Skip to main content

magi/
conduct.rs

1//! The conductor: a single agent seat that arranges the queue.
2//!
3//! [`crate::queue`] orders runnable tasks by `priority` alone; nothing in it
4//! can express that one task should wait for another, or that a task whose
5//! last run stopped short deserves a second look before the loop blindly
6//! retries it. Once per polling cycle, [`Conductor::maybe_run`] shows one
7//! agent seat three things - the runnable tasks, the tasks stuck `running`
8//! with no live daemon behind them, and the `failed`/`held` tasks nobody has
9//! decided about yet - and asks it to decide `blocked_by` for the first and a
10//! recovery for the other two. Everything else about the loop - which
11//! unblocked task runs next, one at a time, in `priority` order - is
12//! unchanged; see `crate::daemon`.
13//!
14//! # What the conductor may not do
15//!
16//! [`Decision`] has no field for `priority`, for deleting a task, or for
17//! touching git, a worktree, or a branch directly. [`Recovery::Review`] only
18//! ever reopens a branch `crate::daemon` itself resolved from the task's own
19//! run record ([`surviving_branch`]) - never a name the model wrote - through
20//! `crate::graph::Runner::review`, which reviews and verifies but never
21//! rewrites history.
22//!
23//! # Non-blocking by construction
24//!
25//! [`crate::ask::ask_and_wait`] is never called from here, and the prompt
26//! tells the model the same: that CLI command blocks until a human answers,
27//! and calling it from inside the conductor's own invocation would park the
28//! whole polling loop behind one task's question. Instead a decision that
29//! wants the operator's judgement carries a `question` field, and [`apply`]
30//! files it with [`Question::new`] and [`Questions::put`] and moves on in the
31//! same call. Something still has to read the owner's reply: the question
32//! carries a [`Deputy`] brief, and `magi serve` starts a short-lived seat of
33//! its own for it ([`crate::deputy`]) - never this loop.
34//!
35//! # Fails soft, always
36//!
37//! [`Conductor::maybe_run`] never returns an error: an unusable roster, a
38//! timed-out invocation, or a reply [`verdict::extract_json`] cannot parse are
39//! all logged and treated as "this cycle changes nothing." `crate::daemon`'s
40//! loop always falls through to its own `Queue::next_runnable` regardless of
41//! what happened here.
42
43use std::collections::{BTreeMap, BTreeSet};
44use std::path::{Path, PathBuf};
45use std::time::Duration;
46
47use anyhow::{Context as _, Result};
48use serde::{Deserialize, Serialize};
49
50use crate::agent::{self, Invocation, SeatState};
51use crate::ask::{ChoiceAction, Deputy, Question, Questions};
52use crate::config::Config;
53use crate::prompt;
54use crate::queue::{Queue, Task, TaskStatus};
55use crate::run::RunState;
56use crate::verdict;
57
58/// Seat name for the conductor's own CLI-side conversation, scoped away from
59/// every other seat magi ever opens - the same rule every other seat follows.
60const SEAT: &str = "conduct";
61
62/// Node name reported to the invoked agent (`MAGI_NODE`) and recorded on any
63/// question it files, so an operator reading the questions list can tell a
64/// conductor's question from one a run's own agent asked.
65pub const NODE: &str = "conduct";
66
67/// Wall-clock limit for one conductor turn. The conductor reads a queue
68/// listing and replies with json; it does not implement anything or run a
69/// build, so this is short - the same order of magnitude as
70/// `crate::chat`'s own single-turn, no-write invocations.
71const TURN_TIMEOUT: Duration = Duration::from_secs(300);
72
73/// What the conductor may choose for a `running`-but-stalled or a
74/// `failed`/`held` task.
75#[derive(Debug, Clone, Copy, PartialEq, Eq, Deserialize)]
76#[serde(rename_all = "lowercase")]
77pub enum Recovery {
78    /// Put it back in line, attempts reset - the same effect as
79    /// `magi task release`.
80    Requeue,
81    /// Leave it for a human, unchanged otherwise - the same effect as
82    /// `magi task hold`.
83    Hold,
84    /// Reopen the task's surviving branch as a review-only pass
85    /// (`crate::graph::Runner::review`) instead of competing from scratch.
86    /// Only takes effect when [`surviving_branch`] can actually name one;
87    /// otherwise `crate::daemon` falls back to [`Recovery::Requeue`].
88    Review,
89    /// Close the task outright, as [`Task::succeed`] - the same effect as
90    /// `magi task done`. For a `failed`/`held` task only, never `running`:
91    /// this is for a task whose own goal is already known to be met outside
92    /// the loop entirely (the branch was merged and the worktree cleaned up
93    /// by hand, say) and competing it again would only spend attempts on
94    /// work that has nothing left to do, not for a task that merely stopped
95    /// mid-competition and might still need to run.
96    Done,
97}
98
99/// The operation for which the operator's decision is requested. These are
100/// fixed protocol values, independent of the language of the displayed choices.
101#[derive(Debug, Clone, Copy, PartialEq, Eq, Deserialize, Serialize)]
102#[serde(rename_all = "snake_case")]
103pub enum QuestionOperation {
104    /// Choose an existing implementation as the recovery source.
105    Recover,
106    /// Continue an existing run from its saved state.
107    Resume,
108    /// Authorize repairing existing work before continuing.
109    Repair,
110    /// Confirm that the task may be closed as complete.
111    Complete,
112    /// Request an operator inspection of the target.
113    Inspect,
114    /// Request a configuration change for the target.
115    Configure,
116}
117
118/// The concrete object of an operator decision. Sources resolve to commit ids;
119/// runs must belong to this task. A generated prose label is never an identity.
120#[derive(Debug, Clone, Deserialize)]
121#[serde(tag = "kind", rename_all = "snake_case")]
122pub enum QuestionTarget {
123    /// The task itself, for decisions without a more specific recorded object.
124    Task,
125    /// A run already owned by this task.
126    Run {
127        /// Full recorded run id.
128        id: String,
129    },
130    /// A concrete TOML setting. Different fields on the same task require
131    /// different decisions even when both operations are `configure`.
132    Setting {
133        /// Repository-relative TOML file containing the setting.
134        path: String,
135        /// Exact table/key components copied from that file.
136        keys: Vec<String>,
137    },
138    /// A recovery source which the repository can resolve to a commit.
139    Source {
140        /// Existing branch name or commit reference, never a display label.
141        reference: String,
142    },
143}
144
145/// The structured meaning of a question, separate from its displayed wording.
146#[derive(Debug, Clone, Deserialize)]
147pub struct QuestionDecision {
148    /// The concrete type of operator decision requested.
149    pub operation: QuestionOperation,
150    /// The recorded object the decision concerns.
151    pub target: QuestionTarget,
152}
153
154impl QuestionDecision {
155    fn identity(&self, task: &Task) -> Result<(QuestionOperation, String)> {
156        let target = match &self.target {
157            QuestionTarget::Task => format!("task:{}", task.id),
158            QuestionTarget::Run { id } => {
159                anyhow::ensure!(
160                    task.runs.contains(id),
161                    "question target is not this task's run"
162                );
163                format!("run:{id}")
164            }
165            QuestionTarget::Setting { path, keys } => {
166                use std::path::Component;
167                anyhow::ensure!(
168                    !path.is_empty()
169                        && Path::new(path)
170                            .components()
171                            .all(|c| matches!(c, Component::Normal(_)))
172                        && !keys.is_empty()
173                        && keys.iter().all(|key| !key.is_empty()),
174                    "question setting must name a repository file and concrete keys"
175                );
176                let root = task.repo.canonicalize()?;
177                let file = root.join(path).canonicalize()?;
178                let relative = file
179                    .strip_prefix(&root)
180                    .context("question setting is outside the repository")?;
181                let content = std::fs::read_to_string(&file)?;
182                let value: toml::Value = toml::from_str(&content)?;
183                let mut setting = &value;
184                for key in keys {
185                    setting = setting
186                        .get(key)
187                        .context("question setting key does not exist")?;
188                }
189                format!("setting:{}", serde_json::to_string(&(relative, keys))?)
190            }
191            QuestionTarget::Source { reference } => {
192                use crate::proc::Quiet as _;
193                let output = std::process::Command::new("git")
194                    .args([
195                        "rev-parse",
196                        "--verify",
197                        "--end-of-options",
198                        &format!("{reference}^{{commit}}"),
199                    ])
200                    .current_dir(&task.repo)
201                    .quiet()
202                    .output()
203                    .context("resolve question source")?;
204                anyhow::ensure!(
205                    output.status.success(),
206                    "question source cannot be resolved"
207                );
208                let commit = String::from_utf8(output.stdout)?.trim().to_owned();
209                anyhow::ensure!(
210                    !commit.is_empty() && commit.bytes().all(|c| c.is_ascii_hexdigit()),
211                    "invalid question source commit"
212                );
213                format!("source:{commit}")
214            }
215        };
216        Ok((self.operation, target))
217    }
218}
219
220/// The conductor's decision for one task. Deliberately has no `priority`
221/// field: see this module's doc.
222#[derive(Debug, Clone, Default, Deserialize)]
223pub struct Decision {
224    /// Task id, expected to be copied verbatim from what it was shown.
225    pub id: String,
226    /// Ids this task should be blocked on. Meaningful only for a runnable
227    /// (`queued`) task, or a machine-held task whose missing-only decision
228    /// is unconfirmed against the queue. Ignored otherwise.
229    #[serde(default)]
230    pub blocked_by: Vec<String>,
231    /// Dependencies claimed missing as the sole reason for this question or
232    /// hold. Must name references in `blocked_by`. Empty (including legacy
233    /// replies) means the question/hold may have an independent purpose and
234    /// must be preserved. Never set for decisions needing separate consent.
235    #[serde(default)]
236    pub missing_dependencies: Vec<String>,
237    /// One line explaining the block or the recovery.
238    #[serde(default)]
239    pub reason: Option<String>,
240    /// Recovery for a stalled or finished task. For a runnable (`queued`)
241    /// one, only [`Recovery::Hold`] has any effect. An independent hold takes
242    /// precedence over `blocked_by`, so dependency completion cannot release it.
243    /// Use a hold when the operator already said the task must not compete,
244    /// rather than asking a `question` that restates that answer. `Requeue`, `Review`,
245    /// and `Done` have no meaning for a task that is already in line.
246    #[serde(default)]
247    pub recovery: Option<Recovery>,
248    /// A question for the operator. When present, [`apply`] files it (unless
249    /// one is already open for this task) and blocks the task on its id
250    /// instead of acting on `blocked_by` or `recovery`. Only a missing-only
251    /// decision explicitly marked by [`Self::missing_dependencies`] may be
252    /// replaced with dependency waiting unless the queue confirms missing blockers.
253    #[serde(default)]
254    pub question: Option<String>,
255    /// Fixed answers for `question`, if it has any. Empty means free text.
256    #[serde(default)]
257    pub choices: Vec<String>,
258    /// Explicit operations keyed by exact choice labels; never inferred from prose.
259    #[serde(default)]
260    pub actions: BTreeMap<String, ChoiceAction>,
261    /// Stable operation and concrete target of a question. Missing on legacy
262    /// decisions: those share the task's existing cause, regardless of prose.
263    #[serde(default)]
264    pub question_decision: Option<QuestionDecision>,
265    /// Task or question ids carrying independently verifiable new blocking evidence.
266    #[serde(default)]
267    pub evidence: Vec<String>,
268}
269
270/// The conductor's whole reply for one cycle.
271///
272/// `decisions` is deliberately **not** `#[serde(default)]`, unlike every
273/// other field in this module. [`verdict::extract_json`] disambiguates
274/// between several balanced `{...}` spans in one reply by trying the type the
275/// caller wants against each of them, last first, and keeping the first that
276/// fits - which only works when a span that is not really the answer can
277/// fail to fit. A `Verdict` with no required field at all would make every
278/// span fit, including a `{}` left by stray trailing prose, and the reply's
279/// real `decisions` - earlier in the text - would never be reached. Requiring
280/// the key costs nothing: the prompt already asks for it on every reply, `[]`
281/// included.
282#[derive(Debug, Clone, Default, Deserialize)]
283pub struct Verdict {
284    /// One entry per task the conductor chose to say something about. A task
285    /// left out of this list is left exactly as it was.
286    pub decisions: Vec<Decision>,
287}
288
289/// A view of a task built for [`prompt::conduct`], shared by the runnable and
290/// stalled sections.
291fn view(t: &Task, max_attempts: usize) -> prompt::ConductTask {
292    prompt::ConductTask {
293        id: t.id.clone(),
294        title: t.title.clone(),
295        instruction: t.instruction.clone(),
296        repo: t.repo.display().to_string(),
297        priority: t.priority,
298        status: t.status.as_str().to_owned(),
299        attempts: t.attempts,
300        max_attempts: t.effective_attempt_limit(max_attempts),
301        last_error: t.last_error.clone(),
302        hold_reason: t.hold_reason.clone(),
303        hold_source: t.hold_source.map(|source| source.label().to_owned()),
304        blocked_by: t.blocked_by.clone(),
305        answers: t
306            .answers
307            .iter()
308            .map(|a| prompt::ConductAnswer {
309                question: a.question.clone(),
310                answer: a.answer.clone(),
311            })
312            .collect(),
313        operator_resume: t.resumption.as_ref().map(|r| format!(
314            "Explicit operator resumption generation {} via {} at {}. Approved instruction: {}. \
315             The first {} historical answers and prior hold ({:?}) are superseded as waiting premises. \
316             To block again cite evidence ids from a later owned question, changed dependency, or new failed run; \
317             wording alone is insufficient.",
318            r.generation, r.source, r.at, r.instruction, r.answers_through, r.cleared_reason
319        )).or_else(|| t.resume_override.as_ref().map(|o| {
320            format!(
321                "the operator explicitly answered \"resume\" at {}; do not hold this \
322                 task again for the same reason unless there is new information",
323                o.at
324            )
325        })),
326    }
327}
328
329/// Did the operator answer a `magi ask` choice with a `resume` action that
330/// this task has not yet acted on? Then `Recovery::Requeue` must not turn the
331/// task into a fresh competition either.
332fn pinned_resume(task: &Task) -> bool {
333    task.resume_override
334        .as_ref()
335        .is_some_and(|o| o.pinned_run.is_some())
336}
337
338/// Verify references independently of the model's reason text. Old evidence
339/// cannot become new merely by being described differently.
340fn new_blocking_evidence(
341    queue: &Queue,
342    questions: &Questions,
343    task: &Task,
344    d: &Decision,
345) -> Vec<String> {
346    let Some(r) = &task.resumption else {
347        return Vec::new();
348    };
349    d.evidence
350        .iter()
351        .filter_map(|id| {
352            let version = if let Ok(q) = questions.get(id) {
353                (q.asked_at > r.at
354                    && (q.run == task.id || task.runs.contains(&q.run))
355                    && q.node != NODE
356                    && q.status != crate::ask::QuestionStatus::Abandoned)
357                    .then(|| {
358                        format!(
359                            "question:{id}:{}:{:?}:{}",
360                            q.asked_at,
361                            q.status,
362                            q.resolution().unwrap_or_default()
363                        )
364                    })
365            } else if d.blocked_by.contains(id) {
366                queue
367                    .get(id)
368                    .ok()
369                    .filter(|dep| {
370                        dep.id != task.id
371                            && dep.status != TaskStatus::Done
372                            && (dep.created_at > r.at
373                                || dep.held_at.is_some_and(|at| at > r.at)
374                                || dep.operator_decided_at.is_some_and(|at| at > r.at)
375                                || dep.passes.iter().any(|p| p.settled_at > r.at)
376                                || (dep.status == TaskStatus::Blocked
377                                    && dep.blocked_at.is_some_and(|at| at > r.at)))
378                    })
379                    .map(|dep| {
380                        format!(
381                            "dependency:{id}:{:?}:{:?}:{:?}",
382                            dep.status,
383                            dep.chat_notice_key(),
384                            dep.hold_notice_cause
385                        )
386                    })
387            } else if task.runs.iter().any(|run| run == id) {
388                RunState::load(id)
389                    .ok()
390                    .filter(|run| {
391                        // A resumed run keeps its id. Only a pass started after
392                        // approval (or a newly created legacy run) is new evidence.
393                        run.updated_at > r.at
394                            && run
395                                .passes
396                                .last()
397                                .map_or(run.created_at > r.at, |p| p.started_at > r.at)
398                            && matches!(
399                                run.status,
400                                crate::run::RunStatus::Failed
401                                    | crate::run::RunStatus::Blocked
402                                    | crate::run::RunStatus::Stalled
403                            )
404                    })
405                    .map(|run| {
406                        format!(
407                            "run:{id}:{:?}:{:?}",
408                            run.status,
409                            run.passes.last().map(|p| p.index)
410                        )
411                    })
412            } else {
413                None
414            };
415            version.filter(|v| !r.blocking_evidence.contains(v))
416        })
417        .collect()
418}
419
420/// Durable approval is enforced by verified evidence at the mutation boundary.
421/// A pinned run stays protected; legacy transient overrides keep their original
422/// bounded contradiction policy until an explicit new approval replaces it.
423fn may_hold(task: &mut Task, reason: &str) -> bool {
424    if pinned_resume(task) {
425        return false;
426    }
427    if task.resumption.is_some() {
428        return true; // The mutation boundary already verified new evidence.
429    }
430    let Some(o) = task.resume_override.as_mut() else {
431        return true;
432    };
433    if o.forced || o.pinned_run.is_some() {
434        tracing::warn!(
435            "conductor tried to hold task {} after the operator forced a resume: {reason}",
436            task.id
437        );
438        return false;
439    }
440    if o.conductor_rehold.is_some() {
441        return false;
442    }
443    o.conductor_rehold = Some(reason.to_owned());
444    true
445}
446
447/// Severity as a lowercase word, matching how `crate::verdict::Severity` is
448/// spelled everywhere else an operator or a model reads it.
449fn severity_str(s: crate::verdict::Severity) -> &'static str {
450    match s {
451        crate::verdict::Severity::Nit => "nit",
452        crate::verdict::Severity::Minor => "minor",
453        crate::verdict::Severity::Major => "major",
454        crate::verdict::Severity::Blocker => "blocker",
455    }
456}
457
458/// The branch a task's last run left behind, if the tally ever ran on it -
459/// what [`Recovery::Review`] reopens. Derived from `crate::run::RunState`
460/// alone, never from anything the conductor wrote, so a hallucinated branch
461/// name can never reach `crate::graph::Runner::review`.
462fn surviving_branch(task: &Task) -> Option<String> {
463    let last = task.runs.last()?;
464    let state = RunState::load(last).ok()?;
465    if !crate::daemon::compatible_run_instruction(&state, task) {
466        return None;
467    }
468    state.winner().map(|c| c.branch.clone())
469}
470
471/// The reason to record when the conductor holds a `failed`/`held` task via
472/// [`Recovery::Hold`]. Always `Some`, never a bare `d.reason.clone()`.
473///
474/// [`Task::hold_machine`] only overwrites [`Task::hold_reason`] when given
475/// one, precisely so a stalled task's existing diagnosis survives a
476/// decision that has nothing new to add. That is the right default when the
477/// task was not already `held` - but a task that *is* already `held`, and
478/// is being held again here, is a different case: its current
479/// `hold_reason` may still read as a cause `crate::triage` knows how to
480/// re-check on its own (a disk-pressure message, say - see
481/// `triage::machine_cause_resolved`), and if the model gives no reason,
482/// leaving that text untouched would make this decision - the conductor
483/// choosing, informed by the operator's own answer, to keep the task held
484/// anyway - indistinguishable from the original, never-reconsidered hold.
485/// `crate::triage` would then read the stale text the next time its one
486/// recognised cause looks resolved and release the task straight through
487/// the decision this call was recording. Prepending (not appending) the new
488/// note keeps the old text as context without leaving the string starting
489/// with whatever pattern `crate::triage` matched before.
490///
491/// Two bounds keep the string from growing, which it once did without limit
492/// (every re-hold nested the whole previous reason one level deeper, and every
493/// change relit the task's notice): a note equal to the current reason - or to
494/// its outermost note - adds nothing, so the result is `None` and the caller
495/// writes nothing; and only ONE `(previously: ...)` level is ever kept, the
496/// prior being cut back to its outermost note before it is embedded.
497fn reaffirmed_hold_reason(task: &Task, d: &Decision) -> Option<String> {
498    let note = match &d.reason {
499        Some(reason) => reason.clone(),
500        None => match task.answers.last() {
501            Some(a) => format!(
502                "conduct held this again with no new reason given; last operator \
503                 answer on record: {}",
504                a.answer
505            ),
506            None => "conduct held this again with no reason given".to_owned(),
507        },
508    };
509    match task.hold_reason.as_deref() {
510        Some(prior) if !prior.is_empty() => {
511            let outer = outermost_hold_note(prior);
512            if note.trim() == prior.trim() || note.trim() == outer {
513                None
514            } else {
515                Some(format!("{note}\n\n(previously: {outer})"))
516            }
517        }
518        _ => Some(note),
519    }
520}
521
522/// Separator [`reaffirmed_hold_reason`] puts between a note and the reason it
523/// replaced.
524const PREVIOUSLY: &str = "\n\n(previously: ";
525
526/// A hold reason without any `(previously: ...)` tail: the text before the
527/// first separator, trimmed.
528fn outermost_hold_note(reason: &str) -> &str {
529    reason.split(PREVIOUSLY).next().unwrap_or(reason).trim()
530}
531
532/// The conductor's stated reason for a hold, or a stand-in when it gave none.
533fn hold_note(d: &Decision) -> String {
534    d.reason
535        .clone()
536        .unwrap_or_else(|| "(no reason given)".to_owned())
537}
538
539/// Everything the conductor is shown about a `failed`/`held` task's last run.
540async fn outcome_for(task: &Task, repo: &Path) -> prompt::ConductOutcome {
541    let Some(run_id) = task.runs.last().cloned() else {
542        return prompt::ConductOutcome {
543            run_id: "(none)".to_owned(),
544            unreadable: Some("this task has not produced a run yet".to_owned()),
545            run_status: None,
546            open_findings: Vec::new(),
547            rounds_used: 0,
548            rounds_max: 0,
549            rounds: Vec::new(),
550            branch: None,
551            branch_head: None,
552            references: None,
553            empty_candidate: false,
554        };
555    };
556    let state = match RunState::load(&run_id) {
557        Ok(s) => s,
558        Err(e) => {
559            // The exact failure this feature exists to stop hiding: a schema
560            // mismatch (or any other unreadable state) must never be treated
561            // as "nothing to recover" - it is surfaced here, verbatim, rather
562            // than swallowed into a quiet re-competition.
563            tracing::warn!(
564                "conductor: could not read run {run_id} for task {}: {e:#}",
565                task.short()
566            );
567            return prompt::ConductOutcome {
568                run_id,
569                unreadable: Some(format!("{e:#}")),
570                run_status: None,
571                open_findings: Vec::new(),
572                rounds_used: 0,
573                rounds_max: 0,
574                rounds: Vec::new(),
575                branch: None,
576                branch_head: None,
577                references: None,
578                empty_candidate: false,
579            };
580        }
581    };
582
583    let finding_view = |f: &crate::verdict::Finding| prompt::ConductFinding {
584        id: f.id.clone(),
585        title: f.title.clone(),
586        severity: severity_str(f.severity).to_owned(),
587    };
588    let open_findings = state
589        .open_findings()
590        .into_iter()
591        .map(finding_view)
592        .collect();
593    let rounds = state
594        .reviews
595        .iter()
596        .map(|r| prompt::ConductRound {
597            round: r.round,
598            findings: r
599                .reviews
600                .iter()
601                .flat_map(|rec| rec.findings.iter())
602                .map(finding_view)
603                .collect(),
604            addressed: r
605                .fix
606                .as_ref()
607                .map(|fx| fx.addressed.clone())
608                .unwrap_or_default(),
609            rejected: r
610                .fix
611                .as_ref()
612                .map(|fx| {
613                    fx.rejected
614                        .iter()
615                        .map(|rej| prompt::ConductRejection {
616                            id: rej.id.clone(),
617                            why: rej.why.clone(),
618                        })
619                        .collect()
620                })
621                .unwrap_or_default(),
622        })
623        .collect();
624    let branch = state.winner().map(|c| c.branch.clone());
625    let branch_head = match &branch {
626        Some(b) => crate::git::rev_parse(repo, b)
627            .await
628            .ok()
629            .map(|h| h.chars().take(8).collect()),
630        None => None,
631    };
632
633    prompt::ConductOutcome {
634        run_id,
635        unreadable: None,
636        run_status: Some(state.status.as_str().to_owned()),
637        open_findings,
638        rounds_used: state.reviews.len(),
639        rounds_max: state.config.graph.review_rounds,
640        rounds,
641        branch,
642        branch_head,
643        references: crate::refs::describe(&state.seeds),
644        empty_candidate: state.merge.as_ref().is_some_and(|m| m.empty),
645    }
646}
647
648/// Whether the task's latest run's winner branch already has an open pull
649/// request into the run's own base - a fact the repository can answer, so the
650/// operator is never asked it. `None` when there is no run or winner to check.
651async fn open_pr_fact(task: &Task, repo: &Path) -> Option<String> {
652    let id = task.runs.last()?;
653    let state = match RunState::load(id) {
654        Ok(s) => s,
655        Err(e) => return Some(format!("could not check open pull requests: {e:#}")),
656    };
657    let branch = state.winner()?.branch.clone();
658    let base = state.base_branch.clone();
659    match crate::land::find_open_pr(repo, &branch, &base).await {
660        Ok(crate::land::OpenPr::None) => None,
661        Ok(crate::land::OpenPr::One { url, .. }) => Some(format!(
662            "pull request {url} is already open for {branch} into {base}"
663        )),
664        Ok(crate::land::OpenPr::Many(urls)) => Some(format!(
665            "several pull requests are already open for {branch} into {base}: {}",
666            urls.join(" ")
667        )),
668        Err(e) => Some(format!("could not check open pull requests: {e:#}")),
669    }
670}
671
672/// Append what the repository says about the branches and commits a task
673/// names to every question the conductor is about to file, so the operator is
674/// never asked a fact git can answer ("is it already on main?"). Done here,
675/// in code, rather than trusted to the model: a question the model wrote
676/// without checking still reaches the operator with the answer attached.
677/// When it cannot be checked the question says so instead of guessing.
678async fn attach_facts(cfg: &Config, repo: &Path, queue: &Queue, verdict: &mut Verdict) {
679    for d in verdict
680        .decisions
681        .iter_mut()
682        .filter(|d| d.question.is_some())
683    {
684        let Ok(task) = queue.get(&d.id) else {
685            continue;
686        };
687        let text = format!("{}\n{}", task.title, task.instruction);
688        let pr_fact = open_pr_fact(&task, repo_for(&task, repo).as_path()).await;
689        if crate::refs::scan(&text).is_empty() {
690            if let (Some(fact), Some(q)) = (pr_fact, d.question.as_mut()) {
691                q.push_str(&format!(
692                    "\n\nChecked against the repository (magi did this, not the model):\n{fact}"
693                ));
694            }
695            continue;
696        }
697        let repo = repo_for(&task, repo);
698        let remote = &cfg.merge.remote;
699        let base = crate::git::merge_base_branch(&repo, remote, cfg.merge.base.as_deref())
700            .await
701            .ok();
702        let facts = match base {
703            Some(base) => {
704                let base_name = base.clone();
705                let tracking = format!("{remote}/{base}");
706                // Best effort: a stale tracking ref would report work that
707                // has since landed as still missing from the base.
708                let refreshed = crate::git::fetch(&repo, remote, &base)
709                    .await
710                    .is_ok_and(|o| o.ok());
711                // Only the remote's tip counts: a local branch of the same
712                // name is whatever the checkout happens to hold.
713                let against = tracking;
714                match crate::git::rev_parse(&repo, &against).await {
715                    Ok(tip) => crate::refs::describe(
716                        &crate::refs::resolve(&repo, &tip, remote, &text).await,
717                    )
718                    .map(|facts| {
719                        if refreshed {
720                            facts
721                        } else {
722                            format!(
723                                "{facts}\n(could not fetch {remote}/{base_name}: this is \
724                                 against the last fetched `{against}`, which may be behind \
725                                 the remote)"
726                            )
727                        }
728                    }),
729                    Err(_) => Some(format!(
730                        "could NOT be checked against the repository: {against} is not \
731                         available here, so whether the work is in the base is unknown"
732                    )),
733                }
734            }
735            None => Some("could not check the repository: no base branch known".to_owned()),
736        };
737        let facts = match (facts, pr_fact) {
738            (Some(f), Some(p)) => Some(format!("{f}\n{p}")),
739            (f, p) => f.or(p),
740        };
741        if let (Some(facts), Some(q)) = (facts, d.question.as_mut()) {
742            q.push_str(&format!(
743                "\n\nChecked against the repository (magi did this, not the model):\n{facts}"
744            ));
745        }
746    }
747}
748
749/// A `failed`/`held` task together with how its last run ended.
750async fn finished_view(t: &Task, repo: &Path, max_attempts: usize) -> prompt::ConductFinished {
751    prompt::ConductFinished {
752        task: view(t, max_attempts),
753        outcome: outcome_for(t, &repo_for(t, repo)).await,
754    }
755}
756
757/// The repository containing a task's branch. A task filed without a
758/// repository uses the daemon's repository, exactly as its later attempt does.
759fn repo_for(task: &Task, fallback: &Path) -> PathBuf {
760    if task.repo.as_os_str().is_empty() || task.repo == Path::new(".") {
761        fallback.to_path_buf()
762    } else {
763        task.repo.clone()
764    }
765}
766
767/// Apply one decision to the queue and the question store.
768///
769/// Takes the task's own claim before touching it: a model call spans a whole
770/// agent turn, and the queue can have moved on by the time its answer comes
771/// back. A claim that cannot be taken means something else owns this task
772/// right now - most often a live daemon mid-competition on it - so the
773/// conductor's now-stale view of it is dropped rather than raced against; see
774/// `crate::queue::Queue::claim`'s own doc on why a claim is proof, not a
775/// guess.
776///
777/// A decision is matched against the task's *current* status, re-read under
778/// the claim, not against whichever section of the prompt it came from: a
779/// `blocked_by` only takes effect on a `queued` task, and `recovery` only on
780/// one `running` (stalled) or `failed`/`held`, so a decision that no longer
781/// matches what the task actually is - it moved on between the read that
782/// built the prompt and this write - changes nothing.
783fn apply_one(
784    queue: &Queue,
785    questions: &Questions,
786    d: &Decision,
787    snapshot: Option<&Task>,
788) -> Result<()> {
789    let _claim = queue
790        .claim(&d.id)
791        .with_context(|| format!("task {} is claimed elsewhere right now", d.id))?;
792    let mut task = queue.get(&d.id).context("no such task")?;
793
794    // Compare the actual input, not a generation echoed by the model. Even
795    // question creation must be discarded after an intervening operator edit.
796    if snapshot.is_some_and(|old| decision_input(old) != decision_input(&task)) {
797        return Ok(());
798    }
799    let blocking =
800        d.question.is_some() || !d.blocked_by.is_empty() || d.recovery == Some(Recovery::Hold);
801    let evidence = new_blocking_evidence(queue, questions, &task, d);
802    if blocking && task.resumption.is_some() && evidence.is_empty() {
803        crate::notices::raise(crate::notices::conductor_resumption(
804            &task.id,
805            task.short(),
806            d.reason.as_deref(),
807            &d.blocked_by,
808            &d.evidence,
809            d.question.is_some(),
810        ));
811        return Ok(());
812    }
813
814    if blocking && !evidence.is_empty() {
815        let mut cause = evidence.clone();
816        cause.sort();
817        cause.dedup();
818        task.notice_cause(
819            &format!("conductor-evidence:{}", cause.join(",")),
820            None,
821            None,
822        );
823    }
824    if blocking {
825        if let Some(r) = task.resumption.as_mut() {
826            r.blocking_evidence.extend(evidence);
827        }
828    }
829
830    // A conductor answer is never operator authorization.  In particular,
831    // do this before questions and blocking too: either would reclassify a
832    // manual hold and let a later deterministic resolver queue it.
833    if task.operator_held() {
834        return Ok(());
835    }
836
837    // A claimed task may have deferred the pre-scheduling triage sweep.
838    // Leave an open or answered-unapplied question to that sweep rather
839    // than move the task out from under it.
840    if task.status == TaskStatus::Held && crate::triage::pending_for(questions, &task) {
841        return Ok(());
842    }
843
844    // The operator's `resume` action is waiting for its turn: a question or a
845    // dependency would park the task before the daemon can run it.
846    let pinned = pinned_resume(&task);
847    if pinned && (d.question.is_some() || !d.blocked_by.is_empty()) {
848        return Ok(());
849    }
850
851    // A missing-only decision with unverifiable aliases is not evidence for
852    // a question or hold. Do not guess that two ambiguous/unknown references
853    // mean the same record, or treat failure to prove that as proof of absence.
854    if matches!(task.status, TaskStatus::Queued | TaskStatus::Held)
855        && !d.missing_dependencies.is_empty()
856        && (d.question.is_some() || d.recovery == Some(Recovery::Hold))
857        && d.missing_dependencies
858            .iter()
859            .any(|id| !d.blocked_by.contains(id))
860        && d.missing_dependencies
861            .iter()
862            .chain(&d.blocked_by)
863            .any(|id| {
864                matches!(
865                    crate::queue::dependency(queue, questions, id),
866                    crate::queue::Dependency::Ambiguous | crate::queue::Dependency::Unknown
867                )
868            })
869    {
870        return Ok(());
871    }
872
873    // Only an explicit missing-only claim can justify replacing a question
874    // or hold. Existing prerequisites alone say nothing about independent
875    // consent, destructive-operation questions, or other reasons to stay held.
876    // Ambiguous or unreadable inventories cannot confirm absence either: keep
877    // the original reference waiting until it can be resolved authoritatively.
878    if matches!(task.status, TaskStatus::Queued | TaskStatus::Held)
879        && !d.missing_dependencies.is_empty()
880        && d.missing_dependencies.iter().all(|id| {
881            let claim = crate::queue::dependency(queue, questions, id);
882            claim != crate::queue::Dependency::Missing
883                && d.blocked_by.iter().any(|blocker| {
884                    blocker == id
885                        || (matches!(
886                            &claim,
887                            crate::queue::Dependency::Task(_)
888                                | crate::queue::Dependency::Question(_)
889                                | crate::queue::Dependency::Deleted(_)
890                        ) && crate::queue::dependency(queue, questions, blocker) == claim)
891                })
892        })
893        && (d.question.is_some() || d.recovery == Some(Recovery::Hold))
894        && crate::queue::missing_blockers(queue, questions, &d.blocked_by).is_empty()
895    {
896        let mut ids = d.blocked_by.clone();
897        crate::queue::normalize_blockers(queue, questions, &mut ids);
898        let reason = format!("Waiting on dependencies: {}", ids.join(", "));
899        task.block(ids, Some(reason));
900        return queue.put(&mut task);
901    }
902
903    if let Some(text) = &d.question {
904        if task.status == TaskStatus::Done {
905            return Ok(());
906        }
907        let mut held = task.clone();
908        held.status = TaskStatus::Held;
909        let targets: BTreeSet<_> = d
910            .blocked_by
911            .iter()
912            .filter_map(|id| {
913                if let Ok(dep) = queue.get(id) {
914                    Some(format!("task:{id}:{:?}", dep.chat_notice_key()))
915                } else if questions.get(id).is_ok() {
916                    Some(format!("question:{id}"))
917                } else {
918                    None
919                }
920            })
921            .collect();
922        let decision = d
923            .question_decision
924            .as_ref()
925            .and_then(|q| match q.identity(&task) {
926                Ok(identity) => Some(identity),
927                Err(error) => {
928                    // Invalid optional metadata must not discard the question
929                    // or leave a runnable task without its decision barrier.
930                    tracing::warn!(task = %task.id, %error, "ignoring invalid question decision target");
931                    None
932                }
933            });
934        let context_value =
935            serde_json::json!([&task.instruction, held.chat_notice_key(), targets, decision]);
936        let context = serde_json::to_string(&context_value)?;
937        // An answer alone does not justify asking the same decision again.
938        // New producer evidence, an instruction edit or a different dependency
939        // target can open another question; rewording cannot.
940        if questions.list().iter().any(|q| {
941            q.node == NODE
942                && q.run == task.id
943                && q.status == crate::ask::QuestionStatus::Answered
944                && q.decision_context.as_ref().is_some_and(|stored| {
945                    if stored == &context {
946                        return true;
947                    }
948                    // Missing metadata cannot manufacture a new target. This
949                    // also baselines older contexts whose last item was labels.
950                    decision.is_none()
951                        && serde_json::from_str::<serde_json::Value>(stored)
952                            .ok()
953                            .and_then(|v| {
954                                v.as_array().map(|items| {
955                                    items.len() == 4
956                                        && items[..3]
957                                            == context_value.as_array().expect("context")[..3]
958                                })
959                            })
960                            == Some(true)
961                })
962        }) {
963            return Ok(());
964        }
965        // `Question::run` is a task id for conductor questions, but ordinary
966        // graph questions use it as a run id. A coincidental equality must
967        // not block this task on an answer meant for another node.
968        let question_id = match questions
969            .list()
970            .into_iter()
971            .find(|q| q.status.open() && q.node == NODE && q.run == task.id)
972        {
973            Some(existing) => existing.id,
974            None => {
975                let mut q = Question::new(
976                    task.id.clone(),
977                    NODE.to_owned(),
978                    SEAT.to_owned(),
979                    text.clone(),
980                    d.reason.clone().unwrap_or_default(),
981                    d.choices.clone(),
982                );
983                q.actions = d.actions.clone();
984                anyhow::ensure!(
985                    q.actions.keys().all(|label| q.choices.contains(label)),
986                    "question action must name an offered choice"
987                );
988                q.decision_context = Some(context);
989                // The wait is handed to a deputy (`crate::deputy`), so a
990                // free-text reply reaches an agent that knows why this was
991                // asked. The conductor still never blocks: this only records
992                // what the deputy will be told, and `serve` starts it.
993                q.deputy = Some(Deputy::new(crate::deputy::brief(
994                    &task.id,
995                    d.reason.as_deref().unwrap_or_default(),
996                    &q.choices,
997                    &q.actions,
998                )));
999                if task.repo.is_dir() && task.repo != Path::new(".") {
1000                    q.cwd = Some(task.repo.to_string_lossy().into_owned());
1001                }
1002                questions.put(&mut q)?;
1003                q.id
1004            }
1005        };
1006        task.block(vec![question_id], d.reason.clone());
1007        return queue.put(&mut task);
1008    }
1009
1010    match task.status {
1011        // See `Decision::recovery`'s doc: the only lever a runnable task has
1012        // besides `blocked_by` is holding it outright, for a task whose
1013        // answers already say it should not compete again.
1014        TaskStatus::Queued if d.recovery == Some(Recovery::Hold) => {
1015            if may_hold(&mut task, &hold_note(d)) {
1016                task.hold_machine(d.reason.clone());
1017                queue.put(&mut task)?;
1018            }
1019        }
1020        TaskStatus::Queued if !d.blocked_by.is_empty() => {
1021            let mut ids = d.blocked_by.clone();
1022            crate::queue::normalize_blockers(queue, questions, &mut ids);
1023            task.block(ids, d.reason.clone());
1024            queue.put(&mut task)?;
1025        }
1026        TaskStatus::Running => match d.recovery {
1027            Some(Recovery::Requeue) if pinned_resume(&task) => {}
1028            Some(Recovery::Requeue) => {
1029                task.requeue();
1030                queue.put(&mut task)?;
1031            }
1032            Some(Recovery::Hold) if may_hold(&mut task, &hold_note(d)) => {
1033                task.hold_machine(d.reason.clone());
1034                queue.put(&mut task)?;
1035            }
1036            // `Review` reopens a branch, which only makes sense once a run
1037            // has actually stopped; a task still `running` has nothing to
1038            // reopen yet.
1039            _ => {}
1040        },
1041        TaskStatus::Failed | TaskStatus::Held => match d.recovery {
1042            // The operator picked a specific run to continue; a fresh
1043            // competition would throw that run away.
1044            Some(Recovery::Requeue) if pinned_resume(&task) => {}
1045            Some(Recovery::Requeue) => {
1046                task.requeue();
1047                queue.put(&mut task)?;
1048            }
1049            Some(Recovery::Hold) => {
1050                // Decided before `may_hold`, which spends the one override: a
1051                // note that adds nothing writes nothing (and so raises no
1052                // notice and moves no queue revision).
1053                if let Some(reason) = reaffirmed_hold_reason(&task, d) {
1054                    if may_hold(&mut task, &hold_note(d)) {
1055                        task.hold_machine(Some(reason));
1056                        queue.put(&mut task)?;
1057                    }
1058                }
1059            }
1060            Some(Recovery::Review) => {
1061                if let Some(branch) = surviving_branch(&task) {
1062                    task.request_review(branch);
1063                    queue.put(&mut task)?;
1064                }
1065                // No survivable branch: a decision naming `review` here is
1066                // simply not actionable, and is dropped rather than guessed
1067                // at - `crate::daemon` applies the same "no branch, no
1068                // review" rule again, from its own read, right before it
1069                // would actually start the run.
1070            }
1071            Some(Recovery::Done) => {
1072                task.succeed();
1073                // Same reason `magi task done` and the web UI's equivalent
1074                // do this: the conductor closing a task by hand is just as
1075                // much a finished story as the loop's own settle path.
1076                crate::daemon::supersede_prior_runs(&task, &crate::run::home());
1077                queue.put(&mut task)?;
1078            }
1079            None => {}
1080        },
1081        // `queued` with nothing to block on, `done`, or already `blocked`:
1082        // nothing for this decision to do.
1083        _ => {}
1084    }
1085    Ok(())
1086}
1087
1088/// Apply every decision in `verdict`. A single bad decision - a task id that
1089/// no longer exists, one already claimed elsewhere - is logged and skipped
1090/// rather than losing every other decision in the same reply.
1091pub fn apply(queue: &Queue, questions: &Questions, verdict: &Verdict) -> Result<()> {
1092    for d in &verdict.decisions {
1093        if let Err(e) = apply_one(queue, questions, d, None) {
1094            tracing::warn!("conductor decision for task {}: {e:#}", d.id);
1095        }
1096    }
1097    Ok(())
1098}
1099
1100/// Where the conductor's seat is kept between turns, so the daemon waiter can
1101/// resume it when the owner answers a question it asked.
1102///
1103/// A graph seat's session lives in its run's `RunState`; the conductor has no
1104/// run, and until this file its seat existed only in memory.
1105pub fn seat_path(home: &Path) -> PathBuf {
1106    home.join("conduct").join("seat.json")
1107}
1108
1109/// The seat the conductor last used, if it has taken a turn.
1110pub fn load_seat(home: &Path) -> Option<SeatState> {
1111    serde_json::from_str(&std::fs::read_to_string(seat_path(home)).ok()?).ok()
1112}
1113
1114fn busy_path(home: &Path) -> PathBuf {
1115    home.join("conduct").join("busy")
1116}
1117
1118/// Is a conductor turn in flight right now?
1119///
1120/// A marker file rather than shared memory because the waiter reads it and the
1121/// conductor is not necessarily in its process. A marker older than one turn's
1122/// timeout (plus slack) was left by a turn that died and counts for nothing.
1123pub fn busy(home: &Path) -> bool {
1124    std::fs::metadata(busy_path(home))
1125        .and_then(|m| m.modified())
1126        .ok()
1127        .and_then(|t| t.elapsed().ok())
1128        .is_some_and(|age| age < TURN_TIMEOUT + Duration::from_secs(30))
1129}
1130
1131/// Removes the busy marker when a turn ends, however it ends.
1132struct Busy(PathBuf);
1133
1134impl Busy {
1135    fn mark(home: &Path) -> Self {
1136        let path = busy_path(home);
1137        if let Some(dir) = path.parent() {
1138            let _ = std::fs::create_dir_all(dir);
1139        }
1140        let _ = std::fs::write(&path, std::process::id().to_string());
1141        Self(path)
1142    }
1143}
1144
1145impl Drop for Busy {
1146    fn drop(&mut self) {
1147        let _ = std::fs::remove_file(&self.0);
1148    }
1149}
1150
1151/// The conductor's state across polling cycles: its own CLI-side conversation
1152/// and the last (revision, stalled ∪ finished ids) pair it actually acted on.
1153#[derive(Debug, Default)]
1154pub struct Conductor {
1155    seat: Option<SeatState>,
1156    last_seen: Option<(u64, BTreeSet<String>)>,
1157    /// Content fingerprint of each held task the model was last shown, see
1158    /// [`input_fingerprint`]. A held task whose fingerprint still matches is
1159    /// settled: the conductor's own rewrites of it do not count as news.
1160    considered: BTreeMap<String, String>,
1161}
1162
1163/// What the conductor's decision about a held task is a function of: the task
1164/// without the fields the conductor itself rewrites (`updated_at`,
1165/// `hold_reason`, the recorded re-hold). Independent of how the model words a
1166/// repeated decision.
1167fn decision_input(task: &Task) -> serde_json::Value {
1168    let mut v = serde_json::to_value(task).unwrap_or_default();
1169    if let Some(o) = v.as_object_mut() {
1170        o.remove("chat_notice");
1171        o.remove("chat_notices");
1172        // A delivery write changes only bookkeeping, not the decision input.
1173        o.remove("updated_at");
1174    }
1175    v
1176}
1177
1178fn input_fingerprint(task: &Task) -> String {
1179    let mut v = decision_input(task);
1180    if let Some(o) = v.as_object_mut() {
1181        o.remove("hold_reason");
1182        if let Some(ro) = o.get_mut("resume_override").and_then(|r| r.as_object_mut()) {
1183            ro.remove("conductor_rehold");
1184        }
1185    }
1186    v.to_string()
1187}
1188
1189impl Conductor {
1190    /// A conductor that has never run.
1191    #[must_use]
1192    pub fn new() -> Self {
1193        Self::default()
1194    }
1195
1196    /// Held tasks among `stalled` / `finished` whose input is unchanged since
1197    /// the model last saw them. Their files are left out of the revision, so
1198    /// the conductor re-holding one (however it words it) is not news, while
1199    /// an owner's answer changes the fingerprint and brings the file back in.
1200    fn settled(
1201        considered: &BTreeMap<String, String>,
1202        stalled: &[Task],
1203        finished: &[Task],
1204    ) -> BTreeSet<String> {
1205        stalled
1206            .iter()
1207            .chain(finished)
1208            .filter(|t| t.status == TaskStatus::Held)
1209            .filter(|t| considered.get(&t.id) == Some(&input_fingerprint(t)))
1210            .map(|t| t.id.clone())
1211            .collect()
1212    }
1213
1214    fn snapshot_with(
1215        considered: &BTreeMap<String, String>,
1216        queue: &Queue,
1217        stalled: &[Task],
1218        finished: &[Task],
1219    ) -> (u64, BTreeSet<String>) {
1220        let ids = stalled
1221            .iter()
1222            .chain(finished)
1223            .map(|t| t.id.clone())
1224            .collect();
1225        let skip = Self::settled(considered, stalled, finished);
1226        (queue.revision_excluding(&skip), ids)
1227    }
1228
1229    fn snapshot(
1230        &self,
1231        queue: &Queue,
1232        stalled: &[Task],
1233        finished: &[Task],
1234    ) -> (u64, BTreeSet<String>) {
1235        Self::snapshot_with(&self.considered, queue, stalled, finished)
1236    }
1237
1238    /// Whether calling the conductor could possibly do anything different
1239    /// from last time: [`Queue::revision`] moved, or the set of stalled and
1240    /// finished task ids changed.
1241    ///
1242    /// Deliberately **not** "stalled or finished is non-empty" - a task
1243    /// sitting stalled or finished with nobody changing anything about it
1244    /// must not be re-shown to the model every single poll forever; only a
1245    /// change in *which* tasks are stalled or finished, or a queue write
1246    /// changing something about a runnable one, is worth another look.
1247    ///
1248    /// Cheap and config-free on purpose, so `crate::daemon`'s poll loop can
1249    /// skip `Config::discover`'s synchronous I/O entirely on a cycle where
1250    /// this says no - which [`Conductor::maybe_run`] would otherwise only
1251    /// discover after paying for that load. Both ask the identical question,
1252    /// from the same [`Conductor::last_seen`], so they can never disagree
1253    /// about whether there is anything to look at.
1254    #[must_use]
1255    pub fn worth_a_look(&self, queue: &Queue, stalled: &[Task], finished: &[Task]) -> bool {
1256        self.last_seen.as_ref() != Some(&self.snapshot(queue, stalled, finished))
1257    }
1258
1259    /// Call the conductor once, unless nothing has changed since the last
1260    /// time it was worth calling - see [`Conductor::worth_a_look`], the exact
1261    /// same test. Never fatal - see this module's doc.
1262    #[allow(clippy::too_many_arguments)]
1263    pub async fn maybe_run(
1264        &mut self,
1265        cfg: &Config,
1266        repo: &Path,
1267        queue: &Queue,
1268        questions: &Questions,
1269        home: &Path,
1270        queued: &[Task],
1271        stalled: &[Task],
1272        finished: &[Task],
1273        max_attempts: usize,
1274    ) {
1275        if self.last_seen.as_ref() == Some(&self.snapshot(queue, stalled, finished)) {
1276            return;
1277        }
1278        // What the model is about to be shown becomes "considered" before the
1279        // snapshot is taken, so its own re-hold writes land in the excluded set.
1280        self.considered = stalled
1281            .iter()
1282            .chain(finished)
1283            .filter(|t| t.status == TaskStatus::Held)
1284            .map(|t| (t.id.clone(), input_fingerprint(t)))
1285            .collect();
1286        self.last_seen = Some(self.snapshot(queue, stalled, finished));
1287        if let Err(e) = self
1288            .run_once(
1289                cfg,
1290                repo,
1291                queue,
1292                questions,
1293                home,
1294                queued,
1295                stalled,
1296                finished,
1297                max_attempts,
1298            )
1299            .await
1300        {
1301            // The conductor's own write moves the revision, so one more cycle
1302            // follows; a repeat decision writes nothing (see
1303            // `reaffirmed_hold_reason`), which is what ends the loop. The
1304            // revision is deliberately never re-read here: it cannot tell our
1305            // write from an owner's concurrent update.
1306            tracing::warn!("conductor: {e:#}");
1307        }
1308    }
1309
1310    #[allow(clippy::too_many_arguments)]
1311    async fn run_once(
1312        &mut self,
1313        cfg: &Config,
1314        repo: &Path,
1315        queue: &Queue,
1316        questions: &Questions,
1317        home: &Path,
1318        queued: &[Task],
1319        stalled: &[Task],
1320        finished: &[Task],
1321        max_attempts: usize,
1322    ) -> Result<()> {
1323        if queued.is_empty() && stalled.is_empty() && finished.is_empty() {
1324            return Ok(());
1325        }
1326
1327        let primary = cfg
1328            .resolve_roles()
1329            .context("resolving the conductor seat")?
1330            .conductor;
1331        // A string keeps today's resolution exactly (no CLI preflight here;
1332        // an unavailable CLI fails at invocation). Only a multi-id array
1333        // builds a chain, whose members are each tried at most once.
1334        let chain = match cfg.roles.conductor.as_ref() {
1335            Some(c) if c.ids().len() > 1 => {
1336                agent::pick_chain(&cfg.agents, Some(c), &agent::installed, "conductor")?
1337            }
1338            _ => vec![primary],
1339        };
1340
1341        let runnable_views: Vec<prompt::ConductTask> =
1342            queued.iter().map(|t| view(t, max_attempts)).collect();
1343        let stalled_views: Vec<prompt::ConductTask> =
1344            stalled.iter().map(|t| view(t, max_attempts)).collect();
1345        let mut finished_views = Vec::with_capacity(finished.len());
1346        for t in finished {
1347            finished_views.push(finished_view(t, repo, max_attempts).await);
1348        }
1349
1350        let body = prompt::with_overlay(
1351            prompt::conduct(
1352                &runnable_views,
1353                &stalled_views,
1354                &finished_views,
1355                &cfg.graph.language,
1356            ),
1357            cfg.prompts.overlay(NODE),
1358        );
1359
1360        let body = format!(
1361            "{body}\n\n{}",
1362            crate::queue::dependency_facts(queue, questions)
1363        );
1364        let artifacts = home.join("conduct").join("artifacts");
1365        // Bound to a local: `Invocation` only borrows the cache path, and the
1366        // `Option<PathBuf>` `cache_dir()` returns has to outlive that borrow.
1367        let cache_dir = cfg.cache_dir();
1368
1369        // One marker for the whole chain: a fallback is still this turn.
1370        let busy = Busy::mark(home);
1371        let mut result: Result<Verdict> = Err(anyhow::anyhow!("no conductor agent ran"));
1372        for spec in &chain {
1373            // The prompt carries the whole queue state every cycle, so a
1374            // fallback agent needs no history - but its seat is its own.
1375            let needs_new_seat = !matches!(&self.seat, Some(s) if s.agent == spec.id);
1376            if needs_new_seat {
1377                self.seat = Some(SeatState::new(SEAT, &spec.id, crate::rng::entropy()));
1378            }
1379            let seat = self.seat.as_mut().expect("just ensured a seat exists");
1380            // The first agent keeps the plain stem; a fallback's own artifacts
1381            // must not overwrite it.
1382            let stem = if std::ptr::eq(spec, &chain[0]) {
1383                format!("turn-{}", seat.turns + 1)
1384            } else {
1385                format!("turn-{}-{}", seat.turns + 1, spec.id)
1386            };
1387            let inv = Invocation {
1388                coordination: cfg.coordination.as_ref(),
1389                fetched_base: None,
1390                cwd: repo,
1391                prompt: &body,
1392                timeout: TURN_TIMEOUT,
1393                // The conductor never edits anything - it only decides what
1394                // blocks a task and what to do about one stuck or finished.
1395                allow_write: false,
1396                unsandboxed: false,
1397                sessions: cfg.graph.sessions,
1398                artifacts: &artifacts,
1399                stem: &stem,
1400                run: NODE,
1401                node: NODE,
1402                cache_dir: cache_dir.as_deref(),
1403                attachments: &[],
1404                writable: &[],
1405            };
1406            let out = agent::invoke(spec, seat, &inv).await;
1407            // Kept after every turn, failed or not: the CLI-side conversation is
1408            // what a resumed question needs, and it exists once the turn ran.
1409            if let Ok(body) = serde_json::to_string(&*seat) {
1410                let path = seat_path(home);
1411                if let Some(dir) = path.parent() {
1412                    let _ = std::fs::create_dir_all(dir);
1413                }
1414                let _ = std::fs::write(path, body);
1415            }
1416            let advance = agent::chain_advances(&out);
1417            result = match out {
1418                Err(e) => Err(e.context("invoking the conductor")),
1419                Ok(out) if advance => Err(anyhow::anyhow!(
1420                    "no usable reply ({})",
1421                    crate::notices::turn_status(
1422                        out.exit_code,
1423                        out.timed_out,
1424                        out.quota_exhausted()
1425                    )
1426                )),
1427                Ok(out) => verdict::extract_json(&out.text)
1428                    .context("the conductor's reply could not be parsed"),
1429            };
1430            if result.is_ok() {
1431                break;
1432            }
1433            if chain.len() > 1 {
1434                tracing::warn!("conductor: `{}` failed, trying the next agent", spec.id);
1435            }
1436        }
1437        drop(busy);
1438        let mut verdict = result?;
1439        attach_facts(cfg, repo, queue, &mut verdict).await;
1440        for d in &verdict.decisions {
1441            let snapshot = queued
1442                .iter()
1443                .chain(stalled)
1444                .chain(finished)
1445                .find(|t| t.id == d.id);
1446            if let Some(snapshot) = snapshot {
1447                if let Err(e) = apply_one(queue, questions, d, Some(snapshot)) {
1448                    tracing::warn!("conductor decision for task {}: {e:#}", d.id);
1449                }
1450            }
1451        }
1452        Ok(())
1453    }
1454}
1455
1456#[cfg(test)]
1457mod tests {
1458    use std::collections::BTreeMap;
1459
1460    use tempfile::tempdir;
1461
1462    use super::*;
1463    use crate::ask::{Answer, QuestionStatus};
1464    use crate::config::{AgentKind, AgentSpec, Graph};
1465    use crate::queue::Source;
1466
1467    fn mock_agent(dir: &Path, script: &str, env: BTreeMap<String, String>) -> AgentSpec {
1468        let path = dir.join("mock-conduct-agent.sh");
1469        std::fs::write(&path, script).expect("write mock");
1470        AgentSpec {
1471            id: "mock".to_owned(),
1472            kind: AgentKind::Command,
1473            model: None,
1474            command: vec!["sh".to_owned(), path.to_string_lossy().into_owned()],
1475            extra_args: Vec::new(),
1476            env,
1477            prompt_delivery: None,
1478        }
1479    }
1480
1481    fn config(spec: AgentSpec) -> Config {
1482        Config {
1483            agents: vec![spec],
1484            graph: Graph {
1485                language: "en".to_owned(),
1486                ..Graph::default()
1487            },
1488            ..Config::default()
1489        }
1490    }
1491
1492    fn task(title: &str) -> Task {
1493        Task::new(
1494            title.to_owned(),
1495            format!("do {title}"),
1496            std::path::PathBuf::from("."),
1497            Source::Human,
1498        )
1499    }
1500
1501    #[test]
1502    fn delivery_bookkeeping_does_not_restart_conductor_deliberation() {
1503        let mut t = task("waiting");
1504        t.hold_machine(Some("waiting for recovery".into()));
1505        let before = input_fingerprint(&t);
1506        let snapshot = decision_input(&t);
1507        t.chat_notice = Some(crate::queue::ChatNotice {
1508            key: t.chat_notice_key().unwrap(),
1509            at: jiff::Timestamp::now(),
1510            outcome: crate::queue::ChatNoticeOutcome::Posted,
1511        });
1512        t.chat_notices.push(t.chat_notice_key().unwrap());
1513        t.updated_at = jiff::Timestamp::now();
1514        assert_eq!(before, input_fingerprint(&t));
1515        assert_eq!(snapshot, decision_input(&t));
1516        t.hold_reason = Some("recovery remains under consideration".into());
1517        assert_eq!(before, input_fingerprint(&t));
1518        t.record_answer("a different question".into(), "a new decision".into());
1519        assert_ne!(before, input_fingerprint(&t));
1520    }
1521
1522    const BROKEN: &str = "#!/bin/sh\ncat >/dev/null\nexit 3\n";
1523    const GARBAGE: &str = "#!/bin/sh\ncat >/dev/null\nprintf 'not json at all\\n'\n";
1524
1525    fn env(reply: &str) -> BTreeMap<String, String> {
1526        BTreeMap::from([("MOCK_REPLY".to_owned(), reply.to_owned())])
1527    }
1528
1529    const REPLY: &str = "#!/bin/sh\ncat >/dev/null\nprintf '%s\\n' \"$MOCK_REPLY\"\n";
1530
1531    /// A throwaway repo with one commit on `main` and a second branch ahead
1532    /// of it, so `outcome_for`'s own `git::rev_parse` call has a real head to
1533    /// resolve.
1534    fn init_repo_with_branch(dir: &Path, branch: &str) {
1535        use crate::proc::Quiet as _;
1536        let run = |args: &[&str]| {
1537            let out = std::process::Command::new("git")
1538                .args(args)
1539                .current_dir(dir)
1540                .quiet()
1541                .output()
1542                .expect("spawn git");
1543            assert!(
1544                out.status.success(),
1545                "git {args:?} failed: {}",
1546                String::from_utf8_lossy(&out.stderr)
1547            );
1548        };
1549        run(&["init", "-b", "main"]);
1550        run(&["config", "user.name", "magi test"]);
1551        run(&["config", "user.email", "magi@example.com"]);
1552        std::fs::write(dir.join("README.md"), "# fixture\n").unwrap();
1553        run(&["add", "-A"]);
1554        run(&["commit", "-m", "init"]);
1555        run(&["checkout", "-b", branch]);
1556        std::fs::write(dir.join("change.txt"), "x\n").unwrap();
1557        run(&["add", "-A"]);
1558        run(&["commit", "-m", "candidate work"]);
1559    }
1560
1561    fn review_round_with_finding(
1562        round: usize,
1563        finding_id: &str,
1564        title: &str,
1565        addressed: &[&str],
1566        rejected: &[(&str, &str)],
1567    ) -> crate::run::ReviewRound {
1568        crate::run::ReviewRound {
1569            round,
1570            head: "deadbeef".to_owned(),
1571            verified_head: None,
1572            verified_at: None,
1573            reviews: vec![crate::run::ReviewRecord {
1574                attempts: 0,
1575                reviewer: 1,
1576                agent: "mock".to_owned(),
1577                summary: String::new(),
1578                findings: vec![crate::verdict::Finding {
1579                    id: finding_id.to_owned(),
1580                    severity: crate::verdict::Severity::Major,
1581                    file: None,
1582                    line: None,
1583                    title: title.to_owned(),
1584                    detail: String::new(),
1585                }],
1586                vote: None,
1587                failed: None,
1588                duration_ms: 0,
1589            }],
1590            e2e: Vec::new(),
1591            verify_retried: false,
1592            e2e_deferred: false,
1593            e2e_defer_reason: None,
1594            fix: Some(crate::run::FixRecord {
1595                agent: "mock".to_owned(),
1596                addressed: addressed.iter().map(|s| (*s).to_owned()).collect(),
1597                rejected: rejected
1598                    .iter()
1599                    .map(|(id, why)| crate::verdict::Rejection {
1600                        id: (*id).to_owned(),
1601                        why: (*why).to_owned(),
1602                    })
1603                    .collect(),
1604                notes: String::new(),
1605                committed: false,
1606                failed: None,
1607                duration_ms: 0,
1608                continuation: None,
1609            }),
1610            blocking: 1,
1611            answered: 1,
1612            expected: 1,
1613            clean: false,
1614            progressed: true,
1615            vote_split: false,
1616            reconsideration: Vec::new(),
1617            verdict: None,
1618        }
1619    }
1620
1621    #[test]
1622    fn explicit_resumption_supersedes_history_and_rejects_stale_decisions() {
1623        crate::run::pin_test_home();
1624        let dir = tempdir().unwrap();
1625        let queue = Queue::at(dir.path().join("queue"));
1626        let questions = Questions::at(dir.path().join("questions"));
1627        let mut t = task("recover");
1628        t.answers.push(crate::queue::AnsweredQuestion {
1629            question: "Inspect remote first?".into(),
1630            answer: "Inspect, then await resumption instructions".into(),
1631        });
1632        t.hold_machine(Some("waiting for remote inspection".into()));
1633        queue.put(&mut t).unwrap();
1634        let snapshot = queue.get(&t.id).unwrap();
1635        t.edit(
1636            "recover".into(),
1637            "Reuse the measured PR head; fix Clippy; review".into(),
1638        )
1639        .unwrap();
1640        t.release_operator("cli", None);
1641        queue.put(&mut t).unwrap();
1642        let approval = t.resumption.clone();
1643        let d = Decision {
1644            id: t.id.clone(),
1645            recovery: Some(Recovery::Hold),
1646            reason: Some("remote inspection still required".into()),
1647            ..Decision::default()
1648        };
1649        apply_one(&queue, &questions, &d, Some(&snapshot)).unwrap();
1650        let stale_question = Decision {
1651            question: Some("Inspect remote?".into()),
1652            ..d.clone()
1653        };
1654        apply_one(&queue, &questions, &stale_question, Some(&snapshot)).unwrap();
1655        assert!(questions.list().is_empty());
1656        for reason in [
1657            "waiting for remote inspection",
1658            "inspection has not been confirmed",
1659        ] {
1660            let repeated = Decision {
1661                reason: Some(reason.into()),
1662                question: Some("How should we proceed?".into()),
1663                ..d.clone()
1664            };
1665            apply(
1666                &queue,
1667                &questions,
1668                &Verdict {
1669                    decisions: vec![repeated],
1670                },
1671            )
1672            .unwrap();
1673            assert_eq!(queue.get(&t.id).unwrap().status, TaskStatus::Queued);
1674        }
1675        assert!(
1676            questions.list().is_empty(),
1677            "a rejected proposal registers no question"
1678        );
1679        let notice = crate::notices::Notices::open()
1680            .get(&crate::notices::id_of(&format!(
1681                "conductor-resumption:{}",
1682                t.id
1683            )))
1684            .unwrap();
1685        assert!(!notice.message.contains("How should we proceed?"));
1686        assert!(!notice.message.contains("Some("));
1687        assert!(notice.message.contains("No reply to this notice is needed"));
1688        let reloaded = Queue::at(dir.path().join("queue")).get(&t.id).unwrap();
1689        assert_eq!(reloaded.resumption, approval);
1690        assert_eq!(reloaded.answers, t.answers);
1691        let v = view(&reloaded, 3);
1692        assert!(v.operator_resume.unwrap().contains("superseded"));
1693        let mut dep = task("new prerequisite");
1694        dep.hold_machine(Some("new authentication failure".into()));
1695        queue.put(&mut dep).unwrap();
1696        let new_blocker = Decision {
1697            blocked_by: vec![dep.id.clone()],
1698            evidence: vec![dep.id.clone()],
1699            // Exercise dependency waiting, without inheriting a separate hold.
1700            recovery: None,
1701            reason: Some("new prerequisite failed".into()),
1702            ..d.clone()
1703        };
1704        apply(
1705            &queue,
1706            &questions,
1707            &Verdict {
1708                decisions: vec![new_blocker.clone()],
1709            },
1710        )
1711        .unwrap();
1712        let mut after = queue.get(&t.id).unwrap();
1713        assert_eq!(after.status, TaskStatus::Blocked);
1714        // Automatic release does not grant fresh operator authority or make
1715        // the same unchanged evidence reusable.
1716        after.release();
1717        queue.put(&mut after).unwrap();
1718        apply(
1719            &queue,
1720            &questions,
1721            &Verdict {
1722                decisions: vec![new_blocker],
1723            },
1724        )
1725        .unwrap();
1726        assert_eq!(queue.get(&t.id).unwrap().status, TaskStatus::Queued);
1727        after.start("new-run".into());
1728        assert_eq!(after.resumption.as_ref().unwrap().generation, 1);
1729        after.hold_manual(Some("operator says wait".into()));
1730        queue.put(&mut after).unwrap();
1731        apply(&queue, &questions, &Verdict { decisions: vec![d] }).unwrap();
1732        assert!(queue.get(&t.id).unwrap().operator_held());
1733    }
1734
1735    #[test]
1736    fn newly_blocked_dependency_is_material_evidence_after_approval() {
1737        let dir = tempdir().unwrap();
1738        let queue = Queue::at(dir.path().join("queue"));
1739        let questions = Questions::at(dir.path().join("questions"));
1740        let mut dep = task("old queued dependency");
1741        queue.put(&mut dep).unwrap();
1742        let mut t = task("approved task");
1743        t.release_operator("test", None);
1744        queue.put(&mut t).unwrap();
1745        let mut q = Question::new(
1746            dep.id.clone(),
1747            "check".into(),
1748            "check".into(),
1749            "New dependency approval".into(),
1750            String::new(),
1751            Vec::new(),
1752        );
1753        questions.put(&mut q).unwrap();
1754        dep.block(vec![q.id.clone()], Some("new decision needed".into()));
1755        let changed = dep.blocked_at;
1756        queue.put(&mut dep).unwrap();
1757        dep.block(vec![q.id], Some("same decision paraphrased".into()));
1758        assert_eq!(
1759            dep.blocked_at, changed,
1760            "prose is not new blocking evidence"
1761        );
1762        queue.put(&mut dep).unwrap();
1763        let d = Decision {
1764            id: t.id.clone(),
1765            blocked_by: vec![dep.id.clone()],
1766            evidence: vec![dep.id],
1767            ..Decision::default()
1768        };
1769        assert_eq!(new_blocking_evidence(&queue, &questions, &t, &d).len(), 1);
1770        apply_one(&queue, &questions, &d, None).unwrap();
1771        assert_eq!(queue.get(&t.id).unwrap().status, TaskStatus::Blocked);
1772    }
1773
1774    #[test]
1775    fn resumed_run_evidence_uses_approval_time_not_run_list_offsets() {
1776        crate::run::pin_test_home();
1777        let dir = tempdir().unwrap();
1778        let queue = Queue::at(dir.path().join("queue"));
1779        let questions = Questions::at(dir.path().join("questions"));
1780        let mut t = task("resume evidence");
1781        let mut run = RunState::new(
1782            t.repo.clone(),
1783            "main".into(),
1784            "base".into(),
1785            "work".into(),
1786            crate::config::Config::default(),
1787        );
1788        run.begin_pass();
1789        run.status = crate::run::RunStatus::Failed;
1790        run.save().unwrap();
1791        t.start(run.id.clone());
1792        // Model legacy duplicate history, which the next start will deduplicate.
1793        t.runs.push(run.id.clone());
1794        t.release_operator("cli", None);
1795        let approval = t.resumption.as_ref().unwrap().at;
1796        let decision = Decision {
1797            evidence: vec![run.id.clone()],
1798            ..Decision::default()
1799        };
1800        assert!(new_blocking_evidence(&queue, &questions, &t, &decision).is_empty());
1801        t.start(run.id.clone());
1802        assert_eq!(t.runs.len(), 1);
1803        run.begin_pass();
1804        run.save().unwrap();
1805        assert!(run.passes.last().unwrap().started_at > approval);
1806        let evidence = new_blocking_evidence(&queue, &questions, &t, &decision);
1807        assert_eq!(evidence.len(), 1);
1808        t.resumption.as_mut().unwrap().blocking_evidence = evidence;
1809        assert!(new_blocking_evidence(&queue, &questions, &t, &decision).is_empty());
1810        run.begin_pass();
1811        run.save().unwrap();
1812        assert_eq!(
1813            new_blocking_evidence(&queue, &questions, &t, &decision).len(),
1814            1
1815        );
1816
1817        let mut q = Question::new(
1818            run.id.clone(),
1819            "review".into(),
1820            "check".into(),
1821            "New failure".into(),
1822            String::new(),
1823            Vec::new(),
1824        );
1825        questions.put(&mut q).unwrap();
1826        let decision = Decision {
1827            evidence: vec![q.id.clone()],
1828            ..Decision::default()
1829        };
1830        assert_eq!(
1831            new_blocking_evidence(&queue, &questions, &t, &decision).len(),
1832            1
1833        );
1834        q.asked_at = approval;
1835        questions.put(&mut q).unwrap();
1836        assert!(new_blocking_evidence(&queue, &questions, &t, &decision).is_empty());
1837        q.asked_at = jiff::Timestamp::now();
1838        q.run = "unrelated".into();
1839        questions.put(&mut q).unwrap();
1840        assert!(new_blocking_evidence(&queue, &questions, &t, &decision).is_empty());
1841
1842        let mut fresh = RunState::new(
1843            t.repo.clone(),
1844            "main".into(),
1845            "base".into(),
1846            "fresh work".into(),
1847            crate::config::Config::default(),
1848        );
1849        fresh.status = crate::run::RunStatus::Failed;
1850        fresh.save().unwrap();
1851        t.start(fresh.id.clone());
1852        let decision = Decision {
1853            evidence: vec![fresh.id.clone()],
1854            ..Decision::default()
1855        };
1856        assert_eq!(
1857            new_blocking_evidence(&queue, &questions, &t, &decision).len(),
1858            1
1859        );
1860    }
1861
1862    #[test]
1863    fn outcome_for_carries_every_rounds_findings_and_the_branch_head() {
1864        crate::run::pin_test_home();
1865        let dir = tempdir().unwrap();
1866        let default_repo = dir.path().join("default");
1867        let task_repo = dir.path().join("task");
1868        std::fs::create_dir_all(&default_repo).unwrap();
1869        std::fs::create_dir_all(&task_repo).unwrap();
1870        init_repo_with_branch(&default_repo, "other-branch");
1871        init_repo_with_branch(&task_repo, "magi/f00d/A");
1872
1873        let mut config = Config::default();
1874        config.graph.review_rounds = 6;
1875        let mut state = crate::run::RunState::new(
1876            task_repo.clone(),
1877            "main".to_owned(),
1878            "deadbeef".to_owned(),
1879            "task".to_owned(),
1880            config,
1881        );
1882        state.status = crate::run::RunStatus::Blocked;
1883        state.candidates.push(crate::run::Candidate {
1884            index: 0,
1885            label: 'A',
1886            agent: "mock".to_owned(),
1887            branch: "magi/f00d/A".to_owned(),
1888            worktree: task_repo.clone(),
1889            summary: String::new(),
1890            stat: String::new(),
1891            files: 1,
1892            commits: 1,
1893            empty: false,
1894            failed: None,
1895            verified_noop: None,
1896            duration_ms: 0,
1897            folded: false,
1898        });
1899        state.tally = Some(crate::run::Tally {
1900            first_choice: std::collections::BTreeMap::new(),
1901            borda: std::collections::BTreeMap::new(),
1902            winner: 'A',
1903            rankings: 0,
1904            unanimous_initial: false,
1905            deliberated: false,
1906            changed_votes: 0,
1907            unanimous_final: false,
1908            tie_break: None,
1909            judges: 0,
1910            present: 0,
1911            quorum: 0,
1912            met_quorum: true,
1913            uncontested: Some("solo".to_owned()),
1914        });
1915        state.reviews = vec![
1916            review_round_with_finding(
1917                1,
1918                "R1-1-2",
1919                "answer content is dropped",
1920                &[],
1921                &[("R1-1-2", "the id leaving blocked_by is enough")],
1922            ),
1923            review_round_with_finding(2, "R2-1-3", "answer content is still dropped", &[], &[]),
1924        ];
1925        state.save().unwrap();
1926
1927        let mut t = task("outcome test");
1928        t.repo = task_repo;
1929        t.runs.push(state.id.clone());
1930
1931        let finished = tokio_test_block_on(finished_view(&t, &default_repo, 2));
1932        let outcome = finished.outcome;
1933
1934        assert!(outcome.unreadable.is_none());
1935        assert_eq!(outcome.run_status.as_deref(), Some("blocked"));
1936        assert_eq!(outcome.rounds_used, 2);
1937        assert_eq!(outcome.rounds_max, 6);
1938        assert_eq!(outcome.rounds.len(), 2);
1939        assert_eq!(outcome.rounds[0].findings[0].id, "R1-1-2");
1940        assert_eq!(outcome.rounds[0].rejected[0].id, "R1-1-2");
1941        assert!(outcome.rounds[1].addressed.is_empty());
1942        assert!(outcome.rounds[1].rejected.is_empty());
1943        assert_eq!(outcome.branch.as_deref(), Some("magi/f00d/A"));
1944        assert!(
1945            outcome.branch_head.is_some(),
1946            "a real branch must resolve a head commit: {outcome:?}"
1947        );
1948    }
1949
1950    /// A tiny single-threaded block-on, so an `async fn` can be exercised
1951    /// from a plain `#[test]` without pulling `tokio::test`'s multi-thread
1952    /// runtime into a test that does no other async work.
1953    fn tokio_test_block_on<F: std::future::Future>(f: F) -> F::Output {
1954        tokio::runtime::Builder::new_current_thread()
1955            .enable_all()
1956            .build()
1957            .unwrap()
1958            .block_on(f)
1959    }
1960
1961    #[test]
1962    fn view_resolves_task_attempt_limit_before_daemon_default() {
1963        let mut t = task("budget");
1964        assert_eq!(view(&t, 7).max_attempts, 7);
1965        t.change_attempt_limit(std::num::NonZeroUsize::new(3))
1966            .unwrap();
1967        assert_eq!(view(&t, 7).max_attempts, 3);
1968        t.change_attempt_limit(None).unwrap();
1969        assert_eq!(view(&t, 7).max_attempts, 7);
1970    }
1971
1972    #[test]
1973    fn view_carries_a_tasks_recorded_answers_into_the_conductor_prompt_input() {
1974        let mut t = task("answered");
1975        t.record_answer("Which backend?".to_owned(), "SQLite".to_owned());
1976        let v = view(&t, 2);
1977        assert_eq!(v.answers.len(), 1);
1978        assert_eq!(v.answers[0].question, "Which backend?");
1979        assert_eq!(v.answers[0].answer, "SQLite");
1980    }
1981
1982    #[tokio::test]
1983    async fn dependency_resolution_rechecks_questions_in_the_production_conductor_path() {
1984        for held in [false, true] {
1985            for reference in ["ca5b", "20261013-193611-ca5b"] {
1986                for question in [false, true] {
1987                    let dir = tempdir().unwrap();
1988                    let queue = Queue::at(dir.path().join("queue"));
1989                    let questions = Questions::at(dir.path().join("questions"));
1990                    let mut dep = task("prerequisite");
1991                    dep.id = "20261013-193611-ca5b".into();
1992                    dep.start("existing-run".into());
1993                    queue.put(&mut dep).unwrap();
1994                    let mut waiting = task("waiting");
1995                    if held {
1996                        waiting.hold_machine(Some("dependency could not be located".into()));
1997                    }
1998                    queue.put(&mut waiting).unwrap();
1999                    let missing_reference = if reference == "ca5b" {
2000                        dep.id.as_str()
2001                    } else {
2002                        "ca5b"
2003                    };
2004                    let reply = serde_json::json!({"decisions": [{
2005                        "id": waiting.id, "blocked_by": [reference],
2006                        "missing_dependencies": [missing_reference],
2007                        "question": question.then_some("ca5b is missing; where is its replacement?"),
2008                        "recovery": (!question).then_some("hold"),
2009                        "reason": "ca5b is missing"
2010                    }]}).to_string();
2011                    let cfg = config(mock_agent(dir.path(), REPLY, env(&reply)));
2012                    let (queued, finished) = if held {
2013                        (vec![], vec![waiting.clone()])
2014                    } else {
2015                        (vec![waiting.clone()], vec![])
2016                    };
2017                    Conductor::new()
2018                        .run_once(
2019                            &cfg,
2020                            dir.path(),
2021                            &queue,
2022                            &questions,
2023                            dir.path(),
2024                            &queued,
2025                            &[],
2026                            &finished,
2027                            2,
2028                        )
2029                        .await
2030                        .unwrap();
2031                    let after = queue.get(&waiting.id).unwrap();
2032                    assert_eq!(after.status, TaskStatus::Blocked);
2033                    assert_eq!(after.blocked_by, [dep.id]);
2034                    assert_eq!(after.blocked_from, Some(waiting.status));
2035                    assert!(!after.block_reason.unwrap().contains("missing"));
2036                    assert!(questions.list().is_empty());
2037                    assert_eq!(after.instruction, waiting.instruction);
2038                }
2039            }
2040        }
2041    }
2042
2043    #[tokio::test]
2044    async fn dependency_resolution_inconclusive_missing_only_decisions_keep_waiting() {
2045        for held in [false, true] {
2046            for unknown in [false, true] {
2047                for question in [false, true] {
2048                    let dir = tempdir().unwrap();
2049                    let queue = Queue::at(dir.path().join("queue"));
2050                    let questions = Questions::at(dir.path().join("questions"));
2051                    let mut dep = task("running prerequisite");
2052                    dep.id = "20261013-193611-ca5b".into();
2053                    dep.start("existing-run".into());
2054                    queue.put(&mut dep).unwrap();
2055                    let reference = if unknown {
2056                        // Even an exact, existing task id is Unknown when
2057                        // the question inventory cannot be scanned.
2058                        std::fs::write(questions.root(), "not a directory").unwrap();
2059                        dep.id.clone()
2060                    } else {
2061                        let mut other = task("another prerequisite");
2062                        other.id = "20261014-193611-ca5b".into();
2063                        queue.put(&mut other).unwrap();
2064                        "ca5b".to_owned()
2065                    };
2066                    assert_eq!(
2067                        crate::queue::dependency(&queue, &questions, &reference),
2068                        if unknown {
2069                            crate::queue::Dependency::Unknown
2070                        } else {
2071                            crate::queue::Dependency::Ambiguous
2072                        }
2073                    );
2074                    let mut waiting = task("waiting");
2075                    if held {
2076                        waiting.hold_machine(Some("dependency could not be located".into()));
2077                    }
2078                    queue.put(&mut waiting).unwrap();
2079                    let reply = serde_json::json!({"decisions": [{
2080                        "id": waiting.id, "blocked_by": [reference],
2081                        "missing_dependencies": [reference],
2082                        "question": question.then_some("ca5b is missing"),
2083                        "recovery": (!question).then_some("hold"),
2084                        "reason": "ca5b is missing"
2085                    }]})
2086                    .to_string();
2087                    let cfg = config(mock_agent(dir.path(), REPLY, env(&reply)));
2088                    let (queued, finished) = if held {
2089                        (vec![], vec![waiting.clone()])
2090                    } else {
2091                        (vec![waiting.clone()], vec![])
2092                    };
2093                    Conductor::new()
2094                        .run_once(
2095                            &cfg,
2096                            dir.path(),
2097                            &queue,
2098                            &questions,
2099                            dir.path(),
2100                            &queued,
2101                            &[],
2102                            &finished,
2103                            2,
2104                        )
2105                        .await
2106                        .unwrap();
2107                    let after = queue.get(&waiting.id).unwrap();
2108                    assert_eq!(after.status, TaskStatus::Blocked);
2109                    assert_eq!(after.blocked_by, [reference]);
2110                    assert_eq!(after.blocked_from, Some(waiting.status));
2111                    assert!(!after.block_reason.unwrap().contains("missing"));
2112                    assert_eq!(after.instruction, waiting.instruction);
2113                    if unknown {
2114                        assert!(questions.root().is_file());
2115                    } else {
2116                        assert!(questions.list().is_empty());
2117                    }
2118                }
2119            }
2120        }
2121    }
2122
2123    #[tokio::test]
2124    async fn dependency_resolution_unverifiable_alias_decisions_are_discarded() {
2125        for held in [false, true] {
2126            for unknown in [false, true] {
2127                for question in [false, true] {
2128                    for short_blocker in [false, true] {
2129                        let dir = tempdir().unwrap();
2130                        let queue = Queue::at(dir.path().join("queue"));
2131                        let questions = Questions::at(dir.path().join("questions"));
2132                        let mut dep = task("running prerequisite");
2133                        dep.id = "20261013-193611-ca5b".into();
2134                        dep.start("existing-run".into());
2135                        queue.put(&mut dep).unwrap();
2136                        if unknown {
2137                            std::fs::write(questions.root(), "not a directory").unwrap();
2138                        } else {
2139                            let mut other = task("another prerequisite");
2140                            other.id = "20261014-193611-ca5b".into();
2141                            queue.put(&mut other).unwrap();
2142                        }
2143                        let (blocker, missing) = if short_blocker {
2144                            ("ca5b", dep.id.as_str())
2145                        } else {
2146                            (dep.id.as_str(), "ca5b")
2147                        };
2148                        let mut waiting = task("waiting");
2149                        if held {
2150                            waiting.hold_machine(Some("dependency could not be located".into()));
2151                        }
2152                        queue.put(&mut waiting).unwrap();
2153                        let reply = serde_json::json!({"decisions": [{
2154                            "id": waiting.id, "blocked_by": [blocker],
2155                            "missing_dependencies": [missing],
2156                            "question": question.then_some("ca5b is missing"),
2157                            "recovery": (!question).then_some("hold"),
2158                            "reason": "ca5b is missing"
2159                        }]})
2160                        .to_string();
2161                        let cfg = config(mock_agent(dir.path(), REPLY, env(&reply)));
2162                        let (queued, finished) = if held {
2163                            (vec![], vec![waiting.clone()])
2164                        } else {
2165                            (vec![waiting.clone()], vec![])
2166                        };
2167                        Conductor::new()
2168                            .run_once(
2169                                &cfg,
2170                                dir.path(),
2171                                &queue,
2172                                &questions,
2173                                dir.path(),
2174                                &queued,
2175                                &[],
2176                                &finished,
2177                                2,
2178                            )
2179                            .await
2180                            .unwrap();
2181                        assert_eq!(
2182                            serde_json::to_value(queue.get(&waiting.id).unwrap()).unwrap(),
2183                            serde_json::to_value(&waiting).unwrap(),
2184                            "unverifiable aliases must leave the entire task unchanged"
2185                        );
2186                        if unknown {
2187                            assert!(questions.root().is_file());
2188                        } else {
2189                            assert!(questions.list().is_empty());
2190                        }
2191                    }
2192                }
2193            }
2194        }
2195    }
2196
2197    #[tokio::test]
2198    async fn dependency_resolution_genuine_missing_reference_still_asks() {
2199        let dir = tempdir().unwrap();
2200        let queue = Queue::at(dir.path().join("queue"));
2201        let questions = Questions::at(dir.path().join("questions"));
2202        let mut waiting = task("missing prerequisite");
2203        queue.put(&mut waiting).unwrap();
2204        let reply = serde_json::json!({"decisions": [{
2205            "id": waiting.id, "blocked_by": ["not-present"],
2206            "missing_dependencies": ["not-present"],
2207            "question": "Which prerequisite replaces the missing task?"
2208        }]})
2209        .to_string();
2210        let cfg = config(mock_agent(dir.path(), REPLY, env(&reply)));
2211        Conductor::new()
2212            .run_once(
2213                &cfg,
2214                dir.path(),
2215                &queue,
2216                &questions,
2217                dir.path(),
2218                &[waiting.clone()],
2219                &[],
2220                &[],
2221                2,
2222            )
2223            .await
2224            .unwrap();
2225        let after = queue.get(&waiting.id).unwrap();
2226        assert_eq!(after.status, TaskStatus::Blocked);
2227        assert_eq!(
2228            questions.get(&after.blocked_by[0]).unwrap().summary,
2229            "Which prerequisite replaces the missing task?"
2230        );
2231    }
2232
2233    #[tokio::test]
2234    async fn dependency_resolution_keeps_decisions_during_unrelated_task_progress() {
2235        let dir = tempdir().unwrap();
2236        let queue = Queue::at(dir.path().join("queue"));
2237        let questions = Questions::at(dir.path().join("questions"));
2238        let mut dep = task("prerequisite");
2239        dep.id = "20261013-193611-ca5b".into();
2240        dep.start("existing-run".into());
2241        queue.put(&mut dep).unwrap();
2242        let mut unrelated = task("unrelated work");
2243        unrelated.start("other-run".into());
2244        queue.put(&mut unrelated).unwrap();
2245        // Make the old and new revision distinguishable without relying on
2246        // the filesystem clock advancing while the mock process starts.
2247        std::fs::File::options()
2248            .write(true)
2249            .open(queue.path_of(&unrelated.id))
2250            .unwrap()
2251            .set_times(std::fs::FileTimes::new().set_modified(std::time::UNIX_EPOCH))
2252            .unwrap();
2253        unrelated.succeed();
2254        let mut waiting = task("waiting");
2255        let mut asking = task("needs a product decision");
2256        let mut holding = task("must stay held");
2257        queue.put(&mut waiting).unwrap();
2258        queue.put(&mut asking).unwrap();
2259        queue.put(&mut holding).unwrap();
2260        let reply = serde_json::json!({"decisions": [
2261            {"id": waiting.id, "blocked_by": ["ca5b"], "reason": "wait for implementation"},
2262            {"id": asking.id, "question": "Which behavior do you want?"},
2263            {"id": holding.id, "recovery": "hold", "reason": "already implemented elsewhere"}
2264        ]})
2265        .to_string();
2266        let mut vars = env(&reply);
2267        vars.insert(
2268            "UNRELATED_PATH".into(),
2269            queue.path_of(&unrelated.id).to_string_lossy().into_owned(),
2270        );
2271        vars.insert(
2272            "UNRELATED_RECORD".into(),
2273            serde_json::to_string(&unrelated).unwrap(),
2274        );
2275        let script = r#"#!/bin/sh
2276cat >/dev/null
2277printf '%s' "$UNRELATED_RECORD" > "$UNRELATED_PATH"
2278printf '%s\n' "$MOCK_REPLY"
2279"#;
2280        let cfg = config(mock_agent(dir.path(), script, vars));
2281        Conductor::new()
2282            .run_once(
2283                &cfg,
2284                dir.path(),
2285                &queue,
2286                &questions,
2287                dir.path(),
2288                &[waiting.clone(), asking.clone(), holding.clone()],
2289                &[],
2290                &[],
2291                2,
2292            )
2293            .await
2294            .unwrap();
2295        assert_eq!(queue.get(&unrelated.id).unwrap().status, TaskStatus::Done);
2296        let after = queue.get(&waiting.id).unwrap();
2297        assert_eq!(after.status, TaskStatus::Blocked);
2298        assert_eq!(after.blocked_by, [dep.id]);
2299        let asked = queue.get(&asking.id).unwrap();
2300        assert_eq!(asked.status, TaskStatus::Blocked);
2301        assert_eq!(
2302            questions.get(&asked.blocked_by[0]).unwrap().summary,
2303            "Which behavior do you want?"
2304        );
2305        assert_eq!(queue.get(&holding.id).unwrap().status, TaskStatus::Held);
2306    }
2307
2308    #[test]
2309    fn a_dependency_decision_blocks_the_task_and_leaves_priority_alone() {
2310        let dir = tempdir().unwrap();
2311        let queue = Queue::at(dir.path().join("queue"));
2312        let questions = Questions::at(dir.path().join("questions"));
2313        let mut a = task("a");
2314        a.priority = 9;
2315        queue.put(&mut a).unwrap();
2316
2317        let verdict = Verdict {
2318            decisions: vec![Decision {
2319                id: a.id.clone(),
2320                blocked_by: vec!["20260101-000000-dead".to_owned()],
2321                reason: Some("waits on the other task".to_owned()),
2322                recovery: None,
2323                question: None,
2324                choices: Vec::new(),
2325                actions: BTreeMap::new(),
2326                question_decision: None,
2327                evidence: Vec::new(),
2328                missing_dependencies: Vec::new(),
2329            }],
2330        };
2331        apply(&queue, &questions, &verdict).unwrap();
2332
2333        let back = queue.get(&a.id).unwrap();
2334        assert_eq!(back.status, TaskStatus::Blocked);
2335        assert_eq!(back.blocked_by, ["20260101-000000-dead"]);
2336        assert_eq!(
2337            back.priority, 9,
2338            "the conductor's reply cannot carry priority"
2339        );
2340    }
2341
2342    #[test]
2343    fn a_question_decision_files_one_and_blocks_on_its_id() {
2344        let dir = tempdir().unwrap();
2345        let queue = Queue::at(dir.path().join("queue"));
2346        let questions = Questions::at(dir.path().join("questions"));
2347        let mut t = task("ambiguous");
2348        queue.put(&mut t).unwrap();
2349
2350        let verdict = Verdict {
2351            decisions: vec![Decision {
2352                id: t.id.clone(),
2353                blocked_by: Vec::new(),
2354                reason: Some("which backend?".to_owned()),
2355                recovery: None,
2356                question: Some("Which storage backend?".to_owned()),
2357                choices: vec!["SQLite".to_owned(), "Redis".to_owned()],
2358                actions: BTreeMap::new(),
2359                question_decision: None,
2360                evidence: Vec::new(),
2361                missing_dependencies: Vec::new(),
2362            }],
2363        };
2364        apply(&queue, &questions, &verdict).unwrap();
2365
2366        let back = queue.get(&t.id).unwrap();
2367        assert_eq!(back.status, TaskStatus::Blocked);
2368        assert_eq!(back.blocked_by.len(), 1);
2369        let q = questions.get(&back.blocked_by[0]).unwrap();
2370        assert_eq!(q.summary, "Which storage backend?");
2371        assert_eq!(q.node, NODE);
2372        assert!(q.status.open());
2373    }
2374
2375    #[test]
2376    fn a_task_with_an_open_question_already_reuses_it_rather_than_filing_a_second_one() {
2377        let dir = tempdir().unwrap();
2378        let queue = Queue::at(dir.path().join("queue"));
2379        let questions = Questions::at(dir.path().join("questions"));
2380        let mut t = task("asked once");
2381        queue.put(&mut t).unwrap();
2382
2383        let decision = Decision {
2384            id: t.id.clone(),
2385            reason: Some("still deciding".to_owned()),
2386            question: Some("Which backend?".to_owned()),
2387            ..Decision::default()
2388        };
2389        apply(
2390            &queue,
2391            &questions,
2392            &Verdict {
2393                decisions: vec![decision.clone()],
2394            },
2395        )
2396        .unwrap();
2397        assert_eq!(questions.list().len(), 1);
2398        let first_question_id = queue.get(&t.id).unwrap().blocked_by[0].clone();
2399
2400        // An operator releasing the blocked task by hand, without answering,
2401        // puts it back at `Queued` while the question stays open - exactly
2402        // the case the guard in `apply_one` exists for: a later cycle
2403        // proposing the very same question must reuse it, not file a second.
2404        let mut released = queue.get(&t.id).unwrap();
2405        released.release();
2406        queue.put(&mut released).unwrap();
2407
2408        apply(
2409            &queue,
2410            &questions,
2411            &Verdict {
2412                decisions: vec![decision],
2413            },
2414        )
2415        .unwrap();
2416        assert_eq!(questions.list().len(), 1, "no duplicate question was filed");
2417        let after = queue.get(&t.id).unwrap();
2418        assert_eq!(
2419            after.blocked_by,
2420            [first_question_id],
2421            "the existing open question is reused, not replaced"
2422        );
2423    }
2424
2425    #[test]
2426    fn a_same_id_question_from_another_node_is_not_reused() {
2427        let dir = tempdir().unwrap();
2428        let queue = Queue::at(dir.path().join("queue"));
2429        let questions = Questions::at(dir.path().join("questions"));
2430        let mut t = task("must ask the conductor");
2431        queue.put(&mut t).unwrap();
2432
2433        let mut unrelated = Question::new(
2434            t.id.clone(),
2435            "review".to_owned(),
2436            "reviewer-1".to_owned(),
2437            "An unrelated review question".to_owned(),
2438            String::new(),
2439            Vec::new(),
2440        );
2441        questions.put(&mut unrelated).unwrap();
2442
2443        apply(
2444            &queue,
2445            &questions,
2446            &Verdict {
2447                decisions: vec![Decision {
2448                    id: t.id.clone(),
2449                    question: Some("Which backend?".to_owned()),
2450                    ..Decision::default()
2451                }],
2452            },
2453        )
2454        .unwrap();
2455
2456        let blocked_by = &queue.get(&t.id).unwrap().blocked_by;
2457        assert_eq!(blocked_by.len(), 1);
2458        assert_ne!(blocked_by[0], unrelated.id);
2459        assert!(questions.get(&unrelated.id).unwrap().status.open());
2460        assert_eq!(questions.get(&blocked_by[0]).unwrap().node, NODE);
2461    }
2462
2463    #[test]
2464    fn answering_the_question_lets_the_resolver_clear_the_block_with_the_answer_kept() {
2465        let dir = tempdir().unwrap();
2466        let queue = Queue::at(dir.path().join("queue"));
2467        let questions = Questions::at(dir.path().join("questions"));
2468        let mut t = task("waits on an answer");
2469        queue.put(&mut t).unwrap();
2470
2471        apply(
2472            &queue,
2473            &questions,
2474            &Verdict {
2475                decisions: vec![Decision {
2476                    id: t.id.clone(),
2477                    blocked_by: Vec::new(),
2478                    reason: None,
2479                    recovery: None,
2480                    question: Some("Which backend?".to_owned()),
2481                    choices: Vec::new(),
2482                    actions: BTreeMap::new(),
2483                    question_decision: None,
2484                    evidence: Vec::new(),
2485                    missing_dependencies: Vec::new(),
2486                }],
2487            },
2488        )
2489        .unwrap();
2490        let blocked = queue.get(&t.id).unwrap();
2491        let question_id = blocked.blocked_by[0].clone();
2492
2493        let mut q = questions.get(&question_id).unwrap();
2494        q.answer(Answer::Text("SQLite".to_owned())).unwrap();
2495        questions.put(&mut q).unwrap();
2496        assert_eq!(q.status, QuestionStatus::Answered);
2497
2498        // `crate::daemon::resolve_blockers` is the deterministic resolver
2499        // that actually does this on the real queue; here it is enough to
2500        // prove the pure steps it is built from behave together.
2501        let mut task_after = queue.get(&t.id).unwrap();
2502        task_after.record_answer(q.summary.clone(), "SQLite".to_owned());
2503        task_after.unblock(&question_id);
2504        assert_eq!(task_after.status, TaskStatus::Queued);
2505        assert_eq!(task_after.answers[0].answer, "SQLite");
2506    }
2507
2508    #[test]
2509    fn a_stalled_task_can_be_requeued_or_held() {
2510        let dir = tempdir().unwrap();
2511        let queue = Queue::at(dir.path().join("queue"));
2512        let questions = Questions::at(dir.path().join("questions"));
2513
2514        let mut requeue_me = task("stuck a");
2515        requeue_me.start("run-1".to_owned());
2516        queue.put(&mut requeue_me).unwrap();
2517
2518        let mut hold_me = task("stuck b");
2519        hold_me.start("run-2".to_owned());
2520        queue.put(&mut hold_me).unwrap();
2521
2522        apply(
2523            &queue,
2524            &questions,
2525            &Verdict {
2526                decisions: vec![
2527                    Decision {
2528                        id: requeue_me.id.clone(),
2529                        recovery: Some(Recovery::Requeue),
2530                        ..Decision::default()
2531                    },
2532                    Decision {
2533                        id: hold_me.id.clone(),
2534                        recovery: Some(Recovery::Hold),
2535                        reason: Some("looks broken".to_owned()),
2536                        ..Decision::default()
2537                    },
2538                ],
2539            },
2540        )
2541        .unwrap();
2542
2543        let requeued = queue.get(&requeue_me.id).unwrap();
2544        assert_eq!(requeued.status, TaskStatus::Queued);
2545        assert_eq!(requeued.attempts, 0);
2546
2547        let held = queue.get(&hold_me.id).unwrap();
2548        assert_eq!(held.status, TaskStatus::Held);
2549        assert_eq!(held.hold_reason.as_deref(), Some("looks broken"));
2550    }
2551
2552    #[test]
2553    fn a_machine_held_task_asked_about_restores_to_held_once_answered() {
2554        // Reproduces the reported bug: a task held after burning its
2555        // attempts, once the conductor asks a follow-up question about it,
2556        // must come back `held` - not `queued` - once the question is
2557        // answered, whatever the answer said.
2558        let dir = tempdir().unwrap();
2559        let queue = Queue::at(dir.path().join("queue"));
2560        let questions = Questions::at(dir.path().join("questions"));
2561        let mut t = task("held out of attempts");
2562        t.hold_machine(Some("out of attempts".to_owned()));
2563        queue.put(&mut t).unwrap();
2564
2565        apply(
2566            &queue,
2567            &questions,
2568            &Verdict {
2569                decisions: vec![Decision {
2570                    id: t.id.clone(),
2571                    reason: Some("what should happen to this one?".to_owned()),
2572                    question: Some("Hold it, or try again?".to_owned()),
2573                    ..Decision::default()
2574                }],
2575            },
2576        )
2577        .unwrap();
2578        let blocked = queue.get(&t.id).unwrap();
2579        assert_eq!(blocked.status, TaskStatus::Blocked);
2580        let question_id = blocked.blocked_by[0].clone();
2581
2582        let mut q = questions.get(&question_id).unwrap();
2583        q.answer(Answer::Text("leave it held".to_owned())).unwrap();
2584        questions.put(&mut q).unwrap();
2585
2586        // What `crate::daemon::resolve_blockers` does on the real queue.
2587        let mut after = queue.get(&t.id).unwrap();
2588        after.record_answer(q.summary.clone(), "leave it held".to_owned());
2589        after.unblock(&question_id);
2590        assert_eq!(
2591            after.status,
2592            TaskStatus::Held,
2593            "must not fall back to queued"
2594        );
2595        assert_eq!(after.hold_reason.as_deref(), Some("out of attempts"));
2596    }
2597
2598    #[test]
2599    fn a_reaffirmed_hold_with_no_new_reason_is_not_silently_auto_released_by_triage() {
2600        // A different path to the same bug round 1 fixed: a task
2601        // machine-held for disk pressure, blocked on a conductor question,
2602        // answered "keep it held", and restored to `held`. If the
2603        // conductor's next decision reconfirms `recovery: hold` with no new
2604        // `reason` - allowed, `Decision::reason` is optional - `hold_machine`
2605        // must not silently leave the stale disk-pressure text in place, or
2606        // `crate::triage::run_once` reads it as an unexamined, auto-resolvable
2607        // hold and releases the task straight through the operator's answer
2608        // the instant disk space looks fine again.
2609        let dir = tempdir().unwrap();
2610        let queue = Queue::at(dir.path().join("queue"));
2611        let questions = Questions::at(dir.path().join("questions"));
2612        let mut t = task("disk pressure, then reconsidered");
2613        t.hold_machine(Some(
2614            "not enough free space to start a run: 10 bytes free, 100 required by \
2615             `[disk] min_free_bytes`"
2616                .to_owned(),
2617        ));
2618        t.record_answer(
2619            "How should this be handled?".to_owned(),
2620            "keep it held, a human will look at it later".to_owned(),
2621        );
2622        queue.put(&mut t).unwrap();
2623
2624        apply(
2625            &queue,
2626            &questions,
2627            &Verdict {
2628                decisions: vec![Decision {
2629                    id: t.id.clone(),
2630                    recovery: Some(Recovery::Hold),
2631                    ..Decision::default()
2632                }],
2633            },
2634        )
2635        .unwrap();
2636
2637        let after = queue.get(&t.id).unwrap();
2638        assert_eq!(after.status, TaskStatus::Held);
2639        assert!(
2640            !after
2641                .hold_reason
2642                .as_deref()
2643                .unwrap_or_default()
2644                .starts_with("not enough free space"),
2645            "the stale disk-pressure text must not survive a reconfirmed hold: {:?}",
2646            after.hold_reason
2647        );
2648
2649        // The disk gate would report space is fine now - `crate::triage`
2650        // must not read the old text and release the task through it.
2651        let cfg_dir = tempdir().unwrap();
2652        let config = cfg_dir.path().join("magi.toml");
2653        std::fs::write(&config, "[disk]\nmin_free_bytes = 0\n").unwrap();
2654        let report =
2655            crate::triage::run_once(&queue, &questions, Some(&config), jiff::Timestamp::now());
2656        assert!(report.resumed.is_empty(), "must not be auto-released");
2657        assert_eq!(queue.get(&t.id).unwrap().status, TaskStatus::Held);
2658    }
2659
2660    #[test]
2661    fn explicit_resumption_rehold_requires_new_evidence_and_another_answer() {
2662        crate::run::pin_test_home();
2663        let dir = tempdir().unwrap();
2664        let queue = Queue::at(dir.path().join("queue"));
2665        let questions = Questions::at(dir.path().join("questions"));
2666        let config = dir.path().join("magi.toml");
2667        std::fs::write(&config, "[disk]\nmin_free_bytes = 0\n").unwrap();
2668        let now = jiff::Timestamp::now();
2669        let triage = || crate::triage::run_once(&queue, &questions, Some(&config), now);
2670        let open = || {
2671            questions
2672                .list()
2673                .into_iter()
2674                .filter(|q| q.node == "triage" && q.status.open())
2675                .collect::<Vec<_>>()
2676        };
2677        let hold = Decision {
2678            recovery: Some(Recovery::Hold),
2679            reason: Some("waiting on manual worktree cleanup".to_owned()),
2680            ..Decision::default()
2681        };
2682
2683        let mut t = task("looping hold");
2684        t.hold_machine(Some("waiting on manual worktree cleanup".to_owned()));
2685        queue.put(&mut t).unwrap();
2686        let hold = Decision {
2687            id: t.id.clone(),
2688            ..hold
2689        };
2690
2691        // 1. triage asks the ordinary machine question; the operator resumes.
2692        assert_eq!(triage().asked.len(), 1);
2693        let first = open().remove(0);
2694        let mut q = questions.get(&first.id).unwrap();
2695        let resume = q.choices[0].clone();
2696        q.answer(Answer::Choice(resume)).unwrap();
2697        questions.put(&mut q).unwrap();
2698        assert_eq!(triage().answered.len(), 1);
2699        assert_eq!(queue.get(&t.id).unwrap().status, TaskStatus::Queued);
2700
2701        // Rewording the old cause cannot undo the first explicit answer.
2702        apply(
2703            &queue,
2704            &questions,
2705            &Verdict {
2706                decisions: vec![hold.clone()],
2707            },
2708        )
2709        .unwrap();
2710        assert_eq!(queue.get(&t.id).unwrap().status, TaskStatus::Queued);
2711        assert!(triage().asked.is_empty());
2712
2713        // A new independently filed check can justify a later hold.
2714        let mut evidence = Question::new(
2715            t.id.clone(),
2716            "preflight".into(),
2717            "check".into(),
2718            "New permission failure".into(),
2719            String::new(),
2720            Vec::new(),
2721        );
2722        questions.put(&mut evidence).unwrap();
2723        let hold = Decision {
2724            evidence: vec![evidence.id.clone()],
2725            reason: Some("new permission check failed".into()),
2726            ..hold
2727        };
2728        apply(
2729            &queue,
2730            &questions,
2731            &Verdict {
2732                decisions: vec![hold.clone()],
2733            },
2734        )
2735        .unwrap();
2736        assert_eq!(queue.get(&t.id).unwrap().status, TaskStatus::Held);
2737
2738        // The new hold earns one question identifying the new evidence.
2739        assert_eq!(triage().asked.len(), 1);
2740        let second = open().remove(0);
2741        assert!(second.detail.contains("new permission check failed"));
2742        assert!(triage().asked.is_empty(), "no duplicate question");
2743        assert_eq!(open().len(), 1);
2744
2745        // Another explicit answer supersedes this later hold too.
2746        let mut q = questions.get(&second.id).unwrap();
2747        let force = q.choices[0].clone();
2748        q.answer(Answer::Choice(force)).unwrap();
2749        questions.put(&mut q).unwrap();
2750        assert_eq!(triage().answered.len(), 1);
2751        apply(
2752            &queue,
2753            &questions,
2754            &Verdict {
2755                decisions: vec![hold],
2756            },
2757        )
2758        .unwrap();
2759        assert_eq!(queue.get(&t.id).unwrap().status, TaskStatus::Queued);
2760    }
2761
2762    #[test]
2763    fn conductor_resumption_notice_does_not_solicit_unregistered_questions() {
2764        crate::run::pin_test_home();
2765        let dir = tempdir().unwrap();
2766        let queue = Queue::at(dir.path().join("queue"));
2767        let questions = Questions::at(dir.path().join("questions"));
2768
2769        let mut t = task("task with explicit resumption");
2770        t.status = TaskStatus::Queued;
2771        t.resumption = Some(crate::queue::Resumption {
2772            generation: 1,
2773            source: "test".into(),
2774            question_id: Some("q1".into()),
2775            at: jiff::Timestamp::now(),
2776            instruction: "resume".into(),
2777            answers_through: 0,
2778            runs_through: 0,
2779            cleared_reason: None,
2780            blocking_evidence: Vec::new(),
2781        });
2782        queue.put(&mut t).unwrap();
2783
2784        // Conductor suggests a question and dependencies without new evidence
2785        let decision = Decision {
2786            id: t.id.clone(),
2787            question: Some("How should we proceed with this task?".into()),
2788            blocked_by: vec!["dep-x".into()],
2789            reason: Some("needs clarification".into()),
2790            recovery: Some(Recovery::Hold),
2791            ..Decision::default()
2792        };
2793
2794        apply(
2795            &queue,
2796            &questions,
2797            &Verdict {
2798                decisions: vec![decision],
2799            },
2800        )
2801        .unwrap();
2802
2803        // Task stays Queued (not re-blocked on old evidence)
2804        assert_eq!(queue.get(&t.id).unwrap().status, TaskStatus::Queued);
2805        // No question was filed in the questions store
2806        assert!(questions.list().is_empty());
2807
2808        // A notice was filed with human-readable text and task link
2809        let notices = crate::notices::Notices::open();
2810        let notice = notices
2811            .get(&crate::notices::id_of(&format!(
2812                "conductor-resumption:{}",
2813                t.id
2814            )))
2815            .expect("conductor-resumption notice filed");
2816
2817        assert_eq!(
2818            notice.link,
2819            Some(crate::notices::Link::Task { id: t.id.clone() })
2820        );
2821        // Message does NOT contain the question prose to avoid asking for non-existent answers
2822        assert!(!notice.message.contains("How should we proceed"));
2823        // Message explicitly notes that no reply is required
2824        assert!(
2825            notice
2826                .message
2827                .contains("No reply to this notice is needed.")
2828        );
2829        // No raw Rust debug representations
2830        assert!(!notice.message.contains("Some("));
2831        assert!(!notice.message.contains("None"));
2832        assert!(!notice.message.contains("[]"));
2833        assert!(!notice.message.contains("{:?}"));
2834        // Mention proposed dependencies cleanly
2835        assert!(notice.message.contains("Proposed dependencies: dep-x."));
2836    }
2837
2838    #[test]
2839    fn a_runnable_task_can_be_held_directly_without_a_question() {
2840        let dir = tempdir().unwrap();
2841        let queue = Queue::at(dir.path().join("queue"));
2842        let questions = Questions::at(dir.path().join("questions"));
2843        let mut t = task("already answered, should stay put");
2844        queue.put(&mut t).unwrap();
2845
2846        apply(
2847            &queue,
2848            &questions,
2849            &Verdict {
2850                decisions: vec![Decision {
2851                    id: t.id.clone(),
2852                    recovery: Some(Recovery::Hold),
2853                    reason: Some("operator already said keep this held".to_owned()),
2854                    ..Decision::default()
2855                }],
2856            },
2857        )
2858        .unwrap();
2859
2860        let after = queue.get(&t.id).unwrap();
2861        assert_eq!(after.status, TaskStatus::Held);
2862        assert_eq!(after.hold_source, Some(crate::queue::HoldSource::Machine));
2863    }
2864
2865    #[test]
2866    fn done_recovery_closes_a_held_task_whose_goal_is_already_met() {
2867        crate::run::pin_test_home();
2868        // Reproduces the other half of the reported bug (task 6081): once
2869        // the operator's answer says the work already happened outside the
2870        // loop - PR merged, worktree cleaned up - the conductor needs an
2871        // actual terminal state to put the task in, not just a hold it will
2872        // keep being re-asked about.
2873        let dir = tempdir().unwrap();
2874        let queue = Queue::at(dir.path().join("queue"));
2875        let questions = Questions::at(dir.path().join("questions"));
2876        let mut t = task("already merged by hand");
2877        t.hold_machine(Some("branch survived, awaiting a decision".to_owned()));
2878        t.record_answer(
2879            "Handle this one?".to_owned(),
2880            "already merged and cleaned up, close it".to_owned(),
2881        );
2882        queue.put(&mut t).unwrap();
2883
2884        apply(
2885            &queue,
2886            &questions,
2887            &Verdict {
2888                decisions: vec![Decision {
2889                    id: t.id.clone(),
2890                    recovery: Some(Recovery::Done),
2891                    reason: Some("operator confirmed this already landed".to_owned()),
2892                    ..Decision::default()
2893                }],
2894            },
2895        )
2896        .unwrap();
2897
2898        let after = queue.get(&t.id).unwrap();
2899        assert_eq!(after.status, TaskStatus::Done);
2900        assert!(after.hold_reason.is_none());
2901        assert_eq!(after.answers.len(), 1, "the record of why is kept");
2902    }
2903
2904    #[test]
2905    fn done_recovery_is_ignored_for_a_runnable_or_running_task() {
2906        let dir = tempdir().unwrap();
2907        let queue = Queue::at(dir.path().join("queue"));
2908        let questions = Questions::at(dir.path().join("questions"));
2909
2910        let mut queued = task("never ran yet");
2911        queue.put(&mut queued).unwrap();
2912
2913        let mut running = task("mid-run");
2914        running.start("run-1".to_owned());
2915        queue.put(&mut running).unwrap();
2916
2917        for id in [queued.id.clone(), running.id.clone()] {
2918            apply(
2919                &queue,
2920                &questions,
2921                &Verdict {
2922                    decisions: vec![Decision {
2923                        id,
2924                        recovery: Some(Recovery::Done),
2925                        ..Decision::default()
2926                    }],
2927                },
2928            )
2929            .unwrap();
2930        }
2931
2932        assert_eq!(queue.get(&queued.id).unwrap().status, TaskStatus::Queued);
2933        assert_eq!(queue.get(&running.id).unwrap().status, TaskStatus::Running);
2934    }
2935
2936    #[test]
2937    fn a_question_after_two_settled_answers_is_still_filed_and_blocks() {
2938        // A genuine question must stay answerable however many were settled
2939        // before it: refusing it left `magi answer --list` empty.
2940        let dir = tempdir().unwrap();
2941        let queue = Queue::at(dir.path().join("queue"));
2942        let questions = Questions::at(dir.path().join("questions"));
2943        let mut t = task("asked about repeatedly");
2944        t.hold_machine(Some("out of attempts".to_owned()));
2945        t.record_answer("Handle this one? (1)".to_owned(), "not yet".to_owned());
2946        t.record_answer(
2947            "Handle this one? (2)".to_owned(),
2948            "still not yet".to_owned(),
2949        );
2950        queue.put(&mut t).unwrap();
2951        assert_eq!(questions.list().len(), 0);
2952
2953        let verdict = Verdict {
2954            decisions: vec![Decision {
2955                id: t.id.clone(),
2956                question: Some("Branch conflicts with origin/main, how do we proceed?".to_owned()),
2957                ..Decision::default()
2958            }],
2959        };
2960        apply(&queue, &questions, &verdict).unwrap();
2961
2962        let filed = questions.list();
2963        assert_eq!(filed.len(), 1, "the question was filed");
2964        assert_eq!(
2965            filed[0].summary,
2966            "Branch conflicts with origin/main, how do we proceed?"
2967        );
2968        let deputy = filed[0]
2969            .deputy
2970            .as_ref()
2971            .expect("the wait is handed to a deputy");
2972        assert!(deputy.brief.contains(&t.id), "{}", deputy.brief);
2973        assert_eq!(
2974            deputy.starts, 0,
2975            "filing never starts anything: the loop does not wait"
2976        );
2977        assert!(filed[0].status.open());
2978        assert_eq!(filed[0].node, NODE);
2979        let after = queue.get(&t.id).unwrap();
2980        assert_eq!(after.status, TaskStatus::Blocked);
2981        assert_eq!(after.blocked_by, vec![filed[0].id.clone()]);
2982        assert_eq!(after.answers.len(), 2, "the prior answers are untouched");
2983        assert!(
2984            !after
2985                .hold_reason
2986                .clone()
2987                .unwrap_or_default()
2988                .contains("conduct tried to ask"),
2989            "no hold was applied"
2990        );
2991
2992        apply(&queue, &questions, &verdict).unwrap();
2993        let again = questions.list();
2994        assert_eq!(again.len(), 1, "the open question is reused");
2995        assert_eq!(again[0].id, filed[0].id);
2996    }
2997
2998    #[test]
2999    fn question_source_identity_resolves_aliases_and_rejects_invented_targets() {
3000        let dir = tempdir().unwrap();
3001        init_repo_with_branch(dir.path(), "source-a");
3002        let mut t = task("source decision");
3003        t.repo = dir.path().into();
3004        let decision = |reference: &str| QuestionDecision {
3005            operation: QuestionOperation::Recover,
3006            target: QuestionTarget::Source {
3007                reference: reference.into(),
3008            },
3009        };
3010        assert_eq!(
3011            decision("source-a").identity(&t).unwrap(),
3012            decision("refs/heads/source-a").identity(&t).unwrap()
3013        );
3014        assert_ne!(
3015            decision("source-a").identity(&t).unwrap(),
3016            decision("main").identity(&t).unwrap()
3017        );
3018        assert!(decision("invented-source-label").identity(&t).is_err());
3019        let unrelated = QuestionDecision {
3020            operation: QuestionOperation::Resume,
3021            target: QuestionTarget::Run {
3022                id: "unrelated-run".into(),
3023            },
3024        };
3025        assert!(unrelated.identity(&t).is_err());
3026    }
3027
3028    #[test]
3029    fn invalid_question_targets_still_file_and_block_queued_tasks() {
3030        let dir = tempdir().unwrap();
3031        init_repo_with_branch(dir.path(), "source-a");
3032        let queue = Queue::at(dir.path().join("queue"));
3033        let questions = Questions::at(dir.path().join("questions"));
3034        for target in [
3035            QuestionTarget::Source {
3036                reference: "missing-source".into(),
3037            },
3038            QuestionTarget::Run {
3039                id: "unrelated-run".into(),
3040            },
3041        ] {
3042            let mut t = task("decision with invalid metadata");
3043            t.repo = dir.path().into();
3044            queue.put(&mut t).unwrap();
3045            let d = Decision {
3046                id: t.id.clone(),
3047                question: Some("Which recovery should proceed?".into()),
3048                choices: vec!["inspect first".into(), "keep waiting".into()],
3049                question_decision: Some(QuestionDecision {
3050                    operation: QuestionOperation::Recover,
3051                    target,
3052                }),
3053                ..Decision::default()
3054            };
3055            apply_one(&queue, &questions, &d, None).unwrap();
3056            let q = questions
3057                .list()
3058                .into_iter()
3059                .find(|q| q.run == t.id)
3060                .unwrap();
3061            assert_eq!(q.summary, d.question.as_deref().unwrap());
3062            assert_eq!(q.choices, d.choices);
3063            assert!(q.status.open());
3064            let context: serde_json::Value =
3065                serde_json::from_str(q.decision_context.as_ref().unwrap()).unwrap();
3066            assert!(context[3].is_null(), "invalid metadata was discarded");
3067            let after = queue.get(&t.id).unwrap();
3068            assert_eq!(after.status, TaskStatus::Blocked);
3069            assert_eq!(after.blocked_by, vec![q.id.clone()]);
3070            apply_one(&queue, &questions, &d, None).unwrap();
3071            assert_eq!(questions.list().iter().filter(|q| q.run == t.id).count(), 1);
3072        }
3073    }
3074
3075    #[test]
3076    fn answered_conductor_question_needs_changed_facts_before_reasking() {
3077        let dir = tempdir().unwrap();
3078        let queue = Queue::at(dir.path().join("queue"));
3079        let questions = Questions::at(dir.path().join("questions"));
3080        let mut t = task("recovery decision");
3081        t.hold_machine(Some("out of attempts".into()));
3082        t.runs = vec!["source-run-a".into(), "source-run-b".into()];
3083        queue.put(&mut t).unwrap();
3084        let mut d = Decision {
3085            id: t.id.clone(),
3086            question: Some("Recover from source A?".into()),
3087            choices: vec!["approve source A".into(), "keep held".into()],
3088            question_decision: Some(QuestionDecision {
3089                operation: QuestionOperation::Recover,
3090                target: QuestionTarget::Run {
3091                    id: "source-run-a".into(),
3092                },
3093            }),
3094            ..Decision::default()
3095        };
3096        apply_one(&queue, &questions, &d, None).unwrap();
3097        let mut q = questions.list().pop().unwrap();
3098        q.answer(crate::ask::Answer::Choice("keep held".into()))
3099            .unwrap();
3100        questions.put(&mut q).unwrap();
3101        queue
3102            .modify(&t.id, |task| {
3103                task.unblock(&q.id);
3104                task.hold_machine(Some("recovery remains under consideration".into()));
3105                true
3106            })
3107            .unwrap();
3108        d.question = Some("Should we keep waiting on recovery?".into());
3109        d.choices = vec!["recover using source A".into(), "remain on hold".into()];
3110        apply_one(&queue, &questions, &d, None).unwrap();
3111        assert_eq!(questions.list().len(), 1, "no repeated operator turn");
3112        d.question = Some("Recover from source B?".into());
3113        // The labels can remain identical while the approval target changes.
3114        d.question_decision.as_mut().unwrap().target = QuestionTarget::Run {
3115            id: "source-run-b".into(),
3116        };
3117        apply_one(&queue, &questions, &d, None).unwrap();
3118        assert_eq!(
3119            questions.list().len(),
3120            2,
3121            "new decision with unchanged task facts"
3122        );
3123        let after = queue.get(&t.id).unwrap();
3124        assert_eq!(after.instruction, t.instruction);
3125        assert_eq!(after.runs, t.runs);
3126        assert_eq!(after.status, TaskStatus::Blocked);
3127    }
3128
3129    #[test]
3130    fn a_second_conductor_question_is_still_allowed_after_one_settled_answer() {
3131        let dir = tempdir().unwrap();
3132        let queue = Queue::at(dir.path().join("queue"));
3133        let questions = Questions::at(dir.path().join("questions"));
3134        let mut t = task("asked about once already");
3135        t.hold_machine(Some("out of attempts".to_owned()));
3136        t.record_answer("Handle this one?".to_owned(), "not yet".to_owned());
3137        queue.put(&mut t).unwrap();
3138
3139        apply(
3140            &queue,
3141            &questions,
3142            &Verdict {
3143                decisions: vec![Decision {
3144                    id: t.id.clone(),
3145                    question: Some("Still not sure - now what?".to_owned()),
3146                    ..Decision::default()
3147                }],
3148            },
3149        )
3150        .unwrap();
3151
3152        assert_eq!(questions.list().len(), 1, "the second question was filed");
3153        assert_eq!(queue.get(&t.id).unwrap().status, TaskStatus::Blocked);
3154    }
3155
3156    #[test]
3157    fn a_held_task_with_an_open_triage_question_is_left_to_triage() {
3158        // `crate::triage` never calls `Task::block`, and only ever looks at
3159        // tasks still `held` - so a conductor question that moved this task
3160        // to `blocked` would orphan triage's own open question. The guard in
3161        // `apply_one` must leave the task alone until triage's question
3162        // settles.
3163        let dir = tempdir().unwrap();
3164        let queue = Queue::at(dir.path().join("queue"));
3165        let questions = Questions::at(dir.path().join("questions"));
3166        let mut t = task("held, triage already asking about it");
3167        t.hold_machine(Some("cause unclear".to_owned()));
3168        queue.put(&mut t).unwrap();
3169
3170        let mut triage_q = Question::new(
3171            t.id.clone(),
3172            crate::triage::NODE.to_owned(),
3173            "triage".to_owned(),
3174            "Still needed?".to_owned(),
3175            String::new(),
3176            vec![
3177                "resume".to_owned(),
3178                "not yet".to_owned(),
3179                "discard".to_owned(),
3180            ],
3181        );
3182        questions.put(&mut triage_q).unwrap();
3183
3184        for decision in [
3185            Decision {
3186                id: t.id.clone(),
3187                question: Some("what now?".to_owned()),
3188                ..Decision::default()
3189            },
3190            Decision {
3191                id: t.id.clone(),
3192                recovery: Some(Recovery::Requeue),
3193                ..Decision::default()
3194            },
3195        ] {
3196            apply(
3197                &queue,
3198                &questions,
3199                &Verdict {
3200                    decisions: vec![decision],
3201                },
3202            )
3203            .unwrap();
3204        }
3205
3206        let after = queue.get(&t.id).unwrap();
3207        assert_eq!(
3208            after.status,
3209            TaskStatus::Held,
3210            "triage still owns this hold"
3211        );
3212        assert!(after.blocked_by.is_empty());
3213        assert_eq!(
3214            questions.list().len(),
3215            1,
3216            "no second, conductor-owned question was filed"
3217        );
3218    }
3219
3220    #[test]
3221    fn a_held_task_with_an_answered_but_unapplied_triage_question_is_still_left_alone() {
3222        // The race `crate::triage::pending_for` exists to close: the
3223        // operator has already answered the triage question (it is no
3224        // longer `open`), but `crate::triage::run_once` - which only runs on
3225        // a fully idle daemon tick, far less often than the conductor polls
3226        // - has not had a turn to apply it yet. `apply_one` must not treat
3227        // "not open" as "settled" here, or it would block the task out from
3228        // under an answer triage has not read back yet.
3229        let dir = tempdir().unwrap();
3230        let queue = Queue::at(dir.path().join("queue"));
3231        let questions = Questions::at(dir.path().join("questions"));
3232        let mut t = task("held, triage question answered but not yet applied");
3233        t.hold_machine(Some("cause unclear".to_owned()));
3234        queue.put(&mut t).unwrap();
3235
3236        let mut triage_q = Question::new(
3237            t.id.clone(),
3238            crate::triage::NODE.to_owned(),
3239            "triage".to_owned(),
3240            "Still needed?".to_owned(),
3241            String::new(),
3242            vec![
3243                "resume".to_owned(),
3244                "not yet".to_owned(),
3245                "discard".to_owned(),
3246            ],
3247        );
3248        questions.put(&mut triage_q).unwrap();
3249        triage_q
3250            .answer(Answer::Choice("not yet".to_owned()))
3251            .unwrap();
3252        questions.put(&mut triage_q).unwrap();
3253        assert!(!triage_q.status.open());
3254
3255        apply(
3256            &queue,
3257            &questions,
3258            &Verdict {
3259                decisions: vec![Decision {
3260                    id: t.id.clone(),
3261                    question: Some("what now?".to_owned()),
3262                    ..Decision::default()
3263                }],
3264            },
3265        )
3266        .unwrap();
3267
3268        let after = queue.get(&t.id).unwrap();
3269        assert_eq!(
3270            after.status,
3271            TaskStatus::Held,
3272            "triage's own answer is not yet applied - conduct must wait"
3273        );
3274        assert_eq!(
3275            questions.list().len(),
3276            1,
3277            "no conductor question was filed over the pending triage answer"
3278        );
3279    }
3280
3281    #[test]
3282    fn manual_hold_rejects_hostile_or_stale_conductor_recovery() {
3283        let dir = tempdir().unwrap();
3284        let queue = Queue::at(dir.path().join("queue"));
3285        let questions = Questions::at(dir.path().join("questions"));
3286        let mut held = task("manual recovery");
3287        held.priority = 300;
3288        held.runs.push("run20260912-224242-daf5".to_owned());
3289        held.hold_manual(Some(
3290            "active manual recovery run20260912-224242-daf5".to_owned(),
3291        ));
3292        queue.put(&mut held).unwrap();
3293
3294        // Every field a conductor may use to alter lifecycle state is ignored:
3295        // requeue/review would dispatch duplicate work, hold could overwrite
3296        // evidence, and a question would turn the hold into `blocked`.
3297        for decision in [
3298            Decision {
3299                id: held.id.clone(),
3300                recovery: Some(Recovery::Requeue),
3301                ..Decision::default()
3302            },
3303            Decision {
3304                id: held.id.clone(),
3305                recovery: Some(Recovery::Hold),
3306                reason: Some("stale replacement reason".to_owned()),
3307                ..Decision::default()
3308            },
3309            Decision {
3310                id: held.id.clone(),
3311                recovery: Some(Recovery::Review),
3312                ..Decision::default()
3313            },
3314            Decision {
3315                id: held.id.clone(),
3316                blocked_by: vec!["other-task".to_owned()],
3317                question: Some("retry now?".to_owned()),
3318                ..Decision::default()
3319            },
3320        ] {
3321            apply(
3322                &queue,
3323                &questions,
3324                &Verdict {
3325                    decisions: vec![decision],
3326                },
3327            )
3328            .unwrap();
3329        }
3330
3331        let after = queue.get(&held.id).unwrap();
3332        assert_eq!(after.status, TaskStatus::Held);
3333        assert!(after.operator_held());
3334        assert_eq!(after.priority, 300);
3335        assert_eq!(after.runs, ["run20260912-224242-daf5"]);
3336        assert_eq!(
3337            after.hold_reason.as_deref(),
3338            Some("active manual recovery run20260912-224242-daf5")
3339        );
3340        assert!(after.blocked_by.is_empty());
3341        assert!(questions.list().is_empty());
3342        assert!(
3343            queue.next_runnable().is_none(),
3344            "must not dispatch a duplicate"
3345        );
3346    }
3347
3348    #[test]
3349    fn machine_holds_remain_recoverable_and_manual_release_is_authorization() {
3350        let dir = tempdir().unwrap();
3351        let queue = Queue::at(dir.path().join("queue"));
3352        let questions = Questions::at(dir.path().join("questions"));
3353
3354        let mut automatic = task("disk gate");
3355        automatic.hold_machine(Some("disk full".to_owned()));
3356        queue.put(&mut automatic).unwrap();
3357        let requeue = || Verdict {
3358            decisions: vec![Decision {
3359                id: automatic.id.clone(),
3360                recovery: Some(Recovery::Requeue),
3361                ..Decision::default()
3362            }],
3363        };
3364        apply(&queue, &questions, &requeue()).unwrap();
3365        assert_eq!(queue.get(&automatic.id).unwrap().status, TaskStatus::Queued);
3366
3367        let mut manual = task("operator gate");
3368        manual.hold_manual(Some("wait for operator".to_owned()));
3369        queue.put(&mut manual).unwrap();
3370        apply(
3371            &queue,
3372            &questions,
3373            &Verdict {
3374                decisions: vec![Decision {
3375                    id: manual.id.clone(),
3376                    recovery: Some(Recovery::Requeue),
3377                    ..Decision::default()
3378                }],
3379            },
3380        )
3381        .unwrap();
3382        assert_eq!(queue.get(&manual.id).unwrap().status, TaskStatus::Held);
3383
3384        // This mirrors the CLI and web release routes: only an explicit
3385        // operator action clears the manual boundary.
3386        let mut released = queue.get(&manual.id).unwrap();
3387        released.release();
3388        queue.put(&mut released).unwrap();
3389        assert_eq!(queue.get(&manual.id).unwrap().status, TaskStatus::Queued);
3390    }
3391
3392    #[test]
3393    fn legacy_reasoned_hold_is_protected_without_losing_its_metadata() {
3394        let dir = tempdir().unwrap();
3395        let queue = Queue::at(dir.path().join("queue"));
3396        let questions = Questions::at(dir.path().join("questions"));
3397        let mut legacy = task("old explicit hold");
3398        legacy.status = TaskStatus::Held;
3399        legacy.hold_reason = Some("manual recovery already active".to_owned());
3400        legacy.hold_source = None;
3401        legacy.blocked_by = vec!["dependency".to_owned()];
3402        queue.put(&mut legacy).unwrap();
3403
3404        apply(
3405            &queue,
3406            &questions,
3407            &Verdict {
3408                decisions: vec![Decision {
3409                    id: legacy.id.clone(),
3410                    recovery: Some(Recovery::Requeue),
3411                    ..Decision::default()
3412                }],
3413            },
3414        )
3415        .unwrap();
3416
3417        let after = queue.get(&legacy.id).unwrap();
3418        assert_eq!(after.status, TaskStatus::Held);
3419        assert_eq!(after.hold_source, None);
3420        assert_eq!(after.hold_reason, legacy.hold_reason);
3421        assert_eq!(after.blocked_by, legacy.blocked_by);
3422    }
3423
3424    #[test]
3425    fn review_recovery_is_a_no_op_without_a_survivable_branch() {
3426        // `surviving_branch` reaches `RunState::load`, which reaches the
3427        // process-global `run::home()` - a `OnceLock`, so this only wins the
3428        // race the first time it runs in the binary; every other test still
3429        // reaches the same directory whichever call won, and this test's own
3430        // run id never collides with another test's.
3431        crate::run::pin_test_home();
3432        let dir = tempdir().unwrap();
3433        let queue = Queue::at(dir.path().join("queue"));
3434        let questions = Questions::at(dir.path().join("questions"));
3435        let mut t = task("blocked with no readable run");
3436        t.start("20260101-000000-dead".to_owned()); // no such run on disk
3437        t.fail("blocked", 5);
3438        queue.put(&mut t).unwrap();
3439
3440        apply(
3441            &queue,
3442            &questions,
3443            &Verdict {
3444                decisions: vec![Decision {
3445                    id: t.id.clone(),
3446                    recovery: Some(Recovery::Review),
3447                    ..Decision::default()
3448                }],
3449            },
3450        )
3451        .unwrap();
3452
3453        let after = queue.get(&t.id).unwrap();
3454        assert_eq!(
3455            after.status,
3456            TaskStatus::Failed,
3457            "with nothing to reopen, the decision is dropped rather than guessed at"
3458        );
3459        assert!(after.review_branch.is_none());
3460    }
3461
3462    #[test]
3463    fn requeue_and_review_recovery_are_ignored_for_a_runnable_task() {
3464        // `Hold` is the one exception - see `Decision::recovery`'s doc and
3465        // `a_runnable_task_can_be_held_directly_without_a_question` - but a
3466        // task already in line has nothing for `requeue` or `review` to do.
3467        let dir = tempdir().unwrap();
3468        let queue = Queue::at(dir.path().join("queue"));
3469        let questions = Questions::at(dir.path().join("questions"));
3470
3471        for recovery in [Recovery::Requeue, Recovery::Review] {
3472            let mut t = task("ordinary");
3473            queue.put(&mut t).unwrap();
3474
3475            apply(
3476                &queue,
3477                &questions,
3478                &Verdict {
3479                    decisions: vec![Decision {
3480                        id: t.id.clone(),
3481                        recovery: Some(recovery),
3482                        ..Decision::default()
3483                    }],
3484                },
3485            )
3486            .unwrap();
3487
3488            assert_eq!(queue.get(&t.id).unwrap().status, TaskStatus::Queued);
3489        }
3490    }
3491
3492    #[tokio::test]
3493    async fn a_broken_agent_leaves_the_queue_untouched_and_does_not_error() {
3494        let dir = tempdir().unwrap();
3495        let cfg = config(mock_agent(dir.path(), BROKEN, BTreeMap::new()));
3496        let queue = Queue::at(dir.path().join("queue"));
3497        let questions = Questions::at(dir.path().join("questions"));
3498        let mut t = task("normal");
3499        queue.put(&mut t).unwrap();
3500
3501        let mut conductor = Conductor::new();
3502        conductor
3503            .maybe_run(
3504                &cfg,
3505                dir.path(),
3506                &queue,
3507                &questions,
3508                dir.path(),
3509                &[t.clone()],
3510                &[],
3511                &[],
3512                2,
3513            )
3514            .await;
3515
3516        assert_eq!(
3517            queue.get(&t.id).unwrap().status,
3518            TaskStatus::Queued,
3519            "a failed invocation must change nothing"
3520        );
3521        assert!(
3522            queue.next_runnable().is_some(),
3523            "the loop must still be able to take the next task"
3524        );
3525    }
3526
3527    #[tokio::test]
3528    async fn a_reply_with_no_json_leaves_the_queue_untouched() {
3529        let dir = tempdir().unwrap();
3530        let cfg = config(mock_agent(dir.path(), GARBAGE, BTreeMap::new()));
3531        let queue = Queue::at(dir.path().join("queue"));
3532        let questions = Questions::at(dir.path().join("questions"));
3533        let mut t = task("normal");
3534        queue.put(&mut t).unwrap();
3535
3536        let mut conductor = Conductor::new();
3537        conductor
3538            .maybe_run(
3539                &cfg,
3540                dir.path(),
3541                &queue,
3542                &questions,
3543                dir.path(),
3544                &[t.clone()],
3545                &[],
3546                &[],
3547                2,
3548            )
3549            .await;
3550
3551        assert_eq!(queue.get(&t.id).unwrap().status, TaskStatus::Queued);
3552    }
3553
3554    #[tokio::test]
3555    async fn a_conductor_chain_falls_to_the_next_agent_once_each() {
3556        let dir = tempdir().unwrap();
3557        let mut t = task("chained");
3558        let reply = format!(
3559            "{{\"decisions\":[{{\"id\":\"{}\",\"blocked_by\":[\"x\"],\"reason\":\"why\"}}]}}",
3560            t.id
3561        );
3562        let counted = |id: &str, tail: &str| {
3563            let calls = dir.path().join(format!("{id}.calls"));
3564            let path = dir.path().join(format!("{id}.sh"));
3565            std::fs::write(
3566                &path,
3567                format!(
3568                    "#!/bin/sh\ncat >/dev/null\necho x >> '{}'\n{tail}\n",
3569                    calls.display()
3570                ),
3571            )
3572            .unwrap();
3573            AgentSpec {
3574                id: id.to_owned(),
3575                kind: AgentKind::Command,
3576                model: None,
3577                command: vec!["sh".to_owned(), path.to_string_lossy().into_owned()],
3578                extra_args: Vec::new(),
3579                env: BTreeMap::new(),
3580                prompt_delivery: None,
3581            }
3582        };
3583        let n_calls = |id: &str| {
3584            std::fs::read_to_string(dir.path().join(format!("{id}.calls")))
3585                .map_or(0, |s| s.lines().count())
3586        };
3587        let mut cfg = config(counted("a", "exit 3"));
3588        cfg.agents = vec![
3589            counted("a", "exit 3"),
3590            counted("b", &format!("printf '%s' '{reply}'")),
3591        ];
3592        cfg.roles.conductor = Some(crate::config::AgentChoice::Chain(vec![
3593            "a".into(),
3594            "b".into(),
3595            "a".into(),
3596        ]));
3597        let queue = Queue::at(dir.path().join("queue"));
3598        let questions = Questions::at(dir.path().join("questions"));
3599        queue.put(&mut t).unwrap();
3600
3601        let mut conductor = Conductor::new();
3602        conductor
3603            .maybe_run(
3604                &cfg,
3605                dir.path(),
3606                &queue,
3607                &questions,
3608                dir.path(),
3609                &[t.clone()],
3610                &[],
3611                &[],
3612                2,
3613            )
3614            .await;
3615
3616        assert_eq!(queue.get(&t.id).unwrap().status, TaskStatus::Blocked);
3617        assert_eq!(n_calls("a"), 1, "each id is tried once");
3618        assert_eq!(n_calls("b"), 1);
3619        let saved = std::fs::read_to_string(seat_path(dir.path())).unwrap();
3620        assert!(
3621            saved.contains("\"b\""),
3622            "the seat on disk is the one that ran: {saved}"
3623        );
3624    }
3625
3626    #[tokio::test]
3627    async fn json_survives_code_fences_and_a_preamble() {
3628        let dir = tempdir().unwrap();
3629        let mut t = task("fenced");
3630        let reply = format!(
3631            "Sure, here is my decision.\n\n```json\n{{\"decisions\":[{{\"id\":\"{}\",\
3632             \"blocked_by\":[\"x\"],\"reason\":\"why\"}}]}}\n```\n",
3633            t.id
3634        );
3635        let cfg = config(mock_agent(dir.path(), REPLY, env(&reply)));
3636        let queue = Queue::at(dir.path().join("queue"));
3637        let questions = Questions::at(dir.path().join("questions"));
3638        queue.put(&mut t).unwrap();
3639
3640        let mut conductor = Conductor::new();
3641        conductor
3642            .maybe_run(
3643                &cfg,
3644                dir.path(),
3645                &queue,
3646                &questions,
3647                dir.path(),
3648                &[t.clone()],
3649                &[],
3650                &[],
3651                2,
3652            )
3653            .await;
3654
3655        let back = queue.get(&t.id).unwrap();
3656        assert_eq!(back.status, TaskStatus::Blocked);
3657        assert_eq!(back.blocked_by, ["x"]);
3658    }
3659
3660    #[tokio::test]
3661    async fn the_conductor_is_not_called_again_when_nothing_worth_looking_at_has_changed() {
3662        // Each real invocation writes its own artifact stem, `turn-<n>`, so
3663        // whether a second one happened is read off the artifacts directory.
3664        let dir = tempdir().unwrap();
3665        let cfg = config(mock_agent(dir.path(), REPLY, env("{\"decisions\":[]}")));
3666        let queue = Queue::at(dir.path().join("queue"));
3667        let questions = Questions::at(dir.path().join("questions"));
3668        let mut t = task("stable");
3669        queue.put(&mut t).unwrap();
3670        let artifacts = dir.path().join("conduct").join("artifacts");
3671        let turn = |n: usize| artifacts.join(format!("turn-{n}.out"));
3672
3673        let mut conductor = Conductor::new();
3674        conductor
3675            .maybe_run(
3676                &cfg,
3677                dir.path(),
3678                &queue,
3679                &questions,
3680                dir.path(),
3681                &[t.clone()],
3682                &[],
3683                &[],
3684                2,
3685            )
3686            .await;
3687        assert!(turn(1).is_file(), "the first cycle must call the conductor");
3688
3689        conductor
3690            .maybe_run(
3691                &cfg,
3692                dir.path(),
3693                &queue,
3694                &questions,
3695                dir.path(),
3696                &[t.clone()],
3697                &[],
3698                &[],
3699                2,
3700            )
3701            .await;
3702        assert!(
3703            !turn(2).is_file(),
3704            "an unchanged revision and an unchanged stalled/finished set must not call the \
3705             conductor twice"
3706        );
3707
3708        // Once the queue actually changes, the next `maybe_run` calls again.
3709        t.priority = 1;
3710        queue.put(&mut t).unwrap();
3711        conductor
3712            .maybe_run(
3713                &cfg,
3714                dir.path(),
3715                &queue,
3716                &questions,
3717                dir.path(),
3718                &[t.clone()],
3719                &[],
3720                &[],
3721                2,
3722            )
3723            .await;
3724        assert!(turn(2).is_file(), "a moved revision calls it again");
3725    }
3726
3727    #[tokio::test]
3728    async fn a_task_turning_stalled_calls_the_conductor_again_despite_an_unchanged_revision() {
3729        // The queue's own revision has not moved - nothing wrote to it - but
3730        // a task now looks stalled, purely because time passed. Calling
3731        // again here, and never again once this exact set has been shown
3732        // once, is the whole point of keying `worth_a_look` on the id set
3733        // rather than on "is it non-empty".
3734        let dir = tempdir().unwrap();
3735        let cfg = config(mock_agent(dir.path(), REPLY, env("{\"decisions\":[]}")));
3736        let queue = Queue::at(dir.path().join("queue"));
3737        let questions = Questions::at(dir.path().join("questions"));
3738        let mut t = task("quiet");
3739        queue.put(&mut t).unwrap();
3740        let artifacts = dir.path().join("conduct").join("artifacts");
3741        let turn = |n: usize| artifacts.join(format!("turn-{n}.out"));
3742
3743        let mut conductor = Conductor::new();
3744        conductor
3745            .maybe_run(
3746                &cfg,
3747                dir.path(),
3748                &queue,
3749                &questions,
3750                dir.path(),
3751                &[t.clone()],
3752                &[],
3753                &[],
3754                2,
3755            )
3756            .await;
3757        assert!(turn(1).is_file());
3758
3759        conductor
3760            .maybe_run(
3761                &cfg,
3762                dir.path(),
3763                &queue,
3764                &questions,
3765                dir.path(),
3766                &[],
3767                &[t.clone()],
3768                &[],
3769                2,
3770            )
3771            .await;
3772        assert!(
3773            turn(2).is_file(),
3774            "a task turning stalled must call the conductor again"
3775        );
3776
3777        // But once shown at this exact revision, showing the *same* stalled
3778        // set again must not call a third time.
3779        conductor
3780            .maybe_run(
3781                &cfg,
3782                dir.path(),
3783                &queue,
3784                &questions,
3785                dir.path(),
3786                &[],
3787                &[t.clone()],
3788                &[],
3789                2,
3790            )
3791            .await;
3792        assert!(
3793            !turn(3).is_file(),
3794            "the same stalled task lingering must not call the conductor every cycle"
3795        );
3796    }
3797
3798    fn rehold(queue: &Queue, questions: &Questions, id: &str, note: &str) {
3799        apply(
3800            queue,
3801            questions,
3802            &Verdict {
3803                decisions: vec![Decision {
3804                    id: id.to_owned(),
3805                    reason: Some(note.to_owned()),
3806                    recovery: Some(Recovery::Hold),
3807                    ..Decision::default()
3808                }],
3809            },
3810        )
3811        .unwrap();
3812    }
3813
3814    #[test]
3815    fn reholding_with_an_identical_note_changes_nothing() {
3816        let dir = tempdir().unwrap();
3817        let queue = Queue::at(dir.path().join("queue"));
3818        let questions = Questions::at(dir.path().join("questions"));
3819        let mut t = task("held");
3820        t.hold_machine(Some("waiting on a human".to_owned()));
3821        queue.put(&mut t).unwrap();
3822
3823        rehold(&queue, &questions, &t.id, "owner said keep it held");
3824        let first = queue.get(&t.id).unwrap();
3825        let rev = queue.revision();
3826        for _ in 0..5 {
3827            rehold(&queue, &questions, &t.id, "owner said keep it held");
3828        }
3829        let after = queue.get(&t.id).unwrap();
3830        assert_eq!(after.hold_reason, first.hold_reason);
3831        assert_eq!(after.updated_at, first.updated_at);
3832        assert_eq!(queue.revision(), rev, "nothing was written");
3833    }
3834
3835    #[test]
3836    fn reholding_with_different_notes_keeps_one_previously_level() {
3837        let dir = tempdir().unwrap();
3838        let queue = Queue::at(dir.path().join("queue"));
3839        let questions = Questions::at(dir.path().join("questions"));
3840        let mut t = task("held");
3841        t.hold_machine(Some("original".to_owned()));
3842        queue.put(&mut t).unwrap();
3843
3844        for i in 0..20 {
3845            rehold(&queue, &questions, &t.id, &format!("note {}", i % 2));
3846        }
3847        let reason = queue.get(&t.id).unwrap().hold_reason.unwrap();
3848        assert_eq!(reason.matches("(previously:").count(), 1, "{reason}");
3849        assert!(reason.starts_with("note "), "the new note leads: {reason}");
3850    }
3851
3852    #[test]
3853    fn an_already_nested_reason_is_matched_by_its_outermost_note() {
3854        let mut t = task("deep");
3855        let nested = "same\n\n(previously: same\n\n(previously: same))";
3856        t.hold_machine(Some(nested.to_owned()));
3857        let d = Decision {
3858            id: t.id.clone(),
3859            reason: Some("same".to_owned()),
3860            ..Decision::default()
3861        };
3862        assert_eq!(reaffirmed_hold_reason(&t, &d), None);
3863        let d2 = Decision {
3864            reason: Some("other".to_owned()),
3865            ..d
3866        };
3867        assert_eq!(
3868            reaffirmed_hold_reason(&t, &d2).as_deref(),
3869            Some("other\n\n(previously: same)")
3870        );
3871    }
3872
3873    #[tokio::test]
3874    async fn a_repeated_hold_decision_settles_instead_of_looping() {
3875        let dir = tempdir().unwrap();
3876        let queue = Queue::at(dir.path().join("queue"));
3877        let questions = Questions::at(dir.path().join("questions"));
3878        let reply = r#"{"decisions":[{"id":"ID","reason":"keep held","recovery":"hold"}]}"#;
3879        let mut t = task("held");
3880        t.hold_machine(Some("first".to_owned()));
3881        queue.put(&mut t).unwrap();
3882        let cfg = config(mock_agent(
3883            dir.path(),
3884            REPLY,
3885            env(&reply.replace("ID", &t.id)),
3886        ));
3887        let mut conductor = Conductor::new();
3888        // Cycle 1 writes the new note; cycle 2 repeats it and writes nothing.
3889        for _ in 0..2 {
3890            let held = queue.get(&t.id).unwrap();
3891            conductor
3892                .maybe_run(
3893                    &cfg,
3894                    dir.path(),
3895                    &queue,
3896                    &questions,
3897                    dir.path(),
3898                    &[],
3899                    &[],
3900                    &[held],
3901                    2,
3902                )
3903                .await;
3904        }
3905        let held = queue.get(&t.id).unwrap();
3906        assert!(
3907            held.hold_reason
3908                .as_deref()
3909                .unwrap()
3910                .starts_with("keep held")
3911        );
3912        assert!(
3913            !conductor.worth_a_look(&queue, &[], &[held]),
3914            "an unchanged held task must stop being reconsidered"
3915        );
3916    }
3917
3918    #[tokio::test]
3919    async fn alternating_wording_does_not_keep_a_held_task_in_the_loop() {
3920        let dir = tempdir().unwrap();
3921        let queue = Queue::at(dir.path().join("queue"));
3922        let questions = Questions::at(dir.path().join("questions"));
3923        let mut t = task("held");
3924        t.hold_machine(Some("first".to_owned()));
3925        queue.put(&mut t).unwrap();
3926        let reply = |note: &str| {
3927            format!(
3928                r#"{{"decisions":[{{"id":"{}","reason":"{note}","recovery":"hold"}}]}}"#,
3929                t.id
3930            )
3931        };
3932        let cfg_a = config(mock_agent(dir.path(), REPLY, env(&reply("keep held"))));
3933        let cfg_b = config(mock_agent(dir.path(), REPLY, env(&reply("leave held"))));
3934        let mut conductor = Conductor::new();
3935        let mut writes = 0;
3936        for i in 0..6 {
3937            let held = queue.get(&t.id).unwrap();
3938            let before = queue.revision();
3939            conductor
3940                .maybe_run(
3941                    if i % 2 == 0 { &cfg_a } else { &cfg_b },
3942                    dir.path(),
3943                    &queue,
3944                    &questions,
3945                    dir.path(),
3946                    &[],
3947                    &[],
3948                    &[held],
3949                    2,
3950                )
3951                .await;
3952            if queue.revision() != before {
3953                writes += 1;
3954            }
3955        }
3956        assert!(writes <= 2, "the loop must settle, saw {writes} writes");
3957        let mut held = queue.get(&t.id).unwrap();
3958        held.instruction.push_str(" (edited)");
3959        queue.put(&mut held).unwrap();
3960        let held = queue.get(&t.id).unwrap();
3961        assert!(conductor.worth_a_look(&queue, &[], &[held]));
3962    }
3963
3964    #[test]
3965    fn worth_a_look_is_config_free_and_matches_maybe_runs_own_gate() {
3966        let dir = tempdir().unwrap();
3967        let queue = Queue::at(dir.path().join("queue"));
3968        let mut t = task("t");
3969        queue.put(&mut t).unwrap();
3970
3971        let mut conductor = Conductor::new();
3972        assert!(
3973            conductor.worth_a_look(&queue, &[], &[]),
3974            "a conductor that has never run has something to look at"
3975        );
3976
3977        conductor.last_seen = Some(conductor.snapshot(&queue, &[], &[]));
3978        assert!(
3979            !conductor.worth_a_look(&queue, &[], &[]),
3980            "nothing changed and nothing is stalled or finished"
3981        );
3982        assert!(
3983            conductor.worth_a_look(&queue, &[t.clone()], &[]),
3984            "a stalled task is worth a look even at the same revision"
3985        );
3986        assert!(
3987            conductor.worth_a_look(&queue, &[], &[t.clone()]),
3988            "a finished task is worth a look even at the same revision"
3989        );
3990    }
3991
3992    #[tokio::test]
3993    async fn the_conduct_path_never_calls_ask_and_wait() {
3994        // Structural: grepping this module and `daemon.rs` for
3995        // `ask_and_wait` is the actual assertion this module's own doc
3996        // promises; this test exists so the promise has a name in the test
3997        // output too. `apply_one`'s question path uses `Questions::put`
3998        // exclusively.
3999        let dir = tempdir().unwrap();
4000        let queue = Queue::at(dir.path().join("queue"));
4001        let questions = Questions::at(dir.path().join("questions"));
4002        let mut t = task("asks without blocking");
4003        queue.put(&mut t).unwrap();
4004
4005        apply(
4006            &queue,
4007            &questions,
4008            &Verdict {
4009                decisions: vec![Decision {
4010                    id: t.id.clone(),
4011                    question: Some("ok?".to_owned()),
4012                    ..Decision::default()
4013                }],
4014            },
4015        )
4016        .unwrap();
4017        // Reaching here at all (no hang) is the assertion.
4018        assert_eq!(queue.get(&t.id).unwrap().status, TaskStatus::Blocked);
4019    }
4020
4021    #[test]
4022    fn a_pinned_resume_is_not_held_or_requeued_by_the_conductor() {
4023        let mut t = Task::new("t".into(), "t".into(), PathBuf::new(), Source::Human);
4024        t.resume_override = Some(crate::queue::OperatorResume {
4025            question_id: "q".into(),
4026            at: jiff::Timestamp::now(),
4027            conductor_rehold: None,
4028            forced: false,
4029            pinned_run: Some("run-1".into()),
4030        });
4031        assert!(pinned_resume(&t));
4032        assert!(!may_hold(&mut t, "waiting for magi resume"));
4033        assert!(
4034            t.resume_override
4035                .as_ref()
4036                .unwrap()
4037                .conductor_rehold
4038                .is_none(),
4039            "a refused hold is not recorded as an override"
4040        );
4041        t.resume_override = None;
4042        assert!(!pinned_resume(&t));
4043        assert!(may_hold(&mut t, "no override, so a hold is allowed"));
4044    }
4045
4046    #[test]
4047    fn a_pinned_resume_is_not_blocked_by_a_conductor_question_or_dependency() {
4048        let dir = tempdir().unwrap();
4049        let queue = Queue::at(dir.path().join("queue"));
4050        let questions = Questions::at(dir.path().join("questions"));
4051        let mut t = task("resume me");
4052        t.resume_override = Some(crate::queue::OperatorResume {
4053            question_id: "q".into(),
4054            at: jiff::Timestamp::now(),
4055            conductor_rehold: None,
4056            forced: true,
4057            pinned_run: Some("run-1".into()),
4058        });
4059        queue.put(&mut t).unwrap();
4060
4061        apply(
4062            &queue,
4063            &questions,
4064            &Verdict {
4065                decisions: vec![
4066                    Decision {
4067                        id: t.id.clone(),
4068                        question: Some("really?".to_owned()),
4069                        ..Decision::default()
4070                    },
4071                    Decision {
4072                        id: t.id.clone(),
4073                        blocked_by: vec!["other".to_owned()],
4074                        ..Decision::default()
4075                    },
4076                ],
4077            },
4078        )
4079        .unwrap();
4080
4081        assert_eq!(queue.get(&t.id).unwrap().status, TaskStatus::Queued);
4082        assert!(questions.list().is_empty());
4083    }
4084}