hyperhive/swarm-controller/src/forge.rs

//! Swarm-controller's own forgejo client. Self-contained — deliberately
//! not sharing code with `hive-c0re::forge` across the crate boundary
//! (forcing that split now, over a handful of idempotent CRUD-ish
//! calls, is premature plumbing; factor out a shared crate later if the
//! duplication actually starts to hurt).
//!
//! The forge's URL comes from `swarm-controller.nix`'s `forgeEnv`
//! (`SWARM_CONTROLLER_FORGE_URL`). The token comes from the swarm's secret
//! store, at `swarm_secret_client::forge::controller_token_path`, read under
//! this daemon's own certificate and held only in memory. The forge mints it
//! inside its container and puts it there (`hive-forge/default.nix`'s
//! `forgejo-swarm-controller-account`).
//!
//! [`enable`] fills the daemon's one client slot: at start when the token is
//! there, later from a background re-read when it is not. Every caller reads
//! the slot and treats an empty one as "no forge access here" — the same
//! graceful-absence shape the queue coordinates use — so the daemon serves
//! from the start and gains forge access without a restart. Once filled, the
//! client reads the token again when the forge answers 401 and sends the
//! request once more with it (`forge/api.rs`), so a replaced token needs no restart
//! either.

use anyhow::{Context, Result};
use backon::{ExponentialBuilder, Retryable as _};
use forgejo_api::structs::{
    AddCollaboratorOption, AddCollaboratorOptionPermission, ChangeFileOperation,
    ChangeFileOperationOperation, ChangeFilesOptions, CommitStatusState,
    CreateBranchProtectionOption, CreateHookOption, CreateHookOptionConfig, CreateHookOptionType,
    CreateRepoOption, CreateStatusOption, CreateUserOption, EditUserOption, GetTreeQuery, GitEntry,
    IssueListIssuesQuery, IssueListIssuesQueryState, IssueListIssuesQueryType,
    RepoDownloadPullDiffOrPatchQuery, RepoGetContentsQuery, RepoGetPullRequestCommitsQuery,
    RepoGetPullRequestFilesQuery, RepoListPullRequestsQuery, RepoListPullRequestsQueryState,
    RepoSearchQuery, StateType,
};
use forgejo_api::{ApiErrorKind, ForgejoError};
use futures_util::{StreamExt as _, TryStreamExt as _};
use reqwest::StatusCode;
use serde::{Deserialize, Serialize};
use std::collections::{BTreeMap, HashMap, HashSet};
use std::sync::{Arc, OnceLock};
use utoipa::ToSchema;

use crate::webhook::DeliveryKind;
use swarm_secret_client::stored_secret::StoredSecret;
use swarm_wire::review_classifier::Commit;

pub mod agent_token;
mod api;
pub mod builder_account;
pub mod ci_metrics_account;
pub mod classifier_account;
pub mod docs_grant;
pub mod hive_token;
pub mod legacy_tokens;
pub mod objects;
pub mod operator_profiles;
pub mod pull_diff;
pub mod runner_token;
pub mod site_admin;

/// The longest commit status description [`Client::create_commit_status`]
/// sends, in characters. The forge keeps more; its PR view shows one line.
pub const STATUS_DESCRIPTION_MAX: usize = 200;

/// One commit status, as [`Client::create_commit_status`] writes it.
pub struct CommitStatus<'a> {
    pub context: &'a str,
    pub state: CommitStatusState,
    /// Cut to [`STATUS_DESCRIPTION_MAX`] characters.
    pub description: &'a str,
    /// Where the status links to, if anywhere.
    pub target_url: Option<&'a str>,
}

/// `description` cut to [`STATUS_DESCRIPTION_MAX`] characters, an ellipsis
/// marking a cut.
fn status_description(description: &str) -> String {
    if description.chars().count() <= STATUS_DESCRIPTION_MAX {
        return description.to_owned();
    }
    let mut cut: String = description
        .chars()
        .take(STATUS_DESCRIPTION_MAX - 1)
        .collect();
    cut.push('…');
    cut
}

/// An agent's open config-PR, as [`Client::list_open_config_prs`] reports it
/// and `GET /api/agents/{name}/config-pr` serves it.
#[derive(Clone, Debug, PartialEq, Serialize, Deserialize, ToSchema)]
pub struct ConfigPrStatus {
    pub pr_number: u64,
    /// Absent only if Forgejo itself omitted the field — every real PR has
    /// one; not worth failing the whole scan over.
    pub html_url: Option<String>,
}

/// An open PR as the review try-out's PR dropdown lists it.
#[derive(Clone, Debug, PartialEq, Eq, Serialize, ToSchema)]
pub struct OpenPull {
    pub number: u64,
    pub title: String,
    /// The head commit the PR is at now. Absent only if Forgejo omitted it.
    pub head_sha: Option<String>,
}

/// `pulls` as the dropdown's rows, newest (highest number) first. A PR
/// without a usable number cannot be named, so it is left out.
/// The merge base of a PR's head and base branch, read off its commits (the
/// forge lists them from the current head and base): the one parent that is
/// not itself a PR commit. `None` when the PR has no commits or its commits
/// reach more than one such parent (a merge of the base branch into the head).
///
/// The PR's own `merge_base` field is not used: the forge stores it and
/// refreshes it asynchronously, so right after a rebase or force-push it
/// still names the old base.
fn merge_base_of(commits: &[forgejo_api::structs::Commit]) -> Option<String> {
    let own: HashSet<&str> = commits.iter().filter_map(|c| c.sha.as_deref()).collect();
    let outside: HashSet<&str> = commits
        .iter()
        .flat_map(|c| c.parents.iter().flatten())
        .filter_map(|p| p.sha.as_deref())
        .filter(|sha| !own.contains(sha))
        .collect();
    let mut outside = outside.into_iter();
    match (outside.next(), outside.next()) {
        (Some(base), None) => Some(base.to_owned()),
        _ => None,
    }
}

fn open_pulls(pulls: Vec<forgejo_api::structs::PullRequest>) -> Vec<OpenPull> {
    let mut rows: Vec<OpenPull> = pulls
        .into_iter()
        .filter_map(|pr| {
            Some(OpenPull {
                number: u64::try_from(pr.number?).ok()?,
                title: pr.title.unwrap_or_default(),
                head_sha: pr.head.and_then(|h| h.sha),
            })
        })
        .collect();
    rows.sort_by_key(|r| std::cmp::Reverse(r.number));
    rows
}

/// A PR as [`Client::pull_for_review`] reads it for the review classifier.
#[derive(Clone, Debug, PartialEq, Eq)]
pub struct PullForReview {
    pub head_sha: String,
    pub commits: Vec<Commit>,
    pub diff: String,
    /// Why `diff` is the forge's `.diff`, with 3 lines of context, rather
    /// than [`pull_diff::wide`]'s. `None` when it is the wide one.
    pub narrowed: Option<String>,
}

/// Files on a repo's default branch, as [`Client::default_branch_files`] reads
/// them.
#[derive(Clone, Debug, PartialEq, Eq)]
pub struct DefaultBranchFiles {
    pub branch: String,
    /// The commit the files were read at.
    pub commit: String,
    /// The repo's page on the forge, as a browser reaches it.
    pub web_url: Option<url::Url>,
    /// Each asked-for path that exists there, in the order asked.
    pub files: Vec<BranchFile>,
}

/// One file of [`DefaultBranchFiles`].
#[derive(Clone, Debug, PartialEq, Eq)]
pub struct BranchFile {
    pub path: String,
    /// The blob's sha, which the forge asks for to replace the file.
    pub sha: String,
    pub bytes: Vec<u8>,
}

/// One row of [`Client::issue_report`] — an open issue plus the two facts
/// derived from its dependency graph that swarm-ui's issue-report page
/// sorts/filters on. See that function's doc comment for how both are
/// computed.
#[derive(Clone, Debug, PartialEq, Serialize, ToSchema)]
pub struct IssueReportRow {
    /// `owner/name` — present on every row (not just the cross-repo
    /// aggregate) so a caller can use the same row type for either
    /// endpoint without a second shape to branch on.
    pub repo: String,
    pub number: i64,
    pub title: String,
    pub labels: Vec<String>,
    /// All assignees, not just one — forgejo's `Issue::assignee` is a
    /// legacy single-value field; `assignee`**s** (plural) is the real
    /// multi-assignee list, and this repo's issues do use more than one.
    pub assignees: Vec<String>,
    pub html_url: Option<String>,
    /// True if any of this issue's own dependencies is still open — same
    /// semantics `hive-forge issue dependency list` already uses (a closed
    /// blocker doesn't count).
    pub blocked: bool,
    /// How many OTHER issues in this same report list this issue as a
    /// dependency — the reverse of `blocked`. Ranks "fix this one to
    /// unblock the most other work" highest.
    pub depended_on_by_count: u32,
    /// How many distinct OTHER issues this issue unblocks, following the
    /// same reverse-dependency edges transitively (a depends on b depends
    /// on c: fixing c eventually frees both a and b, so c's count includes
    /// both). A superset of `depended_on_by_count`, which only counts the
    /// direct edge. See [`Client::issue_report`] for how it's walked.
    pub transitively_blocks_count: u32,
}

/// The `operators` team, whitelisted for the merge gate on every repo
/// this client protects. Provisioned by this daemon's own periodic pass
/// ([`objects`]), and checked again by [`Client::create_repo`] before it
/// applies a rule naming the team.
const OPERATORS_TEAM: &str = "operators";

/// The org owning per-agent config repos — where this daemon creates them,
/// seeds them, and where the `pull_request` hook lives. Same value as
/// `hive-c0re::forge::CONFIG_ORG` — one forge, one org.
///
/// ⚠️ Duplicated across the crate boundary (this crate deliberately does not
/// depend on `hive-c0re`), so nothing makes the two fail together. The
/// failure mode if they drift is quiet: the hook is created on an org that
/// exists, forgejo reports it healthy, and it simply never fires.
///
/// ⚠️ **Not `agents`.** That org exists and is a different thing: it is the
/// namespace repos an *agent asks for* land in, per
/// `hive-c0re::forge::AGENTS_ORG`. Agent **config** repos have always lived
/// here, which is where `hive-c0re` reconciles, merges and mirrors them — a
/// config repo created in `agents` is invisible to every one of those paths,
/// and nothing errors, because both orgs exist and both accept a repo.
///
/// The merge gate survives the distinction: [`objects`] provisions the
/// `operators` team in **both** orgs precisely so branch protection can be
/// applied in either.
pub(crate) const CONFIG_ORG: &str = "agent-configs";

/// The hive-wide knowledge repo, where the `push` hook lives. Same values as
/// `hive-c0re::workers::knowledge::{ORG, REPO}`.
const KNOWLEDGE_ORG: &str = "internal";
const KNOWLEDGE_REPO: &str = "knowledge";

/// This daemon's own forge account — same constant as
/// `swarm-controller.nix`'s `swarmControllerForgeUser` (`hive-forge/default.nix`).
/// Needed in the branch-protection push whitelist (see
/// [`Client::apply_operator_branch_protection`]): being a site admin does
/// not exempt a caller from a protected branch's push check on the
/// content-edit API path (`repo_change_files`, what [`Client::seed_agent_config`]
/// uses) the way it does for a real `git push` — confirmed live against a
/// re-test where the account was already promoted to admin and the seed
/// commit still failed `ErrUserCannotCommit`. Whitelisting the account by
/// name is the same shape a git-push-based seed gets for free from the
/// admin bypass, without adding a `git` shell-out to this daemon.
const SWARM_CONTROLLER_FORGE_USER: &str = "swarm-controller";

/// The env var `swarm-controller.nix`'s `forgeEnv` sets. Named here once so
/// [`url_from_env`] and the nix module can't drift silently.
const URL_ENV: &str = "SWARM_CONTROLLER_FORGE_URL";

/// One login and one read. An empty value is refused like an absent one: it
/// would reach the forge as the token and come back as a 401 on the first
/// call, which names neither the store nor the publisher.
async fn fetch_token() -> Result<String> {
    let path = swarm_secret_client::forge::controller_token_path()?;
    let store = crate::store::connect()
        .await
        .context("logging in to the swarm secret store")?;
    let stored: Option<swarm_secret_client::forge::ControllerToken> = store
        .read_optional(&path)
        .await
        .with_context(|| format!("reading {path}"))?;
    match stored {
        Some(t) if !t.value.trim().is_empty() => Ok(t.value.trim().to_owned()),
        _ => anyhow::bail!(
            "nothing at {path}: forgejo-swarm-controller-account, in the forge's container, has \
             not published this daemon's forge token (if the store lost it, delete /var/lib/forgejo/\
             swarm-controller-token-name in that container and restart the unit)"
        ),
    }
}

/// Typed forgejo client, built once at startup from env.
pub struct Client {
    api: api::Api,
    /// The forge's base URL, kept so a client authenticated as someone
    /// else can be built against the same forge (see
    /// [`agent_token`]'s read-back).
    url: url::Url,
}

/// The forge's URL. `Ok(None)` (not an error) when `SWARM_CONTROLLER_FORGE_URL`
/// is unset — no forge is configured for this daemon.
///
/// # Errors
/// A URL that does not parse. Retrying cannot fix that, so it is reported once
/// rather than waited out.
pub fn url_from_env() -> Result<Option<url::Url>> {
    let Some(url) = std::env::var(URL_ENV).ok() else {
        return Ok(None);
    };
    url::Url::parse(&url)
        .with_context(|| format!("parse {URL_ENV} ({url}) as a URL"))
        .map(Some)
}

/// How long the background re-read waits before its first attempt. The caller
/// has just made one, so an immediate retry would only repeat it.
const BACKGROUND_FIRST: std::time::Duration = std::time::Duration::from_secs(30);

/// The longest the background re-read waits between attempts.
const BACKGROUND_CAP: std::time::Duration = std::time::Duration::from_mins(5);

/// Fill `slot` with the first value `attempt` yields: after
/// [`BACKGROUND_FIRST`], then at 60s, doubling, capped at [`BACKGROUND_CAP`].
/// Never gives up. A failure is logged at `info` once the next wait is the cap,
/// so a controller that never gets its token says so once per
/// [`BACKGROUND_CAP`] without a line per early retry, and at `debug` before
/// that. The caller logs the first failure and the recovery.
async fn fill_when_ready<T, F, Fut>(slot: &OnceLock<T>, attempt: F) -> &T
where
    F: FnMut() -> Fut,
    Fut: std::future::Future<Output = Result<T>>,
{
    tokio::time::sleep(BACKGROUND_FIRST).await;
    let value = attempt
        .retry(
            ExponentialBuilder::default()
                .with_min_delay(BACKGROUND_FIRST * 2)
                .with_max_delay(BACKGROUND_CAP)
                .without_max_times(),
        )
        .notify(|e, wait| {
            if wait >= BACKGROUND_CAP {
                tracing::info!(
                    error = %format!("{e:#}"),
                    "still no forge access: the controller's forge token is unavailable"
                );
            } else {
                tracing::debug!(
                    error = %format!("{e:#}"),
                    "the controller's forge token is still unavailable"
                );
            }
        })
        .await
        .expect("a retry with no cap on its tries returns only once it succeeds");
    slot.get_or_init(|| value)
}

/// The longest one unit of forge work, a periodic pass or a job node, may run
/// before it is abandoned. forgejo-api's client has no request timeout, so
/// without this a forge that accepts a connection and never answers holds the
/// pass's loop, or the node, forever. Under every pass's 5-minute interval,
/// and far over what the longest node (`MintHiveForgeToken`, about fifteen
/// sequential forge and store requests) takes while it is making progress.
pub(crate) const FORGE_DEADLINE: std::time::Duration = std::time::Duration::from_mins(2);

/// Run `fut`, the forge work named `work` in the log, under
/// [`FORGE_DEADLINE`]. `None` means it was abandoned at the deadline; that is
/// logged here.
pub(crate) async fn within_forge_deadline<T>(
    work: &str,
    fut: impl std::future::Future<Output = T>,
) -> Option<T> {
    let result = tokio::time::timeout(FORGE_DEADLINE, fut).await.ok();
    if result.is_none() {
        tracing::warn!(
            work,
            deadline_s = FORGE_DEADLINE.as_secs(),
            "forge work: no answer within the deadline; abandoned"
        );
    }
    result
}

/// Put a client for `url` into `slot` and hand it to `on_ready`: at once when
/// the store has the token, otherwise from a background task that re-reads it
/// until it does. Either way the daemon starts, and every forge-backed path
/// reads `slot`, so it answers "no forge" until the token lands and works
/// from then on.
pub async fn enable(
    url: url::Url,
    slot: Arc<OnceLock<Arc<Client>>>,
    on_ready: impl FnOnce(Arc<Client>) + Send + 'static,
) {
    match Client::connect(&url).await {
        Ok(client) => on_ready(Arc::clone(slot.get_or_init(|| Arc::new(client)))),
        Err(e) => {
            tracing::warn!(
                error = %format!("{e:#}"),
                "no forge access yet; re-reading the forge token from the swarm secret store in \
                 the background"
            );
            tokio::spawn(async move {
                let client = Arc::clone(
                    fill_when_ready(&slot, || async {
                        Client::connect(&url).await.map(Arc::new)
                    })
                    .await,
                );
                tracing::info!("the forge token is in the swarm secret store; forge access is on");
                on_ready(client);
            });
        }
    }
}

impl Client {
    /// One read of the token from the secret store, and a client for `url`
    /// holding it, which reads it again when the forge refuses it.
    ///
    /// # Errors
    /// The store refusing the login, nothing at the path because the forge's
    /// host has not published it yet, or the client failing to build.
    async fn connect(url: &url::Url) -> Result<Self> {
        let token = StoredSecret::new("swarm-controller-forge", fetch_token().await?, || async {
            fetch_token().await.map_err(Into::into)
        });
        let api = api::Api::new(url.clone(), token).context("build forgejo client")?;
        Ok(Self {
            api,
            url: url.clone(),
        })
    }

    /// Creation options for an empty repo defaulting to `main`.
    fn repo_option(name: &str, private: bool) -> CreateRepoOption {
        CreateRepoOption {
            auto_init: Some(false),
            default_branch: Some("main".to_owned()),
            description: None,
            gitignores: None,
            issue_labels: None,
            license: None,
            name: name.to_owned(),
            object_format_name: None,
            private: Some(private),
            readme: None,
            template: None,
            trust_model: None,
        }
    }

    /// Create `name` inside [`CONFIG_ORG`]. Idempotent: an existing repo
    /// (409, or a 422 confirmed by a follow-up `repo_get`) is folded into
    /// success. A 422 that is *not* an existing repo — forgejo also uses
    /// it for other repo-create validation failures — fails with
    /// forgejo's own message rather than being folded silently (see
    /// [`folds_into_success`]).
    async fn ensure_org_repo(&self, name: &str) -> Result<()> {
        match self
            .api
            .call(|api| {
                api.create_org_repo(CONFIG_ORG, Self::repo_option(name, true))
                    .into_future()
            })
            .await
        {
            Ok(_) => {
                tracing::info!(%name, "swarm forge: created repo in {CONFIG_ORG}");
                Ok(())
            }
            Err(e) => {
                let existing = is_ambiguous_validation_failure(&e)
                    && self
                        .api
                        .call(|api| api.repo_get(CONFIG_ORG, name).into_future())
                        .await
                        .is_ok();
                if folds_into_success(&e, existing) {
                    tracing::debug!(%name, "swarm forge: repo already exists");
                    Ok(())
                } else {
                    tracing::error!(%name, error = %e, "swarm forge: forge refused to create repo");
                    Err(e).with_context(|| format!("create repo {CONFIG_ORG}/{name}"))
                }
            }
        }
    }

    /// Add `user` as a collaborator on `CONFIG_ORG/repo` at `permission`.
    async fn add_collaborator(
        &self,
        repo: &str,
        user: &str,
        permission: AddCollaboratorOptionPermission,
    ) -> Result<()> {
        self.add_collaborator_to(CONFIG_ORG, repo, user, permission)
            .await
    }

    /// Add `user` as a collaborator on `owner/repo` at `permission`.
    /// Idempotent: re-adding an existing collaborator just updates its
    /// permission (forgejo answers 204 either way; a 201 from older
    /// server versions is tolerated defensively).
    async fn add_collaborator_to(
        &self,
        owner: &str,
        repo: &str,
        user: &str,
        permission: AddCollaboratorOptionPermission,
    ) -> Result<()> {
        let res = self
            .api
            .call(|api| {
                api.repo_add_collaborator(
                    owner,
                    repo,
                    user,
                    AddCollaboratorOption {
                        permission: Some(permission),
                    },
                )
                .into_future()
            })
            .await;
        match res {
            Ok(()) => {}
            Err(ForgejoError::UnexpectedStatusCode(s)) if s == StatusCode::CREATED => {}
            Err(e) => {
                return Err(e)
                    .with_context(|| format!("add collaborator {user} to {owner}/{repo}"));
            }
        }
        tracing::debug!(%owner, %repo, %user, ?permission, "swarm forge: collaborator set");
        Ok(())
    }

    /// Apply the operator merge-gate branch protection to `repo`'s
    /// default branch — only [`OPERATORS_TEAM`] members can merge, and
    /// an approving review from that team is required, so the agent (a
    /// write-level collaborator, not in the team) cannot merge its own
    /// PR. A create failure is ambiguous (already-exists vs. a silent
    /// reject that created no rule), so on any error this GETs the
    /// `main` rule and only treats it as success if the rule is
    /// actually present — a fail-open merge gate is a security bug,
    /// not a shrug.
    ///
    /// Also push-whitelists [`SWARM_CONTROLLER_FORGE_USER`] alone —
    /// [`Client::seed_agent_config`]'s initial commit is a direct push to
    /// this same protected `main`, and being a site admin does not
    /// exempt the content-edit API path from a protected branch's push
    /// check the way a real `git push` is exempted (confirmed live: an
    /// already-admin-promoted account still hit `ErrUserCannotCommit`
    /// here). `hive-c0re`'s equivalent seed avoids the whole question by
    /// seeding through an actual `git push`, which the admin bypass does
    /// cover — this daemon has no `git` shell-out, so an explicit
    /// whitelist entry is the narrower fix. The agent (a write
    /// collaborator, not on this whitelist) still cannot push directly,
    /// so this doesn't loosen the "can't merge your own PR" guarantee.
    ///
    /// Needs [`OPERATORS_TEAM`] to exist in [`CONFIG_ORG`], which
    /// [`Client::create_repo`] ensures first — if it still doesn't, this
    /// fails loudly rather than silently leaving the repo unprotected.
    async fn apply_operator_branch_protection(&self, repo: &str) -> Result<()> {
        let rule = CreateBranchProtectionOption {
            apply_to_admins: None,
            approvals_whitelist_teams: Some(vec![OPERATORS_TEAM.to_owned()]),
            approvals_whitelist_username: None,
            block_on_official_review_requests: Some(true),
            block_on_outdated_branch: None,
            block_on_rejected_reviews: None,
            branch_name: Some("main".to_owned()),
            dismiss_stale_approvals: None,
            enable_approvals_whitelist: Some(true),
            enable_merge_whitelist: Some(true),
            enable_push: Some(true),
            enable_push_whitelist: Some(true),
            enable_status_check: None,
            ignore_stale_approvals: None,
            merge_whitelist_teams: Some(vec![OPERATORS_TEAM.to_owned()]),
            merge_whitelist_usernames: None,
            protected_file_patterns: None,
            push_whitelist_deploy_keys: None,
            push_whitelist_teams: None,
            push_whitelist_usernames: Some(vec![SWARM_CONTROLLER_FORGE_USER.to_owned()]),
            require_signed_commits: None,
            required_approvals: Some(1),
            rule_name: None,
            status_check_contexts: None,
            unprotected_file_patterns: None,
        };
        let Err(create_err) = self
            .api
            .call(|api| {
                api.repo_create_branch_protection(CONFIG_ORG, repo, rule.clone())
                    .into_future()
            })
            .await
        else {
            tracing::info!(%repo, "swarm forge: applied operator branch protection");
            return Ok(());
        };
        match self
            .api
            .call(|api| {
                api.repo_get_branch_protection(CONFIG_ORG, repo, "main")
                    .into_future()
            })
            .await
        {
            Ok(_) => {
                tracing::debug!(
                    %repo, create_error = %create_err,
                    "swarm forge: operator branch protection already present"
                );
                Ok(())
            }
            Err(check_err) => anyhow::bail!(
                "branch protection for {CONFIG_ORG}/{repo} not applied: create failed \
                 ({create_err}); GET main rule failed ({check_err}), no `main` rule present"
            ),
        }
    }

    /// Create `repo` in [`CONFIG_ORG`] and apply the operator merge gate to
    /// its default branch — the whole job of the `CreateRepo` node.
    /// Branch protection is folded in here rather than a separate node: it
    /// has no independent retry value apart from the repo existing (there
    /// is nothing useful to retry "branch protection on a repo that
    /// doesn't exist yet"), unlike collaborator-add and config-seed, which
    /// are genuinely separate operations against an already-existing repo.
    /// Idempotent — safe to call again.
    ///
    /// Ensures the org and its [`OPERATORS_TEAM`] first: the periodic
    /// [`objects`] pass does too, but an agent can be created before that
    /// pass has succeeded once, and the rule below names the team.
    pub async fn create_repo(&self, repo: &str) -> Result<String> {
        self.ensure_merge_gate_prerequisites().await?;
        self.ensure_org_repo(repo).await?;
        self.apply_operator_branch_protection(repo).await?;
        tracing::info!(%repo, "swarm forge: created repo in {CONFIG_ORG} with operator merge gate");
        Ok(format!("{CONFIG_ORG}/{repo}"))
    }

    /// Add `agent` as a **write** collaborator on `repo` (can push + open
    /// PRs, cannot bypass branch protection) — the whole job of the
    /// `AddRepoMember` node. Idempotent.
    pub async fn add_repo_member(&self, repo: &str, agent: &str) -> Result<()> {
        self.add_collaborator(repo, agent, AddCollaboratorOptionPermission::Write)
            .await
    }

    /// Ensure `agent` exists as a Forgejo user account that cannot create
    /// repos — the whole job of the `CreateForgeUser` node, and the fix for the "user does not
    /// exist" failure `AddRepoMember` hit before this node existed: adding
    /// a nonexistent user as a collaborator is a Forgejo validation error,
    /// not an idempotent no-op, so something has to create the account
    /// first. Mirrors `hive-c0re::forge::users::ensure_user_exists`'s
    /// intent (an agent's Forgejo identity is provisioned once, up front,
    /// authenticates by token thereafter, and its password is never read)
    /// but not its mechanism: that function shells out to the local
    /// `forgejo admin` CLI, which assumes co-location with the forge host.
    /// This daemon has no such assumption — like every other call in this
    /// file, it only ever talks to the forge over HTTP — so this goes through
    /// `admin_create_user` instead. The password is a throwaway: 32 random
    /// bytes, generated once, never persisted or read again (see
    /// [`crate::webhook::generate_hex_secret`], reused here for the same
    /// reason a webhook secret and this password are both "32 random bytes
    /// nothing reads back").
    ///
    /// Idempotent, same as [`Self::ensure_org_repo`]: 409 folds to success
    /// unconditionally, 422 only after a follow-up `user_get` confirms the
    /// user (see [`folds_into_success`] — not every 422 means it exists).
    ///
    /// Then locks the account out of creating repos
    /// ([`Self::disable_repo_creation`]), whether it was just created or
    /// already existed. Nothing else does that for a swarm-created agent, so
    /// a failure fails the node. Leaves an existing account's email as it
    /// finds it.
    pub async fn ensure_agent_user(&self, agent: &str) -> Result<()> {
        self.create_agent_user(agent).await?;
        self.disable_repo_creation(agent).await
    }

    /// Set `max_repo_creation = 0` on `agent`'s account. Agents create repos
    /// through the hive, never with their own token — a write-scoped token
    /// can otherwise create and own a repo and self-merge in it, bypassing
    /// the operator-only-merge policy. Existing repos are untouched; push,
    /// PRs and clone keep working.
    ///
    /// The PATCH itself is unconditional rather than verify-then-patch:
    /// Forgejo's API never returns `max_repo_creation` (it is a field of
    /// `EditUserOption` only, not of `User`), so there is no current value
    /// to compare against, and setting 0 on an account already at 0 is a
    /// no-op. But `create_agent_user`'s idempotent fold only proves *some*
    /// account named `agent` exists — a 409 folds with no read at all, and
    /// a 422's follow-up `user_get` only confirms the username, not that
    /// swarm-controller created it — so before that PATCH this reads the
    /// account once and checks the one thing that *is* comparable: the
    /// email. A match against either [`agent_email`] (what
    /// `create_agent_user` itself sets) or [`legacy_agent_email`] (what
    /// older hive-created agent accounts carry) counts as this agent's
    /// account. A mismatch against both means the
    /// name collided with an account this node didn't create, so it fails
    /// the node rather than sending the PATCH (or silently skipping) to an
    /// account nobody verified is agent-controlled.
    async fn disable_repo_creation(&self, agent: &str) -> Result<()> {
        let user = self
            .api
            .call(|api| api.user_get(agent).into_future())
            .await
            .with_context(|| format!("read forge user {agent} before locking it down"))?;
        let expected = agent_email(agent);
        let legacy = legacy_agent_email(agent);
        if !matches!(user.email.as_deref(), Some(e) if e == expected || e == legacy) {
            anyhow::bail!(
                "forge account `{agent}` exists but is not this agent's (email {:?}, expected \
                 `{expected}` or the legacy `{legacy}`); refusing to lock it down",
                user.email
            );
        }
        self.api
            .call(|api| {
                api.admin_edit_user(agent, repo_creation_lockdown(agent))
                    .into_future()
            })
            .await
            .with_context(|| {
                format!("disable repo creation for forge user {agent} (max_repo_creation = 0)")
            })?;
        tracing::info!(%agent, "swarm forge: disabled repo creation (max_repo_creation = 0)");
        Ok(())
    }

    /// Create `agent`'s account. Idempotent: an existing user (409/422) is
    /// folded into success, same as [`Self::ensure_org_repo`].
    async fn create_agent_user(&self, agent: &str) -> Result<()> {
        self.create_user(agent, &agent_email(agent)).await
    }

    /// Create the account `login` with `email` and a throwaway password.
    /// Idempotent like [`Self::create_agent_user`], and like it proves only
    /// that *some* account named `login` exists: the caller checks the email.
    async fn create_user(&self, login: &str, email: &str) -> Result<()> {
        let password = crate::webhook::generate_hex_secret()
            .context("generating a throwaway password for a forge account")?;
        let res = self
            .api
            .call(|api| {
                api.admin_create_user(CreateUserOption {
                    created_at: None,
                    email: email.to_owned(),
                    full_name: None,
                    login_name: None,
                    must_change_password: Some(false),
                    password: Some(password.clone()),
                    restricted: None,
                    send_notify: None,
                    source_id: None,
                    username: login.to_owned(),
                    visibility: None,
                })
                .into_future()
            })
            .await;
        match res {
            Ok(_) => {
                tracing::info!(%login, "swarm forge: created forge user");
                Ok(())
            }
            Err(e) => {
                let existing = is_ambiguous_validation_failure(&e)
                    && self
                        .api
                        .call(|api| api.user_get(login).into_future())
                        .await
                        .is_ok();
                if folds_into_success(&e, existing) {
                    tracing::debug!(%login, "swarm forge: forge user already exists");
                    Ok(())
                } else {
                    tracing::error!(%login, error = %e, "swarm forge: forge refused to create user");
                    Err(e).with_context(|| format!("create forge user {login}"))
                }
            }
        }
    }

    /// Seed `repo` with the two files every agent config repo needs:
    /// `agent.nix` (the agent's own module) and `flake.nix` (the
    /// boilerplate that lets the meta flake import this repo as a flake
    /// input) — committed here in one atomic `repo_change_files` call
    /// instead of a local `git commit` (this process has no working tree to
    /// commit from — it only ever talks to the forge over HTTP). The whole
    /// job of the `InitAgentConfigRepo` node.
    ///
    /// **This is the agent's config, not a copy of it.** A hive told to
    /// deploy the agent clones this repo
    /// (`hive-c0re::forge::clone_config_into_proposed`) rather than writing
    /// the same files again locally; the byte-identical template in
    /// `hive-c0re::lifecycle::setup::seed_template` is only reached when no
    /// repo exists here to clone.
    ///
    /// Idempotent by construction rather than by catching a conflict: an
    /// unconditional `repo_change_files` against an already-seeded repo
    /// would overwrite `agent.nix` with the seed content again, clobbering
    /// any edits merged since — so this checks for `agent.nix`'s presence
    /// first and treats it as "already seeded, nothing to do" rather than
    /// re-committing over it.
    pub async fn seed_agent_config(&self, repo: &str, agent: &str) -> Result<()> {
        if self.file_exists(repo, "agent.nix").await? {
            tracing::debug!(%repo, "swarm forge: config already seeded");
            return Ok(());
        }
        let files = vec![
            ChangeFileOperation {
                content: Some(base64_encode(&initial_agent_nix(agent))),
                from_path: None,
                operation: ChangeFileOperationOperation::Create,
                path: "agent.nix".to_owned(),
                sha: None,
            },
            ChangeFileOperation {
                content: Some(base64_encode(initial_flake_nix())),
                from_path: None,
                operation: ChangeFileOperationOperation::Create,
                path: "flake.nix".to_owned(),
                sha: None,
            },
        ];
        self.api
            .call(|api| {
                api.repo_change_files(
                    CONFIG_ORG,
                    repo,
                    ChangeFilesOptions {
                        author: None,
                        branch: None,
                        committer: None,
                        dates: None,
                        files: files.clone(),
                        force_overwrite_new_branch: None,
                        message: Some("swarm-controller: seed agent config".to_owned()),
                        new_branch: None,
                        signoff: None,
                    },
                )
                .into_future()
            })
            .await
            .with_context(|| format!("seed config files in {CONFIG_ORG}/{repo}"))?;
        tracing::info!(%repo, %agent, "swarm forge: seeded agent.nix + flake.nix");
        Ok(())
    }

    /// Whether `path` exists on `repo`'s default branch — the idempotency
    /// check [`Self::seed_agent_config`] uses instead of catching a
    /// create-conflict, since `repo_change_files` has no
    /// already-exists-is-fine status the way `create_org_repo` does.
    ///
    /// `create_repo` makes `repo` with `auto_init: false` (see
    /// [`Self::repo_option`]) — no commits, no `main` ref, nothing a
    /// contents lookup can resolve. Forgejo doesn't 404 that the way it
    /// does a missing path on a real ref: it returns `200` with a bare
    /// JSON `[]`, which fails to deserialize into the single-object
    /// `ContentsResponse` the crate expects for a file path ("invalid
    /// length 0, expected struct `ContentsResponse` with 16 elements" —
    /// serde reading the 0-element array as a positional struct). So a
    /// freshly created, not-yet-seeded repo trips this as an opaque
    /// deserialize error, not the `NOT_FOUND` branch below — confirmed
    /// against a live re-test of the swarm agent-create flow tracked on
    /// the forge. Short-circuit on `empty` first: an empty repo
    /// obviously has no `agent.nix`, and skipping
    /// the contents call for it sidesteps the malformed-response shape
    /// entirely rather than trying to special-case parsing it.
    async fn file_exists(&self, repo: &str, path: &str) -> Result<bool> {
        let is_empty = self
            .api
            .call(|api| api.repo_get(CONFIG_ORG, repo).into_future())
            .await
            .with_context(|| format!("get repo {CONFIG_ORG}/{repo}"))?
            .empty
            .unwrap_or(false);
        if is_empty {
            return Ok(false);
        }
        match self
            .api
            .call(|api| {
                api.repo_get_contents(CONFIG_ORG, repo, path, RepoGetContentsQuery::default())
                    .into_future()
            })
            .await
        {
            Ok(_) => Ok(true),
            Err(ForgejoError::UnexpectedStatusCode(s)) if s == StatusCode::NOT_FOUND => Ok(false),
            Err(e) => Err(e).with_context(|| format!("check for {path} in {CONFIG_ORG}/{repo}")),
        }
    }

    /// Every agent in [`CONFIG_ORG`] with an open config PR, keyed by agent
    /// name: list repos in the org, then open PRs per repo. It exists so `GET
    /// /api/agents/{name}/config-pr` has something to answer from,
    /// independent of any one hive being up.
    ///
    /// A single repo's list failing does not fail the whole scan — logged and
    /// skipped, so one flaky repo can't blank out every other agent's status.
    pub async fn list_open_config_prs(
        &self,
    ) -> Result<std::collections::HashMap<String, ConfigPrStatus>> {
        let repos = self
            .api
            .call(|api| Box::pin(api.org_list_repos(CONFIG_ORG).all()))
            .await
            .with_context(|| format!("list repos in {CONFIG_ORG}"))?;

        let mut out = std::collections::HashMap::new();
        for repo in repos {
            let Some(agent) = repo.name.as_deref() else {
                continue;
            };
            let query = RepoListPullRequestsQuery {
                state: Some(RepoListPullRequestsQueryState::Open),
                sort: None,
                milestone: None,
                labels: None,
                poster: None,
                base: None,
                head: None,
            };
            let prs = match self
                .api
                .call(|api| {
                    Box::pin(
                        api.repo_list_pull_requests(CONFIG_ORG, agent, query.clone())
                            .all(),
                    )
                })
                .await
            {
                Ok(prs) => prs,
                Err(e) => {
                    tracing::warn!(%agent, error = %e, "swarm forge: listing config PRs failed, skipping repo");
                    continue;
                }
            };
            // Only the first open PR matters for the panel — a config repo
            // is meant to carry at most one live proposal at a time.
            if let Some(pr) = prs.into_iter().next() {
                let Some(pr_number) = pr.number.and_then(|n| u64::try_from(n).ok()) else {
                    continue;
                };
                out.insert(
                    agent.to_owned(),
                    ConfigPrStatus {
                        pr_number,
                        html_url: pr.html_url.map(|u| u.to_string()),
                    },
                );
            }
        }
        Ok(out)
    }

    /// Pages a search-shaped result (`data: Option<Vec<T>>`, no total-count
    /// header) to exhaustion via `fetch`, which must issue one request per
    /// `page` (1-based) at `page_size` and return its `data`. `repo_search`'s
    /// `SearchResults` is this shape, not the `(Headers, Vec<T>)` shape the
    /// `.all()` helper on `Request` requires, so it can't use that helper —
    /// this is its manual equivalent. Stops at the first page shorter than
    /// `page_size` (forgejo's own last-page signal); a page exactly
    /// `page_size` long always triggers one more fetch to confirm there
    /// isn't a further page. Errors instead of returning a partial result if
    /// `max_pages` is reached without a short page, so a misbehaving or
    /// unexpectedly huge instance fails loudly rather than looping forever
    /// or silently truncating again.
    async fn page_search_results<T, F, Fut>(
        page_size: u32,
        max_pages: u32,
        mut fetch: F,
    ) -> Result<Vec<T>>
    where
        F: FnMut(u32) -> Fut,
        Fut: std::future::Future<Output = Result<Option<Vec<T>>>>,
    {
        let mut out = Vec::new();
        for page in 1..=max_pages {
            let rows = fetch(page).await?.unwrap_or_default();
            let got = rows.len();
            out.extend(rows);
            if got < page_size as usize {
                return Ok(out);
            }
        }
        anyhow::bail!(
            "search paging hit the {max_pages}-page bound (page size {page_size}) without a \
             short page — the forge may hold more results than this scan can safely enumerate"
        )
    }

    /// Repos with at least one open issue, as `owner/name` full names —
    /// the data source for swarm-ui's repo-filter dropdown, which lists
    /// only repos that currently have open issues rather than every repo
    /// the forge hosts. Filtered on `Repository::open_issues_count` from
    /// the search result itself rather than a follow-up
    /// `issue_list_issues` call per repo — forgejo's own repo summary
    /// already carries that count, so this stays one request per page
    /// regardless of how many repos exist.
    pub async fn list_repos_with_open_issues(&self) -> Result<Vec<String>> {
        Ok(Self::full_names_where(self.search_repos().await?, |r| {
            r.open_issues_count.unwrap_or(0) > 0
        }))
    }

    /// Repos with at least one open pull request, as `owner/name` full
    /// names — the data source for the review try-out's repo dropdown. Same
    /// one-request-per-page shape as [`Self::list_repos_with_open_issues`],
    /// filtered on `Repository::open_pr_counter`.
    pub async fn list_repos_with_open_pulls(&self) -> Result<Vec<String>> {
        Ok(Self::full_names_where(self.search_repos().await?, |r| {
            r.open_pr_counter.unwrap_or(0) > 0
        }))
    }

    /// The `owner/name` of each repo in `repos` that `keep` accepts.
    fn full_names_where(
        repos: Vec<forgejo_api::structs::Repository>,
        keep: impl Fn(&forgejo_api::structs::Repository) -> bool,
    ) -> Vec<String> {
        repos
            .into_iter()
            .filter(|r| keep(r))
            .filter_map(|r| r.full_name)
            .collect()
    }

    /// Every repo the forge hosts, from the search endpoint.
    async fn search_repos(&self) -> Result<Vec<forgejo_api::structs::Repository>> {
        const PAGE_SIZE: u32 = 50;
        // 40 pages at 50/page is 2000 repos — comfortably past any repo
        // count this instance has today; see `page_search_results` for why
        // hitting it is an error rather than a truncated result.
        const MAX_PAGES: u32 = 40;
        let repos = Self::page_search_results(PAGE_SIZE, MAX_PAGES, |page| async move {
            Ok(self
                .api
                .call(|api| {
                    api.repo_search(RepoSearchQuery::default())
                        .page(page)
                        .page_size(PAGE_SIZE)
                        .into_future()
                })
                .await
                .context("search repos")?
                .data)
        })
        .await?;
        Ok(repos)
    }

    /// The open pull requests of `owner/repo`, newest first — the data
    /// source for the review try-out's PR dropdown. A Forgejo 404 stays in
    /// the error chain for [`is_repo_not_found`] to find.
    pub async fn list_open_pulls(&self, owner: &str, repo: &str) -> Result<Vec<OpenPull>> {
        let query = RepoListPullRequestsQuery {
            state: Some(RepoListPullRequestsQueryState::Open),
            sort: None,
            milestone: None,
            labels: None,
            poster: None,
            base: None,
            head: None,
        };
        let pulls = self
            .api
            .call(|api| {
                Box::pin(
                    api.repo_list_pull_requests(owner, repo, query.clone())
                        .all(),
                )
            })
            .await
            .with_context(|| format!("list open pull requests in {owner}/{repo}"))?;
        Ok(open_pulls(pulls))
    }

    /// Every open issue in `owner/repo`, each with its dependency-derived
    /// `blocked` / `depended_on_by_count` pre-resolved server-side — the
    /// whole job behind `GET /api/repos/{org}/{repo}/issue-report`. Doing
    /// this resolution client-side (one dependency lookup per issue, on
    /// every page load) doesn't scale, which is the reason this endpoint
    /// exists at all rather than swarm-ui resolving it itself per row.
    ///
    /// Forgejo has no bulk or reverse dependency query — the same
    /// limitation `hive-forge issue dependency` hits — so this is still one
    /// `issue_list_issue_dependencies` call per open issue, just run with
    /// bounded concurrency (`CONCURRENCY`) rather than one-at-a-time or all
    /// at once, so a repo with hundreds of open issues doesn't serialize on
    /// round-trips or fire them all in one burst.
    ///
    /// `depended_on_by_count` is a reverse index built off that SAME
    /// resolve pass, not a second query: every open issue's own dependency
    /// list is already in hand to compute `blocked`, so counting how often
    /// each open issue's number appears across all of them is pure
    /// in-memory aggregation. A dependency on a closed issue, or one
    /// outside this open-issue set, can never accrue a count here — it
    /// isn't a row in the report either, so "how many rows depend on this
    /// row" would have nothing to point at.
    ///
    /// `transitively_blocks_count` walks the same `reverse_adj` (blocker
    /// number → issues naming it as a dependency) from each issue — see
    /// [`transitive_reach`] for how the walk itself is made
    /// diamond/cycle-safe. `depended_on_by_count` is `reverse_adj[n].len()`;
    /// this is everything reachable from `n`, not just the direct edge.
    pub async fn issue_report(&self, owner: &str, repo: &str) -> Result<Vec<IssueReportRow>> {
        const CONCURRENCY: usize = 8;

        let issues = self
            .api
            .call(|api| {
                Box::pin(
                    api.issue_list_issues(
                        owner,
                        repo,
                        IssueListIssuesQuery {
                            state: Some(IssueListIssuesQueryState::Open),
                            r#type: Some(IssueListIssuesQueryType::Issues),
                            ..Default::default()
                        },
                    )
                    .all(),
                )
            })
            .await
            .with_context(|| format!("list open issues in {owner}/{repo}"))?;

        // Numbers collected into an owned `Vec` first, rather than handing
        // `stream::iter` a lazy `filter_map` over borrowed `issues`
        // directly — the latter shape trips a "closure implementation of
        // `FnOnce` is not general enough" HRTB inference error where this
        // function is used as an axum handler (the `routes!` macro checks
        // `Handler` for arbitrary request lifetimes), even though the
        // closure itself is unremarkable.
        let numbers: Vec<i64> = issues.iter().filter_map(|i| i.number).collect();
        let deps: HashMap<i64, Vec<forgejo_api::structs::Issue>> =
            futures_util::stream::iter(numbers)
                .map(|number| async move {
                    let blockers = self
                        .api
                        .call(|api| {
                            api.issue_list_issue_dependencies(owner, repo, number)
                                .into_future()
                        })
                        .await
                        .with_context(|| format!("list dependencies of {owner}/{repo}#{number}"))?;
                    Ok::<_, anyhow::Error>((number, blockers))
                })
                .buffer_unordered(CONCURRENCY)
                .try_collect()
                .await?;

        let open_numbers: HashSet<i64> = issues.iter().filter_map(|i| i.number).collect();
        let mut reverse_adj: HashMap<i64, Vec<i64>> = HashMap::new();
        for (&number, blockers) in &deps {
            for blocker in blockers {
                if let Some(n) = blocker.number
                    && open_numbers.contains(&n)
                {
                    reverse_adj.entry(n).or_default().push(number);
                }
            }
        }

        let repo_full_name = format!("{owner}/{repo}");
        Ok(issues
            .into_iter()
            .filter_map(|issue| {
                let number = issue.number?;
                let blocked = deps
                    .get(&number)
                    .is_some_and(|b| b.iter().any(|d| d.state == Some(StateType::Open)));
                Some(IssueReportRow {
                    repo: repo_full_name.clone(),
                    number,
                    title: issue.title.unwrap_or_default(),
                    labels: issue
                        .labels
                        .unwrap_or_default()
                        .into_iter()
                        .filter_map(|l| l.name)
                        .collect(),
                    assignees: issue
                        .assignees
                        .unwrap_or_default()
                        .into_iter()
                        .filter_map(|a| a.login)
                        .collect(),
                    html_url: issue.html_url.map(|u| u.to_string()),
                    blocked,
                    depended_on_by_count: u32::try_from(
                        reverse_adj.get(&number).map_or(0, Vec::len),
                    )
                    .unwrap_or(u32::MAX),
                    transitively_blocks_count: u32::try_from(transitive_reach(
                        number,
                        &reverse_adj,
                    ))
                    .unwrap_or(u32::MAX),
                })
            })
            .collect())
    }

    /// [`Self::issue_report`], fanned out over every repo
    /// [`Self::list_repos_with_open_issues`] finds — the default,
    /// no-repo-filter view. Each repo's report is resolved independently and
    /// concurrently (same `CONCURRENCY` bound as the per-issue dependency
    /// resolution inside `issue_report`, since this is the same
    /// "N independent forge round-trips" shape one level up), then
    /// flattened into one row set — `IssueReportRow::repo` is what lets a
    /// caller tell which repo a given row came from.
    ///
    /// One repo's report failing does not fail the whole scan — logged and
    /// skipped, same best-effort shape [`Self::list_open_config_prs`] uses:
    /// a flaky repo should not blank out every other repo's rows.
    pub async fn issue_report_all(&self) -> Result<Vec<IssueReportRow>> {
        const CONCURRENCY: usize = 4;

        let repos = self.list_repos_with_open_issues().await?;
        let rows: Vec<Vec<IssueReportRow>> = futures_util::stream::iter(repos)
            .map(|full_name| async move {
                let Some((owner, repo)) = full_name.split_once('/') else {
                    tracing::warn!(%full_name, "issue_report_all: repo full_name has no '/'");
                    return Vec::new();
                };
                match self.issue_report(owner, repo).await {
                    Ok(rows) => rows,
                    Err(e) => {
                        tracing::warn!(
                            error = %format!("{e:#}"),
                            %full_name,
                            "issue_report_all: one repo's report failed, skipping it"
                        );
                        Vec::new()
                    }
                }
            })
            .buffer_unordered(CONCURRENCY)
            .collect()
            .await;
        Ok(rows.into_iter().flatten().collect())
    }

    /// Set `status` on `sha` of `owner/repo`.
    ///
    /// # Errors
    /// The forge refused the status.
    pub async fn create_commit_status(
        &self,
        owner: &str,
        repo: &str,
        sha: &str,
        status: CommitStatus<'_>,
    ) -> Result<()> {
        let context = status.context;
        let status = CreateStatusOption {
            context: Some(context.to_owned()),
            description: Some(status_description(status.description)),
            state: Some(status.state),
            target_url: status.target_url.map(str::to_owned),
        };
        self.api
            .call(|api| {
                api.repo_create_status(owner, repo, sha, status.clone())
                    .into_future()
            })
            .await
            .with_context(|| format!("set the {context} status on {sha} of {owner}/{repo}"))?;
        Ok(())
    }

    /// What the review classifier reads of `owner/repo`'s PR `number`: its
    /// head commit, its commits oldest first with their full messages, and
    /// its unified diff from the merge base ([`merge_base_of`]): [`pull_diff::wide`]'s, or the
    /// forge's `.diff` when the wide one is not read, cannot be read, or is
    /// too large.
    ///
    /// # Errors
    /// A forge failure, with the [`ForgejoError`] underneath; a 404 on the
    /// first call means the repo or the PR does not exist
    /// ([`is_repo_not_found`]).
    pub async fn pull_for_review(
        &self,
        owner: &str,
        repo: &str,
        number: i64,
    ) -> Result<PullForReview> {
        let pull = self
            .api
            .call(|api| api.repo_get_pull_request(owner, repo, number).into_future())
            .await
            .with_context(|| format!("read PR {number} of {owner}/{repo}"))?;
        let changed_files = pull.changed_files;
        let head_sha = pull
            .head
            .and_then(|h| h.sha)
            .with_context(|| format!("PR {number} of {owner}/{repo} names no head commit"))?;
        let listed_commits = self
            .api
            .call(|api| {
                Box::pin(
                    api.repo_get_pull_request_commits(
                        owner,
                        repo,
                        number,
                        RepoGetPullRequestCommitsQuery {
                            verification: Some(false),
                            files: Some(false),
                        },
                    )
                    .all(),
                )
            })
            .await
            .with_context(|| format!("list the commits of PR {number} of {owner}/{repo}"))?;
        let merge_base = merge_base_of(&listed_commits);
        let commits = listed_commits
            .into_iter()
            .map(|c| Commit {
                sha: c.sha.unwrap_or_default(),
                message: c.commit.and_then(|m| m.message).unwrap_or_default(),
            })
            .collect();
        let count = changed_files.map_or(0, |n| usize::try_from(n).unwrap_or(usize::MAX));
        let wide = match (pull_diff::too_many_files(count), merge_base) {
            (Some(reason), _) => Err(reason),
            (None, None) => Err("the PR's commits do not name a single merge base".to_owned()),
            (None, Some(base)) => {
                match self
                    .changed_files(owner, repo, number, &base, &head_sha)
                    .await
                {
                    Ok(files) => tokio::task::spawn_blocking(move || pull_diff::wide(&files))
                        .await
                        .unwrap_or_else(|e| Err(format!("building the wide diff failed: {e}"))),
                    Err(e) => Err(format!("reading the files failed: {e:#}")),
                }
            }
        };
        let (diff, narrowed) = match wide {
            Ok(diff) => (diff, None),
            Err(reason) => {
                tracing::info!(
                    %owner, %repo, pr = number, %reason,
                    "the review classifier reads the forge's 3-line diff"
                );
                (self.forge_diff(owner, repo, number).await?, Some(reason))
            }
        };
        Ok(PullForReview {
            head_sha,
            commits,
            diff,
            narrowed,
        })
    }

    /// The forge's own `.diff` of PR `number`, with 3 lines of context.
    async fn forge_diff(&self, owner: &str, repo: &str, number: i64) -> Result<String> {
        self.api
            .call(|api| {
                api.repo_download_pull_diff_or_patch(
                    owner,
                    repo,
                    number,
                    "diff",
                    RepoDownloadPullDiffOrPatchQuery::default(),
                )
                .into_future()
            })
            .await
            .with_context(|| format!("read the diff of PR {number} of {owner}/{repo}"))
    }

    /// Each file PR `number` changes, with its bytes at `base` and at `head`.
    /// A renamed file is read at its old path on the `base` side.
    async fn changed_files(
        &self,
        owner: &str,
        repo: &str,
        number: i64,
        base: &str,
        head: &str,
    ) -> Result<Vec<pull_diff::FileChange>> {
        let listed = self
            .api
            .call(|api| {
                Box::pin(
                    api.repo_get_pull_request_files(
                        owner,
                        repo,
                        number,
                        RepoGetPullRequestFilesQuery::default(),
                    )
                    .all(),
                )
            })
            .await
            .with_context(|| format!("list the files of PR {number} of {owner}/{repo}"))?;
        let mut trees = HashMap::new();
        let mut files = Vec::with_capacity(listed.len());
        for file in listed {
            let new_path = file
                .filename
                .with_context(|| format!("a file of PR {number} of {owner}/{repo} has no name"))?;
            let old_path = file.previous_filename.unwrap_or_else(|| new_path.clone());
            let old = self
                .file_at(owner, repo, base, &old_path, &mut trees)
                .await?;
            let new = self
                .file_at(owner, repo, head, &new_path, &mut trees)
                .await?;
            files.push(pull_diff::FileChange {
                old_path,
                new_path,
                old: old.map(|(_, bytes)| bytes),
                new: new.map(|(_, bytes)| bytes),
            });
        }
        Ok(files)
    }

    /// The `paths` of `owner/repo` that exist on its default branch, all
    /// read at the branch's head commit. `None` when the default branch has
    /// no commit: an empty repo, or a default branch that doesn't exist.
    ///
    /// The files come through the git tree and blob API, one path segment
    /// per request: forgejo-api escapes the `/` of a nested path in its
    /// file endpoints.
    ///
    /// # Errors
    /// A forge failure, or a tree the forge truncated. A [`ForgejoError`] 404
    /// underneath comes only from the first call, and means the repo does
    /// not exist ([`is_repo_not_found`]).
    pub async fn default_branch_files(
        &self,
        owner: &str,
        repo: &str,
        paths: &[&str],
    ) -> Result<Option<DefaultBranchFiles>> {
        let info = self
            .api
            .call(|api| api.repo_get(owner, repo).into_future())
            .await
            .with_context(|| format!("get repo {owner}/{repo}"))?;
        if info.empty.unwrap_or(false) {
            return Ok(None);
        }
        let branch = info
            .default_branch
            .with_context(|| format!("{owner}/{repo} names no default branch"))?;
        let head = match self
            .api
            .call(|api| api.repo_get_branch(owner, repo, &branch).into_future())
            .await
        {
            Ok(head) => head,
            Err(e) if is_repo_not_found(&e) => return Ok(None),
            Err(e) => {
                return Err(e).with_context(|| format!("read branch {branch} of {owner}/{repo}"));
            }
        };
        let commit = head
            .commit
            .and_then(|c| c.id)
            .with_context(|| format!("branch {branch} of {owner}/{repo} names no commit"))?;
        let mut trees = HashMap::new();
        let mut files = Vec::new();
        for path in paths {
            if let Some((sha, bytes)) = self.file_at(owner, repo, &commit, path, &mut trees).await?
            {
                files.push(BranchFile {
                    path: (*path).to_owned(),
                    sha,
                    bytes,
                });
            }
        }
        Ok(Some(DefaultBranchFiles {
            branch,
            commit,
            web_url: info.html_url,
            files,
        }))
    }

    /// The blob sha and bytes of the file at `path` in `owner/repo` at
    /// `commit`, or `None` when no blob is there. `trees` holds the tree
    /// listings already read, by sha.
    async fn file_at(
        &self,
        owner: &str,
        repo: &str,
        commit: &str,
        path: &str,
        trees: &mut HashMap<String, Vec<GitEntry>>,
    ) -> Result<Option<(String, Vec<u8>)>> {
        let mut tree = commit.to_owned();
        let mut segments = path.split('/').peekable();
        while let Some(name) = segments.next() {
            let kind = if segments.peek().is_some() {
                "tree"
            } else {
                "blob"
            };
            let entries = self.tree(owner, repo, &tree, trees).await?;
            let Some(sha) = entries
                .iter()
                .find(|e| e.path.as_deref() == Some(name) && e.r#type.as_deref() == Some(kind))
                .and_then(|e| e.sha.clone())
            else {
                return Ok(None);
            };
            if kind == "blob" {
                return self
                    .blob(owner, repo, &sha)
                    .await
                    .map(|bytes| Some((sha, bytes)));
            }
            tree = sha;
        }
        Ok(None)
    }

    /// The entries of tree `sha` (or of a commit's root tree), read once
    /// into `trees`.
    async fn tree<'t>(
        &self,
        owner: &str,
        repo: &str,
        sha: &str,
        trees: &'t mut HashMap<String, Vec<GitEntry>>,
    ) -> Result<&'t [GitEntry]> {
        if !trees.contains_key(sha) {
            let listing = self
                .api
                .call(|api| {
                    api.get_tree(owner, repo, sha, GetTreeQuery::default())
                        .into_future()
                })
                .await
                .map_err(|e| anyhow::anyhow!("read tree {sha} of {owner}/{repo}: {e}"))?;
            if listing.truncated.unwrap_or(false) {
                anyhow::bail!(
                    "the forge truncated tree {sha} of {owner}/{repo}, so a missing entry can't \
                     be told from one past the cut"
                );
            }
            trees.insert(sha.to_owned(), listing.tree.unwrap_or_default());
        }
        Ok(&trees[sha])
    }

    /// The bytes of blob `sha`.
    async fn blob(&self, owner: &str, repo: &str, sha: &str) -> Result<Vec<u8>> {
        use base64::Engine as _;

        let blob = self
            .api
            .call(|api| api.get_blob(owner, repo, sha).into_future())
            .await
            .map_err(|e| anyhow::anyhow!("read blob {sha} of {owner}/{repo}: {e}"))?;
        anyhow::ensure!(
            blob.encoding.as_deref() == Some("base64"),
            "blob {sha} of {owner}/{repo} is not base64 but {:?}",
            blob.encoding
        );
        let content: String = blob
            .content
            .unwrap_or_default()
            .split_ascii_whitespace()
            .collect();
        base64::engine::general_purpose::STANDARD
            .decode(content)
            .with_context(|| format!("decode blob {sha} of {owner}/{repo}"))
    }

    /// Register the swarm-wide hooks against this controller, so a real
    /// forge event reaches [`crate::webhook`] instead of the endpoint only
    /// being reachable by hand.
    ///
    /// Adds its own hooks only. A hive's `/webhook/config-pr` hook left on
    /// the `agent-configs` org by an older release stays registered and
    /// delivers to a route no hive serves.
    ///
    /// Idempotent: an existing hook with the same `target_url` is left
    /// alone, so this is safe on every boot.
    ///
    /// # Errors
    ///
    /// Returns the first failure. A listing failure is not fatal — it falls
    /// through to the create attempt, which is idempotent server-side by way
    /// of the already-exists fold. A forge that does not answer within
    /// [`FORGE_DEADLINE`] is a failure too.
    pub async fn ensure_swarm_webhooks(&self, public_base: &str, secret: &str) -> Result<()> {
        within_forge_deadline(
            "register swarm webhooks",
            self.ensure_all_swarm_hooks(public_base, secret),
        )
        .await
        .unwrap_or_else(|| {
            anyhow::bail!(
                "the forge did not answer within {}s",
                FORGE_DEADLINE.as_secs()
            )
        })
    }

    async fn ensure_all_swarm_hooks(&self, public_base: &str, secret: &str) -> Result<()> {
        for kind in DeliveryKind::ALL {
            let target_url = kind.target_url(public_base);
            let (event, scope) = match kind {
                DeliveryKind::Knowledge => (
                    "push",
                    HookScope::Repo {
                        org: KNOWLEDGE_ORG,
                        repo: KNOWLEDGE_REPO,
                    },
                ),
                DeliveryKind::ConfigPr => ("pull_request", HookScope::Org { org: CONFIG_ORG }),
                // Instance-wide: commit/push activity is not scoped to one
                // org (unlike the two above) — it needs observing across
                // every repo the forge hosts, per the "no metric exists for
                // commits or pushes" tracker issue. Forgejo's
                // "global (system) webhook" (`admin_create_hook`) is exactly
                // this — the only scope in this API that fires for a repo
                // in an org created after this hook was registered, with no
                // per-org registration to keep in sync as orgs come and go.
                DeliveryKind::VcsActivity => ("push", HookScope::Instance),
                // Instance-wide too: any repo whose admin gave the
                // classifier's account access can request its review.
                DeliveryKind::ReviewRequest => ("pull_request_review_request", HookScope::Instance),
            };
            self.ensure_hook(&scope, &target_url, event, secret)
                .await
                .with_context(|| format!("register swarm webhook {target_url}"))?;
        }
        Ok(())
    }

    /// One hook, idempotently. Split out of [`Self::ensure_swarm_webhooks`]
    /// so the org-scoped and repo-scoped calls do not each grow their own
    /// copy of the list-then-create logic.
    async fn ensure_hook(
        &self,
        scope: &HookScope<'_>,
        target_url: &str,
        event: &str,
        secret: &str,
    ) -> Result<()> {
        // ⚠️ Both non-matching arms log at a level the journal keeps, because
        // this idempotency check fails *silently*: when the list step does
        // not recognise a hook that is already there, the create below runs
        // again, and that stays harmless only while the forge rejects the
        // duplicate. A forge that accepts it leaves one extra hook per
        // process start, each delivering on every event — and the create
        // arm's line reads exactly like a first, correct registration.
        //
        // The two ways to reach a create have different causes, so they are
        // separated: `listed=0` means the create is not landing in the set
        // the list reads (a bucket or permission problem — see
        // [`HookScope::extra_create_config`], which is where the instance
        // scope's own version of that went wrong), a non-zero count with no
        // match means the recorded url is not the one compared.
        match scope.list_hook_urls(&self.api).await {
            Ok(urls) => {
                if urls.iter().any(|u| u == target_url) {
                    tracing::debug!(%target_url, "swarm forge: webhook already registered");
                    return Ok(());
                }
                tracing::info!(
                    %target_url,
                    listed = urls.len(),
                    "swarm forge: no listed hook matches; registering"
                );
            }
            Err(e) => {
                // Best-effort: a forge that
                // cannot be listed may still accept a create, and a
                // duplicate create is folded into success below. `warn!`
                // because that fold is an assumption about the forge, not a
                // guarantee — see the note above.
                tracing::warn!(error = %e, %target_url, "swarm forge: listing hooks failed; registering blind");
            }
        }

        let mut additional = BTreeMap::new();
        additional.insert("secret".to_owned(), secret.to_owned());
        for (key, value) in scope.extra_create_config() {
            additional.insert((*key).to_owned(), (*value).to_owned());
        }
        let hook = CreateHookOption {
            active: Some(true),
            authorization_header: None,
            branch_filter: None,
            config: CreateHookOptionConfig {
                content_type: "json".to_owned(),
                url: url::Url::parse(target_url)
                    .with_context(|| format!("parse webhook target url {target_url}"))?,
                additional,
            },
            events: Some(vec![event.to_owned()]),
            r#type: CreateHookOptionType::Forgejo,
        };
        match scope.create_hook(&self.api, hook).await {
            Ok(()) => {
                tracing::info!(%target_url, %event, "swarm forge: webhook registered");
                Ok(())
            }
            Err(e) if is_already_exists(&e) => {
                tracing::debug!(%target_url, "swarm forge: webhook already present");
                Ok(())
            }
            Err(e) => Err(e.into()),
        }
    }
}

/// Every distinct node reachable from `start` by following `adj` edges
/// (`adj[n]` = the issues that directly depend on `n`), not counting
/// `start` itself. Plain DFS with a `visited` set doubling as the
/// cycle guard — see [`Client::issue_report`]'s doc comment for why a
/// cycle has to be tolerated rather than assumed impossible.
fn transitive_reach(start: i64, adj: &HashMap<i64, Vec<i64>>) -> usize {
    // `start` goes into `visited` up front, pre-empting the one case a
    // plain "insert on visit" walk gets wrong: a cycle that loops back to
    // `start` would otherwise re-insert it and count it as its own
    // descendant. Pre-seeding makes that re-visit a no-op instead, so the
    // `- 1` below only ever backs out the seed, never a real node.
    let mut visited: HashSet<i64> = HashSet::from([start]);
    let mut stack: Vec<i64> = adj.get(&start).cloned().unwrap_or_default();
    while let Some(node) = stack.pop() {
        if visited.insert(node)
            && let Some(next) = adj.get(&node)
        {
            stack.extend(next);
        }
    }
    visited.len() - 1
}

/// Where a hook lives. The knowledge hook is repo-scoped and the config-PR
/// hook is org-scoped — a hook on the wrong scope would never fire, and forgejo would
/// report that as a perfectly healthy hook with no deliveries.
enum HookScope<'a> {
    Repo {
        org: &'a str,
        repo: &'a str,
    },
    Org {
        org: &'a str,
    },
    /// Forgejo's "global (system) webhook" — fires for every repo in the
    /// instance, in every org, present or future. The `admin_*` API
    /// namespace; needs the same `write:admin` scope
    /// `Client::ensure_agent_user` already requires on this token, **and**
    /// the `is_system_webhook` config key from
    /// [`Self::extra_create_config`] — that endpoint's default is the
    /// *other* kind of admin hook.
    Instance,
}

impl HookScope<'_> {
    /// Config-map keys the create call needs beyond the shared ones.
    ///
    /// Instance scope carries `is_system_webhook`, and it is load-bearing
    /// twice over. `POST /admin/hooks` reads that key **out of the config
    /// map** and defaults it to `false`, which makes a forgejo *default*
    /// webhook — a template copied into repos created later, not a live
    /// hook — while `GET /admin/hooks` returns only webhooks with the flag
    /// set. Omitting it therefore breaks both halves at once: the hook is
    /// not the instance-wide one this scope exists for, and
    /// [`Self::list_hook_urls`] can never see it, so every process start
    /// creates another one — twelve in a day, `listed=0` each time,
    /// before this key was sent.
    fn extra_create_config(&self) -> &'static [(&'static str, &'static str)] {
        match self {
            Self::Repo { .. } | Self::Org { .. } => &[],
            Self::Instance => &[("is_system_webhook", "true")],
        }
    }

    /// The `url` config value of every hook currently on this scope.
    async fn list_hook_urls(&self, api: &api::Api) -> Result<Vec<String>, ForgejoError> {
        let hooks = match self {
            Self::Repo { org, repo } => {
                api.call(|api| Box::pin(api.repo_list_hooks(org, repo).all()))
                    .await?
            }
            Self::Org { org } => {
                api.call(|api| api.org_list_hooks(org).into_future())
                    .await?
            }
            Self::Instance => api.call(|api| api.admin_list_hooks().into_future()).await?,
        };
        Ok(hooks
            .iter()
            .filter_map(|h| h.config.as_ref()?.get("url").cloned())
            .collect())
    }

    async fn create_hook(
        &self,
        api: &api::Api,
        hook: CreateHookOption,
    ) -> Result<(), ForgejoError> {
        match self {
            Self::Repo { org, repo } => api
                .call(|api| api.repo_create_hook(org, repo, hook.clone()).into_future())
                .await
                .map(drop),
            Self::Org { org } => api
                .call(|api| api.org_create_hook(org, hook.clone()).into_future())
                .await
                .map(drop),
            Self::Instance => api
                .call(|api| api.admin_create_hook(hook.clone()).into_future())
                .await
                .map(drop),
        }
    }
}

/// Base64-encode `content` for a forgejo content-API call — every file
/// write on this API takes base64, never raw bytes.
fn base64_encode(content: &str) -> String {
    use base64::Engine as _;
    base64::engine::general_purpose::STANDARD.encode(content)
}

/// Per-agent NixOS module seed — same content
/// `hive-c0re::lifecycle::setup::initial_agent_nix` writes at the per-hive
/// level (this process has no access to that function across the crate
/// boundary, and it's three lines — not worth a shared crate for).
///
/// Deliberately does **not** record which hive the agent belongs to, even
/// though `POST /api/agents` now takes one: an agent's config should carry
/// no reference to the hive it runs on. The hive is an address the swarm
/// routes on, not a property of the agent — a
/// config that named its own hive would be a second place stating where
/// the agent lives, free to disagree with the queue that actually
/// delivers to it.
fn initial_agent_nix(name: &str) -> String {
    format!(
        "{{ config, pkgs, lib, ... }}:\n{{\n  # Per-agent overrides for {name}. This is a regular NixOS module\n  # — add packages, services, modules, imports as needed.\n  #\n  # imports = [ ./extra-module.nix ];\n  # environment.systemPackages = with pkgs; [ ];\n}}\n",
    )
}

/// Module-only flake seed, byte-identical to
/// `hive-c0re::lifecycle::setup::initial_flake_nix` — see that function's
/// doc comment for why the shape looks the way it does (`flakeInputs`
/// forwarding, meta-flake consumption). Not shared across the crate
/// boundary for the same reason [`initial_agent_nix`] isn't.
fn initial_flake_nix() -> &'static str {
    "{\n  description = \"hyperhive agent\";\n  inputs = { };\n  outputs =\n    { self, ... }@inputs:\n    {\n      nixosModules.default = {\n        imports = [ ./agent.nix ];\n        _module.args.flakeInputs = builtins.removeAttrs inputs [ \"self\" ];\n      };\n    };\n}\n"
}

/// Canonical email for an agent's Forgejo account, set by
/// [`Client::create_agent_user`] at creation and checked by
/// [`Client::disable_repo_creation`] before it sends the account's PATCH. Same
/// format string as `hive-c0re::forge::users::agent_email` — not called
/// across the crate boundary (see this module's doc comment), but both
/// creators use this exact value, so it doubles as the
/// marker that distinguishes an agent's own account from an unrelated one
/// that happens to share its username.
fn agent_email(agent: &str) -> String {
    format!("{agent}@hyperhive.local")
}

/// The email older hive-created agent accounts carry, and keep: nothing
/// rewrites it to [`agent_email`]. Accepted by
/// [`Client::disable_repo_creation`]'s guard alongside [`agent_email`] so
/// such an account is still recognised as the agent's own.
fn legacy_agent_email(agent: &str) -> String {
    format!("{agent}@hive.local")
}

/// The `admin_edit_user` body that sets `max_repo_creation = 0` on `agent`
/// and nothing else. `login_name` + `source_id = 0` (local auth,
/// which is what [`Client::create_agent_user`] creates) ride along because
/// Forgejo resets `use_custom_avatar` on an admin edit that omits
/// `login_name`; every other field is left unset so nothing else is touched.
fn repo_creation_lockdown(agent: &str) -> EditUserOption {
    EditUserOption {
        active: None,
        admin: None,
        allow_create_organization: None,
        allow_git_hook: None,
        allow_import_local: None,
        description: None,
        email: None,
        full_name: None,
        hide_email: None,
        location: None,
        login_name: Some(agent.to_owned()),
        max_repo_creation: Some(0),
        must_change_password: None,
        password: None,
        prohibit_login: None,
        pronouns: None,
        restricted: None,
        source_id: Some(0),
        visibility: None,
        website: None,
    }
}

/// Whether a create-style call failed because the object already
/// exists. Forgejo signals this as HTTP 409 (conflict) or 422
/// (validation) — match both defensively, same as
/// `hive-c0re::forge::repos::is_already_exists`.
fn is_already_exists(e: &ForgejoError) -> bool {
    match e {
        ForgejoError::ApiError(api) => match api.error_kind() {
            ApiErrorKind::ValidationFailed => true,
            ApiErrorKind::Other(s) => *s == StatusCode::CONFLICT,
            _ => false,
        },
        ForgejoError::UnexpectedStatusCode(s) => {
            *s == StatusCode::CONFLICT || *s == StatusCode::UNPROCESSABLE_ENTITY
        }
        _ => false,
    }
}

/// Whether a create-call failure is forgejo's unambiguous "already
/// exists" signal — HTTP 409. Unlike a 422 (see
/// [`is_ambiguous_validation_failure`]), nothing else produces a 409 on
/// these create calls, so folding it into success needs no follow-up
/// confirmation.
fn is_confirmed_conflict(e: &ForgejoError) -> bool {
    match e {
        ForgejoError::ApiError(api) => {
            matches!(api.error_kind(), ApiErrorKind::Other(s) if *s == StatusCode::CONFLICT)
        }
        ForgejoError::UnexpectedStatusCode(s) => *s == StatusCode::CONFLICT,
        _ => false,
    }
}

/// Whether a create-call failure is forgejo's 422 "validation failed" —
/// ambiguous on its own. Forgejo folds several distinct causes into this
/// one status; an admin user create alone answers 422 for
/// `ErrUserAlreadyExist`, `ErrEmailAlreadyUsed`, `ErrNameReserved`,
/// `ErrNameCharsNotAllowed`, `ErrEmailInvalid` and
/// `ErrNamePatternNotAllowed` — only the first of those means the object
/// exists. A 422 alone is never enough to fold into success; the caller
/// must confirm with a follow-up read first (see [`folds_into_success`]).
fn is_ambiguous_validation_failure(e: &ForgejoError) -> bool {
    match e {
        ForgejoError::ApiError(api) => matches!(api.error_kind(), ApiErrorKind::ValidationFailed),
        ForgejoError::UnexpectedStatusCode(s) => *s == StatusCode::UNPROCESSABLE_ENTITY,
        _ => false,
    }
}

/// Whether a create-call failure `e` should be folded into success, given
/// whether a follow-up read found the object already present.
/// `existing` is only consulted for the ambiguous 422 case — a 409 is
/// trusted on its own, and any other error always fails.
fn folds_into_success(e: &ForgejoError, existing: bool) -> bool {
    is_confirmed_conflict(e) || (is_ambiguous_validation_failure(e) && existing)
}

/// Whether a forge call failed because the named repo itself doesn't exist —
/// forgejo answers a plain HTTP 404 for `owner/repo` not found, which the
/// generated client surfaces as `ApiErrorKind::NotFound`. Checked
/// defensively against `ApiErrorKind::Other`/`UnexpectedStatusCode` too,
/// same dual-check shape [`is_already_exists`] uses for 409/422 — belt and
/// braces against a response whose body didn't parse as the expected shape.
/// Used by [`crate::issue_report::get_issue_report`] to tell "no such repo"
/// (404, retrying won't help) apart from every other forge failure (503,
/// might clear up) — see that module for why the distinction matters.
pub(crate) fn is_repo_not_found(e: &ForgejoError) -> bool {
    match e {
        ForgejoError::ApiError(api) => match api.error_kind() {
            ApiErrorKind::NotFound { .. } => true,
            ApiErrorKind::Other(s) => *s == StatusCode::NOT_FOUND,
            _ => false,
        },
        ForgejoError::UnexpectedStatusCode(s) => *s == StatusCode::NOT_FOUND,
        _ => false,
    }
}

#[cfg(test)]
fn validation_failed_error() -> ForgejoError {
    ForgejoError::ApiError(ApiErrorKind::ValidationFailed.into())
}

#[cfg(test)]
fn conflict_error() -> ForgejoError {
    ForgejoError::ApiError(ApiErrorKind::Other(StatusCode::CONFLICT).into())
}

#[cfg(test)]
impl Client {
    /// A client for a test's own forge at `url`, holding a placeholder token.
    pub(crate) fn for_test_forge(url: url::Url) -> Self {
        let api = crate::forge::api::tests::fixed(url.clone(), "stub-token");
        Self { api, url }
    }
}

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

    #[test]
    fn open_pulls_are_listed_newest_first_and_skip_one_with_no_number() {
        // The generated type insists on `requested_reviewers` being present.
        let pull = |mut pr: serde_json::Value| {
            pr["requested_reviewers"] = serde_json::json!([]);
            serde_json::from_value(pr).expect("a pull request")
        };
        let pulls = vec![
            pull(serde_json::json!({"number": 7, "title": "older", "head": {"sha": "aaa"}})),
            pull(serde_json::json!({"title": "no number", "head": {"sha": "bbb"}})),
            pull(serde_json::json!({"number": 12, "title": "newer", "head": {"sha": "ccc"}})),
            pull(serde_json::json!({"number": 9, "head": {}})),
        ];
        let want = |number, title: &str, head_sha: Option<&str>| OpenPull {
            number,
            title: title.to_owned(),
            head_sha: head_sha.map(str::to_owned),
        };
        assert_eq!(
            open_pulls(pulls),
            vec![
                want(12, "newer", Some("ccc")),
                want(9, "", None),
                want(7, "older", Some("aaa")),
            ]
        );
    }

    // Both scenarios live in ONE test, not two, deliberately: `cargo
    // test` runs tests in parallel by default, and two tests mutating
    // the same process-global env var would race each other. A single
    // test's body is one thread, so sequential set/clear within it is
    // race-free without needing a cross-test lock.
    #[test]
    fn url_from_env_is_none_without_a_url_and_refuses_a_bad_one() {
        // SAFETY: test-only mutation of the env var this module itself
        // owns (`URL_ENV`), scoped to this one test function so there's no
        // cross-test race (see comment above).
        unsafe {
            std::env::remove_var(URL_ENV);
        }
        assert_eq!(url_from_env().unwrap(), None);

        unsafe {
            std::env::set_var(URL_ENV, "not a url");
        }
        let err = url_from_env().expect_err("an unparseable forge URL");
        assert!(err.to_string().contains(URL_ENV), "{err:#}");

        unsafe {
            std::env::remove_var(URL_ENV);
        }
    }

    // Paused time: each sleep completes at once and advances the clock, so the
    // elapsed time is the backoff schedule itself.
    #[tokio::test(start_paused = true)]
    async fn a_token_absent_at_start_fills_the_slot_once_it_arrives() {
        let slot = OnceLock::new();
        let start = tokio::time::Instant::now();
        let mut calls = 0;
        let got = fill_when_ready(&slot, || {
            calls += 1;
            let n = calls;
            async move {
                if n < 3 {
                    anyhow::bail!("nothing at the controller's forge token path")
                }
                Ok("client")
            }
        })
        .await;
        assert_eq!(*got, "client");
        assert_eq!(slot.get(), Some(&"client"));
        assert_eq!(calls, 3);
        assert_eq!(start.elapsed().as_secs(), 30 + 60 + 120);
    }

    #[tokio::test(start_paused = true)]
    async fn a_token_that_stays_absent_is_retried_every_five_minutes() {
        let slot = OnceLock::new();
        let start = tokio::time::Instant::now();
        let mut calls = 0;
        fill_when_ready(&slot, || {
            calls += 1;
            let n = calls;
            async move {
                if n < 7 {
                    anyhow::bail!("nothing at the controller's forge token path")
                }
                Ok("client")
            }
        })
        .await;
        assert_eq!(calls, 7);
        assert_eq!(
            start.elapsed().as_secs(),
            30 + 60 + 120 + 240 + 300 + 300 + 300
        );
    }

    #[tokio::test(start_paused = true)]
    async fn a_pass_that_never_answers_is_abandoned_at_the_deadline() {
        let start = tokio::time::Instant::now();
        let got = within_forge_deadline("test", std::future::pending::<()>()).await;
        assert_eq!(got, None);
        assert_eq!(start.elapsed(), FORGE_DEADLINE);
    }

    #[tokio::test(start_paused = true)]
    async fn a_pass_that_answers_returns_its_result() {
        assert_eq!(within_forge_deadline("test", async { 7 }).await, Some(7));
    }

    /// A forge that accepts the connection and never answers fails the
    /// registration at the deadline, so the caller's retry advances. The
    /// paused clock jumps to the deadline once nothing else can run.
    #[tokio::test(start_paused = true)]
    async fn swarm_webhook_registration_fails_at_the_deadline_when_the_forge_never_answers() {
        // Bound but never accepted: the handshake completes, no response comes.
        let listener = tokio::net::TcpListener::bind("127.0.0.1:0").await.unwrap();
        let url = url::Url::parse(&format!("http://{}/", listener.local_addr().unwrap())).unwrap();
        let client = Client::for_test_forge(url);
        let start = tokio::time::Instant::now();
        // The outer bound turns a missing deadline into a failure, not a hang.
        let err = tokio::time::timeout(
            FORGE_DEADLINE * 2,
            client.ensure_swarm_webhooks("https://swarm.example", "s3cret"),
        )
        .await
        .expect("registration returns by itself")
        .expect_err("no answer is a failure");
        assert_eq!(start.elapsed(), FORGE_DEADLINE);
        assert!(format!("{err:#}").contains("did not answer"), "{err:#}");
    }

    #[tokio::test]
    async fn page_search_results_pages_past_a_full_first_page() {
        #[derive(Clone)]
        struct FakeRepo {
            full_name: String,
            open_issues: u32,
        }

        // Page 1 is exactly PAGE_SIZE long, which must not be mistaken for
        // the last page — a repo living on page 2 still has to turn up.
        const PAGE_SIZE: u32 = 2;
        let pages = vec![
            vec![
                FakeRepo {
                    full_name: "org/a".to_owned(),
                    open_issues: 0,
                },
                FakeRepo {
                    full_name: "org/b".to_owned(),
                    open_issues: 0,
                },
            ],
            vec![FakeRepo {
                full_name: "org/c".to_owned(),
                open_issues: 3,
            }],
        ];
        let repos = Client::page_search_results(PAGE_SIZE, 10, |page| {
            let pages = pages.clone();
            async move { Ok(pages.get(page as usize - 1).cloned()) }
        })
        .await
        .unwrap();

        let with_open_issues: Vec<_> = repos
            .into_iter()
            .filter(|r| r.open_issues > 0)
            .map(|r| r.full_name)
            .collect();
        assert_eq!(with_open_issues, vec!["org/c".to_owned()]);
    }

    #[tokio::test]
    async fn page_search_results_errors_instead_of_looping_forever() {
        // A server that never returns a short page (misbehaving, or an
        // instance that genuinely outgrew `max_pages`) must fail loudly
        // rather than paging forever or handing back a partial result.
        const PAGE_SIZE: u32 = 2;
        let result =
            Client::page_search_results(
                PAGE_SIZE,
                3,
                |_page| async move { Ok(Some(vec![1u32, 2])) },
            )
            .await;
        assert!(result.is_err());
    }

    #[test]
    fn transitive_reach_walks_a_chain_not_just_the_direct_edge() {
        // 3 depends on 2 depends on 1: reverse_adj is "who depends on me",
        // so 1 -> [2], 2 -> [3]. Fixing 1 eventually frees both 2 and 3.
        let adj = HashMap::from([(1, vec![2]), (2, vec![3])]);
        assert_eq!(transitive_reach(1, &adj), 2);
        assert_eq!(transitive_reach(2, &adj), 1);
        assert_eq!(transitive_reach(3, &adj), 0);
    }

    #[test]
    fn transitive_reach_counts_a_diamond_descendant_once() {
        // 1 -> [2, 3], both 2 and 3 -> [4]: 4 is reachable via two paths
        // but must only be counted once.
        let adj = HashMap::from([(1, vec![2, 3]), (2, vec![4]), (3, vec![4])]);
        assert_eq!(transitive_reach(1, &adj), 3);
    }

    #[test]
    fn transitive_reach_terminates_on_a_cycle() {
        // 1 -> [2] -> [1]: nothing forge-side prevents a dependency cycle,
        // so the walk has to survive one instead of looping forever.
        let adj = HashMap::from([(1, vec![2]), (2, vec![1])]);
        assert_eq!(transitive_reach(1, &adj), 1);
    }

    // The three cases from the `forge-422-fold` issue: a create call's
    // 422 with the object confirmed absent must fail (case 1), the same
    // 422 with the object confirmed present must fold to success (case
    // 2), and a 409 must fold to success without needing a follow-up
    // read at all (case 3).
    #[test]
    fn folds_into_success_case1_422_absent_fails() {
        assert!(!folds_into_success(&validation_failed_error(), false));
    }

    #[test]
    fn folds_into_success_case2_422_present_is_done() {
        assert!(folds_into_success(&validation_failed_error(), true));
    }

    #[test]
    fn folds_into_success_case3_conflict_is_done() {
        // `existing` is never consulted for a 409 — pass `false` to prove
        // that, not just `true`.
        assert!(folds_into_success(&conflict_error(), false));
    }

    // Invert-proof: the old `is_already_exists` folded every 422 into
    // success regardless of whether the object existed — exactly the
    // defect `folds_into_success` fixes. Demonstrate the old logic
    // getting case 1 wrong before trusting the new one.
    #[test]
    fn old_is_already_exists_wrongly_returns_true_for_an_absent_object() {
        let e = validation_failed_error();
        assert!(is_already_exists(&e), "old logic: false positive on 422");
        assert!(
            !folds_into_success(&e, false),
            "new logic: correctly refuses to fold a 422 with no confirmed existence"
        );
    }

    /// What [`stub_forge`] saw: method, path and JSON body of each request.
    pub(super) type Seen =
        std::sync::Arc<std::sync::Mutex<Vec<(String, String, serde_json::Value)>>>;

    /// A stand-in forge on a loopback port — never the real one — answering
    /// the admin user create (`POST`) with `create`, the pre-lockdown
    /// `user_get` (`GET`) with `get`, and the admin user edit (`PATCH`) with
    /// `edit` (status, JSON body each), and recording every request.
    pub(super) async fn stub_forge(
        create: (u16, &'static str),
        get: (u16, &'static str),
        edit: (u16, &'static str),
    ) -> (Client, Seen) {
        use axum::http::{Method, StatusCode, Uri, header::CONTENT_TYPE};

        let seen = Seen::default();
        let recorded = seen.clone();
        let app = axum::Router::new().fallback(
            move |method: Method, uri: Uri, body: axum::body::Bytes| {
                let recorded = recorded.clone();
                async move {
                    let json = serde_json::from_slice(&body).unwrap_or_default();
                    recorded.lock().unwrap().push((
                        method.to_string(),
                        uri.path().to_owned(),
                        json,
                    ));
                    let (status, body) = match method {
                        Method::POST => create,
                        Method::GET => get,
                        Method::PATCH => edit,
                        _ => (404, r#"{"message":"not stubbed"}"#),
                    };
                    let status = StatusCode::from_u16(status).unwrap();
                    (status, [(CONTENT_TYPE, "application/json")], body)
                }
            },
        );
        let listener = tokio::net::TcpListener::bind("127.0.0.1:0").await.unwrap();
        let url = url::Url::parse(&format!("http://{}/", listener.local_addr().unwrap())).unwrap();
        tokio::spawn(async move { axum::serve(listener, app).await });
        let api = crate::forge::api::tests::fixed(url.clone(), "stub-token");
        (Client { api, url }, seen)
    }

    /// The `PATCH` [`Client::ensure_agent_user`] sent, if any.
    pub(super) fn lockdown_patch(seen: &Seen, agent: &str) -> Option<serde_json::Value> {
        let path = format!("/api/v1/admin/users/{agent}");
        seen.lock()
            .unwrap()
            .iter()
            .find(|(method, p, _)| method == "PATCH" && *p == path)
            .map(|(_, _, body)| body.clone())
    }

    /// The email [`stub_forge`]'s `user_get` stub should answer with for an
    /// account this agent actually owns — same value
    /// [`Client::create_agent_user`] itself sets.
    fn own_account_get(agent: &str) -> (u16, &'static str) {
        // Leaked rather than borrowed: `stub_forge`'s `get` param needs
        // `&'static str` (it moves into a `'static` axum handler), and this
        // helper builds the body per-agent instead of hardcoding "alice".
        (
            200,
            Box::leak(format!(r#"{{"email":"{agent}@hyperhive.local"}}"#).into_boxed_str()),
        )
    }

    /// A freshly created agent account is locked out of repo creation by
    /// this node itself — not left for a hive-level pass to catch later —
    /// and the edit sets nothing beyond `max_repo_creation` plus the
    /// `login_name` / `source_id` pair Forgejo needs to leave the rest alone.
    #[tokio::test]
    async fn a_created_agent_user_gets_repo_creation_disabled() {
        let (client, seen) = stub_forge((201, "{}"), own_account_get("alice"), (200, "{}")).await;

        client.ensure_agent_user("alice").await.unwrap();

        let body = lockdown_patch(&seen, "alice").expect("no lockdown PATCH sent");
        assert_eq!(body["max_repo_creation"], 0);
        let set: Vec<&str> = body
            .as_object()
            .unwrap()
            .iter()
            .filter(|(_, v)| !v.is_null())
            .map(|(k, _)| k.as_str())
            .collect();
        assert_eq!(set, ["login_name", "max_repo_creation", "source_id"]);
        assert_eq!(body["login_name"], "alice");
        assert_eq!(body["source_id"], 0);
    }

    /// An account that already existed is locked down too: the node is
    /// re-run, and an account created before this lockdown existed must
    /// not keep repo creation just because the create step folded to
    /// success. 409 rather than 422 so the case doesn't hinge on how a 422
    /// is disambiguated.
    #[tokio::test]
    async fn an_existing_agent_user_still_gets_repo_creation_disabled() {
        let (client, seen) = stub_forge(
            (409, r#"{"message":"user already exists"}"#),
            own_account_get("alice"),
            (200, "{}"),
        )
        .await;

        client.ensure_agent_user("alice").await.unwrap();

        let body = lockdown_patch(&seen, "alice").expect("no lockdown PATCH sent");
        assert_eq!(body["max_repo_creation"], 0);
    }

    /// A refused lockdown fails the node with Forgejo's own message, rather
    /// than warning and reporting the account as provisioned, which would
    /// leave an agent able to create and self-merge with nothing marked
    /// failed.
    #[tokio::test]
    async fn a_refused_lockdown_fails_the_node() {
        let (client, _seen) = stub_forge(
            (201, "{}"),
            own_account_get("alice"),
            (403, r#"{"message":"token does not have required scope"}"#),
        )
        .await;

        let err = client.ensure_agent_user("alice").await.unwrap_err();

        let msg = format!("{err:#}");
        assert!(msg.contains("max_repo_creation"), "{msg}");
        assert!(msg.contains("token does not have required scope"), "{msg}");
    }

    /// The guard this PR adds: a name collision with a pre-existing
    /// non-agent account (or any account whose email isn't the
    /// `{agent}@hyperhive.local` marker) must fail the node rather than
    /// PATCH `max_repo_creation`/`source_id` onto an account nobody
    /// verified this node created. 409 create-fold, same as
    /// `an_existing_agent_user_still_gets_repo_creation_disabled` — the
    /// fold itself is untouched by this PR; the fix lives in the read right
    /// before the PATCH.
    #[tokio::test]
    async fn a_name_collision_with_a_non_agent_account_fails_without_patching() {
        let (client, seen) = stub_forge(
            (409, r#"{"message":"user already exists"}"#),
            (200, r#"{"email":"someone@example.com"}"#),
            (200, "{}"),
        )
        .await;

        let err = client.ensure_agent_user("alice").await.unwrap_err();

        let msg = format!("{err:#}");
        assert!(msg.contains("alice"), "{msg}");
        assert!(
            msg.contains("not this agent's"),
            "must name the collision, got: {msg}"
        );
        assert!(
            lockdown_patch(&seen, "alice").is_none(),
            "a collision must never reach the PATCH"
        );
    }

    /// A legacy hive-created account can carry the same
    /// `{agent}@hyperhive.local` marker, plus `login_name`/`source_id` from
    /// a hive admin edit — this node must still
    /// lock such an account down, not just ones it created itself.
    #[tokio::test]
    async fn a_legacy_hive_created_agent_account_still_gets_locked_down() {
        let (client, seen) = stub_forge(
            (409, r#"{"message":"user already exists"}"#),
            (
                200,
                r#"{"id":42,"login":"alice","login_name":"alice","email":"alice@hyperhive.local","source_id":0}"#,
            ),
            (200, "{}"),
        )
        .await;

        client.ensure_agent_user("alice").await.unwrap();

        let body = lockdown_patch(&seen, "alice").expect("no lockdown PATCH sent");
        assert_eq!(body["max_repo_creation"], 0);
    }

    /// A real, pre-existing agent account can carry the legacy
    /// `@hive.local` email, and nothing rewrites it. The guard must accept
    /// that value too, or such an agent would fail with "not this agent's".
    #[tokio::test]
    async fn a_pre_alignment_legacy_email_still_gets_locked_down() {
        let (client, seen) = stub_forge(
            (409, r#"{"message":"user already exists"}"#),
            (200, r#"{"email":"alice@hive.local"}"#),
            (200, "{}"),
        )
        .await;

        client.ensure_agent_user("alice").await.unwrap();

        let body = lockdown_patch(&seen, "alice").expect("no lockdown PATCH sent");
        assert_eq!(body["max_repo_creation"], 0);
    }

    /// Every `PATCH` body the stub saw.
    fn patches(seen: &Seen) -> Vec<serde_json::Value> {
        seen.lock()
            .unwrap()
            .iter()
            .filter(|(method, _, _)| method == "PATCH")
            .map(|(_, _, body)| body.clone())
            .collect()
    }

    /// Only `admin` goes out: a `source_id` would reset an SSO-made
    /// account's login type, and the other fields would overwrite the user's
    /// own settings.
    #[tokio::test]
    async fn make_site_admin_sets_admin_and_nothing_else() {
        let (client, seen) = stub_forge(
            (404, "{}"),
            (200, r#"{"login":"mara","is_admin":false}"#),
            (200, "{}"),
        )
        .await;

        let plan = client.make_site_admin("mara").await.unwrap();

        assert_eq!(plan, site_admin::Plan::Promote);
        let bodies = patches(&seen);
        assert_eq!(bodies.len(), 1, "{bodies:?}");
        let set: Vec<(&String, &serde_json::Value)> = bodies[0]
            .as_object()
            .unwrap()
            .iter()
            .filter(|(_, v)| !v.is_null())
            .collect();
        assert_eq!(set, [(&"admin".to_owned(), &serde_json::Value::Bool(true))]);
    }

    #[tokio::test]
    async fn make_site_admin_leaves_an_admin_alone() {
        let (client, seen) = stub_forge(
            (404, "{}"),
            (200, r#"{"login":"mara","is_admin":true}"#),
            (200, "{}"),
        )
        .await;

        let plan = client.make_site_admin("mara").await.unwrap();

        assert_eq!(plan, site_admin::Plan::AlreadyAdmin);
        assert!(patches(&seen).is_empty());
    }

    /// A user who has not logged in yet is reported, not created: nothing is
    /// sent but the read.
    #[tokio::test]
    async fn make_site_admin_does_not_create_a_missing_user() {
        let (client, seen) = stub_forge(
            (201, "{}"),
            (404, r#"{"message":"user does not exist"}"#),
            (200, "{}"),
        )
        .await;

        let plan = client.make_site_admin("mara").await.unwrap();

        assert_eq!(plan, site_admin::Plan::NoSuchUser);
        let methods: Vec<String> = seen
            .lock()
            .unwrap()
            .iter()
            .map(|(m, _, _)| m.clone())
            .collect();
        assert_eq!(methods, ["GET"]);
    }

    const RULES_YML: &str = ".hyperhive/review-classifier.yml";
    const RULES_JSON: &str = ".hyperhive/review-classifier.json";

    /// A stand-in forge holding `o/r` with default branch `main` at
    /// `def456`. The branch lookup answers `branch`, and the commit's root
    /// tree answers `root` (a 404 for `None`). The `.hyperhive` tree holds
    /// only `review-classifier.yml`. Records every request's path.
    async fn rules_forge(
        branch: u16,
        root: Option<&'static str>,
    ) -> (Client, Arc<std::sync::Mutex<Vec<String>>>) {
        use axum::http::{StatusCode, Uri};
        use base64::Engine as _;

        let blob = format!(
            r#"{{"sha":"b1","encoding":"base64","content":"{}"}}"#,
            base64::engine::general_purpose::STANDARD.encode("classifications: []\n")
        );
        let seen = Arc::new(std::sync::Mutex::new(Vec::new()));
        let recorded = Arc::clone(&seen);
        let app = axum::Router::new().fallback(move |uri: Uri| {
            let (recorded, blob) = (Arc::clone(&recorded), blob.clone());
            async move {
                recorded.lock().unwrap().push(uri.path().to_owned());
                let (status, body) = match uri.path() {
                    "/api/v1/repos/o/r" => {
                        (200, r#"{"default_branch":"main","empty":false}"#.to_owned())
                    }
                    "/api/v1/repos/o/r/branches/main" if branch == 200 => (
                        200,
                        r#"{"name":"main","commit":{"id":"def456"}}"#.to_owned(),
                    ),
                    "/api/v1/repos/o/r/branches/main" => (branch, r#"{"message":"no"}"#.to_owned()),
                    "/api/v1/repos/o/r/git/trees/def456" if root.is_some() => {
                        (200, root.unwrap_or_default().to_owned())
                    }
                    "/api/v1/repos/o/r/git/trees/t1" => (
                        200,
                        r#"{"sha":"t1","truncated":false,"tree":[
                            {"path":"review-classifier.yml","type":"blob","sha":"b1"}]}"#
                            .to_owned(),
                    ),
                    "/api/v1/repos/o/r/git/blobs/b1" => (200, blob),
                    _ => (404, r#"{"message":"not found"}"#.to_owned()),
                };
                (StatusCode::from_u16(status).unwrap(), body)
            }
        });
        let listener = tokio::net::TcpListener::bind("127.0.0.1:0").await.unwrap();
        let url = url::Url::parse(&format!("http://{}/", listener.local_addr().unwrap())).unwrap();
        tokio::spawn(async move { axum::serve(listener, app).await });
        let api = crate::forge::api::tests::fixed(url.clone(), "stub-token");
        (Client { api, url }, seen)
    }

    /// One PR commit whose parent is `mb`.
    const ON_MB: &str = r#"[{"sha":"c1","parents":[{"sha":"mb"}]}]"#;

    /// A stand-in forge holding PR 7 of `o/r`, head `h1`, which changes
    /// `m.nix` from `old` at merge base `mb` to `new`. `pull` is the PR's own
    /// JSON, `commits` its commit list. Records every request's path.
    async fn pull_forge(
        pull: &'static str,
        commits: &'static str,
        old: &'static str,
        new: &'static str,
    ) -> (Client, Arc<std::sync::Mutex<Vec<String>>>) {
        use axum::http::Uri;
        use base64::Engine as _;

        let blob = |content: &str| {
            format!(
                r#"{{"encoding":"base64","content":"{}"}}"#,
                base64::engine::general_purpose::STANDARD.encode(content)
            )
        };
        let tree = |sha: &str, blob: &str| {
            format!(
                r#"{{"sha":"{sha}","truncated":false,"tree":[
                    {{"path":"m.nix","type":"blob","sha":"{blob}"}}]}}"#
            )
        };
        let seen = Arc::new(std::sync::Mutex::new(Vec::new()));
        let recorded = Arc::clone(&seen);
        let app = axum::Router::new().fallback(move |uri: Uri| {
            let recorded = Arc::clone(&recorded);
            let body = match uri.path() {
                "/api/v1/repos/o/r/pulls/7" => Some(pull.to_owned()),
                "/api/v1/repos/o/r/pulls/7/commits" => Some(commits.to_owned()),
                "/api/v1/repos/o/r/pulls/7/files" => {
                    Some(r#"[{"filename":"m.nix","status":"modified"}]"#.to_owned())
                }
                "/api/v1/repos/o/r/pulls/7.diff" => Some("forge diff\n".to_owned()),
                "/api/v1/repos/o/r/git/trees/mb" => Some(tree("mb", "bo")),
                "/api/v1/repos/o/r/git/trees/h1" => Some(tree("h1", "bn")),
                "/api/v1/repos/o/r/git/blobs/bo" => Some(blob(old)),
                "/api/v1/repos/o/r/git/blobs/bn" => Some(blob(new)),
                _ => None,
            };
            async move {
                recorded.lock().unwrap().push(uri.path().to_owned());
                let builder = axum::http::Response::builder();
                let (builder, body) = match body {
                    Some(body) => {
                        let count = if body == "[]" { "0" } else { "1" };
                        (builder.status(200).header("x-total-count", count), body)
                    }
                    None => (builder.status(404), r#"{"message":"not found"}"#.to_owned()),
                };
                builder.body(axum::body::Body::from(body)).unwrap()
            }
        });
        let listener = tokio::net::TcpListener::bind("127.0.0.1:0").await.unwrap();
        let url = url::Url::parse(&format!("http://{}/", listener.local_addr().unwrap())).unwrap();
        tokio::spawn(async move { axum::serve(listener, app).await });
        let api = crate::forge::api::tests::fixed(url.clone(), "stub-token");
        (Client { api, url }, seen)
    }

    fn commit(sha: &str, parents: &[&str]) -> forgejo_api::structs::Commit {
        let parents = parents
            .iter()
            .map(|p| format!(r#"{{"sha":"{p}"}}"#))
            .collect::<Vec<_>>()
            .join(",");
        serde_json::from_str(&format!(r#"{{"sha":"{sha}","parents":[{parents}]}}"#)).unwrap()
    }

    #[test]
    fn the_merge_base_is_the_one_parent_outside_the_pr() {
        let linear = [commit("c2", &["c1"]), commit("c1", &["mb"])];
        assert_eq!(merge_base_of(&linear).as_deref(), Some("mb"));
        // A merge of the base branch into the head reaches two outside parents.
        let merged = [commit("c2", &["c1", "main-tip"]), commit("c1", &["mb"])];
        assert_eq!(merge_base_of(&merged), None);
        assert_eq!(merge_base_of(&[]), None);
    }

    #[tokio::test]
    async fn a_review_reads_the_wide_diff_from_the_merge_base_or_says_why_not() {
        let old = "{\n  config = lib.mkIf gate {\n    a = 1;\n    b = 2;\n    c = 3;\n    d = 4;\n    e = 5;\n  };\n}\n";
        let new = "{\n  config = lib.mkIf gate {\n    a = 1;\n    b = 2;\n    c = 3;\n    d = 4;\n  };\n}\n";
        let wide = r#"{"number":7,"requested_reviewers":[],"head":{"sha":"h1"},"merge_base":"mb","changed_files":1}"#;
        let (client, _) = pull_forge(wide, ON_MB, old, new).await;
        let read = client.pull_for_review("o", "r", 7).await.unwrap();
        assert_eq!(read.narrowed, None);
        assert!(
            read.diff
                .starts_with("diff --git a/m.nix b/m.nix\n--- a/m.nix\n+++ b/m.nix\n"),
            "{}",
            read.diff
        );
        assert!(
            read.diff.contains("\n   config = lib.mkIf gate {\n"),
            "{}",
            read.diff
        );
        assert!(read.diff.contains("\n-    e = 5;\n"), "{}", read.diff);

        let many = r#"{"number":7,"requested_reviewers":[],"head":{"sha":"h1"},"merge_base":"mb","changed_files":101}"#;
        let (client, seen) = pull_forge(many, ON_MB, old, new).await;
        let read = client.pull_for_review("o", "r", 7).await.unwrap();
        assert_eq!(read.diff, "forge diff\n");
        assert_eq!(
            read.narrowed.as_deref(),
            Some("the PR changes 101 files, more than the 100 a wide diff is read for")
        );
        assert!(
            !seen.lock().unwrap().iter().any(|p| p.ends_with("/files")),
            "no file is read for a PR past the limit"
        );

        // The PR record still names the base it was opened on; the commits
        // name the current one.
        let stale = r#"{"number":7,"requested_reviewers":[],"head":{"sha":"h1"},"merge_base":"old-base","changed_files":1}"#;
        let (client, _) = pull_forge(stale, ON_MB, old, new).await;
        let read = client.pull_for_review("o", "r", 7).await.unwrap();
        assert_eq!(read.narrowed, None, "the diff is read from `mb`");
        assert!(read.diff.contains("\n-    e = 5;\n"), "{}", read.diff);

        let no_commits = r#"{"number":7,"requested_reviewers":[],"head":{"sha":"h1"},"merge_base":"mb","changed_files":1}"#;
        let (client, _) = pull_forge(no_commits, "[]", old, new).await;
        let read = client.pull_for_review("o", "r", 7).await.unwrap();
        assert_eq!(read.diff, "forge diff\n");
        assert_eq!(
            read.narrowed.as_deref(),
            Some("the PR's commits do not name a single merge base")
        );

        let unreadable = r#"{"number":7,"requested_reviewers":[],"head":{"sha":"h2"},"merge_base":"mb","changed_files":1}"#;
        let (client, _) = pull_forge(unreadable, ON_MB, old, new).await;
        let read = client.pull_for_review("o", "r", 7).await.unwrap();
        assert_eq!(read.diff, "forge diff\n");
        let reason = read.narrowed.expect("a failed file read falls back");
        assert!(
            reason.starts_with("reading the files failed: read tree h2 of o/r"),
            "{reason}"
        );
    }

    /// A root tree with the `.hyperhive` tree and a blob beside it.
    const ROOT: &str = r#"{"sha":"def456","truncated":false,"tree":[
        {"path":"README.md","type":"blob","sha":"b0"},
        {"path":".hyperhive","type":"tree","sha":"t1"}]}"#;

    #[tokio::test]
    async fn default_branch_files_reads_each_path_through_the_trees_at_the_branch_head() {
        let (client, seen) = rules_forge(200, Some(ROOT)).await;
        let read = client
            .default_branch_files("o", "r", &[RULES_YML, RULES_JSON])
            .await
            .unwrap()
            .expect("the branch has a commit");
        assert_eq!(read.branch, "main");
        assert_eq!(read.commit, "def456");
        assert_eq!(
            read.files,
            [BranchFile {
                path: RULES_YML.to_owned(),
                sha: "b1".to_owned(),
                bytes: b"classifications: []\n".to_vec(),
            }]
        );
        // Each tree is read once for both paths.
        let seen = seen.lock().unwrap();
        let reads = |p: &str| seen.iter().filter(|s| *s == p).count();
        assert_eq!(reads("/api/v1/repos/o/r/git/trees/def456"), 1, "{seen:?}");
        assert_eq!(reads("/api/v1/repos/o/r/git/trees/t1"), 1, "{seen:?}");
    }

    #[tokio::test]
    async fn a_repo_without_the_directory_has_no_files() {
        let root = r#"{"sha":"def456","truncated":false,"tree":[
            {"path":"README.md","type":"blob","sha":"b0"},
            {"path":".hyperhive","type":"blob","sha":"b2"}]}"#;
        let (client, _) = rules_forge(200, Some(root)).await;
        let read = client
            .default_branch_files("o", "r", &[RULES_YML])
            .await
            .unwrap()
            .expect("the branch has a commit");
        assert_eq!(read.files, []);
    }

    #[tokio::test]
    async fn a_truncated_tree_is_an_error_not_a_missing_file() {
        let root = r#"{"sha":"def456","truncated":true,"tree":[]}"#;
        let (client, _) = rules_forge(200, Some(root)).await;
        let err = client
            .default_branch_files("o", "r", &[RULES_YML])
            .await
            .unwrap_err();
        assert!(format!("{err:#}").contains("truncated"), "{err:#}");
    }

    #[tokio::test]
    async fn a_missing_default_branch_is_no_commit_not_no_repo() {
        let (client, _) = rules_forge(404, Some(ROOT)).await;
        let read = client
            .default_branch_files("o", "r", &[RULES_YML])
            .await
            .unwrap();
        assert_eq!(read, None);
    }

    /// Only the repo lookup's 404 reads as "no such repo"; a failed branch
    /// read or a tree the forge doesn't have is some other failure.
    #[tokio::test]
    async fn only_an_unknown_repo_is_a_forge_404() {
        let no_repo = |e: &anyhow::Error| {
            e.downcast_ref::<ForgejoError>()
                .is_some_and(is_repo_not_found)
        };
        let (client, _) = rules_forge(500, Some(ROOT)).await;
        let unknown = client
            .default_branch_files("o", "elsewhere", &[])
            .await
            .unwrap_err();
        assert!(no_repo(&unknown), "{unknown:#}");
        let branch = client
            .default_branch_files("o", "r", &[])
            .await
            .unwrap_err();
        assert!(!no_repo(&branch), "{branch:#}");
        let (client, _) = rules_forge(200, None).await;
        let tree = client
            .default_branch_files("o", "r", &[RULES_YML])
            .await
            .unwrap_err();
        assert!(!no_repo(&tree), "{tree:#}");
    }

    #[test]
    fn a_long_status_description_is_cut_with_an_ellipsis() {
        assert_eq!(status_description("built"), "built");
        let long = "é".repeat(STATUS_DESCRIPTION_MAX + 5);
        let cut = status_description(&long);
        assert_eq!(cut.chars().count(), STATUS_DESCRIPTION_MAX);
        assert!(cut.ends_with('…'));
    }

    #[tokio::test]
    async fn a_commit_status_is_posted_on_the_commit() {
        let (client, seen) = stub_forge(
            (201, r#"{"id":1}"#),
            (404, r#"{"message":"not stubbed"}"#),
            (404, r#"{"message":"not stubbed"}"#),
        )
        .await;
        client
            .create_commit_status(
                CONFIG_ORG,
                "atlas",
                "abc",
                CommitStatus {
                    context: "swarm-builder",
                    state: CommitStatusState::Failure,
                    description: "error: boom",
                    target_url: Some("https://swarm.example/api/builder/runs/r1"),
                },
            )
            .await
            .expect("the stub accepts the status");
        let seen = seen.lock().unwrap();
        let (method, path, body) = seen.first().expect("one request");
        assert_eq!(method, "POST");
        assert_eq!(path, "/api/v1/repos/agent-configs/atlas/statuses/abc");
        assert_eq!(body["context"], "swarm-builder");
        assert_eq!(body["state"], "failure");
        assert_eq!(body["description"], "error: boom");
        assert_eq!(
            body["target_url"],
            "https://swarm.example/api/builder/runs/r1"
        );
    }
}