harn-cli 0.8.37

//! Clap definitions for `harn eval` and its subcommands.
//!
//! The bare form `harn eval <path>` evaluates a run record, run directory,
//! eval manifest, or `.harn` pipeline (legacy entrypoint, dispatched through
//! `eval_run_record`). The `harn eval prompt <file> --fleet <models>`
//! subcommand renders (and optionally runs / judges) a single
//! `.harn.prompt` template against a fleet of models so authors can compare
//! the wire envelope each capability profile materializes.

use std::path::PathBuf;

use clap::{Args, Subcommand, ValueEnum};

#[derive(Debug, Args)]
#[command(args_conflicts_with_subcommands = true)]
pub struct EvalArgs {
    /// Run record path, run directory, eval manifest path, or `.harn` pipeline.
    /// Required unless a subcommand (e.g. `prompt`) is used.
    pub path: Option<String>,
    /// Optional baseline run record for diffing.
    #[arg(long)]
    pub compare: Option<String>,
    /// Run a pipeline twice and compare the baseline against this structural experiment.
    #[arg(long = "structural-experiment")]
    pub structural_experiment: Option<String>,
    /// Replay LLM responses from a JSONL fixture file when `path` is a `.harn` pipeline.
    #[arg(
        long = "llm-mock",
        value_name = "PATH",
        conflicts_with = "llm_mock_record"
    )]
    pub llm_mock: Option<String>,
    /// Record executed LLM responses into a JSONL fixture file when `path` is a `.harn` pipeline.
    #[arg(
        long = "llm-mock-record",
        value_name = "PATH",
        conflicts_with = "llm_mock"
    )]
    pub llm_mock_record: Option<String>,
    /// Positional arguments forwarded to `harn run <pipeline.harn> -- ...` when
    /// `path` is a pipeline file and `--structural-experiment` is set.
    #[arg(last = true)]
    pub argv: Vec<String>,
    #[command(subcommand)]
    pub command: Option<EvalCommand>,
}

#[derive(Debug, Subcommand)]
pub enum EvalCommand {
    /// Benchmark coding-agent fixtures across providers and tool formats.
    CodingAgent(EvalCodingAgentArgs),
    /// Run deterministic context-engineering modes over task fixtures.
    Context(EvalContextArgs),
    /// Render and optionally run a `.harn.prompt` across a fleet of models.
    Prompt(EvalPromptArgs),
    /// Run tool-call accuracy, latency, and cost evals over a dataset.
    ToolCalls(EvalToolCallsArgs),
}

#[derive(Debug, Args)]
pub struct EvalContextArgs {
    /// Context eval manifest JSON or TOML.
    pub manifest: PathBuf,
    /// Output directory for summary.json, per_run.jsonl, and summary.md.
    #[arg(long)]
    pub output: Option<PathBuf>,
    /// Print the aggregate summary JSON to stdout.
    #[arg(long)]
    pub json: bool,
}

#[derive(Debug, Args)]
pub struct EvalCodingAgentArgs {
    /// Fixture ids to run (comma-separated, repeatable). Use `all` for the full suite.
    #[arg(long = "fixture", value_delimiter = ',', default_value = "all")]
    pub fixtures: Vec<String>,
    /// Model selectors to run (comma-separated, repeatable). Each entry may be
    /// an alias, `provider:model`, or `provider=...,model=...`.
    #[arg(long = "model", value_delimiter = ',', default_value = "mock:mock")]
    pub models: Vec<String>,
    /// Tool-call rendering modes to compare.
    #[arg(
        long = "tool-format",
        value_delimiter = ',',
        default_value = "native,text"
    )]
    pub tool_formats: Vec<String>,
    /// Output directory for summary.json, per_run.jsonl, transcripts, and markdown reports.
    #[arg(long)]
    pub output: Option<PathBuf>,
    /// Optional .env file(s) to load for provider credentials. Values are never written to artifacts.
    #[arg(long = "env-file")]
    pub env_files: Vec<PathBuf>,
    /// Append reachable local Ollama/llama.cpp/MLX/vLLM models to the selected matrix.
    #[arg(long = "include-local")]
    pub include_local: bool,
    /// Restrict local discovery to one provider id. Repeatable.
    #[arg(long = "local-provider")]
    pub local_providers: Vec<String>,
    /// Maximum discovered local models to append.
    #[arg(long = "max-local-models", default_value_t = 2)]
    pub max_local_models: usize,
    /// Leave newly-loaded Ollama models running after each local benchmark run.
    #[arg(long = "keep-local-after-run")]
    pub keep_local_after_run: bool,
    /// Stop after N matrix entries, useful for cost-capped smoke runs.
    #[arg(long = "max-runs")]
    pub max_runs: Option<usize>,
    /// Maximum repair-agent loop iterations per run.
    #[arg(long = "max-iterations", default_value_t = 8)]
    pub max_iterations: usize,
    /// Python executable used by the fixture and verification command.
    #[arg(long, default_value = "python3")]
    pub python: String,
    /// Treat missing credentials as an error instead of skipping the run.
    #[arg(long = "fail-on-unauthorized")]
    pub fail_on_unauthorized: bool,
    /// Print the aggregate summary JSON to stdout.
    #[arg(long)]
    pub json: bool,
    /// Optional step_judge config applied to every run in this invocation.
    /// Accepts a preset name (`symmetric-cheap`, `asymmetric`,
    /// `symmetric-strong`) which expands to a known {model, provider}
    /// pair, or `custom:<json>` for a literal JSON object passed verbatim
    /// to `agent_loop({step_judge: ...})`. Omit (or pass `none` / `off`) to disable.
    /// For matrix sweeps across presets, the step-judge experiment driver
    /// at experiments/step-judge/run.sh invokes the eval runner once per
    /// preset and aggregates.
    #[arg(long = "step-judge")]
    pub step_judge: Option<String>,
    /// Override the on_veto remediation shape for the step-judge config
    /// (`replace` or `retain`). Default is `replace`.
    #[arg(long = "step-judge-on-veto")]
    pub step_judge_on_veto: Option<String>,
    /// Use the adversarial rubric variant.
    #[arg(long = "step-judge-adversarial")]
    pub step_judge_adversarial: bool,
    /// Free-form reason attached when forcing a tool format against catalog guidance.
    #[arg(long = "override-reason")]
    pub override_reason: Option<String>,
    /// Structural-validator config applied to every run in this invocation.
    /// Omit to use the suite default (currently the 4-rule validator).
    /// Accepts `on` / `default`, `off` / `none`, or `custom:<json>` for a
    /// literal JSON object passed to `with_structural_validator(...)`.
    #[arg(long = "structural-validator")]
    pub structural_validator: Option<String>,
    /// Free-form label persisted in summary.json for grouping repeat runs
    /// (e.g. "replicate-1", "probe-judge-arch-gpt"). Defaults to empty.
    #[arg(long = "run-label", default_value = "")]
    pub run_label: String,
    /// Path to a previous coding-agent `summary.json` (or its parent dir).
    /// When present, the new summary embeds a `baseline_comparison` block
    /// listing per-fixture regressions (baseline passed but this cell
    /// failed) and recoveries (baseline failed but this cell passed),
    /// plus aggregate counts and a net lift in percentage points.
    /// Useful for cross-cell A/Bs (provider sweep, prompt change, step
    /// judge on/off) where net pass-rate hides destructive interactions
    /// like the cli-help-flag regression the step-judge experiment
    /// surfaced (harn#2318).
    #[arg(long = "baseline-comparison-against")]
    pub baseline_comparison_against: Option<PathBuf>,
}

#[derive(Debug, Args)]
pub struct EvalToolCallsArgs {
    #[command(subcommand)]
    pub command: Option<EvalToolCallsCommand>,
    /// Dataset directory or JSON file. Directories prefer a `cases/` child.
    #[arg(long, default_value = "conformance/tool-call-eval")]
    pub dataset: PathBuf,
    /// Planner model selector: alias, `provider:model`, or `provider=...,model=...`.
    #[arg(long)]
    pub planner: Option<String>,
    /// Optional binder model selector. When set, a second model canonicalizes
    /// the planner's response into a call/refusal decision before scoring.
    #[arg(long)]
    pub binder: Option<String>,
    /// Judge model used only for predicate cases.
    #[arg(long = "judge-model", default_value = "claude-opus-4-7")]
    pub judge_model: String,
    /// Output directory for `summary.json` and `per_case.jsonl`.
    #[arg(long)]
    pub output: Option<PathBuf>,
    /// Override tool rendering for the planner (`native` or `text`).
    #[arg(long = "tool-format")]
    pub tool_format: Option<String>,
    /// Maximum planner response tokens.
    #[arg(long = "max-tokens", default_value_t = 512)]
    pub max_tokens: i64,
    /// Maximum binder response tokens. Default is sized to leave room for
    /// reasoning-emitting models (e.g. GPT-OSS-120B emits ~200 tokens of
    /// chain-of-thought before the JSON payload); non-reasoning binders
    /// will under-fill this budget at no extra cost.
    #[arg(long = "binder-max-tokens", default_value_t = 1024)]
    pub binder_max_tokens: i64,
    /// Run only cases whose id or tag contains this string.
    #[arg(long)]
    pub filter: Option<String>,
    /// Stop after N selected cases, useful for smoke runs.
    #[arg(long = "max-cases")]
    pub max_cases: Option<usize>,
    /// Treat missing credentials as an immediate preflight error.
    #[arg(long = "fail-on-unauthorized")]
    pub fail_on_unauthorized: bool,
}

#[derive(Debug, Subcommand)]
pub enum EvalToolCallsCommand {
    /// Compare a current summary against a pinned baseline.
    RegressionCheck(EvalToolCallsRegressionArgs),
}

#[derive(Debug, Args)]
pub struct EvalToolCallsRegressionArgs {
    /// Current run summary. Defaults to `.harn-runs/tool-call-eval/latest/summary.json`.
    #[arg(long)]
    pub current: Option<PathBuf>,
    /// Optional planner label for diagnostics.
    #[arg(long)]
    pub planner: Option<String>,
    /// Baseline summary JSON to compare against.
    #[arg(long)]
    pub against: PathBuf,
    /// Maximum allowed pass-rate drop in percentage points.
    #[arg(long = "max-drop-pp", default_value_t = 2.0)]
    pub max_drop_pp: f64,
}

#[derive(Debug, Args)]
pub struct EvalPromptArgs {
    /// Path to a `.harn.prompt` (or `.prompt`) template.
    pub file: PathBuf,
    /// Fleet of model selectors (comma-separated, repeatable).
    /// Each entry is either a model alias (`claude-opus-4-7`) or a
    /// `provider:model` selector (`ollama:qwen3.5`). Mutually exclusive
    /// with `--fleet-name`.
    #[arg(
        long,
        value_delimiter = ',',
        required_unless_present = "fleet_name",
        conflicts_with = "fleet_name"
    )]
    pub fleet: Vec<String>,
    /// Named fleet from `harn.toml` `[eval.fleets.<name>]`.
    #[arg(long = "fleet-name")]
    pub fleet_name: Option<String>,
    /// JSON file with bindings injected into the template scope.
    #[arg(long)]
    pub bindings: Option<PathBuf>,
    /// Prompt context-quality fixture(s) that score artifact selection,
    /// stale/noisy rejection, budget adherence, and logical-section shape.
    #[arg(long = "context-fixture")]
    pub context_fixture: Vec<PathBuf>,
    /// Evaluation mode.
    #[arg(long, value_enum, default_value_t = EvalPromptMode::Render)]
    pub mode: EvalPromptMode,
    /// Output format.
    #[arg(long, value_enum, default_value_t = EvalPromptOutput::Terminal)]
    pub output: EvalPromptOutput,
    /// Output destination for HTML / JSON (defaults to stdout).
    #[arg(long = "out-file", short = 'o')]
    pub out_file: Option<PathBuf>,
    /// Maximum concurrent model invocations in run/judge modes.
    #[arg(long, default_value_t = 4)]
    pub max_concurrent: usize,
    /// Optional judge prompt template. When unset, a built-in equivalence
    /// judge is used.
    #[arg(long = "judge-template")]
    pub judge_template: Option<PathBuf>,
    /// Model used for `--mode judge` evaluation.
    #[arg(long = "judge-model", default_value = "claude-opus-4-7")]
    pub judge_model: String,
    /// Maximum tokens for `--mode run` / `--mode judge` calls.
    #[arg(long = "max-tokens", default_value_t = 1024)]
    pub max_tokens: i64,
    /// Treat unauthenticated providers as errors rather than skipping them.
    #[arg(long = "fail-on-unauthorized")]
    pub fail_on_unauthorized: bool,
}

#[derive(Debug, Clone, Copy, ValueEnum)]
pub enum EvalPromptMode {
    /// Render the template against each model's capability profile.
    Render,
    /// Render + execute against each model and collect outputs.
    Run,
    /// Render + run + LLM-as-judge equivalence scoring.
    Judge,
}

#[derive(Debug, Clone, Copy, ValueEnum)]
pub enum EvalPromptOutput {
    Terminal,
    Json,
    Html,
}