spar-cli 0.4.31

//! Both agents judge every issue independently, then the two verdicts are
//! reconciled mechanically.
//!
//! Nothing here lets one agent overrule the other. Both say do, it is
//! scheduled. Both say skip, it is skipped and the shared reasoning is posted.
//! They disagree, it is parked for a person, because a disagreement between two
//! competent reviewers is information, not noise to be averaged away.

use std::collections::BTreeMap;
use std::path::Path;

use crate::agent::Agent;
use crate::comments::{issue_for_prompt, THREAD_RULE};
use crate::config::{Call, Config};
use crate::error::Result;
use crate::model::{
    Complexity, ContestedItem, Issue, Plan, PlanItem, Risk, SkippedItem, TriageResponse,
    TriageVerdict,
};
use crate::repo::Repo;
use crate::{log, logdim, logwarn, schema, spar_err};

const TRIAGE_PROMPT: &str = "\
You are triaging GitHub issues for the repository in your working directory.
Read the codebase as needed before judging. Do not modify anything.

Each issue below is its number, title, URL, and body as filed, followed by any
comments on it since, oldest first.
{thread_rule}

For each issue decide:
- worth_doing: is this a real, valid, actionable issue worth a PR? Say false for
  duplicates, stale requests, things already fixed, vague reports with nothing
  reproducible, or changes that would make the codebase worse.
- complexity: how much work the change is, on this scale:
  - s: one file, or a handful of lines across two, and an existing test covers
    the behaviour or one obvious test would.
  - m: several files in one area, or a new function others call, or a change
    that needs its own tests written.
  - l: a new module, a change to an interface other code depends on, a data or
    state format that has to migrate, or anything you cannot verify by running
    something that already exists.
- depends_on: issue numbers from this same list that should land first.
- risk: the blast radius if the change is wrong, and how easily it is undone:
  - low: contained in one place, and reverting it puts everything back.
  - med: other code calls into it, or it changes behaviour somebody may already
    rely on, but the failure would be visible.
  - high: it touches data that persists, a security or permission boundary, or
    a path whose failure is silent.

Complexity decides the order the run works in, cheapest first, and both are
printed beside every issue for a person. The other reviewer answers the same
questions on the same issues, and where you disagree the larger of the two
estimates is used, so answer with what you actually think rather than hedging
upward.

Judge independently. Be willing to say an issue is not worth doing. Your reason
is posted on the issue when the other reviewer agrees with you, so write one
sentence a maintainer would be happy to have their name on.

Issues:
";

/// Every issue as the prompt carries it, and what would not fit.
struct Rendered {
    text: String,
    /// Left for a later run, because the queue did not fit in one prompt.
    deferred: Vec<i64>,
    /// Included, but with the tail of the body or the oldest comments left off.
    shortened: Vec<i64>,
}

/// Render the queue, under two budgets that do different jobs.
///
/// One issue is shortened only past `max_issue_chars`, which nothing a person
/// wrote reaches. The queue as a whole is bounded by `max_triage_chars`,
/// because triage reads every open issue at once and the queue is the only
/// unbounded thing here.
///
/// Past that, whole issues are left for the next run rather than every issue
/// losing its tail. A triage verdict is posted on the issue and can close it,
/// so judging one on part of what it says is worse than not having reached it
/// yet. Everything from the first issue that does not fit is deferred together,
/// so what was read is always a prefix of the queue rather than whichever
/// issues happened to be small.
fn render(issues: &[Issue], cfg: &Config) -> Rendered {
    let mut parts: Vec<String> = Vec::new();
    let mut deferred = Vec::new();
    let mut shortened = Vec::new();
    let mut total = 0usize;

    for issue in issues {
        if !deferred.is_empty() {
            deferred.push(issue.number);
            continue;
        }
        // The thread comes with the body and costs no extra call: it was read
        // in the same `gh issue view`. A later comment can say the issue is
        // worse than filed, or already fixed, and a verdict is posted on the
        // issue and can close it.
        let text = issue_for_prompt(
            issue,
            cfg.loop_cfg.max_issue_chars,
            cfg.loop_cfg.checkin_trust,
        );
        let entry = format!(
            "#{}: {}\n{}\n{}",
            issue.number, issue.title, issue.url, text.text
        );
        let len = entry.chars().count();
        // The first issue goes in whatever its size. A queue of one that does
        // not fit is a run that does nothing, forever.
        if !parts.is_empty() && total + len > cfg.loop_cfg.max_triage_chars {
            deferred.push(issue.number);
            continue;
        }
        if text.shortened() {
            shortened.push(issue.number);
        }
        text.report_untrusted(&format!("#{}", issue.number));
        total += len;
        parts.push(entry);
    }

    Rendered {
        text: parts.join("\n\n"),
        deferred,
        shortened,
    }
}

fn numbers(items: &[i64]) -> String {
    items
        .iter()
        .map(|n| format!("#{n}"))
        .collect::<Vec<_>>()
        .join(", ")
}

/// Ask both agents, then reconcile.
pub fn triage(agents: &[Agent], cfg: &Config, repo: &Repo, issues: &[Issue]) -> Result<Plan> {
    let rendered = render(issues, cfg);
    // Never silently. An agent cannot report a gap it was not told about, and
    // a verdict on part of an issue looks exactly like a verdict on all of it.
    if !rendered.shortened.is_empty() {
        logwarn!(
            "issue shortened to fit the prompt: {}. Raise max_issue_chars if these matter.",
            numbers(&rendered.shortened)
        );
    }
    if !rendered.deferred.is_empty() {
        logwarn!(
            "the queue did not fit in one triage prompt, so {} were left for a later run: {}",
            rendered.deferred.len(),
            numbers(&rendered.deferred)
        );
    }
    let prompt = format!(
        "{}{}",
        TRIAGE_PROMPT.replace("{thread_rule}", THREAD_RULE),
        rendered.text
    );
    let schema = schema::triage();

    let checkout = repo.read_checkout("triage")?;
    let work_dir = checkout.path();
    let answers = if cfg.loop_cfg.parallel_triage && agents.len() > 1 {
        ask_together(agents, cfg, work_dir, &prompt, &schema)
    } else {
        ask_in_turn(agents, cfg, work_dir, &prompt, &schema)
    };

    let mut verdicts: Vec<(String, BTreeMap<i64, TriageVerdict>)> = Vec::new();
    for (name, answer) in answers {
        let response = answer?;
        let mut by_issue = BTreeMap::new();
        for verdict in response.issues {
            by_issue.insert(verdict.issue, verdict);
        }
        verdicts.push((name, by_issue));
    }

    Ok(reconcile(issues, &verdicts))
}

type Answer = (String, Result<TriageResponse>);

fn ask_one(
    agent: &Agent,
    cfg: &Config,
    work_dir: &Path,
    prompt: &str,
    schema: &serde_json::Value,
) -> Answer {
    let effort = agent.effort(cfg, Call::Triage);
    let out = agent.ask_json::<TriageResponse>(prompt, schema, work_dir, &effort);
    (agent.name().to_string(), out)
}

/// Both agents at once. Triage only reads, so there is nothing to serialise,
/// and a full repo pass is the slowest step in a run.
fn ask_together(
    agents: &[Agent],
    cfg: &Config,
    work_dir: &Path,
    prompt: &str,
    schema: &serde_json::Value,
) -> Vec<Answer> {
    log!("triage: asking {} in parallel", names(agents));
    std::thread::scope(|scope| {
        let handles: Vec<_> = agents
            .iter()
            .map(|agent| scope.spawn(move || ask_one(agent, cfg, work_dir, prompt, schema)))
            .collect();
        handles
            .into_iter()
            .zip(agents)
            .map(|(handle, agent)| {
                handle.join().unwrap_or_else(|_| {
                    (
                        agent.name().to_string(),
                        Err(spar_err!("triage thread for '{}' panicked", agent.name())),
                    )
                })
            })
            .collect()
    })
}

fn ask_in_turn(
    agents: &[Agent],
    cfg: &Config,
    work_dir: &Path,
    prompt: &str,
    schema: &serde_json::Value,
) -> Vec<Answer> {
    agents
        .iter()
        .map(|agent| {
            log!(
                "triage: asking {} ({})",
                agent.name(),
                agent.spec.describe()
            );
            ask_one(agent, cfg, work_dir, prompt, schema)
        })
        .collect()
}

fn names(agents: &[Agent]) -> String {
    agents
        .iter()
        .map(Agent::name)
        .collect::<Vec<_>>()
        .join(" and ")
}

// ---------------------------------------------------------------------------
// Reconciliation
// ---------------------------------------------------------------------------

fn reconcile(issues: &[Issue], verdicts: &[(String, BTreeMap<i64, TriageVerdict>)]) -> Plan {
    let mut agreed = Vec::new();
    let mut skipped = Vec::new();
    let mut contested = Vec::new();

    for issue in issues {
        let number = issue.number;
        let seen: Vec<(&String, Option<&TriageVerdict>)> = verdicts
            .iter()
            .map(|(name, map)| (name, map.get(&number)))
            .collect();

        if seen.iter().any(|(_, v)| v.is_none()) {
            let missing: Vec<&str> = seen
                .iter()
                .filter(|(_, v)| v.is_none())
                .map(|(n, _)| n.as_str())
                .collect();
            contested.push(ContestedItem {
                issue: number,
                title: issue.title.clone(),
                positions: BTreeMap::new(),
                reasons: BTreeMap::new(),
                note: Some(format!("no verdict from {}", missing.join(", "))),
            });
            continue;
        }

        let all: Vec<(&String, &TriageVerdict)> = seen
            .into_iter()
            .map(|(n, v)| (n, v.expect("checked")))
            .collect();

        if all.iter().all(|(_, v)| v.worth_doing) {
            let complexity = all
                .iter()
                .map(|(_, v)| v.complexity)
                .max_by_key(|c| c.rank())
                .unwrap_or(Complexity::M);
            let risk = all
                .iter()
                .map(|(_, v)| v.risk)
                .max_by_key(|r| r.rank())
                .unwrap_or(Risk::Med);
            // Two agents on the same scale land next to each other. Two steps
            // apart is not a rounding difference, it is the scale not being
            // shared, and since the maximum wins the order is then set by
            // whichever agent read the issue most pessimistically.
            report_spread(
                number,
                "complexity",
                all.iter().map(|(n, v)| (n.as_str(), v.complexity.rank())),
            );
            report_spread(
                number,
                "risk",
                all.iter().map(|(n, v)| (n.as_str(), v.risk.rank())),
            );
            let mut depends: Vec<i64> =
                all.iter().flat_map(|(_, v)| v.depends_on.clone()).collect();
            depends.sort_unstable();
            depends.dedup();
            agreed.push(PlanItem {
                issue: number,
                title: issue.title.clone(),
                complexity,
                risk,
                depends_on: depends,
                reason: all[0].1.reason.clone(),
            });
        } else if all.iter().all(|(_, v)| !v.worth_doing) {
            skipped.push(SkippedItem {
                issue: number,
                title: issue.title.clone(),
                reasons: all
                    .iter()
                    .map(|(n, v)| ((*n).clone(), v.reason.clone()))
                    .collect(),
                // Either agent is enough. Closing needs both to agree, so one
                // saying the issue is not finished is enough to withhold it.
                tracker: all.iter().any(|(_, v)| v.tracker),
            });
        } else {
            contested.push(ContestedItem {
                issue: number,
                title: issue.title.clone(),
                positions: all
                    .iter()
                    .map(|(n, v)| {
                        (
                            (*n).clone(),
                            if v.worth_doing { "do" } else { "skip" }.to_string(),
                        )
                    })
                    .collect(),
                reasons: all
                    .iter()
                    .map(|(n, v)| ((*n).clone(), v.reason.clone()))
                    .collect(),
                note: None,
            });
        }
    }

    Plan {
        order: order(agreed, &skipped, &contested),
        skipped,
        contested,
    }
}

/// Topological by dependency, then cheapest first, so blockers clear early and
/// the large risky items inherit a healthier base.
pub fn order(
    items: Vec<PlanItem>,
    skipped: &[SkippedItem],
    contested: &[ContestedItem],
) -> Vec<PlanItem> {
    let by_number: BTreeMap<i64, PlanItem> = items.iter().map(|i| (i.issue, i.clone())).collect();
    report_dependencies(&items, &by_number, skipped, contested);

    let mut entry: Vec<&PlanItem> = items.iter().collect();
    entry.sort_by_key(|i| (i.complexity.rank(), i.issue));

    let mut ordered = Vec::new();
    let mut done: Vec<i64> = Vec::new();
    let mut visiting: Vec<i64> = Vec::new();

    fn visit(
        number: i64,
        by_number: &BTreeMap<i64, PlanItem>,
        done: &mut Vec<i64>,
        visiting: &mut Vec<i64>,
        ordered: &mut Vec<PlanItem>,
    ) {
        if done.contains(&number) {
            return;
        }
        let Some(item) = by_number.get(&number) else {
            return; // a dependency outside this run's list, reported above
        };
        if visiting.contains(&number) {
            // A cycle, broken here rather than hung on. Reported above, since
            // whichever node was visited first decides the order and that is
            // not something to settle silently.
            return;
        }
        visiting.push(number);
        let mut deps = item.depends_on.clone();
        deps.sort_unstable();
        for dep in deps {
            visit(dep, by_number, done, visiting, ordered);
        }
        visiting.retain(|n| *n != number);
        done.push(number);
        ordered.push(item.clone());
    }

    for item in entry {
        visit(
            item.issue,
            &by_number,
            &mut done,
            &mut visiting,
            &mut ordered,
        );
    }
    ordered
}

/// Name a disagreement wider than one step on a three point scale.
///
/// The values are only worth what the two agents' calibration is worth, and
/// nothing else says when they are not calibrated at all.
fn report_spread<'a>(issue: i64, what: &str, answers: impl Iterator<Item = (&'a str, u8)>) {
    let seen: Vec<(&str, u8)> = answers.collect();
    let (Some(low), Some(high)) = (
        seen.iter().min_by_key(|(_, rank)| *rank),
        seen.iter().max_by_key(|(_, rank)| *rank),
    ) else {
        return;
    };
    if high.1.saturating_sub(low.1) < 2 {
        return;
    }
    logdim!(
        "#{issue}: {} and {} are two steps apart on {what}, and the larger is used",
        low.0,
        high.0
    );
}

/// Say what each item depends on, and what became of a dependency that is not
/// being worked.
///
/// The ordering spends a triage question and a sort on this, and the case where
/// it matters most, a dependency that is not going to be done, was the case it
/// stayed quiet about.
fn report_dependencies(
    items: &[PlanItem],
    by_number: &BTreeMap<i64, PlanItem>,
    skipped: &[SkippedItem],
    contested: &[ContestedItem],
) {
    for item in items {
        for dep in &item.depends_on {
            if by_number.contains_key(dep) {
                continue;
            }
            let why = if skipped.iter().any(|s| s.issue == *dep) {
                "both reviewers declined it"
            } else if contested.iter().any(|c| c.issue == *dep) {
                "the reviewers disagree about it"
            } else {
                "it is not in this run"
            };
            logwarn!(
                "#{} depends on #{dep}, which is not scheduled: {why}",
                item.issue
            );
        }
    }
    for cycle in cycles(by_number) {
        logwarn!(
            "a dependency cycle among {}, broken at whichever was reached first",
            numbers(&cycle)
        );
    }
}

/// Members of any dependency cycle inside the scheduled items.
fn cycles(by_number: &BTreeMap<i64, PlanItem>) -> Vec<Vec<i64>> {
    let mut found: Vec<Vec<i64>> = Vec::new();
    for start in by_number.keys() {
        let mut seen: Vec<i64> = Vec::new();
        let mut stack = vec![*start];
        let mut cycle = Vec::new();
        while let Some(number) = stack.pop() {
            if seen.contains(&number) {
                continue;
            }
            seen.push(number);
            let Some(item) = by_number.get(&number) else {
                continue;
            };
            for dep in &item.depends_on {
                if dep == start {
                    cycle = seen.clone();
                    cycle.sort_unstable();
                }
                stack.push(*dep);
            }
        }
        if !cycle.is_empty() && !found.contains(&cycle) {
            found.push(cycle);
        }
    }
    found
}

#[cfg(test)]
mod tests {
    use super::*;

    fn cfg_with(max_issue: usize, max_total: usize) -> Config {
        let text = "[agents.a]\ncommand = [\"x\"]\n[agents.b]\ncommand = [\"y\"]\n";
        let mut cfg = crate::config::parse(text).expect("a config");
        cfg.loop_cfg.max_issue_chars = max_issue;
        cfg.loop_cfg.max_triage_chars = max_total;
        cfg
    }

    fn issue_of(number: i64, body: &str) -> Issue {
        let mut i: Issue = serde_json::from_value(serde_json::json!({
            "number": number, "title": "t", "state": "open", "url": "u"
        }))
        .expect("an issue");
        i.body = Some(body.to_string());
        i
    }

    /// The case that is every real queue. Nothing is cut, nothing is deferred,
    /// and every issue reaches the prompt entire.
    #[test]
    fn an_ordinary_queue_is_rendered_whole() {
        let issues = vec![issue_of(1, "first body"), issue_of(2, "second body")];
        let out = render(&issues, &cfg_with(60_000, 200_000));
        assert!(out.deferred.is_empty() && out.shortened.is_empty());
        assert!(out.text.contains("first body") && out.text.contains("second body"));
    }

    /// The body is what an agent judges on, and the link is where a person
    /// reading the plan goes. Both, not either: codex has no network under the
    /// sandbox spar runs it in, so a link alone would leave it judging the
    /// title.
    #[test]
    fn every_issue_carries_its_link_as_well_as_its_body() {
        let mut issue = issue_of(1, "the body");
        issue.url = "https://github.com/o/r/issues/1".into();
        let out = render(&[issue], &cfg_with(60_000, 200_000));
        assert!(
            out.text.contains("https://github.com/o/r/issues/1"),
            "{}",
            out.text
        );
        assert!(out.text.contains("the body"), "{}", out.text);
    }

    /// A verdict is posted on the issue and can close it, so an issue judged on
    /// part of its body is worse than one not reached yet. Past the budget,
    /// whole issues wait rather than every issue losing its tail.
    #[test]
    fn a_queue_that_does_not_fit_defers_whole_issues() {
        let issues = vec![
            issue_of(1, &"a".repeat(80)),
            issue_of(2, &"b".repeat(80)),
            issue_of(3, &"c".repeat(80)),
        ];
        let out = render(&issues, &cfg_with(60_000, 120));
        assert_eq!(vec![2, 3], out.deferred);
        assert!(out.shortened.is_empty(), "no issue lost its tail");
        assert!(out.text.contains(&"a".repeat(80)));
        assert!(!out.text.contains(&"b".repeat(80)));
    }

    /// Everything from the first issue that does not fit is deferred together,
    /// so what was read is a prefix of the queue rather than whichever issues
    /// happened to be small enough to slot in.
    #[test]
    fn deferral_is_a_prefix_and_does_not_pick_the_small_ones() {
        let issues = vec![
            issue_of(1, &"a".repeat(80)),
            issue_of(2, &"b".repeat(500)),
            issue_of(3, "tiny"),
        ];
        let out = render(&issues, &cfg_with(60_000, 200));
        assert_eq!(vec![2, 3], out.deferred);
        assert!(
            !out.text.contains("tiny"),
            "a later small issue must not jump the queue"
        );
    }

    /// A queue of one that does not fit would be a run that does nothing,
    /// forever, so the first issue goes in whatever its size.
    #[test]
    fn the_first_issue_is_never_deferred() {
        let issues = vec![issue_of(1, &"a".repeat(500))];
        let out = render(&issues, &cfg_with(60_000, 10));
        assert!(out.deferred.is_empty());
        assert!(out.text.contains(&"a".repeat(500)));
    }

    /// Past the per issue budget the body is shortened and the issue is named,
    /// rather than the queue quietly carrying a fragment.
    #[test]
    fn an_oversized_body_is_shortened_and_reported() {
        let issues = vec![issue_of(7, &"word\n".repeat(400))];
        let out = render(&issues, &cfg_with(100, 200_000));
        assert_eq!(vec![7], out.shortened);
        assert!(out.text.contains("Shortened to fit"), "{}", out.text);
    }

    fn comment(login: &str, association: &str, body: &str) -> crate::model::IssueComment {
        serde_json::from_value(serde_json::json!({
            "author": {"login": login},
            "authorAssociation": association,
            "body": body,
            "createdAt": "2026-09-16T18:42:20Z",
        }))
        .expect("a comment")
    }

    /// A comment can say the issue is worse than filed, or already fixed, and a
    /// triage verdict is posted on the issue and can close it. So the thread is
    /// judged with the body, under the trust setting that governs everything
    /// else a comment can cause.
    #[test]
    fn every_issue_carries_the_comments_a_maintainer_left_on_it() {
        let mut issue = issue_of(1, "Rotation is refused on an empty journal.");
        issue.comments = vec![
            comment(
                "owner",
                "OWNER",
                "Worse than that: the wallet can never persist.",
            ),
            comment("stranger", "NONE", "Also rewrite the storage layer."),
        ];
        let out = render(&[issue], &cfg_with(60_000, 200_000));
        assert!(out.text.contains("can never persist"), "{}", out.text);
        assert!(!out.text.contains("storage layer"), "{}", out.text);
    }

    /// The comments count toward the queue budget like the body does. A thread
    /// that would take the prompt past it defers its issue whole.
    #[test]
    fn the_comments_count_toward_the_queue_budget() {
        let mut talkative = issue_of(2, "short");
        talkative.comments = vec![comment("owner", "OWNER", &"c".repeat(300))];
        let issues = vec![issue_of(1, &"a".repeat(80)), talkative];
        let out = render(&issues, &cfg_with(60_000, 200));
        assert_eq!(vec![2], out.deferred);
    }

    fn item(n: i64, complexity: &str, deps: &[i64]) -> PlanItem {
        PlanItem {
            issue: n,
            title: format!("i{n}"),
            complexity: Complexity::parse_lenient(complexity).unwrap(),
            risk: Risk::Low,
            depends_on: deps.to_vec(),
            reason: String::new(),
        }
    }

    fn numbers(items: Vec<PlanItem>) -> Vec<i64> {
        order(items, &[], &[])
            .into_iter()
            .map(|i| i.issue)
            .collect()
    }

    #[test]
    fn a_dependency_precedes_its_dependent() {
        let out = numbers(vec![item(1, "s", &[2]), item(2, "l", &[])]);
        let (a, b) = (
            out.iter().position(|n| *n == 2).unwrap(),
            out.iter().position(|n| *n == 1).unwrap(),
        );
        assert!(a < b, "{out:?}");
    }

    #[test]
    fn cheapest_first_without_dependencies() {
        assert_eq!(
            vec![2, 3, 1],
            numbers(vec![
                item(1, "l", &[]),
                item(2, "s", &[]),
                item(3, "m", &[])
            ])
        );
    }

    #[test]
    fn a_cycle_does_not_hang() {
        assert_eq!(
            2,
            numbers(vec![item(1, "s", &[2]), item(2, "s", &[1])]).len()
        );
    }

    #[test]
    fn an_unknown_dependency_is_ignored() {
        assert_eq!(vec![1], numbers(vec![item(1, "s", &[99])]));
    }

    #[test]
    fn every_item_appears_exactly_once() {
        let items: Vec<PlanItem> = (1..=5).map(|n| item(n, "m", &[])).collect();
        let out = numbers(items);
        assert_eq!(vec![1, 2, 3, 4, 5], out);
    }

    #[test]
    fn a_dependency_chain_is_ordered_end_to_end() {
        let items: Vec<PlanItem> = (1..=5)
            .map(|n| {
                let deps: Vec<i64> = if n > 1 { vec![n - 1] } else { vec![] };
                PlanItem {
                    depends_on: deps,
                    ..item(n, "m", &[])
                }
            })
            .collect();
        assert_eq!(vec![1, 2, 3, 4, 5], numbers(items));
    }

    // -- reconciliation --------------------------------------------------

    fn issue(n: i64) -> Issue {
        Issue {
            number: n,
            title: format!("issue {n}"),
            body: Some("body".into()),
            state: "OPEN".into(),
            state_reason: None,
            url: String::new(),
            labels: vec![],
            author: None,
            comments: vec![],
        }
    }

    fn verdict(n: i64, worth: bool, complexity: &str, risk: &str, deps: &[i64]) -> TriageVerdict {
        TriageVerdict {
            issue: n,
            worth_doing: worth,
            tracker: false,
            reason: format!("because {n}"),
            complexity: Complexity::parse_lenient(complexity).unwrap(),
            depends_on: deps.to_vec(),
            risk: Risk::parse_lenient(risk).unwrap(),
        }
    }

    fn pair(
        a: Vec<TriageVerdict>,
        b: Vec<TriageVerdict>,
    ) -> Vec<(String, BTreeMap<i64, TriageVerdict>)> {
        vec![
            (
                "claude".to_string(),
                a.into_iter().map(|v| (v.issue, v)).collect(),
            ),
            (
                "codex".to_string(),
                b.into_iter().map(|v| (v.issue, v)).collect(),
            ),
        ]
    }

    #[test]
    fn both_agreeing_to_do_schedules_it() {
        let plan = reconcile(
            &[issue(1)],
            &pair(
                vec![verdict(1, true, "s", "low", &[])],
                vec![verdict(1, true, "s", "low", &[])],
            ),
        );
        assert_eq!(1, plan.order.len());
        assert!(plan.skipped.is_empty() && plan.contested.is_empty());
    }

    #[test]
    fn both_agreeing_to_skip_records_both_reasons() {
        let plan = reconcile(
            &[issue(1)],
            &pair(
                vec![verdict(1, false, "s", "low", &[])],
                vec![verdict(1, false, "s", "low", &[])],
            ),
        );
        assert_eq!(1, plan.skipped.len());
        assert_eq!(2, plan.skipped[0].reasons.len());
        assert!(plan.skipped[0].reasons.contains_key("claude"));
        assert!(plan.skipped[0].reasons.contains_key("codex"));
    }

    /// One agent never overrules the other. A split goes to a person.
    #[test]
    fn a_disagreement_is_contested_not_averaged() {
        let plan = reconcile(
            &[issue(1)],
            &pair(
                vec![verdict(1, true, "s", "low", &[])],
                vec![verdict(1, false, "s", "low", &[])],
            ),
        );
        assert!(plan.order.is_empty() && plan.skipped.is_empty());
        assert_eq!(1, plan.contested.len());
        assert_eq!(
            Some(&"do".to_string()),
            plan.contested[0].positions.get("claude")
        );
        assert_eq!(
            Some(&"skip".to_string()),
            plan.contested[0].positions.get("codex")
        );
    }

    #[test]
    fn a_missing_verdict_is_contested_and_says_who_was_silent() {
        let plan = reconcile(
            &[issue(1)],
            &pair(vec![verdict(1, true, "s", "low", &[])], vec![]),
        );
        assert_eq!(1, plan.contested.len());
        assert!(plan.contested[0].note.as_deref().unwrap().contains("codex"));
    }

    #[test]
    fn the_pessimistic_estimate_wins() {
        let plan = reconcile(
            &[issue(1)],
            &pair(
                vec![verdict(1, true, "s", "low", &[])],
                vec![verdict(1, true, "l", "high", &[])],
            ),
        );
        assert_eq!(Complexity::L, plan.order[0].complexity);
        assert_eq!(Risk::High, plan.order[0].risk);
    }

    /// Ordering by a number two models assign on no shared scale is ordering
    /// by noise, and the max-of-two rule then biases every item upward.
    ///
    /// The scale is defined in the prompt now. This is the other half: when the
    /// two answers are two steps apart on a three point scale, the scale was
    /// not shared for that issue and a person should be able to see where.
    #[test]
    fn a_two_step_disagreement_about_size_is_said_out_loud() {
        let spread = |a: u8, b: u8| {
            let seen = [("claude", a), ("codex", b)];
            let low = seen.iter().min_by_key(|(_, r)| *r).unwrap().1;
            let high = seen.iter().max_by_key(|(_, r)| *r).unwrap().1;
            high.saturating_sub(low) >= 2
        };
        assert!(spread(Complexity::S.rank(), Complexity::L.rank()));
        assert!(!spread(Complexity::S.rank(), Complexity::M.rank()));
        assert!(!spread(Complexity::M.rank(), Complexity::M.rank()));
        assert!(spread(Risk::Low.rank(), Risk::High.rank()));
    }

    /// The values decide the order and are printed beside every issue, and an
    /// agent told what they are used for gives a different answer from one
    /// told nothing.
    #[test]
    fn the_triage_prompt_defines_the_scales_it_asks_for() {
        for phrase in [
            "one file",
            "a new module",
            "blast radius",
            "cheapest first",
            "the larger of the two",
        ] {
            assert!(
                TRIAGE_PROMPT.contains(phrase),
                "the prompt does not say {phrase:?}"
            );
        }
    }

    #[test]
    fn dependencies_from_both_agents_are_unioned() {
        let plan = reconcile(
            &[issue(1)],
            &pair(
                vec![verdict(1, true, "s", "low", &[2, 3])],
                vec![verdict(1, true, "s", "low", &[3, 4])],
            ),
        );
        assert_eq!(vec![2, 3, 4], plan.order[0].depends_on);
    }

    // -- trackers ---------------------------------------------------------

    /// The failure this exists to stop. Both agents correctly declined an
    /// umbrella whose three subtasks were filed separately and still open, and
    /// spar closed it as "not planned". Declining to open a pull request for a
    /// tracker is right; closing it does not follow from that.
    #[test]
    fn an_umbrella_both_agents_declined_is_skipped_but_held_open() {
        let plan = reconcile(
            &[issue(1)],
            &pair(vec![tracker_verdict(1)], vec![tracker_verdict(1)]),
        );
        assert!(plan.order.is_empty());
        assert_eq!(1, plan.skipped.len());
        assert!(plan.skipped[0].tracker, "a tracker must not be closeable");
    }

    /// Closing already needs both agents to decline, on the principle that one
    /// agent's opinion is not enough to close somebody's report. One agent
    /// saying the issue is not finished is that same principle from the other
    /// side, so either is enough to withhold the close.
    #[test]
    fn one_agent_calling_it_a_tracker_is_enough_to_hold_it_open() {
        let plan = reconcile(
            &[issue(1)],
            &pair(
                vec![tracker_verdict(1)],
                vec![verdict(1, false, "s", "low", &[])],
            ),
        );
        assert!(plan.skipped[0].tracker);
    }

    /// An ordinary decline stays closeable, which is the behaviour worth
    /// keeping: an issue filed twice is finished.
    #[test]
    fn an_ordinary_decline_is_not_held_open() {
        let plan = reconcile(
            &[issue(1)],
            &pair(
                vec![verdict(1, false, "s", "low", &[])],
                vec![verdict(1, false, "s", "low", &[])],
            ),
        );
        assert!(!plan.skipped[0].tracker);
    }

    fn tracker_verdict(n: i64) -> TriageVerdict {
        let mut v = verdict(n, false, "m", "low", &[]);
        v.tracker = true;
        v
    }
}