Skip to main content

harn_cli/
lib.rs

1#![recursion_limit = "256"]
2
3pub mod acp;
4pub mod cli;
5pub mod commands;
6pub mod config;
7pub mod env_guard;
8pub mod format;
9pub mod package;
10mod provider_bootstrap;
11pub mod skill_loader;
12pub mod skill_provenance;
13pub mod test_runner;
14#[doc(hidden)]
15pub mod tests;
16
17use clap::{error::ErrorKind, CommandFactory, Parser as ClapParser};
18use std::path::{Path, PathBuf};
19use std::sync::Arc;
20use std::{env, fs, process, thread};
21
22use cli::{
23    Cli, Command, CompletionShell, MergeCaptainCommand, MergeCaptainMockCommand, ModelInfoArgs,
24    PackageArtifactsCommand, PackageCacheCommand, PackageCommand, PersonaCommand,
25    PersonaSupervisionCommand, ProvidersCommand, RunsCommand, ServeCommand, SkillCommand,
26    SkillKeyCommand, SkillTrustCommand, SkillsCommand, ToolCommand,
27};
28use harn_lexer::Lexer;
29use harn_parser::{DiagnosticSeverity, Parser, TypeChecker};
30
31pub const CLI_RUNTIME_STACK_SIZE: usize = 16 * 1024 * 1024;
32
33#[cfg(feature = "hostlib")]
34pub(crate) fn install_default_hostlib(vm: &mut harn_vm::Vm) {
35    let _ = harn_hostlib::install_default(vm);
36}
37
38#[cfg(not(feature = "hostlib"))]
39pub(crate) fn install_default_hostlib(_vm: &mut harn_vm::Vm) {}
40
41/// Entry point used by `src/main.rs`. Hosts the CLI runtime thread and
42/// drives the async dispatcher in `async_main`.
43pub fn run() {
44    let handle = thread::Builder::new()
45        .name("harn-cli".to_string())
46        .stack_size(CLI_RUNTIME_STACK_SIZE)
47        .spawn(|| {
48            let runtime = tokio::runtime::Builder::new_multi_thread()
49                .enable_all()
50                .build()
51                .unwrap_or_else(|error| {
52                    eprintln!("failed to start async runtime: {error}");
53                    process::exit(1);
54                });
55            runtime.block_on(async_main());
56        })
57        .unwrap_or_else(|error| {
58            eprintln!("failed to start CLI runtime thread: {error}");
59            process::exit(1);
60        });
61
62    if let Err(payload) = handle.join() {
63        std::panic::resume_unwind(payload);
64    }
65}
66
67async fn async_main() {
68    let raw_args = normalize_serve_args(env::args().collect());
69    if raw_args.len() == 2 && raw_args[1].ends_with(".harn") {
70        provider_bootstrap::maybe_seed_ollama_for_run_file(Path::new(&raw_args[1]), false, false)
71            .await;
72        commands::run::run_file(
73            &raw_args[1],
74            false,
75            std::collections::HashSet::new(),
76            Vec::new(),
77            commands::run::CliLlmMockMode::Off,
78            None,
79            commands::run::RunProfileOptions::default(),
80        )
81        .await;
82        return;
83    }
84
85    let cli = match Cli::try_parse_from(&raw_args) {
86        Ok(cli) => cli,
87        Err(error) => {
88            if matches!(
89                error.kind(),
90                ErrorKind::DisplayHelp | ErrorKind::DisplayVersion
91            ) {
92                error.exit();
93            }
94            error.exit();
95        }
96    };
97
98    match cli.command.expect("clap requires a command") {
99        Command::Version => print_version(),
100        Command::Skill(args) => match args.command {
101            SkillCommand::Key(key_args) => match key_args.command {
102                SkillKeyCommand::Generate(generate) => commands::skill::run_key_generate(&generate),
103            },
104            SkillCommand::Sign(sign) => commands::skill::run_sign(&sign),
105            SkillCommand::Endorse(endorse) => commands::skill::run_endorse(&endorse),
106            SkillCommand::Verify(verify) => commands::skill::run_verify(&verify),
107            SkillCommand::WhoSigned(who_signed) => {
108                commands::skill::run_who_signed(&who_signed).await
109            }
110            SkillCommand::Trust(trust_args) => match trust_args.command {
111                SkillTrustCommand::Add(add) => commands::skill::run_trust_add(&add),
112                SkillTrustCommand::List(list) => commands::skill::run_trust_list(&list),
113            },
114            SkillCommand::New(new_args) => commands::skills::run_new(&new_args),
115        },
116        Command::Run(args) => {
117            if !args.explain_cost {
118                match (args.eval.as_deref(), args.file.as_deref()) {
119                    (Some(code), None) => {
120                        provider_bootstrap::maybe_seed_ollama_for_inline(
121                            code,
122                            args.yes,
123                            args.llm_mock.is_some(),
124                        )
125                        .await;
126                    }
127                    (None, Some(file)) => {
128                        provider_bootstrap::maybe_seed_ollama_for_run_file(
129                            Path::new(file),
130                            args.yes,
131                            args.llm_mock.is_some(),
132                        )
133                        .await;
134                    }
135                    _ => {}
136                }
137            }
138            let denied =
139                commands::run::build_denied_builtins(args.deny.as_deref(), args.allow.as_deref());
140            let llm_mock_mode = if let Some(path) = args.llm_mock.as_ref() {
141                commands::run::CliLlmMockMode::Replay {
142                    fixture_path: PathBuf::from(path),
143                }
144            } else if let Some(path) = args.llm_mock_record.as_ref() {
145                commands::run::CliLlmMockMode::Record {
146                    fixture_path: PathBuf::from(path),
147                }
148            } else {
149                commands::run::CliLlmMockMode::Off
150            };
151            let attestation = args.attest.then(|| commands::run::RunAttestationOptions {
152                receipt_out: args.receipt_out.as_ref().map(PathBuf::from),
153                agent_id: args.attest_agent.clone(),
154            });
155            let profile_options = run_profile_options(&args.profile);
156
157            match (args.eval.as_deref(), args.file.as_deref()) {
158                (Some(code), None) => {
159                    let (wrapped, tmp) = commands::run::prepare_eval_temp_file(code)
160                        .unwrap_or_else(|e| command_error(&e));
161                    let tmp_path: PathBuf = tmp.path().to_path_buf();
162                    fs::write(&tmp_path, &wrapped).unwrap_or_else(|e| {
163                        command_error(&format!("failed to write temp file for -e: {e}"))
164                    });
165                    let tmp_str = tmp_path.to_string_lossy().into_owned();
166                    if args.explain_cost {
167                        commands::run::run_explain_cost_file_with_skill_dirs(&tmp_str);
168                    } else {
169                        commands::run::run_file_with_skill_dirs(
170                            &tmp_str,
171                            args.trace,
172                            denied,
173                            args.argv.clone(),
174                            args.skill_dir.clone(),
175                            llm_mock_mode.clone(),
176                            attestation.clone(),
177                            profile_options.clone(),
178                        )
179                        .await;
180                    }
181                    drop(tmp);
182                }
183                (None, Some(file)) => {
184                    if args.explain_cost {
185                        commands::run::run_explain_cost_file_with_skill_dirs(file);
186                    } else {
187                        commands::run::run_file_with_skill_dirs(
188                            file,
189                            args.trace,
190                            denied,
191                            args.argv.clone(),
192                            args.skill_dir.clone(),
193                            llm_mock_mode,
194                            attestation,
195                            profile_options,
196                        )
197                        .await
198                    }
199                }
200                (Some(_), Some(_)) => command_error(
201                    "`harn run` accepts either `-e <code>` or `<file.harn>`, not both",
202                ),
203                (None, None) => {
204                    command_error("`harn run` requires either `-e <code>` or `<file.harn>`")
205                }
206            }
207        }
208        Command::Check(args) => {
209            if args.provider_matrix {
210                let cwd = std::env::current_dir().unwrap_or_else(|_| PathBuf::from("."));
211                let extensions = package::load_runtime_extensions(&cwd);
212                package::install_runtime_extensions(&extensions);
213                commands::check::provider_matrix::run(args.format, args.filter.as_deref());
214                return;
215            }
216            if args.connector_matrix {
217                commands::check::connector_matrix::run(
218                    args.format,
219                    args.filter.as_deref(),
220                    &args.targets,
221                );
222                return;
223            }
224            let mut target_strings: Vec<String> = args.targets.clone();
225            if args.workspace {
226                let anchor = target_strings.first().map(Path::new);
227                match package::load_workspace_config(anchor) {
228                    Some((workspace, manifest_dir)) if !workspace.pipelines.is_empty() => {
229                        for pipeline in &workspace.pipelines {
230                            let candidate = Path::new(pipeline);
231                            let resolved = if candidate.is_absolute() {
232                                candidate.to_path_buf()
233                            } else {
234                                manifest_dir.join(candidate)
235                            };
236                            target_strings.push(resolved.to_string_lossy().into_owned());
237                        }
238                    }
239                    Some(_) => command_error(
240                        "--workspace requires `[workspace].pipelines` in the nearest harn.toml",
241                    ),
242                    None => command_error(
243                        "--workspace could not find a harn.toml walking up from the target(s)",
244                    ),
245                }
246            }
247            if target_strings.is_empty() {
248                command_error(
249                    "`harn check` requires at least one target path, or `--workspace` with `[workspace].pipelines`",
250                );
251            }
252            for target in &target_strings {
253                if let Err(error) = package::validate_runtime_manifest_extensions(Path::new(target))
254                {
255                    command_error(&format!("manifest extension validation failed: {error}"));
256                }
257            }
258            let targets: Vec<&str> = target_strings.iter().map(String::as_str).collect();
259            let files = commands::check::collect_harn_targets(&targets);
260            if files.is_empty() {
261                command_error("no .harn files found under the given target(s)");
262            }
263            let module_graph = commands::check::build_module_graph(&files);
264            let cross_file_imports = commands::check::collect_cross_file_imports(&module_graph);
265            let mut should_fail = false;
266            for file in &files {
267                let mut config = package::load_check_config(Some(file));
268                if let Some(path) = args.host_capabilities.as_ref() {
269                    config.host_capabilities_path = Some(path.clone());
270                }
271                if let Some(path) = args.bundle_root.as_ref() {
272                    config.bundle_root = Some(path.clone());
273                }
274                if args.strict_types {
275                    config.strict_types = true;
276                }
277                if let Some(sev) = args.preflight.as_deref() {
278                    config.preflight_severity = Some(sev.to_string());
279                }
280                let outcome = commands::check::check_file_inner(
281                    file,
282                    &config,
283                    &cross_file_imports,
284                    &module_graph,
285                    args.invariants,
286                );
287                should_fail |= outcome.should_fail(config.strict);
288            }
289            if should_fail {
290                process::exit(1);
291            }
292        }
293        Command::Config(args) => {
294            if let Err(error) = commands::config_cmd::run(args).await {
295                command_error(&error);
296            }
297        }
298        Command::Explain(args) => {
299            let code = commands::explain::run_explain(&args);
300            if code != 0 {
301                process::exit(code);
302            }
303        }
304        Command::Contracts(args) => {
305            commands::contracts::handle_contracts_command(args).await;
306        }
307        Command::Connect(args) => {
308            commands::connect::run_connect(*args).await;
309        }
310        Command::Lint(args) => {
311            let targets: Vec<&str> = args.targets.iter().map(String::as_str).collect();
312            let files = commands::check::collect_harn_targets(&targets);
313            if files.is_empty() {
314                command_error("no .harn files found under the given target(s)");
315            }
316            let module_graph = commands::check::build_module_graph(&files);
317            let cross_file_imports = commands::check::collect_cross_file_imports(&module_graph);
318            if args.fix {
319                for file in &files {
320                    let mut config = package::load_check_config(Some(file));
321                    commands::check::apply_harn_lint_config(file, &mut config);
322                    let require_header = args.require_file_header
323                        || commands::check::harn_lint_require_file_header(file);
324                    let complexity_threshold =
325                        commands::check::harn_lint_complexity_threshold(file);
326                    let persona_step_allowlist =
327                        commands::check::harn_lint_persona_step_allowlist(file);
328                    commands::check::lint_fix_file(
329                        file,
330                        &config,
331                        &cross_file_imports,
332                        &module_graph,
333                        require_header,
334                        complexity_threshold,
335                        &persona_step_allowlist,
336                    );
337                }
338            } else {
339                let mut should_fail = false;
340                for file in &files {
341                    let mut config = package::load_check_config(Some(file));
342                    commands::check::apply_harn_lint_config(file, &mut config);
343                    let require_header = args.require_file_header
344                        || commands::check::harn_lint_require_file_header(file);
345                    let complexity_threshold =
346                        commands::check::harn_lint_complexity_threshold(file);
347                    let persona_step_allowlist =
348                        commands::check::harn_lint_persona_step_allowlist(file);
349                    let outcome = commands::check::lint_file_inner(
350                        file,
351                        &config,
352                        &cross_file_imports,
353                        &module_graph,
354                        require_header,
355                        complexity_threshold,
356                        &persona_step_allowlist,
357                    );
358                    should_fail |= outcome.should_fail(config.strict);
359                }
360                if should_fail {
361                    process::exit(1);
362                }
363            }
364        }
365        Command::Fmt(args) => {
366            let targets: Vec<&str> = args.targets.iter().map(String::as_str).collect();
367            // Anchor config resolution on the first target; CLI flags
368            // always win over harn.toml values.
369            let anchor = targets.first().map(Path::new).unwrap_or(Path::new("."));
370            let loaded = match config::load_for_path(anchor) {
371                Ok(c) => c,
372                Err(e) => {
373                    eprintln!("warning: {e}");
374                    config::HarnConfig::default()
375                }
376            };
377            let mut opts = harn_fmt::FmtOptions::default();
378            if let Some(w) = loaded.fmt.line_width {
379                opts.line_width = w;
380            }
381            if let Some(w) = loaded.fmt.separator_width {
382                opts.separator_width = w;
383            }
384            if let Some(w) = args.line_width {
385                opts.line_width = w;
386            }
387            if let Some(w) = args.separator_width {
388                opts.separator_width = w;
389            }
390            commands::check::fmt_targets(
391                &targets,
392                commands::check::FmtMode::from_check_flag(args.check),
393                &opts,
394            );
395        }
396        Command::Test(args) => {
397            if args.target.as_deref() == Some("agents-conformance") {
398                if args.selection.is_some() {
399                    command_error(
400                        "`harn test agents-conformance` does not accept a second positional target; use --category instead",
401                    );
402                }
403                if args.evals || args.determinism || args.record || args.replay || args.watch {
404                    command_error(
405                        "`harn test agents-conformance` cannot be combined with --evals, --determinism, --record, --replay, or --watch",
406                    );
407                }
408                let Some(target_url) = args.agents_target.clone() else {
409                    command_error("`harn test agents-conformance` requires --target <url>");
410                };
411                commands::agents_conformance::run_agents_conformance(
412                    commands::agents_conformance::AgentsConformanceConfig {
413                        target_url,
414                        api_key: args.agents_api_key.clone(),
415                        categories: args.agents_category.clone(),
416                        timeout_ms: args.timeout,
417                        verbose: args.verbose,
418                        json: args.json,
419                        json_out: args.json_out.clone(),
420                        workspace_id: args.agents_workspace_id.clone(),
421                        session_id: args.agents_session_id.clone(),
422                    },
423                )
424                .await;
425                return;
426            }
427            if args.target.as_deref() == Some("protocols") {
428                if args.evals || args.determinism || args.record || args.replay || args.watch {
429                    command_error(
430                        "`harn test protocols` cannot be combined with --evals, --determinism, --record, --replay, or --watch",
431                    );
432                }
433                if args.junit.is_some()
434                    || args.agents_target.is_some()
435                    || args.agents_api_key.is_some()
436                    || !args.agents_category.is_empty()
437                    || args.json
438                    || args.json_out.is_some()
439                    || args.agents_workspace_id.is_some()
440                    || args.agents_session_id.is_some()
441                    || args.parallel
442                    || !args.skill_dir.is_empty()
443                {
444                    command_error(
445                        "`harn test protocols` accepts only --filter, --verbose, --timing, and an optional fixture selection",
446                    );
447                }
448                commands::protocol_conformance::run_protocol_conformance(
449                    args.selection.as_deref(),
450                    args.filter.as_deref(),
451                    args.verbose || args.timing,
452                );
453                return;
454            }
455            if args.evals {
456                if args.determinism || args.record || args.replay || args.watch {
457                    command_error("--evals cannot be combined with --determinism, --record, --replay, or --watch");
458                }
459                if args.target.as_deref() != Some("package") || args.selection.is_some() {
460                    command_error("package evals are run with `harn test package --evals`");
461                }
462                run_package_evals();
463            } else if args.determinism {
464                if args.watch {
465                    command_error("--determinism cannot be combined with --watch");
466                }
467                if args.record || args.replay {
468                    command_error("--determinism manages its own record/replay cycle");
469                }
470                if let Some(t) = args.target.as_deref() {
471                    if t == "conformance" {
472                        commands::test::run_conformance_determinism_tests(
473                            t,
474                            args.selection.as_deref(),
475                            args.filter.as_deref(),
476                            args.timeout,
477                        )
478                        .await;
479                    } else if args.selection.is_some() {
480                        command_error(
481                            "only `harn test conformance` accepts a second positional target",
482                        );
483                    } else {
484                        commands::test::run_determinism_tests(
485                            t,
486                            args.filter.as_deref(),
487                            args.timeout,
488                        )
489                        .await;
490                    }
491                } else {
492                    let test_dir = if PathBuf::from("tests").is_dir() {
493                        "tests".to_string()
494                    } else {
495                        command_error("no path specified and no tests/ directory found");
496                    };
497                    if args.selection.is_some() {
498                        command_error(
499                            "only `harn test conformance` accepts a second positional target",
500                        );
501                    }
502                    commands::test::run_determinism_tests(
503                        &test_dir,
504                        args.filter.as_deref(),
505                        args.timeout,
506                    )
507                    .await;
508                }
509            } else {
510                if args.record {
511                    harn_vm::llm::set_replay_mode(
512                        harn_vm::llm::LlmReplayMode::Record,
513                        ".harn-fixtures",
514                    );
515                } else if args.replay {
516                    harn_vm::llm::set_replay_mode(
517                        harn_vm::llm::LlmReplayMode::Replay,
518                        ".harn-fixtures",
519                    );
520                }
521
522                if let Some(t) = args.target.as_deref() {
523                    if t == "conformance" {
524                        commands::test::run_conformance_tests(
525                            t,
526                            args.selection.as_deref(),
527                            args.filter.as_deref(),
528                            args.junit.as_deref(),
529                            args.timeout,
530                            args.verbose,
531                            args.timing,
532                            args.differential_optimizations,
533                        )
534                        .await;
535                    } else if args.selection.is_some() {
536                        command_error(
537                            "only `harn test conformance` accepts a second positional target",
538                        );
539                    } else if args.watch {
540                        commands::test::run_watch_tests(
541                            t,
542                            args.filter.as_deref(),
543                            args.timeout,
544                            args.parallel,
545                        )
546                        .await;
547                    } else {
548                        commands::test::run_user_tests(
549                            t,
550                            args.filter.as_deref(),
551                            args.timeout,
552                            args.parallel,
553                        )
554                        .await;
555                    }
556                } else {
557                    let test_dir = if PathBuf::from("tests").is_dir() {
558                        "tests".to_string()
559                    } else {
560                        command_error("no path specified and no tests/ directory found");
561                    };
562                    if args.selection.is_some() {
563                        command_error(
564                            "only `harn test conformance` accepts a second positional target",
565                        );
566                    }
567                    if args.watch {
568                        commands::test::run_watch_tests(
569                            &test_dir,
570                            args.filter.as_deref(),
571                            args.timeout,
572                            args.parallel,
573                        )
574                        .await;
575                    } else {
576                        commands::test::run_user_tests(
577                            &test_dir,
578                            args.filter.as_deref(),
579                            args.timeout,
580                            args.parallel,
581                        )
582                        .await;
583                    }
584                }
585            }
586        }
587        Command::Init(args) => commands::init::init_project(args.name.as_deref(), args.template),
588        Command::New(args) => match commands::init::resolve_new_args(&args) {
589            Ok((name, template)) => commands::init::init_project(name.as_deref(), template),
590            Err(error) => {
591                eprintln!("error: {error}");
592                process::exit(1);
593            }
594        },
595        Command::Doctor(args) => {
596            commands::doctor::run_doctor_with_options(commands::doctor::DoctorOptions {
597                network: !args.no_network,
598                json: args.json,
599            })
600            .await
601        }
602        Command::Models(args) => commands::models::run(args).await,
603        Command::Providers(args) => match args.command {
604            ProvidersCommand::Refresh(refresh) => {
605                if let Err(error) = commands::providers::run_refresh(&refresh).await {
606                    command_error(&error);
607                }
608            }
609            ProvidersCommand::Validate(validate) => {
610                if let Err(error) = commands::providers::run_validate(&validate) {
611                    command_error(&error);
612                }
613            }
614            ProvidersCommand::Export(export) => {
615                if let Err(error) = commands::providers::run_export(&export) {
616                    command_error(&error);
617                }
618            }
619        },
620        Command::Try(args) => commands::try_cmd::run(args).await,
621        Command::Quickstart(args) => {
622            if let Err(error) = commands::quickstart::run_quickstart(&args).await {
623                command_error(&error);
624            }
625        }
626        Command::Serve(args) => match args.command {
627            ServeCommand::Acp(args) => {
628                if let Err(error) = commands::serve::run_acp_server(&args).await {
629                    command_error(&error);
630                }
631            }
632            ServeCommand::A2a(args) => {
633                if let Err(error) = commands::serve::run_a2a_server(&args).await {
634                    command_error(&error);
635                }
636            }
637            ServeCommand::Api(args) => {
638                if let Err(error) = commands::serve::run_api_server(&args).await {
639                    command_error(&error);
640                }
641            }
642            ServeCommand::Mcp(args) => {
643                if let Err(error) = commands::serve::run_mcp_server(&args).await {
644                    command_error(&error);
645                }
646            }
647        },
648        Command::Connector(args) => {
649            if let Err(error) = commands::connector::handle_connector_command(args).await {
650                eprintln!("error: {error}");
651                process::exit(1);
652            }
653        }
654        Command::Mcp(args) => commands::mcp::handle_mcp_command(&args.command).await,
655        Command::Watch(args) => {
656            let denied =
657                commands::run::build_denied_builtins(args.deny.as_deref(), args.allow.as_deref());
658            commands::run::run_watch(&args.file, denied).await;
659        }
660        Command::Portal(args) => {
661            commands::portal::run_portal(
662                &args.dir,
663                args.manifest,
664                args.persona_state_dir,
665                &args.host,
666                args.port,
667                args.open,
668            )
669            .await
670        }
671        Command::Trigger(args) => {
672            if let Err(error) = commands::trigger::handle(args).await {
673                eprintln!("error: {error}");
674                process::exit(1);
675            }
676        }
677        Command::Flow(args) => match commands::flow::run_flow(&args) {
678            Ok(code) => {
679                if code != 0 {
680                    process::exit(code);
681                }
682            }
683            Err(error) => command_error(&error),
684        },
685        Command::Workflow(args) => match commands::workflow::handle(args) {
686            Ok(code) => {
687                if code != 0 {
688                    process::exit(code);
689                }
690            }
691            Err(error) => command_error(&error),
692        },
693        Command::Supervisor(args) => {
694            if let Err(error) = commands::supervisor::handle(args).await {
695                eprintln!("error: {error}");
696                process::exit(1);
697            }
698        }
699        Command::Trace(args) => {
700            if let Err(error) = commands::trace::handle(args).await {
701                eprintln!("error: {error}");
702                process::exit(1);
703            }
704        }
705        Command::Crystallize(args) => {
706            if let Err(error) = commands::crystallize::run(args) {
707                eprintln!("error: {error}");
708                process::exit(1);
709            }
710        }
711        Command::Trust(args) | Command::TrustGraph(args) => {
712            if let Err(error) = commands::trust::handle(args).await {
713                eprintln!("error: {error}");
714                process::exit(1);
715            }
716        }
717        Command::Verify(args) => {
718            if let Err(error) = verify_provenance_receipt(&args.receipt, args.json) {
719                eprintln!("error: {error}");
720                process::exit(1);
721            }
722        }
723        Command::Completions(args) => print_completions(args.shell),
724        Command::Orchestrator(args) => {
725            if let Err(error) = commands::orchestrator::handle(args).await {
726                eprintln!("error: {error}");
727                process::exit(1);
728            }
729        }
730        Command::Playground(args) => {
731            provider_bootstrap::maybe_seed_ollama_for_playground(
732                Path::new(&args.host),
733                Path::new(&args.script),
734                args.yes,
735                args.llm.is_some(),
736                args.llm_mock.is_some(),
737            )
738            .await;
739            let llm_mock_mode = if let Some(path) = args.llm_mock.as_ref() {
740                commands::run::CliLlmMockMode::Replay {
741                    fixture_path: PathBuf::from(path),
742                }
743            } else if let Some(path) = args.llm_mock_record.as_ref() {
744                commands::run::CliLlmMockMode::Record {
745                    fixture_path: PathBuf::from(path),
746                }
747            } else {
748                commands::run::CliLlmMockMode::Off
749            };
750            if let Err(error) = commands::playground::run_command(args, llm_mock_mode).await {
751                eprint!("{error}");
752                process::exit(1);
753            }
754        }
755        Command::Runs(args) => match args.command {
756            RunsCommand::Inspect(inspect) => {
757                inspect_run_record(&inspect.path, inspect.compare.as_deref())
758            }
759        },
760        Command::Session(args) => commands::session::run(args),
761        Command::Replay(args) => replay_run_record(&args.path),
762        Command::Eval(args) => {
763            let llm_mock_mode = if let Some(path) = args.llm_mock.as_ref() {
764                commands::run::CliLlmMockMode::Replay {
765                    fixture_path: PathBuf::from(path),
766                }
767            } else if let Some(path) = args.llm_mock_record.as_ref() {
768                commands::run::CliLlmMockMode::Record {
769                    fixture_path: PathBuf::from(path),
770                }
771            } else {
772                commands::run::CliLlmMockMode::Off
773            };
774            eval_run_record(
775                &args.path,
776                args.compare.as_deref(),
777                args.structural_experiment.as_deref(),
778                &args.argv,
779                &llm_mock_mode,
780            )
781        }
782        Command::Repl => commands::repl::run_repl().await,
783        Command::Bench(args) => commands::bench::run(args).await,
784        Command::TestBench(args) => commands::test_bench::run(args.command).await,
785        Command::Viz(args) => commands::viz::run_viz(&args.file, args.output.as_deref()),
786        Command::Install(args) => package::install_packages(
787            args.frozen || args.locked || args.offline,
788            args.refetch.as_deref(),
789            args.offline,
790            args.json,
791        ),
792        Command::Add(args) => package::add_package_with_registry(
793            &args.name_or_spec,
794            args.alias.as_deref(),
795            args.git.as_deref(),
796            args.tag.as_deref(),
797            args.rev.as_deref(),
798            args.branch.as_deref(),
799            args.path.as_deref(),
800            args.registry.as_deref(),
801        ),
802        Command::Update(args) => {
803            package::update_packages(args.alias.as_deref(), args.all, args.json)
804        }
805        Command::Remove(args) => package::remove_package(&args.alias),
806        Command::Lock => package::lock_packages(),
807        Command::Package(args) => match args.command {
808            PackageCommand::List(list) => package::list_packages(list.json),
809            PackageCommand::Doctor(doctor) => package::doctor_packages(doctor.json),
810            PackageCommand::Search(search) => package::search_package_registry(
811                search.query.as_deref(),
812                search.registry.as_deref(),
813                search.json,
814            ),
815            PackageCommand::Info(info) => {
816                package::show_package_registry_info(&info.name, info.registry.as_deref(), info.json)
817            }
818            PackageCommand::Check(check) => {
819                package::check_package(check.package.as_deref(), check.json)
820            }
821            PackageCommand::Pack(pack) => package::pack_package(
822                pack.package.as_deref(),
823                pack.output.as_deref(),
824                pack.dry_run,
825                pack.json,
826            ),
827            PackageCommand::Docs(docs) => package::generate_package_docs(
828                docs.package.as_deref(),
829                docs.output.as_deref(),
830                docs.check,
831            ),
832            PackageCommand::Cache(cache) => match cache.command {
833                PackageCacheCommand::List => package::list_package_cache(),
834                PackageCacheCommand::Clean(clean) => package::clean_package_cache(clean.all),
835                PackageCacheCommand::Verify(verify) => {
836                    package::verify_package_cache(verify.materialized)
837                }
838            },
839            PackageCommand::Outdated(args) => package::outdated_packages(
840                args.refresh,
841                args.remote,
842                args.registry.as_deref(),
843                args.json,
844            ),
845            PackageCommand::Audit(args) => {
846                package::audit_packages(args.registry.as_deref(), args.skip_materialized, args.json)
847            }
848            PackageCommand::Artifacts(args) => match args.command {
849                PackageArtifactsCommand::Manifest(manifest) => {
850                    package::artifacts_manifest(manifest.output.as_deref())
851                }
852                PackageArtifactsCommand::Check(check) => {
853                    package::artifacts_check(&check.manifest, check.json)
854                }
855            },
856        },
857        Command::Publish(args) => package::publish_package(
858            args.package.as_deref(),
859            args.dry_run,
860            args.registry.as_deref(),
861            args.json,
862        ),
863        Command::MergeCaptain(args) => match args.command {
864            MergeCaptainCommand::Run(run) => {
865                let code = commands::merge_captain::run_driver(&run);
866                if code != 0 {
867                    process::exit(code);
868                }
869            }
870            MergeCaptainCommand::Ladder(ladder) => {
871                let code = commands::merge_captain::run_ladder(&ladder);
872                if code != 0 {
873                    process::exit(code);
874                }
875            }
876            MergeCaptainCommand::Iterate(iterate) => {
877                let code = commands::merge_captain::run_iterate(&iterate);
878                if code != 0 {
879                    process::exit(code);
880                }
881            }
882            MergeCaptainCommand::Audit(audit) => {
883                let code = commands::merge_captain::run_audit(&audit);
884                if code != 0 {
885                    process::exit(code);
886                }
887            }
888            MergeCaptainCommand::Mock(mock) => {
889                let code = match mock {
890                    MergeCaptainMockCommand::Init(args) => {
891                        commands::merge_captain_mock::run_init(&args)
892                    }
893                    MergeCaptainMockCommand::Step(args) => {
894                        commands::merge_captain_mock::run_step(&args)
895                    }
896                    MergeCaptainMockCommand::Status(args) => {
897                        commands::merge_captain_mock::run_status(&args)
898                    }
899                    MergeCaptainMockCommand::Serve(args) => {
900                        commands::merge_captain_mock::run_serve(&args).await
901                    }
902                    MergeCaptainMockCommand::Cleanup(args) => {
903                        commands::merge_captain_mock::run_cleanup(&args)
904                    }
905                    MergeCaptainMockCommand::Scenarios => {
906                        commands::merge_captain_mock::run_scenarios()
907                    }
908                };
909                if code != 0 {
910                    process::exit(code);
911                }
912            }
913        },
914        Command::Persona(args) => match args.command {
915            PersonaCommand::New(new) => {
916                if let Err(error) = commands::persona_scaffold::run_new(&new) {
917                    eprintln!("error: {error}");
918                    process::exit(1);
919                }
920            }
921            PersonaCommand::Doctor(doctor) => {
922                if let Err(error) =
923                    commands::persona_doctor::run_doctor(args.manifest.as_deref(), &doctor).await
924                {
925                    eprintln!("error: {error}");
926                    process::exit(1);
927                }
928            }
929            PersonaCommand::Check(check) => {
930                commands::persona::run_check(args.manifest.as_deref(), &check)
931            }
932            PersonaCommand::List(list) => {
933                commands::persona::run_list(args.manifest.as_deref(), &list)
934            }
935            PersonaCommand::Inspect(inspect) => {
936                commands::persona::run_inspect(args.manifest.as_deref(), &inspect)
937            }
938            PersonaCommand::Status(status) => {
939                if let Err(error) = commands::persona::run_status(
940                    args.manifest.as_deref(),
941                    &args.state_dir,
942                    &status,
943                )
944                .await
945                {
946                    eprintln!("error: {error}");
947                    process::exit(1);
948                }
949            }
950            PersonaCommand::Pause(control) => {
951                if let Err(error) = commands::persona::run_pause(
952                    args.manifest.as_deref(),
953                    &args.state_dir,
954                    &control,
955                )
956                .await
957                {
958                    eprintln!("error: {error}");
959                    process::exit(1);
960                }
961            }
962            PersonaCommand::Resume(control) => {
963                if let Err(error) = commands::persona::run_resume(
964                    args.manifest.as_deref(),
965                    &args.state_dir,
966                    &control,
967                )
968                .await
969                {
970                    eprintln!("error: {error}");
971                    process::exit(1);
972                }
973            }
974            PersonaCommand::Disable(control) => {
975                if let Err(error) = commands::persona::run_disable(
976                    args.manifest.as_deref(),
977                    &args.state_dir,
978                    &control,
979                )
980                .await
981                {
982                    eprintln!("error: {error}");
983                    process::exit(1);
984                }
985            }
986            PersonaCommand::Tick(tick) => {
987                if let Err(error) =
988                    commands::persona::run_tick(args.manifest.as_deref(), &args.state_dir, &tick)
989                        .await
990                {
991                    eprintln!("error: {error}");
992                    process::exit(1);
993                }
994            }
995            PersonaCommand::Trigger(trigger) => {
996                if let Err(error) = commands::persona::run_trigger(
997                    args.manifest.as_deref(),
998                    &args.state_dir,
999                    &trigger,
1000                )
1001                .await
1002                {
1003                    eprintln!("error: {error}");
1004                    process::exit(1);
1005                }
1006            }
1007            PersonaCommand::Spend(spend) => {
1008                if let Err(error) =
1009                    commands::persona::run_spend(args.manifest.as_deref(), &args.state_dir, &spend)
1010                        .await
1011                {
1012                    eprintln!("error: {error}");
1013                    process::exit(1);
1014                }
1015            }
1016            PersonaCommand::Supervision(supervision) => match supervision.command {
1017                PersonaSupervisionCommand::Tail(tail) => {
1018                    if let Err(error) = commands::persona_supervision::run_tail(
1019                        args.manifest.as_deref(),
1020                        &args.state_dir,
1021                        &tail,
1022                    )
1023                    .await
1024                    {
1025                        eprintln!("error: {error}");
1026                        process::exit(1);
1027                    }
1028                }
1029            },
1030        },
1031        Command::ModelInfo(args) => {
1032            if !print_model_info(&args).await {
1033                process::exit(1);
1034            }
1035        }
1036        Command::ProviderCatalog(args) => print_provider_catalog(args.available_only),
1037        Command::ProviderReady(args) => {
1038            run_provider_ready(
1039                &args.provider,
1040                args.model.as_deref(),
1041                args.base_url.as_deref(),
1042                args.json,
1043            )
1044            .await
1045        }
1046        Command::Skills(args) => match args.command {
1047            SkillsCommand::List(list) => commands::skills::run_list(&list),
1048            SkillsCommand::Inspect(inspect) => commands::skills::run_inspect(&inspect),
1049            SkillsCommand::Match(matcher) => commands::skills::run_match(&matcher),
1050            SkillsCommand::Install(install) => commands::skills::run_install(&install),
1051            SkillsCommand::New(new_args) => commands::skills::run_new(&new_args),
1052        },
1053        Command::Tool(args) => match args.command {
1054            ToolCommand::New(new_args) => {
1055                if let Err(error) = commands::tool::run_new(&new_args) {
1056                    eprintln!("error: {error}");
1057                    process::exit(1);
1058                }
1059            }
1060        },
1061        Command::DumpHighlightKeywords(args) => {
1062            commands::dump_highlight_keywords::run(&args.output, args.check);
1063        }
1064        Command::DumpTriggerQuickref(args) => {
1065            commands::dump_trigger_quickref::run(&args.output, args.check);
1066        }
1067        Command::DumpConnectorMatrix(args) => {
1068            commands::check::connector_matrix::run_docs(&args.output, &args.sources, args.check);
1069        }
1070        Command::DumpProtocolArtifacts(args) => {
1071            commands::dump_protocol_artifacts::run(&args.output_dir, args.check);
1072        }
1073    }
1074}
1075
1076fn run_profile_options(args: &cli::ProfileArgs) -> commands::run::RunProfileOptions {
1077    commands::run::RunProfileOptions {
1078        text: args.text,
1079        json_path: args.json_path.clone(),
1080    }
1081}
1082
1083fn print_completions(shell: CompletionShell) {
1084    let mut command = Cli::command();
1085    let shell = clap_complete::Shell::from(shell);
1086    clap_complete::generate(shell, &mut command, "harn", &mut std::io::stdout());
1087}
1088
1089fn normalize_serve_args(mut raw_args: Vec<String>) -> Vec<String> {
1090    if raw_args.len() > 2
1091        && raw_args.get(1).is_some_and(|arg| arg == "serve")
1092        && !matches!(
1093            raw_args.get(2).map(String::as_str),
1094            Some("acp" | "a2a" | "api" | "mcp" | "-h" | "--help")
1095        )
1096    {
1097        raw_args.insert(2, "a2a".to_string());
1098    }
1099    raw_args
1100}
1101
1102fn print_version() {
1103    println!(
1104        r#"
1105 ╱▔▔╲
1106 ╱    ╲    harn v{}
1107 │ ◆  │    the agent harness language
1108 │    │
1109 ╰──╯╱
1110   ╱╱
1111"#,
1112        env!("CARGO_PKG_VERSION")
1113    );
1114}
1115
1116async fn print_model_info(args: &ModelInfoArgs) -> bool {
1117    let resolved = harn_vm::llm_config::resolve_model_info(&args.model);
1118    let api_key_result = harn_vm::llm::resolve_api_key(&resolved.provider);
1119    let api_key_set = api_key_result.is_ok();
1120    let api_key = api_key_result.unwrap_or_default();
1121    let context_window =
1122        harn_vm::llm::fetch_provider_max_context(&resolved.provider, &resolved.id, &api_key).await;
1123    let readiness = local_openai_readiness(&resolved.provider, &resolved.id, &api_key).await;
1124    let catalog = harn_vm::llm_config::model_catalog_entry(&resolved.id);
1125    let runtime_context_window = catalog
1126        .as_ref()
1127        .and_then(|entry| entry.runtime_context_window);
1128    let capabilities = harn_vm::llm::capabilities::lookup(&resolved.provider, &resolved.id);
1129    let mut payload = serde_json::json!({
1130        "alias": args.model,
1131        "id": resolved.id,
1132        "provider": resolved.provider,
1133        "resolved_alias": resolved.alias,
1134        "tool_format": resolved.tool_format,
1135        "tier": resolved.tier,
1136        "api_key_set": api_key_set,
1137        "context_window": context_window,
1138        "runtime_context_window": runtime_context_window,
1139        "readiness": readiness,
1140        "catalog": catalog,
1141        "capabilities": {
1142            "native_tools": capabilities.native_tools,
1143            "defer_loading": capabilities.defer_loading,
1144            "tool_search": capabilities.tool_search,
1145            "max_tools": capabilities.max_tools,
1146            "prompt_caching": capabilities.prompt_caching,
1147            "vision": capabilities.vision,
1148            "vision_supported": capabilities.vision_supported,
1149            "audio": capabilities.audio,
1150            "pdf": capabilities.pdf,
1151            "files_api_supported": capabilities.files_api_supported,
1152            "json_schema": capabilities.json_schema,
1153            "thinking": !capabilities.thinking_modes.is_empty(),
1154            "thinking_modes": capabilities.thinking_modes,
1155            "interleaved_thinking_supported": capabilities.interleaved_thinking_supported,
1156            "anthropic_beta_features": capabilities.anthropic_beta_features,
1157            "preserve_thinking": capabilities.preserve_thinking,
1158            "server_parser": capabilities.server_parser,
1159            "honors_chat_template_kwargs": capabilities.honors_chat_template_kwargs,
1160            "recommended_endpoint": capabilities.recommended_endpoint,
1161            "text_tool_wire_format_supported": capabilities.text_tool_wire_format_supported,
1162        },
1163        "qc_default_model": harn_vm::llm_config::qc_default_model(&resolved.provider),
1164    });
1165
1166    let should_verify = args.verify || args.warm;
1167    let mut ok = true;
1168    if should_verify {
1169        if resolved.provider == "ollama" {
1170            let mut readiness = harn_vm::llm::OllamaReadinessOptions::new(resolved.id.clone());
1171            readiness.warm = args.warm;
1172            readiness.keep_alive = args
1173                .keep_alive
1174                .as_deref()
1175                .and_then(harn_vm::llm::normalize_ollama_keep_alive);
1176            let result = harn_vm::llm::ollama_readiness(readiness).await;
1177            ok = result.valid;
1178            payload["readiness"] = serde_json::to_value(&result).unwrap_or_else(|error| {
1179                serde_json::json!({
1180                    "valid": false,
1181                    "status": "serialization_error",
1182                    "message": format!("failed to serialize readiness result: {error}"),
1183                })
1184            });
1185        } else {
1186            ok = false;
1187            payload["readiness"] = serde_json::json!({
1188                "valid": false,
1189                "status": "unsupported_provider",
1190                "message": format!(
1191                    "model-info --verify is only supported for Ollama models; resolved provider is '{}'",
1192                    resolved.provider
1193                ),
1194                "provider": resolved.provider,
1195            });
1196        }
1197    }
1198
1199    println!(
1200        "{}",
1201        serde_json::to_string(&payload).unwrap_or_else(|error| {
1202            command_error(&format!("failed to serialize model info: {error}"))
1203        })
1204    );
1205    ok
1206}
1207
1208async fn local_openai_readiness(
1209    provider: &str,
1210    model: &str,
1211    api_key: &str,
1212) -> Option<serde_json::Value> {
1213    let def = harn_vm::llm_config::provider_config(provider)?;
1214    if def.auth_style != "none" || !harn_vm::llm::supports_model_readiness_probe(&def) {
1215        return None;
1216    }
1217    let readiness = harn_vm::llm::probe_openai_compatible_model(provider, model, api_key).await;
1218    Some(serde_json::json!({
1219        "valid": readiness.valid,
1220        "category": readiness.category,
1221        "message": readiness.message,
1222        "provider": readiness.provider,
1223        "model": readiness.model,
1224        "url": readiness.url,
1225        "status": readiness.status,
1226        "available_models": readiness.available_models,
1227    }))
1228}
1229
1230fn print_provider_catalog(available_only: bool) {
1231    let provider_names = if available_only {
1232        harn_vm::llm_config::available_provider_names()
1233    } else {
1234        harn_vm::llm_config::provider_names()
1235    };
1236    let providers: Vec<_> = provider_names
1237        .into_iter()
1238        .filter_map(|name| {
1239            harn_vm::llm_config::provider_config(&name).map(|def| {
1240                serde_json::json!({
1241                    "name": name,
1242                    "display_name": def.display_name,
1243                    "icon": def.icon,
1244                    "base_url": harn_vm::llm_config::resolve_base_url(&def),
1245                    "base_url_env": def.base_url_env,
1246                    "auth_style": def.auth_style,
1247                    "auth_envs": harn_vm::llm_config::auth_env_names(&def.auth_env),
1248                    "auth_available": harn_vm::llm_config::provider_key_available(&name),
1249                    "features": def.features,
1250                    "cost_per_1k_in": def.cost_per_1k_in,
1251                    "cost_per_1k_out": def.cost_per_1k_out,
1252                    "latency_p50_ms": def.latency_p50_ms,
1253                })
1254            })
1255        })
1256        .collect();
1257    let models: Vec<_> = harn_vm::llm_config::model_catalog_entries()
1258        .into_iter()
1259        .map(|(id, model)| {
1260            serde_json::json!({
1261                "id": id,
1262                "name": model.name,
1263                "provider": model.provider,
1264                "context_window": model.context_window,
1265                "runtime_context_window": model.runtime_context_window,
1266                "stream_timeout": model.stream_timeout,
1267                "capabilities": model.capabilities,
1268                "pricing": model.pricing,
1269            })
1270        })
1271        .collect();
1272    let aliases: Vec<_> = harn_vm::llm_config::alias_entries()
1273        .into_iter()
1274        .map(|(name, alias)| {
1275            serde_json::json!({
1276                "name": name,
1277                "id": alias.id,
1278                "provider": alias.provider,
1279                "tool_format": alias.tool_format,
1280            })
1281        })
1282        .collect();
1283    let payload = serde_json::json!({
1284        "providers": providers,
1285        "known_model_names": harn_vm::llm_config::known_model_names(),
1286        "available_providers": harn_vm::llm_config::available_provider_names(),
1287        "aliases": aliases,
1288        "models": models,
1289        "qc_defaults": harn_vm::llm_config::qc_defaults(),
1290    });
1291    println!(
1292        "{}",
1293        serde_json::to_string(&payload).unwrap_or_else(|error| {
1294            command_error(&format!("failed to serialize provider catalog: {error}"))
1295        })
1296    );
1297}
1298
1299async fn run_provider_ready(
1300    provider: &str,
1301    model: Option<&str>,
1302    base_url: Option<&str>,
1303    json: bool,
1304) {
1305    let readiness =
1306        harn_vm::llm::readiness::probe_provider_readiness(provider, model, base_url).await;
1307    if json {
1308        match serde_json::to_string_pretty(&readiness) {
1309            Ok(payload) => println!("{payload}"),
1310            Err(error) => command_error(&format!("failed to serialize readiness result: {error}")),
1311        }
1312    } else if readiness.ok {
1313        println!("{}", readiness.message);
1314    } else {
1315        eprintln!("{}", readiness.message);
1316    }
1317    if !readiness.ok {
1318        process::exit(1);
1319    }
1320}
1321
1322fn command_error(message: &str) -> ! {
1323    Cli::command()
1324        .error(ErrorKind::ValueValidation, message)
1325        .exit()
1326}
1327
1328fn verify_provenance_receipt(path: &str, json: bool) -> Result<(), String> {
1329    let raw =
1330        fs::read_to_string(path).map_err(|error| format!("failed to read {path}: {error}"))?;
1331    let receipt: harn_vm::ProvenanceReceipt = serde_json::from_str(&raw)
1332        .map_err(|error| format!("failed to parse provenance receipt {path}: {error}"))?;
1333    let report = harn_vm::verify_receipt(&receipt);
1334    if json {
1335        println!(
1336            "{}",
1337            serde_json::to_string_pretty(&report).map_err(|error| error.to_string())?
1338        );
1339    } else if report.verified {
1340        println!(
1341            "verified receipt={} events={} receipt_hash={} event_root_hash={}",
1342            report.receipt_id.unwrap_or_else(|| "-".to_string()),
1343            report.event_count,
1344            report.receipt_hash.unwrap_or_else(|| "-".to_string()),
1345            report.event_root_hash.unwrap_or_else(|| "-".to_string())
1346        );
1347    } else {
1348        println!(
1349            "failed receipt={} events={}",
1350            report.receipt_id.unwrap_or_else(|| "-".to_string()),
1351            report.event_count
1352        );
1353        for error in &report.errors {
1354            println!("  {error}");
1355        }
1356        return Err("provenance receipt verification failed".to_string());
1357    }
1358    Ok(())
1359}
1360
1361fn load_run_record_or_exit(path: &Path) -> harn_vm::orchestration::RunRecord {
1362    match harn_vm::orchestration::load_run_record(path) {
1363        Ok(run) => run,
1364        Err(error) => {
1365            eprintln!("Failed to load run record: {error}");
1366            process::exit(1);
1367        }
1368    }
1369}
1370
1371fn load_eval_suite_manifest_or_exit(path: &Path) -> harn_vm::orchestration::EvalSuiteManifest {
1372    harn_vm::orchestration::load_eval_suite_manifest(path).unwrap_or_else(|error| {
1373        eprintln!("Failed to load eval manifest {}: {error}", path.display());
1374        process::exit(1);
1375    })
1376}
1377
1378fn load_eval_pack_manifest_or_exit(path: &Path) -> harn_vm::orchestration::EvalPackManifest {
1379    harn_vm::orchestration::load_eval_pack_manifest(path).unwrap_or_else(|error| {
1380        eprintln!("Failed to load eval pack {}: {error}", path.display());
1381        process::exit(1);
1382    })
1383}
1384
1385fn load_persona_eval_ladder_manifest_or_exit(
1386    path: &Path,
1387) -> harn_vm::orchestration::PersonaEvalLadderManifest {
1388    harn_vm::orchestration::load_persona_eval_ladder_manifest(path).unwrap_or_else(|error| {
1389        eprintln!(
1390            "Failed to load persona eval ladder {}: {error}",
1391            path.display()
1392        );
1393        process::exit(1);
1394    })
1395}
1396
1397fn file_looks_like_eval_manifest(path: &Path) -> bool {
1398    if path.file_name().and_then(|name| name.to_str()) == Some("harn.eval.toml") {
1399        return true;
1400    }
1401    if path.extension().and_then(|ext| ext.to_str()) == Some("toml") {
1402        let Ok(content) = fs::read_to_string(path) else {
1403            return false;
1404        };
1405        return toml::from_str::<harn_vm::orchestration::EvalPackManifest>(&content)
1406            .is_ok_and(|manifest| !manifest.cases.is_empty() || !manifest.ladders.is_empty());
1407    }
1408    let Ok(content) = fs::read_to_string(path) else {
1409        return false;
1410    };
1411    let Ok(json) = serde_json::from_str::<serde_json::Value>(&content) else {
1412        return false;
1413    };
1414    json.get("_type").and_then(|value| value.as_str()) == Some("eval_suite_manifest")
1415        || json.get("cases").is_some()
1416}
1417
1418fn file_looks_like_eval_pack_manifest(path: &Path) -> bool {
1419    if path.file_name().and_then(|name| name.to_str()) == Some("harn.eval.toml") {
1420        return true;
1421    }
1422    if path.extension().and_then(|ext| ext.to_str()) == Some("toml") {
1423        return file_looks_like_eval_manifest(path);
1424    }
1425    let Ok(content) = fs::read_to_string(path) else {
1426        return false;
1427    };
1428    let Ok(json) = serde_json::from_str::<serde_json::Value>(&content) else {
1429        return false;
1430    };
1431    json.get("version").is_some()
1432        && (json.get("cases").is_some() || json.get("ladders").is_some())
1433        && json.get("_type").and_then(|value| value.as_str()) != Some("eval_suite_manifest")
1434}
1435
1436fn file_looks_like_persona_eval_ladder_manifest(path: &Path) -> bool {
1437    let Ok(content) = fs::read_to_string(path) else {
1438        return false;
1439    };
1440    if path.extension().and_then(|ext| ext.to_str()) == Some("json") {
1441        let Ok(json) = serde_json::from_str::<serde_json::Value>(&content) else {
1442            return false;
1443        };
1444        return json.get("_type").and_then(|value| value.as_str())
1445            == Some("persona_eval_ladder_manifest")
1446            || json.get("timeout_tiers").is_some()
1447            || json.get("timeout-tiers").is_some();
1448    }
1449    toml::from_str::<harn_vm::orchestration::PersonaEvalLadderManifest>(&content).is_ok_and(
1450        |manifest| {
1451            manifest
1452                .type_name
1453                .eq_ignore_ascii_case("persona_eval_ladder_manifest")
1454                || (!manifest.timeout_tiers.is_empty() && manifest.backend.path.is_some())
1455        },
1456    )
1457}
1458
1459fn collect_run_record_paths(path: &str) -> Vec<PathBuf> {
1460    let path = Path::new(path);
1461    if path.is_file() {
1462        return vec![path.to_path_buf()];
1463    }
1464    if path.is_dir() {
1465        let mut entries: Vec<PathBuf> = fs::read_dir(path)
1466            .unwrap_or_else(|error| {
1467                eprintln!("Failed to read run directory {}: {error}", path.display());
1468                process::exit(1);
1469            })
1470            .filter_map(|entry| entry.ok().map(|entry| entry.path()))
1471            .filter(|entry| entry.extension().and_then(|ext| ext.to_str()) == Some("json"))
1472            .collect();
1473        entries.sort();
1474        return entries;
1475    }
1476    eprintln!("Run path does not exist: {}", path.display());
1477    process::exit(1);
1478}
1479
1480fn print_run_diff(diff: &harn_vm::orchestration::RunDiffReport) {
1481    println!(
1482        "Diff: {} -> {} [{} -> {}]",
1483        diff.left_run_id, diff.right_run_id, diff.left_status, diff.right_status
1484    );
1485    println!("Identical: {}", diff.identical);
1486    println!("Stage diffs: {}", diff.stage_diffs.len());
1487    println!("Tool diffs: {}", diff.tool_diffs.len());
1488    println!("Observability diffs: {}", diff.observability_diffs.len());
1489    println!("Transition delta: {}", diff.transition_count_delta);
1490    println!("Artifact delta: {}", diff.artifact_count_delta);
1491    println!("Checkpoint delta: {}", diff.checkpoint_count_delta);
1492    for stage in &diff.stage_diffs {
1493        println!("- {} [{}]", stage.node_id, stage.change);
1494        for detail in &stage.details {
1495            println!("  {}", detail);
1496        }
1497    }
1498    for tool in &diff.tool_diffs {
1499        println!("- tool {} [{}]", tool.tool_name, tool.args_hash);
1500        println!("  left: {:?}", tool.left_result);
1501        println!("  right: {:?}", tool.right_result);
1502    }
1503    for item in &diff.observability_diffs {
1504        println!("- {} [{}]", item.label, item.section);
1505        for detail in &item.details {
1506            println!("  {}", detail);
1507        }
1508    }
1509}
1510
1511fn inspect_run_record(path: &str, compare: Option<&str>) {
1512    let run = load_run_record_or_exit(Path::new(path));
1513    println!("Run: {}", run.id);
1514    println!(
1515        "Workflow: {}",
1516        run.workflow_name
1517            .clone()
1518            .unwrap_or_else(|| run.workflow_id.clone())
1519    );
1520    println!("Status: {}", run.status);
1521    println!("Task: {}", run.task);
1522    println!("Stages: {}", run.stages.len());
1523    println!("Artifacts: {}", run.artifacts.len());
1524    println!("Transitions: {}", run.transitions.len());
1525    println!("Checkpoints: {}", run.checkpoints.len());
1526    println!("HITL questions: {}", run.hitl_questions.len());
1527    if let Some(observability) = &run.observability {
1528        println!("Planner rounds: {}", observability.planner_rounds.len());
1529        println!("Research facts: {}", observability.research_fact_count);
1530        println!("Workers: {}", observability.worker_lineage.len());
1531        println!(
1532            "Action graph: {} nodes / {} edges",
1533            observability.action_graph_nodes.len(),
1534            observability.action_graph_edges.len()
1535        );
1536        println!(
1537            "Transcript pointers: {}",
1538            observability.transcript_pointers.len()
1539        );
1540        println!("Daemon events: {}", observability.daemon_events.len());
1541    }
1542    if let Some(parent_worker_id) = run
1543        .metadata
1544        .get("parent_worker_id")
1545        .and_then(|value| value.as_str())
1546    {
1547        println!("Parent worker: {}", parent_worker_id);
1548    }
1549    if let Some(parent_stage_id) = run
1550        .metadata
1551        .get("parent_stage_id")
1552        .and_then(|value| value.as_str())
1553    {
1554        println!("Parent stage: {}", parent_stage_id);
1555    }
1556    if run
1557        .metadata
1558        .get("delegated")
1559        .and_then(|value| value.as_bool())
1560        .unwrap_or(false)
1561    {
1562        println!("Delegated: true");
1563    }
1564    println!(
1565        "Pending nodes: {}",
1566        if run.pending_nodes.is_empty() {
1567            "-".to_string()
1568        } else {
1569            run.pending_nodes.join(", ")
1570        }
1571    );
1572    println!(
1573        "Replay fixture: {}",
1574        if run.replay_fixture.is_some() {
1575            "embedded"
1576        } else {
1577            "derived"
1578        }
1579    );
1580    for stage in &run.stages {
1581        let worker = stage.metadata.get("worker");
1582        let worker_suffix = worker
1583            .and_then(|value| value.get("name"))
1584            .and_then(|value| value.as_str())
1585            .map(|name| format!(" worker={name}"))
1586            .unwrap_or_default();
1587        println!(
1588            "- {} [{}] status={} outcome={} branch={}{}",
1589            stage.node_id,
1590            stage.kind,
1591            stage.status,
1592            stage.outcome,
1593            stage.branch.clone().unwrap_or_else(|| "-".to_string()),
1594            worker_suffix,
1595        );
1596        if let Some(worker) = worker {
1597            if let Some(worker_id) = worker.get("id").and_then(|value| value.as_str()) {
1598                println!("  worker_id: {}", worker_id);
1599            }
1600            if let Some(child_run_id) = worker.get("child_run_id").and_then(|value| value.as_str())
1601            {
1602                println!("  child_run_id: {}", child_run_id);
1603            }
1604            if let Some(child_run_path) = worker
1605                .get("child_run_path")
1606                .and_then(|value| value.as_str())
1607            {
1608                println!("  child_run_path: {}", child_run_path);
1609            }
1610        }
1611    }
1612    if let Some(observability) = &run.observability {
1613        for round in &observability.planner_rounds {
1614            println!(
1615                "- planner {} iterations={} llm_calls={} tools={} research_facts={}",
1616                round.node_id,
1617                round.iteration_count,
1618                round.llm_call_count,
1619                round.tool_execution_count,
1620                round.research_facts.len()
1621            );
1622        }
1623        for pointer in &observability.transcript_pointers {
1624            println!(
1625                "- transcript {} [{}] available={} {}",
1626                pointer.label,
1627                pointer.kind,
1628                pointer.available,
1629                pointer
1630                    .path
1631                    .clone()
1632                    .unwrap_or_else(|| pointer.location.clone())
1633            );
1634        }
1635        for event in &observability.daemon_events {
1636            println!(
1637                "- daemon {} [{:?}] at {}",
1638                event.name, event.kind, event.timestamp
1639            );
1640            println!("  id: {}", event.daemon_id);
1641            println!("  persist_path: {}", event.persist_path);
1642            if let Some(summary) = &event.payload_summary {
1643                println!("  payload: {}", summary);
1644            }
1645        }
1646    }
1647    if let Some(compare_path) = compare {
1648        let baseline = load_run_record_or_exit(Path::new(compare_path));
1649        print_run_diff(&harn_vm::orchestration::diff_run_records(&baseline, &run));
1650    }
1651}
1652
1653fn replay_run_record(path: &str) {
1654    let run = load_run_record_or_exit(Path::new(path));
1655    println!("Replay: {}", run.id);
1656    for stage in &run.stages {
1657        println!(
1658            "[{}] status={} outcome={} branch={}",
1659            stage.node_id,
1660            stage.status,
1661            stage.outcome,
1662            stage.branch.clone().unwrap_or_else(|| "-".to_string())
1663        );
1664        if let Some(text) = &stage.visible_text {
1665            println!("  visible: {}", text);
1666        }
1667        if let Some(verification) = &stage.verification {
1668            println!("  verification: {}", verification);
1669        }
1670    }
1671    if let Some(transcript) = &run.transcript {
1672        println!(
1673            "Transcript events persisted: {}",
1674            transcript["events"]
1675                .as_array()
1676                .map(|v| v.len())
1677                .unwrap_or(0)
1678        );
1679    }
1680    let fixture = run
1681        .replay_fixture
1682        .clone()
1683        .unwrap_or_else(|| harn_vm::orchestration::replay_fixture_from_run(&run));
1684    let report = harn_vm::orchestration::evaluate_run_against_fixture(&run, &fixture);
1685    println!(
1686        "Embedded replay fixture: {}",
1687        if report.pass { "PASS" } else { "FAIL" }
1688    );
1689    for transition in &run.transitions {
1690        println!(
1691            "transition {} -> {} ({})",
1692            transition
1693                .from_node_id
1694                .clone()
1695                .unwrap_or_else(|| "start".to_string()),
1696            transition.to_node_id,
1697            transition
1698                .branch
1699                .clone()
1700                .unwrap_or_else(|| "default".to_string())
1701        );
1702    }
1703}
1704
1705fn eval_run_record(
1706    path: &str,
1707    compare: Option<&str>,
1708    structural_experiment: Option<&str>,
1709    argv: &[String],
1710    llm_mock_mode: &commands::run::CliLlmMockMode,
1711) {
1712    if let Some(experiment) = structural_experiment {
1713        let path_buf = PathBuf::from(path);
1714        if !path_buf.is_file() || path_buf.extension().and_then(|ext| ext.to_str()) != Some("harn")
1715        {
1716            eprintln!(
1717                "--structural-experiment currently requires a .harn pipeline path, got {}",
1718                path
1719            );
1720            process::exit(1);
1721        }
1722        if compare.is_some() {
1723            eprintln!("--compare cannot be combined with --structural-experiment");
1724            process::exit(1);
1725        }
1726        if matches!(llm_mock_mode, commands::run::CliLlmMockMode::Record { .. }) {
1727            eprintln!("--llm-mock-record cannot be combined with --structural-experiment");
1728            process::exit(1);
1729        }
1730        let path_buf = fs::canonicalize(&path_buf).unwrap_or_else(|error| {
1731            command_error(&format!(
1732                "failed to canonicalize structural eval pipeline {}: {error}",
1733                path_buf.display()
1734            ))
1735        });
1736        run_structural_experiment_eval(&path_buf, experiment, argv, llm_mock_mode);
1737        return;
1738    }
1739
1740    let path_buf = PathBuf::from(path);
1741    if path_buf.is_file() && file_looks_like_persona_eval_ladder_manifest(&path_buf) {
1742        if compare.is_some() {
1743            eprintln!("--compare is not supported with persona eval ladder manifests");
1744            process::exit(1);
1745        }
1746        let manifest = load_persona_eval_ladder_manifest_or_exit(&path_buf);
1747        let report =
1748            harn_vm::orchestration::run_persona_eval_ladder(&manifest).unwrap_or_else(|error| {
1749                eprintln!(
1750                    "Failed to evaluate persona eval ladder {}: {error}",
1751                    path_buf.display()
1752                );
1753                process::exit(1);
1754            });
1755        print_persona_ladder_report(&report);
1756        if !report.pass {
1757            process::exit(1);
1758        }
1759        return;
1760    }
1761
1762    if path_buf.is_file() && file_looks_like_eval_pack_manifest(&path_buf) {
1763        if compare.is_some() {
1764            eprintln!("--compare is not supported with eval pack manifests");
1765            process::exit(1);
1766        }
1767        let manifest = load_eval_pack_manifest_or_exit(&path_buf);
1768        let report = harn_vm::orchestration::evaluate_eval_pack_manifest(&manifest).unwrap_or_else(
1769            |error| {
1770                eprintln!(
1771                    "Failed to evaluate eval pack {}: {error}",
1772                    path_buf.display()
1773                );
1774                process::exit(1);
1775            },
1776        );
1777        print_eval_pack_report(&report);
1778        if !report.pass {
1779            process::exit(1);
1780        }
1781        return;
1782    }
1783
1784    if path_buf.is_file() && file_looks_like_eval_manifest(&path_buf) {
1785        if compare.is_some() {
1786            eprintln!("--compare is not supported with eval suite manifests");
1787            process::exit(1);
1788        }
1789        let manifest = load_eval_suite_manifest_or_exit(&path_buf);
1790        let suite = harn_vm::orchestration::evaluate_run_suite_manifest(&manifest).unwrap_or_else(
1791            |error| {
1792                eprintln!(
1793                    "Failed to evaluate manifest {}: {error}",
1794                    path_buf.display()
1795                );
1796                process::exit(1);
1797            },
1798        );
1799        println!(
1800            "{} {} passed, {} failed, {} total",
1801            if suite.pass { "PASS" } else { "FAIL" },
1802            suite.passed,
1803            suite.failed,
1804            suite.total
1805        );
1806        for case in &suite.cases {
1807            println!(
1808                "- {} [{}] {}",
1809                case.label.clone().unwrap_or_else(|| case.run_id.clone()),
1810                case.workflow_id,
1811                if case.pass { "PASS" } else { "FAIL" }
1812            );
1813            if let Some(path) = &case.source_path {
1814                println!("  path: {}", path);
1815            }
1816            if let Some(comparison) = &case.comparison {
1817                println!("  baseline identical: {}", comparison.identical);
1818                if !comparison.identical {
1819                    println!(
1820                        "  baseline status: {} -> {}",
1821                        comparison.left_status, comparison.right_status
1822                    );
1823                }
1824            }
1825            for failure in &case.failures {
1826                println!("  {}", failure);
1827            }
1828        }
1829        if !suite.pass {
1830            process::exit(1);
1831        }
1832        return;
1833    }
1834
1835    let paths = collect_run_record_paths(path);
1836    if paths.len() > 1 {
1837        let mut cases = Vec::new();
1838        for path in &paths {
1839            let run = load_run_record_or_exit(path);
1840            let fixture = run
1841                .replay_fixture
1842                .clone()
1843                .unwrap_or_else(|| harn_vm::orchestration::replay_fixture_from_run(&run));
1844            cases.push((run, fixture, Some(path.display().to_string())));
1845        }
1846        let suite = harn_vm::orchestration::evaluate_run_suite(cases);
1847        println!(
1848            "{} {} passed, {} failed, {} total",
1849            if suite.pass { "PASS" } else { "FAIL" },
1850            suite.passed,
1851            suite.failed,
1852            suite.total
1853        );
1854        for case in &suite.cases {
1855            println!(
1856                "- {} [{}] {}",
1857                case.run_id,
1858                case.workflow_id,
1859                if case.pass { "PASS" } else { "FAIL" }
1860            );
1861            if let Some(path) = &case.source_path {
1862                println!("  path: {}", path);
1863            }
1864            if let Some(comparison) = &case.comparison {
1865                println!("  baseline identical: {}", comparison.identical);
1866            }
1867            for failure in &case.failures {
1868                println!("  {}", failure);
1869            }
1870        }
1871        if !suite.pass {
1872            process::exit(1);
1873        }
1874        return;
1875    }
1876
1877    let run = load_run_record_or_exit(&paths[0]);
1878    let fixture = run
1879        .replay_fixture
1880        .clone()
1881        .unwrap_or_else(|| harn_vm::orchestration::replay_fixture_from_run(&run));
1882    let report = harn_vm::orchestration::evaluate_run_against_fixture(&run, &fixture);
1883    println!("{}", if report.pass { "PASS" } else { "FAIL" });
1884    println!("Stages: {}", report.stage_count);
1885    if let Some(compare_path) = compare {
1886        let baseline = load_run_record_or_exit(Path::new(compare_path));
1887        print_run_diff(&harn_vm::orchestration::diff_run_records(&baseline, &run));
1888    }
1889    if !report.failures.is_empty() {
1890        for failure in &report.failures {
1891            println!("- {}", failure);
1892        }
1893    }
1894    if !report.pass {
1895        process::exit(1);
1896    }
1897}
1898
1899fn print_eval_pack_report(report: &harn_vm::orchestration::EvalPackReport) {
1900    println!(
1901        "{} {} passed, {} blocking failed, {} warning, {} informational, {} total",
1902        if report.pass { "PASS" } else { "FAIL" },
1903        report.passed,
1904        report.blocking_failed,
1905        report.warning_failed,
1906        report.informational_failed,
1907        report.total
1908    );
1909    for case in &report.cases {
1910        println!(
1911            "- {} [{}] {} ({})",
1912            case.label,
1913            case.workflow_id,
1914            if case.pass { "PASS" } else { "FAIL" },
1915            case.severity
1916        );
1917        if let Some(path) = &case.source_path {
1918            println!("  path: {}", path);
1919        }
1920        if let Some(comparison) = &case.comparison {
1921            println!("  baseline identical: {}", comparison.identical);
1922            if !comparison.identical {
1923                println!(
1924                    "  baseline status: {} -> {}",
1925                    comparison.left_status, comparison.right_status
1926                );
1927            }
1928        }
1929        for failure in &case.failures {
1930            println!("  {}", failure);
1931        }
1932        for warning in &case.warnings {
1933            println!("  warning: {}", warning);
1934        }
1935        for item in &case.informational {
1936            println!("  info: {}", item);
1937        }
1938    }
1939    for ladder in &report.ladders {
1940        println!(
1941            "- ladder {} [{}] {} ({}) first_correct={}/{}",
1942            ladder.id,
1943            ladder.persona,
1944            if ladder.pass { "PASS" } else { "FAIL" },
1945            ladder.severity,
1946            ladder.first_correct_route.as_deref().unwrap_or("<none>"),
1947            ladder.first_correct_tier.as_deref().unwrap_or("<none>")
1948        );
1949        println!("  artifacts: {}", ladder.artifact_root);
1950        for tier in &ladder.tiers {
1951            println!(
1952                "  - {} [{}] {} tools={} models={} latency={}ms cost=${:.6}",
1953                tier.timeout_tier,
1954                tier.route_id,
1955                tier.outcome,
1956                tier.tool_calls,
1957                tier.model_calls,
1958                tier.latency_ms,
1959                tier.cost_usd
1960            );
1961            for reason in &tier.degradation_reasons {
1962                println!("    {}", reason);
1963            }
1964        }
1965    }
1966}
1967
1968fn print_persona_ladder_report(report: &harn_vm::orchestration::PersonaEvalLadderReport) {
1969    println!(
1970        "{} ladder {} passed, {} degraded/looped, {} total",
1971        if report.pass { "PASS" } else { "FAIL" },
1972        report.passed,
1973        report.failed,
1974        report.total
1975    );
1976    println!(
1977        "first_correct: {}/{}",
1978        report.first_correct_route.as_deref().unwrap_or("<none>"),
1979        report.first_correct_tier.as_deref().unwrap_or("<none>")
1980    );
1981    println!("artifacts: {}", report.artifact_root);
1982    for tier in &report.tiers {
1983        println!(
1984            "- {} [{}] {} tools={} models={} latency={}ms cost=${:.6}",
1985            tier.timeout_tier,
1986            tier.route_id,
1987            tier.outcome,
1988            tier.tool_calls,
1989            tier.model_calls,
1990            tier.latency_ms,
1991            tier.cost_usd
1992        );
1993        for reason in &tier.degradation_reasons {
1994            println!("  {}", reason);
1995        }
1996    }
1997}
1998
1999fn run_package_evals() {
2000    let paths = package::load_package_eval_pack_paths(None).unwrap_or_else(|error| {
2001        eprintln!("{error}");
2002        process::exit(1);
2003    });
2004    let mut all_pass = true;
2005    for path in &paths {
2006        println!("Eval pack: {}", path.display());
2007        let manifest = load_eval_pack_manifest_or_exit(path);
2008        let report = harn_vm::orchestration::evaluate_eval_pack_manifest(&manifest).unwrap_or_else(
2009            |error| {
2010                eprintln!("Failed to evaluate eval pack {}: {error}", path.display());
2011                process::exit(1);
2012            },
2013        );
2014        print_eval_pack_report(&report);
2015        all_pass &= report.pass;
2016    }
2017    if !all_pass {
2018        process::exit(1);
2019    }
2020}
2021
2022fn run_structural_experiment_eval(
2023    path: &Path,
2024    experiment: &str,
2025    argv: &[String],
2026    llm_mock_mode: &commands::run::CliLlmMockMode,
2027) {
2028    let baseline_dir = tempfile::Builder::new()
2029        .prefix("harn-eval-baseline-")
2030        .tempdir()
2031        .unwrap_or_else(|error| {
2032            command_error(&format!("failed to create baseline tempdir: {error}"))
2033        });
2034    let variant_dir = tempfile::Builder::new()
2035        .prefix("harn-eval-variant-")
2036        .tempdir()
2037        .unwrap_or_else(|error| {
2038            command_error(&format!("failed to create variant tempdir: {error}"))
2039        });
2040
2041    let baseline = spawn_eval_pipeline_run(path, baseline_dir.path(), None, argv, llm_mock_mode);
2042    if !baseline.status.success() {
2043        relay_subprocess_failure("baseline", &baseline);
2044    }
2045
2046    let variant = spawn_eval_pipeline_run(
2047        path,
2048        variant_dir.path(),
2049        Some(experiment),
2050        argv,
2051        llm_mock_mode,
2052    );
2053    if !variant.status.success() {
2054        relay_subprocess_failure("variant", &variant);
2055    }
2056
2057    let baseline_runs = collect_structural_eval_runs(baseline_dir.path());
2058    let variant_runs = collect_structural_eval_runs(variant_dir.path());
2059    if baseline_runs.is_empty() || variant_runs.is_empty() {
2060        eprintln!(
2061            "structural eval expected workflow run records under {} and {}, but one side was empty",
2062            baseline_dir.path().display(),
2063            variant_dir.path().display()
2064        );
2065        process::exit(1);
2066    }
2067    if baseline_runs.len() != variant_runs.len() {
2068        eprintln!(
2069            "structural eval produced different run counts: baseline={} variant={}",
2070            baseline_runs.len(),
2071            variant_runs.len()
2072        );
2073        process::exit(1);
2074    }
2075
2076    let mut baseline_ok = 0usize;
2077    let mut variant_ok = 0usize;
2078    let mut any_failures = false;
2079
2080    println!("Structural experiment: {}", experiment);
2081    println!("Cases: {}", baseline_runs.len());
2082    for (baseline_run, variant_run) in baseline_runs.iter().zip(variant_runs.iter()) {
2083        let baseline_fixture = baseline_run
2084            .replay_fixture
2085            .clone()
2086            .unwrap_or_else(|| harn_vm::orchestration::replay_fixture_from_run(baseline_run));
2087        let variant_fixture = variant_run
2088            .replay_fixture
2089            .clone()
2090            .unwrap_or_else(|| harn_vm::orchestration::replay_fixture_from_run(variant_run));
2091        let baseline_report =
2092            harn_vm::orchestration::evaluate_run_against_fixture(baseline_run, &baseline_fixture);
2093        let variant_report =
2094            harn_vm::orchestration::evaluate_run_against_fixture(variant_run, &variant_fixture);
2095        let diff = harn_vm::orchestration::diff_run_records(baseline_run, variant_run);
2096        if baseline_report.pass {
2097            baseline_ok += 1;
2098        }
2099        if variant_report.pass {
2100            variant_ok += 1;
2101        }
2102        any_failures |= !baseline_report.pass || !variant_report.pass;
2103        println!(
2104            "- {} [{}]",
2105            variant_run
2106                .workflow_name
2107                .clone()
2108                .unwrap_or_else(|| variant_run.workflow_id.clone()),
2109            variant_run.task
2110        );
2111        println!(
2112            "  baseline: {}",
2113            if baseline_report.pass { "PASS" } else { "FAIL" }
2114        );
2115        for failure in &baseline_report.failures {
2116            println!("    {}", failure);
2117        }
2118        println!(
2119            "  variant: {}",
2120            if variant_report.pass { "PASS" } else { "FAIL" }
2121        );
2122        for failure in &variant_report.failures {
2123            println!("    {}", failure);
2124        }
2125        println!("  diff identical: {}", diff.identical);
2126        println!("  stage diffs: {}", diff.stage_diffs.len());
2127        println!("  tool diffs: {}", diff.tool_diffs.len());
2128        println!("  observability diffs: {}", diff.observability_diffs.len());
2129    }
2130
2131    println!("Baseline {} / {} passed", baseline_ok, baseline_runs.len());
2132    println!("Variant {} / {} passed", variant_ok, variant_runs.len());
2133
2134    if any_failures {
2135        process::exit(1);
2136    }
2137}
2138
2139fn spawn_eval_pipeline_run(
2140    path: &Path,
2141    run_dir: &Path,
2142    structural_experiment: Option<&str>,
2143    argv: &[String],
2144    llm_mock_mode: &commands::run::CliLlmMockMode,
2145) -> std::process::Output {
2146    let exe = env::current_exe().unwrap_or_else(|error| {
2147        command_error(&format!("failed to resolve current executable: {error}"))
2148    });
2149    let mut command = std::process::Command::new(exe);
2150    command.current_dir(path.parent().unwrap_or_else(|| Path::new(".")));
2151    command.arg("run");
2152    match llm_mock_mode {
2153        commands::run::CliLlmMockMode::Off => {}
2154        commands::run::CliLlmMockMode::Replay { fixture_path } => {
2155            command
2156                .arg("--llm-mock")
2157                .arg(absolute_cli_path(fixture_path));
2158        }
2159        commands::run::CliLlmMockMode::Record { fixture_path } => {
2160            command
2161                .arg("--llm-mock-record")
2162                .arg(absolute_cli_path(fixture_path));
2163        }
2164    }
2165    command.arg(path);
2166    if !argv.is_empty() {
2167        command.arg("--");
2168        command.args(argv);
2169    }
2170    command.env(harn_vm::runtime_paths::HARN_RUN_DIR_ENV, run_dir);
2171    if let Some(experiment) = structural_experiment {
2172        command.env("HARN_STRUCTURAL_EXPERIMENT", experiment);
2173    }
2174    command.output().unwrap_or_else(|error| {
2175        command_error(&format!(
2176            "failed to spawn `harn run {}` for structural eval: {error}",
2177            path.display()
2178        ))
2179    })
2180}
2181
2182fn absolute_cli_path(path: &Path) -> PathBuf {
2183    if path.is_absolute() {
2184        return path.to_path_buf();
2185    }
2186    env::current_dir()
2187        .unwrap_or_else(|_| PathBuf::from("."))
2188        .join(path)
2189}
2190
2191fn relay_subprocess_failure(label: &str, output: &std::process::Output) -> ! {
2192    let stdout = String::from_utf8_lossy(&output.stdout);
2193    let stderr = String::from_utf8_lossy(&output.stderr);
2194    if !stdout.trim().is_empty() {
2195        eprintln!("[{label}] stdout:\n{stdout}");
2196    }
2197    if !stderr.trim().is_empty() {
2198        eprintln!("[{label}] stderr:\n{stderr}");
2199    }
2200    process::exit(output.status.code().unwrap_or(1));
2201}
2202
2203fn collect_structural_eval_runs(dir: &Path) -> Vec<harn_vm::orchestration::RunRecord> {
2204    let mut paths: Vec<PathBuf> = fs::read_dir(dir)
2205        .unwrap_or_else(|error| {
2206            command_error(&format!(
2207                "failed to read structural eval run dir {}: {error}",
2208                dir.display()
2209            ))
2210        })
2211        .filter_map(|entry| entry.ok().map(|entry| entry.path()))
2212        .filter(|entry| entry.extension().and_then(|ext| ext.to_str()) == Some("json"))
2213        .collect();
2214    paths.sort();
2215    let mut runs: Vec<_> = paths
2216        .iter()
2217        .map(|path| load_run_record_or_exit(path))
2218        .collect();
2219    runs.sort_by(|left, right| {
2220        (
2221            left.started_at.as_str(),
2222            left.workflow_id.as_str(),
2223            left.task.as_str(),
2224        )
2225            .cmp(&(
2226                right.started_at.as_str(),
2227                right.workflow_id.as_str(),
2228                right.task.as_str(),
2229            ))
2230    });
2231    runs
2232}
2233
2234/// Parse a .harn file, returning (source, AST). Exits on error.
2235pub(crate) fn parse_source_file(path: &str) -> (String, Vec<harn_parser::SNode>) {
2236    let source = match fs::read_to_string(path) {
2237        Ok(s) => s,
2238        Err(e) => {
2239            eprintln!("Error reading {path}: {e}");
2240            process::exit(1);
2241        }
2242    };
2243
2244    let mut lexer = Lexer::new(&source);
2245    let tokens = match lexer.tokenize() {
2246        Ok(t) => t,
2247        Err(e) => {
2248            let diagnostic = harn_parser::diagnostic::render_diagnostic(
2249                &source,
2250                path,
2251                &error_span_from_lex(&e),
2252                "error",
2253                &e.to_string(),
2254                Some("here"),
2255                None,
2256            );
2257            eprint!("{diagnostic}");
2258            process::exit(1);
2259        }
2260    };
2261
2262    let mut parser = Parser::new(tokens);
2263    let program = match parser.parse() {
2264        Ok(p) => p,
2265        Err(err) => {
2266            if parser.all_errors().is_empty() {
2267                let span = error_span_from_parse(&err);
2268                let diagnostic = harn_parser::diagnostic::render_diagnostic(
2269                    &source,
2270                    path,
2271                    &span,
2272                    "error",
2273                    &harn_parser::diagnostic::parser_error_message(&err),
2274                    Some(harn_parser::diagnostic::parser_error_label(&err)),
2275                    harn_parser::diagnostic::parser_error_help(&err),
2276                );
2277                eprint!("{diagnostic}");
2278            } else {
2279                for e in parser.all_errors() {
2280                    let span = error_span_from_parse(e);
2281                    let diagnostic = harn_parser::diagnostic::render_diagnostic(
2282                        &source,
2283                        path,
2284                        &span,
2285                        "error",
2286                        &harn_parser::diagnostic::parser_error_message(e),
2287                        Some(harn_parser::diagnostic::parser_error_label(e)),
2288                        harn_parser::diagnostic::parser_error_help(e),
2289                    );
2290                    eprint!("{diagnostic}");
2291                }
2292            }
2293            process::exit(1);
2294        }
2295    };
2296
2297    (source, program)
2298}
2299
2300fn error_span_from_lex(e: &harn_lexer::LexerError) -> harn_lexer::Span {
2301    match e {
2302        harn_lexer::LexerError::UnexpectedCharacter(_, span)
2303        | harn_lexer::LexerError::UnterminatedString(span)
2304        | harn_lexer::LexerError::UnterminatedBlockComment(span) => *span,
2305    }
2306}
2307
2308fn error_span_from_parse(e: &harn_parser::ParserError) -> harn_lexer::Span {
2309    match e {
2310        harn_parser::ParserError::Unexpected { span, .. } => *span,
2311        harn_parser::ParserError::UnexpectedEof { span, .. } => *span,
2312    }
2313}
2314
2315/// Execute source code and return the output. Used by REPL and conformance tests.
2316pub(crate) async fn execute(source: &str, source_path: Option<&Path>) -> Result<String, String> {
2317    let mut lexer = Lexer::new(source);
2318    let tokens = lexer.tokenize().map_err(|e| e.to_string())?;
2319    let mut parser = Parser::new(tokens);
2320    let program = parser.parse().map_err(|e| e.to_string())?;
2321
2322    // Static cross-module resolution: when executed from a file, derive the
2323    // import graph so `execute` catches undefined calls at typecheck time.
2324    // The REPL / `-e` path invokes this without `source_path`, where there
2325    // is no importing file context; we fall back to no-imports checking.
2326    let mut checker = TypeChecker::new();
2327    if let Some(path) = source_path {
2328        let graph = harn_modules::build(&[path.to_path_buf()]);
2329        if let Some(imported) = graph.imported_names_for_file(path) {
2330            checker = checker.with_imported_names(imported);
2331        }
2332        if let Some(imported) = graph.imported_type_declarations_for_file(path) {
2333            checker = checker.with_imported_type_decls(imported);
2334        }
2335        if let Some(imported) = graph.imported_callable_declarations_for_file(path) {
2336            checker = checker.with_imported_callable_decls(imported);
2337        }
2338    }
2339    let type_diagnostics = checker.check(&program);
2340    let mut warning_lines = Vec::new();
2341    for diag in &type_diagnostics {
2342        match diag.severity {
2343            DiagnosticSeverity::Error => return Err(diag.message.clone()),
2344            DiagnosticSeverity::Warning => {
2345                warning_lines.push(format!("warning: {}", diag.message));
2346            }
2347        }
2348    }
2349
2350    let chunk = harn_vm::Compiler::new()
2351        .compile(&program)
2352        .map_err(|e| e.to_string())?;
2353
2354    let local = tokio::task::LocalSet::new();
2355    local
2356        .run_until(async {
2357            let mut vm = harn_vm::Vm::new();
2358            harn_vm::register_vm_stdlib(&mut vm);
2359            install_default_hostlib(&mut vm);
2360            let source_parent = source_path
2361                .and_then(|p| p.parent())
2362                .unwrap_or(std::path::Path::new("."));
2363            let project_root = harn_vm::stdlib::process::find_project_root(source_parent);
2364            let store_base = project_root.as_deref().unwrap_or(source_parent);
2365            let execution_cwd = std::env::current_dir()
2366                .unwrap_or_else(|_| std::path::PathBuf::from("."))
2367                .to_string_lossy()
2368                .into_owned();
2369            let source_dir = source_parent.to_string_lossy().into_owned();
2370            if source_path.is_some_and(is_conformance_path) {
2371                harn_vm::event_log::install_memory_for_current_thread(64);
2372            }
2373            harn_vm::register_store_builtins(&mut vm, store_base);
2374            harn_vm::register_metadata_builtins(&mut vm, store_base);
2375            let pipeline_name = source_path
2376                .and_then(|p| p.file_stem())
2377                .and_then(|s| s.to_str())
2378                .unwrap_or("default");
2379            harn_vm::register_checkpoint_builtins(&mut vm, store_base, pipeline_name);
2380            harn_vm::stdlib::process::set_thread_execution_context(Some(
2381                harn_vm::orchestration::RunExecutionRecord {
2382                    cwd: Some(execution_cwd),
2383                    source_dir: Some(source_dir),
2384                    env: std::collections::BTreeMap::new(),
2385                    adapter: None,
2386                    repo_path: None,
2387                    worktree_path: None,
2388                    branch: None,
2389                    base_ref: None,
2390                    cleanup: None,
2391                },
2392            ));
2393            if let Some(ref root) = project_root {
2394                vm.set_project_root(root);
2395            }
2396            if let Some(path) = source_path {
2397                if let Some(parent) = path.parent() {
2398                    if !parent.as_os_str().is_empty() {
2399                        vm.set_source_dir(parent);
2400                    }
2401                }
2402            }
2403            // Conformance tests land here via `run_conformance_tests`; for
2404            // `skill_fs_*` fixtures to see the bundled `skills/` folder
2405            // we run the same layered discovery as `harn run`.
2406            let loaded = skill_loader::load_skills(&skill_loader::SkillLoaderInputs {
2407                cli_dirs: Vec::new(),
2408                source_path: source_path.map(Path::to_path_buf),
2409            });
2410            skill_loader::emit_loader_warnings(&loaded.loader_warnings);
2411            skill_loader::install_skills_global(&mut vm, &loaded);
2412            if let Some(path) = source_path {
2413                let extensions = package::load_runtime_extensions(path);
2414                package::install_runtime_extensions(&extensions);
2415                package::install_manifest_triggers(&mut vm, &extensions)
2416                    .await
2417                    .map_err(|error| format!("failed to install manifest triggers: {error}"))?;
2418                package::install_manifest_hooks(&mut vm, &extensions)
2419                    .await
2420                    .map_err(|error| format!("failed to install manifest hooks: {error}"))?;
2421            }
2422            let _event_log = harn_vm::event_log::active_event_log()
2423                .unwrap_or_else(|| harn_vm::event_log::install_memory_for_current_thread(64));
2424            let connector_clients_installed =
2425                should_install_default_connector_clients(source, source_path);
2426            if connector_clients_installed {
2427                install_default_connector_clients(store_base)
2428                    .await
2429                    .map_err(|error| format!("failed to initialize connector clients: {error}"))?;
2430            }
2431            let execution_result = vm.execute(&chunk).await.map_err(|e| e.to_string());
2432            harn_vm::egress::reset_egress_policy_for_host();
2433            if connector_clients_installed {
2434                harn_vm::clear_active_connector_clients();
2435            }
2436            harn_vm::stdlib::process::set_thread_execution_context(None);
2437            execution_result?;
2438            let mut output = String::new();
2439            for wl in &warning_lines {
2440                output.push_str(wl);
2441                output.push('\n');
2442            }
2443            output.push_str(vm.output());
2444            Ok(output)
2445        })
2446        .await
2447}
2448
2449fn should_install_default_connector_clients(source: &str, source_path: Option<&Path>) -> bool {
2450    if !source_path.is_some_and(is_conformance_path) {
2451        return true;
2452    }
2453    source.contains("connector_call")
2454        || source.contains("std/connectors")
2455        || source.contains("connectors/")
2456}
2457
2458fn is_conformance_path(path: &Path) -> bool {
2459    path.components()
2460        .any(|component| component.as_os_str() == "conformance")
2461}
2462
2463async fn install_default_connector_clients(base_dir: &Path) -> Result<(), String> {
2464    let event_log = harn_vm::event_log::active_event_log()
2465        .unwrap_or_else(|| harn_vm::event_log::install_memory_for_current_thread(64));
2466    let secret_namespace = connector_secret_namespace(base_dir);
2467    let secrets: Arc<dyn harn_vm::secrets::SecretProvider> = Arc::new(
2468        harn_vm::secrets::configured_default_chain(secret_namespace)
2469            .map_err(|error| format!("failed to configure secret providers: {error}"))?,
2470    );
2471
2472    let registry = harn_vm::ConnectorRegistry::default();
2473    let metrics = Arc::new(harn_vm::MetricsRegistry::default());
2474    let inbox = Arc::new(
2475        harn_vm::InboxIndex::new(event_log.clone(), metrics.clone())
2476            .await
2477            .map_err(|error| error.to_string())?,
2478    );
2479    registry
2480        .init_all(harn_vm::ConnectorCtx {
2481            event_log,
2482            secrets,
2483            inbox,
2484            metrics,
2485            rate_limiter: Arc::new(harn_vm::RateLimiterFactory::default()),
2486        })
2487        .await
2488        .map_err(|error| error.to_string())?;
2489    let clients = registry.client_map().await;
2490    harn_vm::install_active_connector_clients(clients);
2491    Ok(())
2492}
2493
2494fn connector_secret_namespace(base_dir: &Path) -> String {
2495    match std::env::var("HARN_SECRET_NAMESPACE") {
2496        Ok(namespace) if !namespace.trim().is_empty() => namespace,
2497        _ => {
2498            let leaf = base_dir
2499                .file_name()
2500                .and_then(|name| name.to_str())
2501                .filter(|name| !name.is_empty())
2502                .unwrap_or("workspace");
2503            format!("harn/{leaf}")
2504        }
2505    }
2506}
2507
2508#[cfg(test)]
2509mod main_tests {
2510    use super::{normalize_serve_args, should_install_default_connector_clients};
2511    use std::path::Path;
2512
2513    #[test]
2514    fn normalize_serve_args_inserts_a2a_for_legacy_shape() {
2515        let args = normalize_serve_args(vec![
2516            "harn".to_string(),
2517            "serve".to_string(),
2518            "--port".to_string(),
2519            "3000".to_string(),
2520            "agent.harn".to_string(),
2521        ]);
2522        assert_eq!(
2523            args,
2524            vec![
2525                "harn".to_string(),
2526                "serve".to_string(),
2527                "a2a".to_string(),
2528                "--port".to_string(),
2529                "3000".to_string(),
2530                "agent.harn".to_string(),
2531            ]
2532        );
2533    }
2534
2535    #[test]
2536    fn normalize_serve_args_preserves_explicit_subcommands() {
2537        let args = normalize_serve_args(vec![
2538            "harn".to_string(),
2539            "serve".to_string(),
2540            "acp".to_string(),
2541            "server.harn".to_string(),
2542        ]);
2543        assert_eq!(
2544            args,
2545            vec![
2546                "harn".to_string(),
2547                "serve".to_string(),
2548                "acp".to_string(),
2549                "server.harn".to_string(),
2550            ]
2551        );
2552    }
2553
2554    #[test]
2555    fn conformance_skips_connector_clients_unless_fixture_uses_connectors() {
2556        let path = Path::new("conformance/tests/language/basic.harn");
2557        assert!(!should_install_default_connector_clients(
2558            "println(1)",
2559            Some(path)
2560        ));
2561        assert!(!should_install_default_connector_clients(
2562            "trust_graph_verify_chain()",
2563            Some(path)
2564        ));
2565        assert!(should_install_default_connector_clients(
2566            "import { post_message } from \"std/connectors/slack\"",
2567            Some(path)
2568        ));
2569        assert!(should_install_default_connector_clients(
2570            "println(1)",
2571            Some(Path::new("examples/demo.harn"))
2572        ));
2573    }
2574}