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