Create Harness

Create an LLM Harness

A harness bridges a game environment and an LLM. It translates game observations into prompts, sends them to the model, and parses the model's response back into a legal game action. The framework in kaggle_environments/core_harness.py handles the LLM call, retry loop, telemetry, and agent lifecycle — you only implement the game-specific logic.

What you're building

A harness module that implements three functions (the GameHarness protocol):

Production wires these three module-level functions into an agent via an external wrapper template — do not write an adapter class or a create_agent_fn(...) line in harness.py. Tests construct their own local adapter when they want end-to-end coverage; see Step 6 for the pattern.

Step 1: Understand the game

Before writing any code, understand the game you're building a harness for:

1. Read the game's interpreter to understand actions, observations, and turn structure.
2. For OpenSpiel games, check for a proxy (*_proxy.py) — if one exists, the observation will be structured JSON rather than raw OpenSpiel text. The proxy's state_dict() method shows exactly what fields the LLM will see. Non-OpenSpiel environments already provide structured JSON observations directly, so no proxy is needed.
3. Identify the action space type:
   • Enumerable (most games): a fixed set of legal moves per turn (e.g., board coordinates, directions, bids). get_legal_moves returns dict[int, str].
   • Free-form (rare): the agent can submit any structured object (e.g., a clue in a word game). get_legal_moves returns None.
   • Mixed (rarest): some turns are enumerable, others are free-form, determined at runtime.
4. Understand the observation dict. The harness receives an observation dict with these standard fields:
   • observationString — the game state (often JSON from a proxy)
   • legalActions — list of action IDs (ints)
   • legalActionStrings — list of human-readable action strings
   • currentPlayer — whose turn it is (-2 means simultaneous)
   • playerId — this agent's player ID
   • isTerminal — whether the game is over
   • serializedGameAndState — full pyspiel state (OpenSpiel games only)

Step 2: Implement get_legal_moves

This function extracts legal actions from the observation and returns them as {action_id: action_string}.

from typing import Any, Mapping, Sequence
from kaggle_environments.core_harness import ParseResult

def get_legal_moves(observation: Mapping[str, Any]) -> dict[int, str] | None:
    """Return {action_id: action_string} for enumerable games, or None for free-form."""
    legal_actions = observation.get("legalActions")
    legal_action_strings = observation.get("legalActionStrings")
    if legal_actions and legal_action_strings:
        return dict(zip(legal_actions, legal_action_strings))
    return {}

For OpenSpiel games with a fallback (when the environment doesn't always provide legalActions):

def get_legal_moves(observation: Mapping[str, Any]) -> dict[int, str]:
    legal_actions = observation.get("legalActions")
    legal_action_strings = observation.get("legalActionStrings")
    if legal_actions and legal_action_strings:
        return dict(zip(legal_actions, legal_action_strings))
    # Fallback: deserialize pyspiel state
    serialized = observation.get("serializedGameAndState", "")
    if not serialized:
        return {}
    _, state = pyspiel.deserialize_game_and_state(serialized)
    return {a: state.action_to_string(a) for a in state.legal_actions()}

For mixed action spaces (like word_association), return None on free-form turns:

def get_legal_moves(observation: Mapping[str, Any]) -> dict[int, str] | None:
    if observation.get("isCluemaster"):
        return None  # Free-form: LLM submits any valid clue
    # Enumerable: return word choices
    words = observation.get("words", [])
    return {i: f"{i}: {words[i]}" for i in range(len(words))}

Step 3: Write the prompt

The prompt is the most important part of a harness — it determines how well the LLM plays. The make_prompt function (sometimes named generate_prompt in older code) builds the full text sent to the model.

Signature

def generate_prompt(
    observation: Mapping[str, Any],
    move_history: list[str],
    previous_response: str | None = None,
    previous_action: str | None = None,
) -> str:

• observation — the current game state
• move_history — list of this agent's past action strings (persists across turns)
• previous_response — the LLM's last response if this is a retry after an illegal move
• previous_action — the illegal move that was attempted

Prompt structure

A good prompt follows this pattern:

1. Game rules — concise explanation of how the game works
2. Current state — the board/situation, parsed from observation["observationString"]
3. Move history — what's happened so far
4. Player identity — which side the LLM is playing
5. Output format — tell the LLM exactly how to format its move
6. Rethink suffix — appended on retries to explain what went wrong

Do not enumerate the legal moves in the prompt. The framework already validates the LLM's response against legalActionStrings and triggers a retry with the rethink suffix on an illegal move, so listing every legal action just bloats the prompt, encourages the model to pick mechanically from the list instead of reasoning about the position, and trivializes the benchmark for games whose challenge is partly finding the legal moves (e.g. checkers' forced captures). Instead, describe the move-legality rules clearly and let the model derive legality from the state. The lone exception is when the legal-action set is not derivable from the visible state at all (e.g. a hidden hand of cards the LLM owns but the rules text cannot enumerate) — in that case, list only what the model cannot infer.

Prompt template example

Read kaggle_environments/envs/open_spiel_env/games/checkers/harness.py — it's the canonical shape: a single *_PROMPT_TEMPLATE constant with explicit {field} placeholders, plus two named rethink templates (RETHINK_ILLEGAL, RETHINK_UNPARSABLE) selected by render_rethink_suffix from core_harness.

The required pieces:

• A main template that interpolates game state, player identity, move history, and a JSON output spec with a concrete For example: line.
• RETHINK_ILLEGAL — fires when the parser extracted a move that wasn't legal. Leads with {previous_action} ("You suggested … but this is not a legal move"). Does NOT include the previous response.
• RETHINK_UNPARSABLE — fires when the parser extracted nothing. Leads with {previous_response} and restates the JSON format with a clean placeholder + concrete example.
• A generate_prompt that builds the main prompt and appends render_rethink_suffix(RETHINK_ILLEGAL, RETHINK_UNPARSABLE, previous_response, previous_action). The helper returns "" on the first attempt, picks the right branch otherwise, and truncates previous_response to the last 500 chars.

For an imperfect-information variant (per-player observation, custom matcher=), use dark_hex/harness.py as the model instead.

Prompt writing tips

• Parse the observation string. If the game has a proxy, observation["observationString"] is JSON — parse it and present the state clearly rather than dumping raw JSON.
• Be specific about the output format. The LLM's response needs to be parseable. JSON with a clear field name (like "move") works well.
• Explain the coordinate system. If the game uses a board, explain notation clearly (e.g., "columns are letters a-h, rows are numbers 1-9").
• Include rules that affect strategy. Don't just list rules mechanically — highlight the ones that matter for making good decisions.
• Don't give strategy advice. The prompt should explain rules and mechanics, not coach the model on how to play. Saying "capture pieces to win" is fine (that's a rule); saying "control the center early" or "prefer defensive moves" is not (that's strategy). The LLM should reason about strategy on its own from the rules and game state.
• Keep the rethink suffix concise. Truncate the previous response to ~500 characters. Include the illegal move attempt and remind the LLM to re-derive a legal move from the state and rules (do not paste in a legal-moves list on retry either — same reasoning as above).
• Branch the rethink on whether previous_action was extracted. Each case wants a different signal back at the model:
  • previous_action is set (parser pulled a value from the JSON) → the action string itself is the most useful signal. Lead with "You suggested move {previous_action} but this is not legal." Do NOT also include the full previous response — that's noise the model has to skim past to find the actual correction signal. Tail with a brief reminder to keep using the same JSON format.
  • previous_action is None (parser found nothing) → there's no action string to show. Lead with the previous response (last 500 chars, not first 500 — the model's conclusion is at the end, not the preamble) so the model can see what it tried, then restate the JSON format with a concrete example. Tail with a brief reminder that the move must also be legal.
  • Pick the suffix at render time based on previous_action. A single one-size-fits-all suffix that always restates the format teaches the wrong fix on illegal-move retries; a single suffix that never restates the format leaves the model in the dark on unparseable retries.
• Make the JSON example unambiguous. In the rethink-unparseable suffix, write the JSON example so the placeholder and the concrete example are clearly separated. ```json `{"move": "<notation>, e.g. 24/23 24/22"}` ``` reads like the value should literally be that whole string. Instead, use a clean placeholder in the fenced block and put a concrete example on its own line right after:

  ```json
  {"move": "<your_move>"}
  ```
  For example: `{"move": "24/23 24/22"}`

• Every concrete claim must trace to a code path. Before you ship a prompt, take each specific field, rule, or value it references and walk it backwards: (a) which proxy state_dict() key produces it? (b) which engine call produces THAT? (c) what params is the env actually loaded with? If you can't trace a claim to a source, it's a phantom — either implement the missing path or delete the claim. The two recurring failure shapes are drift (the field was renamed or the env switched params underneath the prompt) and aspirational copy (the author described a feature they planned but never wired up — e.g. oshi-zumo's docstring once claimed opponent coin counts were "encoded as a hidden suffix" of move_history entries, but no code surfaced them). A prompt that lies is worse than one that says less.
• Cross-check rule claims against the env's actual load_game params. Don't trust the game's name or docstring — read the env factory. Gin rummy's prompt described the "Oklahoma variant" because the author assumed the default, but the env loads gin_rummy with oklahoma=false. Whenever the prompt says "in this variant…" or names a ruleset, confirm the params at the pyspiel.load_game(name, params) call site match.
• Parameterize directional/orientation language on player_id. Sentences like "lower is your goal", "toward the top edge", "first move", or "your stones are at the bottom" are almost always asymmetric — correct for one player and wrong for the other. Render the prompt for player_id=0 AND player_id=1 and diff them; any directional text that's byte-identical between the two is probably the bug. Factor through a per-player helper (e.g. _player_info(player_id) -> (label, code, direction_text) like dark_hex does) so the asymmetry is in one place. Oshi-zumo shipped with "lower is your goal" baked in once; 7.7% of P0 turns echoed the wrong direction.
• For multi-phase games, dispatch on an explicit phase identifier, not legals-shape. When phases share {Pass, Knock}-shaped legals (gin rummy's Wall and Layoff) or any other coincidental legal-action signature, a legals-based fallback will silently misroute. Build a {phase_name: template} table keyed on the engine's own phase string/enum, assert at construction time that every engine phase is covered, and raise on an unknown phase instead of falling through to a default — a new phase should fail loudly, not silently get the wrong prompt.
• Read board dimensions from the observation, not hardcoded constants. Many games support multiple board sizes via env configuration (havannah's board_size, dark_hex's num_rows/num_cols, amazons' build-dependent defaults). A prompt that bakes in "10x10 grid", hardcodes column letters a–j, or assumes a fixed coordinate range will silently lie to the model whenever the env is loaded with a different size. Source dims from the parsed observationString (proxy state_dict() usually exposes board plus num_rows/num_cols or board_size), or — as a fallback — deserialize serializedGameAndState and read state.get_game().get_parameters(). See amazons/harness.py:113 (_board_dims) for the canonical pattern: prefer the actual board grid, fall back to explicit dimension fields, only then a default. Interpolate num_rows / num_cols into the prompt template ("on a {num_rows}x{num_cols} grid") and derive any coordinate-system text from those dims (e.g. compute the column-letter range from num_cols rather than literal "a-j"). Render the prompt at every configured size the env supports and verify each renders correctly.
• Include the full game's move history, not just this agent's moves — and render the actual moves, not a count of them. Two distinct failure modes:
  1. Per-agent only. The framework-provided move_history argument is this agent's past actions only — it omits the opponent's moves entirely. A prompt that uses only this is showing the model half the game.
  2. Count instead of moves. The proxy exposes move_number (or moves_played, turn_count) and the prompt interpolates that — "Moves played so far: 14" — instead of the actual move list. This was the clobber bug: the prompt rendered the count and the model had no way to reconstruct what had been played. A count tells the model "we're 14 moves in" and nothing else; the model needs the actual sequence (a1b1, b3a3, c2c3, ...) to reason about threats, repetitions, and what the opponent has been doing. Sources for the full move list, in order of preference:
  • The proxy's state_dict() exposes a move_history (or action_history, moves, played_moves) field covering both players — see coin_game/harness.py and ant_foraging_arena/harness.py for proxies that surface this and harnesses that render it.
  • If the proxy doesn't expose one, deserialize serializedGameAndState and walk state.history() / state.full_history() to reconstruct it (see chess/harness.py:36 _build_pgn_movetext for a worked example that emits PGN-style movetext from the pyspiel state). For games with no chance phase, play actions alternate from player 0, so per-move player_ids fall out of index parity.
  • Last resort: have the proxy add a move_history field (clobber's proxy did this — see clobber_proxy.py's state_dict()). Don't fall back to the per-agent move_history argument and call it "history" — that's the "Move history framing wrong" anti-pattern. Render the moves themselves ("a1b1, b3a3, c2c3, ..." or PGN-style for chess), readably and with player labels if alternation isn't obvious from order, and label the line accurately — "Moves played so far this game (both players, oldest first): a1b1, b3a3, ..." reads true; "Moves played so far: 14" does not. Keep move_number as a separate field if useful, but never as a substitute for the move list.
• Move history formatting. Show it as a readable list or "None" if empty. Don't let an empty string confuse the model.

move_history_str = ", ".join(move_history) if move_history else "None"

Second pass: compact the prose

Shorter prompts have matched or beaten longer ones across games; padding buries the signal. After the first working draft, do a dedicated compaction pass at a real rendered observation:

• Collapse restatements (same rule in the rules section and output-format section — keep one).
• Cut filler ("Note that…", "Please…") and hedging ("Try your best to…").
• Replace paragraphs with a single declarative sentence where possible.
• Keep every concrete rule, phase distinction, coordinate-system explanation, and format example.

Start compact; only add clarifying prose or worked examples back in when a later evaluation shows the model actually struggling.

Step 4: Implement parse_response

The parser extracts the LLM's chosen move from its text response and matches it to a legal action. This is where robustness matters — LLMs don't always follow instructions perfectly.

Signature and return type

def parse_response(
    response: str,
    legal_action_strings: Sequence[str] | None,
) -> ParseResult:

ParseResult is a frozen dataclass with three fields:

@dataclasses.dataclass(frozen=True)
class ParseResult:
    legal_action: str | None = None   # Matched legal move string (enumerable)
    raw_action: str | None = None     # What the model actually said (for rethink context)
    submission: Any = None            # Parsed object (free-form only)

For enumerable actions: set legal_action to the matched string from legal_action_strings, and raw_action to what the model originally said.

For free-form actions: set submission to the parsed object (e.g., a dict), and raw_action to a string representation.

On failure: return ParseResult(raw_action=<what_was_attempted>) — the framework will retry with the rethink prompt.