magi-cli 0.10.1

Blind multi-agent implementation competition: N agents implement, M judges rank blind, deliberate, vote privately, winner survives double review + E2E gate
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
868
869
870
871
872
873
874
875
876
877
878
879
880
881
882
883
884
885
886
887
//! The headless design-deliberation stage between `magi plan`'s interview and
//! the task file it files.
//!
//! [`crate::graph`]'s judge / vote / deliberate machinery only runs when more
//! than one candidate survives - and `[graph] candidates` defaults to 1 now
//! (see [`crate::config::Graph::candidates`]'s doc for the cost numbers
//! behind that default), so on an ordinary task none of that machinery ever
//! fires any more. Diversity did not stop paying for itself; it moved. A
//! design sketch is a few paragraphs an agent can write without touching the
//! repository, where a full implementation is a hundred-plus-turn tool loop
//! that re-reads the codebase on every turn - so three sketches, gathered
//! once between the interview and the file, cost a fraction of a third
//! implementation and buy back the same disagreement the judges used to
//! surface, on every task rather than only the ones run with `--candidates`.
//!
//! Three independent, read-only advisors ([`gather`]) each propose one
//! design. The planner seat (`[roles] planner`) then reads all three and
//! [`prompt::synthesize`]s them into the task file's `## Context` and
//! `## Change`, naming which advisor's idea it kept where - it is told, in so
//! many words, not to pick a winner. Nothing here talks to the operator: the
//! interview already did that, and turning this stage into three more
//! conversations would be exactly the cost this module exists to avoid.
//!
//! # Disposability narrows the blast radius, it does not forbid writing
//!
//! `allow_write: false` alone is not a guarantee: opencode has no read-only
//! mode at all, and Claude's `--disallowed-tools` stops its edit tools but not
//! a `rm` or a redirect run through its Bash tool. That is the existing,
//! accepted risk model for every read-only seat in this codebase - judge and
//! reviewer seats in [`crate::graph`] carry exactly the same weak guarantee,
//! and get away with it because their `cwd` is already a worktree the run
//! treats as disposable. This stage runs before any run exists, so it was
//! pointed at the operator's own checkout until [`checkout_worktrees`] gave
//! every advisor seat, and the planner's synthesis, a `git worktree add
//! --detach` checkout at `HEAD` of their own, thrown away when [`run`]
//! returns - the same protection judges and reviewers already had, closing
//! the one gap unique to this stage rather than inventing a stronger
//! guarantee nothing else here provides.
//!
//! What this buys: a relative-path write from a seat that ignores its
//! instructions lands in that seat's own disposable checkout, not in the
//! operator's repository and not in another seat's. What it does not buy: an
//! absolute-path write, or an edit to the shared `.git` metadata a linked
//! worktree does not copy (its own config extension aside), can still reach
//! outside the checkout - the same as it always could for a judge or a
//! reviewer. Closing that would mean sandboxing the process itself (a
//! container, a chroot, an OS-level read-only mount), which no seat of any
//! kind in this codebase has today; adding one is a different, much larger
//! change than a design-deliberation stage, not something this module can
//! give an advisor seat on its own.
//!
//! # The draft survives every failure short of success
//!
//! [`run`] never writes to `draft` until it holds a complete, synthesized
//! replacement. Every early return - an advisor roster that produced nothing
//! usable, a planner seat that crashed or answered with no fenced `task`
//! block - leaves the interview's own draft exactly as the leader wrote it,
//! and the error names its path, the same contract [`crate::plan::vet`]
//! keeps for a validation failure. The raw advisor records are written to
//! disk unconditionally, before that check even runs, so a total failure
//! still leaves something for the operator to read.
use std::path::{Path, PathBuf};
use std::time::{Duration, Instant};

use anyhow::{Context as _, Result, bail};
use serde::{Deserialize, Serialize};

use crate::agent::{self, Invocation, SeatState};
use crate::chat;
use crate::config::{AgentSpec, Config};
use crate::git;
use crate::plan;
use crate::prompt;
use crate::verdict::{self, Proposal};

/// One advisor seat's outcome, kept even on failure so a synthesis that only
/// had two of three proposals to work with is not a mystery later - see
/// [`run`]'s doc.
#[derive(Debug, Clone, Serialize, Deserialize)]
pub struct AdvisorRecord {
    /// Seat name, e.g. `advisor-1`.
    pub seat: String,
    /// Agent id occupying the seat.
    pub agent: String,
    /// The proposal, when the seat produced a usable one.
    #[serde(default, skip_serializing_if = "Option::is_none")]
    pub proposal: Option<Proposal>,
    /// Why there is no proposal, when there is not one.
    #[serde(default, skip_serializing_if = "Option::is_none")]
    pub error: Option<String>,
    /// Wall-clock duration.
    pub duration_ms: u64,
}

/// The whole deliberation: one record per advisor seat, written to
/// `<id>.advisors.json` next to the draft so the operator can read every
/// seat's reasoning - including a seat that failed - not only whichever parts
/// synthesis kept.
#[derive(Debug, Clone, Serialize, Deserialize, Default)]
pub struct Advice {
    /// One record per advisor seat asked.
    pub records: Vec<AdvisorRecord>,
}

impl Advice {
    /// The seats that produced a usable proposal, in seat order.
    pub fn proposals(&self) -> Vec<(&str, &Proposal)> {
        self.records
            .iter()
            .filter_map(|r| r.proposal.as_ref().map(|p| (r.seat.as_str(), p)))
            .collect()
    }
}

/// Run the design-deliberation stage: gather independent proposals, write the
/// raw record, synthesize them under the planner seat, and overwrite `draft`
/// with the result.
///
/// `dir` is the drafts directory the raw record and the seats' artifacts are
/// written under; `id` names this interview's draft, so its siblings
/// (`<id>.advisors.json`, `<id>.advisors/`) sit next to `<id>.md` the same way
/// `<id>.briefing.md` already does.
pub async fn run(
    config: &Config,
    repo: &Path,
    draft: &Path,
    dir: &Path,
    id: &str,
) -> Result<Advice> {
    let requirements = std::fs::read_to_string(draft).with_context(|| {
        format!(
            "no task file at {} - the leader was asked to write one there",
            draft.display()
        )
    })?;

    let seats = config.advisors().context("resolving advisor seats")?;
    if seats.is_empty() {
        bail!(
            "`[graph] advisors` is 0, so there is nobody to deliberate with; \
             the interview draft is unchanged at {0} - file it as-is with \
             `magi task add --file {0}`, or set `[graph] advisors` above 0 \
             and re-run `magi plan`.",
            draft.display()
        );
    }

    let worktrees = checkout_worktrees(repo, dir, id, seats.len())
        .await
        .with_context(|| {
            format!(
                "could not prepare a disposable checkout for the advisor \
                 seats; the interview draft is unchanged at {d} - file it \
                 as-is with `magi task add --file {d}`, or retry `magi plan`.",
                d = draft.display(),
            )
        })?;

    // Split out so the worktrees are removed on every path out of here,
    // success or failure - Rust has no `try`/`finally` to hang this off of.
    let outcome = deliberate(
        &requirements,
        &seats,
        &worktrees,
        &DeliberationCtx {
            config,
            draft,
            dir,
            id,
            language: &config.graph.language,
            // Read + reason, no write: the same shape of work `[graph]
            // timeout_judge` already budgets for judges, so a stage-specific
            // timeout nobody asked for would be one more number to tune for
            // no benefit.
            timeout: Duration::from_secs(config.graph.timeout_judge.max(1)),
            seed: crate::rng::entropy(),
        },
    )
    .await;

    remove_worktrees(repo, &worktrees).await;

    outcome
}

/// Disposable, detached worktrees at `HEAD`, one per advisor seat - see the
/// module doc's "Disposability narrows the blast radius" section for why a
/// seat needs one of these rather than the operator's own checkout, and for
/// what it does and does not protect against.
///
/// Sequential, not parallel: `git worktree add` takes a lock on the
/// repository's own `.git` metadata, and setup is a one-time cost paid once
/// per `magi plan` invocation, not on the hot path a parallel seat wave
/// exists to keep cheap.
async fn checkout_worktrees(repo: &Path, dir: &Path, id: &str, n: usize) -> Result<Vec<PathBuf>> {
    let root = dir.join(format!("{id}.repo"));
    let mut paths = Vec::with_capacity(n);
    for i in 0..n {
        let wt = root.join(format!("advisor-{}", i + 1));
        if let Err(e) = git::worktree_add_detached(repo, &wt, "HEAD").await {
            // Partial setup must not leak the worktrees it did manage to
            // register before the failure that stopped it.
            remove_worktrees(repo, &paths).await;
            return Err(e);
        }
        paths.push(wt);
    }
    Ok(paths)
}

/// Best-effort teardown. A worktree `magi plan` fails to remove costs the
/// operator disk, not correctness - the advisor stage already answered or
/// already failed by the time this runs - so a removal error is logged and
/// moved past rather than turned into a second error on top of whatever
/// [`run`] is already returning.
async fn remove_worktrees(repo: &Path, worktrees: &[PathBuf]) {
    for wt in worktrees {
        if let Err(e) = git::worktree_remove(repo, wt).await {
            tracing::warn!(
                "could not remove disposable advisor worktree {}: {e:#}",
                wt.display()
            );
        }
    }
    // Only succeeds once every child above is gone; harmless otherwise.
    if let Some(root) = worktrees.first().and_then(|w| w.parent()) {
        let _ = std::fs::remove_dir(root);
    }
}

/// Everything [`deliberate`] needs once a disposable checkout exists per
/// seat, bundled so the function takes one borrow instead of a parameter per
/// field - the same reason [`crate::graph`]'s wave takes a `WaveCtx`.
struct DeliberationCtx<'a> {
    config: &'a Config,
    draft: &'a Path,
    dir: &'a Path,
    id: &'a str,
    language: &'a str,
    timeout: Duration,
    seed: u64,
}

/// The body of [`run`]: gather, record, synthesize, validate, write. Split out
/// only so [`run`] can guarantee `worktrees` are removed on every exit from
/// this, not so it can be called independently of a checkout existing.
async fn deliberate(
    requirements: &str,
    seats: &[AgentSpec],
    worktrees: &[PathBuf],
    ctx: &DeliberationCtx<'_>,
) -> Result<Advice> {
    let draft = ctx.draft;
    let artifacts = ctx.dir.join(format!("{}.advisors", ctx.id));

    let advice = gather(
        seats,
        requirements,
        worktrees,
        &GatherCtx {
            artifacts: &artifacts,
            run: ctx.id,
            language: ctx.language,
            timeout: ctx.timeout,
            seed: ctx.seed,
        },
    )
    .await;

    // Written before the checks below can bail: a total failure must still
    // leave the raw attempts on disk, or "nobody produced a proposal" is a
    // claim the operator has no way to check.
    let advice_path = ctx.dir.join(format!("{}.advisors.json", ctx.id));
    std::fs::write(
        &advice_path,
        serde_json::to_string_pretty(&advice).context("serialize the advisor records")?,
    )
    .with_context(|| format!("write {}", advice_path.display()))?;

    let proposals = advice.proposals();
    if proposals.is_empty() {
        bail!(
            "none of {n} advisor seat(s) produced a usable design proposal \
             (see {record}); the interview draft is unchanged at {d} - file \
             it as-is with `magi task add --file {d}`, or retry `magi plan`.",
            n = seats.len(),
            record = advice_path.display(),
            d = draft.display(),
        );
    }

    let planner = plan::pick(
        &ctx.config.agents,
        ctx.config.roles.planner.as_deref(),
        &plan::installed,
    )
    .context("resolving the planner seat for design synthesis")?;
    let mut seat = SeatState::new("plan-synthesis", &planner.id, ctx.seed);
    let synth_prompt = prompt::synthesize(requirements, &proposals, ctx.language);
    let out = agent::invoke(
        &planner,
        &mut seat,
        &Invocation {
            // The first advisor's disposable checkout, reused: every advisor
            // task has already finished by this point (`gather` awaited them
            // all), so there is nothing left to race with, and a fourth
            // checkout just for synthesis would buy nothing this one does
            // not already give it - a read-only view of the repository that
            // is not the operator's own.
            cwd: &worktrees[0],
            prompt: &synth_prompt,
            timeout: ctx.timeout,
            allow_write: false,
            sessions: false,
            artifacts: &artifacts,
            stem: "synthesis",
            run: ctx.id,
            node: "plan-advise",
            cache_dir: None,
        },
    )
    .await
    .with_context(|| {
        format!(
            "the planner seat could not synthesize the design proposals; the \
             interview draft is unchanged at {0} - file it as-is with `magi \
             task add --file {0}`, or retry `magi plan`.",
            draft.display()
        )
    })?;

    if !out.usable() {
        bail!(
            "the planner seat produced nothing usable while synthesizing the \
             design proposals; the interview draft is unchanged at {0} - file \
             it as-is with `magi task add --file {0}`, or retry `magi plan`.",
            draft.display()
        );
    }

    let synthesized = chat::extract_draft(&out.text).with_context(|| {
        format!(
            "the planner seat's reply had no fenced ```task block; the \
             interview draft is unchanged at {0} - file it as-is with `magi \
             task add --file {0}`, or retry `magi plan`.",
            draft.display()
        )
    })?;

    // Checked before a single byte reaches `draft`: `extract_draft` accepts an
    // unclosed fence as "whatever came before end of stream", so a synthesis
    // that stopped mid-sentence - a truncated reply, not a system timeout -
    // would otherwise overwrite a perfectly good interview draft with
    // something `vet` rejects one call later, at which point the operator's
    // requirements are already gone. The same shape `vet` itself uses: length
    // alone warns rather than refuses, everything else must hold.
    if let Err(problems) = plan::review_draft(&synthesized) {
        let hard: Vec<&String> = problems
            .iter()
            .filter(|p| p.as_str() != plan::SHORT_DRAFT)
            .collect();
        if !hard.is_empty() {
            let list = hard
                .iter()
                .map(|p| format!("  - {p}"))
                .collect::<Vec<_>>()
                .join("\n");
            bail!(
                "the planner seat's synthesis is not a usable task file:\n{list}\n\n\
                 the interview draft is unchanged at {d} - file it as-is with \
                 `magi task add --file {d}`, or retry `magi plan`.",
                d = draft.display(),
            );
        }
    }

    std::fs::write(draft, &synthesized).with_context(|| format!("write {}", draft.display()))?;

    Ok(advice)
}

/// The parts of [`deliberate`]'s setup every advisor seat needs, bundled so
/// [`gather`] takes one borrow instead of a parameter per field - the same
/// reason [`crate::graph`]'s wave takes a `WaveCtx`.
struct GatherCtx<'a> {
    artifacts: &'a Path,
    run: &'a str,
    language: &'a str,
    timeout: Duration,
    seed: u64,
}

/// Ask every seat for a design proposal, in parallel, headless and read-only,
/// each in its own disposable worktree (`worktrees[i]` for `seats[i]`).
///
/// Failures are per-seat, not fatal to the wave: a seat that crashes or
/// answers unparsably still produces an [`AdvisorRecord`], so one bad seat
/// does not cost the operator the other two.
async fn gather(
    seats: &[AgentSpec],
    requirements: &str,
    worktrees: &[PathBuf],
    ctx: &GatherCtx<'_>,
) -> Advice {
    let n = seats.len();
    let mut set = tokio::task::JoinSet::new();
    for (i, spec) in seats.iter().cloned().enumerate() {
        let cwd = worktrees[i].clone();
        let requirements = requirements.to_owned();
        let artifacts = ctx.artifacts.to_owned();
        let run = ctx.run.to_owned();
        let language = ctx.language.to_owned();
        let timeout = ctx.timeout;
        let seed = ctx.seed;
        let key = format!("advisor-{}", i + 1);
        set.spawn(async move {
            let mut seat = SeatState::new(&key, &spec.id, seed ^ (i as u64 + 1));
            let prompt = prompt::advisor(&requirements, i + 1, n, &language);
            let started = Instant::now();
            let outcome = agent::invoke(
                &spec,
                &mut seat,
                &Invocation {
                    cwd: &cwd,
                    prompt: &prompt,
                    timeout,
                    allow_write: false,
                    sessions: false,
                    artifacts: &artifacts,
                    stem: &key,
                    run: &run,
                    node: "plan-advise",
                    cache_dir: None,
                },
            )
            .await;
            to_record(key, spec.id, started.elapsed(), outcome)
        });
    }
    let mut records = Vec::with_capacity(n);
    while let Some(res) = set.join_next().await {
        records.push(match res {
            Ok(rec) => rec,
            Err(e) => AdvisorRecord {
                seat: "?".to_owned(),
                agent: "?".to_owned(),
                proposal: None,
                error: Some(format!("advisor task panicked: {e}")),
                duration_ms: 0,
            },
        });
    }
    // Stable seat order for a readable record: a `JoinSet` completes in
    // whichever order the seats actually answered, not seat 1, 2, 3.
    records.sort_by(|a, b| a.seat.cmp(&b.seat));
    Advice { records }
}

fn to_record(
    seat: String,
    agent_id: String,
    elapsed: Duration,
    outcome: Result<agent::AgentOutput>,
) -> AdvisorRecord {
    match outcome {
        Ok(out) if out.usable() => {
            match verdict::extract_json::<Proposal>(&out.text)
                .and_then(|p| p.validate().map(|()| p))
            {
                Ok(proposal) => AdvisorRecord {
                    seat,
                    agent: agent_id,
                    proposal: Some(proposal),
                    error: None,
                    duration_ms: out.duration_ms,
                },
                Err(e) => AdvisorRecord {
                    seat,
                    agent: agent_id,
                    proposal: None,
                    error: Some(e.to_string()),
                    duration_ms: out.duration_ms,
                },
            }
        }
        Ok(out) => AdvisorRecord {
            seat,
            agent: agent_id,
            proposal: None,
            error: Some(if out.timed_out {
                "timed out".to_owned()
            } else {
                format!("exit {:?}: {}", out.exit_code, out.text.trim())
            }),
            duration_ms: out.duration_ms,
        },
        Err(e) => AdvisorRecord {
            seat,
            agent: agent_id,
            proposal: None,
            error: Some(e.to_string()),
            duration_ms: elapsed.as_millis() as u64,
        },
    }
}

#[cfg(test)]
mod tests {
    use super::*;
    use crate::config::{AgentKind, Graph, Roles};

    /// A `kind = "command"` agent that discards its prompt and prints `output`
    /// verbatim: `cat` drains stdin so `invoke`'s writer never blocks, and the
    /// heredoc's quoted delimiter keeps `sh` from expanding anything inside
    /// `output` - the same trick a JSON block full of `{`/`}` and a `task`
    /// fence full of markdown both need.
    fn command(id: &str, output: &str) -> AgentSpec {
        AgentSpec {
            id: id.to_owned(),
            kind: AgentKind::Command,
            model: None,
            command: vec![
                "sh".to_owned(),
                "-c".to_owned(),
                format!("cat >/dev/null && cat <<'EOF'\n{output}\nEOF"),
            ],
            extra_args: Vec::new(),
            env: Default::default(),
            prompt_delivery: None,
        }
    }

    fn proposal_json(approach: &str) -> String {
        format!(
            "```json\n{{\"approach\":\"{approach}\",\"key_tradeoff\":\"t\",\
             \"risks\":[\"r\"],\"touches\":[\"src/a.rs\"],\
             \"why_not_naive\":\"w\"}}\n```"
        )
    }

    fn good_draft() -> String {
        "# Rework the config loader\n\
         \n\
         ## Context\n\
         \n\
         placeholder context.\n\
         \n\
         ## Change\n\
         \n\
         placeholder change.\n\
         \n\
         ## Constraints\n\
         \n\
         No new dependencies.\n\
         \n\
         ## Completion criteria\n\
         \n\
         - [ ] it works\n\
         \n\
         ## Out of scope\n\
         \n\
         nothing\n"
            .to_owned()
    }

    fn synthesized_task_block() -> String {
        format!(
            "```task\n{}```",
            good_draft().replace("placeholder", "synthesized")
        )
    }

    /// A real git repository with one commit, so `checkout_worktrees` has a
    /// `HEAD` to detach from. A plain temp directory is enough for the tests
    /// that fail before that point (no draft, zero advisors); only the ones
    /// that reach the disposable checkout need this.
    fn init_repo(dir: &Path) {
        let run = |args: &[&str]| {
            let out = std::process::Command::new("git")
                .args(args)
                .current_dir(dir)
                .output()
                .expect("spawn git");
            assert!(
                out.status.success(),
                "git {args:?} failed: {}",
                String::from_utf8_lossy(&out.stderr)
            );
        };
        std::fs::create_dir_all(dir).unwrap();
        run(&["init", "-b", "main"]);
        run(&["config", "user.name", "magi test"]);
        run(&["config", "user.email", "magi@example.com"]);
        std::fs::write(dir.join("README.md"), "# fixture\n").unwrap();
        run(&["add", "-A"]);
        run(&["commit", "-m", "init"]);
    }

    #[tokio::test]
    async fn gather_records_every_seat_including_one_that_fails() {
        let seats = vec![
            command("sage-a", &proposal_json("do X")),
            command("sage-b", "not json at all"),
        ];
        let dir = tempfile::tempdir().unwrap();
        let worktrees = vec![dir.path().join("wt-1"), dir.path().join("wt-2")];
        for wt in &worktrees {
            std::fs::create_dir_all(wt).unwrap();
        }
        let advice = gather(
            &seats,
            "the requirements",
            &worktrees,
            &GatherCtx {
                artifacts: &dir.path().join("artifacts"),
                run: "test-run",
                language: "en",
                timeout: Duration::from_secs(30),
                seed: 7,
            },
        )
        .await;

        assert_eq!(advice.records.len(), 2);
        assert_eq!(advice.records[0].seat, "advisor-1");
        assert_eq!(advice.records[1].seat, "advisor-2");
        let ok = advice.records[0]
            .proposal
            .as_ref()
            .expect("advisor-1 parses");
        assert_eq!(ok.approach, "do X");
        assert!(advice.records[1].proposal.is_none());
        assert!(advice.records[1].error.is_some());
    }

    fn config(agents: Vec<AgentSpec>, advisors: usize) -> Config {
        Config {
            agents,
            roles: Roles {
                advisors: vec!["sage-a".to_owned(), "sage-b".to_owned()],
                planner: Some("planner".to_owned()),
                ..Roles::default()
            },
            graph: Graph {
                advisors,
                ..Graph::default()
            },
            ..Config::default()
        }
    }

    /// This does not prove a seat *cannot* write - `allow_write: false` is
    /// not enforced by every CLI kind (opencode has no read-only mode at
    /// all), and nothing here forbids an absolute-path write either. What it
    /// proves is the gap that was unique to this stage: a relative-path
    /// write from a seat that ignores its instructions used to land in the
    /// operator's own repository - the one directory this stage must never
    /// touch - and now lands in that seat's own disposable checkout instead.
    #[tokio::test]
    async fn a_relative_path_write_from_an_advisor_lands_in_its_worktree_not_the_operators_repository()
     {
        let tmp = tempfile::tempdir().unwrap();
        let repo = tmp.path().join("repo");
        init_repo(&repo);
        let dir = tmp.path().join("drafts");
        std::fs::create_dir_all(&dir).unwrap();
        let draft = dir.join("20260906-000000-kl12.md");
        std::fs::write(&draft, good_draft()).unwrap();

        // Ignores its own read-only instruction and writes a file anyway -
        // standing in for a CLI kind (or a Bash tool) `allow_write: false`
        // does not actually stop.
        let writer = AgentSpec {
            id: "sage-a".to_owned(),
            kind: AgentKind::Command,
            model: None,
            command: vec![
                "sh".to_owned(),
                "-c".to_owned(),
                format!(
                    "cat >/dev/null && touch leaked-by-advisor.txt && cat <<'EOF'\n{}\nEOF",
                    proposal_json("do X")
                ),
            ],
            extra_args: Vec::new(),
            env: Default::default(),
            prompt_delivery: None,
        };

        let cfg = config(
            vec![
                writer,
                command("sage-b", &proposal_json("do Y")),
                command("planner", &synthesized_task_block()),
            ],
            2,
        );

        run(&cfg, &repo, &draft, &dir, "20260906-000000-kl12")
            .await
            .expect("deliberation still succeeds even though a seat wrote something");

        assert!(
            !repo.join("leaked-by-advisor.txt").exists(),
            "an advisor's write must land in its disposable worktree, never in the operator's repository"
        );
    }

    #[tokio::test]
    async fn run_writes_the_raw_records_and_overwrites_the_draft_with_the_synthesis() {
        let tmp = tempfile::tempdir().unwrap();
        let repo = tmp.path().join("repo");
        init_repo(&repo);
        let dir = tmp.path().join("drafts");
        std::fs::create_dir_all(&dir).unwrap();
        let draft = dir.join("20260906-000000-ab12.md");
        std::fs::write(&draft, good_draft()).unwrap();

        let cfg = config(
            vec![
                command("sage-a", &proposal_json("do X")),
                command("sage-b", &proposal_json("do Y")),
                command("planner", &synthesized_task_block()),
            ],
            2,
        );

        let advice = run(&cfg, &repo, &draft, &dir, "20260906-000000-ab12")
            .await
            .expect("deliberation succeeds");
        assert_eq!(advice.proposals().len(), 2);

        let advice_path = dir.join("20260906-000000-ab12.advisors.json");
        let raw = std::fs::read_to_string(&advice_path).expect("raw record on disk");
        let reread: Advice = serde_json::from_str(&raw).expect("parses back");
        assert_eq!(reread.records.len(), 2);

        let final_draft = std::fs::read_to_string(&draft).unwrap();
        assert!(
            final_draft.contains("synthesized context"),
            "the draft must be overwritten with the synthesis: {final_draft}"
        );
        assert!(final_draft.contains("## Completion criteria"));

        // The disposable checkouts must not survive a successful run - a
        // seat's worktree left behind would be exactly the write surface this
        // whole isolation exists to avoid leaving around.
        assert!(
            !dir.join("20260906-000000-ab12.repo").exists(),
            "advisor worktrees must be cleaned up after the run"
        );
    }

    #[tokio::test]
    async fn run_leaves_the_draft_untouched_when_no_advisor_produces_a_proposal() {
        let tmp = tempfile::tempdir().unwrap();
        let repo = tmp.path().join("repo");
        init_repo(&repo);
        let dir = tmp.path().join("drafts");
        std::fs::create_dir_all(&dir).unwrap();
        let draft = dir.join("20260906-000000-cd34.md");
        let original = good_draft();
        std::fs::write(&draft, &original).unwrap();

        let cfg = config(
            vec![
                command("sage-a", "garbage"),
                command("sage-b", "also garbage"),
                command("planner", &synthesized_task_block()),
            ],
            2,
        );

        let err = run(&cfg, &repo, &draft, &dir, "20260906-000000-cd34")
            .await
            .expect_err("no proposal must fail the stage");
        let msg = err.to_string();
        assert!(msg.contains(&draft.display().to_string()), "{msg}");
        assert!(msg.contains("magi task add --file"), "{msg}");
        assert_eq!(
            std::fs::read_to_string(&draft).unwrap(),
            original,
            "the interview draft must survive a total advisor failure"
        );
        // The raw attempt is still on disk, garbage and all - the operator can
        // read why every seat failed even though nothing was usable.
        assert!(dir.join("20260906-000000-cd34.advisors.json").is_file());
    }

    #[tokio::test]
    async fn run_leaves_the_draft_untouched_when_the_planner_replies_with_no_task_block() {
        let tmp = tempfile::tempdir().unwrap();
        let repo = tmp.path().join("repo");
        init_repo(&repo);
        let dir = tmp.path().join("drafts");
        std::fs::create_dir_all(&dir).unwrap();
        let draft = dir.join("20260906-000000-ef56.md");
        let original = good_draft();
        std::fs::write(&draft, &original).unwrap();

        let cfg = config(
            vec![
                command("sage-a", &proposal_json("do X")),
                command("sage-b", &proposal_json("do Y")),
                command("planner", "sure, here is my answer with no fence"),
            ],
            2,
        );

        let err = run(&cfg, &repo, &draft, &dir, "20260906-000000-ef56")
            .await
            .expect_err("a synthesis with no task block must fail the stage");
        let msg = err.to_string();
        assert!(msg.contains(&draft.display().to_string()), "{msg}");
        assert_eq!(std::fs::read_to_string(&draft).unwrap(), original);
    }

    /// Reported: `chat::extract_draft` accepts an unclosed fence as whatever
    /// came before end of stream, so a planner reply that stops mid-sentence
    /// - not a timeout, `out.usable()` is still true - used to overwrite a
    /// perfectly good interview draft with a stub `vet` then rejected one
    /// call later, by which point the original requirements were gone.
    #[tokio::test]
    async fn run_leaves_the_draft_untouched_when_the_synthesis_fence_never_closes() {
        let tmp = tempfile::tempdir().unwrap();
        let repo = tmp.path().join("repo");
        init_repo(&repo);
        let dir = tmp.path().join("drafts");
        std::fs::create_dir_all(&dir).unwrap();
        let draft = dir.join("20260906-000000-ij90.md");
        let original = good_draft();
        std::fs::write(&draft, &original).unwrap();

        let cfg = config(
            vec![
                command("sage-a", &proposal_json("do X")),
                command("sage-b", &proposal_json("do Y")),
                command("planner", "```task\n# incomplete"),
            ],
            2,
        );

        let err = run(&cfg, &repo, &draft, &dir, "20260906-000000-ij90")
            .await
            .expect_err("an incomplete synthesis must not become the task file");
        let msg = err.to_string();
        assert!(msg.contains(&draft.display().to_string()), "{msg}");
        assert!(msg.contains("not a usable task file"), "{msg}");
        assert_eq!(
            std::fs::read_to_string(&draft).unwrap(),
            original,
            "the interview draft must survive an incomplete synthesis"
        );
    }

    #[tokio::test]
    async fn run_reports_a_missing_draft_against_the_path_the_leader_was_given() {
        let tmp = tempfile::tempdir().unwrap();
        let dir = tmp.path().join("drafts");
        std::fs::create_dir_all(&dir).unwrap();
        let draft = dir.join("never-written.md");

        let cfg = config(vec![command("sage-a", &proposal_json("x"))], 1);
        let msg = run(&cfg, tmp.path(), &draft, &dir, "never-written")
            .await
            .expect_err("nothing to deliberate over")
            .to_string();
        assert!(msg.contains(&draft.display().to_string()), "{msg}");
    }

    #[tokio::test]
    async fn zero_advisors_is_an_error_that_still_names_the_draft() {
        let tmp = tempfile::tempdir().unwrap();
        let dir = tmp.path().join("drafts");
        std::fs::create_dir_all(&dir).unwrap();
        let draft = dir.join("20260906-000000-gh78.md");
        std::fs::write(&draft, good_draft()).unwrap();

        let cfg = config(vec![command("sage-a", &proposal_json("x"))], 0);
        let msg = run(&cfg, tmp.path(), &draft, &dir, "20260906-000000-gh78")
            .await
            .expect_err("nobody to deliberate with")
            .to_string();
        assert!(msg.contains("advisors` is 0"), "{msg}");
        assert!(msg.contains(&draft.display().to_string()), "{msg}");
    }
}