/** A chat message in an OpenAI-compatible trajectory. */
export interface ChatMessage {
    role: 'system' | 'user' | 'assistant' | 'tool';
    /** Text content. May be null on an assistant turn that only emits tool_calls. */
    content: string | null;
    /**
     * Assistant tool-call requests (the ReAct action step). PRESERVED verbatim
     * into the SFT set — the model must learn real tool-use trajectories, so this
     * structure is NEVER flattened to plain text.
     */
    tool_calls?: ToolCall[];
    /** On a `role:'tool'` message, the id of the tool_call it answers. */
    tool_call_id?: string;
    /** Optional name (tool name on a tool message; function name otherwise). */
    name?: string;
}
/** An OpenAI-style tool call (the ReAct action). */
export interface ToolCall {
    id: string;
    type: 'function';
    function: {
        name: string;
        arguments: string;
    };
}
/**
 * The tier that produced a trajectory. 'cheap' = the open model the cascade
 * runs first (GLM/Qwen/DeepSeek). 'frontier' = the escalation tier
 * (Opus/GPT/Sonnet). This is the on/off-policy discriminator: only cheap-tier
 * trajectories are on-policy for DPO.
 */
export type PolicyTier = 'cheap' | 'frontier';
/**
 * One trajectory in the Darwin archive: a single solve attempt on a single
 * SWE-bench instance by a single model. This is the clear, documented input
 * contract the exporter codes against. It is reconstructable from the
 * Firestore `darwin_runs` docs + the local prediction/trajectory artifacts
 * (predictions-*.jsonl rows carry instance_id + model_patch; the agentic loop
 * carries the `messages` array with tool_calls — see solve-agentic.mjs).
 */
export interface DarwinTrajectory {
    /** SWE-bench instance id, e.g. "astropy__astropy-14182". The contamination key. */
    instance_id: string;
    /** The model that produced this trajectory, e.g. "z-ai/glm-5.2". */
    model: string;
    /** Which cascade tier this model belongs to (on/off-policy discriminator). */
    tier: PolicyTier;
    /**
     * Gold-resolved status from the OFFICIAL swebench harness (resolved_ids).
     * NEVER from in-loop oracle signals — only conformant gold eval counts as a
     * success for SFT distillation.
     */
    resolved: boolean;
    /**
     * The ReAct message trajectory: system + user(issue) + (assistant tool_calls
     * → tool results)* + final assistant(patch). May be empty for a failed/empty
     * attempt (those become DPO `rejected` candidates).
     */
    messages: ChatMessage[];
    /** The unified diff the attempt produced (may be '' for an empty/failed attempt). */
    model_patch: string;
    /**
     * Best-of-N sample index on this instance (BoN-derived). Lets the DPO pairer
     * find a chosen (resolved) and a rejected (empty/failed) sample from the SAME
     * model on the SAME instance. Default 0 when single-sample.
     */
    sample?: number;
    /** Optional source tag (e.g. "darwin_runs", "predictions-mm25"). Provenance only. */
    source?: string;
}
/**
 * One SFT row — OpenAI chat JSONL. The tool_calls structure is preserved so
 * the model learns real tool-use trajectories (not a flattened transcript).
 */
export interface SftRow {
    messages: ChatMessage[];
}
/**
 * One DPO row — TRL/HF CONVERSATIONAL preference schema. ReAct diverges from
 * the first action, so `prompt` = the shared system+issue messages, and
 * chosen/rejected are FULL trajectories from that divergence point.
 */
export interface DpoRow {
    prompt: ChatMessage[];
    chosen: ChatMessage[];
    rejected: ChatMessage[];
}
export interface ExportOptions {
    /**
     * THE CONTAMINATION GUARD (non-negotiable). Instance ids reserved for
     * evaluation. ANY trajectory whose instance_id is in this set is excluded,
     * and an overlap is asserted against — training on eval instances is fake
     * lift, the exact contamination we debunk elsewhere. Required (may be []).
     */
    evalHoldout: string[];
    /**
     * Max token budget per trajectory (rough word/char heuristic — see
     * `estimateTokens`). Trajectories over budget are DROPPED (or truncated when
     * `truncateOverLength` is set) and REPORTED — never silently lost. Targets a
     * 7-14B context window. Default 28000 (headroom under a 32k window).
     */
    maxTokens?: number;
    /**
     * When true, over-length trajectories are TRUNCATED (oldest tool round-trips
     * dropped, keeping system+issue+final) instead of dropped entirely. The drop
     * is still reported. Default false (drop, the safe default).
     */
    truncateOverLength?: boolean;
    /**
     * THE REWARD-HACKING FILTER (Ornith-1.0 borrow). When true (the DEFAULT), a
     * deterministic monitor runs over each trajectory and DROPS any that read a
     * withheld gold/test path, modified the verification harness, or escaped the
     * sandbox — an archived "success" that secretly reward-hacked would teach the
     * model to reward-hack. This is the training-data analog of the conformance
     * firewall. Set false ONLY for debugging; the count is always reported.
     */
    dropRewardHacked?: boolean;
}
/** What the exporter produced + an honest accounting of what it dropped. */
export interface ExportResult {
    sft: SftRow[];
    dpo: DpoRow[];
    report: ExportReport;
}
export interface ExportReport {
    totalTrajectories: number;
    /** Excluded because their instance_id is in the eval holdout (contamination guard). */
    excludedByHoldout: number;
    /** Dropped because over the token budget (the long-context filter). */
    droppedOverLength: number;
    /** Truncated to fit the token budget (only when truncateOverLength). */
    truncatedOverLength: number;
    /** Dropped by the reward-hacking monitor (gold-read / verification-tamper / sandbox-escape). */
    droppedRewardHacked: number;
    sftRows: number;
    dpoRows: number;
    /** Per-instance ids that ended up in the SFT set (for an audit trail). */
    sftInstanceIds: string[];
    /** Per-instance ids that ended up in the DPO set. */
    dpoInstanceIds: string[];
    /** Human-readable notes (e.g. which trajectories were dropped + why). */
    notes: string[];
}
//# sourceMappingURL=types.d.ts.map