit less energy", "make moose flee more at distance 2" — and you need to confirm it has the intended effect without breaking something else.

Launch, list, stop, and inspect games on bitworld's local games_server without using the web UI. Use whenever the user asks to start/create/launch a game, stop/kill/cleanup a game, check what's running, tail container logs, or hit /healthz on a game container. Common manifests include cogs_vs_clips (v2 Coworld) and among_them (v1 bitworld).

Example skill for testing ccpm file asset copying