llmlint 0.3.3

LLM-as-judge linter: enforce code-quality checks deterministic linters can't express, by driving real coding harnesses through oneharness.
Documentation
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
440
441
442
443
444
445
446
447
448
449
450
451
452
453
454
455
456
457
458
459
460
461
462
463
464
465
466
467
468
469
470
471
472
473
474
475
476
477
478
479
480
481
482
483
484
485
486
487
488
489
490
491
492
493
494
495
496
497
498
499
500
501
502
503
504
505
506
507
508
509
510
511
512
513
514
515
516
517
518
519
520
521
522
523
524
525
526
527
528
529
530
531
532
533
534
535
536
537
538
539
540
541
542
543
544
545
546
547
548
549
550
551
552
553
554
555
556
557
558
559
560
561
562
563
564
565
566
567
568
569
570
571
572
573
574
575
576
577
578
579
580
581
582
583
584
585
586
587
588
589
590
591
592
593
594
595
596
597
598
599
600
601
602
603
604
605
606
607
608
609
610
611
612
613
614
615
616
617
618
619
620
621
622
623
624
625
626
627
628
629
630
631
632
633
634
635
636
637
638
639
640
641
642
643
644
645
646
647
648
649
650
651
652
653
654
655
656
657
658
659
660
661
662
663
664
665
666
667
668
669
670
671
672
673
674
675
676
677
678
679
680
681
682
683
684
685
686
687
688
689
690
691
692
693
694
695
696
697
698
699
700
701
702
703
704
705
706
707
708
709
710
711
712
713
714
715
716
717
718
719
720
721
722
723
724
725
726
727
728
729
730
731
732
733
734
735
736
737
738
739
740
741
742
743
744
745
746
747
748
749
750
751
752
753
754
755
756
757
758
759
760
761
762
763
764
765
766
767
768
769
770
771
772
773
774
775
776
777
778
779
780
781
782
783
784
785
786
787
788
789
790
791
792
793
794
795
796
797
798
799
800
801
802
803
804
805
806
807
808
809
810
811
812
813
814
815
816
817
818
819
820
821
822
823
824
825
826
827
828
829
830
831
832
833
834
835
836
837
838
839
840
841
842
843
844
845
846
847
848
849
850
851
852
853
854
855
856
857
858
859
860
861
862
863
864
865
866
867
//! `llmlint lint` (the default): load config, resolve files, plan judge runs,
//! drive oneharness in parallel, aggregate votes, and report.

use std::collections::{BTreeMap, BTreeSet, HashSet};
use std::io::{IsTerminal, Write};
use std::path::{Path, PathBuf};
use std::thread;

use indicatif::ProgressDrawTarget;

use crate::cli::{ColorChoice, LintArgs, OutputFormat};
use crate::commands::ignores;
use crate::commands::progress::{LiveStatus, ProgressView};
use crate::domain::config::{validate, Config, RelevanceMode, Rule};
use crate::domain::ignore::Suppressions;
use crate::domain::plan::{self, JudgeRun};
use crate::domain::report::Report;
use crate::domain::template::{self};
use crate::domain::verdict::{Outcome, RuleOutcome, RuleVerdict};
use crate::domain::{applicability, attribution, ignore, schema, vote};
use crate::errors::{io_err, Error, Result};
use crate::io::configfs::RuleScope;
use crate::io::{assets, configfs, diff, files, oneharness};

const DEFAULT_BATCH_SIZE: usize = 20;
const DEFAULT_TIMEOUT: u64 = 120;
const DEFAULT_MAX_PARALLEL: usize = 8;
const PROMPT_TRIGGER: &str =
    "Evaluate each rule against the target files and respond with the structured verdict object.";
/// How many times a judge whose verdict strays outside a rule's file scope is
/// asked to rework its answer before llmlint drops the wrong-file violations
/// deterministically. One corrective round catches an honest slip without
/// looping on a judge that won't comply.
const MAX_REWORKS: usize = 1;

pub fn run(args: LintArgs) -> Result<i32> {
    let cwd = resolve_cwd(&args.cwd)?;

    // Explicit CLI files relevance-gate the subtree cascade: linting specific
    // files never loads an unrelated subtree's config (see `load_with_targets`).
    let loaded = configfs::load_with_targets(&args.config, &cwd, &args.files)?;
    run_loaded(loaded, cwd, args)
}

/// Resolve the working directory for a run: the `--cwd` override, else the
/// process cwd. Shared by `lint` and the `lint-config` subcommand.
pub(crate) fn resolve_cwd(arg: &Option<PathBuf>) -> Result<PathBuf> {
    match arg {
        Some(d) => Ok(d.clone()),
        None => std::env::current_dir().map_err(|e| Error::Io(e.to_string())),
    }
}

/// Drive a lint from an already-loaded config. Split out from [`run`] so the
/// `lint-config` subcommand can hand in the bundled config-lint config (loaded
/// without discovery) and reuse the entire engine — validation, planning,
/// judging, voting, and reporting — unchanged.
pub(crate) fn run_loaded(loaded: configfs::Loaded, cwd: PathBuf, args: LintArgs) -> Result<i32> {
    let scopes = loaded.scopes;
    let mut config = loaded.config;
    validate(&config)?;
    validate_filters(&config, &args)?;
    // Overlay CLI overrides onto the merged config so every top-level arg can be
    // set on the command line, with the CLI winning over the config (which in
    // turn won over its plugins). After this, downstream (planning, schema,
    // template) reads the single effective config.
    apply_cli_overrides(&mut config, &args)?;
    let session_rationales = config.rationales_default();

    let selected = select_rules(&config, &args);
    if selected.is_empty() {
        let report = Report::new(Vec::new(), Vec::new());
        emit(&report, args.format, args.verbose, args.color);
        return Ok(report.exit_code());
    }

    let master_template = config
        .prompt_template
        .clone()
        .unwrap_or_else(|| assets::DEFAULT_TEMPLATE.to_string());
    let cli_files = files::from_cli(&cwd, &args.files);

    let mut resolved = Vec::new();
    // Rules declared statically not relevant (`relevance: false`) never reach a
    // judge — they are reported as not relevant directly.
    let mut not_relevant: Vec<String> = Vec::new();
    for rule in &selected {
        let relevance = match rule.relevance_mode() {
            RelevanceMode::Never => {
                not_relevant.push(rule.name.clone());
                continue;
            }
            RelevanceMode::Always => None,
            RelevanceMode::Conditional(cond) => Some(cond),
        };
        let agent_name = rule.agent.clone().unwrap_or_else(|| "default".to_string());
        let fallback;
        let scope = match scopes.get(&rule.name) {
            Some(s) => s,
            // Every selected rule has a scope; fall back to cwd defensively.
            None => {
                fallback = RuleScope {
                    dir: cwd.clone(),
                    files: config.files.clone(),
                };
                &fallback
            }
        };
        let target = ignores::resolve_files(&cwd, rule, &cli_files, scope)?;
        resolved.push(plan::ResolvedRule {
            name: rule.name.clone(),
            description: rule.description.clone(),
            judges: rule.judges(),
            agent: agent_name,
            files: target,
            rationale: rule.wants_rationale(session_rationales),
            relevance,
            require_line_attribution: rule.requires_line_attribution(),
        });
    }

    // Reject malformed inline `llmlint: ignore` directives in the target files
    // before spending any judge calls. Honoring well-formed ones is the judge's
    // job (the default template tells it how); their *structure* is enforced here
    // so a typo'd or reason-less ignore fails loudly instead of silently doing
    // nothing. This is the same check `llmlint check-ignores` runs standalone, so
    // the fast static loop and the full run never disagree.
    let targets: BTreeSet<PathBuf> = resolved
        .iter()
        .flat_map(|r| r.files.iter().cloned())
        .collect();
    let known = ignores::known_rules(&config);
    ignores::check(&cwd, &targets, &known)?;

    // Parse each target file's well-formed inline ignores into line-span
    // suppressions, keyed by the file's slash path. After a judge answers, any
    // violation an ignore covers is dropped deterministically — llmlint honors
    // the directives itself rather than trusting the judge to (the default
    // template still documents line/block ignores as a backstop).
    let mut suppressions: BTreeMap<String, Suppressions> = BTreeMap::new();
    for rel in &targets {
        if let Some(text) = files::read_text(&cwd, rel)? {
            let s = ignore::suppressions(&text, &known);
            if !s.is_empty() {
                suppressions.insert(files::to_slash(rel), s);
            }
        }
    }

    // Under `--diff`, compute each target file's changed-line diff once (at the
    // I/O boundary) so every judge prompt can show exactly what changed. The
    // backend is selected behind the `DiffProvider` trait, comparing against
    // `--diff-base` (a branch/tag/commit/range) or the backend default; an
    // unchanged file simply has no entry. Absent the flag this is empty and
    // nothing renders.
    let diffs: BTreeMap<PathBuf, String> = match args.diff {
        Some(backend) => {
            let targets: Vec<PathBuf> = targets.iter().cloned().collect();
            // The base is the effective config value: `--diff-base` already won
            // over a config `diff_base` in `apply_cli_overrides`; `None` leaves
            // the backend's built-in default (`HEAD` for git).
            diff::provider(backend, config.diff_base.clone()).diffs(&cwd, &targets)?
        }
        None => BTreeMap::new(),
    };

    // Rules whose rationale is disabled: llmlint is authoritative, so we drop any
    // rationale a harness returns anyway, keeping `--no-rationales` deterministic
    // regardless of harness behavior.
    let rationale_off: HashSet<String> = resolved
        .iter()
        .filter(|r| !r.rationale)
        .map(|r| r.name.clone())
        .collect();

    // Rules that opted into line attribution: every violation they surface must
    // cite a file+line. The schema makes oneharness re-prompt for it; this set
    // backs the deterministic post-vote backstop below.
    let require_attribution: BTreeSet<String> = resolved
        .iter()
        .filter(|r| r.require_line_attribution)
        .map(|r| r.name.clone())
        .collect();

    let the_plan = plan::build(&config, &master_template, DEFAULT_BATCH_SIZE, resolved);

    let bin = args
        .oneharness_bin
        .clone()
        .or_else(|| {
            std::env::var("LLMLINT_ONEHARNESS_BIN")
                .ok()
                .filter(|s| !s.is_empty())
        })
        .or_else(|| config.oneharness.bin.clone());
    let client = oneharness::Client::new(bin.as_deref());
    // Pre-flight: read-only mode (so the harness never edits target files)
    // requires oneharness >= MIN_VERSION. Check once up front and fail with a
    // clear message rather than letting every judge's `--mode read-only` error.
    client.check_min_version()?;
    let timeout = args
        .timeout
        .or(config.oneharness.timeout)
        .unwrap_or(DEFAULT_TIMEOUT);
    let oh_config = resolve_oneharness_config(&args, &config);
    let oh_config_ref = oh_config.as_deref();
    let global_model = config.oneharness.model.as_deref();
    let max_parallel = args.max_parallel.unwrap_or(DEFAULT_MAX_PARALLEL).max(1);

    let mut verdicts: BTreeMap<String, Vec<RuleVerdict>> = BTreeMap::new();
    let mut run_errors: Vec<String> = Vec::new();
    // At `-v` we collect each oneharness invocation's exact command + raw result
    // and print them to stderr (the debug view); off by default to avoid the
    // cost of formatting the full command line on every judge.
    let want_trace = args.verbose >= 1;
    let mut traces: Vec<(String, oneharness::RunTrace)> = Vec::new();

    // The live-progress view (rules resolving as their judges return), drawn to
    // stderr for an interactive human. Decide whether to animate from the audience
    // (`--progress` + TTY/CI/agent), then build the view with a real or hidden
    // draw target. It is *always* constructed — hidden when not interactive — so
    // the driving glue below runs on every path (a piped/CI/agent run just draws
    // to nothing, emitting no escape codes). The report on stdout is unaffected.
    let ctx = crate::io::terminal::detect();
    let show_progress = args.format == OutputFormat::Human
        && args
            .progress
            .resolve(ctx.stderr_tty, ctx.is_ci, ctx.is_agent, ctx.color_ok());
    // Rules shown in the view, in a stable (sorted) order: every judged rule plus
    // the ones already resolved without a judge (skipped / statically not
    // relevant), which we finish immediately below.
    let mut view_rules: BTreeSet<String> = BTreeSet::new();
    let mut expected: BTreeMap<String, usize> = BTreeMap::new();
    for run in &the_plan.runs {
        for rs in &run.rules {
            view_rules.insert(rs.name.clone());
            *expected.entry(rs.name.clone()).or_default() += 1;
        }
    }
    view_rules.extend(the_plan.skipped.iter().cloned());
    view_rules.extend(not_relevant.iter().cloned());
    let view_rules: Vec<String> = view_rules.into_iter().collect();
    let view = ProgressView::new(
        progress_target(show_progress),
        &view_rules,
        the_plan.runs.len(),
        show_progress,
    );
    for name in &the_plan.skipped {
        view.finish_rule(name, LiveStatus::Skipped);
    }
    for name in &not_relevant {
        view.finish_rule(name, LiveStatus::NotRelevant);
    }

    // Bind Copy references so the per-judge `move` closures capture borrows
    // (which outlive the scope) rather than moving the owned `client`/`cwd`.
    let client_ref = &client;
    let cwd_ref = cwd.as_path();
    let diffs_ref = &diffs;
    let suppressions_ref = &suppressions;
    for wave in the_plan.runs.chunks(max_parallel) {
        // The wave's rules are now in flight — spin their lines.
        for run in wave {
            for rs in &run.rules {
                view.set_running(&rs.name);
            }
        }
        thread::scope(|s| {
            let handles: Vec<_> = wave
                .iter()
                .map(|run| {
                    s.spawn(move || {
                        execute(
                            client_ref,
                            run,
                            cwd_ref,
                            timeout,
                            oh_config_ref,
                            global_model,
                            want_trace,
                            diffs_ref,
                            suppressions_ref,
                        )
                    })
                })
                .collect();
            for (run, handle) in wave.iter().zip(handles) {
                let (trace, result) = handle.join().expect("judge thread panicked");
                if let Some(trace) = trace {
                    traces.push((judge_label(run), trace));
                }
                view.tick_run();
                match result {
                    Ok(map) => {
                        for (name, verdict) in map {
                            verdicts.entry(name).or_default().push(verdict);
                        }
                    }
                    Err(e) => run_errors.push(format!("{}: {}", judge_label(run), e)),
                }
                // Resolve each of this run's rules whose judges are now all in:
                // tally with the same `vote::tally` the report uses, so the live
                // ✓/✗ can never disagree with the final verdict. A rule with no
                // usable verdict (all its runs errored) shows an error glyph.
                for rs in &run.rules {
                    let remaining = expected.get_mut(rs.name.as_str());
                    if let Some(remaining) = remaining {
                        *remaining = remaining.saturating_sub(1);
                        if *remaining == 0 {
                            let status = match verdicts.get(rs.name.as_str()) {
                                Some(vs) if !vs.is_empty() => {
                                    live_status(&vote::tally(&rs.name, vs))
                                }
                                _ => LiveStatus::Error,
                            };
                            view.finish_rule(&rs.name, status);
                        }
                    }
                }
            }
        });
    }
    // Clear the whole view from stderr before any report/trace output, so nothing
    // is interleaved with progress fragments.
    view.finish();

    if want_trace {
        print_traces(&traces);
    }

    let mut outcomes: Vec<RuleOutcome> = verdicts
        .iter()
        .map(|(name, vs)| vote::tally(name, vs))
        .collect();
    for o in &mut outcomes {
        if rationale_off.contains(&o.name) {
            o.rationale = None;
            for j in &mut o.judges {
                j.rationale = None;
            }
        }
    }
    for name in &the_plan.skipped {
        outcomes.push(RuleOutcome::skipped(name));
    }
    for name in &not_relevant {
        outcomes.push(RuleOutcome::not_relevant(name));
    }

    // Backstop the `require_line_attribution` contract: any failing rule that
    // opted in but still surfaced a violation without a file+line is a hard error
    // (the schema already had oneharness re-prompt for it in one batched turn).
    if !require_attribution.is_empty() {
        run_errors.extend(attribution::unlocalized_errors(
            &outcomes,
            &require_attribution,
        ));
    }

    let report = Report::new(outcomes, run_errors);
    emit(&report, args.format, args.verbose, args.color);
    Ok(report.exit_code())
}

/// Reject `--rule`/`--agent` targets that name nothing in the config. Without
/// this a typo (`--rule no_todos` vs `no_todo`) selects zero rules and exits 0 —
/// a false green. A genuinely-empty-but-valid selection (real names that just
/// don't intersect) is allowed through to exit 0 by `select_rules`. `default`
/// is always a valid agent target (it runs rules with no explicit agent).
fn validate_filters(config: &Config, args: &LintArgs) -> Result<()> {
    let mut problems: Vec<String> = Vec::new();

    if !args.rule.is_empty() {
        let known: HashSet<&str> = config.rules.iter().map(|r| r.name.as_str()).collect();
        let mut unknown: Vec<&str> = args
            .rule
            .iter()
            .map(String::as_str)
            .filter(|n| !known.contains(n))
            .collect();
        if !unknown.is_empty() {
            unknown.sort_unstable();
            unknown.dedup();
            let mut available: Vec<&str> = known.into_iter().collect();
            available.sort_unstable();
            problems.push(format!(
                "no rule named {}; available rules: {}",
                unknown.join(", "),
                join_or_none(&available)
            ));
        }
    }

    if let Some(agent) = &args.agent {
        if agent != "default" && !config.agents.contains_key(agent) {
            let mut available: Vec<&str> = config.agents.keys().map(String::as_str).collect();
            if !available.contains(&"default") {
                available.push("default");
            }
            available.sort_unstable();
            problems.push(format!(
                "no agent named {}; available agents: {}",
                agent,
                join_or_none(&available)
            ));
        }
    }

    if problems.is_empty() {
        Ok(())
    } else {
        Err(Error::UnknownFilter(problems.join("; ")))
    }
}

fn join_or_none(names: &[&str]) -> String {
    if names.is_empty() {
        "(none)".to_string()
    } else {
        names.join(", ")
    }
}

fn select_rules<'a>(config: &'a Config, args: &LintArgs) -> Vec<&'a Rule> {
    let rule_filter: Option<HashSet<&str>> = if args.rule.is_empty() {
        None
    } else {
        Some(args.rule.iter().map(String::as_str).collect())
    };
    config
        .rules
        .iter()
        .filter(|r| {
            let agent_ok = args.agent.as_deref().is_none_or(|a| {
                r.agent.as_deref() == Some(a) || (a == "default" && r.agent.is_none())
            });
            let name_ok = rule_filter
                .as_ref()
                .is_none_or(|set| set.contains(r.name.as_str()));
            agent_ok && name_ok
        })
        .collect()
}

/// Overlay the lint CLI's top-level overrides onto the merged config so the CLI
/// wins over the config. Each knob also has a config field and (transitively) a
/// plugin precedence; this is the final, highest-priority layer. `--oneharness-bin`,
/// `--timeout`, `--oneharness-config`, and the file/agent/rule selectors are
/// resolved at their use sites (they fold in env/discovery too) and are not
/// overlaid here.
fn apply_cli_overrides(config: &mut Config, args: &LintArgs) -> Result<()> {
    if let Some(path) = &args.prompt_template {
        let text = std::fs::read_to_string(path)
            .map_err(|e| io_err(format!("reading prompt template {}", path.display()), e))?;
        config.prompt_template = Some(text);
    }
    if let Some(model) = &args.model {
        config.oneharness.model = Some(model.clone());
    }
    if let Some(n) = args.schema_max_retries {
        config.oneharness.schema_max_retries = Some(n);
    }
    if let Some(b) = args.rationales() {
        config.rationales = Some(b);
    }
    if args.diff_base.is_some() {
        config.diff_base = args.diff_base.clone();
    }
    Ok(())
}

fn resolve_oneharness_config(args: &LintArgs, config: &Config) -> Option<PathBuf> {
    let mut all: Vec<PathBuf> = args.oneharness_config.clone();
    all.extend(config.oneharness.config.iter().map(PathBuf::from));
    if all.len() > 1 {
        eprintln!(
            "llmlint: warning: oneharness `--config` takes a single file; using {} and ignoring \
             {} other(s)",
            all[0].display(),
            all.len() - 1
        );
    }
    all.into_iter().next()
}

/// The draw target for the live view: real stderr when it should show (indicatif
/// still self-hides if stderr isn't a terminal), else a hidden target that draws
/// nothing. Split out so the selection is unit-testable off a real TTY.
fn progress_target(show: bool) -> ProgressDrawTarget {
    if show {
        ProgressDrawTarget::stderr()
    } else {
        ProgressDrawTarget::hidden()
    }
}

/// Map a tallied rule outcome to the live view's status glyph.
fn live_status(o: &RuleOutcome) -> LiveStatus {
    match o.outcome {
        Outcome::Pass => LiveStatus::Pass,
        Outcome::Fail => LiveStatus::Fail,
        Outcome::Skipped => LiveStatus::Skipped,
        Outcome::NotRelevant => LiveStatus::NotRelevant,
    }
}

/// A stable per-judge label used for both error messages and the `-v` debug
/// trace headers: `agent <name> judge <i> [<rule>, ...]`.
fn judge_label(run: &JudgeRun) -> String {
    format!(
        "agent {} judge {} [{}]",
        run.agent,
        run.judge_index,
        run.rules
            .iter()
            .map(|r| r.name.as_str())
            .collect::<Vec<_>>()
            .join(", ")
    )
}

/// Print the oneharness debug view (exact command + raw result per judge) to
/// stderr, keeping the parseable report on stdout. Shown at `-v`.
fn print_traces(traces: &[(String, oneharness::RunTrace)]) {
    for (label, trace) in traces {
        eprintln!("\n# oneharness: {label}");
        eprintln!("$ {}", trace.command);
        if let Some(code) = trace.exit_code {
            eprintln!("exit: {code}");
        }
        let stdout = trace.stdout.trim();
        if !stdout.is_empty() {
            eprintln!("result:\n{stdout}");
        }
        let stderr = trace.stderr.trim();
        if !stderr.is_empty() {
            eprintln!("stderr:\n{stderr}");
        }
    }
}

#[allow(clippy::too_many_arguments)]
fn execute(
    client: &oneharness::Client,
    run: &JudgeRun,
    cwd: &Path,
    timeout: u64,
    oh_config: Option<&Path>,
    global_model: Option<&str>,
    want_trace: bool,
    diffs: &BTreeMap<PathBuf, String>,
    suppressions: &BTreeMap<String, Suppressions>,
) -> (
    Option<oneharness::RunTrace>,
    Result<BTreeMap<String, RuleVerdict>>,
) {
    let files_str: Vec<String> = run.files.iter().map(|p| files::to_slash(p)).collect();
    // Per-file diffs for this run's files, in the same order as `files_str`. Only
    // files that actually changed appear in `diffs`, so unchanged ones are
    // skipped here; the slice is empty when `--diff` wasn't passed.
    let file_diffs: Vec<template::FileDiff> = run
        .files
        .iter()
        .filter_map(|p| {
            diffs.get(p).map(|d| template::FileDiff {
                file: files::to_slash(p),
                diff: d.clone(),
            })
        })
        .collect();
    // Show the rationale guidance when any rule in this batch wants a rationale,
    // and the relevance guidance when any rule is conditional on relevance.
    let want_rationale = run.rules.iter().any(|r| r.rationale);
    let want_relevance = run.rules.iter().any(|r| r.relevance.is_some());
    let want_line_attribution = run.rules.iter().any(|r| r.require_line_attribution);
    let system = match template::render(
        &run.template,
        &run.rules,
        &files_str,
        &file_diffs,
        want_rationale,
        want_relevance,
        want_line_attribution,
    ) {
        Ok(s) => s,
        // A render failure happens before any oneharness call, so there is no
        // command to trace.
        Err(e) => return (None, Err(e)),
    };
    let specs: Vec<schema::SchemaRule> = run
        .rules
        .iter()
        .map(|r| schema::SchemaRule {
            name: r.name.as_str(),
            rationale: r.rationale,
            relevance: r.relevance.is_some(),
            require_line_attribution: r.require_line_attribution,
        })
        .collect();
    let schema = schema::build(&specs);

    // Per-rule scope (which files each rule covers) + the per-file applicability
    // shown to the judge — reused to validate the verdict and to phrase a rework.
    let pairs: Vec<(String, Vec<String>)> = run
        .rules
        .iter()
        .map(|r| (r.name.clone(), r.files.clone()))
        .collect();
    let scope = applicability::scope_map(&pairs);
    let file_rules = applicability::per_file(&pairs, &files_str);

    // One judge call, plus up to MAX_REWORKS corrective rounds: if the verdict
    // pins a violation to a file outside that rule's scope (a "wrong rule in
    // wrong file"), re-ask with the exact per-file rule lists before falling back
    // to deterministic cleanup.
    let mut prompt = PROMPT_TRIGGER.to_string();
    let mut last_trace;
    let mut verdicts;
    let mut attempt = 0;
    loop {
        let req = oneharness::RunRequest {
            harness: run.harness.as_deref(),
            model: run.model.as_deref().or(global_model),
            system: &system,
            prompt: &prompt,
            schema: &schema,
            schema_max_retries: run.schema_max_retries,
            cwd,
            timeout_secs: timeout,
            oneharness_config: oh_config,
            no_config: false,
        };
        let (trace, result) = if want_trace {
            let (t, r) = client.run_with_trace(&req);
            (Some(t), r)
        } else {
            (None, client.run(&req))
        };
        last_trace = trace;
        verdicts = match result {
            Ok(v) => v,
            Err(e) => return (last_trace, Err(e)),
        };
        let problems = applicability::scope_problems(&scope, &verdicts);
        if problems.is_empty() || attempt >= MAX_REWORKS {
            break;
        }
        prompt = applicability::rework_prompt(&problems, &file_rules);
        attempt += 1;
    }

    // Deterministic cleanup, applied whether or not a rework ran: drop any
    // remaining wrong-file violation and any an inline ignore covers, flipping a
    // fail to a pass when that removes its entire basis.
    let empty_scope = BTreeSet::new();
    for (name, verdict) in verdicts.iter_mut() {
        let rule_scope = scope.get(name.as_str()).unwrap_or(&empty_scope);
        applicability::clean_verdict(verdict, rule_scope, |file, line| {
            suppressions.get(file).is_some_and(|s| s.covers(name, line))
        });
    }
    (last_trace, Ok(verdicts))
}

fn emit(report: &Report, format: OutputFormat, verbosity: u8, color: ColorChoice) {
    match format {
        OutputFormat::Human => {
            // Resolve `--color` against the live stdout: `auto` colors only an
            // interactive terminal with `NO_COLOR`/`TERM=dumb` unset that is not an
            // AI agent (captured ANSI is unreliable there). The decision is made
            // here (the I/O boundary) and handed to the pure formatter as a bool.
            let ctx = crate::io::terminal::detect();
            let on = color.resolve(std::io::stdout().is_terminal(), ctx.no_color, ctx.is_agent);
            // Write through anstream's `AutoStream` so the ANSI we emit renders
            // everywhere, not just on Unix. On a *legacy* Windows console (no
            // virtual-terminal processing) raw ANSI prints as `←[31m` garbage;
            // `AutoStream` enables VT when it can and otherwise translates the SGR
            // codes into Win32 console attribute calls (`anstyle-wincon`). The
            // `--color` decision already lives in `on`, so feed the stream a
            // concrete choice rather than letting it re-detect: `Always` when
            // color is on (never second-guess `--color always`), `Never`
            // otherwise — a no-op strip over already-plain text that keeps the
            // sink uniform. On a pipe/redirect (no console) `Always` still emits
            // plain ANSI on every OS, so `--color always` and the screenshot/e2e
            // byte assertions stay byte-identical.
            let choice = if on {
                anstream::ColorChoice::Always
            } else {
                anstream::ColorChoice::Never
            };
            let mut out = anstream::AutoStream::new(std::io::stdout().lock(), choice);
            let _ = write!(out, "{}", report.to_human(verbosity, on));
            let _ = out.flush();
        }
        OutputFormat::Json => {
            println!(
                "{}",
                serde_json::to_string_pretty(&report.to_json()).unwrap_or_default()
            )
        }
    }
}

#[cfg(test)]
mod tests {
    use super::*;
    use crate::domain::config::Agent;

    fn rule(name: &str, agent: Option<&str>) -> Rule {
        Rule {
            name: name.into(),
            description: "true when ok; false otherwise.".into(),
            r#override: false,
            agent: agent.map(Into::into),
            judges: None,
            files: None,
            rationale: None,
            relevance: None,
            require_line_attribution: None,
        }
    }

    fn config_with(rules: Vec<Rule>, agents: &[&str]) -> Config {
        let mut c = Config {
            rules,
            ..Default::default()
        };
        for a in agents {
            c.agents.insert((*a).into(), Agent::default());
        }
        c
    }

    fn args(rules: &[&str], agent: Option<&str>) -> LintArgs {
        LintArgs {
            rule: rules.iter().map(|s| (*s).to_string()).collect(),
            agent: agent.map(Into::into),
            ..Default::default()
        }
    }

    #[test]
    fn no_filters_is_ok() {
        let cfg = config_with(vec![rule("a_rule", None)], &[]);
        assert!(validate_filters(&cfg, &args(&[], None)).is_ok());
    }

    #[test]
    fn progress_target_hidden_when_not_showing() {
        // The suppressed path is always a hidden target (draws nothing).
        assert!(progress_target(false).is_hidden());
        // The showing path builds a real stderr target; under the test runner
        // stderr isn't a TTY, so indicatif still resolves it to hidden — the point
        // is the construction line runs. (A real terminal makes it visible.)
        let _ = progress_target(true);
    }

    #[test]
    fn live_status_maps_every_outcome() {
        let outcome = |o: Outcome| RuleOutcome {
            name: "r".into(),
            rationale: None,
            outcome: o,
            votes_total: 1,
            votes_hold: 0,
            judges: vec![],
            violations: vec![],
        };
        assert!(matches!(
            live_status(&outcome(Outcome::Pass)),
            LiveStatus::Pass
        ));
        assert!(matches!(
            live_status(&outcome(Outcome::Fail)),
            LiveStatus::Fail
        ));
        assert!(matches!(
            live_status(&outcome(Outcome::Skipped)),
            LiveStatus::Skipped
        ));
        assert!(matches!(
            live_status(&outcome(Outcome::NotRelevant)),
            LiveStatus::NotRelevant
        ));
    }

    #[test]
    fn known_rule_and_agent_are_ok() {
        let cfg = config_with(vec![rule("a_rule", Some("special"))], &["special"]);
        assert!(validate_filters(&cfg, &args(&["a_rule"], Some("special"))).is_ok());
    }

    #[test]
    fn default_agent_is_always_valid() {
        let cfg = config_with(vec![rule("a_rule", None)], &[]);
        assert!(validate_filters(&cfg, &args(&[], Some("default"))).is_ok());
    }

    #[test]
    fn unknown_rule_lists_available_sorted() {
        let cfg = config_with(vec![rule("beta", None), rule("alpha", None)], &[]);
        let err = validate_filters(&cfg, &args(&["typo", "alpha"], None)).unwrap_err();
        let msg = err.to_string();
        assert!(msg.contains("no rule named typo"), "got: {msg}");
        assert!(msg.contains("available rules: alpha, beta"), "got: {msg}");
    }

    #[test]
    fn unknown_agent_lists_available_with_default() {
        let cfg = config_with(vec![rule("a_rule", Some("special"))], &["special"]);
        let err = validate_filters(&cfg, &args(&[], Some("ghost"))).unwrap_err();
        let msg = err.to_string();
        assert!(msg.contains("no agent named ghost"), "got: {msg}");
        assert!(
            msg.contains("available agents: default, special"),
            "got: {msg}"
        );
    }

    #[test]
    fn unknown_rule_with_no_rules_says_none() {
        let cfg = config_with(vec![], &[]);
        let err = validate_filters(&cfg, &args(&["x"], None)).unwrap_err();
        assert!(err.to_string().contains("available rules: (none)"));
    }

    #[test]
    fn cli_overrides_win_over_config() {
        let mut cfg = Config {
            rationales: Some(true),
            ..Default::default()
        };
        cfg.oneharness.model = Some("config-model".into());
        let args = LintArgs {
            model: Some("cli-model".into()),
            schema_max_retries: Some(5),
            no_rationales: true,
            ..Default::default()
        };
        apply_cli_overrides(&mut cfg, &args).unwrap();
        assert_eq!(cfg.oneharness.model.as_deref(), Some("cli-model"));
        assert_eq!(cfg.oneharness.schema_max_retries, Some(5));
        assert_eq!(cfg.rationales, Some(false));
        assert!(!cfg.rationales_default());
    }

    #[test]
    fn no_rationale_flags_leaves_config_untouched() {
        let mut cfg = Config {
            rationales: Some(false),
            ..Default::default()
        };
        apply_cli_overrides(&mut cfg, &LintArgs::default()).unwrap();
        // No --model/--schema-max-retries/--rationales: config value survives.
        assert_eq!(cfg.rationales, Some(false));
    }

    #[test]
    fn both_unknown_rule_and_agent_are_reported() {
        let cfg = config_with(vec![rule("a_rule", None)], &[]);
        let err = validate_filters(&cfg, &args(&["nope"], Some("ghost"))).unwrap_err();
        let msg = err.to_string();
        assert!(msg.contains("no rule named nope"), "got: {msg}");
        assert!(msg.contains("no agent named ghost"), "got: {msg}");
    }
}