Merge pull request 'Add agentws prune' (#13) from benvin/agentws-prune into main

Reviewed-on: #13

AGENTS.md

@@ -19,7 +19,7 @@ like repospawner can run these tools as itself.
   repos into the source root (`~/src/prodenv/<repo>`), creates worktrees under
   the worktree root (`~/.cache/agentws/<repo>__<branch>`), and authenticates
   clone/fetch/push via an ephemeral credential helper. Subcommands: `new`,
-  `list`, `rm`, `clean`, `token`, `credential`.
+  `list`, `rm`, `prune`, `clean`, `token`, `credential`.
 
 All tools are separate `main` packages under `cmd/` and share the
 `internal/agent` package (Vault AppRole login, Gitea REST client, PR-ref
@@ -31,6 +31,7 @@ parsing, watch-state comparison, git worktree helpers).
 cmd/agentpr/main.go      # agentpr CLI (pr create / pr comment / whoami)
 cmd/watchpr/main.go      # watchpr CLI (poll + meaningful-change exit)
 cmd/agentws/main.go      # agentws CLI (new / list / rm / clean / token / credential)
+cmd/agentws/prune.go     # agentws prune (classify worktrees, remove the safe ones)
 cmd/agentvault/main.go   # agentvault CLI (seed-outpost / seed-oauth)
 internal/agent/          # shared plumbing:
   token.go               #   env config + in-process Gitea-token cache
@@ -186,5 +187,15 @@ wrapped per stage (login / read denied / write denied) via `ErrVaultDenied`.
   `seed-oauth` reports key names only.
 - `--rotate` regenerates the `client_secret` too, which then no longer matches
   the IdP provider unless that is rotated alongside.
+- `agentws prune` is a dry run unless `--yes`. It matches a branch to its PR on
+  `head.label`: Gitea rewrites `head.ref` to `refs/pull/<n>/head` once the branch
+  is deleted, which merging does, so `head.ref` matching misses every merged PR.
+  Git signals (`merge-base --is-ancestor`, `git cherry`) are authoritative and
+  offline-safe; an unreachable Gitea only means no branch gets deleted without
+  git proof. A PR's state never authorises a branch delete on its own — HEAD
+  must be contained in the PR's head commit or in `origin/<branch>`, otherwise
+  the worktree goes and the branch stays. `origin/<branch>` is only evidence when
+  this run's pruning fetch succeeded; a failed fetch leaves stale tracking refs,
+  so those verdicts fall back to keeping the branch.
 - CI "combined status" comes from `/commits/{sha}/status`; an empty head SHA
   yields an empty state without an API call.

README.md

@@ -104,6 +104,11 @@ agentws list
 agentws rm benvin/my-change
 agentws rm ~/.cache/agentws/argocd-apps__benvin-my-change --delete-branch
 
+# Classify every managed worktree; dry run unless --yes is given
+agentws prune
+agentws prune --yes
+agentws prune --yes --keep-branches
+
 # Remove every managed worktree and prune each source repo
 agentws clean
 
@@ -111,6 +116,47 @@ agentws clean
 agentws token
 ```
 
+### prune
+
+`agentws prune` decides, per worktree, whether its work is safely upstream:
+
+| Signal (first match wins) | Verdict |
+|---|---|
+| uncommitted or untracked changes | keep |
+| branch has an open PR | keep |
+| tip contained in `origin/<default>` | remove worktree + local branch |
+| every commit patch-equivalent to one in `origin/<default>`'s history | remove worktree + local branch |
+| PR merged **and** HEAD contained in the PR's head commit (or in a verified `origin/<branch>`) | remove worktree + local branch |
+| PR closed **and** HEAD contained in a verified `origin/<branch>` | remove worktree + local branch |
+| anything else | remove worktree, keep the branch |
+
+A branch is deleted only where git proves its commits survive elsewhere. PR
+state alone never authorises that: a merged or closed PR whose branch picked up
+commits since keeps its branch, because those commits exist nowhere but here.
+The delete runs `git branch -d` first so git's own unmerged check is a backstop,
+falling back to `-D` only for a proven branch — squash merges keep the guard
+tripping even once the work has landed.
+
+Patch equivalence comes from `git cherry`, which these squash-merging repos need
+because a merged branch's commits carry different SHAs upstream. It proves the
+patches reached the default branch's history at some point — a later revert
+still counts — not that they stand at its tip.
+
+`origin/<branch>` counts as evidence only when this run's `git fetch --prune`
+succeeded. A tracking ref left over from an earlier fetch may name a branch that
+is already gone upstream and is itself due for deletion, so a failed fetch
+downgrades those verdicts to `remove` and keeps the branch. Proofs that read
+only local objects — containment in `origin/<default>`, patch equivalence, and
+containment in a merged PR's head SHA — stand on their own.
+
+Gitea PR state only adds to the git answer: when it cannot be reached, prune
+says so and never deletes a branch it could not prove, and a PR listing that
+hits the pagination cap is reported rather than read as "no PR". Matching a
+branch to its PR uses `head.label`, since Gitea rewrites `head.ref` to
+`refs/pull/<n>/head` once the branch is deleted on merge.
+
+`--keep-branches` removes worktrees only, and its verdicts print as `remove`.
+
 ### Auth / credential-helper design
 
 Gitea tokens minted from Vault are short-lived (~1h), so `agentws` never

cmd/agentws/main.go

@@ -13,6 +13,7 @@
 //	agentws new <repo> [--branch benvin/<name>] [--from <base-branch>]
 //	agentws list
 //	agentws rm <path-or-branch> [--delete-branch]
+//	agentws prune [--yes] [--keep-branches]
 //	agentws clean
 //	agentws token
 //	agentws credential get   # git credential-helper protocol on stdin
@@ -54,6 +55,7 @@ func newRootCmd() *cobra.Command {
 		newNewCmd(),
 		newListCmd(),
 		newRmCmd(),
+		newPruneCmd(),
 		newCleanCmd(),
 		newTokenCmd(),
 		newCredentialCmd(),
@@ -294,7 +296,8 @@ func newRmCmd() *cobra.Command {
 			if err != nil {
 				return err
 			}
-			return removeWorktree(cmd.OutOrStdout(), wt, deleteBranch)
+			// Naming one worktree to delete is explicit, so rm keeps the force fallback.
+			return removeWorktree(cmd.OutOrStdout(), wt, deleteBranch, true)
 		},
 	}
 	cmd.Flags().BoolVar(&deleteBranch, "delete-branch", false, "Also delete the local branch after removing the worktree")
@@ -316,13 +319,16 @@ func resolveWorktree(target string) (managedWt, error) {
 	return managedWt{}, fmt.Errorf("no managed worktree matching %q (try `agentws list`)", target)
 }
 
-func removeWorktree(out io.Writer, wt managedWt, deleteBranch bool) error {
+// removeWorktree removes a managed worktree and, when asked, its local branch.
+// forceBranch overrides git's unmerged-branch guard, so only a caller that
+// proved the commits survive elsewhere may set it.
+func removeWorktree(out io.Writer, wt managedWt, deleteBranch, forceBranch bool) error {
 	if err := agent.GitWorktreeRemove(wt.srcDir, wt.path, true); err != nil {
 		return err
 	}
 	_, _ = fmt.Fprintf(out, "removed worktree %s\n", wt.path)
 	if deleteBranch {
-		if err := agent.GitDeleteBranch(wt.srcDir, wt.branch, true); err != nil {
+		if err := deleteLocalBranch(wt, forceBranch); err != nil {
 			return err
 		}
 		_, _ = fmt.Fprintf(out, "deleted branch %s\n", wt.branch)
@@ -334,6 +340,16 @@ func removeWorktree(out io.Writer, wt managedWt, deleteBranch bool) error {
 	return agent.GitWorktreePrune(wt.srcDir)
 }
 
+// deleteLocalBranch tries the guarded delete first so git refuses to drop
+// unmerged commits on its own; force is a fallback, never the first attempt.
+func deleteLocalBranch(wt managedWt, force bool) error {
+	err := agent.GitDeleteBranch(wt.srcDir, wt.branch, false)
+	if err == nil || !force {
+		return err
+	}
+	return agent.GitDeleteBranch(wt.srcDir, wt.branch, true)
+}
+
 // --- clean ----------------------------------------------------------------
 
 func newCleanCmd() *cobra.Command {
@@ -352,7 +368,7 @@ func newCleanCmd() *cobra.Command {
 				return nil
 			}
 			for _, w := range managed {
-				if err := removeWorktree(out, w, false); err != nil {
+				if err := removeWorktree(out, w, false, false); err != nil {
 					return err
 				}
 			}

cmd/agentws/prune.go

@@ -0,0 +1,295 @@
+package main
+
+import (
+	"errors"
+	"fmt"
+	"io"
+	"path/filepath"
+	"sort"
+
+	"git.unkin.net/unkin/agent-tools/internal/agent"
+
+	"github.com/spf13/cobra"
+)
+
+// Verdicts a worktree can be classified into.
+const (
+	verdictKeep         = "keep"
+	verdictRemove       = "remove"
+	verdictRemoveBranch = "remove+branch"
+)
+
+// prLister is the slice of the Gitea client prune needs, so tests can drive
+// classification without a live server.
+type prLister interface {
+	ListPRs(repoPath, state string) ([]agent.PullRequest, error)
+}
+
+type pruneResult struct {
+	wt      managedWt
+	verdict string
+	reason  string
+	// proven records that git itself confirmed the branch's commits survive
+	// elsewhere; only then may a branch delete override git's own guard.
+	proven bool
+}
+
+// repoCtx is the per-repo state classification is decided against.
+type repoCtx struct {
+	srcDir    string
+	defBranch string
+	prs       map[string]agent.PullRequest
+	prsKnown  bool
+	// fetched records that this run's pruning fetch succeeded; without it an
+	// origin/<branch> ref may be stale and due for deletion, so it proves nothing.
+	fetched bool
+}
+
+func newPruneCmd() *cobra.Command {
+	var apply, keepBranches bool
+	cmd := &cobra.Command{
+		Use:          "prune",
+		Short:        "Classify managed worktrees and remove the ones whose work is safely upstream",
+		Long:         "prune inspects every managed worktree, classifies it against git and its Gitea\npull request, and removes the ones whose work is provably upstream. It is a dry\nrun unless --yes is given.",
+		SilenceUsage: true,
+		RunE: func(cmd *cobra.Command, args []string) error {
+			return runPrune(cmd.OutOrStdout(), pruneClient(), apply, keepBranches)
+		},
+	}
+	f := cmd.Flags()
+	f.BoolVar(&apply, "yes", false, "Actually remove worktrees (default is a dry run)")
+	f.BoolVar(&keepBranches, "keep-branches", false, "Never delete a local branch, whatever the classification")
+	return cmd
+}
+
+// pruneClient builds a Gitea client, falling back to anonymous access when no
+// token can be minted; prune degrades to git-only signals if that fails too.
+func pruneClient() prLister {
+	tok, err := agent.GiteaToken()
+	if err != nil {
+		tok = ""
+	}
+	return agent.NewGiteaClient(tok)
+}
+
+func runPrune(out io.Writer, prs prLister, apply, keepBranches bool) error {
+	managed, err := managedWorktrees()
+	if err != nil {
+		return err
+	}
+	if len(managed) == 0 {
+		_, _ = fmt.Fprintln(out, "no managed worktrees")
+		return nil
+	}
+
+	byRepo := map[string][]managedWt{}
+	for _, w := range managed {
+		byRepo[w.srcDir] = append(byRepo[w.srcDir], w)
+	}
+	srcDirs := make([]string, 0, len(byRepo))
+	for dir := range byRepo {
+		srcDirs = append(srcDirs, dir)
+	}
+	sort.Strings(srcDirs)
+
+	var results []pruneResult
+	for _, srcDir := range srcDirs {
+		ctx, err := newRepoCtx(out, prs, srcDir)
+		if err != nil {
+			for _, w := range byRepo[srcDir] {
+				results = append(results, pruneResult{wt: w, verdict: verdictKeep, reason: "repo state unknown: " + err.Error()})
+			}
+			continue
+		}
+		for _, w := range byRepo[srcDir] {
+			res, err := classify(w, ctx)
+			if err != nil {
+				res = pruneResult{wt: w, verdict: verdictKeep, reason: "inspection failed: " + err.Error()}
+			}
+			results = append(results, res)
+		}
+	}
+
+	for _, r := range results {
+		_, _ = fmt.Fprintf(out, "%-44s %-34s %-14s %s\n", filepath.Base(r.wt.path), r.wt.branch, plannedVerdict(r, keepBranches), r.reason)
+	}
+	if !apply {
+		_, _ = fmt.Fprintln(out, "dry run: nothing removed (pass --yes to apply)")
+		return nil
+	}
+
+	var errs []error
+	for _, r := range results {
+		if r.verdict == verdictKeep {
+			continue
+		}
+		deleteBranch := r.verdict == verdictRemoveBranch && !keepBranches
+		if err := removeWorktree(out, r.wt, deleteBranch, r.proven); err != nil {
+			errs = append(errs, fmt.Errorf("%s: %w", r.wt.path, err))
+		}
+	}
+	return errors.Join(errs...)
+}
+
+// plannedVerdict is what will actually happen, so --keep-branches does not
+// print a branch deletion it will not perform.
+func plannedVerdict(r pruneResult, keepBranches bool) string {
+	if keepBranches && r.verdict == verdictRemoveBranch {
+		return verdictRemove
+	}
+	return r.verdict
+}
+
+// newRepoCtx refreshes a source repo and collects the signals prune classifies
+// against. A failed fetch or an unreachable Gitea is reported and tolerated:
+// the signals that hold offline still work, and the rest are recorded as
+// unverified.
+func newRepoCtx(out io.Writer, prs prLister, srcDir string) (repoCtx, error) {
+	ctx := repoCtx{srcDir: srcDir, prs: map[string]agent.PullRequest{}}
+	repo := filepath.Base(srcDir)
+	if err := agent.GitFetchPrune(srcDir, "origin", credentialHelperArgs()...); err != nil {
+		_, _ = fmt.Fprintf(out, "warn: fetch %s: %v (remote state unverified)\n", repo, err)
+	} else {
+		ctx.fetched = true
+	}
+	def, err := agent.GitRemoteDefaultBranch(srcDir, "origin")
+	if err != nil {
+		return repoCtx{}, err
+	}
+	ctx.defBranch = def
+
+	if prs == nil {
+		return ctx, nil
+	}
+	list, err := prs.ListPRs(repoPath(srcDir, repo), "all")
+	switch {
+	case errors.Is(err, agent.ErrPRListTruncated):
+		// A branch missing from a partial listing must not read as "no PR".
+		_, _ = fmt.Fprintf(out, "warn: list PRs for %s: %v (older PRs unseen)\n", repo, err)
+		ctx.prs = prsByBranch(list)
+	case err != nil:
+		_, _ = fmt.Fprintf(out, "warn: list PRs for %s: %v (git signals only)\n", repo, err)
+	default:
+		ctx.prs = prsByBranch(list)
+		ctx.prsKnown = true
+	}
+	return ctx, nil
+}
+
+// repoPath is the Gitea "owner/repo" for a checkout, read from origin's URL
+// because not every managed repo lives under AGENTWS_OWNER.
+func repoPath(srcDir, repo string) string {
+	url, err := agent.GitRemoteURL(srcDir, "origin")
+	if err == nil && agent.RemoteHost(url) == giteaHost() {
+		if path, err := agent.RepoPathFromRemoteURL(url); err == nil {
+			return path
+		}
+	}
+	return owner() + "/" + repo
+}
+
+// prsByBranch indexes PRs by head branch, preferring an open PR and otherwise
+// the most recent one when a branch has been used more than once.
+func prsByBranch(list []agent.PullRequest) map[string]agent.PullRequest {
+	out := map[string]agent.PullRequest{}
+	for _, pr := range list {
+		branch := agent.PRHeadBranch(pr)
+		if branch == "" {
+			continue
+		}
+		if cur, ok := out[branch]; ok && !supersedes(pr, cur) {
+			continue
+		}
+		out[branch] = pr
+	}
+	return out
+}
+
+func supersedes(a, b agent.PullRequest) bool {
+	if a.IsOpen() != b.IsOpen() {
+		return a.IsOpen()
+	}
+	if a.Merged != b.Merged {
+		return a.Merged
+	}
+	return a.Number > b.Number
+}
+
+// headContainedIn reports whether the worktree's HEAD is reachable from ref. A
+// ref that cannot be resolved proves nothing, so it reads as not contained.
+func headContainedIn(dir, ref string) bool {
+	if ref == "" {
+		return false
+	}
+	ok, err := agent.GitIsAncestor(dir, "HEAD", ref)
+	return err == nil && ok
+}
+
+// classify applies the prune precedence: dirty and open-PR worktrees are kept,
+// provably-upstream work loses its branch too, and anything unproven keeps its
+// branch so no commits become unreachable. A PR's state alone never authorises
+// deleting a branch — git must confirm HEAD is contained in what merged or in
+// what origin still holds, and origin's refs only count when this run's pruning
+// fetch refreshed them.
+func classify(wt managedWt, ctx repoCtx) (pruneResult, error) {
+	res := pruneResult{wt: wt}
+	dirty, err := agent.GitIsDirty(wt.path)
+	if err != nil {
+		return res, err
+	}
+	if dirty {
+		res.verdict, res.reason = verdictKeep, "dirty"
+		return res, nil
+	}
+
+	pr, hasPR := ctx.prs[wt.branch]
+	if hasPR && pr.IsOpen() {
+		res.verdict, res.reason = verdictKeep, fmt.Sprintf("PR open #%d", pr.Number)
+		return res, nil
+	}
+
+	upstream := "origin/" + ctx.defBranch
+	contained, err := agent.GitIsAncestor(wt.path, "HEAD", upstream)
+	if err != nil {
+		return res, err
+	}
+	if contained {
+		res.verdict, res.reason, res.proven = verdictRemoveBranch, "contained in "+upstream, true
+		return res, nil
+	}
+	unmerged, err := agent.GitUnmergedCommits(wt.path, upstream, "HEAD")
+	if err != nil {
+		return res, err
+	}
+	if unmerged == 0 {
+		// git cherry proves the patches reached that history, not that they stand at its tip.
+		res.verdict, res.reason, res.proven = verdictRemoveBranch, "patch-equivalent commits in "+upstream+" history", true
+		return res, nil
+	}
+
+	remote := "origin/" + wt.branch
+	onOrigin := hasPR && ctx.fetched && agent.GitRemoteBranchExists(ctx.srcDir, "origin", wt.branch)
+	switch {
+	case hasPR && pr.Merged && headContainedIn(wt.path, pr.Head.Sha):
+		res.verdict, res.reason, res.proven = verdictRemoveBranch, fmt.Sprintf("PR merged #%d, HEAD contained in the merged head", pr.Number), true
+	case hasPR && pr.Merged && onOrigin && headContainedIn(wt.path, remote):
+		res.verdict, res.reason, res.proven = verdictRemoveBranch, fmt.Sprintf("PR merged #%d, HEAD contained in %s", pr.Number, remote), true
+	case hasPR && pr.Merged && !ctx.fetched:
+		res.verdict, res.reason = verdictRemove, fmt.Sprintf("PR merged #%d, fetch failed so %s is unverified", pr.Number, remote)
+	case hasPR && pr.Merged:
+		res.verdict, res.reason = verdictRemove, fmt.Sprintf("PR merged #%d, local commits not in the merged head", pr.Number)
+	case hasPR && onOrigin && headContainedIn(wt.path, remote):
+		res.verdict, res.reason, res.proven = verdictRemoveBranch, fmt.Sprintf("PR closed #%d, HEAD contained in %s", pr.Number, remote), true
+	case hasPR && onOrigin:
+		res.verdict, res.reason = verdictRemove, fmt.Sprintf("PR closed #%d, local commits not on %s", pr.Number, remote)
+	case hasPR && !ctx.fetched:
+		res.verdict, res.reason = verdictRemove, fmt.Sprintf("PR closed #%d, fetch failed so %s is unverified", pr.Number, remote)
+	case hasPR:
+		res.verdict, res.reason = verdictRemove, fmt.Sprintf("PR closed #%d, branch gone", pr.Number)
+	case ctx.prsKnown:
+		res.verdict, res.reason = verdictRemove, "no PR"
+	default:
+		res.verdict, res.reason = verdictRemove, "PR state unknown"
+	}
+	return res, nil
+}

cmd/agentws/prune_test.go

@@ -0,0 +1,687 @@
+package main
+
+import (
+	"bytes"
+	"encoding/json"
+	"fmt"
+	"net/http"
+	"net/http/httptest"
+	"os"
+	"os/exec"
+	"path/filepath"
+	"strconv"
+	"strings"
+	"testing"
+
+	"git.unkin.net/unkin/agent-tools/internal/agent"
+)
+
+// fixture is a bare origin plus a source checkout named "repo" and a worktree
+// root, wired so managedWorktrees() finds the worktrees created here.
+type fixture struct {
+	root   string
+	bare   string
+	srcDir string
+	wtRoot string
+}
+
+func git(t *testing.T, dir string, args ...string) string {
+	t.Helper()
+	cmd := exec.Command("git", args...)
+	cmd.Dir = dir
+	out, err := cmd.CombinedOutput()
+	if err != nil {
+		t.Fatalf("git %s (in %s): %v: %s", strings.Join(args, " "), dir, err, out)
+	}
+	return strings.TrimSpace(string(out))
+}
+
+func newFixture(t *testing.T) *fixture {
+	t.Helper()
+	root := t.TempDir()
+	f := &fixture{
+		root:   root,
+		bare:   filepath.Join(root, "origin.git"),
+		srcDir: filepath.Join(root, "src", "repo"),
+		wtRoot: filepath.Join(root, "worktrees"),
+	}
+	if err := os.MkdirAll(filepath.Join(root, "src"), 0o755); err != nil {
+		t.Fatal(err)
+	}
+	if err := os.MkdirAll(f.wtRoot, 0o755); err != nil {
+		t.Fatal(err)
+	}
+	git(t, root, "init", "--bare", "-b", "main", f.bare)
+
+	seed := filepath.Join(root, "seed")
+	git(t, root, "init", "-b", "main", seed)
+	identity(t, seed)
+	writeCommit(t, seed, "README.md", "hi\n", "init")
+	git(t, seed, "remote", "add", "origin", f.bare)
+	git(t, seed, "push", "-u", "origin", "main")
+
+	git(t, filepath.Join(root, "src"), "clone", f.bare, f.srcDir)
+	identity(t, f.srcDir)
+
+	t.Setenv("AGENTWS_ROOT", f.wtRoot)
+	t.Setenv("AGENTWS_SRC_ROOT", filepath.Join(root, "src"))
+	t.Setenv("AGENTWS_OWNER", "unkin")
+	return f
+}
+
+func identity(t *testing.T, dir string) {
+	t.Helper()
+	git(t, dir, "config", "user.email", "test@example.com")
+	git(t, dir, "config", "user.name", "Test")
+}
+
+func writeCommit(t *testing.T, dir, name, content, msg string) {
+	t.Helper()
+	if err := os.WriteFile(filepath.Join(dir, name), []byte(content), 0o644); err != nil {
+		t.Fatal(err)
+	}
+	git(t, dir, "add", ".")
+	git(t, dir, "commit", "-m", msg)
+}
+
+// addWorktree creates a managed worktree for branch and returns its path.
+func (f *fixture) addWorktree(t *testing.T, branch string) string {
+	t.Helper()
+	path := filepath.Join(f.wtRoot, agent.WorktreeDirName("repo", branch))
+	git(t, f.srcDir, "worktree", "add", path, "-b", branch, "origin/main")
+	identity(t, path)
+	return path
+}
+
+// landUpstream commits content on origin's main, mimicking a squash merge: the
+// same patch arrives upstream under a different SHA.
+func (f *fixture) landUpstream(t *testing.T, name, content, msg string) {
+	t.Helper()
+	seed := filepath.Join(f.root, "seed")
+	git(t, seed, "pull", "--ff-only", "origin", "main")
+	writeCommit(t, seed, name, content, msg)
+	git(t, seed, "push", "origin", "main")
+}
+
+// fakeGitea serves the pulls listing for unkin/repo with the given PR bodies.
+func fakeGitea(t *testing.T, prs ...map[string]any) *httptest.Server {
+	t.Helper()
+	mux := http.NewServeMux()
+	mux.HandleFunc("/api/v1/repos/unkin/repo/pulls", func(w http.ResponseWriter, r *http.Request) {
+		if r.URL.Query().Get("page") != "1" {
+			_, _ = w.Write([]byte("[]"))
+			return
+		}
+		body, err := json.Marshal(prs)
+		if err != nil {
+			t.Errorf("marshal PRs: %v", err)
+		}
+		_, _ = w.Write(body)
+	})
+	srv := httptest.NewServer(mux)
+	t.Cleanup(srv.Close)
+	return srv
+}
+
+func client(srv *httptest.Server) prLister {
+	return &agent.GiteaClient{BaseURL: srv.URL, HTTP: srv.Client()}
+}
+
+// mergedPR mimics Gitea after a merge: the branch is deleted, so head.ref
+// becomes refs/pull/<n>/head and only head.label still names the branch.
+func mergedPR(number int, branch string) map[string]any {
+	return map[string]any{
+		"number": number,
+		"state":  "closed",
+		"merged": true,
+		"head": map[string]any{
+			"ref":   "refs/pull/" + strconv.Itoa(number) + "/head",
+			"label": branch,
+		},
+	}
+}
+
+func withHeadSha(pr map[string]any, sha string) map[string]any {
+	pr["head"].(map[string]any)["sha"] = sha
+	return pr
+}
+
+func openPR(number int, branch string) map[string]any {
+	return map[string]any{
+		"number": number,
+		"state":  "open",
+		"merged": false,
+		"head":   map[string]any{"ref": branch, "label": branch},
+	}
+}
+
+func closedPR(number int, branch string) map[string]any {
+	return map[string]any{
+		"number": number,
+		"state":  "closed",
+		"merged": false,
+		"head":   map[string]any{"ref": branch, "label": branch},
+	}
+}
+
+func run(t *testing.T, prs prLister, apply, keepBranches bool) string {
+	t.Helper()
+	var out bytes.Buffer
+	if err := runPrune(&out, prs, apply, keepBranches); err != nil {
+		t.Fatalf("runPrune: %v\n%s", err, out.String())
+	}
+	return out.String()
+}
+
+func lineFor(t *testing.T, out, branch string) string {
+	t.Helper()
+	for _, line := range strings.Split(out, "\n") {
+		if strings.Contains(line, " "+branch+" ") {
+			return line
+		}
+	}
+	t.Fatalf("no line for branch %q in:\n%s", branch, out)
+	return ""
+}
+
+func assertVerdict(t *testing.T, out, branch, verdict, reason string) {
+	t.Helper()
+	line := lineFor(t, out, branch)
+	fields := strings.Fields(line)
+	if len(fields) < 3 || fields[2] != verdict {
+		t.Errorf("branch %s: verdict line %q, want verdict %q", branch, line, verdict)
+	}
+	if reason != "" && !strings.Contains(line, reason) {
+		t.Errorf("branch %s: line %q, want reason containing %q", branch, line, reason)
+	}
+}
+
+func exists(path string) bool {
+	_, err := os.Stat(path)
+	return err == nil
+}
+
+// A merged PR's branch is deleted on merge, so its head.ref reads
+// refs/pull/<n>/head; classification must still see the merge (via head.label)
+// and remove the branch, not fall through to "no PR".
+func TestPruneMergedPRWithDeletedBranch(t *testing.T) {
+	f := newFixture(t)
+	wt := f.addWorktree(t, "benvin/merged")
+	writeCommit(t, wt, "m.txt", "m\n", "work")
+	head := git(t, wt, "rev-parse", "HEAD")
+
+	srv := fakeGitea(t, withHeadSha(mergedPR(3, "benvin/merged"), head))
+	out := run(t, client(srv), true, false)
+	assertVerdict(t, out, "benvin/merged", verdictRemoveBranch, "PR merged #3")
+
+	if exists(wt) {
+		t.Errorf("worktree %s should have been removed", wt)
+	}
+	if agent.GitBranchExists(f.srcDir, "benvin/merged") {
+		t.Error("branch of a merged PR should be deleted")
+	}
+}
+
+// The same PR without head.label: matching falls back to head.ref, which no
+// longer names the branch, so prune must not guess it is merged — the worktree
+// goes but the branch stays.
+func TestPruneMergedPRWithoutLabelKeepsBranch(t *testing.T) {
+	f := newFixture(t)
+	wt := f.addWorktree(t, "benvin/merged")
+	writeCommit(t, wt, "m.txt", "m\n", "work")
+
+	pr := mergedPR(3, "benvin/merged")
+	pr["head"].(map[string]any)["label"] = ""
+	srv := fakeGitea(t, pr)
+	out := run(t, client(srv), true, false)
+	assertVerdict(t, out, "benvin/merged", verdictRemove, "no PR")
+
+	if !agent.GitBranchExists(f.srcDir, "benvin/merged") {
+		t.Error("branch must survive when the PR could not be matched")
+	}
+}
+
+func TestPruneContainedBranch(t *testing.T) {
+	f := newFixture(t)
+	wt := f.addWorktree(t, "benvin/contained")
+
+	srv := fakeGitea(t)
+	out := run(t, client(srv), true, false)
+	assertVerdict(t, out, "benvin/contained", verdictRemoveBranch, "contained in origin/main")
+
+	if exists(wt) {
+		t.Error("contained worktree should have been removed")
+	}
+	if agent.GitBranchExists(f.srcDir, "benvin/contained") {
+		t.Error("contained branch should be deleted")
+	}
+}
+
+// Squash-merged work keeps a local SHA that is not upstream, so only the
+// patch-equivalence check proves it landed.
+func TestPruneCherryCleanBranch(t *testing.T) {
+	f := newFixture(t)
+	wt := f.addWorktree(t, "benvin/squashed")
+	writeCommit(t, wt, "s.txt", "same\n", "add s")
+	f.landUpstream(t, "s.txt", "same\n", "squashed s")
+
+	srv := fakeGitea(t)
+	out := run(t, client(srv), true, false)
+	assertVerdict(t, out, "benvin/squashed", verdictRemoveBranch, "patch-equivalent commits in origin/main history")
+
+	if agent.GitBranchExists(f.srcDir, "benvin/squashed") {
+		t.Error("cherry-clean branch should be deleted")
+	}
+}
+
+// Uncommitted work outranks every other signal, including a branch that is
+// otherwise fully contained upstream.
+func TestPruneNeverTouchesDirtyWorktree(t *testing.T) {
+	f := newFixture(t)
+	wt := f.addWorktree(t, "benvin/dirty")
+	if err := os.WriteFile(filepath.Join(wt, "wip.txt"), []byte("wip\n"), 0o644); err != nil {
+		t.Fatal(err)
+	}
+
+	srv := fakeGitea(t, mergedPR(4, "benvin/dirty"))
+	out := run(t, client(srv), true, false)
+	assertVerdict(t, out, "benvin/dirty", verdictKeep, "dirty")
+
+	if !exists(wt) {
+		t.Error("dirty worktree must not be removed")
+	}
+	if !agent.GitBranchExists(f.srcDir, "benvin/dirty") {
+		t.Error("dirty worktree's branch must survive")
+	}
+}
+
+// An open PR is kept even when its commits are already upstream.
+func TestPruneKeepsOpenPR(t *testing.T) {
+	f := newFixture(t)
+	wt := f.addWorktree(t, "benvin/open")
+
+	srv := fakeGitea(t, openPR(5, "benvin/open"))
+	out := run(t, client(srv), true, false)
+	assertVerdict(t, out, "benvin/open", verdictKeep, "PR open #5")
+
+	if !exists(wt) {
+		t.Error("worktree with an open PR must not be removed")
+	}
+	if !agent.GitBranchExists(f.srcDir, "benvin/open") {
+		t.Error("branch with an open PR must not be deleted")
+	}
+}
+
+// Closed-unmerged with the branch still on origin: the work is not lost, so the
+// local branch goes too.
+func TestPruneClosedPRWithBranchOnOrigin(t *testing.T) {
+	f := newFixture(t)
+	wt := f.addWorktree(t, "benvin/closed")
+	writeCommit(t, wt, "c.txt", "c\n", "work")
+	git(t, wt, "push", "origin", "benvin/closed")
+
+	srv := fakeGitea(t, closedPR(6, "benvin/closed"))
+	out := run(t, client(srv), true, false)
+	assertVerdict(t, out, "benvin/closed", verdictRemoveBranch, "PR closed #6, HEAD contained in origin/benvin/closed")
+
+	if exists(wt) {
+		t.Error("worktree should have been removed")
+	}
+	if agent.GitBranchExists(f.srcDir, "benvin/closed") {
+		t.Error("branch should be deleted while origin still has it")
+	}
+}
+
+// Closed-unmerged with nothing on origin: the commits exist only here, so the
+// branch is kept and only the worktree goes.
+func TestPruneClosedPRWithBranchGone(t *testing.T) {
+	f := newFixture(t)
+	wt := f.addWorktree(t, "benvin/orphan")
+	writeCommit(t, wt, "o.txt", "o\n", "work")
+
+	srv := fakeGitea(t, closedPR(7, "benvin/orphan"))
+	out := run(t, client(srv), true, false)
+	assertVerdict(t, out, "benvin/orphan", verdictRemove, "PR closed #7, branch gone")
+
+	if exists(wt) {
+		t.Error("worktree should have been removed")
+	}
+	if !agent.GitBranchExists(f.srcDir, "benvin/orphan") {
+		t.Error("branch must survive when origin does not have the commits")
+	}
+}
+
+func TestPruneNoPRKeepsBranch(t *testing.T) {
+	f := newFixture(t)
+	wt := f.addWorktree(t, "benvin/unpushed")
+	writeCommit(t, wt, "u.txt", "u\n", "work")
+
+	srv := fakeGitea(t)
+	out := run(t, client(srv), true, false)
+	assertVerdict(t, out, "benvin/unpushed", verdictRemove, "no PR")
+
+	if exists(wt) {
+		t.Error("worktree should have been removed")
+	}
+	if !agent.GitBranchExists(f.srcDir, "benvin/unpushed") {
+		t.Error("branch with unproven work must survive")
+	}
+}
+
+// The default run reports and changes nothing.
+func TestPruneDryRunChangesNothing(t *testing.T) {
+	f := newFixture(t)
+	contained := f.addWorktree(t, "benvin/contained")
+	merged := f.addWorktree(t, "benvin/merged")
+	writeCommit(t, merged, "m.txt", "m\n", "work")
+	head := git(t, merged, "rev-parse", "HEAD")
+
+	srv := fakeGitea(t, withHeadSha(mergedPR(8, "benvin/merged"), head))
+	out := run(t, client(srv), false, false)
+
+	if !strings.Contains(out, "dry run") {
+		t.Errorf("dry-run output should say so:\n%s", out)
+	}
+	assertVerdict(t, out, "benvin/contained", verdictRemoveBranch, "contained")
+	assertVerdict(t, out, "benvin/merged", verdictRemoveBranch, "PR merged #8")
+	if !exists(contained) || !exists(merged) {
+		t.Error("dry run must not remove worktrees")
+	}
+	if !agent.GitBranchExists(f.srcDir, "benvin/contained") || !agent.GitBranchExists(f.srcDir, "benvin/merged") {
+		t.Error("dry run must not delete branches")
+	}
+}
+
+// --keep-branches removes worktrees but leaves every branch alone, and the
+// printed verdict says so.
+func TestPruneKeepBranches(t *testing.T) {
+	f := newFixture(t)
+	wt := f.addWorktree(t, "benvin/contained")
+
+	srv := fakeGitea(t)
+	out := run(t, client(srv), true, true)
+	assertVerdict(t, out, "benvin/contained", verdictRemove, "contained")
+	if strings.Contains(out, verdictRemoveBranch) {
+		t.Errorf("--keep-branches must not print a branch-deleting verdict:\n%s", out)
+	}
+
+	if exists(wt) {
+		t.Error("worktree should have been removed")
+	}
+	if !agent.GitBranchExists(f.srcDir, "benvin/contained") {
+		t.Error("--keep-branches must not delete the branch")
+	}
+}
+
+// With Gitea unreachable prune falls back to the git signals: provably-upstream
+// work is still cleaned up, and anything unproven keeps its branch.
+func TestPruneDegradesWhenGiteaUnreachable(t *testing.T) {
+	f := newFixture(t)
+	contained := f.addWorktree(t, "benvin/contained")
+	unproven := f.addWorktree(t, "benvin/unproven")
+	writeCommit(t, unproven, "u.txt", "u\n", "work")
+
+	dead := httptest.NewServer(http.NewServeMux())
+	c := &agent.GiteaClient{BaseURL: dead.URL, HTTP: dead.Client()}
+	dead.Close()
+
+	out := run(t, c, true, false)
+	if !strings.Contains(out, "git signals only") {
+		t.Errorf("output should note the Gitea failure:\n%s", out)
+	}
+	assertVerdict(t, out, "benvin/contained", verdictRemoveBranch, "contained")
+	assertVerdict(t, out, "benvin/unproven", verdictRemove, "PR state unknown")
+
+	if exists(contained) || exists(unproven) {
+		t.Error("both worktrees should have been removed")
+	}
+	if agent.GitBranchExists(f.srcDir, "benvin/contained") {
+		t.Error("contained branch is safe to delete without Gitea")
+	}
+	if !agent.GitBranchExists(f.srcDir, "benvin/unproven") {
+		t.Error("unproven branch must survive an unreachable Gitea")
+	}
+}
+
+func TestPruneNoWorktrees(t *testing.T) {
+	newFixture(t)
+	srv := fakeGitea(t)
+	if out := run(t, client(srv), true, false); !strings.Contains(out, "no managed worktrees") {
+		t.Errorf("output = %q", out)
+	}
+}
+
+// Commits made after the PR merged exist nowhere else, so a merged PR alone
+// must not authorise deleting the branch.
+func TestPruneMergedPRWithCommitsAfterMerge(t *testing.T) {
+	f := newFixture(t)
+	wt := f.addWorktree(t, "benvin/ahead")
+	writeCommit(t, wt, "a.txt", "a\n", "merged work")
+	merged := git(t, wt, "rev-parse", "HEAD")
+	writeCommit(t, wt, "b.txt", "b\n", "work after the merge")
+
+	srv := fakeGitea(t, withHeadSha(mergedPR(9, "benvin/ahead"), merged))
+	out := run(t, client(srv), true, false)
+	assertVerdict(t, out, "benvin/ahead", verdictRemove, "PR merged #9")
+
+	if exists(wt) {
+		t.Error("worktree should have been removed")
+	}
+	if !agent.GitBranchExists(f.srcDir, "benvin/ahead") {
+		t.Error("branch with commits beyond the merged head must survive")
+	}
+}
+
+// HEAD proven contained in the merged head still loses its branch.
+func TestPruneMergedPRContainedInMergedHead(t *testing.T) {
+	f := newFixture(t)
+	wt := f.addWorktree(t, "benvin/landed")
+	writeCommit(t, wt, "a.txt", "a\n", "work")
+	head := git(t, wt, "rev-parse", "HEAD")
+
+	srv := fakeGitea(t, withHeadSha(mergedPR(10, "benvin/landed"), head))
+	out := run(t, client(srv), true, false)
+	assertVerdict(t, out, "benvin/landed", verdictRemoveBranch, "HEAD contained in the merged head")
+
+	if exists(wt) {
+		t.Error("worktree should have been removed")
+	}
+	if agent.GitBranchExists(f.srcDir, "benvin/landed") {
+		t.Error("branch contained in the merged head should be deleted")
+	}
+}
+
+// A surviving remote branch only covers what was pushed to it; later local
+// commits keep the branch.
+func TestPruneClosedPRWithCommitsBeyondOrigin(t *testing.T) {
+	f := newFixture(t)
+	wt := f.addWorktree(t, "benvin/beyond")
+	writeCommit(t, wt, "c.txt", "c\n", "pushed work")
+	git(t, wt, "push", "origin", "benvin/beyond")
+	writeCommit(t, wt, "d.txt", "d\n", "local only")
+
+	srv := fakeGitea(t, closedPR(11, "benvin/beyond"))
+	out := run(t, client(srv), true, false)
+	assertVerdict(t, out, "benvin/beyond", verdictRemove, "local commits not on origin/benvin/beyond")
+
+	if exists(wt) {
+		t.Error("worktree should have been removed")
+	}
+	if !agent.GitBranchExists(f.srcDir, "benvin/beyond") {
+		t.Error("branch with commits beyond origin must survive")
+	}
+}
+
+// truncatedLister stands in for a repo with more PRs than the listing cap.
+type truncatedLister struct{ prs []agent.PullRequest }
+
+func (l truncatedLister) ListPRs(string, string) ([]agent.PullRequest, error) {
+	return l.prs, fmt.Errorf("unkin/repo: %w after 1000 pull requests", agent.ErrPRListTruncated)
+}
+
+// A truncated listing still classifies the PRs it saw, but a branch missing
+// from it reads as unknown rather than as having no PR.
+func TestPruneWarnsOnTruncatedPRListing(t *testing.T) {
+	f := newFixture(t)
+	listed := f.addWorktree(t, "benvin/listed")
+	writeCommit(t, listed, "a.txt", "a\n", "work")
+	head := git(t, listed, "rev-parse", "HEAD")
+	unlisted := f.addWorktree(t, "benvin/unlisted")
+	writeCommit(t, unlisted, "b.txt", "b\n", "work")
+
+	var pr agent.PullRequest
+	pr.Number, pr.State, pr.Merged = 12, "closed", true
+	pr.Head.Label, pr.Head.Sha = "benvin/listed", head
+
+	out := run(t, truncatedLister{prs: []agent.PullRequest{pr}}, true, false)
+	if !strings.Contains(out, "truncated") || !strings.Contains(out, "older PRs unseen") {
+		t.Errorf("output should warn about the truncated listing:\n%s", out)
+	}
+	assertVerdict(t, out, "benvin/listed", verdictRemoveBranch, "PR merged #12")
+	assertVerdict(t, out, "benvin/unlisted", verdictRemove, "PR state unknown")
+
+	if agent.GitBranchExists(f.srcDir, "benvin/listed") {
+		t.Error("branch of a merged PR seen in the listing should be deleted")
+	}
+	if !agent.GitBranchExists(f.srcDir, "benvin/unlisted") {
+		t.Error("branch missing from a truncated listing must survive")
+	}
+}
+
+// breakRemote points origin at a path that does not exist, so every fetch fails.
+func (f *fixture) breakRemote(t *testing.T) {
+	t.Helper()
+	git(t, f.srcDir, "remote", "set-url", "origin", filepath.Join(f.root, "missing.git"))
+}
+
+// applyUnreachable applies the plan against a repo whose remote is unreachable.
+// The post-removal refresh fetch fails, so runPrune must report an error; the
+// classification and the removals it authorised happen regardless.
+func applyUnreachable(t *testing.T, prs prLister) string {
+	t.Helper()
+	var out bytes.Buffer
+	err := runPrune(&out, prs, true, false)
+	if err == nil {
+		t.Fatalf("expected the refresh fetch to fail against a missing remote:\n%s", out.String())
+	}
+	return out.String()
+}
+
+// staleFixture builds a repo where origin/<branch> covers HEAD for a closed and
+// a merged PR. Deleting the branches on origin makes those tracking refs stale:
+// a pruning fetch would drop them, so they only prove anything while a fetch
+// this run confirms they are still there.
+func staleFixture(t *testing.T, deleteUpstream bool) (*fixture, *httptest.Server, string, string) {
+	t.Helper()
+	f := newFixture(t)
+	closed := f.addWorktree(t, "benvin/stale-closed")
+	writeCommit(t, closed, "c.txt", "c\n", "work")
+	git(t, closed, "push", "origin", "benvin/stale-closed")
+
+	merged := f.addWorktree(t, "benvin/stale-merged")
+	writeCommit(t, merged, "m.txt", "m\n", "work")
+	git(t, merged, "push", "origin", "benvin/stale-merged")
+
+	if deleteUpstream {
+		git(t, f.bare, "update-ref", "-d", "refs/heads/benvin/stale-closed")
+		git(t, f.bare, "update-ref", "-d", "refs/heads/benvin/stale-merged")
+	}
+
+	// The merged head is an older commit, so only origin/<branch> covers HEAD.
+	base := git(t, f.srcDir, "rev-parse", "origin/main")
+	srv := fakeGitea(t,
+		closedPR(20, "benvin/stale-closed"),
+		withHeadSha(mergedPR(21, "benvin/stale-merged"), base),
+	)
+	return f, srv, closed, merged
+}
+
+// A tracking ref this run's fetch could not confirm is not evidence: the branch
+// may already be gone upstream, and the next successful --prune deletes the ref.
+// Both worktrees go, both branches stay.
+func TestPruneFailedFetchDistrustsStaleRemoteBranch(t *testing.T) {
+	f, srv, closed, merged := staleFixture(t, true)
+	f.breakRemote(t)
+
+	out := run(t, client(srv), false, false)
+	if !strings.Contains(out, "remote state unverified") {
+		t.Errorf("output should report the failed fetch:\n%s", out)
+	}
+	assertVerdict(t, out, "benvin/stale-closed", verdictRemove, "PR closed #20, fetch failed so origin/benvin/stale-closed is unverified")
+	assertVerdict(t, out, "benvin/stale-merged", verdictRemove, "PR merged #21, fetch failed so origin/benvin/stale-merged is unverified")
+	for _, b := range []string{"benvin/stale-closed", "benvin/stale-merged"} {
+		if !agent.GitRemoteBranchExists(f.srcDir, "origin", b) {
+			t.Fatalf("fixture: origin/%s should still be present as a stale ref", b)
+		}
+	}
+
+	applyUnreachable(t, client(srv))
+	if exists(closed) || exists(merged) {
+		t.Error("both worktrees should have been removed")
+	}
+	for _, b := range []string{"benvin/stale-closed", "benvin/stale-merged"} {
+		if !agent.GitBranchExists(f.srcDir, b) {
+			t.Errorf("branch %s must survive an unverified origin", b)
+		}
+	}
+}
+
+// The control for the case above: with the fetch working and the branches still
+// on origin, the same shape still loses both branches.
+func TestPruneSuccessfulFetchTrustsRemoteBranch(t *testing.T) {
+	f, srv, closed, merged := staleFixture(t, false)
+
+	out := run(t, client(srv), true, false)
+	assertVerdict(t, out, "benvin/stale-closed", verdictRemoveBranch, "PR closed #20, HEAD contained in origin/benvin/stale-closed")
+	assertVerdict(t, out, "benvin/stale-merged", verdictRemoveBranch, "PR merged #21, HEAD contained in origin/benvin/stale-merged")
+
+	if exists(closed) || exists(merged) {
+		t.Error("both worktrees should have been removed")
+	}
+	for _, b := range []string{"benvin/stale-closed", "benvin/stale-merged"} {
+		if agent.GitBranchExists(f.srcDir, b) {
+			t.Errorf("branch %s should be deleted while origin still has it", b)
+		}
+	}
+}
+
+// Proofs that read only local objects and the merged head SHA from the API do
+// not depend on the fetch, so a failed fetch must not suppress them.
+func TestPruneFailedFetchKeepsFetchIndependentProofs(t *testing.T) {
+	f := newFixture(t)
+	contained := f.addWorktree(t, "benvin/contained")
+	landed := f.addWorktree(t, "benvin/landed")
+	writeCommit(t, landed, "a.txt", "a\n", "work")
+	head := git(t, landed, "rev-parse", "HEAD")
+	f.breakRemote(t)
+
+	srv := fakeGitea(t, withHeadSha(mergedPR(22, "benvin/landed"), head))
+	out := applyUnreachable(t, client(srv))
+	assertVerdict(t, out, "benvin/contained", verdictRemoveBranch, "contained in origin/main")
+	assertVerdict(t, out, "benvin/landed", verdictRemoveBranch, "PR merged #22, HEAD contained in the merged head")
+
+	if exists(contained) || exists(landed) {
+		t.Error("both worktrees should have been removed")
+	}
+	for _, b := range []string{"benvin/contained", "benvin/landed"} {
+		if agent.GitBranchExists(f.srcDir, b) {
+			t.Errorf("branch %s is proven without the fetch and should be deleted", b)
+		}
+	}
+}
+
+// Managed repos are not all under AGENTWS_OWNER, so the Gitea path comes from
+// origin's URL; a non-Gitea remote falls back to the configured owner.
+func TestRepoPathFollowsOrigin(t *testing.T) {
+	t.Setenv("AGENTWS_OWNER", "unkin")
+	dir := t.TempDir()
+	git(t, dir, "init", "-b", "main", ".")
+	git(t, dir, "remote", "add", "origin", "https://git.unkin.net/unkinben/dotfiles.git")
+	if got := repoPath(dir, "dotfiles"); got != "unkinben/dotfiles" {
+		t.Errorf("repoPath = %q, want unkinben/dotfiles", got)
+	}
+	git(t, dir, "remote", "set-url", "origin", filepath.Join(dir, "origin.git"))
+	if got := repoPath(dir, "dotfiles"); got != "unkin/dotfiles" {
+		t.Errorf("repoPath for a local remote = %q, want unkin/dotfiles", got)
+	}
+}

internal/agent/client_test.go

@@ -3,6 +3,7 @@ package agent
 import (
 	"encoding/json"
 	"errors"
+	"fmt"
 	"io"
 	"net/http"
 	"net/http/httptest"
@@ -219,6 +220,100 @@ func TestFetchStateFailsOnNon404StatusError(t *testing.T) {
 	}
 }
 
+// Gitea rewrites head.ref to "refs/pull/<n>/head" once the PR's branch is
+// deleted, which merging does in these repos. Matching a branch against
+// head.ref alone therefore finds nothing for every merged PR; head.label keeps
+// the original name.
+func TestPRHeadBranch(t *testing.T) {
+	tests := []struct {
+		name       string
+		ref, label string
+		want       string
+	}{
+		{"merged, branch deleted", "refs/pull/12/head", "benvin/merged", "benvin/merged"},
+		{"open PR", "benvin/open", "benvin/open", "benvin/open"},
+		{"fully qualified ref", "refs/heads/benvin/x", "", "benvin/x"},
+		{"no label falls back to ref", "benvin/y", "", "benvin/y"},
+		{"cross-repo label", "benvin/z", "someone:benvin/z", "benvin/z"},
+		{"nothing usable", "refs/pull/12/head", "", ""},
+	}
+	for _, tt := range tests {
+		t.Run(tt.name, func(t *testing.T) {
+			var pr PullRequest
+			pr.Head.Ref = tt.ref
+			pr.Head.Label = tt.label
+			if got := PRHeadBranch(pr); got != tt.want {
+				t.Errorf("PRHeadBranch(ref=%q,label=%q) = %q, want %q", tt.ref, tt.label, got, tt.want)
+			}
+		})
+	}
+}
+
+func TestListPRsPaginates(t *testing.T) {
+	var pages []string
+	mux := http.NewServeMux()
+	mux.HandleFunc("/api/v1/repos/unkin/repo/pulls", func(w http.ResponseWriter, r *http.Request) {
+		q := r.URL.Query()
+		pages = append(pages, q.Get("page"))
+		if q.Get("state") != "all" {
+			t.Errorf("state = %q, want all", q.Get("state"))
+		}
+		if q.Get("page") == "1" {
+			full := make([]string, 0, prPageSize)
+			for i := 0; i < prPageSize; i++ {
+				full = append(full, fmt.Sprintf(`{"number":%d,"state":"closed","merged":true,"head":{"ref":"refs/pull/%d/head","label":"benvin/b%d"}}`, i+1, i+1, i+1))
+			}
+			_, _ = io.WriteString(w, "["+strings.Join(full, ",")+"]")
+			return
+		}
+		_, _ = io.WriteString(w, `[{"number":99,"state":"open","head":{"ref":"benvin/last","label":"benvin/last"}}]`)
+	})
+	srv := httptest.NewServer(mux)
+	defer srv.Close()
+
+	c := &GiteaClient{BaseURL: srv.URL, HTTP: srv.Client()}
+	prs, err := c.ListPRs("unkin/repo", "all")
+	if err != nil {
+		t.Fatalf("ListPRs: %v", err)
+	}
+	if len(prs) != prPageSize+1 {
+		t.Fatalf("got %d PRs, want %d", len(prs), prPageSize+1)
+	}
+	if len(pages) != 2 || pages[0] != "1" || pages[1] != "2" {
+		t.Errorf("pages requested = %v, want [1 2]", pages)
+	}
+	if got := PRHeadBranch(prs[0]); got != "benvin/b1" {
+		t.Errorf("first PR head branch = %q, want benvin/b1", got)
+	}
+	if !prs[len(prs)-1].IsOpen() {
+		t.Error("last PR should be open")
+	}
+}
+
+// A listing that fills every page is truncated: the caller must be told rather
+// than treating a partial view as the whole repo.
+func TestListPRsReportsTruncation(t *testing.T) {
+	mux := http.NewServeMux()
+	mux.HandleFunc("/api/v1/repos/unkin/repo/pulls", func(w http.ResponseWriter, r *http.Request) {
+		full := make([]string, 0, prPageSize)
+		for i := 0; i < prPageSize; i++ {
+			full = append(full, fmt.Sprintf(`{"number":%s,"state":"open"}`, r.URL.Query().Get("page")))
+		}
+		_, _ = io.WriteString(w, "["+strings.Join(full, ",")+"]")
+	})
+	srv := httptest.NewServer(mux)
+	defer srv.Close()
+
+	c := &GiteaClient{BaseURL: srv.URL, HTTP: srv.Client()}
+	prs, err := c.ListPRs("unkin/repo", "all")
+	if !errors.Is(err, ErrPRListTruncated) {
+		t.Fatalf("ListPRs err = %v, want ErrPRListTruncated", err)
+	}
+	if len(prs) != maxPRPages*prPageSize {
+		t.Errorf("got %d PRs, want %d", len(prs), maxPRPages*prPageSize)
+	}
+}
+
 func TestGiteaAPIError(t *testing.T) {
 	mux := http.NewServeMux()
 	mux.HandleFunc("/api/v1/repos/unkin/repo/pulls", func(w http.ResponseWriter, r *http.Request) {

internal/agent/git.go

@@ -2,6 +2,7 @@ package agent
 
 import (
 	"bytes"
+	"errors"
 	"fmt"
 	"os"
 	"os/exec"
@@ -67,6 +68,14 @@ func GitFetch(repoDir, remote string, globalArgs ...string) error {
 	return err
 }
 
+// GitFetchPrune runs `git fetch --prune <remote>` in repoDir so remote-tracking
+// refs for branches deleted on the remote (e.g. after a merge) disappear.
+func GitFetchPrune(repoDir, remote string, globalArgs ...string) error {
+	args := append(append([]string{}, globalArgs...), "fetch", "--prune", remote)
+	_, err := runGit(repoDir, args...)
+	return err
+}
+
 // GitRemoteDefaultBranch returns the short name of remote's default branch
 // (e.g. "main") by resolving refs/remotes/<remote>/HEAD.
 func GitRemoteDefaultBranch(repoDir, remote string) (string, error) {
@@ -83,6 +92,64 @@ func GitBranchExists(repoDir, branch string) bool {
 	return err == nil
 }
 
+// GitRemoteURL returns the configured URL for a remote.
+func GitRemoteURL(repoDir, remote string) (string, error) {
+	return runGit(repoDir, "remote", "get-url", remote)
+}
+
+// GitRemoteBranchExists reports whether a remote-tracking ref for branch exists
+// (accurate only after a pruning fetch).
+func GitRemoteBranchExists(repoDir, remote, branch string) bool {
+	_, err := runGit(repoDir, "show-ref", "--verify", "--quiet", "refs/remotes/"+remote+"/"+branch)
+	return err == nil
+}
+
+// GitIsDirty reports whether the checkout at dir has uncommitted or untracked
+// changes.
+func GitIsDirty(dir string) (bool, error) {
+	out, err := runGit(dir, "status", "--porcelain")
+	if err != nil {
+		return false, err
+	}
+	return strings.TrimSpace(out) != "", nil
+}
+
+// GitIsAncestor reports whether ancestor is reachable from descendant.
+func GitIsAncestor(repoDir, ancestor, descendant string) (bool, error) {
+	cmd := exec.Command("git", "merge-base", "--is-ancestor", ancestor, descendant)
+	cmd.Dir = repoDir
+	var stderr bytes.Buffer
+	cmd.Stderr = &stderr
+	if err := cmd.Run(); err != nil {
+		// Exit 1 is the documented "not an ancestor" answer; anything else is a
+		// real failure (bad ref, not a repo).
+		var exitErr *exec.ExitError
+		if errors.As(err, &exitErr) && exitErr.ExitCode() == 1 {
+			return false, nil
+		}
+		return false, fmt.Errorf("git merge-base --is-ancestor %s %s: %w: %s",
+			ancestor, descendant, err, strings.TrimSpace(stderr.String()))
+	}
+	return true, nil
+}
+
+// GitUnmergedCommits counts commits on head whose patch has no equivalent on
+// upstream, using `git cherry` so squash- and rebase-merged work is recognised
+// despite its rewritten SHAs.
+func GitUnmergedCommits(repoDir, upstream, head string) (int, error) {
+	out, err := runGit(repoDir, "cherry", upstream, head)
+	if err != nil {
+		return 0, err
+	}
+	n := 0
+	for _, line := range strings.Split(out, "\n") {
+		if strings.HasPrefix(strings.TrimSpace(line), "+") {
+			n++
+		}
+	}
+	return n, nil
+}
+
 // GitWorktreeAdd adds a worktree at path checked out to branch. When the branch
 // already exists it is reused; otherwise it is created from startPoint.
 func GitWorktreeAdd(repoDir, path, branch, startPoint string) error {

internal/agent/git_test.go

@@ -185,6 +185,134 @@ func TestGitWorktreeLifecycle(t *testing.T) {
 	}
 }
 
+// commit writes a file and commits it, returning the new HEAD sha.
+func commit(t *testing.T, dir, name, content, msg string) string {
+	t.Helper()
+	if err := os.WriteFile(filepath.Join(dir, name), []byte(content), 0o644); err != nil {
+		t.Fatal(err)
+	}
+	if _, err := runGit(dir, "add", "."); err != nil {
+		t.Fatalf("add: %v", err)
+	}
+	if _, err := runGit(dir, "commit", "-m", msg); err != nil {
+		t.Fatalf("commit: %v", err)
+	}
+	sha, err := runGit(dir, "rev-parse", "HEAD")
+	if err != nil {
+		t.Fatalf("rev-parse: %v", err)
+	}
+	return sha
+}
+
+func TestGitIsAncestor(t *testing.T) {
+	srcDir := newTempRepos(t)
+	base, err := runGit(srcDir, "rev-parse", "HEAD")
+	if err != nil {
+		t.Fatal(err)
+	}
+	tip := commit(t, srcDir, "a.txt", "a\n", "add a")
+
+	if ok, err := GitIsAncestor(srcDir, base, tip); err != nil || !ok {
+		t.Errorf("GitIsAncestor(base, tip) = %v, %v; want true", ok, err)
+	}
+	if ok, err := GitIsAncestor(srcDir, tip, base); err != nil || ok {
+		t.Errorf("GitIsAncestor(tip, base) = %v, %v; want false with no error", ok, err)
+	}
+	if _, err := GitIsAncestor(srcDir, "no-such-ref", tip); err == nil {
+		t.Error("GitIsAncestor with a bogus ref should error, not report false")
+	}
+}
+
+// These repos squash-merge, so merged work keeps its local SHA while the
+// upstream commit is a different one carrying the same patch. `git cherry` must
+// see that as merged even though the SHAs differ.
+func TestGitUnmergedCommitsIgnoresRewrittenSHAs(t *testing.T) {
+	srcDir := newTempRepos(t)
+
+	if _, err := runGit(srcDir, "checkout", "-b", "feature"); err != nil {
+		t.Fatalf("checkout: %v", err)
+	}
+	commit(t, srcDir, "f.txt", "hello\n", "add f")
+
+	n, err := GitUnmergedCommits(srcDir, "origin/main", "HEAD")
+	if err != nil {
+		t.Fatalf("GitUnmergedCommits: %v", err)
+	}
+	if n != 1 {
+		t.Fatalf("unmerged before upstream landing = %d, want 1", n)
+	}
+
+	// Land the same patch upstream under a different SHA.
+	if _, err := runGit(srcDir, "checkout", "main"); err != nil {
+		t.Fatalf("checkout main: %v", err)
+	}
+	commit(t, srcDir, "f.txt", "hello\n", "squashed f")
+	if _, err := runGit(srcDir, "push", "origin", "main"); err != nil {
+		t.Fatalf("push: %v", err)
+	}
+	if err := GitFetchPrune(srcDir, "origin"); err != nil {
+		t.Fatalf("GitFetchPrune: %v", err)
+	}
+
+	if ok, err := GitIsAncestor(srcDir, "feature", "origin/main"); err != nil || ok {
+		t.Fatalf("squash-merged branch must not be an ancestor: %v, %v", ok, err)
+	}
+	n, err = GitUnmergedCommits(srcDir, "origin/main", "feature")
+	if err != nil {
+		t.Fatalf("GitUnmergedCommits: %v", err)
+	}
+	if n != 0 {
+		t.Errorf("unmerged after upstream landing = %d, want 0", n)
+	}
+}
+
+func TestGitIsDirty(t *testing.T) {
+	srcDir := newTempRepos(t)
+	if dirty, err := GitIsDirty(srcDir); err != nil || dirty {
+		t.Fatalf("clean checkout reported dirty=%v, err=%v", dirty, err)
+	}
+	if err := os.WriteFile(filepath.Join(srcDir, "scratch.txt"), []byte("wip\n"), 0o644); err != nil {
+		t.Fatal(err)
+	}
+	if dirty, err := GitIsDirty(srcDir); err != nil || !dirty {
+		t.Errorf("untracked file must count as dirty: dirty=%v, err=%v", dirty, err)
+	}
+}
+
+func TestGitRemoteBranchExists(t *testing.T) {
+	srcDir := newTempRepos(t)
+	if !GitRemoteBranchExists(srcDir, "origin", "main") {
+		t.Error("origin/main should exist")
+	}
+	if GitRemoteBranchExists(srcDir, "origin", "benvin/nope") {
+		t.Error("origin/benvin/nope should not exist")
+	}
+
+	if _, err := runGit(srcDir, "checkout", "-b", "benvin/pushed"); err != nil {
+		t.Fatalf("checkout: %v", err)
+	}
+	commit(t, srcDir, "p.txt", "p\n", "add p")
+	if _, err := runGit(srcDir, "push", "origin", "benvin/pushed"); err != nil {
+		t.Fatalf("push: %v", err)
+	}
+	if err := GitFetchPrune(srcDir, "origin"); err != nil {
+		t.Fatalf("GitFetchPrune: %v", err)
+	}
+	if !GitRemoteBranchExists(srcDir, "origin", "benvin/pushed") {
+		t.Error("pushed branch should have a remote-tracking ref")
+	}
+
+	if _, err := runGit(srcDir, "push", "origin", "--delete", "benvin/pushed"); err != nil {
+		t.Fatalf("delete remote branch: %v", err)
+	}
+	if err := GitFetchPrune(srcDir, "origin"); err != nil {
+		t.Fatalf("GitFetchPrune: %v", err)
+	}
+	if GitRemoteBranchExists(srcDir, "origin", "benvin/pushed") {
+		t.Error("a pruning fetch must drop the tracking ref for a deleted remote branch")
+	}
+}
+
 // resolve canonicalizes a path (temp dirs may live behind symlinks like /var).
 func resolve(t *testing.T, p string) string {
 	t.Helper()

internal/agent/gitea.go

@@ -137,10 +137,68 @@ type PullRequest struct {
 	Mergeable bool   `json:"mergeable"`
 	HTMLURL   string `json:"html_url"`
 	Head      struct {
-		Sha string `json:"sha"`
+		Sha   string `json:"sha"`
+		Ref   string `json:"ref"`
+		Label string `json:"label"`
 	} `json:"head"`
 }
 
+// prPageSize is the per-page limit for the pulls listing; maxPRPages caps how
+// far back a listing walks.
+const (
+	prPageSize = 50
+	maxPRPages = 20
+)
+
+// ErrPRListTruncated reports that a listing hit the page cap, so the returned
+// pull requests are only the most recent ones and older PRs went unseen.
+var ErrPRListTruncated = errors.New("pull request listing truncated at the page cap")
+
+// ListPRs lists a repo's pull requests in the given state ("open", "closed" or
+// "all"), following pagination. A repo with more PRs than the page cap returns
+// the PRs it did read alongside ErrPRListTruncated.
+func (c *GiteaClient) ListPRs(repoPath, state string) ([]PullRequest, error) {
+	if state == "" {
+		state = "all"
+	}
+	var all []PullRequest
+	for page := 1; page <= maxPRPages; page++ {
+		var batch []PullRequest
+		path := fmt.Sprintf("/api/v1/repos/%s/pulls?state=%s&limit=%d&page=%d", repoPath, state, prPageSize, page)
+		if err := c.do(http.MethodGet, path, nil, &batch); err != nil {
+			return nil, err
+		}
+		all = append(all, batch...)
+		if len(batch) < prPageSize {
+			return all, nil
+		}
+	}
+	return all, fmt.Errorf("%s: %w after %d pull requests", repoPath, ErrPRListTruncated, len(all))
+}
+
+// PRHeadBranch returns the branch a PR was opened from. Gitea rewrites head.ref
+// to "refs/pull/<n>/head" once the branch is deleted (which merging does), so
+// head.label — which keeps the original name — is authoritative.
+func PRHeadBranch(pr PullRequest) string {
+	if label := pr.Head.Label; label != "" && !strings.HasPrefix(label, "refs/pull/") {
+		// Cross-repo PRs label as "<owner>:<branch>".
+		if _, branch, ok := strings.Cut(label, ":"); ok {
+			return branch
+		}
+		return label
+	}
+	ref := pr.Head.Ref
+	if strings.HasPrefix(ref, "refs/pull/") {
+		return ""
+	}
+	return strings.TrimPrefix(ref, "refs/heads/")
+}
+
+// IsOpen reports whether a PR is still open (not merged, not closed).
+func (pr PullRequest) IsOpen() bool {
+	return pr.State == "open" && !pr.Merged
+}
+
 // CreatePROptions are the fields for opening a PR.
 type CreatePROptions struct {
 	Base  string `json:"base"`

internal/agent/parse.go

@@ -68,6 +68,52 @@ func ParseDurationFlag(flag, value string) (time.Duration, error) {
 	return d, nil
 }
 
+// RemoteHost returns the host a git remote URL points at, or "" for a local
+// path remote.
+func RemoteHost(remote string) string {
+	s := strings.TrimSpace(remote)
+	if _, after, ok := strings.Cut(s, "://"); ok {
+		host, _, _ := strings.Cut(after, "/")
+		if _, bare, ok := strings.Cut(host, "@"); ok {
+			host = bare
+		}
+		return host
+	}
+	if strings.HasPrefix(s, "/") || strings.HasPrefix(s, ".") {
+		return ""
+	}
+	host, _, ok := strings.Cut(s, ":")
+	if !ok {
+		return ""
+	}
+	if _, bare, ok := strings.Cut(host, "@"); ok {
+		host = bare
+	}
+	return host
+}
+
+// RepoPathFromRemoteURL extracts the "owner/repo" API path from a git remote
+// URL, accepting both https and scp-style ssh forms.
+func RepoPathFromRemoteURL(remote string) (string, error) {
+	s := strings.TrimSuffix(strings.TrimSuffix(strings.TrimSpace(remote), "/"), ".git")
+	switch {
+	case strings.Contains(s, "://"):
+		_, after, _ := strings.Cut(s, "://")
+		_, path, ok := strings.Cut(after, "/")
+		if !ok {
+			return "", fmt.Errorf("remote URL %q has no repo path", remote)
+		}
+		s = path
+	case strings.Contains(s, ":"):
+		_, s, _ = strings.Cut(s, ":")
+	}
+	parts := strings.Split(strings.Trim(s, "/"), "/")
+	if len(parts) < 2 || parts[len(parts)-2] == "" || parts[len(parts)-1] == "" {
+		return "", fmt.Errorf("remote URL %q is not owner/repo shaped", remote)
+	}
+	return parts[len(parts)-2] + "/" + parts[len(parts)-1], nil
+}
+
 // ParseRepo validates and splits an "owner/repo" string.
 func ParseRepo(s string) (owner, repo string, err error) {
 	s = strings.TrimSpace(s)

internal/agent/parse_test.go

@@ -132,3 +132,52 @@ func TestParseDurationFlagErrorMessage(t *testing.T) {
 		}
 	}
 }
+
+// Not every managed repo lives under the default owner, so the API path comes
+// from origin's URL rather than the directory name.
+func TestRepoPathFromRemoteURL(t *testing.T) {
+	tests := []struct {
+		in   string
+		want string
+	}{
+		{"https://git.unkin.net/unkin/agent-tools.git", "unkin/agent-tools"},
+		{"https://git.unkin.net/unkinben/dotfiles.git", "unkinben/dotfiles"},
+		{"https://git.unkin.net/unkin/agent-tools", "unkin/agent-tools"},
+		{"https://user@git.unkin.net/unkin/agent-tools.git", "unkin/agent-tools"},
+		{"ssh://git@git.unkin.net:2222/unkin/agent-tools.git", "unkin/agent-tools"},
+		{"git@git.unkin.net:unkin/agent-tools.git", "unkin/agent-tools"},
+	}
+	for _, tt := range tests {
+		got, err := RepoPathFromRemoteURL(tt.in)
+		if err != nil {
+			t.Errorf("RepoPathFromRemoteURL(%q): %v", tt.in, err)
+			continue
+		}
+		if got != tt.want {
+			t.Errorf("RepoPathFromRemoteURL(%q) = %q, want %q", tt.in, got, tt.want)
+		}
+	}
+	for _, bad := range []string{"", "https://git.unkin.net", "agent-tools"} {
+		if got, err := RepoPathFromRemoteURL(bad); err == nil {
+			t.Errorf("RepoPathFromRemoteURL(%q) = %q, want error", bad, got)
+		}
+	}
+}
+
+func TestRemoteHost(t *testing.T) {
+	tests := []struct {
+		in, want string
+	}{
+		{"https://git.unkin.net/unkin/repo.git", "git.unkin.net"},
+		{"https://user@git.unkin.net/unkin/repo.git", "git.unkin.net"},
+		{"ssh://git@git.unkin.net:2222/unkin/repo.git", "git.unkin.net:2222"},
+		{"git@git.unkin.net:unkin/repo.git", "git.unkin.net"},
+		{"/tmp/fixture/origin.git", ""},
+		{"../other/origin.git", ""},
+	}
+	for _, tt := range tests {
+		if got := RemoteHost(tt.in); got != tt.want {
+			t.Errorf("RemoteHost(%q) = %q, want %q", tt.in, got, tt.want)
+		}
+	}
+}