Skip to main content

harn_cli/
lib.rs

1#![recursion_limit = "256"]
2
3pub mod acp;
4pub mod cli;
5mod cli_bytecode;
6pub mod commands;
7pub mod config;
8#[doc(hidden)]
9pub mod dispatch;
10pub mod env_guard;
11pub mod format;
12pub mod json_envelope;
13pub mod package;
14mod provider_bootstrap;
15pub mod skill_loader;
16pub mod skill_provenance;
17pub mod test_report;
18pub mod test_runner;
19#[doc(hidden)]
20pub mod tests;
21
22pub use harn_skills::{get_embedded_skill, list_embedded_skills, EmbeddedSkill, SkillFrontmatter};
23
24use clap::{error::ErrorKind, CommandFactory, Parser as ClapParser};
25use std::path::{Path, PathBuf};
26use std::sync::{Arc, Once};
27use std::{env, fs, panic, process, thread};
28
29use cli::{
30    Cli, Command, CompletionShell, EvalCommand, MergeCaptainCommand, MergeCaptainMockCommand,
31    ModelInfoArgs, PackageArtifactsCommand, PackageCacheCommand, PackageCommand,
32    PackageScaffoldCommand, PersonaCommand, PersonaSupervisionCommand, PgCommand, ProvidersCommand,
33    RunsCommand, ServeCommand, SkillCommand, SkillKeyCommand, SkillTrustCommand, SkillsCommand,
34    TimeCommand, ToolCommand,
35};
36use harn_lexer::Lexer;
37use harn_parser::{DiagnosticSeverity, Parser, TypeChecker};
38
39pub const CLI_RUNTIME_STACK_SIZE: usize = 16 * 1024 * 1024;
40
41static BROKEN_PIPE_PANIC_HOOK: Once = Once::new();
42
43/// Install the macro-emitted builtin signature slice into the
44/// `harn_parser` registry the first time any harn-cli entry point parses
45/// or typechecks a script.
46///
47/// Every code path that drives the parser — `run()`, `execute_run()`,
48/// `parse_source_file()`, `analyze_file()`, every test harness — funnels
49/// through this single helper so the registry is always populated by the
50/// time the typechecker reads it. `install_builtin_signatures` is
51/// idempotent on identical `&'static` slices, so repeat calls are
52/// cheap (a `OnceLock::set` that no-ops after the first success).
53///
54/// Tests cannot rely on `run()` having executed, so they must reach the
55/// parser via one of these entry points (which always do call this).
56pub(crate) fn ensure_builtin_signatures_installed() {
57    harn_parser::install_builtin_signatures(harn_vm::stdlib::all_builtin_signatures());
58}
59
60#[cfg(feature = "hostlib")]
61pub(crate) fn install_default_hostlib(vm: &mut harn_vm::Vm) {
62    let _ = harn_hostlib::install_default(vm);
63}
64
65#[cfg(not(feature = "hostlib"))]
66pub(crate) fn install_default_hostlib(_vm: &mut harn_vm::Vm) {}
67
68/// Entry point used by `src/main.rs`. Hosts the CLI runtime thread and
69/// drives the async dispatcher in `async_main`.
70pub fn run() {
71    install_broken_pipe_panic_hook();
72
73    // Defeat rlib dead-code stripping of `#[harn_builtin]`-emitted statics
74    // (linkme issue #36). Without this touch the linker can drop every
75    // builtin's distributed-slice entry, leaving `ALL_BUILTIN_DEFS` empty
76    // and surfacing as a swarm of `HARN-NAM-002` errors at first call.
77    harn_vm::stdlib::force_link();
78
79    ensure_builtin_signatures_installed();
80
81    let handle = thread::Builder::new()
82        .name("harn-cli".to_string())
83        .stack_size(CLI_RUNTIME_STACK_SIZE)
84        .spawn(|| {
85            let runtime = tokio::runtime::Builder::new_multi_thread()
86                .enable_all()
87                .build()
88                .unwrap_or_else(|error| {
89                    eprintln!("failed to start async runtime: {error}");
90                    process::exit(1);
91                });
92            runtime.block_on(async_main());
93            // Drain any queued OTLP exports while the tokio runtime
94            // is still alive. The auto-registered `OtelSink` uses a
95            // batch processor with `runtime::Tokio`; if we let the
96            // runtime drop before this call, in-flight spans never
97            // reach the configured collector. No-op when OTel is not
98            // configured.
99            if let Err(error) = harn_vm::events::shutdown_otel_sink() {
100                eprintln!("[harn] OTel exporter shutdown failed: {error}");
101            }
102        })
103        .unwrap_or_else(|error| {
104            eprintln!("failed to start CLI runtime thread: {error}");
105            process::exit(1);
106        });
107
108    if let Err(payload) = handle.join() {
109        if is_broken_pipe_panic_payload(payload.as_ref()) {
110            process::exit(0);
111        }
112        std::panic::resume_unwind(payload);
113    }
114}
115
116fn install_broken_pipe_panic_hook() {
117    BROKEN_PIPE_PANIC_HOOK.call_once(|| {
118        let previous = panic::take_hook();
119        panic::set_hook(Box::new(move |info| {
120            if is_broken_pipe_panic_payload(info.payload()) {
121                return;
122            }
123            previous(info);
124        }));
125    });
126}
127
128fn is_broken_pipe_panic_payload(payload: &(dyn std::any::Any + Send)) -> bool {
129    let message = if let Some(message) = payload.downcast_ref::<String>() {
130        message.as_str()
131    } else if let Some(message) = payload.downcast_ref::<&str>() {
132        message
133    } else {
134        return false;
135    };
136
137    let print_failure = message.contains("failed printing to stdout")
138        || message.contains("failed printing to stderr");
139    let broken_pipe = message.contains("Broken pipe")
140        || message.contains("os error 32")
141        || message.contains("EPIPE");
142    print_failure && broken_pipe
143}
144
145#[allow(clippy::large_stack_frames)] // dispatch entrypoint owns full Args + per-feature locals.
146async fn async_main() {
147    // Install the OTLP exporter sink before any subcommand runs so a
148    // 20+ minute autonomous session has spans streaming to the
149    // configured collector from the first turn. When neither
150    // `HARN_OTEL_ENDPOINT` nor `OTEL_EXPORTER_OTLP_ENDPOINT` is set
151    // this is a no-op. A misconfigured endpoint logs and continues —
152    // local observability is opt-in and must never fail the run.
153    if let Err(error) = harn_vm::events::install_otel_sink_from_env() {
154        eprintln!("[harn] OTel exporter disabled: {error}");
155    }
156
157    let raw_args = normalize_serve_args(env::args().collect());
158    if raw_args.len() == 2 && raw_args[1].ends_with(".harn") {
159        provider_bootstrap::maybe_seed_ollama_for_run_file(Path::new(&raw_args[1]), false, false)
160            .await;
161        commands::run::run_file(
162            &raw_args[1],
163            false,
164            std::collections::HashSet::new(),
165            Vec::new(),
166            commands::run::CliLlmMockMode::Off,
167            None,
168            commands::run::RunProfileOptions::default(),
169        )
170        .await;
171        return;
172    }
173
174    let cli = match Cli::try_parse_from(&raw_args) {
175        Ok(cli) => cli,
176        Err(error) => {
177            if matches!(
178                error.kind(),
179                ErrorKind::DisplayHelp | ErrorKind::DisplayVersion
180            ) {
181                error.exit();
182            }
183            error.exit();
184        }
185    };
186
187    if cli.json_schemas {
188        commands::json_schemas::run(cli.schema_command.as_deref());
189        return;
190    }
191
192    let Some(subcommand) = cli.command else {
193        // `arg_required_else_help` already shows help when no args are
194        // supplied. We only land here if a top-level flag (e.g. a
195        // future `--version` long flag) parsed without a subcommand.
196        let mut cmd = Cli::command();
197        cmd.print_help().ok();
198        return;
199    };
200    match subcommand {
201        Command::Version(args) => {
202            let exit = run_version(args).await;
203            if exit != 0 {
204                process::exit(exit);
205            }
206        }
207        Command::Upgrade(args) => {
208            if let Err(error) = commands::upgrade::run(args).await {
209                eprintln!("error: {error}");
210                process::exit(1);
211            }
212        }
213        Command::Skill(args) => match args.command {
214            SkillCommand::Key(key_args) => match key_args.command {
215                SkillKeyCommand::Generate(generate) => commands::skill::run_key_generate(&generate),
216            },
217            SkillCommand::Sign(sign) => commands::skill::run_sign(&sign),
218            SkillCommand::Endorse(endorse) => commands::skill::run_endorse(&endorse),
219            SkillCommand::Verify(verify) => commands::skill::run_verify(&verify),
220            SkillCommand::WhoSigned(who_signed) => {
221                commands::skill::run_who_signed(&who_signed).await;
222            }
223            SkillCommand::Trust(trust_args) => match trust_args.command {
224                SkillTrustCommand::Add(add) => commands::skill::run_trust_add(&add),
225                SkillTrustCommand::List(list) => commands::skill::run_trust_list(&list),
226            },
227            SkillCommand::New(new_args) => commands::skills::run_new(&new_args),
228        },
229        Command::Run(args) => {
230            if !args.explain_cost {
231                match (args.eval.as_deref(), args.file.as_deref()) {
232                    (Some(code), None) => {
233                        provider_bootstrap::maybe_seed_ollama_for_inline(
234                            code,
235                            args.yes,
236                            args.llm_mock.is_some(),
237                        )
238                        .await;
239                    }
240                    (None, Some(file)) => {
241                        provider_bootstrap::maybe_seed_ollama_for_run_file(
242                            Path::new(file),
243                            args.yes,
244                            args.llm_mock.is_some(),
245                        )
246                        .await;
247                    }
248                    _ => {}
249                }
250            }
251            let denied =
252                commands::run::build_denied_builtins(args.deny.as_deref(), args.allow.as_deref());
253            let llm_mock_mode = if let Some(path) = args.llm_mock.as_ref() {
254                commands::run::CliLlmMockMode::Replay {
255                    fixture_path: PathBuf::from(path),
256                }
257            } else if let Some(path) = args.llm_mock_record.as_ref() {
258                commands::run::CliLlmMockMode::Record {
259                    fixture_path: PathBuf::from(path),
260                }
261            } else {
262                commands::run::CliLlmMockMode::Off
263            };
264            let attestation = args.attest.then(|| commands::run::RunAttestationOptions {
265                receipt_out: args.receipt_out.as_ref().map(PathBuf::from),
266                agent_id: args.attest_agent.clone(),
267            });
268            let profile_options = run_profile_options(&args.profile);
269            let sandbox_options = if args.no_sandbox {
270                commands::run::RunSandboxOptions::disabled()
271            } else {
272                commands::run::RunSandboxOptions::default()
273            };
274            let json_options = args
275                .json
276                .then_some(commands::run::RunJsonOptions { quiet: args.quiet });
277            let aux_options = commands::run::run_aux_options_from_args(&args);
278            let harnpack_options = commands::run::harnpack::HarnpackRunOptions {
279                allow_unsigned: args.allow_unsigned,
280                dry_run_verify: args.dry_run_verify,
281            };
282
283            if let Some(resume_target) = args.resume.as_deref() {
284                commands::run::run_resume_with_skill_dirs(
285                    resume_target,
286                    args.trace,
287                    denied,
288                    args.argv.clone(),
289                    args.skill_dir.clone(),
290                    llm_mock_mode,
291                    attestation,
292                    profile_options,
293                    sandbox_options.clone(),
294                    json_options,
295                    aux_options,
296                )
297                .await;
298                return;
299            }
300
301            match (args.eval.as_deref(), args.file.as_deref()) {
302                (Some(code), None) => {
303                    if args.allow_unsigned || args.dry_run_verify {
304                        command_error(
305                            "`--allow-unsigned` and `--dry-run-verify` apply to `.harnpack` inputs; \
306                             they cannot be combined with `-e`",
307                        );
308                    }
309                    let (wrapped, tmp) = commands::run::prepare_eval_temp_file(code)
310                        .unwrap_or_else(|e| command_error(&e));
311                    let tmp_path: PathBuf = tmp.path().to_path_buf();
312                    fs::write(&tmp_path, &wrapped).unwrap_or_else(|e| {
313                        command_error(&format!("failed to write temp file for -e: {e}"))
314                    });
315                    let tmp_str = tmp_path.to_string_lossy().into_owned();
316                    if args.explain_cost {
317                        commands::run::run_explain_cost_file_with_skill_dirs(&tmp_str);
318                    } else {
319                        commands::run::run_file_with_skill_dirs(
320                            &tmp_str,
321                            args.trace,
322                            denied,
323                            args.argv.clone(),
324                            args.skill_dir.clone(),
325                            llm_mock_mode.clone(),
326                            attestation.clone(),
327                            profile_options.clone(),
328                            sandbox_options.clone(),
329                            json_options.clone(),
330                            aux_options.clone(),
331                            harnpack_options.clone(),
332                        )
333                        .await;
334                    }
335                    drop(tmp);
336                }
337                (None, Some(file)) => {
338                    if args.explain_cost {
339                        commands::run::run_explain_cost_file_with_skill_dirs(file);
340                    } else {
341                        commands::run::run_file_with_skill_dirs(
342                            file,
343                            args.trace,
344                            denied,
345                            args.argv.clone(),
346                            args.skill_dir.clone(),
347                            llm_mock_mode,
348                            attestation,
349                            profile_options,
350                            sandbox_options,
351                            json_options,
352                            aux_options,
353                            harnpack_options,
354                        )
355                        .await;
356                    }
357                }
358                (Some(_), Some(_)) => command_error(
359                    "`harn run` accepts either `-e <code>` or `<file.harn>`, not both",
360                ),
361                (None, None) => command_error(
362                    "`harn run` requires `--resume <snapshot>`, `-e <code>`, or `<file.harn>`",
363                ),
364            }
365        }
366        Command::Check(args) => {
367            let json_format_alias =
368                !args.json && matches!(args.format, cli::CheckOutputFormat::Json);
369            let matrix_format = if args.json {
370                if !matches!(args.format, cli::CheckOutputFormat::Text) {
371                    command_error("`harn check` accepts either `--json` or `--format`, not both");
372                }
373                cli::CheckOutputFormat::Json
374            } else {
375                args.format
376            };
377            if args.provider_matrix {
378                let cwd = std::env::current_dir().unwrap_or_else(|_| PathBuf::from("."));
379                let extensions = package::load_runtime_extensions(&cwd);
380                package::install_runtime_extensions(&extensions);
381                commands::check::provider_matrix::run(
382                    matrix_format,
383                    args.filter.as_deref(),
384                    json_format_alias,
385                );
386                return;
387            }
388            if args.connector_matrix {
389                commands::check::connector_matrix::run(
390                    matrix_format,
391                    args.filter.as_deref(),
392                    &args.targets,
393                    json_format_alias,
394                );
395                return;
396            }
397            let mut target_strings: Vec<String> = args.targets.clone();
398            if args.workspace {
399                let anchor = target_strings.first().map(Path::new);
400                match package::load_workspace_config(anchor) {
401                    Some((workspace, manifest_dir)) if !workspace.pipelines.is_empty() => {
402                        for pipeline in &workspace.pipelines {
403                            let candidate = Path::new(pipeline);
404                            let resolved = if candidate.is_absolute() {
405                                candidate.to_path_buf()
406                            } else {
407                                manifest_dir.join(candidate)
408                            };
409                            target_strings.push(resolved.to_string_lossy().into_owned());
410                        }
411                    }
412                    Some(_) => command_error(
413                        "--workspace requires `[workspace].pipelines` in the nearest harn.toml",
414                    ),
415                    None => command_error(
416                        "--workspace could not find a harn.toml walking up from the target(s)",
417                    ),
418                }
419            }
420            if target_strings.is_empty() {
421                if args.json {
422                    print_check_error(
423                        "missing_targets",
424                        "`harn check` requires at least one target path, or `--workspace` with `[workspace].pipelines`",
425                    );
426                }
427                command_error(
428                    "`harn check` requires at least one target path, or `--workspace` with `[workspace].pipelines`",
429                );
430            }
431            for target in &target_strings {
432                if let Err(error) = package::validate_runtime_manifest_extensions(Path::new(target))
433                {
434                    if args.json {
435                        print_check_error(
436                            "manifest_extension_error",
437                            &format!("manifest extension validation failed: {error}"),
438                        );
439                    }
440                    command_error(&format!("manifest extension validation failed: {error}"));
441                }
442            }
443            let targets: Vec<&str> = target_strings.iter().map(String::as_str).collect();
444            let files = commands::check::collect_harn_targets(&targets);
445            if files.is_empty() {
446                if args.json {
447                    print_check_error(
448                        "no_harn_files",
449                        "no .harn files found under the given target(s)",
450                    );
451                }
452                command_error("no .harn files found under the given target(s)");
453            }
454            let module_graph = commands::check::build_module_graph(&files);
455            let cross_file_imports = commands::check::collect_cross_file_imports(&module_graph);
456            let mut analysis = harn_parser::analysis::AnalysisDatabase::new();
457            let mut should_fail = false;
458            let mut json_files = Vec::new();
459            for file in &files {
460                let mut config = package::load_check_config(Some(file));
461                if let Some(path) = args.host_capabilities.as_ref() {
462                    config.host_capabilities_path = Some(path.clone());
463                }
464                if let Some(path) = args.bundle_root.as_ref() {
465                    config.bundle_root = Some(path.clone());
466                }
467                if args.strict_types {
468                    config.strict_types = true;
469                }
470                if let Some(sev) = args.preflight.as_deref() {
471                    config.preflight_severity = Some(sev.to_string());
472                }
473                if args.json {
474                    let report = commands::check::check_file_report(
475                        &mut analysis,
476                        file,
477                        &config,
478                        &cross_file_imports,
479                        &module_graph,
480                        args.invariants,
481                    );
482                    should_fail |= report.outcome().should_fail(config.strict);
483                    json_files.push(report);
484                } else {
485                    let outcome = commands::check::check_file_inner(
486                        &mut analysis,
487                        file,
488                        &config,
489                        &cross_file_imports,
490                        &module_graph,
491                        args.invariants,
492                    );
493                    should_fail |= outcome.should_fail(config.strict);
494                }
495            }
496            if args.json {
497                let report = commands::check::CheckReport::from_files(json_files);
498                let envelope = if should_fail {
499                    json_envelope::JsonEnvelope {
500                        schema_version: commands::check::CHECK_SCHEMA_VERSION,
501                        ok: false,
502                        data: Some(report),
503                        error: Some(json_envelope::JsonError {
504                            code: "check_failed".to_string(),
505                            message: "one or more files failed `harn check`".to_string(),
506                            details: serde_json::Value::Null,
507                        }),
508                        warnings: Vec::new(),
509                    }
510                } else {
511                    json_envelope::JsonEnvelope::ok(commands::check::CHECK_SCHEMA_VERSION, report)
512                };
513                println!("{}", json_envelope::to_string_pretty(&envelope));
514                if should_fail {
515                    process::exit(1);
516                }
517                return;
518            }
519            if should_fail {
520                process::exit(1);
521            }
522        }
523        Command::Parse(args) => {
524            if let Err(error) = commands::parse_tokens::run_parse(&args) {
525                command_error(&error);
526            }
527        }
528        Command::Tokens(args) => {
529            if let Err(error) = commands::parse_tokens::run_tokens(&args) {
530                command_error(&error);
531            }
532        }
533        Command::Config(args) => {
534            if let Err(error) = commands::config_cmd::run(args).await {
535                command_error(&error);
536            }
537        }
538        Command::Explain(args) => {
539            let code = commands::explain::run_explain(&args).await;
540            if code != 0 {
541                process::exit(code);
542            }
543        }
544        Command::Fix(args) => {
545            if let Err(error) = commands::fix::run(&args) {
546                if error.is_partial_failure() {
547                    eprintln!("error: {}", error.message());
548                    process::exit(1);
549                }
550                command_error(error.message());
551            }
552        }
553        Command::Contracts(args) => {
554            commands::contracts::handle_contracts_command(args).await;
555        }
556        Command::Connect(args) => {
557            commands::connect::run_connect(*args).await;
558        }
559        Command::Lint(args) => {
560            let targets: Vec<&str> = args.targets.iter().map(String::as_str).collect();
561            let files = commands::check::collect_harn_targets(&targets);
562            let prompt_files = commands::check::collect_prompt_targets(&targets);
563            if files.is_empty() && prompt_files.is_empty() {
564                if args.json {
565                    print_lint_error(
566                        "no_lint_targets",
567                        "no .harn or .harn.prompt files found under the given target(s)",
568                    );
569                }
570                command_error("no .harn or .harn.prompt files found under the given target(s)");
571            }
572            let module_graph = commands::check::build_module_graph(&files);
573            let cross_file_imports = commands::check::collect_cross_file_imports(&module_graph);
574            let mut analysis = harn_parser::analysis::AnalysisDatabase::new();
575            if args.json {
576                // `--json` always reports without modifying source — `--fix`
577                // is intentionally orthogonal to structured output so agents
578                // can plan repairs from the report and apply them in a
579                // follow-up `harn lint --fix` (or `harn fix apply`).
580                let mut should_fail = false;
581                let mut json_files: Vec<commands::check::LintFileReport> = Vec::new();
582                for file in &files {
583                    let mut config = package::load_check_config(Some(file));
584                    commands::check::apply_harn_lint_config(file, &mut config);
585                    let require_header = args.require_file_header
586                        || commands::check::harn_lint_require_file_header(file);
587                    let complexity_threshold =
588                        commands::check::harn_lint_complexity_threshold(file);
589                    let persona_step_allowlist =
590                        commands::check::harn_lint_persona_step_allowlist(file);
591                    let report = commands::check::lint_file_report(
592                        &mut analysis,
593                        file,
594                        &config,
595                        &cross_file_imports,
596                        &module_graph,
597                        require_header,
598                        complexity_threshold,
599                        &persona_step_allowlist,
600                    );
601                    should_fail |= report.outcome().should_fail(config.strict);
602                    json_files.push(report);
603                }
604                let report = commands::check::LintReport::from_files(json_files);
605                let envelope = if should_fail {
606                    json_envelope::JsonEnvelope {
607                        schema_version: commands::check::LINT_SCHEMA_VERSION,
608                        ok: false,
609                        data: Some(report),
610                        error: Some(json_envelope::JsonError {
611                            code: "lint_failed".to_string(),
612                            message: "one or more files failed `harn lint`".to_string(),
613                            details: serde_json::Value::Null,
614                        }),
615                        warnings: Vec::new(),
616                    }
617                } else {
618                    json_envelope::JsonEnvelope::ok(commands::check::LINT_SCHEMA_VERSION, report)
619                };
620                println!("{}", json_envelope::to_string_pretty(&envelope));
621                if should_fail {
622                    process::exit(1);
623                }
624                return;
625            }
626            if args.fix {
627                for file in &files {
628                    let mut config = package::load_check_config(Some(file));
629                    commands::check::apply_harn_lint_config(file, &mut config);
630                    let require_header = args.require_file_header
631                        || commands::check::harn_lint_require_file_header(file);
632                    let complexity_threshold =
633                        commands::check::harn_lint_complexity_threshold(file);
634                    let persona_step_allowlist =
635                        commands::check::harn_lint_persona_step_allowlist(file);
636                    commands::check::lint_fix_file(
637                        &mut analysis,
638                        file,
639                        &config,
640                        &cross_file_imports,
641                        &module_graph,
642                        require_header,
643                        complexity_threshold,
644                        &persona_step_allowlist,
645                    );
646                }
647                for file in &prompt_files {
648                    let threshold =
649                        commands::check::harn_lint_template_variant_branch_threshold(file);
650                    let disabled = commands::check::harn_lint_disabled_rules(file);
651                    // The template lint rules don't carry autofix
652                    // edits yet (intentionally — see
653                    // `template_provider_identity::make_diagnostic`),
654                    // so `--fix` is equivalent to a regular run.
655                    commands::check::lint_prompt_file_inner(file, threshold, &disabled);
656                }
657            } else {
658                let mut should_fail = false;
659                for file in &files {
660                    let mut config = package::load_check_config(Some(file));
661                    commands::check::apply_harn_lint_config(file, &mut config);
662                    let require_header = args.require_file_header
663                        || commands::check::harn_lint_require_file_header(file);
664                    let complexity_threshold =
665                        commands::check::harn_lint_complexity_threshold(file);
666                    let persona_step_allowlist =
667                        commands::check::harn_lint_persona_step_allowlist(file);
668                    let outcome = commands::check::lint_file_inner(
669                        &mut analysis,
670                        file,
671                        &config,
672                        &cross_file_imports,
673                        &module_graph,
674                        require_header,
675                        complexity_threshold,
676                        &persona_step_allowlist,
677                    );
678                    should_fail |= outcome.should_fail(config.strict);
679                }
680                for file in &prompt_files {
681                    let threshold =
682                        commands::check::harn_lint_template_variant_branch_threshold(file);
683                    let disabled = commands::check::harn_lint_disabled_rules(file);
684                    let config = package::load_check_config(Some(file));
685                    let outcome =
686                        commands::check::lint_prompt_file_inner(file, threshold, &disabled);
687                    should_fail |= outcome.should_fail(config.strict);
688                }
689                if should_fail {
690                    process::exit(1);
691                }
692            }
693        }
694        Command::Fmt(args) => {
695            let targets: Vec<&str> = args.targets.iter().map(String::as_str).collect();
696            // Anchor config resolution on the first target; CLI flags
697            // always win over harn.toml values.
698            let anchor = targets.first().map(Path::new).unwrap_or(Path::new("."));
699            let loaded = match config::load_for_path(anchor) {
700                Ok(c) => c,
701                Err(e) => {
702                    eprintln!("warning: {e}");
703                    config::HarnConfig::default()
704                }
705            };
706            let mut opts = harn_fmt::FmtOptions::default();
707            if let Some(w) = loaded.fmt.line_width {
708                opts.line_width = w;
709            }
710            if let Some(w) = loaded.fmt.separator_width {
711                opts.separator_width = w;
712            }
713            if let Some(w) = args.line_width {
714                opts.line_width = w;
715            }
716            if let Some(w) = args.separator_width {
717                opts.separator_width = w;
718            }
719            let mode = commands::check::FmtMode::from_check_flag(args.check);
720            if args.json {
721                let envelope = commands::check::fmt_targets_json(&targets, mode, &opts);
722                let failed = !envelope.ok;
723                println!("{}", json_envelope::to_string_pretty(&envelope));
724                if failed {
725                    process::exit(1);
726                }
727            } else {
728                commands::check::fmt_targets(&targets, mode, &opts);
729            }
730        }
731        Command::Test(args) => {
732            if args.watch && (args.junit.is_some() || args.json_out.is_some()) {
733                command_error(
734                    "`harn test --watch` cannot combine with --junit or --json-out; the watch loop never terminates so the report would never be written",
735                );
736            }
737            if args.target.as_deref() == Some("agents-conformance") {
738                if args.selection.is_some() {
739                    command_error(
740                        "`harn test agents-conformance` does not accept a second positional target; use --category instead",
741                    );
742                }
743                if args.evals || args.determinism || args.record || args.replay || args.watch {
744                    command_error(
745                        "`harn test agents-conformance` cannot be combined with --evals, --determinism, --record, --replay, or --watch",
746                    );
747                }
748                let Some(target_url) = args.agents_target.clone() else {
749                    command_error("`harn test agents-conformance` requires --target <url>");
750                };
751                commands::agents_conformance::run_agents_conformance(
752                    commands::agents_conformance::AgentsConformanceConfig {
753                        target_url,
754                        api_key: args.agents_api_key.clone(),
755                        categories: args.agents_category.clone(),
756                        timeout_ms: args.timeout,
757                        verbose: args.verbose,
758                        json: args.json,
759                        json_out: args.json_out.clone(),
760                        workspace_id: args.agents_workspace_id.clone(),
761                        session_id: args.agents_session_id.clone(),
762                    },
763                )
764                .await;
765                return;
766            }
767            if args.target.as_deref() == Some("protocols") {
768                if args.evals || args.determinism || args.record || args.replay || args.watch {
769                    command_error(
770                        "`harn test protocols` cannot be combined with --evals, --determinism, --record, --replay, or --watch",
771                    );
772                }
773                if args.junit.is_some()
774                    || args.agents_target.is_some()
775                    || args.agents_api_key.is_some()
776                    || !args.agents_category.is_empty()
777                    || args.json
778                    || args.json_out.is_some()
779                    || args.agents_workspace_id.is_some()
780                    || args.agents_session_id.is_some()
781                    || args.parallel
782                    || !args.skill_dir.is_empty()
783                {
784                    command_error(
785                        "`harn test protocols` accepts only --filter, --verbose, --timing, and an optional fixture selection",
786                    );
787                }
788                commands::protocol_conformance::run_protocol_conformance(
789                    args.selection.as_deref(),
790                    args.filter.as_deref(),
791                    args.verbose || args.timing,
792                );
793                return;
794            }
795            if args.evals {
796                if args.determinism || args.record || args.replay || args.watch {
797                    command_error("--evals cannot be combined with --determinism, --record, --replay, or --watch");
798                }
799                if args.target.as_deref() != Some("package") || args.selection.is_some() {
800                    command_error("package evals are run with `harn test package --evals`");
801                }
802                run_package_evals();
803            } else if args.determinism {
804                let cli_skill_dirs: Vec<PathBuf> =
805                    args.skill_dir.iter().map(PathBuf::from).collect();
806                if args.watch {
807                    command_error("--determinism cannot be combined with --watch");
808                }
809                if args.record || args.replay {
810                    command_error("--determinism manages its own record/replay cycle");
811                }
812                if let Some(t) = args.target.as_deref() {
813                    if t == "conformance" {
814                        commands::test::run_conformance_determinism_tests(
815                            t,
816                            args.selection.as_deref(),
817                            args.filter.as_deref(),
818                            args.timeout,
819                            &cli_skill_dirs,
820                        )
821                        .await;
822                    } else if args.selection.is_some() {
823                        command_error(
824                            "only `harn test conformance` accepts a second positional target",
825                        );
826                    } else {
827                        commands::test::run_determinism_tests(
828                            t,
829                            args.filter.as_deref(),
830                            args.timeout,
831                            &cli_skill_dirs,
832                        )
833                        .await;
834                    }
835                } else {
836                    let test_dir = if PathBuf::from("tests").is_dir() {
837                        "tests".to_string()
838                    } else {
839                        command_error("no path specified and no tests/ directory found");
840                    };
841                    if args.selection.is_some() {
842                        command_error(
843                            "only `harn test conformance` accepts a second positional target",
844                        );
845                    }
846                    commands::test::run_determinism_tests(
847                        &test_dir,
848                        args.filter.as_deref(),
849                        args.timeout,
850                        &cli_skill_dirs,
851                    )
852                    .await;
853                }
854            } else {
855                let cli_skill_dirs: Vec<PathBuf> =
856                    args.skill_dir.iter().map(PathBuf::from).collect();
857                if args.record {
858                    harn_vm::llm::set_replay_mode(
859                        harn_vm::llm::LlmReplayMode::Record,
860                        ".harn-fixtures",
861                    );
862                } else if args.replay {
863                    harn_vm::llm::set_replay_mode(
864                        harn_vm::llm::LlmReplayMode::Replay,
865                        ".harn-fixtures",
866                    );
867                }
868
869                if let Some(t) = args.target.as_deref() {
870                    if t == "conformance" {
871                        commands::test::run_conformance_tests(
872                            t,
873                            args.selection.as_deref(),
874                            args.filter.as_deref(),
875                            args.junit.as_deref(),
876                            args.timeout,
877                            commands::test::ConformanceRunOptions {
878                                verbose: args.verbose,
879                                timing: args.timing,
880                                differential_optimizations: args.differential_optimizations,
881                                json: args.json,
882                                cli_skill_dirs: &cli_skill_dirs,
883                            },
884                        )
885                        .await;
886                    } else if args.selection.is_some() {
887                        command_error(
888                            "only `harn test conformance` accepts a second positional target",
889                        );
890                    } else {
891                        let run_args = commands::test::UserTestRunArgs {
892                            filter: args.filter.as_deref(),
893                            timeout_ms: args.timeout,
894                            parallel: args.parallel,
895                            jobs: args.jobs,
896                            verbose: args.verbose,
897                            timing: args.timing,
898                            diagnose: args.diagnose,
899                            cli_skill_dirs: &cli_skill_dirs,
900                        };
901                        if args.watch {
902                            commands::test::run_watch_tests(t, run_args).await;
903                        } else {
904                            commands::test::run_user_tests(
905                                t,
906                                run_args,
907                                commands::test::UserTestReportConfig {
908                                    junit_path: args.junit.as_deref(),
909                                    json_out_path: args.json_out.as_deref(),
910                                },
911                            )
912                            .await;
913                        }
914                    }
915                } else {
916                    let test_dir = if PathBuf::from("tests").is_dir() {
917                        "tests".to_string()
918                    } else {
919                        command_error("no path specified and no tests/ directory found");
920                    };
921                    if args.selection.is_some() {
922                        command_error(
923                            "only `harn test conformance` accepts a second positional target",
924                        );
925                    }
926                    let run_args = commands::test::UserTestRunArgs {
927                        filter: args.filter.as_deref(),
928                        timeout_ms: args.timeout,
929                        parallel: args.parallel,
930                        jobs: args.jobs,
931                        verbose: args.verbose,
932                        timing: args.timing,
933                        diagnose: args.diagnose,
934                        cli_skill_dirs: &cli_skill_dirs,
935                    };
936                    if args.watch {
937                        commands::test::run_watch_tests(&test_dir, run_args).await;
938                    } else {
939                        commands::test::run_user_tests(
940                            &test_dir,
941                            run_args,
942                            commands::test::UserTestReportConfig {
943                                junit_path: args.junit.as_deref(),
944                                json_out_path: args.json_out.as_deref(),
945                            },
946                        )
947                        .await;
948                    }
949                }
950            }
951        }
952        Command::Init(args) => {
953            commands::init::init_project(args.name.as_deref(), args.template).await;
954        }
955        Command::New(args) => match commands::init::resolve_new_args(&args) {
956            Ok((name, template)) => commands::init::init_project(name.as_deref(), template).await,
957            Err(error) => {
958                eprintln!("error: {error}");
959                process::exit(1);
960            }
961        },
962        Command::Doctor(args) => {
963            commands::doctor::run_doctor_with_options(commands::doctor::DoctorOptions {
964                json: args.json,
965                check_providers: args.check_providers,
966                check_targets: args.check_targets,
967            })
968            .await;
969        }
970        Command::Models(args) => commands::models::run(args).await,
971        Command::Local(args) => commands::local::run(args).await,
972        Command::Providers(args) => match args.command {
973            ProvidersCommand::Refresh(refresh) => {
974                if let Err(error) = commands::providers::run_refresh(&refresh).await {
975                    command_error(&error);
976                }
977            }
978            ProvidersCommand::Validate(validate) => {
979                if let Err(error) = commands::providers::run_validate(&validate) {
980                    command_error(&error);
981                }
982            }
983            ProvidersCommand::Export(export) => {
984                if let Err(error) = commands::providers::run_export(&export) {
985                    command_error(&error);
986                }
987            }
988            ProvidersCommand::Matrix(matrix) => {
989                if let Err(error) = commands::providers::run_matrix(&matrix) {
990                    command_error(&error);
991                }
992            }
993            ProvidersCommand::Support(support) => {
994                if let Err(error) = commands::provider_support::run(&support) {
995                    command_error(&error);
996                }
997            }
998            ProvidersCommand::Recommend(recommend) => {
999                if let Err(error) = commands::providers::run_recommend(&recommend).await {
1000                    command_error(&error);
1001                }
1002            }
1003        },
1004        Command::Provider(args) => commands::provider_capabilities::run_or_exit(args),
1005        Command::Try(args) => commands::try_cmd::run(args).await,
1006        Command::Quickstart(args) => {
1007            if let Err(error) = commands::quickstart::run_quickstart(&args).await {
1008                command_error(&error);
1009            }
1010        }
1011        Command::Demo(args) => {
1012            let code = commands::demo::run(args).await;
1013            if code != 0 {
1014                process::exit(code);
1015            }
1016        }
1017        Command::Serve(args) => match args.command {
1018            ServeCommand::Acp(args) => {
1019                if let Err(error) = commands::serve::run_acp_server(&args).await {
1020                    command_error(&error);
1021                }
1022            }
1023            ServeCommand::A2a(args) => {
1024                if let Err(error) = commands::serve::run_a2a_server(&args).await {
1025                    command_error(&error);
1026                }
1027            }
1028            ServeCommand::Api(args) => {
1029                if let Err(error) = commands::serve::run_api_server(&args).await {
1030                    command_error(&error);
1031                }
1032            }
1033            ServeCommand::Mcp(args) => {
1034                if let Err(error) = commands::serve::run_mcp_server(&args).await {
1035                    command_error(&error);
1036                }
1037            }
1038        },
1039        Command::Connector(args) => {
1040            if let Err(error) = commands::connector::handle_connector_command(args).await {
1041                eprintln!("error: {error}");
1042                process::exit(1);
1043            }
1044        }
1045        Command::Mcp(args) => commands::mcp::handle_mcp_command(&args.command).await,
1046        Command::Watch(args) => {
1047            let denied =
1048                commands::run::build_denied_builtins(args.deny.as_deref(), args.allow.as_deref());
1049            commands::run::run_watch(&args.file, denied).await;
1050        }
1051        Command::Dev(args) => {
1052            commands::dev::run(args).await;
1053        }
1054        Command::Portal(args) => {
1055            commands::portal::run_portal(
1056                &args.dir,
1057                args.manifest,
1058                args.persona_state_dir,
1059                &args.host,
1060                args.port,
1061                args.open,
1062                args.allow_remote_launch,
1063            )
1064            .await;
1065        }
1066        Command::Trigger(args) => {
1067            if let Err(error) = commands::trigger::handle(args).await {
1068                eprintln!("error: {error}");
1069                process::exit(1);
1070            }
1071        }
1072        Command::Graph(args) => {
1073            let code = commands::graph::run(args).await;
1074            if code != 0 {
1075                process::exit(code);
1076            }
1077        }
1078        Command::Routes(args) => {
1079            let code = commands::routes::run(args).await;
1080            if code != 0 {
1081                process::exit(code);
1082            }
1083        }
1084        Command::Flow(args) => match commands::flow::run_flow(&args) {
1085            Ok(code) => {
1086                if code != 0 {
1087                    process::exit(code);
1088                }
1089            }
1090            Err(error) => command_error(&error),
1091        },
1092        Command::Workflow(args) => match commands::workflow::handle(args) {
1093            Ok(code) => {
1094                if code != 0 {
1095                    process::exit(code);
1096                }
1097            }
1098            Err(error) => command_error(&error),
1099        },
1100        Command::Supervisor(args) => {
1101            if let Err(error) = commands::supervisor::handle(args).await {
1102                eprintln!("error: {error}");
1103                process::exit(1);
1104            }
1105        }
1106        Command::Trace(args) => {
1107            if let Err(error) = commands::trace::handle(args).await {
1108                eprintln!("error: {error}");
1109                process::exit(1);
1110            }
1111        }
1112        Command::Crystallize(args) => {
1113            if let Err(error) = commands::crystallize::run(args) {
1114                eprintln!("error: {error}");
1115                process::exit(1);
1116            }
1117        }
1118        Command::Trust(args) | Command::TrustGraph(args) => {
1119            if let Err(error) = commands::trust::handle(args).await {
1120                eprintln!("error: {error}");
1121                process::exit(1);
1122            }
1123        }
1124        Command::Verify(args) => {
1125            if let Err(error) = verify_provenance_receipt(&args.receipt, args.json) {
1126                eprintln!("error: {error}");
1127                process::exit(1);
1128            }
1129        }
1130        Command::Completions(args) => print_completions(args.shell),
1131        Command::Orchestrator(args) => {
1132            if let Err(error) = commands::orchestrator::handle(args).await {
1133                eprintln!("error: {error}");
1134                process::exit(1);
1135            }
1136        }
1137        Command::Playground(args) => {
1138            provider_bootstrap::maybe_seed_ollama_for_playground(
1139                Path::new(&args.host),
1140                Path::new(&args.script),
1141                args.yes,
1142                args.llm.is_some(),
1143                args.llm_mock.is_some(),
1144            )
1145            .await;
1146            let llm_mock_mode = if let Some(path) = args.llm_mock.as_ref() {
1147                commands::run::CliLlmMockMode::Replay {
1148                    fixture_path: PathBuf::from(path),
1149                }
1150            } else if let Some(path) = args.llm_mock_record.as_ref() {
1151                commands::run::CliLlmMockMode::Record {
1152                    fixture_path: PathBuf::from(path),
1153                }
1154            } else {
1155                commands::run::CliLlmMockMode::Off
1156            };
1157            if let Err(error) = commands::playground::run_command(args, llm_mock_mode).await {
1158                eprint!("{error}");
1159                process::exit(1);
1160            }
1161        }
1162        Command::Runs(args) => match args.command {
1163            RunsCommand::Inspect(inspect) => {
1164                inspect_run_record(&inspect.path, inspect.compare.as_deref());
1165            }
1166        },
1167        Command::Session(args) => commands::session::run(args),
1168        Command::Replay(args) => {
1169            let exit = commands::replay::run(args);
1170            if exit != 0 {
1171                process::exit(exit);
1172            }
1173        }
1174        Command::Eval(args) => match args.command {
1175            Some(EvalCommand::CodingAgent(coding_agent_args)) => {
1176                let code = commands::eval_coding_agent::run(coding_agent_args).await;
1177                if code != 0 {
1178                    process::exit(code);
1179                }
1180            }
1181            Some(EvalCommand::Context(context_args)) => {
1182                let code = commands::eval_context::run(context_args).await;
1183                if code != 0 {
1184                    process::exit(code);
1185                }
1186            }
1187            Some(EvalCommand::Prompt(prompt_args)) => {
1188                let code = commands::eval_prompt::run(prompt_args).await;
1189                if code != 0 {
1190                    process::exit(code);
1191                }
1192            }
1193            Some(EvalCommand::ScopeTriage(scope_args)) => {
1194                process::exit(commands::eval_scope_triage::run(scope_args).await)
1195            }
1196            Some(EvalCommand::ToolCalls(tool_calls_args)) => {
1197                let code = commands::eval_tool_calls::run(tool_calls_args).await;
1198                if code != 0 {
1199                    process::exit(code);
1200                }
1201            }
1202            None => {
1203                let Some(path) = args.path else {
1204                    eprintln!("error: `harn eval` requires a path or a subcommand (e.g. `prompt`).\nSee `harn eval --help`.");
1205                    process::exit(2);
1206                };
1207                let llm_mock_mode = if let Some(path) = args.llm_mock.as_ref() {
1208                    commands::run::CliLlmMockMode::Replay {
1209                        fixture_path: PathBuf::from(path),
1210                    }
1211                } else if let Some(path) = args.llm_mock_record.as_ref() {
1212                    commands::run::CliLlmMockMode::Record {
1213                        fixture_path: PathBuf::from(path),
1214                    }
1215                } else {
1216                    commands::run::CliLlmMockMode::Off
1217                };
1218                eval_run_record(
1219                    &path,
1220                    args.compare.as_deref(),
1221                    args.structural_experiment.as_deref(),
1222                    &args.argv,
1223                    &llm_mock_mode,
1224                );
1225            }
1226        },
1227        Command::Repl => commands::repl::run_repl().await,
1228        Command::Bench(args) => commands::bench::run(args).await,
1229        Command::Precompile(args) => commands::precompile::run(args).await,
1230        Command::Pack(args) => commands::pack::run(args),
1231        Command::TestBench(args) => commands::test_bench::run(args.command).await,
1232        Command::Viz(args) => commands::viz::run_viz(&args.file, args.output.as_deref()),
1233        Command::Install(args) => package::install_packages(
1234            args.frozen || args.locked || args.offline,
1235            args.refetch.as_deref(),
1236            args.offline,
1237            args.json,
1238        ),
1239        Command::Add(args) => package::add_package_with_registry(
1240            &args.name_or_spec,
1241            args.alias.as_deref(),
1242            args.git.as_deref(),
1243            args.tag.as_deref(),
1244            args.rev.as_deref(),
1245            args.branch.as_deref(),
1246            args.path.as_deref(),
1247            args.registry.as_deref(),
1248        ),
1249        Command::Update(args) => {
1250            package::update_packages(args.alias.as_deref(), args.all, args.json);
1251        }
1252        Command::Remove(args) => package::remove_package(&args.alias),
1253        Command::Lock => package::lock_packages(),
1254        Command::Package(args) => match args.command {
1255            PackageCommand::List(list) => package::list_packages(list.json),
1256            PackageCommand::Doctor(doctor) => package::doctor_packages(doctor.json),
1257            PackageCommand::Search(search) => package::search_package_registry(
1258                search.query.as_deref(),
1259                search.registry.as_deref(),
1260                search.json,
1261            ),
1262            PackageCommand::Info(info) => {
1263                package::show_package_registry_info(
1264                    &info.name,
1265                    info.registry.as_deref(),
1266                    info.json,
1267                );
1268            }
1269            PackageCommand::Check(check) => {
1270                package::check_package(check.package.as_deref(), check.json);
1271            }
1272            PackageCommand::Pack(pack) => package::pack_package(
1273                pack.package.as_deref(),
1274                pack.output.as_deref(),
1275                pack.dry_run,
1276                pack.json,
1277            ),
1278            PackageCommand::Docs(docs) => package::generate_package_docs(
1279                docs.package.as_deref(),
1280                docs.output.as_deref(),
1281                docs.check,
1282            ),
1283            PackageCommand::Cache(cache) => match cache.command {
1284                PackageCacheCommand::List => package::list_package_cache(),
1285                PackageCacheCommand::Clean(clean) => package::clean_package_cache(clean.all),
1286                PackageCacheCommand::Verify(verify) => {
1287                    package::verify_package_cache(verify.materialized);
1288                }
1289            },
1290            PackageCommand::Outdated(args) => package::outdated_packages(
1291                args.refresh,
1292                args.remote,
1293                args.registry.as_deref(),
1294                args.json,
1295            ),
1296            PackageCommand::Audit(args) => {
1297                package::audit_packages(
1298                    args.registry.as_deref(),
1299                    args.skip_materialized,
1300                    args.json,
1301                );
1302            }
1303            PackageCommand::Artifacts(args) => match args.command {
1304                PackageArtifactsCommand::Manifest(manifest) => {
1305                    package::artifacts_manifest(manifest.output.as_deref());
1306                }
1307                PackageArtifactsCommand::Check(check) => {
1308                    package::artifacts_check(&check.manifest, check.json);
1309                }
1310            },
1311            PackageCommand::Scaffold(args) => match args.command {
1312                PackageScaffoldCommand::Openapi(openapi) => {
1313                    if let Err(error) = commands::package_scaffold::run_openapi(&openapi).await {
1314                        eprintln!("error: {error}");
1315                        process::exit(1);
1316                    }
1317                }
1318            },
1319        },
1320        Command::Publish(args) => package::publish_package(
1321            args.package.as_deref(),
1322            args.dry_run,
1323            &args.remote,
1324            &args.index_repo,
1325            &args.index_path,
1326            args.registry_name.as_deref(),
1327            args.skip_index_pr,
1328            args.registry.as_deref(),
1329            args.json,
1330        ),
1331        Command::MergeCaptain(args) => match args.command {
1332            MergeCaptainCommand::Run(run) => {
1333                let code = commands::merge_captain::run_driver(&run);
1334                if code != 0 {
1335                    process::exit(code);
1336                }
1337            }
1338            MergeCaptainCommand::Ladder(ladder) => {
1339                let code = commands::merge_captain::run_ladder(&ladder);
1340                if code != 0 {
1341                    process::exit(code);
1342                }
1343            }
1344            MergeCaptainCommand::Iterate(iterate) => {
1345                let code = commands::merge_captain::run_iterate(&iterate);
1346                if code != 0 {
1347                    process::exit(code);
1348                }
1349            }
1350            MergeCaptainCommand::Audit(audit) => {
1351                let code = commands::merge_captain::run_audit(&audit);
1352                if code != 0 {
1353                    process::exit(code);
1354                }
1355            }
1356            MergeCaptainCommand::Mock(mock) => {
1357                let code = match mock {
1358                    MergeCaptainMockCommand::Init(args) => {
1359                        commands::merge_captain_mock::run_init(&args)
1360                    }
1361                    MergeCaptainMockCommand::Step(args) => {
1362                        commands::merge_captain_mock::run_step(&args)
1363                    }
1364                    MergeCaptainMockCommand::Status(args) => {
1365                        commands::merge_captain_mock::run_status(&args)
1366                    }
1367                    MergeCaptainMockCommand::Serve(args) => {
1368                        commands::merge_captain_mock::run_serve(&args).await
1369                    }
1370                    MergeCaptainMockCommand::Cleanup(args) => {
1371                        commands::merge_captain_mock::run_cleanup(&args)
1372                    }
1373                    MergeCaptainMockCommand::Scenarios => {
1374                        commands::merge_captain_mock::run_scenarios()
1375                    }
1376                };
1377                if code != 0 {
1378                    process::exit(code);
1379                }
1380            }
1381        },
1382        Command::Pg(args) => match args.command {
1383            PgCommand::Codegen(codegen) => {
1384                let code = commands::pg_codegen::run(&codegen);
1385                if code != 0 {
1386                    process::exit(code);
1387                }
1388            }
1389        },
1390        Command::Persona(args) => match args.command {
1391            PersonaCommand::New(new) => {
1392                if let Err(error) = commands::persona_scaffold::run_new(&new) {
1393                    eprintln!("error: {error}");
1394                    process::exit(1);
1395                }
1396            }
1397            PersonaCommand::Doctor(doctor) => {
1398                if let Err(error) =
1399                    commands::persona_doctor::run_doctor(args.manifest.as_deref(), &doctor).await
1400                {
1401                    eprintln!("error: {error}");
1402                    process::exit(1);
1403                }
1404            }
1405            PersonaCommand::Check(check) => {
1406                commands::persona::run_check(args.manifest.as_deref(), &check);
1407            }
1408            PersonaCommand::List(list) => {
1409                commands::persona::run_list(args.manifest.as_deref(), &list);
1410            }
1411            PersonaCommand::Inspect(inspect) => {
1412                commands::persona::run_inspect(args.manifest.as_deref(), &inspect);
1413            }
1414            PersonaCommand::Status(status) => {
1415                if let Err(error) = commands::persona::run_status(
1416                    args.manifest.as_deref(),
1417                    &args.state_dir,
1418                    &status,
1419                )
1420                .await
1421                {
1422                    eprintln!("error: {error}");
1423                    process::exit(1);
1424                }
1425            }
1426            PersonaCommand::Pause(control) => {
1427                if let Err(error) = commands::persona::run_pause(
1428                    args.manifest.as_deref(),
1429                    &args.state_dir,
1430                    &control,
1431                )
1432                .await
1433                {
1434                    eprintln!("error: {error}");
1435                    process::exit(1);
1436                }
1437            }
1438            PersonaCommand::Resume(control) => {
1439                if let Err(error) = commands::persona::run_resume(
1440                    args.manifest.as_deref(),
1441                    &args.state_dir,
1442                    &control,
1443                )
1444                .await
1445                {
1446                    eprintln!("error: {error}");
1447                    process::exit(1);
1448                }
1449            }
1450            PersonaCommand::Disable(control) => {
1451                if let Err(error) = commands::persona::run_disable(
1452                    args.manifest.as_deref(),
1453                    &args.state_dir,
1454                    &control,
1455                )
1456                .await
1457                {
1458                    eprintln!("error: {error}");
1459                    process::exit(1);
1460                }
1461            }
1462            PersonaCommand::Tick(tick) => {
1463                if let Err(error) =
1464                    commands::persona::run_tick(args.manifest.as_deref(), &args.state_dir, &tick)
1465                        .await
1466                {
1467                    eprintln!("error: {error}");
1468                    process::exit(1);
1469                }
1470            }
1471            PersonaCommand::Trigger(trigger) => {
1472                if let Err(error) = commands::persona::run_trigger(
1473                    args.manifest.as_deref(),
1474                    &args.state_dir,
1475                    &trigger,
1476                )
1477                .await
1478                {
1479                    eprintln!("error: {error}");
1480                    process::exit(1);
1481                }
1482            }
1483            PersonaCommand::Spend(spend) => {
1484                if let Err(error) =
1485                    commands::persona::run_spend(args.manifest.as_deref(), &args.state_dir, &spend)
1486                        .await
1487                {
1488                    eprintln!("error: {error}");
1489                    process::exit(1);
1490                }
1491            }
1492            PersonaCommand::Supervision(supervision) => match supervision.command {
1493                PersonaSupervisionCommand::Tail(tail) => {
1494                    if let Err(error) = commands::persona_supervision::run_tail(
1495                        args.manifest.as_deref(),
1496                        &args.state_dir,
1497                        &tail,
1498                    )
1499                    .await
1500                    {
1501                        eprintln!("error: {error}");
1502                        process::exit(1);
1503                    }
1504                }
1505            },
1506        },
1507        Command::ModelInfo(args) => {
1508            if !print_model_info(&args).await {
1509                process::exit(1);
1510            }
1511        }
1512        Command::ProviderCatalog(args) => {
1513            if std::env::var("HARN_CLI_IMPL").as_deref() == Ok("rust") {
1514                print_provider_catalog(args.available_only);
1515            } else {
1516                let exit_code = dispatch_provider_catalog(args.available_only).await;
1517                if exit_code != 0 {
1518                    process::exit(exit_code);
1519                }
1520            }
1521        }
1522        Command::ProviderReady(args) => {
1523            run_provider_ready(
1524                &args.provider,
1525                args.model.as_deref(),
1526                args.base_url.as_deref(),
1527                args.json,
1528            )
1529            .await;
1530        }
1531        Command::ProviderProbe(args) => commands::provider::run_provider_probe(args).await,
1532        Command::ProviderToolProbe(args) => commands::provider::run_provider_tool_probe(args).await,
1533        Command::Skills(args) => match args.command {
1534            SkillsCommand::List(list) => commands::skills::run_list(&list),
1535            SkillsCommand::Get(get) => commands::skills::run_get(&get),
1536            SkillsCommand::Dump(dump) => commands::skills::run_dump(&dump),
1537            SkillsCommand::Resolved(resolved) => commands::skills::run_resolved(&resolved),
1538            SkillsCommand::Inspect(inspect) => commands::skills::run_inspect(&inspect),
1539            SkillsCommand::Match(matcher) => commands::skills::run_match(&matcher),
1540            SkillsCommand::Install(install) => commands::skills::run_install(&install),
1541            SkillsCommand::New(new_args) => commands::skills::run_new(&new_args),
1542        },
1543        Command::Tool(args) => match args.command {
1544            ToolCommand::New(new_args) => {
1545                if let Err(error) = commands::tool::run_new(&new_args).await {
1546                    eprintln!("error: {error}");
1547                    process::exit(1);
1548                }
1549            }
1550        },
1551        Command::DumpHighlightKeywords(args) => {
1552            commands::dump_highlight_keywords::run(&args.output, args.check);
1553        }
1554        Command::DumpTriggerQuickref(args) => {
1555            commands::dump_trigger_quickref::run(&args.output, args.check);
1556        }
1557        Command::DumpConnectorMatrix(args) => {
1558            commands::check::connector_matrix::run_docs(&args.output, &args.sources, args.check);
1559        }
1560        Command::DumpProtocolArtifacts(args) => {
1561            commands::dump_protocol_artifacts::run(&args.output_dir, args.check);
1562        }
1563        Command::Time(args) => match args.command {
1564            TimeCommand::Run(time_args) => commands::time::run(time_args).await,
1565        },
1566    }
1567}
1568
1569fn run_profile_options(args: &cli::ProfileArgs) -> commands::run::RunProfileOptions {
1570    commands::run::RunProfileOptions {
1571        text: args.text,
1572        json_path: args.json_path.clone(),
1573    }
1574}
1575
1576fn print_completions(shell: CompletionShell) {
1577    let mut command = Cli::command();
1578    let shell = clap_complete::Shell::from(shell);
1579    clap_complete::generate(shell, &mut command, "harn", &mut std::io::stdout());
1580}
1581
1582fn normalize_serve_args(mut raw_args: Vec<String>) -> Vec<String> {
1583    if raw_args.len() > 2
1584        && raw_args.get(1).is_some_and(|arg| arg == "serve")
1585        && !matches!(
1586            raw_args.get(2).map(String::as_str),
1587            Some("acp" | "a2a" | "api" | "mcp" | "-h" | "--help")
1588        )
1589    {
1590        raw_args.insert(2, "a2a".to_string());
1591    }
1592    raw_args
1593}
1594
1595fn print_version() {
1596    println!(
1597        r"
1598 ╱▔▔╲
1599 ╱    ╲    harn v{}
1600 │ ◆  │    the agent harness language
1601 │    │
1602 ╰──╯╱
1603   ╱╱
1604",
1605        env!("CARGO_PKG_VERSION")
1606    );
1607}
1608
1609/// Schema version for `harn version --json`. Bump when the data shape
1610/// changes; new optional fields can be added freely.
1611pub(crate) const VERSION_SCHEMA_VERSION: u32 = 1;
1612
1613#[derive(serde::Serialize)]
1614struct VersionInfo {
1615    name: &'static str,
1616    version: &'static str,
1617    description: &'static str,
1618}
1619
1620fn print_version_json() {
1621    let payload = VersionInfo {
1622        name: env!("CARGO_PKG_NAME"),
1623        version: env!("CARGO_PKG_VERSION"),
1624        description: env!("CARGO_PKG_DESCRIPTION"),
1625    };
1626    let envelope = json_envelope::JsonEnvelope::ok(VERSION_SCHEMA_VERSION, payload);
1627    println!("{}", json_envelope::to_string_pretty(&envelope));
1628}
1629
1630/// Run `harn version`. Dispatches to the embedded `.harn` script by
1631/// default; set `HARN_CLI_IMPL=rust` to keep the legacy Rust handlers
1632/// (used by the parity-snapshot harness to compare both impls).
1633async fn run_version(args: cli::VersionArgs) -> i32 {
1634    if std::env::var("HARN_CLI_IMPL").as_deref() == Ok("rust") {
1635        if args.json {
1636            print_version_json();
1637        } else {
1638            print_version();
1639        }
1640        return 0;
1641    }
1642    // Build-time constants travel to the script via scoped env vars
1643    // rather than a new builtin — the script reads them with
1644    // `env_or("HARN_BUILD_VERSION", "unknown")`.
1645    let _name = env_guard::ScopedEnvVar::set("HARN_BUILD_NAME", env!("CARGO_PKG_NAME"));
1646    let _version = env_guard::ScopedEnvVar::set("HARN_BUILD_VERSION", env!("CARGO_PKG_VERSION"));
1647    let _description =
1648        env_guard::ScopedEnvVar::set("HARN_BUILD_DESCRIPTION", env!("CARGO_PKG_DESCRIPTION"));
1649    let argv = if args.json {
1650        vec!["--json".to_string()]
1651    } else {
1652        Vec::new()
1653    };
1654    dispatch::dispatch_to_embedded_script("version", argv, args.json).await
1655}
1656
1657async fn print_model_info(args: &ModelInfoArgs) -> bool {
1658    let resolved = harn_vm::llm_config::resolve_model_info(&args.model);
1659    let api_key_result = harn_vm::llm::resolve_api_key(&resolved.provider);
1660    let api_key_set = api_key_result.is_ok();
1661    let api_key = api_key_result.unwrap_or_default();
1662    let context_window =
1663        harn_vm::llm::fetch_provider_max_context(&resolved.provider, &resolved.id, &api_key).await;
1664    let readiness = local_openai_readiness(&resolved.provider, &resolved.id, &api_key).await;
1665    let catalog = harn_vm::llm_config::model_catalog_entry(&resolved.id);
1666    let runtime_context_window = catalog
1667        .as_ref()
1668        .and_then(|entry| entry.runtime_context_window);
1669    let capabilities = harn_vm::llm::capabilities::lookup(&resolved.provider, &resolved.id);
1670    let mut payload = serde_json::json!({
1671        "alias": args.model,
1672        "id": resolved.id,
1673        "provider": resolved.provider,
1674        "resolved_alias": resolved.alias,
1675        "tool_format": resolved.tool_format,
1676        "tier": resolved.tier,
1677        "api_key_set": api_key_set,
1678        "context_window": context_window,
1679        "runtime_context_window": runtime_context_window,
1680        "readiness": readiness,
1681        "catalog": catalog,
1682        "capabilities": {
1683            "native_tools": capabilities.native_tools,
1684            "defer_loading": capabilities.defer_loading,
1685            "tool_search": capabilities.tool_search,
1686            "max_tools": capabilities.max_tools,
1687            "prompt_caching": capabilities.prompt_caching,
1688            "vision": capabilities.vision,
1689            "vision_supported": capabilities.vision_supported,
1690            "audio": capabilities.audio,
1691            "pdf": capabilities.pdf,
1692            "files_api_supported": capabilities.files_api_supported,
1693            "json_schema": capabilities.json_schema,
1694            "prefers_xml_scaffolding": capabilities.prefers_xml_scaffolding,
1695            "prefers_markdown_scaffolding": capabilities.prefers_markdown_scaffolding,
1696            "structured_output_mode": capabilities.structured_output_mode,
1697            "supports_assistant_prefill": capabilities.supports_assistant_prefill,
1698            "prefers_role_developer": capabilities.prefers_role_developer,
1699            "prefers_xml_tools": capabilities.prefers_xml_tools,
1700            "thinking": !capabilities.thinking_modes.is_empty(),
1701            "thinking_block_style": capabilities.thinking_block_style,
1702            "thinking_modes": capabilities.thinking_modes,
1703            "interleaved_thinking_supported": capabilities.interleaved_thinking_supported,
1704            "anthropic_beta_features": capabilities.anthropic_beta_features,
1705            "preserve_thinking": capabilities.preserve_thinking,
1706            "server_parser": capabilities.server_parser,
1707            "honors_chat_template_kwargs": capabilities.honors_chat_template_kwargs,
1708            "recommended_endpoint": capabilities.recommended_endpoint,
1709            "text_tool_wire_format_supported": capabilities.text_tool_wire_format_supported,
1710            "preferred_tool_format": capabilities.preferred_tool_format,
1711            "tool_mode_parity": capabilities.tool_mode_parity,
1712            "tool_mode_parity_notes": capabilities.tool_mode_parity_notes,
1713        },
1714        "qc_default_model": harn_vm::llm_config::qc_default_model(&resolved.provider),
1715    });
1716
1717    let should_verify = args.verify || args.warm;
1718    let mut ok = true;
1719    if should_verify {
1720        if resolved.provider == "ollama" {
1721            let mut readiness = harn_vm::llm::OllamaReadinessOptions::new(resolved.id.clone());
1722            readiness.warm = args.warm;
1723            readiness.observe_loaded = true;
1724            readiness.keep_alive = args
1725                .keep_alive
1726                .as_deref()
1727                .and_then(harn_vm::llm::normalize_ollama_keep_alive);
1728            let result = harn_vm::llm::ollama_readiness(readiness).await;
1729            ok = result.valid;
1730            payload["readiness"] = serde_json::to_value(&result).unwrap_or_else(|error| {
1731                serde_json::json!({
1732                    "valid": false,
1733                    "status": "serialization_error",
1734                    "message": format!("failed to serialize readiness result: {error}"),
1735                })
1736            });
1737        } else {
1738            ok = false;
1739            payload["readiness"] = serde_json::json!({
1740                "valid": false,
1741                "status": "unsupported_provider",
1742                "message": format!(
1743                    "model-info --verify is only supported for Ollama models; resolved provider is '{}'",
1744                    resolved.provider
1745                ),
1746                "provider": resolved.provider,
1747            });
1748        }
1749    }
1750
1751    println!(
1752        "{}",
1753        serde_json::to_string(&payload).unwrap_or_else(|error| {
1754            command_error(&format!("failed to serialize model info: {error}"))
1755        })
1756    );
1757    ok
1758}
1759
1760async fn local_openai_readiness(
1761    provider: &str,
1762    model: &str,
1763    api_key: &str,
1764) -> Option<serde_json::Value> {
1765    let def = harn_vm::llm_config::provider_config(provider)?;
1766    if def.auth_style != "none" || !harn_vm::llm::supports_model_readiness_probe(&def) {
1767        return None;
1768    }
1769    let readiness = harn_vm::llm::probe_openai_compatible_model(provider, model, api_key).await;
1770    Some(serde_json::json!({
1771        "valid": readiness.valid,
1772        "category": readiness.category,
1773        "message": readiness.message,
1774        "provider": readiness.provider,
1775        "model": readiness.model,
1776        "url": readiness.url,
1777        "status": readiness.status,
1778        "available_models": readiness.available_models,
1779    }))
1780}
1781
1782fn build_provider_catalog_payload(available_only: bool) -> serde_json::Value {
1783    let provider_names = if available_only {
1784        harn_vm::llm_config::available_provider_names()
1785    } else {
1786        harn_vm::llm_config::provider_names()
1787    };
1788    let providers: Vec<_> = provider_names
1789        .into_iter()
1790        .filter_map(|name| {
1791            harn_vm::llm_config::provider_config(&name).map(|def| {
1792                serde_json::json!({
1793                    "name": name,
1794                    "display_name": def.display_name,
1795                    "icon": def.icon,
1796                    "base_url": harn_vm::llm_config::resolve_base_url(&def),
1797                    "base_url_env": def.base_url_env,
1798                    "auth_style": def.auth_style,
1799                    "auth_envs": harn_vm::llm_config::auth_env_names(&def.auth_env),
1800                    "auth_available": harn_vm::llm_config::provider_key_available(&name),
1801                    "features": def.features,
1802                    "cost_per_1k_in": def.cost_per_1k_in,
1803                    "cost_per_1k_out": def.cost_per_1k_out,
1804                    "latency_p50_ms": def.latency_p50_ms,
1805                })
1806            })
1807        })
1808        .collect();
1809    let models: Vec<_> = harn_vm::llm_config::model_catalog_entries()
1810        .into_iter()
1811        .map(|(id, model)| {
1812            serde_json::json!({
1813                "id": id,
1814                "name": model.name,
1815                "provider": model.provider,
1816                "context_window": model.context_window,
1817                "runtime_context_window": model.runtime_context_window,
1818                "stream_timeout": model.stream_timeout,
1819                "capabilities": model.capabilities,
1820                "pricing": model.pricing,
1821            })
1822        })
1823        .collect();
1824    let aliases: Vec<_> = harn_vm::llm_config::alias_entries()
1825        .into_iter()
1826        .map(|(name, alias)| {
1827            serde_json::json!({
1828                "name": name,
1829                "id": alias.id,
1830                "provider": alias.provider,
1831                "tool_format": alias.tool_format,
1832                "tool_calling": harn_vm::llm_config::alias_tool_calling_entry(&name),
1833            })
1834        })
1835        .collect();
1836    serde_json::json!({
1837        "providers": providers,
1838        "known_model_names": harn_vm::llm_config::known_model_names(),
1839        "available_providers": harn_vm::llm_config::available_provider_names(),
1840        "aliases": aliases,
1841        "models": models,
1842        "qc_defaults": harn_vm::llm_config::qc_defaults(),
1843    })
1844}
1845
1846fn print_provider_catalog(available_only: bool) {
1847    let payload = build_provider_catalog_payload(available_only);
1848    println!(
1849        "{}",
1850        serde_json::to_string(&payload).unwrap_or_else(|error| {
1851            command_error(&format!("failed to serialize provider catalog: {error}"))
1852        })
1853    );
1854}
1855
1856/// Dispatch shim for `harn provider-catalog`. Aggregation stays in
1857/// Rust (the script can't reach `llm_config` for the catalog walk);
1858/// the .harn renderer in `stdlib/cli/providers/catalog.harn` only
1859/// re-emits the JSON envelope.
1860///
1861/// Lock keeps concurrent in-process callers from racing on the global
1862/// env var the dispatch wedge reads — same pattern as the other
1863/// partial-port commands (see harn#2305 / #2309).
1864async fn dispatch_provider_catalog(available_only: bool) -> i32 {
1865    static DISPATCH_LOCK: tokio::sync::Mutex<()> = tokio::sync::Mutex::const_new(());
1866    let payload = build_provider_catalog_payload(available_only);
1867    let payload_json = match serde_json::to_string(&payload) {
1868        Ok(json) => json,
1869        Err(error) => {
1870            eprintln!("error: failed to serialise provider catalog payload: {error}");
1871            return 1;
1872        }
1873    };
1874    let _guard = DISPATCH_LOCK.lock().await;
1875    let _payload_guard =
1876        crate::env_guard::ScopedEnvVar::set("HARN_PROVIDER_CATALOG_PAYLOAD_JSON", &payload_json);
1877    // `--available-only` doesn't enable JSON; the catalog dump is JSON-
1878    // only on both impls, but pass `true` so the dispatch wedge sets
1879    // HARN_OUTPUT_JSON for symmetry with peer scripts.
1880    crate::dispatch::dispatch_to_embedded_script("providers/catalog", Vec::new(), true).await
1881}
1882
1883async fn run_provider_ready(
1884    provider: &str,
1885    model: Option<&str>,
1886    base_url: Option<&str>,
1887    json: bool,
1888) {
1889    let readiness =
1890        harn_vm::llm::readiness::probe_provider_readiness(provider, model, base_url).await;
1891    if json {
1892        match serde_json::to_string_pretty(&readiness) {
1893            Ok(payload) => println!("{payload}"),
1894            Err(error) => command_error(&format!("failed to serialize readiness result: {error}")),
1895        }
1896    } else if readiness.ok {
1897        println!("{}", readiness.message);
1898    } else {
1899        eprintln!("{}", readiness.message);
1900    }
1901    if !readiness.ok {
1902        process::exit(1);
1903    }
1904}
1905
1906fn command_error(message: &str) -> ! {
1907    Cli::command()
1908        .error(ErrorKind::ValueValidation, message)
1909        .exit()
1910}
1911
1912fn print_check_error(code: &str, message: &str) -> ! {
1913    let envelope: json_envelope::JsonEnvelope<commands::check::CheckReport> =
1914        json_envelope::JsonEnvelope::err(commands::check::CHECK_SCHEMA_VERSION, code, message);
1915    println!("{}", json_envelope::to_string_pretty(&envelope));
1916    process::exit(1);
1917}
1918
1919fn print_lint_error(code: &str, message: &str) -> ! {
1920    let envelope: json_envelope::JsonEnvelope<commands::check::LintReport> =
1921        json_envelope::JsonEnvelope::err(commands::check::LINT_SCHEMA_VERSION, code, message);
1922    println!("{}", json_envelope::to_string_pretty(&envelope));
1923    process::exit(1);
1924}
1925
1926fn verify_provenance_receipt(path: &str, json: bool) -> Result<(), String> {
1927    let raw =
1928        fs::read_to_string(path).map_err(|error| format!("failed to read {path}: {error}"))?;
1929    let receipt: harn_vm::ProvenanceReceipt = serde_json::from_str(&raw)
1930        .map_err(|error| format!("failed to parse provenance receipt {path}: {error}"))?;
1931    let report = harn_vm::verify_receipt(&receipt);
1932    if json {
1933        println!(
1934            "{}",
1935            serde_json::to_string_pretty(&report).map_err(|error| error.to_string())?
1936        );
1937    } else if report.verified {
1938        println!(
1939            "verified receipt={} events={} receipt_hash={} event_root_hash={}",
1940            report.receipt_id.unwrap_or_else(|| "-".to_string()),
1941            report.event_count,
1942            report.receipt_hash.unwrap_or_else(|| "-".to_string()),
1943            report.event_root_hash.unwrap_or_else(|| "-".to_string())
1944        );
1945    } else {
1946        println!(
1947            "failed receipt={} events={}",
1948            report.receipt_id.unwrap_or_else(|| "-".to_string()),
1949            report.event_count
1950        );
1951        for error in &report.errors {
1952            println!("  {error}");
1953        }
1954        return Err("provenance receipt verification failed".to_string());
1955    }
1956    Ok(())
1957}
1958
1959fn load_run_record_or_exit(path: &Path) -> harn_vm::orchestration::RunRecord {
1960    match harn_vm::orchestration::load_run_record(path) {
1961        Ok(run) => run,
1962        Err(error) => {
1963            eprintln!("Failed to load run record: {error}");
1964            process::exit(1);
1965        }
1966    }
1967}
1968
1969fn load_eval_suite_manifest_or_exit(path: &Path) -> harn_vm::orchestration::EvalSuiteManifest {
1970    harn_vm::orchestration::load_eval_suite_manifest(path).unwrap_or_else(|error| {
1971        eprintln!("Failed to load eval manifest {}: {error}", path.display());
1972        process::exit(1);
1973    })
1974}
1975
1976fn load_eval_pack_manifest_or_exit(path: &Path) -> harn_vm::orchestration::EvalPackManifest {
1977    harn_vm::orchestration::load_eval_pack_manifest(path).unwrap_or_else(|error| {
1978        eprintln!("Failed to load eval pack {}: {error}", path.display());
1979        process::exit(1);
1980    })
1981}
1982
1983fn load_persona_eval_ladder_manifest_or_exit(
1984    path: &Path,
1985) -> harn_vm::orchestration::PersonaEvalLadderManifest {
1986    harn_vm::orchestration::load_persona_eval_ladder_manifest(path).unwrap_or_else(|error| {
1987        eprintln!(
1988            "Failed to load persona eval ladder {}: {error}",
1989            path.display()
1990        );
1991        process::exit(1);
1992    })
1993}
1994
1995fn file_looks_like_eval_manifest(path: &Path) -> bool {
1996    if path.file_name().and_then(|name| name.to_str()) == Some("harn.eval.toml") {
1997        return true;
1998    }
1999    if path.extension().and_then(|ext| ext.to_str()) == Some("toml") {
2000        let Ok(content) = fs::read_to_string(path) else {
2001            return false;
2002        };
2003        return toml::from_str::<harn_vm::orchestration::EvalPackManifest>(&content)
2004            .is_ok_and(|manifest| !manifest.cases.is_empty() || !manifest.ladders.is_empty());
2005    }
2006    let Ok(content) = fs::read_to_string(path) else {
2007        return false;
2008    };
2009    let Ok(json) = serde_json::from_str::<serde_json::Value>(&content) else {
2010        return false;
2011    };
2012    json.get("_type").and_then(|value| value.as_str()) == Some("eval_suite_manifest")
2013        || json.get("cases").is_some()
2014}
2015
2016fn file_looks_like_eval_pack_manifest(path: &Path) -> bool {
2017    if path.file_name().and_then(|name| name.to_str()) == Some("harn.eval.toml") {
2018        return true;
2019    }
2020    if path.extension().and_then(|ext| ext.to_str()) == Some("toml") {
2021        return file_looks_like_eval_manifest(path);
2022    }
2023    let Ok(content) = fs::read_to_string(path) else {
2024        return false;
2025    };
2026    let Ok(json) = serde_json::from_str::<serde_json::Value>(&content) else {
2027        return false;
2028    };
2029    json.get("version").is_some()
2030        && (json.get("cases").is_some() || json.get("ladders").is_some())
2031        && json.get("_type").and_then(|value| value.as_str()) != Some("eval_suite_manifest")
2032}
2033
2034fn file_looks_like_persona_eval_ladder_manifest(path: &Path) -> bool {
2035    let Ok(content) = fs::read_to_string(path) else {
2036        return false;
2037    };
2038    if path.extension().and_then(|ext| ext.to_str()) == Some("json") {
2039        let Ok(json) = serde_json::from_str::<serde_json::Value>(&content) else {
2040            return false;
2041        };
2042        return json.get("_type").and_then(|value| value.as_str())
2043            == Some("persona_eval_ladder_manifest")
2044            || json.get("timeout_tiers").is_some()
2045            || json.get("timeout-tiers").is_some();
2046    }
2047    toml::from_str::<harn_vm::orchestration::PersonaEvalLadderManifest>(&content).is_ok_and(
2048        |manifest| {
2049            manifest
2050                .type_name
2051                .eq_ignore_ascii_case("persona_eval_ladder_manifest")
2052                || (!manifest.timeout_tiers.is_empty() && manifest.backend.path.is_some())
2053        },
2054    )
2055}
2056
2057fn collect_run_record_paths(path: &str) -> Vec<PathBuf> {
2058    let path = Path::new(path);
2059    if path.is_file() {
2060        return vec![path.to_path_buf()];
2061    }
2062    if path.is_dir() {
2063        let mut entries: Vec<PathBuf> = fs::read_dir(path)
2064            .unwrap_or_else(|error| {
2065                eprintln!("Failed to read run directory {}: {error}", path.display());
2066                process::exit(1);
2067            })
2068            .filter_map(|entry| entry.ok().map(|entry| entry.path()))
2069            .filter(|entry| entry.extension().and_then(|ext| ext.to_str()) == Some("json"))
2070            .collect();
2071        entries.sort();
2072        return entries;
2073    }
2074    eprintln!("Run path does not exist: {}", path.display());
2075    process::exit(1);
2076}
2077
2078fn print_run_diff(diff: &harn_vm::orchestration::RunDiffReport) {
2079    println!(
2080        "Diff: {} -> {} [{} -> {}]",
2081        diff.left_run_id, diff.right_run_id, diff.left_status, diff.right_status
2082    );
2083    println!("Identical: {}", diff.identical);
2084    println!("Stage diffs: {}", diff.stage_diffs.len());
2085    println!("Tool diffs: {}", diff.tool_diffs.len());
2086    println!("Observability diffs: {}", diff.observability_diffs.len());
2087    println!("Transition delta: {}", diff.transition_count_delta);
2088    println!("Artifact delta: {}", diff.artifact_count_delta);
2089    println!("Checkpoint delta: {}", diff.checkpoint_count_delta);
2090    for stage in &diff.stage_diffs {
2091        println!("- {} [{}]", stage.node_id, stage.change);
2092        for detail in &stage.details {
2093            println!("  {detail}");
2094        }
2095    }
2096    for tool in &diff.tool_diffs {
2097        println!("- tool {} [{}]", tool.tool_name, tool.args_hash);
2098        println!("  left: {:?}", tool.left_result);
2099        println!("  right: {:?}", tool.right_result);
2100    }
2101    for item in &diff.observability_diffs {
2102        println!("- {} [{}]", item.label, item.section);
2103        for detail in &item.details {
2104            println!("  {detail}");
2105        }
2106    }
2107}
2108
2109fn inspect_run_record(path: &str, compare: Option<&str>) {
2110    let run = load_run_record_or_exit(Path::new(path));
2111    println!("Run: {}", run.id);
2112    println!(
2113        "Workflow: {}",
2114        run.workflow_name
2115            .clone()
2116            .unwrap_or_else(|| run.workflow_id.clone())
2117    );
2118    println!("Status: {}", run.status);
2119    println!("Task: {}", run.task);
2120    println!("Stages: {}", run.stages.len());
2121    println!("Artifacts: {}", run.artifacts.len());
2122    println!("Transitions: {}", run.transitions.len());
2123    println!("Checkpoints: {}", run.checkpoints.len());
2124    println!("HITL questions: {}", run.hitl_questions.len());
2125    if let Some(observability) = &run.observability {
2126        println!("Planner rounds: {}", observability.planner_rounds.len());
2127        println!("Research facts: {}", observability.research_fact_count);
2128        println!("Workers: {}", observability.worker_lineage.len());
2129        println!(
2130            "Action graph: {} nodes / {} edges",
2131            observability.action_graph_nodes.len(),
2132            observability.action_graph_edges.len()
2133        );
2134        println!(
2135            "Transcript pointers: {}",
2136            observability.transcript_pointers.len()
2137        );
2138        println!("Daemon events: {}", observability.daemon_events.len());
2139    }
2140    if let Some(parent_worker_id) = run
2141        .metadata
2142        .get("parent_worker_id")
2143        .and_then(|value| value.as_str())
2144    {
2145        println!("Parent worker: {parent_worker_id}");
2146    }
2147    if let Some(parent_stage_id) = run
2148        .metadata
2149        .get("parent_stage_id")
2150        .and_then(|value| value.as_str())
2151    {
2152        println!("Parent stage: {parent_stage_id}");
2153    }
2154    if run
2155        .metadata
2156        .get("delegated")
2157        .and_then(|value| value.as_bool())
2158        .unwrap_or(false)
2159    {
2160        println!("Delegated: true");
2161    }
2162    println!(
2163        "Pending nodes: {}",
2164        if run.pending_nodes.is_empty() {
2165            "-".to_string()
2166        } else {
2167            run.pending_nodes.join(", ")
2168        }
2169    );
2170    println!(
2171        "Replay fixture: {}",
2172        if run.replay_fixture.is_some() {
2173            "embedded"
2174        } else {
2175            "derived"
2176        }
2177    );
2178    for stage in &run.stages {
2179        let worker = stage.metadata.get("worker");
2180        let worker_suffix = worker
2181            .and_then(|value| value.get("name"))
2182            .and_then(|value| value.as_str())
2183            .map(|name| format!(" worker={name}"))
2184            .unwrap_or_default();
2185        println!(
2186            "- {} [{}] status={} outcome={} branch={}{}",
2187            stage.node_id,
2188            stage.kind,
2189            stage.status,
2190            stage.outcome,
2191            stage.branch.clone().unwrap_or_else(|| "-".to_string()),
2192            worker_suffix,
2193        );
2194        if let Some(worker) = worker {
2195            if let Some(worker_id) = worker.get("id").and_then(|value| value.as_str()) {
2196                println!("  worker_id: {worker_id}");
2197            }
2198            if let Some(child_run_id) = worker.get("child_run_id").and_then(|value| value.as_str())
2199            {
2200                println!("  child_run_id: {child_run_id}");
2201            }
2202            if let Some(child_run_path) = worker
2203                .get("child_run_path")
2204                .and_then(|value| value.as_str())
2205            {
2206                println!("  child_run_path: {child_run_path}");
2207            }
2208        }
2209    }
2210    if let Some(observability) = &run.observability {
2211        for round in &observability.planner_rounds {
2212            println!(
2213                "- planner {} iterations={} llm_calls={} tools={} research_facts={}",
2214                round.node_id,
2215                round.iteration_count,
2216                round.llm_call_count,
2217                round.tool_execution_count,
2218                round.research_facts.len()
2219            );
2220        }
2221        for pointer in &observability.transcript_pointers {
2222            println!(
2223                "- transcript {} [{}] available={} {}",
2224                pointer.label,
2225                pointer.kind,
2226                pointer.available,
2227                pointer
2228                    .path
2229                    .clone()
2230                    .unwrap_or_else(|| pointer.location.clone())
2231            );
2232        }
2233        for event in &observability.daemon_events {
2234            println!(
2235                "- daemon {} [{:?}] at {}",
2236                event.name, event.kind, event.timestamp
2237            );
2238            println!("  id: {}", event.daemon_id);
2239            println!("  persist_path: {}", event.persist_path);
2240            if let Some(summary) = &event.payload_summary {
2241                println!("  payload: {summary}");
2242            }
2243        }
2244    }
2245    if let Some(compare_path) = compare {
2246        let baseline = load_run_record_or_exit(Path::new(compare_path));
2247        print_run_diff(&harn_vm::orchestration::diff_run_records(&baseline, &run));
2248    }
2249}
2250
2251fn eval_run_record(
2252    path: &str,
2253    compare: Option<&str>,
2254    structural_experiment: Option<&str>,
2255    argv: &[String],
2256    llm_mock_mode: &commands::run::CliLlmMockMode,
2257) {
2258    if let Some(experiment) = structural_experiment {
2259        let path_buf = PathBuf::from(path);
2260        if !path_buf.is_file() || path_buf.extension().and_then(|ext| ext.to_str()) != Some("harn")
2261        {
2262            eprintln!(
2263                "--structural-experiment currently requires a .harn pipeline path, got {path}"
2264            );
2265            process::exit(1);
2266        }
2267        if compare.is_some() {
2268            eprintln!("--compare cannot be combined with --structural-experiment");
2269            process::exit(1);
2270        }
2271        if matches!(llm_mock_mode, commands::run::CliLlmMockMode::Record { .. }) {
2272            eprintln!("--llm-mock-record cannot be combined with --structural-experiment");
2273            process::exit(1);
2274        }
2275        let path_buf = fs::canonicalize(&path_buf).unwrap_or_else(|error| {
2276            command_error(&format!(
2277                "failed to canonicalize structural eval pipeline {}: {error}",
2278                path_buf.display()
2279            ))
2280        });
2281        run_structural_experiment_eval(&path_buf, experiment, argv, llm_mock_mode);
2282        return;
2283    }
2284
2285    let path_buf = PathBuf::from(path);
2286    if path_buf.is_file() && file_looks_like_persona_eval_ladder_manifest(&path_buf) {
2287        if compare.is_some() {
2288            eprintln!("--compare is not supported with persona eval ladder manifests");
2289            process::exit(1);
2290        }
2291        let manifest = load_persona_eval_ladder_manifest_or_exit(&path_buf);
2292        let report =
2293            harn_vm::orchestration::run_persona_eval_ladder(&manifest).unwrap_or_else(|error| {
2294                eprintln!(
2295                    "Failed to evaluate persona eval ladder {}: {error}",
2296                    path_buf.display()
2297                );
2298                process::exit(1);
2299            });
2300        print_persona_ladder_report(&report);
2301        if !report.pass {
2302            process::exit(1);
2303        }
2304        return;
2305    }
2306
2307    if path_buf.is_file() && file_looks_like_eval_pack_manifest(&path_buf) {
2308        if compare.is_some() {
2309            eprintln!("--compare is not supported with eval pack manifests");
2310            process::exit(1);
2311        }
2312        let manifest = load_eval_pack_manifest_or_exit(&path_buf);
2313        let report = harn_vm::orchestration::evaluate_eval_pack_manifest(&manifest).unwrap_or_else(
2314            |error| {
2315                eprintln!(
2316                    "Failed to evaluate eval pack {}: {error}",
2317                    path_buf.display()
2318                );
2319                process::exit(1);
2320            },
2321        );
2322        print_eval_pack_report(&report);
2323        if !report.pass {
2324            process::exit(1);
2325        }
2326        return;
2327    }
2328
2329    if path_buf.is_file() && file_looks_like_eval_manifest(&path_buf) {
2330        if compare.is_some() {
2331            eprintln!("--compare is not supported with eval suite manifests");
2332            process::exit(1);
2333        }
2334        let manifest = load_eval_suite_manifest_or_exit(&path_buf);
2335        let suite = harn_vm::orchestration::evaluate_run_suite_manifest(&manifest).unwrap_or_else(
2336            |error| {
2337                eprintln!(
2338                    "Failed to evaluate manifest {}: {error}",
2339                    path_buf.display()
2340                );
2341                process::exit(1);
2342            },
2343        );
2344        println!(
2345            "{} {} passed, {} failed, {} total",
2346            if suite.pass { "PASS" } else { "FAIL" },
2347            suite.passed,
2348            suite.failed,
2349            suite.total
2350        );
2351        for case in &suite.cases {
2352            println!(
2353                "- {} [{}] {}",
2354                case.label.clone().unwrap_or_else(|| case.run_id.clone()),
2355                case.workflow_id,
2356                if case.pass { "PASS" } else { "FAIL" }
2357            );
2358            if let Some(path) = &case.source_path {
2359                println!("  path: {path}");
2360            }
2361            if let Some(comparison) = &case.comparison {
2362                println!("  baseline identical: {}", comparison.identical);
2363                if !comparison.identical {
2364                    println!(
2365                        "  baseline status: {} -> {}",
2366                        comparison.left_status, comparison.right_status
2367                    );
2368                }
2369            }
2370            for failure in &case.failures {
2371                println!("  {failure}");
2372            }
2373        }
2374        if !suite.pass {
2375            process::exit(1);
2376        }
2377        return;
2378    }
2379
2380    let paths = collect_run_record_paths(path);
2381    if paths.len() > 1 {
2382        let mut cases = Vec::new();
2383        for path in &paths {
2384            let run = load_run_record_or_exit(path);
2385            let fixture = run
2386                .replay_fixture
2387                .clone()
2388                .unwrap_or_else(|| harn_vm::orchestration::replay_fixture_from_run(&run));
2389            cases.push((run, fixture, Some(path.display().to_string())));
2390        }
2391        let suite = harn_vm::orchestration::evaluate_run_suite(cases);
2392        println!(
2393            "{} {} passed, {} failed, {} total",
2394            if suite.pass { "PASS" } else { "FAIL" },
2395            suite.passed,
2396            suite.failed,
2397            suite.total
2398        );
2399        for case in &suite.cases {
2400            println!(
2401                "- {} [{}] {}",
2402                case.run_id,
2403                case.workflow_id,
2404                if case.pass { "PASS" } else { "FAIL" }
2405            );
2406            if let Some(path) = &case.source_path {
2407                println!("  path: {path}");
2408            }
2409            if let Some(comparison) = &case.comparison {
2410                println!("  baseline identical: {}", comparison.identical);
2411            }
2412            for failure in &case.failures {
2413                println!("  {failure}");
2414            }
2415        }
2416        if !suite.pass {
2417            process::exit(1);
2418        }
2419        return;
2420    }
2421
2422    let run = load_run_record_or_exit(&paths[0]);
2423    let fixture = run
2424        .replay_fixture
2425        .clone()
2426        .unwrap_or_else(|| harn_vm::orchestration::replay_fixture_from_run(&run));
2427    let report = harn_vm::orchestration::evaluate_run_against_fixture(&run, &fixture);
2428    println!("{}", if report.pass { "PASS" } else { "FAIL" });
2429    println!("Stages: {}", report.stage_count);
2430    if let Some(compare_path) = compare {
2431        let baseline = load_run_record_or_exit(Path::new(compare_path));
2432        print_run_diff(&harn_vm::orchestration::diff_run_records(&baseline, &run));
2433    }
2434    if !report.failures.is_empty() {
2435        for failure in &report.failures {
2436            println!("- {failure}");
2437        }
2438    }
2439    if !report.pass {
2440        process::exit(1);
2441    }
2442}
2443
2444fn print_eval_pack_report(report: &harn_vm::orchestration::EvalPackReport) {
2445    println!(
2446        "{} {} passed, {} blocking failed, {} warning, {} informational, {} total",
2447        if report.pass { "PASS" } else { "FAIL" },
2448        report.passed,
2449        report.blocking_failed,
2450        report.warning_failed,
2451        report.informational_failed,
2452        report.total
2453    );
2454    for case in &report.cases {
2455        println!(
2456            "- {} [{}] {} ({})",
2457            case.label,
2458            case.workflow_id,
2459            if case.pass { "PASS" } else { "FAIL" },
2460            case.severity
2461        );
2462        if let Some(path) = &case.source_path {
2463            println!("  path: {path}");
2464        }
2465        if let Some(comparison) = &case.comparison {
2466            println!("  baseline identical: {}", comparison.identical);
2467            if !comparison.identical {
2468                println!(
2469                    "  baseline status: {} -> {}",
2470                    comparison.left_status, comparison.right_status
2471                );
2472            }
2473        }
2474        for failure in &case.failures {
2475            println!("  {failure}");
2476        }
2477        for warning in &case.warnings {
2478            println!("  warning: {warning}");
2479        }
2480        for item in &case.informational {
2481            println!("  info: {item}");
2482        }
2483    }
2484    for ladder in &report.ladders {
2485        println!(
2486            "- ladder {} [{}] {} ({}) first_correct={}/{}",
2487            ladder.id,
2488            ladder.persona,
2489            if ladder.pass { "PASS" } else { "FAIL" },
2490            ladder.severity,
2491            ladder.first_correct_route.as_deref().unwrap_or("<none>"),
2492            ladder.first_correct_tier.as_deref().unwrap_or("<none>")
2493        );
2494        println!("  artifacts: {}", ladder.artifact_root);
2495        for tier in &ladder.tiers {
2496            println!(
2497                "  - {} [{}] {} tools={} models={} latency={}ms cost=${:.6}",
2498                tier.timeout_tier,
2499                tier.route_id,
2500                tier.outcome,
2501                tier.tool_calls,
2502                tier.model_calls,
2503                tier.latency_ms,
2504                tier.cost_usd
2505            );
2506            for reason in &tier.degradation_reasons {
2507                println!("    {reason}");
2508            }
2509        }
2510    }
2511}
2512
2513fn print_persona_ladder_report(report: &harn_vm::orchestration::PersonaEvalLadderReport) {
2514    println!(
2515        "{} ladder {} passed, {} degraded/looped, {} total",
2516        if report.pass { "PASS" } else { "FAIL" },
2517        report.passed,
2518        report.failed,
2519        report.total
2520    );
2521    println!(
2522        "first_correct: {}/{}",
2523        report.first_correct_route.as_deref().unwrap_or("<none>"),
2524        report.first_correct_tier.as_deref().unwrap_or("<none>")
2525    );
2526    println!("artifacts: {}", report.artifact_root);
2527    for tier in &report.tiers {
2528        println!(
2529            "- {} [{}] {} tools={} models={} latency={}ms cost=${:.6}",
2530            tier.timeout_tier,
2531            tier.route_id,
2532            tier.outcome,
2533            tier.tool_calls,
2534            tier.model_calls,
2535            tier.latency_ms,
2536            tier.cost_usd
2537        );
2538        for reason in &tier.degradation_reasons {
2539            println!("  {reason}");
2540        }
2541    }
2542}
2543
2544fn run_package_evals() {
2545    let paths = package::load_package_eval_pack_paths(None).unwrap_or_else(|error| {
2546        eprintln!("{error}");
2547        process::exit(1);
2548    });
2549    let mut all_pass = true;
2550    for path in &paths {
2551        println!("Eval pack: {}", path.display());
2552        let manifest = load_eval_pack_manifest_or_exit(path);
2553        let report = harn_vm::orchestration::evaluate_eval_pack_manifest(&manifest).unwrap_or_else(
2554            |error| {
2555                eprintln!("Failed to evaluate eval pack {}: {error}", path.display());
2556                process::exit(1);
2557            },
2558        );
2559        print_eval_pack_report(&report);
2560        all_pass &= report.pass;
2561    }
2562    if !all_pass {
2563        process::exit(1);
2564    }
2565}
2566
2567fn run_structural_experiment_eval(
2568    path: &Path,
2569    experiment: &str,
2570    argv: &[String],
2571    llm_mock_mode: &commands::run::CliLlmMockMode,
2572) {
2573    let baseline_dir = tempfile::Builder::new()
2574        .prefix("harn-eval-baseline-")
2575        .tempdir()
2576        .unwrap_or_else(|error| {
2577            command_error(&format!("failed to create baseline tempdir: {error}"))
2578        });
2579    let variant_dir = tempfile::Builder::new()
2580        .prefix("harn-eval-variant-")
2581        .tempdir()
2582        .unwrap_or_else(|error| {
2583            command_error(&format!("failed to create variant tempdir: {error}"))
2584        });
2585
2586    let baseline = spawn_eval_pipeline_run(path, baseline_dir.path(), None, argv, llm_mock_mode);
2587    if !baseline.status.success() {
2588        relay_subprocess_failure("baseline", &baseline);
2589    }
2590
2591    let variant = spawn_eval_pipeline_run(
2592        path,
2593        variant_dir.path(),
2594        Some(experiment),
2595        argv,
2596        llm_mock_mode,
2597    );
2598    if !variant.status.success() {
2599        relay_subprocess_failure("variant", &variant);
2600    }
2601
2602    let baseline_runs = collect_structural_eval_runs(baseline_dir.path());
2603    let variant_runs = collect_structural_eval_runs(variant_dir.path());
2604    if baseline_runs.is_empty() || variant_runs.is_empty() {
2605        eprintln!(
2606            "structural eval expected workflow run records under {} and {}, but one side was empty",
2607            baseline_dir.path().display(),
2608            variant_dir.path().display()
2609        );
2610        process::exit(1);
2611    }
2612    if baseline_runs.len() != variant_runs.len() {
2613        eprintln!(
2614            "structural eval produced different run counts: baseline={} variant={}",
2615            baseline_runs.len(),
2616            variant_runs.len()
2617        );
2618        process::exit(1);
2619    }
2620
2621    let mut baseline_ok = 0usize;
2622    let mut variant_ok = 0usize;
2623    let mut any_failures = false;
2624
2625    println!("Structural experiment: {experiment}");
2626    println!("Cases: {}", baseline_runs.len());
2627    for (baseline_run, variant_run) in baseline_runs.iter().zip(variant_runs.iter()) {
2628        let baseline_fixture = baseline_run
2629            .replay_fixture
2630            .clone()
2631            .unwrap_or_else(|| harn_vm::orchestration::replay_fixture_from_run(baseline_run));
2632        let variant_fixture = variant_run
2633            .replay_fixture
2634            .clone()
2635            .unwrap_or_else(|| harn_vm::orchestration::replay_fixture_from_run(variant_run));
2636        let baseline_report =
2637            harn_vm::orchestration::evaluate_run_against_fixture(baseline_run, &baseline_fixture);
2638        let variant_report =
2639            harn_vm::orchestration::evaluate_run_against_fixture(variant_run, &variant_fixture);
2640        let diff = harn_vm::orchestration::diff_run_records(baseline_run, variant_run);
2641        if baseline_report.pass {
2642            baseline_ok += 1;
2643        }
2644        if variant_report.pass {
2645            variant_ok += 1;
2646        }
2647        any_failures |= !baseline_report.pass || !variant_report.pass;
2648        println!(
2649            "- {} [{}]",
2650            variant_run
2651                .workflow_name
2652                .clone()
2653                .unwrap_or_else(|| variant_run.workflow_id.clone()),
2654            variant_run.task
2655        );
2656        println!(
2657            "  baseline: {}",
2658            if baseline_report.pass { "PASS" } else { "FAIL" }
2659        );
2660        for failure in &baseline_report.failures {
2661            println!("    {failure}");
2662        }
2663        println!(
2664            "  variant: {}",
2665            if variant_report.pass { "PASS" } else { "FAIL" }
2666        );
2667        for failure in &variant_report.failures {
2668            println!("    {failure}");
2669        }
2670        println!("  diff identical: {}", diff.identical);
2671        println!("  stage diffs: {}", diff.stage_diffs.len());
2672        println!("  tool diffs: {}", diff.tool_diffs.len());
2673        println!("  observability diffs: {}", diff.observability_diffs.len());
2674    }
2675
2676    println!("Baseline {} / {} passed", baseline_ok, baseline_runs.len());
2677    println!("Variant {} / {} passed", variant_ok, variant_runs.len());
2678
2679    if any_failures {
2680        process::exit(1);
2681    }
2682}
2683
2684fn spawn_eval_pipeline_run(
2685    path: &Path,
2686    run_dir: &Path,
2687    structural_experiment: Option<&str>,
2688    argv: &[String],
2689    llm_mock_mode: &commands::run::CliLlmMockMode,
2690) -> std::process::Output {
2691    let exe = env::current_exe().unwrap_or_else(|error| {
2692        command_error(&format!("failed to resolve current executable: {error}"))
2693    });
2694    let mut command = std::process::Command::new(exe);
2695    command.current_dir(path.parent().unwrap_or_else(|| Path::new(".")));
2696    command.arg("run");
2697    match llm_mock_mode {
2698        commands::run::CliLlmMockMode::Off => {}
2699        commands::run::CliLlmMockMode::Replay { fixture_path } => {
2700            command
2701                .arg("--llm-mock")
2702                .arg(absolute_cli_path(fixture_path));
2703        }
2704        commands::run::CliLlmMockMode::Record { fixture_path } => {
2705            command
2706                .arg("--llm-mock-record")
2707                .arg(absolute_cli_path(fixture_path));
2708        }
2709    }
2710    command.arg(path);
2711    if !argv.is_empty() {
2712        command.arg("--");
2713        command.args(argv);
2714    }
2715    command.env(harn_vm::runtime_paths::HARN_RUN_DIR_ENV, run_dir);
2716    if let Some(experiment) = structural_experiment {
2717        command.env("HARN_STRUCTURAL_EXPERIMENT", experiment);
2718    }
2719    command.output().unwrap_or_else(|error| {
2720        command_error(&format!(
2721            "failed to spawn `harn run {}` for structural eval: {error}",
2722            path.display()
2723        ))
2724    })
2725}
2726
2727fn absolute_cli_path(path: &Path) -> PathBuf {
2728    if path.is_absolute() {
2729        return path.to_path_buf();
2730    }
2731    env::current_dir()
2732        .unwrap_or_else(|_| PathBuf::from("."))
2733        .join(path)
2734}
2735
2736fn relay_subprocess_failure(label: &str, output: &std::process::Output) -> ! {
2737    let stdout = String::from_utf8_lossy(&output.stdout);
2738    let stderr = String::from_utf8_lossy(&output.stderr);
2739    if !stdout.trim().is_empty() {
2740        eprintln!("[{label}] stdout:\n{stdout}");
2741    }
2742    if !stderr.trim().is_empty() {
2743        eprintln!("[{label}] stderr:\n{stderr}");
2744    }
2745    process::exit(output.status.code().unwrap_or(1));
2746}
2747
2748fn collect_structural_eval_runs(dir: &Path) -> Vec<harn_vm::orchestration::RunRecord> {
2749    let mut paths: Vec<PathBuf> = fs::read_dir(dir)
2750        .unwrap_or_else(|error| {
2751            command_error(&format!(
2752                "failed to read structural eval run dir {}: {error}",
2753                dir.display()
2754            ))
2755        })
2756        .filter_map(|entry| entry.ok().map(|entry| entry.path()))
2757        .filter(|entry| entry.extension().and_then(|ext| ext.to_str()) == Some("json"))
2758        .collect();
2759    paths.sort();
2760    let mut runs: Vec<_> = paths
2761        .iter()
2762        .map(|path| load_run_record_or_exit(path))
2763        .collect();
2764    runs.sort_by(|left, right| {
2765        (
2766            left.started_at.as_str(),
2767            left.workflow_id.as_str(),
2768            left.task.as_str(),
2769        )
2770            .cmp(&(
2771                right.started_at.as_str(),
2772                right.workflow_id.as_str(),
2773                right.task.as_str(),
2774            ))
2775    });
2776    runs
2777}
2778
2779/// Exits on error.
2780pub(crate) fn parse_source_file(path: &str) -> (String, Vec<harn_parser::SNode>) {
2781    ensure_builtin_signatures_installed();
2782
2783    let source = match fs::read_to_string(path) {
2784        Ok(s) => s,
2785        Err(e) => {
2786            eprintln!("Error reading {path}: {e}");
2787            process::exit(1);
2788        }
2789    };
2790
2791    let mut lexer = Lexer::new(&source);
2792    let tokens = match lexer.tokenize() {
2793        Ok(t) => t,
2794        Err(e) => {
2795            let diagnostic = harn_parser::diagnostic::render_diagnostic_with_code(
2796                &source,
2797                path,
2798                &error_span_from_lex(&e),
2799                "error",
2800                harn_parser::diagnostic::lexer_error_code(&e),
2801                &e.to_string(),
2802                Some("here"),
2803                None,
2804            );
2805            eprint!("{diagnostic}");
2806            process::exit(1);
2807        }
2808    };
2809
2810    let mut parser = Parser::new(tokens);
2811    let program = match parser.parse() {
2812        Ok(p) => p,
2813        Err(err) => {
2814            if parser.all_errors().is_empty() {
2815                let span = error_span_from_parse(&err);
2816                let diagnostic = harn_parser::diagnostic::render_diagnostic_with_code(
2817                    &source,
2818                    path,
2819                    &span,
2820                    "error",
2821                    harn_parser::diagnostic::parser_error_code(&err),
2822                    &harn_parser::diagnostic::parser_error_message(&err),
2823                    Some(harn_parser::diagnostic::parser_error_label(&err)),
2824                    harn_parser::diagnostic::parser_error_help(&err),
2825                );
2826                eprint!("{diagnostic}");
2827            } else {
2828                for e in parser.all_errors() {
2829                    let span = error_span_from_parse(e);
2830                    let diagnostic = harn_parser::diagnostic::render_diagnostic_with_code(
2831                        &source,
2832                        path,
2833                        &span,
2834                        "error",
2835                        harn_parser::diagnostic::parser_error_code(e),
2836                        &harn_parser::diagnostic::parser_error_message(e),
2837                        Some(harn_parser::diagnostic::parser_error_label(e)),
2838                        harn_parser::diagnostic::parser_error_help(e),
2839                    );
2840                    eprint!("{diagnostic}");
2841                }
2842            }
2843            process::exit(1);
2844        }
2845    };
2846
2847    (source, program)
2848}
2849
2850fn error_span_from_lex(e: &harn_lexer::LexerError) -> harn_lexer::Span {
2851    match e {
2852        harn_lexer::LexerError::UnexpectedCharacter(_, span)
2853        | harn_lexer::LexerError::UnterminatedString(span)
2854        | harn_lexer::LexerError::UnterminatedBlockComment(span) => *span,
2855    }
2856}
2857
2858fn error_span_from_parse(e: &harn_parser::ParserError) -> harn_lexer::Span {
2859    match e {
2860        harn_parser::ParserError::Unexpected { span, .. } => *span,
2861        harn_parser::ParserError::UnexpectedEof { span, .. } => *span,
2862    }
2863}
2864
2865/// Used by REPL and conformance tests.
2866pub(crate) async fn execute(source: &str, source_path: Option<&Path>) -> Result<String, String> {
2867    execute_with_skill_dirs(source, source_path, &[]).await
2868}
2869
2870pub(crate) async fn execute_with_skill_dirs(
2871    source: &str,
2872    source_path: Option<&Path>,
2873    cli_skill_dirs: &[PathBuf],
2874) -> Result<String, String> {
2875    execute_with_skill_dirs_and_optional_harness(source, source_path, cli_skill_dirs, None).await
2876}
2877
2878pub(crate) async fn execute_with_skill_dirs_and_harness(
2879    source: &str,
2880    source_path: Option<&Path>,
2881    cli_skill_dirs: &[PathBuf],
2882    harness: harn_vm::Harness,
2883) -> Result<String, String> {
2884    execute_with_skill_dirs_and_optional_harness(source, source_path, cli_skill_dirs, Some(harness))
2885        .await
2886}
2887
2888async fn execute_with_skill_dirs_and_optional_harness(
2889    source: &str,
2890    source_path: Option<&Path>,
2891    cli_skill_dirs: &[PathBuf],
2892    harness: Option<harn_vm::Harness>,
2893) -> Result<String, String> {
2894    let mut lexer = Lexer::new(source);
2895    let tokens = lexer.tokenize().map_err(|e| e.to_string())?;
2896    let mut parser = Parser::new(tokens);
2897    let program = parser.parse().map_err(|e| e.to_string())?;
2898
2899    // Static cross-module resolution: when executed from a file, derive the
2900    // import graph so `execute` catches undefined calls at typecheck time.
2901    // The REPL / `-e` path invokes this without `source_path`, where there
2902    // is no importing file context; we fall back to no-imports checking.
2903    let mut checker = TypeChecker::new();
2904    if let Some(path) = source_path {
2905        let graph = harn_modules::build(&[path.to_path_buf()]);
2906        if let Some(imported) = graph.imported_names_for_file(path) {
2907            checker = checker.with_imported_names(imported);
2908        }
2909        if let Some(imported) = graph.imported_type_declarations_for_file(path) {
2910            checker = checker.with_imported_type_decls(imported);
2911        }
2912        if let Some(imported) = graph.imported_callable_declarations_for_file(path) {
2913            checker = checker.with_imported_callable_decls(imported);
2914        }
2915    }
2916    let type_diagnostics = checker.check(&program);
2917    let mut warning_lines = Vec::new();
2918    for diag in &type_diagnostics {
2919        match diag.severity {
2920            DiagnosticSeverity::Error => return Err(diag.message.clone()),
2921            DiagnosticSeverity::Warning => {
2922                warning_lines.push(format!("warning: {}", diag.message));
2923            }
2924        }
2925    }
2926
2927    let chunk = harn_vm::Compiler::new()
2928        .compile(&program)
2929        .map_err(|e| e.to_string())?;
2930
2931    let local = tokio::task::LocalSet::new();
2932    local
2933        .run_until(async {
2934            let mut vm = harn_vm::Vm::new();
2935            harn_vm::register_vm_stdlib(&mut vm);
2936            install_default_hostlib(&mut vm);
2937            let source_parent = source_path
2938                .and_then(|p| p.parent())
2939                .unwrap_or(std::path::Path::new("."));
2940            let project_root = harn_vm::stdlib::process::find_project_root(source_parent);
2941            let store_base = project_root.as_deref().unwrap_or(source_parent);
2942            let execution_cwd = std::env::current_dir()
2943                .unwrap_or_else(|_| std::path::PathBuf::from("."))
2944                .to_string_lossy()
2945                .into_owned();
2946            let source_dir = source_parent.to_string_lossy().into_owned();
2947            if source_path.is_some_and(is_conformance_path) {
2948                harn_vm::event_log::install_memory_for_current_thread(64);
2949            }
2950            harn_vm::register_store_builtins(&mut vm, store_base);
2951            harn_vm::register_metadata_builtins(&mut vm, store_base);
2952            let pipeline_name = source_path
2953                .and_then(|p| p.file_stem())
2954                .and_then(|s| s.to_str())
2955                .unwrap_or("default");
2956            harn_vm::register_checkpoint_builtins(&mut vm, store_base, pipeline_name);
2957            harn_vm::stdlib::process::set_thread_execution_context(Some(
2958                harn_vm::orchestration::RunExecutionRecord {
2959                    cwd: Some(execution_cwd),
2960                    source_dir: Some(source_dir),
2961                    env: std::collections::BTreeMap::new(),
2962                    adapter: None,
2963                    repo_path: None,
2964                    worktree_path: None,
2965                    branch: None,
2966                    base_ref: None,
2967                    cleanup: None,
2968                },
2969            ));
2970            if let Some(ref root) = project_root {
2971                vm.set_project_root(root);
2972            }
2973            if let Some(path) = source_path {
2974                if let Some(parent) = path.parent() {
2975                    if !parent.as_os_str().is_empty() {
2976                        vm.set_source_dir(parent);
2977                    }
2978                }
2979            }
2980            // Conformance tests land here via `run_conformance_tests`; for
2981            // `skill_fs_*` fixtures to see the bundled `skills/` folder
2982            // we run the same layered discovery as `harn run`.
2983            let loaded = skill_loader::load_skills(&skill_loader::SkillLoaderInputs {
2984                cli_dirs: cli_skill_dirs.to_vec(),
2985                source_path: source_path.map(Path::to_path_buf),
2986            });
2987            skill_loader::emit_loader_warnings(&loaded.loader_warnings);
2988            skill_loader::install_skills_global(&mut vm, &loaded);
2989            vm.set_harness(harness.unwrap_or_else(harn_vm::Harness::real));
2990            if let Some(path) = source_path {
2991                let extensions = package::load_runtime_extensions(path);
2992                package::install_runtime_extensions(&extensions);
2993                package::install_manifest_triggers(&mut vm, &extensions)
2994                    .await
2995                    .map_err(|error| format!("failed to install manifest triggers: {error}"))?;
2996                package::install_manifest_hooks(&mut vm, &extensions)
2997                    .await
2998                    .map_err(|error| format!("failed to install manifest hooks: {error}"))?;
2999            }
3000            let _event_log = harn_vm::event_log::active_event_log()
3001                .unwrap_or_else(|| harn_vm::event_log::install_memory_for_current_thread(64));
3002            let connector_clients_installed =
3003                should_install_default_connector_clients(source, source_path);
3004            if connector_clients_installed {
3005                install_default_connector_clients(store_base)
3006                    .await
3007                    .map_err(|error| format!("failed to initialize connector clients: {error}"))?;
3008            }
3009            let execution_result = vm.execute(&chunk).await.map_err(|e| e.to_string());
3010            harn_vm::egress::reset_egress_policy_for_host();
3011            if connector_clients_installed {
3012                harn_vm::clear_active_connector_clients();
3013            }
3014            harn_vm::stdlib::process::set_thread_execution_context(None);
3015            execution_result?;
3016            let mut output = String::new();
3017            for wl in &warning_lines {
3018                output.push_str(wl);
3019                output.push('\n');
3020            }
3021            output.push_str(vm.output());
3022            Ok(output)
3023        })
3024        .await
3025}
3026
3027fn should_install_default_connector_clients(source: &str, source_path: Option<&Path>) -> bool {
3028    if !source_path.is_some_and(is_conformance_path) {
3029        return true;
3030    }
3031    source.contains("connector_call")
3032        || source.contains("std/connectors")
3033        || source.contains("connectors/")
3034}
3035
3036fn is_conformance_path(path: &Path) -> bool {
3037    path.components()
3038        .any(|component| component.as_os_str() == "conformance")
3039}
3040
3041async fn install_default_connector_clients(base_dir: &Path) -> Result<(), String> {
3042    let event_log = harn_vm::event_log::active_event_log()
3043        .unwrap_or_else(|| harn_vm::event_log::install_memory_for_current_thread(64));
3044    let secret_namespace = connector_secret_namespace(base_dir);
3045    let secrets: Arc<dyn harn_vm::secrets::SecretProvider> = Arc::new(
3046        harn_vm::secrets::configured_default_chain(secret_namespace)
3047            .map_err(|error| format!("failed to configure secret providers: {error}"))?,
3048    );
3049
3050    let registry = harn_vm::ConnectorRegistry::default();
3051    let metrics = Arc::new(harn_vm::MetricsRegistry::default());
3052    let inbox = Arc::new(
3053        harn_vm::InboxIndex::new(event_log.clone(), metrics.clone())
3054            .await
3055            .map_err(|error| error.to_string())?,
3056    );
3057    registry
3058        .init_all(harn_vm::ConnectorCtx {
3059            event_log,
3060            secrets,
3061            inbox,
3062            metrics,
3063            rate_limiter: Arc::new(harn_vm::RateLimiterFactory::default()),
3064        })
3065        .await
3066        .map_err(|error| error.to_string())?;
3067    let clients = registry.client_map().await;
3068    harn_vm::install_active_connector_clients(clients);
3069    Ok(())
3070}
3071
3072fn connector_secret_namespace(base_dir: &Path) -> String {
3073    match std::env::var("HARN_SECRET_NAMESPACE") {
3074        Ok(namespace) if !namespace.trim().is_empty() => namespace,
3075        _ => {
3076            let leaf = base_dir
3077                .file_name()
3078                .and_then(|name| name.to_str())
3079                .filter(|name| !name.is_empty())
3080                .unwrap_or("workspace");
3081            format!("harn/{leaf}")
3082        }
3083    }
3084}
3085
3086#[cfg(test)]
3087mod main_tests {
3088    use super::{
3089        is_broken_pipe_panic_payload, normalize_serve_args,
3090        should_install_default_connector_clients,
3091    };
3092    use std::path::Path;
3093
3094    #[test]
3095    fn normalize_serve_args_inserts_a2a_for_legacy_shape() {
3096        let args = normalize_serve_args(vec![
3097            "harn".to_string(),
3098            "serve".to_string(),
3099            "--port".to_string(),
3100            "3000".to_string(),
3101            "agent.harn".to_string(),
3102        ]);
3103        assert_eq!(
3104            args,
3105            vec![
3106                "harn".to_string(),
3107                "serve".to_string(),
3108                "a2a".to_string(),
3109                "--port".to_string(),
3110                "3000".to_string(),
3111                "agent.harn".to_string(),
3112            ]
3113        );
3114    }
3115
3116    #[test]
3117    fn normalize_serve_args_preserves_explicit_subcommands() {
3118        let args = normalize_serve_args(vec![
3119            "harn".to_string(),
3120            "serve".to_string(),
3121            "acp".to_string(),
3122            "server.harn".to_string(),
3123        ]);
3124        assert_eq!(
3125            args,
3126            vec![
3127                "harn".to_string(),
3128                "serve".to_string(),
3129                "acp".to_string(),
3130                "server.harn".to_string(),
3131            ]
3132        );
3133    }
3134
3135    #[test]
3136    fn conformance_skips_connector_clients_unless_fixture_uses_connectors() {
3137        let path = Path::new("conformance/tests/language/basic.harn");
3138        assert!(!should_install_default_connector_clients(
3139            "__io_println(1)",
3140            Some(path)
3141        ));
3142        assert!(!should_install_default_connector_clients(
3143            "trust_graph_verify_chain()",
3144            Some(path)
3145        ));
3146        assert!(should_install_default_connector_clients(
3147            "import { post_message } from \"std/connectors/slack\"",
3148            Some(path)
3149        ));
3150        assert!(should_install_default_connector_clients(
3151            "__io_println(1)",
3152            Some(Path::new("examples/demo.harn"))
3153        ));
3154    }
3155
3156    #[test]
3157    fn broken_pipe_print_panic_is_classified_as_clean_consumer_close() {
3158        let payload = String::from("failed printing to stdout: Broken pipe (os error 32)");
3159        assert!(is_broken_pipe_panic_payload(&payload));
3160    }
3161
3162    #[test]
3163    fn unrelated_panic_is_not_classified_as_broken_pipe() {
3164        let payload = String::from("assertion failed: expected true");
3165        assert!(!is_broken_pipe_panic_payload(&payload));
3166    }
3167}