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::Local(args) => commands::local::run(args).await,
604        Command::Providers(args) => match args.command {
605            ProvidersCommand::Refresh(refresh) => {
606                if let Err(error) = commands::providers::run_refresh(&refresh).await {
607                    command_error(&error);
608                }
609            }
610            ProvidersCommand::Validate(validate) => {
611                if let Err(error) = commands::providers::run_validate(&validate) {
612                    command_error(&error);
613                }
614            }
615            ProvidersCommand::Export(export) => {
616                if let Err(error) = commands::providers::run_export(&export) {
617                    command_error(&error);
618                }
619            }
620        },
621        Command::Try(args) => commands::try_cmd::run(args).await,
622        Command::Quickstart(args) => {
623            if let Err(error) = commands::quickstart::run_quickstart(&args).await {
624                command_error(&error);
625            }
626        }
627        Command::Serve(args) => match args.command {
628            ServeCommand::Acp(args) => {
629                if let Err(error) = commands::serve::run_acp_server(&args).await {
630                    command_error(&error);
631                }
632            }
633            ServeCommand::A2a(args) => {
634                if let Err(error) = commands::serve::run_a2a_server(&args).await {
635                    command_error(&error);
636                }
637            }
638            ServeCommand::Api(args) => {
639                if let Err(error) = commands::serve::run_api_server(&args).await {
640                    command_error(&error);
641                }
642            }
643            ServeCommand::Mcp(args) => {
644                if let Err(error) = commands::serve::run_mcp_server(&args).await {
645                    command_error(&error);
646                }
647            }
648        },
649        Command::Connector(args) => {
650            if let Err(error) = commands::connector::handle_connector_command(args).await {
651                eprintln!("error: {error}");
652                process::exit(1);
653            }
654        }
655        Command::Mcp(args) => commands::mcp::handle_mcp_command(&args.command).await,
656        Command::Watch(args) => {
657            let denied =
658                commands::run::build_denied_builtins(args.deny.as_deref(), args.allow.as_deref());
659            commands::run::run_watch(&args.file, denied).await;
660        }
661        Command::Portal(args) => {
662            commands::portal::run_portal(
663                &args.dir,
664                args.manifest,
665                args.persona_state_dir,
666                &args.host,
667                args.port,
668                args.open,
669                args.allow_remote_launch,
670            )
671            .await
672        }
673        Command::Trigger(args) => {
674            if let Err(error) = commands::trigger::handle(args).await {
675                eprintln!("error: {error}");
676                process::exit(1);
677            }
678        }
679        Command::Flow(args) => match commands::flow::run_flow(&args) {
680            Ok(code) => {
681                if code != 0 {
682                    process::exit(code);
683                }
684            }
685            Err(error) => command_error(&error),
686        },
687        Command::Workflow(args) => match commands::workflow::handle(args) {
688            Ok(code) => {
689                if code != 0 {
690                    process::exit(code);
691                }
692            }
693            Err(error) => command_error(&error),
694        },
695        Command::Supervisor(args) => {
696            if let Err(error) = commands::supervisor::handle(args).await {
697                eprintln!("error: {error}");
698                process::exit(1);
699            }
700        }
701        Command::Trace(args) => {
702            if let Err(error) = commands::trace::handle(args).await {
703                eprintln!("error: {error}");
704                process::exit(1);
705            }
706        }
707        Command::Crystallize(args) => {
708            if let Err(error) = commands::crystallize::run(args) {
709                eprintln!("error: {error}");
710                process::exit(1);
711            }
712        }
713        Command::Trust(args) | Command::TrustGraph(args) => {
714            if let Err(error) = commands::trust::handle(args).await {
715                eprintln!("error: {error}");
716                process::exit(1);
717            }
718        }
719        Command::Verify(args) => {
720            if let Err(error) = verify_provenance_receipt(&args.receipt, args.json) {
721                eprintln!("error: {error}");
722                process::exit(1);
723            }
724        }
725        Command::Completions(args) => print_completions(args.shell),
726        Command::Orchestrator(args) => {
727            if let Err(error) = commands::orchestrator::handle(args).await {
728                eprintln!("error: {error}");
729                process::exit(1);
730            }
731        }
732        Command::Playground(args) => {
733            provider_bootstrap::maybe_seed_ollama_for_playground(
734                Path::new(&args.host),
735                Path::new(&args.script),
736                args.yes,
737                args.llm.is_some(),
738                args.llm_mock.is_some(),
739            )
740            .await;
741            let llm_mock_mode = if let Some(path) = args.llm_mock.as_ref() {
742                commands::run::CliLlmMockMode::Replay {
743                    fixture_path: PathBuf::from(path),
744                }
745            } else if let Some(path) = args.llm_mock_record.as_ref() {
746                commands::run::CliLlmMockMode::Record {
747                    fixture_path: PathBuf::from(path),
748                }
749            } else {
750                commands::run::CliLlmMockMode::Off
751            };
752            if let Err(error) = commands::playground::run_command(args, llm_mock_mode).await {
753                eprint!("{error}");
754                process::exit(1);
755            }
756        }
757        Command::Runs(args) => match args.command {
758            RunsCommand::Inspect(inspect) => {
759                inspect_run_record(&inspect.path, inspect.compare.as_deref())
760            }
761        },
762        Command::Session(args) => commands::session::run(args),
763        Command::Replay(args) => replay_run_record(&args.path),
764        Command::Eval(args) => {
765            let llm_mock_mode = if let Some(path) = args.llm_mock.as_ref() {
766                commands::run::CliLlmMockMode::Replay {
767                    fixture_path: PathBuf::from(path),
768                }
769            } else if let Some(path) = args.llm_mock_record.as_ref() {
770                commands::run::CliLlmMockMode::Record {
771                    fixture_path: PathBuf::from(path),
772                }
773            } else {
774                commands::run::CliLlmMockMode::Off
775            };
776            eval_run_record(
777                &args.path,
778                args.compare.as_deref(),
779                args.structural_experiment.as_deref(),
780                &args.argv,
781                &llm_mock_mode,
782            )
783        }
784        Command::Repl => commands::repl::run_repl().await,
785        Command::Bench(args) => commands::bench::run(args).await,
786        Command::TestBench(args) => commands::test_bench::run(args.command).await,
787        Command::Viz(args) => commands::viz::run_viz(&args.file, args.output.as_deref()),
788        Command::Install(args) => package::install_packages(
789            args.frozen || args.locked || args.offline,
790            args.refetch.as_deref(),
791            args.offline,
792            args.json,
793        ),
794        Command::Add(args) => package::add_package_with_registry(
795            &args.name_or_spec,
796            args.alias.as_deref(),
797            args.git.as_deref(),
798            args.tag.as_deref(),
799            args.rev.as_deref(),
800            args.branch.as_deref(),
801            args.path.as_deref(),
802            args.registry.as_deref(),
803        ),
804        Command::Update(args) => {
805            package::update_packages(args.alias.as_deref(), args.all, args.json)
806        }
807        Command::Remove(args) => package::remove_package(&args.alias),
808        Command::Lock => package::lock_packages(),
809        Command::Package(args) => match args.command {
810            PackageCommand::List(list) => package::list_packages(list.json),
811            PackageCommand::Doctor(doctor) => package::doctor_packages(doctor.json),
812            PackageCommand::Search(search) => package::search_package_registry(
813                search.query.as_deref(),
814                search.registry.as_deref(),
815                search.json,
816            ),
817            PackageCommand::Info(info) => {
818                package::show_package_registry_info(&info.name, info.registry.as_deref(), info.json)
819            }
820            PackageCommand::Check(check) => {
821                package::check_package(check.package.as_deref(), check.json)
822            }
823            PackageCommand::Pack(pack) => package::pack_package(
824                pack.package.as_deref(),
825                pack.output.as_deref(),
826                pack.dry_run,
827                pack.json,
828            ),
829            PackageCommand::Docs(docs) => package::generate_package_docs(
830                docs.package.as_deref(),
831                docs.output.as_deref(),
832                docs.check,
833            ),
834            PackageCommand::Cache(cache) => match cache.command {
835                PackageCacheCommand::List => package::list_package_cache(),
836                PackageCacheCommand::Clean(clean) => package::clean_package_cache(clean.all),
837                PackageCacheCommand::Verify(verify) => {
838                    package::verify_package_cache(verify.materialized)
839                }
840            },
841            PackageCommand::Outdated(args) => package::outdated_packages(
842                args.refresh,
843                args.remote,
844                args.registry.as_deref(),
845                args.json,
846            ),
847            PackageCommand::Audit(args) => {
848                package::audit_packages(args.registry.as_deref(), args.skip_materialized, args.json)
849            }
850            PackageCommand::Artifacts(args) => match args.command {
851                PackageArtifactsCommand::Manifest(manifest) => {
852                    package::artifacts_manifest(manifest.output.as_deref())
853                }
854                PackageArtifactsCommand::Check(check) => {
855                    package::artifacts_check(&check.manifest, check.json)
856                }
857            },
858        },
859        Command::Publish(args) => package::publish_package(
860            args.package.as_deref(),
861            args.dry_run,
862            args.registry.as_deref(),
863            args.json,
864        ),
865        Command::MergeCaptain(args) => match args.command {
866            MergeCaptainCommand::Run(run) => {
867                let code = commands::merge_captain::run_driver(&run);
868                if code != 0 {
869                    process::exit(code);
870                }
871            }
872            MergeCaptainCommand::Ladder(ladder) => {
873                let code = commands::merge_captain::run_ladder(&ladder);
874                if code != 0 {
875                    process::exit(code);
876                }
877            }
878            MergeCaptainCommand::Iterate(iterate) => {
879                let code = commands::merge_captain::run_iterate(&iterate);
880                if code != 0 {
881                    process::exit(code);
882                }
883            }
884            MergeCaptainCommand::Audit(audit) => {
885                let code = commands::merge_captain::run_audit(&audit);
886                if code != 0 {
887                    process::exit(code);
888                }
889            }
890            MergeCaptainCommand::Mock(mock) => {
891                let code = match mock {
892                    MergeCaptainMockCommand::Init(args) => {
893                        commands::merge_captain_mock::run_init(&args)
894                    }
895                    MergeCaptainMockCommand::Step(args) => {
896                        commands::merge_captain_mock::run_step(&args)
897                    }
898                    MergeCaptainMockCommand::Status(args) => {
899                        commands::merge_captain_mock::run_status(&args)
900                    }
901                    MergeCaptainMockCommand::Serve(args) => {
902                        commands::merge_captain_mock::run_serve(&args).await
903                    }
904                    MergeCaptainMockCommand::Cleanup(args) => {
905                        commands::merge_captain_mock::run_cleanup(&args)
906                    }
907                    MergeCaptainMockCommand::Scenarios => {
908                        commands::merge_captain_mock::run_scenarios()
909                    }
910                };
911                if code != 0 {
912                    process::exit(code);
913                }
914            }
915        },
916        Command::Persona(args) => match args.command {
917            PersonaCommand::New(new) => {
918                if let Err(error) = commands::persona_scaffold::run_new(&new) {
919                    eprintln!("error: {error}");
920                    process::exit(1);
921                }
922            }
923            PersonaCommand::Doctor(doctor) => {
924                if let Err(error) =
925                    commands::persona_doctor::run_doctor(args.manifest.as_deref(), &doctor).await
926                {
927                    eprintln!("error: {error}");
928                    process::exit(1);
929                }
930            }
931            PersonaCommand::Check(check) => {
932                commands::persona::run_check(args.manifest.as_deref(), &check)
933            }
934            PersonaCommand::List(list) => {
935                commands::persona::run_list(args.manifest.as_deref(), &list)
936            }
937            PersonaCommand::Inspect(inspect) => {
938                commands::persona::run_inspect(args.manifest.as_deref(), &inspect)
939            }
940            PersonaCommand::Status(status) => {
941                if let Err(error) = commands::persona::run_status(
942                    args.manifest.as_deref(),
943                    &args.state_dir,
944                    &status,
945                )
946                .await
947                {
948                    eprintln!("error: {error}");
949                    process::exit(1);
950                }
951            }
952            PersonaCommand::Pause(control) => {
953                if let Err(error) = commands::persona::run_pause(
954                    args.manifest.as_deref(),
955                    &args.state_dir,
956                    &control,
957                )
958                .await
959                {
960                    eprintln!("error: {error}");
961                    process::exit(1);
962                }
963            }
964            PersonaCommand::Resume(control) => {
965                if let Err(error) = commands::persona::run_resume(
966                    args.manifest.as_deref(),
967                    &args.state_dir,
968                    &control,
969                )
970                .await
971                {
972                    eprintln!("error: {error}");
973                    process::exit(1);
974                }
975            }
976            PersonaCommand::Disable(control) => {
977                if let Err(error) = commands::persona::run_disable(
978                    args.manifest.as_deref(),
979                    &args.state_dir,
980                    &control,
981                )
982                .await
983                {
984                    eprintln!("error: {error}");
985                    process::exit(1);
986                }
987            }
988            PersonaCommand::Tick(tick) => {
989                if let Err(error) =
990                    commands::persona::run_tick(args.manifest.as_deref(), &args.state_dir, &tick)
991                        .await
992                {
993                    eprintln!("error: {error}");
994                    process::exit(1);
995                }
996            }
997            PersonaCommand::Trigger(trigger) => {
998                if let Err(error) = commands::persona::run_trigger(
999                    args.manifest.as_deref(),
1000                    &args.state_dir,
1001                    &trigger,
1002                )
1003                .await
1004                {
1005                    eprintln!("error: {error}");
1006                    process::exit(1);
1007                }
1008            }
1009            PersonaCommand::Spend(spend) => {
1010                if let Err(error) =
1011                    commands::persona::run_spend(args.manifest.as_deref(), &args.state_dir, &spend)
1012                        .await
1013                {
1014                    eprintln!("error: {error}");
1015                    process::exit(1);
1016                }
1017            }
1018            PersonaCommand::Supervision(supervision) => match supervision.command {
1019                PersonaSupervisionCommand::Tail(tail) => {
1020                    if let Err(error) = commands::persona_supervision::run_tail(
1021                        args.manifest.as_deref(),
1022                        &args.state_dir,
1023                        &tail,
1024                    )
1025                    .await
1026                    {
1027                        eprintln!("error: {error}");
1028                        process::exit(1);
1029                    }
1030                }
1031            },
1032        },
1033        Command::ModelInfo(args) => {
1034            if !print_model_info(&args).await {
1035                process::exit(1);
1036            }
1037        }
1038        Command::ProviderCatalog(args) => print_provider_catalog(args.available_only),
1039        Command::ProviderReady(args) => {
1040            run_provider_ready(
1041                &args.provider,
1042                args.model.as_deref(),
1043                args.base_url.as_deref(),
1044                args.json,
1045            )
1046            .await
1047        }
1048        Command::Skills(args) => match args.command {
1049            SkillsCommand::List(list) => commands::skills::run_list(&list),
1050            SkillsCommand::Inspect(inspect) => commands::skills::run_inspect(&inspect),
1051            SkillsCommand::Match(matcher) => commands::skills::run_match(&matcher),
1052            SkillsCommand::Install(install) => commands::skills::run_install(&install),
1053            SkillsCommand::New(new_args) => commands::skills::run_new(&new_args),
1054        },
1055        Command::Tool(args) => match args.command {
1056            ToolCommand::New(new_args) => {
1057                if let Err(error) = commands::tool::run_new(&new_args) {
1058                    eprintln!("error: {error}");
1059                    process::exit(1);
1060                }
1061            }
1062        },
1063        Command::DumpHighlightKeywords(args) => {
1064            commands::dump_highlight_keywords::run(&args.output, args.check);
1065        }
1066        Command::DumpTriggerQuickref(args) => {
1067            commands::dump_trigger_quickref::run(&args.output, args.check);
1068        }
1069        Command::DumpConnectorMatrix(args) => {
1070            commands::check::connector_matrix::run_docs(&args.output, &args.sources, args.check);
1071        }
1072        Command::DumpProtocolArtifacts(args) => {
1073            commands::dump_protocol_artifacts::run(&args.output_dir, args.check);
1074        }
1075    }
1076}
1077
1078fn run_profile_options(args: &cli::ProfileArgs) -> commands::run::RunProfileOptions {
1079    commands::run::RunProfileOptions {
1080        text: args.text,
1081        json_path: args.json_path.clone(),
1082    }
1083}
1084
1085fn print_completions(shell: CompletionShell) {
1086    let mut command = Cli::command();
1087    let shell = clap_complete::Shell::from(shell);
1088    clap_complete::generate(shell, &mut command, "harn", &mut std::io::stdout());
1089}
1090
1091fn normalize_serve_args(mut raw_args: Vec<String>) -> Vec<String> {
1092    if raw_args.len() > 2
1093        && raw_args.get(1).is_some_and(|arg| arg == "serve")
1094        && !matches!(
1095            raw_args.get(2).map(String::as_str),
1096            Some("acp" | "a2a" | "api" | "mcp" | "-h" | "--help")
1097        )
1098    {
1099        raw_args.insert(2, "a2a".to_string());
1100    }
1101    raw_args
1102}
1103
1104fn print_version() {
1105    println!(
1106        r#"
1107 ╱▔▔╲
1108 ╱    ╲    harn v{}
1109 │ ◆  │    the agent harness language
1110 │    │
1111 ╰──╯╱
1112   ╱╱
1113"#,
1114        env!("CARGO_PKG_VERSION")
1115    );
1116}
1117
1118async fn print_model_info(args: &ModelInfoArgs) -> bool {
1119    let resolved = harn_vm::llm_config::resolve_model_info(&args.model);
1120    let api_key_result = harn_vm::llm::resolve_api_key(&resolved.provider);
1121    let api_key_set = api_key_result.is_ok();
1122    let api_key = api_key_result.unwrap_or_default();
1123    let context_window =
1124        harn_vm::llm::fetch_provider_max_context(&resolved.provider, &resolved.id, &api_key).await;
1125    let readiness = local_openai_readiness(&resolved.provider, &resolved.id, &api_key).await;
1126    let catalog = harn_vm::llm_config::model_catalog_entry(&resolved.id);
1127    let runtime_context_window = catalog
1128        .as_ref()
1129        .and_then(|entry| entry.runtime_context_window);
1130    let capabilities = harn_vm::llm::capabilities::lookup(&resolved.provider, &resolved.id);
1131    let mut payload = serde_json::json!({
1132        "alias": args.model,
1133        "id": resolved.id,
1134        "provider": resolved.provider,
1135        "resolved_alias": resolved.alias,
1136        "tool_format": resolved.tool_format,
1137        "tier": resolved.tier,
1138        "api_key_set": api_key_set,
1139        "context_window": context_window,
1140        "runtime_context_window": runtime_context_window,
1141        "readiness": readiness,
1142        "catalog": catalog,
1143        "capabilities": {
1144            "native_tools": capabilities.native_tools,
1145            "defer_loading": capabilities.defer_loading,
1146            "tool_search": capabilities.tool_search,
1147            "max_tools": capabilities.max_tools,
1148            "prompt_caching": capabilities.prompt_caching,
1149            "vision": capabilities.vision,
1150            "vision_supported": capabilities.vision_supported,
1151            "audio": capabilities.audio,
1152            "pdf": capabilities.pdf,
1153            "files_api_supported": capabilities.files_api_supported,
1154            "json_schema": capabilities.json_schema,
1155            "thinking": !capabilities.thinking_modes.is_empty(),
1156            "thinking_modes": capabilities.thinking_modes,
1157            "interleaved_thinking_supported": capabilities.interleaved_thinking_supported,
1158            "anthropic_beta_features": capabilities.anthropic_beta_features,
1159            "preserve_thinking": capabilities.preserve_thinking,
1160            "server_parser": capabilities.server_parser,
1161            "honors_chat_template_kwargs": capabilities.honors_chat_template_kwargs,
1162            "recommended_endpoint": capabilities.recommended_endpoint,
1163            "text_tool_wire_format_supported": capabilities.text_tool_wire_format_supported,
1164        },
1165        "qc_default_model": harn_vm::llm_config::qc_default_model(&resolved.provider),
1166    });
1167
1168    let should_verify = args.verify || args.warm;
1169    let mut ok = true;
1170    if should_verify {
1171        if resolved.provider == "ollama" {
1172            let mut readiness = harn_vm::llm::OllamaReadinessOptions::new(resolved.id.clone());
1173            readiness.warm = args.warm;
1174            readiness.observe_loaded = true;
1175            readiness.keep_alive = args
1176                .keep_alive
1177                .as_deref()
1178                .and_then(harn_vm::llm::normalize_ollama_keep_alive);
1179            let result = harn_vm::llm::ollama_readiness(readiness).await;
1180            ok = result.valid;
1181            payload["readiness"] = serde_json::to_value(&result).unwrap_or_else(|error| {
1182                serde_json::json!({
1183                    "valid": false,
1184                    "status": "serialization_error",
1185                    "message": format!("failed to serialize readiness result: {error}"),
1186                })
1187            });
1188        } else {
1189            ok = false;
1190            payload["readiness"] = serde_json::json!({
1191                "valid": false,
1192                "status": "unsupported_provider",
1193                "message": format!(
1194                    "model-info --verify is only supported for Ollama models; resolved provider is '{}'",
1195                    resolved.provider
1196                ),
1197                "provider": resolved.provider,
1198            });
1199        }
1200    }
1201
1202    println!(
1203        "{}",
1204        serde_json::to_string(&payload).unwrap_or_else(|error| {
1205            command_error(&format!("failed to serialize model info: {error}"))
1206        })
1207    );
1208    ok
1209}
1210
1211async fn local_openai_readiness(
1212    provider: &str,
1213    model: &str,
1214    api_key: &str,
1215) -> Option<serde_json::Value> {
1216    let def = harn_vm::llm_config::provider_config(provider)?;
1217    if def.auth_style != "none" || !harn_vm::llm::supports_model_readiness_probe(&def) {
1218        return None;
1219    }
1220    let readiness = harn_vm::llm::probe_openai_compatible_model(provider, model, api_key).await;
1221    Some(serde_json::json!({
1222        "valid": readiness.valid,
1223        "category": readiness.category,
1224        "message": readiness.message,
1225        "provider": readiness.provider,
1226        "model": readiness.model,
1227        "url": readiness.url,
1228        "status": readiness.status,
1229        "available_models": readiness.available_models,
1230    }))
1231}
1232
1233fn print_provider_catalog(available_only: bool) {
1234    let provider_names = if available_only {
1235        harn_vm::llm_config::available_provider_names()
1236    } else {
1237        harn_vm::llm_config::provider_names()
1238    };
1239    let providers: Vec<_> = provider_names
1240        .into_iter()
1241        .filter_map(|name| {
1242            harn_vm::llm_config::provider_config(&name).map(|def| {
1243                serde_json::json!({
1244                    "name": name,
1245                    "display_name": def.display_name,
1246                    "icon": def.icon,
1247                    "base_url": harn_vm::llm_config::resolve_base_url(&def),
1248                    "base_url_env": def.base_url_env,
1249                    "auth_style": def.auth_style,
1250                    "auth_envs": harn_vm::llm_config::auth_env_names(&def.auth_env),
1251                    "auth_available": harn_vm::llm_config::provider_key_available(&name),
1252                    "features": def.features,
1253                    "cost_per_1k_in": def.cost_per_1k_in,
1254                    "cost_per_1k_out": def.cost_per_1k_out,
1255                    "latency_p50_ms": def.latency_p50_ms,
1256                })
1257            })
1258        })
1259        .collect();
1260    let models: Vec<_> = harn_vm::llm_config::model_catalog_entries()
1261        .into_iter()
1262        .map(|(id, model)| {
1263            serde_json::json!({
1264                "id": id,
1265                "name": model.name,
1266                "provider": model.provider,
1267                "context_window": model.context_window,
1268                "runtime_context_window": model.runtime_context_window,
1269                "stream_timeout": model.stream_timeout,
1270                "capabilities": model.capabilities,
1271                "pricing": model.pricing,
1272            })
1273        })
1274        .collect();
1275    let aliases: Vec<_> = harn_vm::llm_config::alias_entries()
1276        .into_iter()
1277        .map(|(name, alias)| {
1278            serde_json::json!({
1279                "name": name,
1280                "id": alias.id,
1281                "provider": alias.provider,
1282                "tool_format": alias.tool_format,
1283            })
1284        })
1285        .collect();
1286    let payload = serde_json::json!({
1287        "providers": providers,
1288        "known_model_names": harn_vm::llm_config::known_model_names(),
1289        "available_providers": harn_vm::llm_config::available_provider_names(),
1290        "aliases": aliases,
1291        "models": models,
1292        "qc_defaults": harn_vm::llm_config::qc_defaults(),
1293    });
1294    println!(
1295        "{}",
1296        serde_json::to_string(&payload).unwrap_or_else(|error| {
1297            command_error(&format!("failed to serialize provider catalog: {error}"))
1298        })
1299    );
1300}
1301
1302async fn run_provider_ready(
1303    provider: &str,
1304    model: Option<&str>,
1305    base_url: Option<&str>,
1306    json: bool,
1307) {
1308    let readiness =
1309        harn_vm::llm::readiness::probe_provider_readiness(provider, model, base_url).await;
1310    if json {
1311        match serde_json::to_string_pretty(&readiness) {
1312            Ok(payload) => println!("{payload}"),
1313            Err(error) => command_error(&format!("failed to serialize readiness result: {error}")),
1314        }
1315    } else if readiness.ok {
1316        println!("{}", readiness.message);
1317    } else {
1318        eprintln!("{}", readiness.message);
1319    }
1320    if !readiness.ok {
1321        process::exit(1);
1322    }
1323}
1324
1325fn command_error(message: &str) -> ! {
1326    Cli::command()
1327        .error(ErrorKind::ValueValidation, message)
1328        .exit()
1329}
1330
1331fn verify_provenance_receipt(path: &str, json: bool) -> Result<(), String> {
1332    let raw =
1333        fs::read_to_string(path).map_err(|error| format!("failed to read {path}: {error}"))?;
1334    let receipt: harn_vm::ProvenanceReceipt = serde_json::from_str(&raw)
1335        .map_err(|error| format!("failed to parse provenance receipt {path}: {error}"))?;
1336    let report = harn_vm::verify_receipt(&receipt);
1337    if json {
1338        println!(
1339            "{}",
1340            serde_json::to_string_pretty(&report).map_err(|error| error.to_string())?
1341        );
1342    } else if report.verified {
1343        println!(
1344            "verified receipt={} events={} receipt_hash={} event_root_hash={}",
1345            report.receipt_id.unwrap_or_else(|| "-".to_string()),
1346            report.event_count,
1347            report.receipt_hash.unwrap_or_else(|| "-".to_string()),
1348            report.event_root_hash.unwrap_or_else(|| "-".to_string())
1349        );
1350    } else {
1351        println!(
1352            "failed receipt={} events={}",
1353            report.receipt_id.unwrap_or_else(|| "-".to_string()),
1354            report.event_count
1355        );
1356        for error in &report.errors {
1357            println!("  {error}");
1358        }
1359        return Err("provenance receipt verification failed".to_string());
1360    }
1361    Ok(())
1362}
1363
1364fn load_run_record_or_exit(path: &Path) -> harn_vm::orchestration::RunRecord {
1365    match harn_vm::orchestration::load_run_record(path) {
1366        Ok(run) => run,
1367        Err(error) => {
1368            eprintln!("Failed to load run record: {error}");
1369            process::exit(1);
1370        }
1371    }
1372}
1373
1374fn load_eval_suite_manifest_or_exit(path: &Path) -> harn_vm::orchestration::EvalSuiteManifest {
1375    harn_vm::orchestration::load_eval_suite_manifest(path).unwrap_or_else(|error| {
1376        eprintln!("Failed to load eval manifest {}: {error}", path.display());
1377        process::exit(1);
1378    })
1379}
1380
1381fn load_eval_pack_manifest_or_exit(path: &Path) -> harn_vm::orchestration::EvalPackManifest {
1382    harn_vm::orchestration::load_eval_pack_manifest(path).unwrap_or_else(|error| {
1383        eprintln!("Failed to load eval pack {}: {error}", path.display());
1384        process::exit(1);
1385    })
1386}
1387
1388fn load_persona_eval_ladder_manifest_or_exit(
1389    path: &Path,
1390) -> harn_vm::orchestration::PersonaEvalLadderManifest {
1391    harn_vm::orchestration::load_persona_eval_ladder_manifest(path).unwrap_or_else(|error| {
1392        eprintln!(
1393            "Failed to load persona eval ladder {}: {error}",
1394            path.display()
1395        );
1396        process::exit(1);
1397    })
1398}
1399
1400fn file_looks_like_eval_manifest(path: &Path) -> bool {
1401    if path.file_name().and_then(|name| name.to_str()) == Some("harn.eval.toml") {
1402        return true;
1403    }
1404    if path.extension().and_then(|ext| ext.to_str()) == Some("toml") {
1405        let Ok(content) = fs::read_to_string(path) else {
1406            return false;
1407        };
1408        return toml::from_str::<harn_vm::orchestration::EvalPackManifest>(&content)
1409            .is_ok_and(|manifest| !manifest.cases.is_empty() || !manifest.ladders.is_empty());
1410    }
1411    let Ok(content) = fs::read_to_string(path) else {
1412        return false;
1413    };
1414    let Ok(json) = serde_json::from_str::<serde_json::Value>(&content) else {
1415        return false;
1416    };
1417    json.get("_type").and_then(|value| value.as_str()) == Some("eval_suite_manifest")
1418        || json.get("cases").is_some()
1419}
1420
1421fn file_looks_like_eval_pack_manifest(path: &Path) -> bool {
1422    if path.file_name().and_then(|name| name.to_str()) == Some("harn.eval.toml") {
1423        return true;
1424    }
1425    if path.extension().and_then(|ext| ext.to_str()) == Some("toml") {
1426        return file_looks_like_eval_manifest(path);
1427    }
1428    let Ok(content) = fs::read_to_string(path) else {
1429        return false;
1430    };
1431    let Ok(json) = serde_json::from_str::<serde_json::Value>(&content) else {
1432        return false;
1433    };
1434    json.get("version").is_some()
1435        && (json.get("cases").is_some() || json.get("ladders").is_some())
1436        && json.get("_type").and_then(|value| value.as_str()) != Some("eval_suite_manifest")
1437}
1438
1439fn file_looks_like_persona_eval_ladder_manifest(path: &Path) -> bool {
1440    let Ok(content) = fs::read_to_string(path) else {
1441        return false;
1442    };
1443    if path.extension().and_then(|ext| ext.to_str()) == Some("json") {
1444        let Ok(json) = serde_json::from_str::<serde_json::Value>(&content) else {
1445            return false;
1446        };
1447        return json.get("_type").and_then(|value| value.as_str())
1448            == Some("persona_eval_ladder_manifest")
1449            || json.get("timeout_tiers").is_some()
1450            || json.get("timeout-tiers").is_some();
1451    }
1452    toml::from_str::<harn_vm::orchestration::PersonaEvalLadderManifest>(&content).is_ok_and(
1453        |manifest| {
1454            manifest
1455                .type_name
1456                .eq_ignore_ascii_case("persona_eval_ladder_manifest")
1457                || (!manifest.timeout_tiers.is_empty() && manifest.backend.path.is_some())
1458        },
1459    )
1460}
1461
1462fn collect_run_record_paths(path: &str) -> Vec<PathBuf> {
1463    let path = Path::new(path);
1464    if path.is_file() {
1465        return vec![path.to_path_buf()];
1466    }
1467    if path.is_dir() {
1468        let mut entries: Vec<PathBuf> = fs::read_dir(path)
1469            .unwrap_or_else(|error| {
1470                eprintln!("Failed to read run directory {}: {error}", path.display());
1471                process::exit(1);
1472            })
1473            .filter_map(|entry| entry.ok().map(|entry| entry.path()))
1474            .filter(|entry| entry.extension().and_then(|ext| ext.to_str()) == Some("json"))
1475            .collect();
1476        entries.sort();
1477        return entries;
1478    }
1479    eprintln!("Run path does not exist: {}", path.display());
1480    process::exit(1);
1481}
1482
1483fn print_run_diff(diff: &harn_vm::orchestration::RunDiffReport) {
1484    println!(
1485        "Diff: {} -> {} [{} -> {}]",
1486        diff.left_run_id, diff.right_run_id, diff.left_status, diff.right_status
1487    );
1488    println!("Identical: {}", diff.identical);
1489    println!("Stage diffs: {}", diff.stage_diffs.len());
1490    println!("Tool diffs: {}", diff.tool_diffs.len());
1491    println!("Observability diffs: {}", diff.observability_diffs.len());
1492    println!("Transition delta: {}", diff.transition_count_delta);
1493    println!("Artifact delta: {}", diff.artifact_count_delta);
1494    println!("Checkpoint delta: {}", diff.checkpoint_count_delta);
1495    for stage in &diff.stage_diffs {
1496        println!("- {} [{}]", stage.node_id, stage.change);
1497        for detail in &stage.details {
1498            println!("  {}", detail);
1499        }
1500    }
1501    for tool in &diff.tool_diffs {
1502        println!("- tool {} [{}]", tool.tool_name, tool.args_hash);
1503        println!("  left: {:?}", tool.left_result);
1504        println!("  right: {:?}", tool.right_result);
1505    }
1506    for item in &diff.observability_diffs {
1507        println!("- {} [{}]", item.label, item.section);
1508        for detail in &item.details {
1509            println!("  {}", detail);
1510        }
1511    }
1512}
1513
1514fn inspect_run_record(path: &str, compare: Option<&str>) {
1515    let run = load_run_record_or_exit(Path::new(path));
1516    println!("Run: {}", run.id);
1517    println!(
1518        "Workflow: {}",
1519        run.workflow_name
1520            .clone()
1521            .unwrap_or_else(|| run.workflow_id.clone())
1522    );
1523    println!("Status: {}", run.status);
1524    println!("Task: {}", run.task);
1525    println!("Stages: {}", run.stages.len());
1526    println!("Artifacts: {}", run.artifacts.len());
1527    println!("Transitions: {}", run.transitions.len());
1528    println!("Checkpoints: {}", run.checkpoints.len());
1529    println!("HITL questions: {}", run.hitl_questions.len());
1530    if let Some(observability) = &run.observability {
1531        println!("Planner rounds: {}", observability.planner_rounds.len());
1532        println!("Research facts: {}", observability.research_fact_count);
1533        println!("Workers: {}", observability.worker_lineage.len());
1534        println!(
1535            "Action graph: {} nodes / {} edges",
1536            observability.action_graph_nodes.len(),
1537            observability.action_graph_edges.len()
1538        );
1539        println!(
1540            "Transcript pointers: {}",
1541            observability.transcript_pointers.len()
1542        );
1543        println!("Daemon events: {}", observability.daemon_events.len());
1544    }
1545    if let Some(parent_worker_id) = run
1546        .metadata
1547        .get("parent_worker_id")
1548        .and_then(|value| value.as_str())
1549    {
1550        println!("Parent worker: {}", parent_worker_id);
1551    }
1552    if let Some(parent_stage_id) = run
1553        .metadata
1554        .get("parent_stage_id")
1555        .and_then(|value| value.as_str())
1556    {
1557        println!("Parent stage: {}", parent_stage_id);
1558    }
1559    if run
1560        .metadata
1561        .get("delegated")
1562        .and_then(|value| value.as_bool())
1563        .unwrap_or(false)
1564    {
1565        println!("Delegated: true");
1566    }
1567    println!(
1568        "Pending nodes: {}",
1569        if run.pending_nodes.is_empty() {
1570            "-".to_string()
1571        } else {
1572            run.pending_nodes.join(", ")
1573        }
1574    );
1575    println!(
1576        "Replay fixture: {}",
1577        if run.replay_fixture.is_some() {
1578            "embedded"
1579        } else {
1580            "derived"
1581        }
1582    );
1583    for stage in &run.stages {
1584        let worker = stage.metadata.get("worker");
1585        let worker_suffix = worker
1586            .and_then(|value| value.get("name"))
1587            .and_then(|value| value.as_str())
1588            .map(|name| format!(" worker={name}"))
1589            .unwrap_or_default();
1590        println!(
1591            "- {} [{}] status={} outcome={} branch={}{}",
1592            stage.node_id,
1593            stage.kind,
1594            stage.status,
1595            stage.outcome,
1596            stage.branch.clone().unwrap_or_else(|| "-".to_string()),
1597            worker_suffix,
1598        );
1599        if let Some(worker) = worker {
1600            if let Some(worker_id) = worker.get("id").and_then(|value| value.as_str()) {
1601                println!("  worker_id: {}", worker_id);
1602            }
1603            if let Some(child_run_id) = worker.get("child_run_id").and_then(|value| value.as_str())
1604            {
1605                println!("  child_run_id: {}", child_run_id);
1606            }
1607            if let Some(child_run_path) = worker
1608                .get("child_run_path")
1609                .and_then(|value| value.as_str())
1610            {
1611                println!("  child_run_path: {}", child_run_path);
1612            }
1613        }
1614    }
1615    if let Some(observability) = &run.observability {
1616        for round in &observability.planner_rounds {
1617            println!(
1618                "- planner {} iterations={} llm_calls={} tools={} research_facts={}",
1619                round.node_id,
1620                round.iteration_count,
1621                round.llm_call_count,
1622                round.tool_execution_count,
1623                round.research_facts.len()
1624            );
1625        }
1626        for pointer in &observability.transcript_pointers {
1627            println!(
1628                "- transcript {} [{}] available={} {}",
1629                pointer.label,
1630                pointer.kind,
1631                pointer.available,
1632                pointer
1633                    .path
1634                    .clone()
1635                    .unwrap_or_else(|| pointer.location.clone())
1636            );
1637        }
1638        for event in &observability.daemon_events {
1639            println!(
1640                "- daemon {} [{:?}] at {}",
1641                event.name, event.kind, event.timestamp
1642            );
1643            println!("  id: {}", event.daemon_id);
1644            println!("  persist_path: {}", event.persist_path);
1645            if let Some(summary) = &event.payload_summary {
1646                println!("  payload: {}", summary);
1647            }
1648        }
1649    }
1650    if let Some(compare_path) = compare {
1651        let baseline = load_run_record_or_exit(Path::new(compare_path));
1652        print_run_diff(&harn_vm::orchestration::diff_run_records(&baseline, &run));
1653    }
1654}
1655
1656fn replay_run_record(path: &str) {
1657    let run = load_run_record_or_exit(Path::new(path));
1658    println!("Replay: {}", run.id);
1659    for stage in &run.stages {
1660        println!(
1661            "[{}] status={} outcome={} branch={}",
1662            stage.node_id,
1663            stage.status,
1664            stage.outcome,
1665            stage.branch.clone().unwrap_or_else(|| "-".to_string())
1666        );
1667        if let Some(text) = &stage.visible_text {
1668            println!("  visible: {}", text);
1669        }
1670        if let Some(verification) = &stage.verification {
1671            println!("  verification: {}", verification);
1672        }
1673    }
1674    if let Some(transcript) = &run.transcript {
1675        println!(
1676            "Transcript events persisted: {}",
1677            transcript["events"]
1678                .as_array()
1679                .map(|v| v.len())
1680                .unwrap_or(0)
1681        );
1682    }
1683    let fixture = run
1684        .replay_fixture
1685        .clone()
1686        .unwrap_or_else(|| harn_vm::orchestration::replay_fixture_from_run(&run));
1687    let report = harn_vm::orchestration::evaluate_run_against_fixture(&run, &fixture);
1688    println!(
1689        "Embedded replay fixture: {}",
1690        if report.pass { "PASS" } else { "FAIL" }
1691    );
1692    for transition in &run.transitions {
1693        println!(
1694            "transition {} -> {} ({})",
1695            transition
1696                .from_node_id
1697                .clone()
1698                .unwrap_or_else(|| "start".to_string()),
1699            transition.to_node_id,
1700            transition
1701                .branch
1702                .clone()
1703                .unwrap_or_else(|| "default".to_string())
1704        );
1705    }
1706}
1707
1708fn eval_run_record(
1709    path: &str,
1710    compare: Option<&str>,
1711    structural_experiment: Option<&str>,
1712    argv: &[String],
1713    llm_mock_mode: &commands::run::CliLlmMockMode,
1714) {
1715    if let Some(experiment) = structural_experiment {
1716        let path_buf = PathBuf::from(path);
1717        if !path_buf.is_file() || path_buf.extension().and_then(|ext| ext.to_str()) != Some("harn")
1718        {
1719            eprintln!(
1720                "--structural-experiment currently requires a .harn pipeline path, got {}",
1721                path
1722            );
1723            process::exit(1);
1724        }
1725        if compare.is_some() {
1726            eprintln!("--compare cannot be combined with --structural-experiment");
1727            process::exit(1);
1728        }
1729        if matches!(llm_mock_mode, commands::run::CliLlmMockMode::Record { .. }) {
1730            eprintln!("--llm-mock-record cannot be combined with --structural-experiment");
1731            process::exit(1);
1732        }
1733        let path_buf = fs::canonicalize(&path_buf).unwrap_or_else(|error| {
1734            command_error(&format!(
1735                "failed to canonicalize structural eval pipeline {}: {error}",
1736                path_buf.display()
1737            ))
1738        });
1739        run_structural_experiment_eval(&path_buf, experiment, argv, llm_mock_mode);
1740        return;
1741    }
1742
1743    let path_buf = PathBuf::from(path);
1744    if path_buf.is_file() && file_looks_like_persona_eval_ladder_manifest(&path_buf) {
1745        if compare.is_some() {
1746            eprintln!("--compare is not supported with persona eval ladder manifests");
1747            process::exit(1);
1748        }
1749        let manifest = load_persona_eval_ladder_manifest_or_exit(&path_buf);
1750        let report =
1751            harn_vm::orchestration::run_persona_eval_ladder(&manifest).unwrap_or_else(|error| {
1752                eprintln!(
1753                    "Failed to evaluate persona eval ladder {}: {error}",
1754                    path_buf.display()
1755                );
1756                process::exit(1);
1757            });
1758        print_persona_ladder_report(&report);
1759        if !report.pass {
1760            process::exit(1);
1761        }
1762        return;
1763    }
1764
1765    if path_buf.is_file() && file_looks_like_eval_pack_manifest(&path_buf) {
1766        if compare.is_some() {
1767            eprintln!("--compare is not supported with eval pack manifests");
1768            process::exit(1);
1769        }
1770        let manifest = load_eval_pack_manifest_or_exit(&path_buf);
1771        let report = harn_vm::orchestration::evaluate_eval_pack_manifest(&manifest).unwrap_or_else(
1772            |error| {
1773                eprintln!(
1774                    "Failed to evaluate eval pack {}: {error}",
1775                    path_buf.display()
1776                );
1777                process::exit(1);
1778            },
1779        );
1780        print_eval_pack_report(&report);
1781        if !report.pass {
1782            process::exit(1);
1783        }
1784        return;
1785    }
1786
1787    if path_buf.is_file() && file_looks_like_eval_manifest(&path_buf) {
1788        if compare.is_some() {
1789            eprintln!("--compare is not supported with eval suite manifests");
1790            process::exit(1);
1791        }
1792        let manifest = load_eval_suite_manifest_or_exit(&path_buf);
1793        let suite = harn_vm::orchestration::evaluate_run_suite_manifest(&manifest).unwrap_or_else(
1794            |error| {
1795                eprintln!(
1796                    "Failed to evaluate manifest {}: {error}",
1797                    path_buf.display()
1798                );
1799                process::exit(1);
1800            },
1801        );
1802        println!(
1803            "{} {} passed, {} failed, {} total",
1804            if suite.pass { "PASS" } else { "FAIL" },
1805            suite.passed,
1806            suite.failed,
1807            suite.total
1808        );
1809        for case in &suite.cases {
1810            println!(
1811                "- {} [{}] {}",
1812                case.label.clone().unwrap_or_else(|| case.run_id.clone()),
1813                case.workflow_id,
1814                if case.pass { "PASS" } else { "FAIL" }
1815            );
1816            if let Some(path) = &case.source_path {
1817                println!("  path: {}", path);
1818            }
1819            if let Some(comparison) = &case.comparison {
1820                println!("  baseline identical: {}", comparison.identical);
1821                if !comparison.identical {
1822                    println!(
1823                        "  baseline status: {} -> {}",
1824                        comparison.left_status, comparison.right_status
1825                    );
1826                }
1827            }
1828            for failure in &case.failures {
1829                println!("  {}", failure);
1830            }
1831        }
1832        if !suite.pass {
1833            process::exit(1);
1834        }
1835        return;
1836    }
1837
1838    let paths = collect_run_record_paths(path);
1839    if paths.len() > 1 {
1840        let mut cases = Vec::new();
1841        for path in &paths {
1842            let run = load_run_record_or_exit(path);
1843            let fixture = run
1844                .replay_fixture
1845                .clone()
1846                .unwrap_or_else(|| harn_vm::orchestration::replay_fixture_from_run(&run));
1847            cases.push((run, fixture, Some(path.display().to_string())));
1848        }
1849        let suite = harn_vm::orchestration::evaluate_run_suite(cases);
1850        println!(
1851            "{} {} passed, {} failed, {} total",
1852            if suite.pass { "PASS" } else { "FAIL" },
1853            suite.passed,
1854            suite.failed,
1855            suite.total
1856        );
1857        for case in &suite.cases {
1858            println!(
1859                "- {} [{}] {}",
1860                case.run_id,
1861                case.workflow_id,
1862                if case.pass { "PASS" } else { "FAIL" }
1863            );
1864            if let Some(path) = &case.source_path {
1865                println!("  path: {}", path);
1866            }
1867            if let Some(comparison) = &case.comparison {
1868                println!("  baseline identical: {}", comparison.identical);
1869            }
1870            for failure in &case.failures {
1871                println!("  {}", failure);
1872            }
1873        }
1874        if !suite.pass {
1875            process::exit(1);
1876        }
1877        return;
1878    }
1879
1880    let run = load_run_record_or_exit(&paths[0]);
1881    let fixture = run
1882        .replay_fixture
1883        .clone()
1884        .unwrap_or_else(|| harn_vm::orchestration::replay_fixture_from_run(&run));
1885    let report = harn_vm::orchestration::evaluate_run_against_fixture(&run, &fixture);
1886    println!("{}", if report.pass { "PASS" } else { "FAIL" });
1887    println!("Stages: {}", report.stage_count);
1888    if let Some(compare_path) = compare {
1889        let baseline = load_run_record_or_exit(Path::new(compare_path));
1890        print_run_diff(&harn_vm::orchestration::diff_run_records(&baseline, &run));
1891    }
1892    if !report.failures.is_empty() {
1893        for failure in &report.failures {
1894            println!("- {}", failure);
1895        }
1896    }
1897    if !report.pass {
1898        process::exit(1);
1899    }
1900}
1901
1902fn print_eval_pack_report(report: &harn_vm::orchestration::EvalPackReport) {
1903    println!(
1904        "{} {} passed, {} blocking failed, {} warning, {} informational, {} total",
1905        if report.pass { "PASS" } else { "FAIL" },
1906        report.passed,
1907        report.blocking_failed,
1908        report.warning_failed,
1909        report.informational_failed,
1910        report.total
1911    );
1912    for case in &report.cases {
1913        println!(
1914            "- {} [{}] {} ({})",
1915            case.label,
1916            case.workflow_id,
1917            if case.pass { "PASS" } else { "FAIL" },
1918            case.severity
1919        );
1920        if let Some(path) = &case.source_path {
1921            println!("  path: {}", path);
1922        }
1923        if let Some(comparison) = &case.comparison {
1924            println!("  baseline identical: {}", comparison.identical);
1925            if !comparison.identical {
1926                println!(
1927                    "  baseline status: {} -> {}",
1928                    comparison.left_status, comparison.right_status
1929                );
1930            }
1931        }
1932        for failure in &case.failures {
1933            println!("  {}", failure);
1934        }
1935        for warning in &case.warnings {
1936            println!("  warning: {}", warning);
1937        }
1938        for item in &case.informational {
1939            println!("  info: {}", item);
1940        }
1941    }
1942    for ladder in &report.ladders {
1943        println!(
1944            "- ladder {} [{}] {} ({}) first_correct={}/{}",
1945            ladder.id,
1946            ladder.persona,
1947            if ladder.pass { "PASS" } else { "FAIL" },
1948            ladder.severity,
1949            ladder.first_correct_route.as_deref().unwrap_or("<none>"),
1950            ladder.first_correct_tier.as_deref().unwrap_or("<none>")
1951        );
1952        println!("  artifacts: {}", ladder.artifact_root);
1953        for tier in &ladder.tiers {
1954            println!(
1955                "  - {} [{}] {} tools={} models={} latency={}ms cost=${:.6}",
1956                tier.timeout_tier,
1957                tier.route_id,
1958                tier.outcome,
1959                tier.tool_calls,
1960                tier.model_calls,
1961                tier.latency_ms,
1962                tier.cost_usd
1963            );
1964            for reason in &tier.degradation_reasons {
1965                println!("    {}", reason);
1966            }
1967        }
1968    }
1969}
1970
1971fn print_persona_ladder_report(report: &harn_vm::orchestration::PersonaEvalLadderReport) {
1972    println!(
1973        "{} ladder {} passed, {} degraded/looped, {} total",
1974        if report.pass { "PASS" } else { "FAIL" },
1975        report.passed,
1976        report.failed,
1977        report.total
1978    );
1979    println!(
1980        "first_correct: {}/{}",
1981        report.first_correct_route.as_deref().unwrap_or("<none>"),
1982        report.first_correct_tier.as_deref().unwrap_or("<none>")
1983    );
1984    println!("artifacts: {}", report.artifact_root);
1985    for tier in &report.tiers {
1986        println!(
1987            "- {} [{}] {} tools={} models={} latency={}ms cost=${:.6}",
1988            tier.timeout_tier,
1989            tier.route_id,
1990            tier.outcome,
1991            tier.tool_calls,
1992            tier.model_calls,
1993            tier.latency_ms,
1994            tier.cost_usd
1995        );
1996        for reason in &tier.degradation_reasons {
1997            println!("  {}", reason);
1998        }
1999    }
2000}
2001
2002fn run_package_evals() {
2003    let paths = package::load_package_eval_pack_paths(None).unwrap_or_else(|error| {
2004        eprintln!("{error}");
2005        process::exit(1);
2006    });
2007    let mut all_pass = true;
2008    for path in &paths {
2009        println!("Eval pack: {}", path.display());
2010        let manifest = load_eval_pack_manifest_or_exit(path);
2011        let report = harn_vm::orchestration::evaluate_eval_pack_manifest(&manifest).unwrap_or_else(
2012            |error| {
2013                eprintln!("Failed to evaluate eval pack {}: {error}", path.display());
2014                process::exit(1);
2015            },
2016        );
2017        print_eval_pack_report(&report);
2018        all_pass &= report.pass;
2019    }
2020    if !all_pass {
2021        process::exit(1);
2022    }
2023}
2024
2025fn run_structural_experiment_eval(
2026    path: &Path,
2027    experiment: &str,
2028    argv: &[String],
2029    llm_mock_mode: &commands::run::CliLlmMockMode,
2030) {
2031    let baseline_dir = tempfile::Builder::new()
2032        .prefix("harn-eval-baseline-")
2033        .tempdir()
2034        .unwrap_or_else(|error| {
2035            command_error(&format!("failed to create baseline tempdir: {error}"))
2036        });
2037    let variant_dir = tempfile::Builder::new()
2038        .prefix("harn-eval-variant-")
2039        .tempdir()
2040        .unwrap_or_else(|error| {
2041            command_error(&format!("failed to create variant tempdir: {error}"))
2042        });
2043
2044    let baseline = spawn_eval_pipeline_run(path, baseline_dir.path(), None, argv, llm_mock_mode);
2045    if !baseline.status.success() {
2046        relay_subprocess_failure("baseline", &baseline);
2047    }
2048
2049    let variant = spawn_eval_pipeline_run(
2050        path,
2051        variant_dir.path(),
2052        Some(experiment),
2053        argv,
2054        llm_mock_mode,
2055    );
2056    if !variant.status.success() {
2057        relay_subprocess_failure("variant", &variant);
2058    }
2059
2060    let baseline_runs = collect_structural_eval_runs(baseline_dir.path());
2061    let variant_runs = collect_structural_eval_runs(variant_dir.path());
2062    if baseline_runs.is_empty() || variant_runs.is_empty() {
2063        eprintln!(
2064            "structural eval expected workflow run records under {} and {}, but one side was empty",
2065            baseline_dir.path().display(),
2066            variant_dir.path().display()
2067        );
2068        process::exit(1);
2069    }
2070    if baseline_runs.len() != variant_runs.len() {
2071        eprintln!(
2072            "structural eval produced different run counts: baseline={} variant={}",
2073            baseline_runs.len(),
2074            variant_runs.len()
2075        );
2076        process::exit(1);
2077    }
2078
2079    let mut baseline_ok = 0usize;
2080    let mut variant_ok = 0usize;
2081    let mut any_failures = false;
2082
2083    println!("Structural experiment: {}", experiment);
2084    println!("Cases: {}", baseline_runs.len());
2085    for (baseline_run, variant_run) in baseline_runs.iter().zip(variant_runs.iter()) {
2086        let baseline_fixture = baseline_run
2087            .replay_fixture
2088            .clone()
2089            .unwrap_or_else(|| harn_vm::orchestration::replay_fixture_from_run(baseline_run));
2090        let variant_fixture = variant_run
2091            .replay_fixture
2092            .clone()
2093            .unwrap_or_else(|| harn_vm::orchestration::replay_fixture_from_run(variant_run));
2094        let baseline_report =
2095            harn_vm::orchestration::evaluate_run_against_fixture(baseline_run, &baseline_fixture);
2096        let variant_report =
2097            harn_vm::orchestration::evaluate_run_against_fixture(variant_run, &variant_fixture);
2098        let diff = harn_vm::orchestration::diff_run_records(baseline_run, variant_run);
2099        if baseline_report.pass {
2100            baseline_ok += 1;
2101        }
2102        if variant_report.pass {
2103            variant_ok += 1;
2104        }
2105        any_failures |= !baseline_report.pass || !variant_report.pass;
2106        println!(
2107            "- {} [{}]",
2108            variant_run
2109                .workflow_name
2110                .clone()
2111                .unwrap_or_else(|| variant_run.workflow_id.clone()),
2112            variant_run.task
2113        );
2114        println!(
2115            "  baseline: {}",
2116            if baseline_report.pass { "PASS" } else { "FAIL" }
2117        );
2118        for failure in &baseline_report.failures {
2119            println!("    {}", failure);
2120        }
2121        println!(
2122            "  variant: {}",
2123            if variant_report.pass { "PASS" } else { "FAIL" }
2124        );
2125        for failure in &variant_report.failures {
2126            println!("    {}", failure);
2127        }
2128        println!("  diff identical: {}", diff.identical);
2129        println!("  stage diffs: {}", diff.stage_diffs.len());
2130        println!("  tool diffs: {}", diff.tool_diffs.len());
2131        println!("  observability diffs: {}", diff.observability_diffs.len());
2132    }
2133
2134    println!("Baseline {} / {} passed", baseline_ok, baseline_runs.len());
2135    println!("Variant {} / {} passed", variant_ok, variant_runs.len());
2136
2137    if any_failures {
2138        process::exit(1);
2139    }
2140}
2141
2142fn spawn_eval_pipeline_run(
2143    path: &Path,
2144    run_dir: &Path,
2145    structural_experiment: Option<&str>,
2146    argv: &[String],
2147    llm_mock_mode: &commands::run::CliLlmMockMode,
2148) -> std::process::Output {
2149    let exe = env::current_exe().unwrap_or_else(|error| {
2150        command_error(&format!("failed to resolve current executable: {error}"))
2151    });
2152    let mut command = std::process::Command::new(exe);
2153    command.current_dir(path.parent().unwrap_or_else(|| Path::new(".")));
2154    command.arg("run");
2155    match llm_mock_mode {
2156        commands::run::CliLlmMockMode::Off => {}
2157        commands::run::CliLlmMockMode::Replay { fixture_path } => {
2158            command
2159                .arg("--llm-mock")
2160                .arg(absolute_cli_path(fixture_path));
2161        }
2162        commands::run::CliLlmMockMode::Record { fixture_path } => {
2163            command
2164                .arg("--llm-mock-record")
2165                .arg(absolute_cli_path(fixture_path));
2166        }
2167    }
2168    command.arg(path);
2169    if !argv.is_empty() {
2170        command.arg("--");
2171        command.args(argv);
2172    }
2173    command.env(harn_vm::runtime_paths::HARN_RUN_DIR_ENV, run_dir);
2174    if let Some(experiment) = structural_experiment {
2175        command.env("HARN_STRUCTURAL_EXPERIMENT", experiment);
2176    }
2177    command.output().unwrap_or_else(|error| {
2178        command_error(&format!(
2179            "failed to spawn `harn run {}` for structural eval: {error}",
2180            path.display()
2181        ))
2182    })
2183}
2184
2185fn absolute_cli_path(path: &Path) -> PathBuf {
2186    if path.is_absolute() {
2187        return path.to_path_buf();
2188    }
2189    env::current_dir()
2190        .unwrap_or_else(|_| PathBuf::from("."))
2191        .join(path)
2192}
2193
2194fn relay_subprocess_failure(label: &str, output: &std::process::Output) -> ! {
2195    let stdout = String::from_utf8_lossy(&output.stdout);
2196    let stderr = String::from_utf8_lossy(&output.stderr);
2197    if !stdout.trim().is_empty() {
2198        eprintln!("[{label}] stdout:\n{stdout}");
2199    }
2200    if !stderr.trim().is_empty() {
2201        eprintln!("[{label}] stderr:\n{stderr}");
2202    }
2203    process::exit(output.status.code().unwrap_or(1));
2204}
2205
2206fn collect_structural_eval_runs(dir: &Path) -> Vec<harn_vm::orchestration::RunRecord> {
2207    let mut paths: Vec<PathBuf> = fs::read_dir(dir)
2208        .unwrap_or_else(|error| {
2209            command_error(&format!(
2210                "failed to read structural eval run dir {}: {error}",
2211                dir.display()
2212            ))
2213        })
2214        .filter_map(|entry| entry.ok().map(|entry| entry.path()))
2215        .filter(|entry| entry.extension().and_then(|ext| ext.to_str()) == Some("json"))
2216        .collect();
2217    paths.sort();
2218    let mut runs: Vec<_> = paths
2219        .iter()
2220        .map(|path| load_run_record_or_exit(path))
2221        .collect();
2222    runs.sort_by(|left, right| {
2223        (
2224            left.started_at.as_str(),
2225            left.workflow_id.as_str(),
2226            left.task.as_str(),
2227        )
2228            .cmp(&(
2229                right.started_at.as_str(),
2230                right.workflow_id.as_str(),
2231                right.task.as_str(),
2232            ))
2233    });
2234    runs
2235}
2236
2237/// Parse a .harn file, returning (source, AST). Exits on error.
2238pub(crate) fn parse_source_file(path: &str) -> (String, Vec<harn_parser::SNode>) {
2239    let source = match fs::read_to_string(path) {
2240        Ok(s) => s,
2241        Err(e) => {
2242            eprintln!("Error reading {path}: {e}");
2243            process::exit(1);
2244        }
2245    };
2246
2247    let mut lexer = Lexer::new(&source);
2248    let tokens = match lexer.tokenize() {
2249        Ok(t) => t,
2250        Err(e) => {
2251            let diagnostic = harn_parser::diagnostic::render_diagnostic(
2252                &source,
2253                path,
2254                &error_span_from_lex(&e),
2255                "error",
2256                &e.to_string(),
2257                Some("here"),
2258                None,
2259            );
2260            eprint!("{diagnostic}");
2261            process::exit(1);
2262        }
2263    };
2264
2265    let mut parser = Parser::new(tokens);
2266    let program = match parser.parse() {
2267        Ok(p) => p,
2268        Err(err) => {
2269            if parser.all_errors().is_empty() {
2270                let span = error_span_from_parse(&err);
2271                let diagnostic = harn_parser::diagnostic::render_diagnostic(
2272                    &source,
2273                    path,
2274                    &span,
2275                    "error",
2276                    &harn_parser::diagnostic::parser_error_message(&err),
2277                    Some(harn_parser::diagnostic::parser_error_label(&err)),
2278                    harn_parser::diagnostic::parser_error_help(&err),
2279                );
2280                eprint!("{diagnostic}");
2281            } else {
2282                for e in parser.all_errors() {
2283                    let span = error_span_from_parse(e);
2284                    let diagnostic = harn_parser::diagnostic::render_diagnostic(
2285                        &source,
2286                        path,
2287                        &span,
2288                        "error",
2289                        &harn_parser::diagnostic::parser_error_message(e),
2290                        Some(harn_parser::diagnostic::parser_error_label(e)),
2291                        harn_parser::diagnostic::parser_error_help(e),
2292                    );
2293                    eprint!("{diagnostic}");
2294                }
2295            }
2296            process::exit(1);
2297        }
2298    };
2299
2300    (source, program)
2301}
2302
2303fn error_span_from_lex(e: &harn_lexer::LexerError) -> harn_lexer::Span {
2304    match e {
2305        harn_lexer::LexerError::UnexpectedCharacter(_, span)
2306        | harn_lexer::LexerError::UnterminatedString(span)
2307        | harn_lexer::LexerError::UnterminatedBlockComment(span) => *span,
2308    }
2309}
2310
2311fn error_span_from_parse(e: &harn_parser::ParserError) -> harn_lexer::Span {
2312    match e {
2313        harn_parser::ParserError::Unexpected { span, .. } => *span,
2314        harn_parser::ParserError::UnexpectedEof { span, .. } => *span,
2315    }
2316}
2317
2318/// Execute source code and return the output. Used by REPL and conformance tests.
2319pub(crate) async fn execute(source: &str, source_path: Option<&Path>) -> Result<String, String> {
2320    let mut lexer = Lexer::new(source);
2321    let tokens = lexer.tokenize().map_err(|e| e.to_string())?;
2322    let mut parser = Parser::new(tokens);
2323    let program = parser.parse().map_err(|e| e.to_string())?;
2324
2325    // Static cross-module resolution: when executed from a file, derive the
2326    // import graph so `execute` catches undefined calls at typecheck time.
2327    // The REPL / `-e` path invokes this without `source_path`, where there
2328    // is no importing file context; we fall back to no-imports checking.
2329    let mut checker = TypeChecker::new();
2330    if let Some(path) = source_path {
2331        let graph = harn_modules::build(&[path.to_path_buf()]);
2332        if let Some(imported) = graph.imported_names_for_file(path) {
2333            checker = checker.with_imported_names(imported);
2334        }
2335        if let Some(imported) = graph.imported_type_declarations_for_file(path) {
2336            checker = checker.with_imported_type_decls(imported);
2337        }
2338        if let Some(imported) = graph.imported_callable_declarations_for_file(path) {
2339            checker = checker.with_imported_callable_decls(imported);
2340        }
2341    }
2342    let type_diagnostics = checker.check(&program);
2343    let mut warning_lines = Vec::new();
2344    for diag in &type_diagnostics {
2345        match diag.severity {
2346            DiagnosticSeverity::Error => return Err(diag.message.clone()),
2347            DiagnosticSeverity::Warning => {
2348                warning_lines.push(format!("warning: {}", diag.message));
2349            }
2350        }
2351    }
2352
2353    let chunk = harn_vm::Compiler::new()
2354        .compile(&program)
2355        .map_err(|e| e.to_string())?;
2356
2357    let local = tokio::task::LocalSet::new();
2358    local
2359        .run_until(async {
2360            let mut vm = harn_vm::Vm::new();
2361            harn_vm::register_vm_stdlib(&mut vm);
2362            install_default_hostlib(&mut vm);
2363            let source_parent = source_path
2364                .and_then(|p| p.parent())
2365                .unwrap_or(std::path::Path::new("."));
2366            let project_root = harn_vm::stdlib::process::find_project_root(source_parent);
2367            let store_base = project_root.as_deref().unwrap_or(source_parent);
2368            let execution_cwd = std::env::current_dir()
2369                .unwrap_or_else(|_| std::path::PathBuf::from("."))
2370                .to_string_lossy()
2371                .into_owned();
2372            let source_dir = source_parent.to_string_lossy().into_owned();
2373            if source_path.is_some_and(is_conformance_path) {
2374                harn_vm::event_log::install_memory_for_current_thread(64);
2375            }
2376            harn_vm::register_store_builtins(&mut vm, store_base);
2377            harn_vm::register_metadata_builtins(&mut vm, store_base);
2378            let pipeline_name = source_path
2379                .and_then(|p| p.file_stem())
2380                .and_then(|s| s.to_str())
2381                .unwrap_or("default");
2382            harn_vm::register_checkpoint_builtins(&mut vm, store_base, pipeline_name);
2383            harn_vm::stdlib::process::set_thread_execution_context(Some(
2384                harn_vm::orchestration::RunExecutionRecord {
2385                    cwd: Some(execution_cwd),
2386                    source_dir: Some(source_dir),
2387                    env: std::collections::BTreeMap::new(),
2388                    adapter: None,
2389                    repo_path: None,
2390                    worktree_path: None,
2391                    branch: None,
2392                    base_ref: None,
2393                    cleanup: None,
2394                },
2395            ));
2396            if let Some(ref root) = project_root {
2397                vm.set_project_root(root);
2398            }
2399            if let Some(path) = source_path {
2400                if let Some(parent) = path.parent() {
2401                    if !parent.as_os_str().is_empty() {
2402                        vm.set_source_dir(parent);
2403                    }
2404                }
2405            }
2406            // Conformance tests land here via `run_conformance_tests`; for
2407            // `skill_fs_*` fixtures to see the bundled `skills/` folder
2408            // we run the same layered discovery as `harn run`.
2409            let loaded = skill_loader::load_skills(&skill_loader::SkillLoaderInputs {
2410                cli_dirs: Vec::new(),
2411                source_path: source_path.map(Path::to_path_buf),
2412            });
2413            skill_loader::emit_loader_warnings(&loaded.loader_warnings);
2414            skill_loader::install_skills_global(&mut vm, &loaded);
2415            if let Some(path) = source_path {
2416                let extensions = package::load_runtime_extensions(path);
2417                package::install_runtime_extensions(&extensions);
2418                package::install_manifest_triggers(&mut vm, &extensions)
2419                    .await
2420                    .map_err(|error| format!("failed to install manifest triggers: {error}"))?;
2421                package::install_manifest_hooks(&mut vm, &extensions)
2422                    .await
2423                    .map_err(|error| format!("failed to install manifest hooks: {error}"))?;
2424            }
2425            let _event_log = harn_vm::event_log::active_event_log()
2426                .unwrap_or_else(|| harn_vm::event_log::install_memory_for_current_thread(64));
2427            let connector_clients_installed =
2428                should_install_default_connector_clients(source, source_path);
2429            if connector_clients_installed {
2430                install_default_connector_clients(store_base)
2431                    .await
2432                    .map_err(|error| format!("failed to initialize connector clients: {error}"))?;
2433            }
2434            let execution_result = vm.execute(&chunk).await.map_err(|e| e.to_string());
2435            harn_vm::egress::reset_egress_policy_for_host();
2436            if connector_clients_installed {
2437                harn_vm::clear_active_connector_clients();
2438            }
2439            harn_vm::stdlib::process::set_thread_execution_context(None);
2440            execution_result?;
2441            let mut output = String::new();
2442            for wl in &warning_lines {
2443                output.push_str(wl);
2444                output.push('\n');
2445            }
2446            output.push_str(vm.output());
2447            Ok(output)
2448        })
2449        .await
2450}
2451
2452fn should_install_default_connector_clients(source: &str, source_path: Option<&Path>) -> bool {
2453    if !source_path.is_some_and(is_conformance_path) {
2454        return true;
2455    }
2456    source.contains("connector_call")
2457        || source.contains("std/connectors")
2458        || source.contains("connectors/")
2459}
2460
2461fn is_conformance_path(path: &Path) -> bool {
2462    path.components()
2463        .any(|component| component.as_os_str() == "conformance")
2464}
2465
2466async fn install_default_connector_clients(base_dir: &Path) -> Result<(), String> {
2467    let event_log = harn_vm::event_log::active_event_log()
2468        .unwrap_or_else(|| harn_vm::event_log::install_memory_for_current_thread(64));
2469    let secret_namespace = connector_secret_namespace(base_dir);
2470    let secrets: Arc<dyn harn_vm::secrets::SecretProvider> = Arc::new(
2471        harn_vm::secrets::configured_default_chain(secret_namespace)
2472            .map_err(|error| format!("failed to configure secret providers: {error}"))?,
2473    );
2474
2475    let registry = harn_vm::ConnectorRegistry::default();
2476    let metrics = Arc::new(harn_vm::MetricsRegistry::default());
2477    let inbox = Arc::new(
2478        harn_vm::InboxIndex::new(event_log.clone(), metrics.clone())
2479            .await
2480            .map_err(|error| error.to_string())?,
2481    );
2482    registry
2483        .init_all(harn_vm::ConnectorCtx {
2484            event_log,
2485            secrets,
2486            inbox,
2487            metrics,
2488            rate_limiter: Arc::new(harn_vm::RateLimiterFactory::default()),
2489        })
2490        .await
2491        .map_err(|error| error.to_string())?;
2492    let clients = registry.client_map().await;
2493    harn_vm::install_active_connector_clients(clients);
2494    Ok(())
2495}
2496
2497fn connector_secret_namespace(base_dir: &Path) -> String {
2498    match std::env::var("HARN_SECRET_NAMESPACE") {
2499        Ok(namespace) if !namespace.trim().is_empty() => namespace,
2500        _ => {
2501            let leaf = base_dir
2502                .file_name()
2503                .and_then(|name| name.to_str())
2504                .filter(|name| !name.is_empty())
2505                .unwrap_or("workspace");
2506            format!("harn/{leaf}")
2507        }
2508    }
2509}
2510
2511#[cfg(test)]
2512mod main_tests {
2513    use super::{normalize_serve_args, should_install_default_connector_clients};
2514    use std::path::Path;
2515
2516    #[test]
2517    fn normalize_serve_args_inserts_a2a_for_legacy_shape() {
2518        let args = normalize_serve_args(vec![
2519            "harn".to_string(),
2520            "serve".to_string(),
2521            "--port".to_string(),
2522            "3000".to_string(),
2523            "agent.harn".to_string(),
2524        ]);
2525        assert_eq!(
2526            args,
2527            vec![
2528                "harn".to_string(),
2529                "serve".to_string(),
2530                "a2a".to_string(),
2531                "--port".to_string(),
2532                "3000".to_string(),
2533                "agent.harn".to_string(),
2534            ]
2535        );
2536    }
2537
2538    #[test]
2539    fn normalize_serve_args_preserves_explicit_subcommands() {
2540        let args = normalize_serve_args(vec![
2541            "harn".to_string(),
2542            "serve".to_string(),
2543            "acp".to_string(),
2544            "server.harn".to_string(),
2545        ]);
2546        assert_eq!(
2547            args,
2548            vec![
2549                "harn".to_string(),
2550                "serve".to_string(),
2551                "acp".to_string(),
2552                "server.harn".to_string(),
2553            ]
2554        );
2555    }
2556
2557    #[test]
2558    fn conformance_skips_connector_clients_unless_fixture_uses_connectors() {
2559        let path = Path::new("conformance/tests/language/basic.harn");
2560        assert!(!should_install_default_connector_clients(
2561            "println(1)",
2562            Some(path)
2563        ));
2564        assert!(!should_install_default_connector_clients(
2565            "trust_graph_verify_chain()",
2566            Some(path)
2567        ));
2568        assert!(should_install_default_connector_clients(
2569            "import { post_message } from \"std/connectors/slack\"",
2570            Some(path)
2571        ));
2572        assert!(should_install_default_connector_clients(
2573            "println(1)",
2574            Some(Path::new("examples/demo.harn"))
2575        ));
2576    }
2577}