onevcs/

host.rs

//! The remote-host side of the seam.
//!
//! Host-neutral vocabulary: the review unit is a [`ChangeRequest`]. GitHub maps it
//! to a pull request; a later host maps it to whatever it calls the same thing.

use std::collections::{BTreeMap, BTreeSet};

use serde::{Deserialize, Serialize};
use url::Url;

use crate::boundary::Visibility;
use crate::error::{invalid, Error, Result};
use crate::event::ArtifactId;
use crate::publish::DraftReason;
use crate::review::{CommentId, PostedReply, ReadCost, ReadMarker, ReviewRead};
use crate::rules::MergePolicy;
use crate::standings::{ChangeStandings, ReviewStandings};
use crate::{gh, git, review, standings, stream};

/// What `gh pr checks` exits with when a check it has just reported has not
/// settled. The rollup it printed is still the answer, so this status is read
/// alongside `0` rather than as a failure.
const CHECKS_PENDING: i32 = 8;

/// How `gh pr checks --required` opens its stderr when the repository declares no
/// required check at all — a repository with no branch protection genuinely has
/// none. The rest of that line names the branch, so only the prefix is fixed.
const NO_REQUIRED_CHECKS: &str = "no required checks reported on the ";

/// Whether a `gh pr checks` answer is one to read.
fn usable(answer: &gh::Answer) -> bool {
    matches!(answer.code, Some(0) | Some(CHECKS_PENDING))
}

/// How `gh pr checks` opens its stderr when the head carries no check **at all** —
/// which is every head for the first seconds after it is pushed. The rest of that
/// line names the branch, so only the prefix is fixed.
const NO_CHECKS_YET: &str = "no checks reported on the ";

/// Whether a failing `gh pr checks --required` failed by *answering* "none".
///
/// The whole shape, not a substring anywhere in the output: `gh` exits `1`, writes
/// nothing to stdout, and opens stderr with [`NO_REQUIRED_CHECKS`]. Measured against
/// real GitHub — a repository with no branch protection answers exactly that. It is
/// held to all three because this is the one place a *failed* `gh` call is read as a
/// meaningful answer, and a looser test would swallow an unrelated failure whose
/// message happened to contain the phrase, reporting "nothing blocks the merge"
/// about a host that never said so.
fn no_required_checks(answer: &gh::Answer) -> bool {
    answer.code == Some(1)
        && answer.stdout.trim().is_empty()
        && answer.stderr.trim_start().starts_with(NO_REQUIRED_CHECKS)
}

/// Whether a failing `gh pr checks` failed by reporting **no check at all** on the
/// head, which is not an answer about what blocks the merge.
///
/// Held to the same three facts as [`no_required_checks`], and one word of `gh`'s own
/// wording apart from it. Reading this one as that one would wave a merge through on
/// a head no verification has begun on.
fn no_checks_yet(answer: &gh::Answer) -> bool {
    answer.code == Some(1)
        && answer.stdout.trim().is_empty()
        && answer.stderr.trim_start().starts_with(NO_CHECKS_YET)
}

/// How `gh api` reports a branch that has no classic protection at all — GitHub's
/// own message, which is an answer (the source names nothing) and not a refusal.
const NOT_PROTECTED: &str = "Branch not protected";

/// Which of the host's check sources a call may consult.
#[derive(Debug, Clone, Copy, PartialEq, Eq)]
enum Consult {
    /// The rollup, and Actions when the credential is not allowed to read it.
    Either,
    /// The complete status-check rollup only.
    StatusChecks,
    /// Actions jobs together with the branch rules that say which ones block.
    Actions,
}

/// What [`gh::CHECK_SOURCE_ENV`] narrows a call to, or [`Consult::Either`].
///
/// The knob exists because the fallback is not free: a credential that will never
/// be allowed to read check runs is refused *every* time the rollup is asked,
/// which for a publication polling a change request is one failed call every few
/// seconds. An operator who knows their token is a fine-grained one says so once.
/// The value is checked here rather than acted on loosely — a misspelling that
/// silently meant "auto" would look like the fallback working.
fn consult() -> Result<Consult> {
    let Some(raw) = std::env::var_os(gh::CHECK_SOURCE_ENV) else {
        return Ok(Consult::Either);
    };
    match raw.to_string_lossy().trim().to_ascii_lowercase().as_str() {
        "" | "auto" => Ok(Consult::Either),
        "status-checks" => Ok(Consult::StatusChecks),
        "actions" => Ok(Consult::Actions),
        other => Err(invalid(format!(
            "{} names {other:?}, which is not a check source this build can read: it must be \
             \"auto\", \"status-checks\", or \"actions\"",
            gh::CHECK_SOURCE_ENV
        ))),
    }
}

/// Read the operator's check-source knob and answer whether it names a source.
///
/// Exposed so a publication can refuse a misspelling of it **before** it pushes: the
/// knob is input, and input is rejected at its boundary.
pub(crate) fn check_source_names_a_source() -> Result<()> {
    consult().map(|_| ())
}

/// How many entries an Actions listing is asked for at once, which is the most
/// GitHub will return in one page.
const PAGE: u32 = 100;

/// How GitHub declines something the credential is not allowed to reach, in both
/// the shapes it declines it in: a GraphQL error naming the node it would not
/// produce, and a REST 403. The wording is GitHub's, and it is the whole of the
/// test because it is the whole of what distinguishes the two cases below.
const CHECK_RUN_REFUSAL: &str = "GraphQL: Resource not accessible by personal access token";

/// Whether a refusal is one this credential will *always* get.
///
/// The one distinction the fallback turns on. A credential that may not read check
/// runs will never read them, so asking a narrower source instead is the only way
/// to answer at all — but a rollup that came back garbled, or a host that would not
/// say which of its checks block, is a *complete* answer having gone wrong, and
/// quietly answering from GitHub Actions alone would drop whatever a third-party
/// integration posted. A required check nobody looked at is how a merge that was
/// never gated ends up looking like one that was.
fn check_rollup_refused_for_pat(error: &Error) -> bool {
    let Error::Invalid { reason } = error else {
        return false;
    };
    reason.contains(CHECK_RUN_REFUSAL) && reason.contains("statusCheckRollup")
}

/// Everything `onevcs` asks of a repository's remote host.
pub trait RemoteHost {
    /// Who the host believes is calling.
    fn authenticated_user(&self) -> Result<String>;

    /// Open a change request.
    fn open_change(&self, req: ChangeSpec) -> Result<ChangeRequest>;

    /// Every open change request from `head` into `base`.
    fn find_changes(&self, head: &str, base: &str) -> Result<Vec<ChangeRequest>>;

    /// The checks the host is reporting on a change request, and which of its
    /// sources that answer was read from.
    fn change_checks(&self, cr: &ChangeRequest) -> Result<ChangeChecks>;

    /// Store one check's log as an artifact and return its id.
    fn check_log(&self, cr: &ChangeRequest, check: &Check) -> Result<ArtifactId>;

    /// Merge a change request under a policy.
    fn merge(&self, cr: &ChangeRequest, policy: MergePolicy) -> Result<MergeOutcome>;

    /// The commit a change request reached its base at, or `None` while the host
    /// has not merged it.
    ///
    /// What a publication under `change-auto` watches. The host performs that merge
    /// on its own clock — `merge` arms it and answers
    /// [`MergeOutcome::Queued`] — so the only way to learn that it happened, and at
    /// which commit, is to ask again. `None` is "not yet" and never "somewhere
    /// else": an implementation that cannot say must refuse rather than answer it.
    ///
    /// Defaulted so that an implementation written against the earlier surface
    /// still compiles, and defaulted to the refusal this crate reserves for a seam
    /// with no body rather than to `None` — a host that was never taught to answer
    /// this has not said a change is unmerged, and a publication that read it that
    /// way would watch until its bound and then report checks that were never the
    /// reason.
    fn merged_at(&self, _cr: &ChangeRequest) -> Result<Option<Sha>> {
        Err(Error::NotImplemented {
            operation: "RemoteHost::merged_at",
        })
    }

    /// The commit a change request reached its base at **and the head the host merged
    /// it from**, or `None` while the host has not merged it.
    ///
    /// What a retirement asks before it may call a branch merged: a merge proves the
    /// *branch's* work landed only where the head that merged carries that work, and a
    /// change request whose number happens to match, or whose branch moved on after it
    /// merged, says nothing about what the branch holds now. Both halves come from one
    /// read of the host, so the head is the one the merge was of.
    ///
    /// Defaulted for the reason [`merged_at`](RemoteHost::merged_at) is — the seam
    /// stays additive — and to the same refusal: a host never taught to say which head
    /// merged has not said that this branch's did, and a retirement reads the refusal
    /// as a question nobody answered, which keeps the branch.
    fn merged_head(&self, _cr: &ChangeRequest) -> Result<Option<MergedChange>> {
        Err(Error::NotImplemented {
            operation: "RemoteHost::merged_head",
        })
    }

    /// When the host says a merged change request reached its base, as the host
    /// spelled it, or `None` while it has not merged it.
    ///
    /// What the record of a landing carries as `landed_at`, and the one way to learn
    /// it for a merge the host performed on its own clock — above all one this crate
    /// reconciles on a later read, whose landing time is the host's and never the
    /// moment somebody asked.
    ///
    /// Defaulted for the reason [`merged_at`](RemoteHost::merged_at) is — the seam
    /// stays additive — and to the same refusal. A caller reads that refusal as a host
    /// that cannot say, and records the landing's time as unknown.
    fn merge_time(&self, _cr: &ChangeRequest) -> Result<Option<String>> {
        Err(Error::NotImplemented {
            operation: "RemoteHost::merge_time",
        })
    }

    /// Whether a change request can still merge into its base, as the host has
    /// computed it.
    ///
    /// What a publication under `change-auto` reads beside
    /// [`merged_at`](RemoteHost::merged_at) while it waits for the merge it armed: a
    /// base that moved under the change so that it conflicts is a merge the host can
    /// never perform, and [`Mergeability::Conflicting`] ends that wait at once as a
    /// sync conflict rather than at the bound. [`Mergeability::Unknown`] is "not
    /// computed yet" and is never read as a conflict.
    ///
    /// Defaulted for the reason [`merged_at`](RemoteHost::merged_at) is — the seam
    /// stays additive — and to the same refusal. A publication reads that refusal as a
    /// host that said nothing, which is no conflict, so a host written before this
    /// method watches exactly as it did.
    fn mergeability(&self, _cr: &ChangeRequest) -> Result<Mergeability> {
        Err(Error::NotImplemented {
            operation: "RemoteHost::mergeability",
        })
    }

    /// Take a change request out of its draft state, so the host will let it land.
    ///
    /// What lifts a draft: a publication whose required checks came back green on it
    /// (or whose grace window elapsed with none run), `onevcs change ready`, and a
    /// publication adopting a draft with the lifecycle off. It is always asked
    /// **before** the host is asked to merge or arm a merge, because a host does
    /// neither to a draft.
    ///
    /// Defaulted for the reason [`merged_at`](RemoteHost::merged_at) is — the seam
    /// stays additive — and to the same refusal: a host that was never taught to
    /// lift a draft has not lifted one, and answering `Ok(())` would report a change
    /// as ready for review while the host goes on holding it.
    fn ready_for_review(&self, _cr: &ChangeRequest) -> Result<()> {
        Err(Error::NotImplemented {
            operation: "RemoteHost::ready_for_review",
        })
    }

    /// Whether the host is holding this change request as a draft.
    ///
    /// Asked on both sides of a draft's life: after one is opened, because a host
    /// that ignored the request would otherwise leave a landable change reported as
    /// held back; and before one is lifted, so a change nobody drafted is asked for
    /// nothing and a second lift changes nothing.
    ///
    /// Defaulted to the refusal, never to `false`: "this host was never taught to
    /// answer" and "this change is not a draft" are different facts, and only the
    /// second is a reason to go on and merge — which is what the `Result` around the
    /// answer carries, rather than a third value inside it.
    ///
    /// The answer is a `bool` because the approved amendment in `docs/contract.md`
    /// declares this seam verbatim as `-> Result<bool>`: a named domain type here would
    /// be a public item the contract does not name, which this repository reports rather
    /// than takes. Nothing is lost to the shape — the *third* state a caller must not
    /// confuse with "not a draft" is "this host would not say", and that is the `Err`
    /// arm rather than a value inside the `Ok`. Both call sites read it as one of three,
    /// and they are `hold_as_draft` and `lift_any_draft` in `publish.rs`.
    // llmlint: ignore[invalid_states_unrepresentable] the doc above: the contract fixes this seam.
    fn is_draft(&self, _cr: &ChangeRequest) -> Result<bool> {
        Err(Error::NotImplemented {
            operation: "RemoteHost::is_draft",
        })
    }

    /// The checks the repository requires before anything merges into `base`, asked
    /// of the branch itself rather than of a change request targeting it.
    ///
    /// What `onevcs repos --audit-gates` names per identity. Which checks a
    /// repository requires is a setting on that repository, and a consumer that
    /// sequences its own work behind another repository's merge path had to keep a
    /// copy of that list — which drifted the day the repository renamed a check, and
    /// cost a whole gate to discover. The host holds the list; this is how it is
    /// read.
    ///
    /// The answer says which protection source it could not consult, because "this
    /// source found nothing" and "this merge path requires nothing" are opposite
    /// facts and only the second is safe to act on: an answer with a source in
    /// [`RequiredChecks::unconsulted`] is incomplete, and an empty one is *unknown*
    /// rather than none. A host that could read no source at all must refuse rather
    /// than answer, for the reason [`change_checks`](RemoteHost::change_checks) may
    /// not answer with an empty source list.
    ///
    /// Defaulted for the reason [`merged_at`](RemoteHost::merged_at) is — the seam
    /// stays additive — and to the same refusal, which the audit reports as the list
    /// being unreadable rather than as no check being required.
    fn required_checks_on(&self, _base: &str) -> Result<RequiredChecks> {
        Err(Error::NotImplemented {
            operation: "RemoteHost::required_checks_on",
        })
    }

    /// Whether the host requires a branch to be up to date with `base` before it
    /// merges a change request into it — GitHub branch protection's `strict`.
    ///
    /// What a branch whose identity merges its base `on-conflict` asks before leaving
    /// a moved base out of it: a host that requires an up-to-date branch would refuse
    /// to merge one that is not, so the base is merged in after all. Read from the
    /// same protection sources [`required_checks_on`](RemoteHost::required_checks_on)
    /// reads; `true` where any of them requires it, `false` only where every source
    /// was read and none does.
    ///
    /// Defaulted to the refusal for the reason every other added method is. A caller
    /// reads any refusal as "could not be read" and merges the base, which is what
    /// every build before this method did.
    fn requires_up_to_date(&self, _base: &str) -> Result<bool> {
        Err(Error::NotImplemented {
            operation: "RemoteHost::requires_up_to_date",
        })
    }

    /// Whether the repository lets a change request be armed to merge itself once its
    /// required checks pass — GitHub's `allow_auto_merge`.
    ///
    /// What a `change-auto` or `change-direct` publication asks, **before it changes
    /// anything**, when its wait for the checks elapses with none of them red: a host
    /// that allows it has the draft lifted and its merge armed, so the change lands on
    /// the host's clock once the checks pass; one that does not is left as it was.
    ///
    /// Defaulted to the refusal for the reason every other added method is. A caller
    /// reads any refusal as "not allowed" and changes nothing.
    fn allows_auto_merge(&self) -> Result<bool> {
        Err(Error::NotImplemented {
            operation: "RemoteHost::allows_auto_merge",
        })
    }

    /// Replace a change request's description: its body, and its title when one is
    /// given.
    ///
    /// The one write to a change request's prose after it exists. It writes exactly
    /// the caller's bytes — nothing is composed around the body and nothing of a
    /// draft's reason is rendered into it — because a body is what a reviewer reads
    /// and what a drafting caller may rewrite, and `onevcs change describe` is where
    /// a session's own evidence reaches it.
    ///
    /// Defaulted for the reason [`merged_at`](RemoteHost::merged_at) is — the seam
    /// stays additive — and to the same refusal: a host that was never taught to
    /// describe a change has not described one.
    fn describe_change(
        &self,
        _cr: &ChangeRequest,
        _title: Option<&str>,
        _body: &str,
    ) -> Result<()> {
        Err(Error::NotImplemented {
            operation: "RemoteHost::describe_change",
        })
    }

    /// Whether the repository is public, as the host says.
    ///
    /// Asked at every write that could reach a public destination, because which
    /// terms a write is checked against depends on it, and recorded in the registry.
    /// An organisation-internal repository is [`Visibility::Private`]: it is not
    /// readable by everybody.
    ///
    /// Defaulted for the reason [`merged_at`](RemoteHost::merged_at) is — the seam
    /// stays additive — and to the same refusal, which is recorded as
    /// [`Visibility::Unknown`] and treated as private.
    fn visibility(&self) -> Result<Visibility> {
        Err(Error::NotImplemented {
            operation: "RemoteHost::visibility",
        })
    }

    /// What a change request's description is now: its title and its body, as the
    /// host holds them.
    ///
    /// The read beside [`describe_change`](RemoteHost::describe_change), and what
    /// `onevcs change show` answers with: a caller that wrote a description needs to
    /// see what it wrote, and a caller finishing one off needs to see what it starts
    /// from.
    ///
    /// Defaulted to the refusal for the reason every other added method is.
    fn change_description(&self, _cr: &ChangeRequest) -> Result<Description> {
        Err(Error::NotImplemented {
            operation: "RemoteHost::change_description",
        })
    }

    /// A change request's review feedback: every review-thread comment, review
    /// summary and conversation comment created or edited after `since` (all of them
    /// when `since` is `None`), and every comment whose thread was resolved,
    /// unresolved or outdated since it, with the marker to pass next time.
    ///
    /// What [`review_comments`](crate::review_comments) asks. `ours` and
    /// `reply_marker` are set from the reply marker in a comment's body and from
    /// nothing else, and the marker is the implementation's own to spell.
    ///
    /// Defaulted to the refusal for the reason every other added method is.
    fn review_comments(
        &self,
        _change: &ChangeId,
        _since: Option<&ReadMarker>,
    ) -> Result<ReviewRead> {
        Err(Error::NotImplemented {
            operation: "RemoteHost::review_comments",
        })
    }

    /// Post `body` as a reply to `comment`: in its thread where it is in one, and
    /// otherwise as a new conversation comment whose first line links it. Never
    /// resolves a thread.
    ///
    /// `body` is the whole of what is posted, the reply marker already its last line,
    /// and `key` is the idempotency key that marker carries, handed over for an
    /// implementation that keeps a record of its own. Finding a reply already
    /// carrying `key` is [`reply_to_comment`](crate::reply_to_comment)'s, under its
    /// lock, so this posts.
    ///
    /// Defaulted to the refusal for the reason every other added method is.
    fn reply_to_comment(
        &self,
        _change: &ChangeId,
        _comment: &CommentId,
        _body: &str,
        _key: &str,
    ) -> Result<PostedReply> {
        Err(Error::NotImplemented {
            operation: "RemoteHost::reply_to_comment",
        })
    }

    /// Where each of `changes` stands — its state and the branch it merged into, its
    /// base, head, description, draft flag and its required checks' verdict — one per
    /// change request asked about, in the order asked, with what the read cost.
    ///
    /// `landing` on a merge is true exactly when it merged into `root`; `base_deleted`
    /// on a closed one is whether the origin still has its base. The verdict is the one
    /// a publication's watch would reach from the same rollup, against what the host
    /// declares a merge into each base requires. What [`change_standings`] asks, and
    /// what [`retarget_change`] and [`reopen_change`] read before and after they write.
    ///
    /// [`change_standings`]: crate::change_standings
    /// [`retarget_change`]: crate::retarget_change
    /// [`reopen_change`]: crate::reopen_change
    ///
    /// Defaulted to the refusal for the reason every other added method is.
    fn change_standings(&self, _root: &str, _changes: &[ChangeId]) -> Result<ChangeStandings> {
        Err(Error::NotImplemented {
            operation: "RemoteHost::change_standings",
        })
    }

    /// [`change_standings`](RemoteHost::change_standings) carrying each change
    /// request's review feedback since the marker beside it, in one read: one
    /// [`ReviewRead`] per change request, in the order asked, each the one
    /// [`review_comments`](RemoteHost::review_comments) would answer from the same
    /// marker.
    ///
    /// Defaulted to the refusal for the reason every other added method is.
    fn review_standings(
        &self,
        _root: &str,
        _changes: &[(ChangeId, Option<ReadMarker>)],
    ) -> Result<ReviewStandings> {
        Err(Error::NotImplemented {
            operation: "RemoteHost::review_standings",
        })
    }

    /// Move a change request onto `base`, writing `body` as its description in the
    /// same request where one is given, and answer what the write cost. Never changes
    /// whether it is a draft.
    ///
    /// Defaulted to the refusal for the reason every other added method is.
    fn retarget_change(
        &self,
        _change: &ChangeId,
        _base: &str,
        _body: Option<&str>,
    ) -> Result<ReadCost> {
        Err(Error::NotImplemented {
            operation: "RemoteHost::retarget_change",
        })
    }

    /// Reopen a closed change request, as the draft it was where it was one, and answer
    /// what the write cost. Its base branch is already back on the origin.
    ///
    /// Defaulted to the refusal for the reason every other added method is.
    fn reopen_change(&self, _change: &ChangeId) -> Result<ReadCost> {
        Err(Error::NotImplemented {
            operation: "RemoteHost::reopen_change",
        })
    }

    /// Whether work the host runs on the change request's head commit is still in
    /// progress: a workflow run, or a check suite that has at least one check run,
    /// not yet completed. Asked by a draft's watch only when its reported checks
    /// show nothing incomplete at an early-lift decision or at the after-lift
    /// "did not re-run" ending. Defaulted to the refusal, as `mergeability`
    /// is, and a refusal is read as no answer, never as "idle".
    ///
    /// The watch asks it about the change request at the commit it pushed, so
    /// `cr.head_sha` is that commit; an answer whose [`HeadActivity::head`] names
    /// another commit is read as no answer too.
    fn head_activity(&self, _cr: &ChangeRequest) -> Result<HeadActivity> {
        Err(Error::NotImplemented {
            operation: "RemoteHost::head_activity",
        })
    }
}

/// What a host says is still running on one commit.
#[derive(Debug, Clone, PartialEq, Eq, Serialize, Deserialize)]
pub struct HeadActivity {
    /// The commit the host answered about, read off the host's own records.
    pub head: Sha,
    /// The host's name for each piece of work still in progress on it (a workflow
    /// run's name, an app's check suite); empty when nothing is.
    // llmlint: ignore[invalid_states_unrepresentable] the contract fixes this field as
    // the host's own names, carried for a reader and decided on only by emptiness.
    pub in_progress: Vec<String>,
    /// Whether every source of such work was consulted. `false` means only GitHub
    /// Actions' workflow runs were (a credential that cannot read check suites), so an
    /// external app's work is invisible to this answer.
    // llmlint: ignore[invalid_states_unrepresentable] the approved contract fixes this
    // field as `complete: bool` for the three consumers that restate it; it has exactly
    // two states, both meaningful, and no third a named type could exclude.
    pub complete: bool,
}

/// Whether a change request can merge into its base, as
/// [`RemoteHost::mergeability`] answers it.
///
/// `#[non_exhaustive]` so that a later state a host reports — a change that merges
/// cleanly and is blocked for another reason, say — is not a breaking change. A caller
/// matching on it reads anything it does not name as "not a conflict".
#[derive(Debug, Clone, PartialEq, Eq)]
#[non_exhaustive]
pub enum Mergeability {
    /// The host computed it, and the change merges into its base as it stands.
    Mergeable,
    /// The host confirms the change conflicts with its base, so it cannot merge until
    /// the branch is reconciled with it.
    Conflicting {
        /// What the host reported, in its own words — on GitHub `mergeable` and
        /// `mergeStateStatus` — which is what a refusal about it quotes.
        reported: String,
    },
    /// The host has not computed it yet, which is never a conflict.
    Unknown,
}

/// A change request the host reports merged, as
/// [`RemoteHost::merged_head`] answers it: the commit it reached its base at, and the
/// head commit the host merged it from.
#[derive(Debug, Clone, PartialEq, Eq, Serialize, Deserialize)]
pub struct MergedChange {
    /// The commit the base reached — the squash or merge commit the host made.
    pub commit: Sha,
    /// The head the host reports the change request carried when it merged.
    pub head: Sha,
}

/// A change request's description as the host holds it.
///
/// What [`RemoteHost::change_description`] answers, and the other half of
/// [`ChangeDescription`](crate::ChangeDescription): a caller writes a body and
/// optionally a title, and reads back both. Both are plain strings, because both are
/// prose a host places no shape on — a title *this crate lands* is a
/// [`Subject`](crate::Subject), but a title a host already holds was composed by
/// whoever opened the change request, and is reported as it is rather than refused.
// llmlint: ignore[invalid_states_unrepresentable] the doc above: what a host answers,
// reported rather than judged.
#[derive(Debug, Clone, PartialEq, Eq, Serialize, Deserialize)]
pub struct Description {
    /// The title, which under squash-merge becomes the commit subject.
    pub title: String,
    /// The body, verbatim. Empty where the change request carries none.
    pub body: String,
}

/// One of the ways a host protects a branch, each of which may name required checks.
#[derive(Debug, Clone, Copy, PartialEq, Eq, PartialOrd, Ord, Serialize, Deserialize)]
#[serde(rename_all = "kebab-case")]
pub enum ProtectionSource {
    /// The repository's rulesets, which every credential that can read the
    /// repository can read.
    Rulesets,
    /// Classic branch protection, which only a credential with administration
    /// rights can read — so it is the source an audit is most often refused.
    BranchProtection,
}

impl ProtectionSource {
    /// Every source, in the order a report names them.
    pub fn every() -> impl Iterator<Item = ProtectionSource> {
        [
            ProtectionSource::Rulesets,
            ProtectionSource::BranchProtection,
        ]
        .into_iter()
    }

    /// How a report names it.
    pub fn describe(self) -> &'static str {
        match self {
            ProtectionSource::Rulesets => "the repository's rulesets",
            ProtectionSource::BranchProtection => "classic branch protection",
        }
    }
}

/// What a host requires before a merge into a branch, and how much of the host
/// that answer covers.
#[derive(Debug, Clone, PartialEq, Eq, Serialize, Deserialize)]
pub struct RequiredChecks {
    /// The names, from every source that answered.
    pub checks: BTreeSet<String>,
    /// Each protection source the host holds that could not be read, with the
    /// host's reason. Empty means every source answered and `checks` is the whole
    /// list; otherwise `checks` is what the sources that did answer require, and an
    /// empty `checks` is not "nothing is required".
    #[serde(default, skip_serializing_if = "BTreeMap::is_empty")]
    pub unconsulted: BTreeMap<ProtectionSource, String>,
}

impl RequiredChecks {
    /// Whether every protection source answered, so `checks` is the whole list.
    pub fn complete(&self) -> bool {
        self.unconsulted.is_empty()
    }
}

/// Where a [`RemoteHost`] for one repository comes from.
///
/// A host is addressed at a repository rather than at an installation — every
/// call `gh` makes carries a slug — so the seam a caller supplies is the factory
/// rather than one host object. This is what makes the interface reachable: a
/// publication asks the factory it was handed for the repository it is publishing
/// to, and never names an implementation.
pub trait Hosting {
    /// The host that answers for the repository named `owner/name`.
    // llmlint: ignore[invalid_states_unrepresentable] a validated slug newtype would be a
    // public item beyond the one this seam is specified as, and it would have exactly one
    // constructor — the check below. That check is where the value is decided: `GitHub::new`
    // refuses anything that does not name one repository as `owner/name`, before the slug is
    // interpolated into a single `gh --repo`. A caller cannot reach a host any other way.
    fn for_repo(&self, slug: &str) -> Result<Box<dyn RemoteHost>>;
}

/// The factory that produces [`GitHub`] hosts, which is what a real run uses.
///
/// Private: the way to hold one is [`crate::Providers::real`], because a caller
/// mixing a real GitHub into a run whose repository side is not git is asking for
/// a combination neither half was written for.
#[derive(Debug, Clone, Copy, Default, PartialEq, Eq)]
pub(crate) struct GitHubHosting;

impl Hosting for GitHubHosting {
    fn for_repo(&self, slug: &str) -> Result<Box<dyn RemoteHost>> {
        Ok(Box::new(GitHub::new(slug)?))
    }
}

/// What to open a change request for.
#[derive(Debug, Clone, PartialEq, Eq, Serialize, Deserialize)]
pub struct ChangeSpec {
    /// The branch carrying the change.
    // llmlint: ignore[invalid_states_unrepresentable] a validated branch-ref newtype is
    // exactly what the publication path carries internally, and it cannot surface here:
    // the contract fixes the matching `ChangeRequest.base` as `String` and names no ref
    // type, so spelling one on this side of the same trait would add a public item the
    // contract does not name and disagree with the type it does. Both names are handed to
    // git before they reach a host, which is where the parser that decides them lives.
    pub head: String,
    /// The branch it targets, which for a stacked change is the branch below it.
    pub base: String,
    /// The title, which under squash-merge becomes the commit subject.
    pub title: String,
    /// The body. Absent means the host's default from the repository template.
    pub body: Option<String>,
    /// Open it as a **draft**, and why it is not ready. Absent opens an ordinary
    /// change request, which is every publication that came before this field.
    ///
    /// The host is handed the whole reason and asked for one thing with it: open the
    /// change as a draft. Nothing here renders the reason into the change request —
    /// see [`DraftReason`] for the ruling that keeps it out of the body — so a host
    /// implementation that reads past this field's presence is reading further than
    /// the contract asks it to.
    #[serde(default, skip_serializing_if = "Option::is_none")]
    pub draft: Option<DraftReason>,
    /// Open it as a draft **while its required checks run**, with no
    /// [`DraftReason`]: what a lifecycle publication asks for when its caller asked
    /// for no draft. A host lifts or keeps it once the checks settle, by the rules
    /// the publication states; nothing here is a reason a caller gave.
    ///
    /// Beside `draft` rather than a third kind of it, because [`DraftReason`] is what
    /// a *caller* asks for and this is what the lifecycle asks for when the caller
    /// asked for none. Defaulted and omitted when false, so a spec written before it
    /// reads and writes as it did. Where both are set the reason is what the host
    /// records: either way the change opens as a draft.
    // llmlint: ignore[invalid_states_unrepresentable] one sum type over both would
    // change `draft`, which the approved contract declares verbatim as
    // `Option<DraftReason>`, and the manager's ruling for this amendment forbids widening
    // `DraftReason` itself. So the lifecycle's draft is an additive, defaulted field beside
    // it, and the one combination both allow is given a meaning above rather than refused.
    #[serde(default, skip_serializing_if = "std::ops::Not::not")]
    pub draft_awaiting_checks: bool,
}

/// An open change request on the host.
#[derive(Debug, Clone, PartialEq, Eq, Serialize, Deserialize)]
pub struct ChangeRequest {
    /// The host's identifier for it.
    pub id: ChangeId,
    /// Where a human reads it.
    pub url: Url,
    /// The commit its checks are reported against.
    pub head_sha: Sha,
    /// The branch it targets.
    // llmlint: ignore[invalid_states_unrepresentable] the contract declares this field
    // verbatim as `pub base: String` and names no ref type to narrow it to; changing it
    // is changing the interface, which is reported rather than done. Every one this crate
    // constructs is read out of a host response naming a branch the host already has.
    pub base: String,
}

/// A host's identifier for a change request.
// llmlint: ignore[invalid_states_unrepresentable] the contract declares this as
// `id: ChangeId` and fixes nothing about its content — GitHub numbers its pull
// requests, and another host may not. Every one this crate constructs is read out of
// a host response that is required to carry it: `find_changes` rejects an entry with
// no number, and `open_change` rejects output that printed no URL to take one from.
#[derive(Debug, Clone, PartialEq, Eq, PartialOrd, Ord, Hash, Serialize, Deserialize)]
#[serde(transparent)]
pub struct ChangeId(pub String);

/// A commit hash.
// llmlint: ignore[invalid_states_unrepresentable] the contract declares this field as
// `head_sha: Sha` and fixes nothing about its content, and a hash's shape is the host's
// to state — a SHA-1 hex string today, something else on a repository that has moved.
// Every one this crate constructs is validated where it enters: `head_sha` and
// `merged_sha` below both reject a response that names no commit rather than
// constructing an empty one, so no code path here can produce a blank `Sha`.
#[derive(Debug, Clone, PartialEq, Eq, PartialOrd, Ord, Hash, Serialize, Deserialize)]
#[serde(transparent)]
pub struct Sha(pub String);

/// One check the host reports on a change request.
#[derive(Debug, Clone, PartialEq, Eq, Serialize, Deserialize)]
pub struct Check {
    /// The check's name, as branch protection lists it.
    pub name: String,
    /// Where the check is: the host's own status vocabulary, passed through.
    // llmlint: ignore[invalid_states_unrepresentable] the contract fixes this field name
    // and enumerates no value set for it, and the vocabulary differs per host — which is
    // the thing this crate exists to abstract. Inventing an enum here would add a public
    // item the contract does not name. The one word this build reads out of it —
    // `completed` — is read in `settled` below and nowhere else, so a host that spells
    // its states differently is understood by a second implementation of that predicate
    // rather than by a new variant here. Recorded as open question 1 in
    // docs/inferred-surface.md for the planner to settle across the three repositories.
    pub status: String,
    /// How it ended, once it has. Absent while it is still running.
    // llmlint: ignore[invalid_states_unrepresentable] `conclusion` is the other half of the
    // same open question as `status` above, for the same reason: the contract names the
    // field and enumerates no conclusion vocabulary, and each host spells its own. The
    // three this build treats as not blocking a merge are read in `green` below, which is
    // the one place a second host's vocabulary would be taught.
    pub conclusion: Option<String>,
    /// Whether it blocks the merge.
    pub required: bool,
    /// The commit the host attached this check to, or `None` where the host did not
    /// say which one it is about.
    ///
    /// What makes a check's verdict addressable to a commit rather than to a change
    /// request. A change request is a moving target — every push gives it a new head,
    /// and the host attaches the new head's checks seconds to minutes later — so a
    /// caller reading "the checks on this change request" moments after a push reads
    /// the *previous* head's answer and cannot tell. `None` is the host declining to
    /// say and never a commit inferred from the change request: a caller that must
    /// distinguish them has to be able to, and one that cannot is told nothing rather
    /// than told wrongly.
    ///
    /// Defaulted, so a check some earlier build serialized still reads — as one whose
    /// commit that build never recorded, which is what it is.
    #[serde(default, skip_serializing_if = "Option::is_none")]
    pub head: Option<Sha>,
    /// Where this check is on the host, so whoever is handed its refusal can go and
    /// look at it. `None` where the host reported no address for it.
    #[serde(default, skip_serializing_if = "Option::is_none")]
    pub url: Option<Url>,
    /// When the host says this run of the check started, spelled as the host spelled
    /// it, or `None` where it did not say.
    ///
    /// The run's own identity: a check re-run under the same name — the run a
    /// workflow starts when a draft is lifted — is a new run with a new start, even
    /// where it concludes exactly as the one before it did. That is what a
    /// publication tells a run the host attached after an early lift from the
    /// draft's by, so it is compared for equality and never parsed or ordered: two
    /// clocks never meet in it.
    ///
    /// Defaulted and omitted when empty, as `head` and `url` are, so a check an
    /// earlier build serialized still reads and one without a start serializes as
    /// it always did.
    // llmlint: ignore[invalid_states_unrepresentable] the host's own timestamp, carried
    // as an opaque token: this crate only ever asks whether two runs report the same
    // one, so parsing it would add a failure mode and decide nothing. `reported_time`
    // is the one place it enters, and it drops GitHub's zero time for a run that has
    // not started rather than letting every queued run share one identity.
    #[serde(default, skip_serializing_if = "Option::is_none")]
    pub started_at: Option<String>,
    /// When the host says this run of the check completed, spelled as the host
    /// spelled it, or `None` where it did not say — which is every run still going.
    ///
    /// Carried for the record of a gate run and nothing else: a `required-checks`
    /// run lists each required check with the times the host reported, so a reader
    /// can tell a long check from a long wait for one. Never compared or ordered
    /// here. Defaulted and omitted when empty, as `started_at` is, so a check an
    /// earlier build serialized still reads.
    // llmlint: ignore[invalid_states_unrepresentable] the host's own timestamp, carried
    // through to the record verbatim for the reason `started_at` is: this crate decides
    // nothing by it, and parsing it would add a failure mode to a field nothing reads.
    // `reported_time` is where it enters, and drops GitHub's zero time as it does there.
    #[serde(default, skip_serializing_if = "Option::is_none")]
    pub completed_at: Option<String>,
}

impl Check {
    /// Whether this check has reached a terminal state.
    pub fn settled(&self) -> bool {
        self.status.eq_ignore_ascii_case("completed")
    }

    /// Which of the five states this check is in — the one classifier every reading
    /// and every rendering of a check follows.
    ///
    /// A skipped check is its own state and never a passed one: a job a workflow
    /// gated on the change not being a draft reports `skipped` without having run,
    /// so reading it as green would call a change verified that nothing verified.
    /// What the publication watch does with a skipped *required* check is its own
    /// rule, stated where the watch is.
    pub fn state(&self) -> CheckState {
        if !self.settled() {
            return CheckState::Pending;
        }
        match self
            .conclusion
            .as_deref()
            .map(str::to_ascii_lowercase)
            .as_deref()
        {
            Some("success" | "neutral") => CheckState::Passed,
            Some("skipped") => CheckState::Skipped,
            Some("cancelled" | "stale") => CheckState::NoVerdict,
            _ => CheckState::Failed,
        }
    }

    /// Whether a settled check passed. A skipped one did not: see [`Check::state`].
    pub fn green(&self) -> bool {
        self.state() == CheckState::Passed
    }

    /// Whether a settled check ended in a way that blocks a merge.
    pub fn red(&self) -> bool {
        self.state() == CheckState::Failed
    }
}

/// Which of five states one check is in, as [`Check::state`] classifies it.
///
/// Exported so a consumer holds its own copy of "what reads as green" to this
/// declaration rather than to a list it keeps beside it. Serialized kebab-case, which
/// is how the `change-check` event's `state`, the `checks-settled` event and `onevcs
/// status` spell it.
#[derive(Debug, Clone, Copy, PartialEq, Eq, PartialOrd, Ord, Hash, Serialize, Deserialize)]
#[serde(rename_all = "kebab-case")]
pub enum CheckState {
    /// Concluded `success` or `neutral`.
    Passed,
    /// Concluded in a way that blocks a merge: every settled conclusion not named by
    /// another state.
    Failed,
    /// Concluded `skipped`: it did not run, and it is never read as passed.
    Skipped,
    /// Not settled yet.
    Pending,
    /// Concluded `cancelled` or `stale`: the host ended the run with no verdict.
    NoVerdict,
}

/// One place a host's answer about a change request's checks was read from.
///
/// A credential decides which of these can be reached, so which were consulted is
/// part of the answer rather than a property of the build: a fine-grained personal
/// access token cannot read check runs *at all* — GitHub offers no `Checks`
/// permission for one, and the permission it briefly had was withdrawn — so for
/// such a token the rollup is unreadable and Actions is everything it can see.
#[derive(Debug, Clone, Copy, PartialEq, Eq, PartialOrd, Ord, Hash, Serialize, Deserialize)]
#[serde(rename_all = "kebab-case")]
pub enum CheckSource {
    /// The host's own rollup of every check on the change request: workflow jobs,
    /// check runs a third-party integration posted, and commit statuses alike,
    /// each carrying whether it blocks the merge. Complete, and the one source a
    /// fine-grained token cannot read.
    StatusChecks,
    /// GitHub Actions' own view of the head commit: the jobs of the workflow runs
    /// on it and nothing else. A check that ran no workflow — anything a
    /// third-party integration posted as a check run or a commit status — is not
    /// in it and cannot be seen from it.
    Actions,
    /// The repository's rulesets, for which checks block a merge into the base.
    /// Narrower than the rollup's own answer: it reports rulesets and not classic
    /// branch protection, so a repository protected the classic way reports
    /// nothing blocking here — which the publication path fails closed on, waiting
    /// for a required check that never arrives rather than merging without one.
    BranchRules,
}

/// What a host reports about a change request's checks, and where it looked.
///
/// The sources travel with the checks because a caller deciding whether a change
/// may merge has to be able to tell "these are all the checks" from "these are the
/// workflow checks, and anything a third-party integration posted is invisible to
/// this credential". Reporting the second as the first is how a merge gets waved
/// through by a check nobody could see.
#[derive(Debug, Clone, PartialEq, Eq, Serialize, Deserialize)]
pub struct ChangeChecks {
    /// The checks themselves.
    pub checks: Vec<Check>,
    /// Every source consulted to produce them. Never empty: a host that could read
    /// none of its sources is a refusal rather than an answer.
    // llmlint: ignore[invalid_states_unrepresentable] the approved contract fixes
    // this as a public BTreeSet; implementations refuse an empty answer.
    pub sources: BTreeSet<CheckSource>,
}

impl ChangeChecks {
    /// Whether this answer covers every check the host reports.
    ///
    /// False means the checks above are GitHub Actions' and only those.
    pub fn complete(&self) -> bool {
        self.sources.contains(&CheckSource::StatusChecks)
    }
}

/// What merging a change request did.
#[derive(Debug, Clone, PartialEq, Eq, Serialize, Deserialize)]
#[serde(rename_all = "kebab-case")]
pub enum MergeOutcome {
    /// It merged, at this commit.
    Merged(Sha),
    /// The host queued it and will merge it once its checks pass.
    Queued,
    /// It was left open for review, which the policy asked for.
    Open,
}

/// The GitHub implementation of [`RemoteHost`], driven through `gh`.
///
/// Every method is one `gh` invocation, and the program that answers as `gh` is the
/// seam a journey substitutes: GitHub's decisioning is what a test replaces, never
/// the git underneath it.
#[derive(Debug, Clone, PartialEq, Eq)]
pub struct GitHub {
    /// The `owner/name` slug every call is addressed to. Private, so the only way
    /// to hold one is to have had it accepted by [`GitHub::new`] — every `gh`
    /// invocation below trusts it as the repository it addresses.
    repo: String,
}

impl GitHub {
    /// Address this host at one repository, named `owner/name`.
    ///
    /// Checked here rather than at each call: the slug is interpolated into every
    /// `gh --repo` invocation, and a value that is not one repository is a value
    /// that addresses something nobody asked for.
    pub fn new(repo: impl Into<String>) -> Result<Self> {
        let repo = repo.into();
        let mut parts = repo.split('/');
        let named = matches!(
            (parts.next(), parts.next(), parts.next()),
            (Some(owner), Some(name), None)
                if !owner.is_empty()
                    && !name.is_empty()
                    && !repo.starts_with('-')
                    && !repo.contains(char::is_whitespace)
        );
        if !named {
            return Err(invalid(format!(
                "{repo:?} does not name one repository as owner/name"
            )));
        }
        Ok(Self { repo })
    }

    /// The names of the checks the host says block the merge.
    ///
    /// A second call, and it has to be one: `gh pr view`'s `statusCheckRollup`
    /// carries a check's name, status, and conclusion and *nothing* about whether it
    /// blocks anything. This build asked it for `isRequired` anyway and read the
    /// field's absence as a host answering partially, so every `change_checks`
    /// against real GitHub failed — a field `gh pr view` has never returned, checked
    /// only ever against a stand-in that returned it. `gh pr checks --required` is
    /// where `gh` reports it.
    fn required_checks(&self, cr: &ChangeRequest) -> Result<BTreeSet<String>> {
        addressable(&cr.id.0, "change request id")?;
        let answer = gh::attempt(&[
            "pr",
            "checks",
            &cr.id.0,
            "--repo",
            &self.repo,
            "--required",
            "--json",
            "name",
        ])?;
        if !usable(&answer) {
            // A repository that declares no required check at all is answering, not
            // failing: refusing here would make every unprotected repository
            // unreadable, and "nothing blocks the merge" is what the publication
            // path already knows how to act on.
            if no_required_checks(&answer) {
                return Ok(BTreeSet::new());
            }
            if no_checks_yet(&answer) {
                return Err(unregistered(&cr.url, answer.stderr.trim()));
            }
            return Err(unsaid(&cr.url, &answer.detail()));
        }
        let value = gh::json(&answer.stdout)?;
        let entries = value
            .as_array()
            .ok_or_else(|| unsaid(&cr.url, &value.to_string()))?;
        let mut names = BTreeSet::new();
        for entry in entries {
            let name = entry
                .get("name")
                .and_then(|value| value.as_str())
                .filter(|value| !value.is_empty())
                .ok_or_else(|| unsaid(&cr.url, &entry.to_string()))?;
            names.insert(name.to_owned());
        }
        Ok(names)
    }

    /// The log of the job one check ran in.
    ///
    /// Two calls, because `gh` addresses a job's log by the job's *id* and a check's
    /// name is not one. This build passed the name to `gh run view --job`, which has
    /// never accepted anything but a number — so the artifact every settled check
    /// carried held a refusal to produce a log rather than the log, and no offline
    /// journey could tell, because the program standing in for `gh` answered to the
    /// name. The id is the last segment of the details URL `gh pr checks` reports.
    fn job_log(&self, cr: &ChangeRequest, name: &str) -> Result<String> {
        addressable(&cr.id.0, "change request id")?;
        let answer = gh::attempt(&[
            "pr",
            "checks",
            &cr.id.0,
            "--repo",
            &self.repo,
            "--json",
            "name,link",
        ])?;
        if !usable(&answer) {
            return Err(invalid(format!(
                "gh pr checks would not say where check {name:?} on {} ran: {}",
                cr.url,
                answer.detail()
            )));
        }
        let value = gh::json(&answer.stdout)?;
        // Rejected here rather than searched through: a host that answered with
        // something other than a list of checks has not said this check has no job,
        // and reporting it as though it had would put the wrong reason in the
        // artifact an operator reads to find out why there is no log.
        let entries = value.as_array().ok_or_else(|| {
            invalid(format!(
                "gh pr checks returned a non-list of checks on {}, so where check {name:?} ran \
                 cannot be read from it: {value}",
                cr.url
            ))
        })?;
        let link = entries
            .iter()
            .find(|entry| entry.get("name").and_then(|value| value.as_str()) == Some(name))
            .and_then(|entry| entry.get("link"))
            .and_then(|value| value.as_str())
            .ok_or_else(|| {
                invalid(format!(
                    "the host reports no job for check {name:?} on {}",
                    cr.url
                ))
            })?;
        let job = link
            .rsplit_once("/job/")
            .map(|(_, id)| id)
            .filter(|id| !id.is_empty() && id.chars().all(|c| c.is_ascii_digit()))
            .ok_or_else(|| {
                invalid(format!(
                    "the host says check {name:?} ran at {link:?}, which names no job this build \
                     can ask for a log"
                ))
            })?;
        gh::invoke_content(&["run", "view", "--repo", &self.repo, "--log", "--job", job])
    }

    /// One check's log, from whichever source this credential can read.
    ///
    /// The same order [`RemoteHost::change_checks`] consults them in, for the same
    /// reason: `gh pr checks` reports where *every* check ran, including one no
    /// workflow produced, and the Actions API answers for the ones it owns when the
    /// credential may not read the rollup at all.
    fn log_of(&self, cr: &ChangeRequest, name: &str) -> Result<String> {
        match consult()? {
            Consult::StatusChecks => self.job_log(cr, name),
            Consult::Actions => self.actions_log(cr, name),
            // As above, and for the same reason: where a check ran is a question
            // `gh pr checks` can answer about every check, and only a credential
            // that may not ask it at all has cause to ask a narrower source.
            Consult::Either => self.job_log(cr, name).or_else(|reported| {
                if !check_rollup_refused_for_pat(&reported) {
                    return Err(reported);
                }
                self.actions_log(cr, name).map_err(|actions| {
                    invalid(format!(
                        "neither of GitHub's check sources would produce the log of check \
                         {name:?} on {}: gh pr checks answered {reported}, and the Actions API \
                         answered {actions}",
                        cr.url
                    ))
                })
            }),
        }
    }

    /// Every check the host's own rollup reports on a change request.
    ///
    /// Complete — a workflow job, a check run some integration posted, and a commit
    /// status all appear here — and readable only by a credential the repository
    /// allows to read its check runs. A fine-grained personal access token is not
    /// one and cannot become one, which is what [`GitHub::actions_checks`] answers
    /// instead.
    fn rollup_checks(&self, cr: &ChangeRequest) -> Result<Vec<Check>> {
        // `headRefOid` in the same call, because the rollup GitHub renders is the
        // rollup of the change request's *last commit* as that one response saw it —
        // and which commit that was is the whole of what makes the answer
        // addressable. Asked together rather than in a second call for the same
        // reason a second call could not answer it: the head moves, so a commit read
        // afterwards is not necessarily the one those checks were about. It costs
        // this call nothing, because the field needs no more than the read access
        // that found the change request while `statusCheckRollup` needs far more —
        // a credential refused for one of them was already refused for the call.
        let value = self.view(&cr.id.0, "headRefOid,statusCheckRollup")?;
        let head = reported_sha(&value, "headRefOid")?;
        let reported = value.get("statusCheckRollup").ok_or_else(|| {
            invalid(format!(
                "gh pr view reported no checks at all on {}",
                cr.url
            ))
        })?;
        if reported.is_null() {
            return Ok(Vec::new());
        }
        let rollup = reported
            .as_array()
            .ok_or_else(|| invalid(format!("gh pr view returned a non-list rollup: {reported}")))?;
        if rollup.is_empty() {
            return Ok(Vec::new());
        }
        let required = self.required_checks(cr)?;
        rollup
            .iter()
            .map(|entry| check(entry, cr, &required, head.as_ref()))
            .collect()
    }

    /// The checks GitHub Actions reports for the change request's head commit.
    ///
    /// One workflow job is one check, which is the same unit the rollup reports
    /// them in — GitHub posts a check run per job — so what a caller waits for is
    /// unchanged. What it cannot see is anything that ran no workflow, which is why
    /// the answer says which sources it came from rather than passing itself off as
    /// the whole picture.
    fn actions_checks(&self, cr: &ChangeRequest) -> Result<ChangeChecks> {
        let jobs = self.actions_jobs(cr)?;
        if jobs.is_empty() {
            // Nothing has run yet, which is the first seconds of every change
            // request. There is no check to ask the rulesets about, so they are not
            // asked and the answer does not claim they were.
            return Ok(ChangeChecks {
                checks: Vec::new(),
                sources: [CheckSource::Actions].into_iter().collect(),
            });
        }
        let required = self.ruled_checks(cr)?;
        Ok(ChangeChecks {
            checks: jobs
                .into_iter()
                .map(|job| Check {
                    required: required.contains(&job.name),
                    name: job.name,
                    status: job.status,
                    conclusion: job.conclusion,
                    head: job.head,
                    url: job.url,
                    started_at: job.started_at,
                    completed_at: job.completed_at,
                })
                .collect(),
            sources: [CheckSource::Actions, CheckSource::BranchRules]
                .into_iter()
                .collect(),
        })
    }

    /// Every job of every workflow run on the change request's head commit.
    ///
    /// Two calls per run, because the Actions API reports a *run*'s status at the
    /// run and a job's at the job, and a run that is still going says nothing about
    /// which of its jobs have finished — and one before them, for which commit that
    /// head is *now*. Asked rather than taken off the [`ChangeRequest`] the caller
    /// holds, because that one was read when the change request was found: a
    /// publication finds it moments after pushing, when the host may not have
    /// processed the push, and a build that kept asking about the head it first saw
    /// would go on asking about a commit this run replaced for as long as it
    /// watched. The rollup source re-reads the same field on every call for the same
    /// reason; this is that, on the source a fine-grained token is left with.
    fn actions_jobs(&self, cr: &ChangeRequest) -> Result<Vec<Job>> {
        let current = head_sha(&self.view(&cr.id.0, "headRefOid")?)?;
        let sha = commit(&current)?;
        let runs = self.api(&format!(
            "repos/{}/actions/runs?head_sha={sha}&per_page={PAGE}",
            self.repo
        ))?;
        let mut jobs = Vec::new();
        for run in listed(&runs, "workflow_runs", &format!("the commit {sha}"))? {
            let id = run
                .get("id")
                .and_then(serde_json::Value::as_u64)
                .ok_or_else(|| {
                    invalid(format!(
                        "the Actions API reported a workflow run on {} with no id: {run}",
                        cr.url
                    ))
                })?;
            // The commit off the run's own record rather than the one this query
            // filtered on: what the host says it ran against is the host's answer,
            // and a build that stamped the question onto the answer could not tell a
            // host that disagreed.
            let head = reported_sha(run, "head_sha")?;
            let listing = self.api(&format!(
                "repos/{}/actions/runs/{id}/jobs?per_page={PAGE}",
                self.repo
            ))?;
            for entry in listed(&listing, "jobs", &format!("workflow run {id}"))? {
                jobs.push(job(entry, cr, head.as_ref())?);
            }
        }
        Ok(jobs)
    }

    /// The names of the checks the repository's rulesets say block a merge into the
    /// change request's base.
    ///
    /// The second question `gh pr checks --required` answers and a credential that
    /// cannot read check runs cannot ask, because that command reads the same
    /// rollup. This is what such a credential *can* read — and it is narrower, in a
    /// direction that is safe rather than quiet: it reports rulesets and not classic
    /// branch protection, so a repository protected the classic way answers "nothing
    /// blocks" and the publication path then waits for a required check that never
    /// arrives and stops. Under-reporting what blocks holds a merge; over-reporting
    /// what has passed lets one through, and only one of those is recoverable.
    fn ruled_checks(&self, cr: &ChangeRequest) -> Result<BTreeSet<String>> {
        self.required_on(&cr.base, cr.url.as_ref())
    }

    /// The names of the checks the repository's rulesets require before a merge into
    /// `base` — the same read as [`GitHub::ruled_checks`], addressed to a branch
    /// rather than to a change request targeting one, and what
    /// [`RemoteHost::required_checks_on`] answers with. It is the rulesets' answer
    /// and carries the rulesets' limit: a repository protected the classic way names
    /// nothing here.
    ///
    fn required_on(&self, base: &str, about: &str) -> Result<BTreeSet<String>> {
        addressable_branch(base, "the base branch")?;
        let value = self.api(&format!(
            "repos/{}/rules/branches/{}",
            self.repo,
            path_segment(base)
        ))?;
        let rules = value.as_array().ok_or_else(|| {
            invalid(format!(
                "the rules on {about}'s base branch {base:?} came back as something that is \
                 not a list of them, so which of its checks block the merge cannot be read \
                 from it: {value}"
            ))
        })?;
        let mut names = BTreeSet::new();
        for rule in rules {
            if rule.get("type").and_then(|value| value.as_str()) != Some("required_status_checks") {
                continue;
            }
            // A rule that says it requires status checks and then will not say
            // which is the one shape that must not be read as "none": that is the
            // difference between a merge that was gated and one that only looked
            // like it.
            let required = rule
                .get("parameters")
                .and_then(|parameters| parameters.get("required_status_checks"))
                .and_then(|value| value.as_array())
                .ok_or_else(|| unsaid_about(about, &rule.to_string()))?;
            for entry in required {
                let name = entry
                    .get("context")
                    .and_then(|value| value.as_str())
                    .filter(|value| !value.is_empty())
                    .ok_or_else(|| unsaid_about(about, &entry.to_string()))?;
                names.insert(name.to_owned());
            }
        }
        Ok(names)
    }

    /// The contexts classic branch protection requires on `base`, or why the host
    /// would not say.
    ///
    /// Three answers, told apart by what `gh` wrote. A protected branch answers with
    /// its contexts; a branch with no classic protection answers `Branch not
    /// protected`, a 404 that *is* an answer — the source was consulted and names
    /// nothing; and anything else is the source not consulted, which the inner `Err`
    /// carries so the caller records it beside what the rulesets did say rather than
    /// losing that too. The outer `Err` is `gh` itself failing to run.
    fn classic_protection(
        &self,
        base: &str,
    ) -> Result<std::result::Result<BTreeSet<String>, String>> {
        addressable_branch(base, "the base branch")?;
        let path = format!(
            "repos/{}/branches/{}/protection/required_status_checks",
            self.repo,
            path_segment(base)
        );
        let answer = gh::attempt(&["api", &path])?;
        if answer.code != Some(0) {
            if answer.stderr.contains(NOT_PROTECTED) {
                return Ok(Ok(BTreeSet::new()));
            }
            return Ok(Err(answer.detail()));
        }
        let value = gh::json(&answer.stdout)?;
        let contexts = value
            .get("contexts")
            .and_then(|value| value.as_array())
            .ok_or_else(|| {
                invalid(format!(
                    "the classic branch protection on {}'s {base:?} came back without its \
                     contexts, so which checks it requires cannot be read from it: {value}",
                    self.repo
                ))
            })?;
        let mut names = BTreeSet::new();
        for context in contexts {
            let name = context
                .as_str()
                .filter(|value| !value.is_empty())
                .ok_or_else(|| unsaid_about(&self.repo, &context.to_string()))?;
            names.insert(name.to_owned());
        }
        Ok(Ok(names))
    }

    /// Whether the repository's rulesets require a branch to be up to date with `base`
    /// before a merge into it: any `required_status_checks` rule on it whose
    /// `strict_required_status_checks_policy` is set.
    ///
    /// A rule that does not say is refused rather than read as `false`, for the reason
    /// a rule that will not name its checks is: it is the difference between a merge
    /// the host would take and one it would refuse.
    fn ruled_up_to_date(&self, base: &str) -> Result<bool> {
        addressable_branch(base, "the base branch")?;
        let value = self.api(&format!(
            "repos/{}/rules/branches/{}",
            self.repo,
            path_segment(base)
        ))?;
        let rules = value.as_array().ok_or_else(|| {
            invalid(format!(
                "the rules on {}'s base branch {base:?} came back as something that is not a \
                 list of them, so whether it requires an up-to-date branch cannot be read from \
                 it: {value}",
                self.repo
            ))
        })?;
        let mut strict = false;
        for rule in rules {
            if rule.get("type").and_then(|value| value.as_str()) != Some("required_status_checks") {
                continue;
            }
            strict |= rule
                .get("parameters")
                .and_then(|parameters| parameters.get("strict_required_status_checks_policy"))
                .and_then(serde_json::Value::as_bool)
                .ok_or_else(|| {
                    invalid(format!(
                        "a required-status-checks rule on {}'s base branch {base:?} does not say \
                         whether it requires an up-to-date branch: {rule}",
                        self.repo
                    ))
                })?;
        }
        Ok(strict)
    }

    /// Whether classic branch protection on `base` requires an up-to-date branch, or
    /// why the host would not say — the same three answers
    /// [`classic_protection`](GitHub::classic_protection) tells apart.
    fn classic_up_to_date(&self, base: &str) -> Result<std::result::Result<bool, String>> {
        addressable_branch(base, "the base branch")?;
        let path = format!(
            "repos/{}/branches/{}/protection/required_status_checks",
            self.repo,
            path_segment(base)
        );
        let answer = gh::attempt(&["api", &path])?;
        if answer.code != Some(0) {
            if answer.stderr.contains(NOT_PROTECTED) {
                return Ok(Ok(false));
            }
            return Ok(Err(answer.detail()));
        }
        let value = gh::json(&answer.stdout)?;
        value
            .get("strict")
            .and_then(serde_json::Value::as_bool)
            .map(Ok)
            .ok_or_else(|| {
                invalid(format!(
                    "the classic branch protection on {}'s {base:?} came back without saying \
                     whether it requires an up-to-date branch: {value}",
                    self.repo
                ))
            })
    }

    /// The log of the Actions job one check ran in, addressed by the job's id.
    ///
    /// The id comes from the same listing the checks did, so the name matched here
    /// is the name reported there — the two cannot drift apart the way a check's
    /// name and a details URL parsed out of a second command can.
    fn actions_log(&self, cr: &ChangeRequest, name: &str) -> Result<String> {
        let job = self
            .actions_jobs(cr)?
            .into_iter()
            .find(|job| job.name == name)
            .ok_or_else(|| {
                invalid(format!(
                    "GitHub Actions reports no job named {name:?} on the head commit of {}",
                    cr.url
                ))
            })?;
        gh::invoke_content(&[
            "api",
            &format!("repos/{}/actions/jobs/{}/logs", self.repo, job.id),
        ])
    }

    fn api(&self, path: &str) -> Result<serde_json::Value> {
        gh::json(&gh::invoke(&["api", path])?)
    }

    /// One change request as `gh` describes it, in the fields the caller reads and
    /// no others.
    ///
    /// The field list is the caller's, not a constant, because `gh pr view` fails
    /// *whole* when GitHub refuses one field of the several it was asked for — and
    /// the fields differ in what a credential has to be allowed to see.
    /// `statusCheckRollup` resolves the change request's check runs, which a
    /// fine-grained token may not read without the repository's Checks permission,
    /// while `state`, `headRefOid`, and `mergeCommit` need no more than the read
    /// access that found the change request at all. This build asked for all of them
    /// at every call, so a token that could open and merge a change request was
    /// refused at both over a field neither reads — and only from the moment a check
    /// first appeared on it, so the same credential merged a young change request and
    /// failed on an older one. Ask for what you read.
    fn view(&self, id: &str, fields: &str) -> Result<serde_json::Value> {
        addressable(id, "change request id")?;
        let raw = gh::invoke(&["pr", "view", id, "--repo", &self.repo, "--json", fields])?;
        gh::json(&raw)
    }
}

/// One value bound for `gh`'s argument vector, checked before it gets there.
///
/// Every string in these methods arrives from outside — off a caller embedding this
/// crate, or out of the host's own answer — and each becomes a positional or an
/// option's value. `gh` reads a leading `-` as an option of its own and an empty
/// string as a present-but-blank value, so a name shaped like either is refused
/// here rather than silently addressing something other than what it names.
///
/// Those two and no others: the vector is handed to `gh` as a vector, so no shell
/// splits an argument this build wrote as one. Whitespace was refused here too, and
/// [`matchable`] records what that cost — a value with a space in it reaches `gh`
/// exactly as it was written, so it is nothing this boundary has to guard against.
fn addressable(value: &str, what: &str) -> Result<()> {
    if value.is_empty() || value.starts_with('-') {
        return Err(invalid(format!(
            "{what} {value:?} cannot address anything on the host: it must be non-empty and must \
             not begin with '-'"
        )));
    }
    Ok(())
}

/// A check's name, checked before it is matched against the ones the host reports.
///
/// Deliberately not [`addressable`], because a check's name never becomes an
/// argument to `gh`: [`GitHub::job_log`] matches it against the `name` field of the
/// JSON `gh pr checks` returns, and [`GitHub::actions_log`] against the job names
/// the Actions API lists — the job's *id* is what either one then asks for a log by.
/// So `gh`'s argument grammar has no say over this value, and the shapes that
/// grammar refuses are ordinary job names here. Every GitHub matrix job is named
/// with whitespace — `check (macos-latest)`, `test-os (windows-latest)` — and
/// refusing it left every matrix check on every change request this crate published
/// recorded with no log at all.
///
/// What is refused is a name that can match nothing the host reports: the host names
/// every check it answers about, so an empty name is a value a caller made up rather
/// than a check anything has.
fn matchable(name: &str, cr: &ChangeRequest) -> Result<()> {
    if name.is_empty() {
        return Err(invalid(format!(
            "onevcs will not ask for the log of a check with no name on {}: every check the host \
             reports is named, so an empty name matches none of them. This build refused the \
             request; the host was not asked.",
            cr.url
        )));
    }
    Ok(())
}

/// One value bound for a URL path segment, percent-encoded.
///
/// A branch name is the only one of these, and git allows characters in one that a
/// path does not — a `/` above all, which every stacked branch carries and which
/// would otherwise address a resource nobody asked for. Encoded rather than
/// refused: `feature/thing` is a branch a repository really has.
fn path_segment(value: &str) -> String {
    let mut encoded = String::with_capacity(value.len());
    for byte in value.bytes() {
        if byte.is_ascii_alphanumeric() || matches!(byte, b'-' | b'.' | b'_' | b'~') {
            encoded.push(char::from(byte));
        } else {
            encoded.push_str(&format!("%{byte:02X}"));
        }
    }
    encoded
}

/// A commit hash bound for a query string, checked before it gets there.
///
/// It arrives off a host response and goes on to be one parameter of several, so a
/// value carrying anything but hex digits could address a query nobody wrote.
fn commit(sha: &Sha) -> Result<&str> {
    if !is_commit_hash(&sha.0) {
        return Err(invalid(format!(
            "{:?} is not a commit hash, so the checks reported against it cannot be asked for",
            sha.0
        )));
    }
    Ok(&sha.0)
}

/// Whether text the host handed over is the shape of a commit hash.
///
/// The one shape rule, so the two boundaries that read a commit off this host agree
/// about what one is: the commit an Actions query is built from ([`commit`]) and the
/// commit a host says it attached a check to ([`reported_sha`]). Shared as the rule
/// rather than as the refusal because each of those refusals costs something
/// different and says so.
fn is_commit_hash(sha: &str) -> bool {
    !sha.is_empty() && sha.chars().all(|c| c.is_ascii_hexdigit())
}

/// One workflow job, as the Actions API reports it.
struct Job {
    /// What `gh api .../actions/jobs/<id>/logs` addresses its log by.
    id: u64,
    /// The job's name, which is the name GitHub posts its check run under.
    name: String,
    /// Where it is, in the host's own vocabulary.
    status: String,
    /// How it ended, once it has.
    conclusion: Option<String>,
    /// The commit its workflow run was reported against, off that run's own record.
    head: Option<Sha>,
    /// Where a human reads this job on the host.
    url: Option<Url>,
    /// When the host says this job started.
    started_at: Option<String>,
    /// When the host says this job completed.
    completed_at: Option<String>,
}

/// One job of an Actions listing, required to say what it is.
///
/// Defaulting a missing field is what would let a job that is still running be read
/// as a green check, which is the shape that must never be inferred — the same rule
/// [`check`] below is held to, for the same reason.
fn job(entry: &serde_json::Value, cr: &ChangeRequest, head: Option<&Sha>) -> Result<Job> {
    let field = |name: &str| -> Result<&str> {
        entry
            .get(name)
            .and_then(|value| value.as_str())
            .filter(|value| !value.is_empty())
            .ok_or_else(|| {
                invalid(format!(
                    "the Actions API returned a job on {} with no {name}: {entry}",
                    cr.url
                ))
            })
    };
    Ok(Job {
        id: entry
            .get("id")
            .and_then(serde_json::Value::as_u64)
            .ok_or_else(|| {
                invalid(format!(
                    "the Actions API returned a job on {} with no id, so its log cannot be asked \
                     for: {entry}",
                    cr.url
                ))
            })?,
        name: field("name")?.to_owned(),
        status: field("status")?.to_ascii_lowercase(),
        // Genuinely absent while a job is still running.
        conclusion: entry
            .get("conclusion")
            .and_then(|value| value.as_str())
            .filter(|value| !value.is_empty())
            .map(str::to_ascii_lowercase),
        head: head.cloned(),
        url: reported_url(entry, "html_url"),
        started_at: reported_time(entry, "started_at"),
        completed_at: reported_time(entry, "completed_at"),
    })
}

/// The commit a host response names, off whichever field that response spells it
/// in, or `None` where it named none.
///
/// Lenient about *absence* where [`head_sha`] refuses, and the difference is what
/// each answer is for: a change request with no head is one this build cannot go on
/// to ask the Actions API about at all, while a check whose source would not say
/// which commit it is about is still a check, and reading it is still the answer.
/// Inventing one is the thing that must not happen — which is why an absent field
/// reads `None` rather than falling back to the head the caller already holds.
///
/// Strict about *shape* exactly where it is lenient about absence, and for the same
/// reason. This is host-supplied text arriving at a trust boundary, and it is held
/// to [`is_commit_hash`] like every other commit read off this host. Passed through
/// unchecked it would be worse than a refusal rather than more forgiving than one:
/// what a publication does with this value is compare it to the commit it pushed,
/// so anything that is not a commit hash matches nothing and reads as "the host has
/// said nothing about your commit yet" — silence — for as long as the watch runs,
/// and then times out naming a commit the host never had trouble with. A malformed
/// answer must not be able to wear the one shape this path exists to keep distinct
/// from an answer.
fn reported_sha(value: &serde_json::Value, field: &str) -> Result<Option<Sha>> {
    let Some(reported) = value
        .get(field)
        .and_then(|value| value.as_str())
        .filter(|value| !value.is_empty())
    else {
        return Ok(None);
    };
    if !is_commit_hash(reported) {
        return Err(invalid(format!(
            "gh reported {reported:?} as the {field} of a check, and that is not a commit hash, \
             so which commit the check is attached to cannot be decided"
        )));
    }
    Ok(Some(Sha(reported.to_owned())))
}

/// Where the host says something is, off whichever field that response spells it
/// in, or `None` where it named nowhere this build can point a person at.
///
/// A URL that will not parse is `None` rather than a refusal: this is the address a
/// refusal hands over so somebody can go and look, and failing a publication over a
/// link would put the check's verdict behind a formatting complaint.
pub(crate) fn reported_url(entry: &serde_json::Value, field: &str) -> Option<Url> {
    entry
        .get(field)
        .and_then(|value| value.as_str())
        .filter(|value| !value.is_empty())
        .and_then(|value| Url::parse(value).ok())
}

/// The entries of one Actions listing, refused unless the page carried them all.
///
/// These endpoints page, and a page that held entries back is a partial answer
/// about what ran — read as the whole of it, a check nobody saw is a check nobody
/// waited for. `total_count` is how the answer says how many there are, so the two
/// are compared rather than the list being trusted to be complete.
fn listed<'a>(
    value: &'a serde_json::Value,
    field: &str,
    what: &str,
) -> Result<&'a Vec<serde_json::Value>> {
    let entries = value
        .get(field)
        .and_then(|value| value.as_array())
        .ok_or_else(|| {
            invalid(format!(
                "the Actions API answered about {what} with something that lists no {field}: \
                 {value}"
            ))
        })?;
    let total = value
        .get("total_count")
        .and_then(serde_json::Value::as_u64)
        .ok_or_else(|| {
            invalid(format!(
                "the Actions API did not say how many {field} there are on {what}: {value}"
            ))
        })?;
    if total > entries.len() as u64 {
        return Err(invalid(format!(
            "the Actions API reports {total} {field} on {what} and returned {}, so this build has \
             not been shown all of them",
            entries.len()
        )));
    }
    Ok(entries)
}

/// Whether a workflow run or a check suite GitHub reported has completed: its
/// `status` is `completed`. Anything else — `queued`, `in_progress`, `waiting`,
/// `requested`, `pending`, or no status at all — is work still going.
fn is_completed(record: &serde_json::Value) -> bool {
    record
        .get("status")
        .and_then(serde_json::Value::as_str)
        .is_some_and(|status| status.eq_ignore_ascii_case("completed"))
}

/// What GitHub calls one workflow run or check suite, off the field `path` names, or
/// `what` with its id where it names nothing — the name is for a reader, and a record
/// without one is still work in progress.
fn named(record: &serde_json::Value, path: &[&str], what: &str) -> String {
    path.iter()
        .try_fold(record, |value, field| value.get(field))
        .and_then(serde_json::Value::as_str)
        .filter(|name| !name.is_empty())
        .map_or_else(
            || {
                let id = record
                    .get("id")
                    .map_or_else(String::new, |id| format!(" {id}"));
                format!("{what}{id}")
            },
            str::to_owned,
        )
}

/// Whether a REST call failed because GitHub will not let this credential make it —
/// the 403 it answers a fine-grained token the endpoint's permission was not granted
/// to — rather than because the call went wrong.
fn refused_to_this_credential(error: &Error) -> bool {
    let Error::Invalid { reason } = error else {
        return false;
    };
    reason.contains("(HTTP 403)") || reason.contains("Resource not accessible")
}

/// The refusal for a credential that can read none of the host's check sources.
///
/// Both refusals, verbatim, and the permission that answers them — because the
/// thing an operator cannot tell from GitHub's own words is that one of these is
/// not a scope to widen: a fine-grained personal access token has no `Checks`
/// permission to grant, so the rollup is out of reach for that credential class
/// however it is scoped, and `Actions: Read` is what makes the other source work.
fn unreadable(repo: &str, cr: &ChangeRequest, rollup: &Error, actions: &Error) -> Error {
    invalid(format!(
        "neither of GitHub's check sources could be read for {}, so what its checks say is \
         unknown rather than empty. Its check rollup answered: {rollup}. Its Actions API \
         answered: {actions}. Grant this credential `Actions: Read` on {repo} to read the \
         repository's workflow checks — a fine-grained personal access token cannot read the \
         rollup at all, whatever it is scoped to, because GitHub offers no Checks permission for \
         one.",
        cr.url
    ))
}

/// A branch name a caller supplied, checked by the parser that decides branch names.
fn addressable_branch(value: &str, what: &str) -> Result<()> {
    if !git::is_valid_branch_name(value) {
        return Err(invalid(format!(
            "{what} {value:?} is a name git would not accept"
        )));
    }
    Ok(())
}

impl RemoteHost for GitHub {
    fn authenticated_user(&self) -> Result<String> {
        let login = gh::invoke(&["api", "user", "--jq", ".login"])?
            .trim()
            .to_owned();
        if login.is_empty() {
            return Err(Error::Invalid {
                reason: "gh reported no authenticated user".to_owned(),
            });
        }
        Ok(login)
    }

    fn open_change(&self, req: ChangeSpec) -> Result<ChangeRequest> {
        addressable_branch(&req.head, "the head branch")?;
        addressable_branch(&req.base, "the base branch")?;
        // The publication rule the crate states, applied rather than restated —
        // and applied here as well as where a publication takes a request in, because
        // [`ChangeSpec`] is public and this trait is reachable directly. A reason no
        // publication could have carried must not open a change request the record
        // then cannot say anything readable about.
        if let Some(reason) = &req.draft {
            reason.checked()?;
        }
        let body = req.body.unwrap_or_default();
        let mut args = vec![
            "pr", "create", "--repo", &self.repo, "--head", &req.head, "--base", &req.base,
            "--title", &req.title, "--body", &body,
        ];
        // The whole of what the reason does at the host: it opens as a draft. Nothing
        // of the reason itself is written there — see `DraftReason`.
        if req.draft.is_some() || req.draft_awaiting_checks {
            args.push("--draft");
        }
        let raw = gh::invoke(&args)?;
        let url = raw
            .lines()
            .map(str::trim)
            .rfind(|line| line.starts_with("http"))
            .ok_or_else(|| invalid(format!("gh pr create printed no URL: {raw:?}")))?;
        let parsed = Url::parse(url)
            .map_err(|e| invalid(format!("gh pr create printed {url:?}, not a URL: {e}")))?;
        // The host numbers its change requests, and the number is the last segment
        // of the URL it printed. Anything else in that position is `gh` having
        // printed something other than a change request's URL, which is not an
        // identifier to go on addressing it by.
        let id = parsed
            .path_segments()
            .and_then(|mut segments| segments.next_back())
            .filter(|segment| !segment.is_empty() && segment.chars().all(|c| c.is_ascii_digit()))
            .ok_or_else(|| {
                invalid(format!(
                    "gh pr create printed {url:?}, which names no change"
                ))
            })?
            .to_owned();
        Ok(ChangeRequest {
            // The commit the new change request's checks will be reported against,
            // which `gh pr create` does not print.
            head_sha: head_sha(&self.view(&id, "headRefOid")?)?,
            id: ChangeId(id),
            url: parsed,
            base: req.base,
        })
    }

    fn find_changes(&self, head: &str, base: &str) -> Result<Vec<ChangeRequest>> {
        addressable_branch(head, "the head branch")?;
        addressable_branch(base, "the base branch")?;
        let raw = gh::invoke(&[
            "pr",
            "list",
            "--repo",
            &self.repo,
            "--head",
            head,
            "--base",
            base,
            "--state",
            "open",
            "--json",
            "number,url,state,headRefOid",
        ])?;
        let value = gh::json(&raw)?;
        let items = value
            .as_array()
            .ok_or_else(|| invalid(format!("gh pr list returned {raw:?}, not a list")))?;
        let mut changes = Vec::new();
        for item in items {
            let url = item.get("url").and_then(|v| v.as_str()).unwrap_or_default();
            let parsed = Url::parse(url)
                .map_err(|e| invalid(format!("gh pr list returned {url:?}, not a URL: {e}")))?;
            let number = item
                .get("number")
                .and_then(|value| value.as_u64())
                .ok_or_else(|| invalid(format!("gh pr list returned no number: {raw:?}")))?
                .to_string();
            changes.push(ChangeRequest {
                id: ChangeId(number),
                url: parsed,
                head_sha: head_sha(item)?,
                base: base.to_owned(),
            });
        }
        Ok(changes)
    }

    fn change_checks(&self, cr: &ChangeRequest) -> Result<ChangeChecks> {
        let rollup = |checks| ChangeChecks {
            checks,
            sources: [CheckSource::StatusChecks].into_iter().collect(),
        };
        match consult()? {
            Consult::StatusChecks => self.rollup_checks(cr).map(rollup),
            Consult::Actions => self.actions_checks(cr),
            // The rollup first, because it is the complete answer and this build
            // cannot know what the credential is until it is refused. A refusal is
            // never read as "no checks": the fallback either produces an answer of
            // its own or both refusals are reported together.
            Consult::Either => match self.rollup_checks(cr) {
                Ok(checks) => Ok(rollup(checks)),
                Err(refused) if check_rollup_refused_for_pat(&refused) => self
                    .actions_checks(cr)
                    .map_err(|actions| unreadable(&self.repo, cr, &refused, &actions)),
                Err(refused) => Err(refused),
            },
        }
    }

    /// Store one check's log as an artifact, or refuse.
    ///
    /// A fetch that did not produce the log is an error and no artifact: an
    /// artifact reads as what the check printed, so one holding the reason there is
    /// none is a lie in the only record a consumer has. A caller that must not fail
    /// over a missing log declines to fail over this error, as [`crate::publish`]
    /// does.
    fn check_log(&self, cr: &ChangeRequest, check: &Check) -> Result<ArtifactId> {
        // The two are told apart, because they are two different things to whoever
        // reads the refusal: a name this build would not ask about is this build's
        // own doing and says so, while a host that would not produce the log is the
        // host's and quotes what it said. They used to arrive as one message
        // attributing both to the host, which is how a refusal this crate made read
        // as GitHub keeping its logs — and why it went unnoticed for as long as it
        // did. The kind stays `Invalid` for both, because the failure vocabulary is
        // fixed across the libraries that route on it; the message is where the
        // actor is named.
        matchable(&check.name, cr)?;
        let log = self.log_of(cr, &check.name).map_err(|error| {
            invalid(format!(
                "the host could not produce a log for check {:?} on {}: {error}",
                check.name, cr.url
            ))
        })?;
        Ok(ArtifactId(stream::store_artifact("log", &log)?.id))
    }

    fn merge(&self, cr: &ChangeRequest, policy: MergePolicy) -> Result<MergeOutcome> {
        match policy {
            MergePolicy::LocalDirect | MergePolicy::ChangeOpen => Ok(MergeOutcome::Open),
            MergePolicy::ChangeAuto => {
                addressable(&cr.id.0, "change request id")?;
                gh::invoke(&[
                    "pr", "merge", &cr.id.0, "--repo", &self.repo, "--squash", "--auto",
                ])?;
                let view = self.view(&cr.id.0, MERGE_FIELDS)?;
                Ok(match merged_sha(&view, cr)? {
                    Some(sha) => MergeOutcome::Merged(sha),
                    None => MergeOutcome::Queued,
                })
            }
            MergePolicy::ChangeDirect => {
                addressable(&cr.id.0, "change request id")?;
                gh::invoke(&["pr", "merge", &cr.id.0, "--repo", &self.repo, "--squash"])?;
                let view = self.view(&cr.id.0, MERGE_FIELDS)?;
                match merged_sha(&view, cr)? {
                    Some(sha) => Ok(MergeOutcome::Merged(sha)),
                    None => Err(Error::GateFailed {
                        reason: format!(
                            "the host accepted the merge of {} but reports it unmerged",
                            cr.url
                        ),
                    }),
                }
            }
        }
    }

    fn ready_for_review(&self, cr: &ChangeRequest) -> Result<()> {
        addressable(&cr.id.0, "change request id")?;
        gh::invoke(&["pr", "ready", &cr.id.0, "--repo", &self.repo]).map(|_| ())
    }

    /// One `gh pr edit`, handing the body over as a file.
    ///
    /// A file rather than `--body`, because a body is prose of unbounded size and an
    /// argument vector is not where multi-line Markdown belongs: the same reason
    /// `onevcs publish` takes `--body-file`. The file is written under a scratch
    /// directory that goes when this returns, and `gh` reads it before then.
    fn describe_change(&self, cr: &ChangeRequest, title: Option<&str>, body: &str) -> Result<()> {
        addressable(&cr.id.0, "change request id")?;
        let scratch = tempfile::tempdir().map_err(|failure| {
            invalid(format!(
                "could not create a scratch directory to hand the body of {} to gh: {failure}",
                cr.url
            ))
        })?;
        let file = scratch.path().join("body.md");
        std::fs::write(&file, body).map_err(crate::error::at(
            "write the change request's body to",
            &file,
        ))?;
        let file = file.to_string_lossy().into_owned();
        let mut args = vec![
            "pr",
            "edit",
            &cr.id.0,
            "--repo",
            &self.repo,
            "--body-file",
            &file,
        ];
        if let Some(title) = title {
            args.extend(["--title", title]);
        }
        gh::invoke(&args).map(|_| ())
    }

    /// One `gh pr view`, reading exactly the two fields a description is.
    ///
    /// A response missing either is refused rather than read as empty: "the change
    /// request has no body" and "the host would not say what the body is" are
    /// different facts, and a caller finishing a description off what it reads here
    /// would otherwise start from a blank the host never answered.
    fn change_description(&self, cr: &ChangeRequest) -> Result<Description> {
        addressable(&cr.id.0, "change request id")?;
        let view = self.view(&cr.id.0, "title,body")?;
        let read = |name: &str| {
            view.get(name)
                .and_then(serde_json::Value::as_str)
                .map(str::to_owned)
                .ok_or_else(|| {
                    invalid(format!(
                        "gh pr view answered about {} without its {name}, so what the change \
                         request says cannot be read",
                        cr.url
                    ))
                })
        };
        Ok(Description {
            title: read("title")?,
            body: read("body")?,
        })
    }

    /// One `gh api repos/{owner}/{name}`, reading the two fields GitHub states a
    /// repository's visibility in. `visibility` is `public`, `private` or `internal`,
    /// and an internal repository is private to everybody outside its organisation;
    /// an answer carrying neither field is refused rather than read as either.
    fn visibility(&self) -> Result<Visibility> {
        let answer = gh::json(&gh::invoke(&["api", &format!("repos/{}", self.repo)])?)?;
        match answer.get("visibility").and_then(serde_json::Value::as_str) {
            Some("public") => return Ok(Visibility::Public),
            Some("private" | "internal") => return Ok(Visibility::Private),
            _ => {}
        }
        match answer.get("private").and_then(serde_json::Value::as_bool) {
            Some(true) => Ok(Visibility::Private),
            Some(false) => Ok(Visibility::Public),
            None => Err(invalid(format!(
                "gh api answered about {} without its visibility",
                self.repo
            ))),
        }
    }

    /// One `gh pr view`, reading the one field the host decides a draft by.
    ///
    /// A response that does not carry it is refused rather than read as "not a
    /// draft": what this decides is whether a change may be asked to merge, and
    /// "could not look" reported as "nothing is holding it" is the one answer that
    /// lands work somebody held back.
    fn is_draft(&self, cr: &ChangeRequest) -> Result<bool> {
        addressable(&cr.id.0, "change request id")?;
        let view = self.view(&cr.id.0, "isDraft")?;
        view.get("isDraft")
            .and_then(serde_json::Value::as_bool)
            .ok_or_else(|| {
                invalid(format!(
                    "gh pr view answered about {} without saying whether it is a draft, so \
                     whether the host is holding it cannot be read",
                    cr.url
                ))
            })
    }

    /// One `gh pr view`, reading exactly the two fields a merge is decided from —
    /// the same read `merge` makes, through the same parser, so watching a change
    /// request and merging one cannot come to disagree about what merged means.
    fn merged_at(&self, cr: &ChangeRequest) -> Result<Option<Sha>> {
        addressable(&cr.id.0, "change request id")?;
        merged_sha(&self.view(&cr.id.0, MERGE_FIELDS)?, cr)
    }

    /// One `gh pr view`, reading the two fields a merge is decided from and the head
    /// it was of — the read [`merged_at`](RemoteHost::merged_at) makes, through the same
    /// parser, with `headRefOid` beside it rather than in a second request.
    fn merged_head(&self, cr: &ChangeRequest) -> Result<Option<MergedChange>> {
        addressable(&cr.id.0, "change request id")?;
        let view = self.view(&cr.id.0, MERGED_HEAD_FIELDS)?;
        let Some(commit) = merged_sha(&view, cr)? else {
            return Ok(None);
        };
        Ok(Some(MergedChange {
            commit,
            head: head_sha(&view)?,
        }))
    }

    /// One `gh pr view`, reading exactly the field GitHub records a merge's time in.
    ///
    /// `mergedAt` is null — or GitHub's zero time — on a change request that has not
    /// merged, and both read as `None`.
    fn merge_time(&self, cr: &ChangeRequest) -> Result<Option<String>> {
        addressable(&cr.id.0, "change request id")?;
        Ok(reported_time(
            &self.view(&cr.id.0, MERGE_TIME_FIELDS)?,
            "mergedAt",
        ))
    }

    /// One `gh pr view`, reading exactly the two fields GitHub decides a merge's
    /// mergeability by.
    ///
    /// `mergeable` is the answer — `MERGEABLE`, `CONFLICTING`, or `UNKNOWN` while
    /// GitHub has not computed it — and `mergeStateStatus` is quoted beside it, since
    /// `DIRTY` is how GitHub spells the same conflict there. A response that carries
    /// no `mergeable`, or a value GitHub has never answered, is refused rather than
    /// read as either: a conflict nobody reported ends a publication, and a mergeable
    /// change nobody reported hides one.
    fn mergeability(&self, cr: &ChangeRequest) -> Result<Mergeability> {
        addressable(&cr.id.0, "change request id")?;
        let view = self.view(&cr.id.0, MERGEABILITY_FIELDS)?;
        let mergeable = view
            .get("mergeable")
            .and_then(serde_json::Value::as_str)
            .ok_or_else(|| {
                invalid(format!(
                    "gh pr view answered about {} without saying whether it can merge: {view}",
                    cr.url
                ))
            })?;
        let status = view
            .get("mergeStateStatus")
            .and_then(serde_json::Value::as_str)
            .unwrap_or("unreported");
        match mergeable.to_ascii_uppercase().as_str() {
            "MERGEABLE" => Ok(Mergeability::Mergeable),
            "CONFLICTING" => Ok(Mergeability::Conflicting {
                reported: format!("mergeable: {mergeable}, mergeStateStatus: {status}"),
            }),
            "UNKNOWN" => Ok(Mergeability::Unknown),
            _ => Err(invalid(format!(
                "gh pr view reports {} as mergeable {mergeable:?}, which is none of \
                 MERGEABLE, CONFLICTING or UNKNOWN, so whether it can merge cannot be read",
                cr.url
            ))),
        }
    }

    /// Two REST reads about the commit the caller names: the workflow runs on it, and
    /// its check suites.
    ///
    /// A run counts while its `status` is anything but `completed`; a suite counts
    /// while it is not `completed` **and** has at least one check run, because GitHub
    /// leaves an installed app's empty `queued` suite on every commit for ever. A
    /// credential refused the suites answers from the runs alone, saying so with
    /// `complete: false`; one refused both reads is the runs' refusal.
    ///
    /// The commit asked about is the caller's — a watch names the one it pushed — and
    /// the commit answered about is read off the records GitHub returned, so a host
    /// whose records name another commit says so rather than having this build's
    /// question stamped onto its answer. Reading the change request's head first
    /// would cost a third request on every read and decide nothing these two do not.
    fn head_activity(&self, cr: &ChangeRequest) -> Result<HeadActivity> {
        let sha = commit(&cr.head_sha)?;
        let runs = self.api(&format!(
            "repos/{}/actions/runs?head_sha={sha}&per_page={PAGE}",
            self.repo
        ))?;
        let mut answered: Option<Sha> = None;
        let mut in_progress = Vec::new();
        let mut about = |record: &serde_json::Value| -> Result<()> {
            if let Some(head) = reported_sha(record, "head_sha")? {
                if answered.as_ref().is_some_and(|seen| *seen != head) {
                    return Err(invalid(format!(
                        "GitHub answered about the work on {sha} with records naming two \
                         commits, {} and {}, so which one is running cannot be told",
                        answered.as_ref().map_or("", |seen| seen.0.as_str()),
                        head.0
                    )));
                }
                answered = Some(head);
            }
            Ok(())
        };
        for run in listed(&runs, "workflow_runs", &format!("the commit {sha}"))? {
            about(run)?;
            if !is_completed(run) {
                in_progress.push(named(run, &["name"], "workflow run"));
            }
        }
        let suites = gh::invoke(&[
            "api",
            &format!(
                "repos/{}/commits/{sha}/check-suites?per_page={PAGE}",
                self.repo
            ),
        ]);
        let complete = match suites {
            Ok(raw) => {
                let suites = gh::json(&raw)?;
                for suite in listed(&suites, "check_suites", &format!("the commit {sha}"))? {
                    about(suite)?;
                    let runs = suite
                        .get("latest_check_runs_count")
                        .and_then(serde_json::Value::as_u64)
                        .ok_or_else(|| {
                            invalid(format!(
                                "GitHub reported a check suite on {sha} without saying how many \
                                 check runs it has: {suite}"
                            ))
                        })?;
                    if runs > 0 && !is_completed(suite) {
                        in_progress.push(named(suite, &["app", "name"], "check suite"));
                    }
                }
                true
            }
            Err(refused) if refused_to_this_credential(&refused) => false,
            Err(other) => return Err(other),
        };
        Ok(HeadActivity {
            head: answered.unwrap_or_else(|| Sha(sha.to_owned())),
            in_progress,
            complete,
        })
    }

    /// One GraphQL query document, paged — see `review::github`.
    fn review_comments(&self, change: &ChangeId, since: Option<&ReadMarker>) -> Result<ReviewRead> {
        review::github::read(&self.repo, change, since)
    }

    /// One REST request — see `review::github`.
    fn reply_to_comment(
        &self,
        change: &ChangeId,
        comment: &CommentId,
        body: &str,
        _key: &str,
    ) -> Result<PostedReply> {
        review::github::reply(&self.repo, change, comment, body)
    }

    /// One aliased GraphQL query per 50 change requests — see `standings::github`.
    fn change_standings(&self, root: &str, changes: &[ChangeId]) -> Result<ChangeStandings> {
        standings::github::read(&self.repo, root, changes)
    }

    /// One aliased GraphQL query per 150 change requests — see `standings::github`.
    fn review_standings(
        &self,
        root: &str,
        changes: &[(ChangeId, Option<ReadMarker>)],
    ) -> Result<ReviewStandings> {
        standings::github::read_with_reviews(&self.repo, root, changes)
    }

    /// One REST request — see `standings::github`.
    fn retarget_change(
        &self,
        change: &ChangeId,
        base: &str,
        body: Option<&str>,
    ) -> Result<ReadCost> {
        standings::github::retarget(&self.repo, change, base, body)
    }

    /// One REST request — see `standings::github`.
    fn reopen_change(&self, change: &ChangeId) -> Result<ReadCost> {
        standings::github::reopen(&self.repo, change)
    }

    fn required_checks_on(&self, base: &str) -> Result<RequiredChecks> {
        // The rulesets first, and their refusal is the whole read's: every
        // credential that can see the repository can read them, so one that cannot
        // has not answered about anything. Classic protection is asked second and
        // its refusal is recorded rather than raised, because the credential GitHub
        // steers people toward is refused it on every repository, and the rulesets'
        // answer is still worth having — marked as the part it is.
        let checks = self.required_on(base, &self.repo)?;
        let mut answer = RequiredChecks {
            checks,
            unconsulted: BTreeMap::new(),
        };
        match self.classic_protection(base)? {
            Ok(contexts) => answer.checks.extend(contexts),
            Err(refusal) => {
                answer
                    .unconsulted
                    .insert(ProtectionSource::BranchProtection, refusal);
            }
        }
        Ok(answer)
    }

    fn requires_up_to_date(&self, base: &str) -> Result<bool> {
        // The two sources `required_checks_on` reads, in the same order and with the
        // same weight: either one requiring it is the answer, and "neither does" is an
        // answer only once both were read.
        if self.ruled_up_to_date(base)? {
            return Ok(true);
        }
        match self.classic_up_to_date(base)? {
            Ok(strict) => Ok(strict),
            Err(refusal) => Err(invalid(format!(
                "whether {}'s {base:?} requires an up-to-date branch could not be read from \
                 classic branch protection: {refusal}",
                self.repo
            ))),
        }
    }

    /// One `gh repo view`, reading the one field GitHub decides it by.
    fn allows_auto_merge(&self) -> Result<bool> {
        let answer = gh::json(&gh::invoke(&[
            "repo",
            "view",
            &self.repo,
            "--json",
            "autoMergeAllowed",
        ])?)?;
        answer
            .get("autoMergeAllowed")
            .and_then(serde_json::Value::as_bool)
            .ok_or_else(|| {
                invalid(format!(
                    "gh repo view answered about {} without saying whether it allows auto-merge: \
                     {answer}",
                    self.repo
                ))
            })
    }
}

/// The refusal for a host that will not say which of its checks block the merge.
///
/// Inferring it is the one shape that must never be guessed: it is the difference
/// between a merge that was gated and one that only looked like it. What changed
/// with the real backend is *where* the question is asked, not that it may be
/// skipped — a host that answers `gh pr checks --required` with something other
/// than the checks it requires is refused here rather than read as requiring none.
fn unsaid(url: &Url, detail: &str) -> Error {
    invalid(format!(
        "gh pr checks --required answered about {url} with {detail}, so a check it reported does \
         not say whether it blocks the merge"
    ))
}

/// The same refusal from the rulesets read, which is addressed to a base branch and
/// names whatever asked about it — a change request's URL, or an identity.
fn unsaid_about(about: &str, detail: &str) -> Error {
    invalid(format!(
        "the repository's rulesets answered about {about} with {detail}, so a check they \
         reported does not say whether it blocks the merge"
    ))
}

/// The refusal for a host that has no check on the head to report yet.
///
/// Separate from [`unsaid`] because the next move differs: this one is a clock, and
/// whoever reads it waits rather than re-runs.
fn unregistered(url: &Url, said: &str) -> Error {
    invalid(format!(
        "the host reports no check at all yet on the head of {url} ({said}), which is not the \
         same as a repository that declares none required — it is what a head pushed moments ago \
         looks like — so whether anything blocks the merge cannot be read from it"
    ))
}

/// One entry of the host's check rollup, required to say what it is.
///
/// Defaulting a missing field here is what would let a host that answered
/// partially be read as a green check — the one shape that must never be inferred.
/// Whether the check *blocks* is not in this entry at all and never was: it comes
/// from [`GitHub::required_checks`], which refuses a host that will not say.
///
/// Two shapes, told apart by the `__typename` GitHub writes on each: a check run
/// (`CheckRun` — every Actions job, and any app's run), and a commit status an
/// integration posted (`StatusContext`), which [`status_context`] reads.
fn check(
    entry: &serde_json::Value,
    cr: &ChangeRequest,
    required: &BTreeSet<String>,
    head: Option<&Sha>,
) -> Result<Check> {
    let field = |name: &str| -> Result<&str> {
        entry
            .get(name)
            .and_then(|value| value.as_str())
            .filter(|value| !value.is_empty())
            .ok_or_else(|| {
                invalid(format!(
                    "gh pr view returned a check on {} with no {name}: {entry}",
                    cr.url
                ))
            })
    };
    if entry.get("__typename").and_then(|value| value.as_str()) == Some("StatusContext") {
        return status_context(entry, cr, required, head);
    }
    let name = field("name").or_else(|_| field("context"))?.to_owned();
    let blocks = required.contains(&name);
    Ok(Check {
        name,
        status: field("status")?.to_ascii_lowercase(),
        // Genuinely absent while a check is still running, which is the one thing
        // the host cannot yet know.
        conclusion: entry
            .get("conclusion")
            .and_then(|value| value.as_str())
            .filter(|value| !value.is_empty())
            .map(str::to_ascii_lowercase),
        required: blocks,
        head: head.cloned(),
        // A check run's own address; a commit status spells its own `targetUrl`, and
        // is read by `status_context`.
        url: reported_url(entry, "detailsUrl"),
        started_at: reported_time(entry, "startedAt"),
        completed_at: reported_time(entry, "completedAt"),
    })
}

/// One commit status in the host's rollup, read as the check it stands for.
///
/// A commit status has one `state` where a check run has a status and a conclusion:
/// `PENDING` (and `EXPECTED`, a required context nothing has posted yet) is a check
/// still running, and `SUCCESS`, `FAILURE` and `ERROR` are a completed one concluding
/// that way — so an `ERROR` is red, as every conclusion `Check::state` does not name
/// is. A state none of those five is refused rather than guessed at: reading an
/// unknown one as either would wave a merge through or hold one for ever. It is named
/// by its `context`, which is what branch protection requires it by, and its start is
/// when the integration posted it — a status posted again after a lift is a new one.
fn status_context(
    entry: &serde_json::Value,
    cr: &ChangeRequest,
    required: &BTreeSet<String>,
    head: Option<&Sha>,
) -> Result<Check> {
    let named = |name: &str| {
        entry
            .get(name)
            .and_then(|value| value.as_str())
            .filter(|value| !value.is_empty())
            .ok_or_else(|| {
                invalid(format!(
                    "gh pr view returned a commit status on {} with no {name}: {entry}",
                    cr.url
                ))
            })
    };
    let name = named("context")?.to_owned();
    let state = named("state")?;
    let (status, conclusion) = match state.to_ascii_uppercase().as_str() {
        "PENDING" | "EXPECTED" => ("pending", None),
        "SUCCESS" => ("completed", Some("success")),
        "FAILURE" => ("completed", Some("failure")),
        "ERROR" => ("completed", Some("error")),
        _ => {
            return Err(invalid(format!(
                "gh pr view reports commit status {name:?} on {} as {state:?}, which is none of \
                 PENDING, EXPECTED, SUCCESS, FAILURE or ERROR, so where it stands cannot be read",
                cr.url
            )))
        }
    };
    Ok(Check {
        required: required.contains(&name),
        name,
        status: status.to_owned(),
        conclusion: conclusion.map(str::to_owned),
        head: head.cloned(),
        url: reported_url(entry, "targetUrl"),
        started_at: reported_time(entry, "startedAt"),
        completed_at: None,
    })
}

/// When a host response says a run started — or completed, or a change merged — or
/// `None` where it did not say.
///
/// GitHub answers a run that is queued and has not started with the zero time rather
/// than with nothing, and that is read as nothing: a start every queued run shares is
/// no run's identity, and a time every unfinished thing shares is no time at all.
pub(crate) fn reported_time(entry: &serde_json::Value, field: &str) -> Option<String> {
    entry
        .get(field)
        .and_then(|value| value.as_str())
        .filter(|value| !value.is_empty() && !value.starts_with("0001-01-01"))
        .map(str::to_owned)
}

/// What [`merged_sha`] reads, which is what a merge asks `gh pr view` for.
const MERGE_FIELDS: &str = "state,mergeCommit";

/// What [`GitHub`]'s `merged_head` reads: [`MERGE_FIELDS`] and the head it merged.
const MERGED_HEAD_FIELDS: &str = "state,mergeCommit,headRefOid";

/// What [`GitHub`]'s `merge_time` reads.
const MERGE_TIME_FIELDS: &str = "mergedAt";

/// What [`GitHub`]'s `mergeability` reads.
const MERGEABILITY_FIELDS: &str = "mergeable,mergeStateStatus";

/// The commit a merged change request landed as, or `None` while it is still open.
///
/// A host reporting `MERGED` without naming the commit is answering wrongly rather
/// than answering "not yet": the commit is the whole evidence that the change
/// reached its base, and reporting a merge with no SHA is the one thing that must
/// not be passed through.
fn merged_sha(view: &serde_json::Value, cr: &ChangeRequest) -> Result<Option<Sha>> {
    let state = view
        .get("state")
        .and_then(|value| value.as_str())
        .ok_or_else(|| invalid(format!("gh pr view returned no state for {}", cr.url)))?;
    if !state.eq_ignore_ascii_case("merged") {
        return Ok(None);
    }
    view.get("mergeCommit")
        .and_then(|commit| commit.get("oid"))
        .and_then(|value| value.as_str())
        .filter(|value| !value.is_empty())
        .map(|sha| Some(Sha(sha.to_owned())))
        .ok_or_else(|| {
            invalid(format!(
                "gh pr view reports {} merged without naming the commit it merged as",
                cr.url
            ))
        })
}

/// The commit a change request's checks are reported against.
fn head_sha(view: &serde_json::Value) -> Result<Sha> {
    view.get("headRefOid")
        .and_then(|value| value.as_str())
        .filter(|value| !value.is_empty())
        .map(|sha| Sha(sha.to_owned()))
        .ok_or_else(|| invalid(format!("gh returned a change request with no head: {view}")))
}