Merge pull request 'Add agentpr pr edit subcommand' (#15) from benvin/agentpr-pr-edit into main

Reviewed-on: #15

AGENTS.md

@@ -8,9 +8,9 @@ from Vault, so actions are attributed to the agent rather than to whoever runs
 the tool. Setting `AGENT_LOGIN` selects a different agent identity, so a service
 like repospawner can run these tools as itself.
 
-- **`agentpr`** — create pull requests and post PR comments as `unkin-agent`
-  (fixes the "tea posts as Ben" attribution problem). Subcommands:
-  `pr create`, `pr comment`, `whoami`.
+- **`agentpr`** — create and edit pull requests, and post PR comments, as
+  `unkin-agent` (fixes the "tea posts as Ben" attribution problem).
+  Subcommands: `pr create`, `pr comment`, `pr edit`, `whoami`.
 - **`watchpr`** — poll one or more PRs and exit when a tracked PR changes
   meaningfully: it merges/closes, gets a new non-agent comment, its CI fails,
   or it loses mergeability. Benign transitions (CI pending→success, the agent's
@@ -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
@@ -28,20 +28,22 @@ parsing, watch-state comparison, git worktree helpers).
 ## Structure
 
 ```
-cmd/agentpr/main.go      # agentpr CLI (pr create / pr comment / whoami)
+cmd/agentpr/main.go      # agentpr CLI (pr create / pr comment / pr edit / whoami)
 cmd/watchpr/main.go      # watchpr CLI (poll + meaningful-change exit)
 cmd/agentws/main.go      # agentws CLI (new / list / rm / clean / token / credential)
-cmd/agentvault/main.go   # agentvault CLI (seed-outpost)
+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
   vault.go               #   AppRole login + read the gitea creds path
-  gitea.go               #   Gitea REST client (PR create/get, comments, status, whoami)
+  gitea.go               #   Gitea REST client (PR create/edit/get, comments, status, whoami)
   parse.go               #   owner/repo#N and owner/repo parsing
   watch.go               #   PRState snapshot + MeaningfulChange comparison
   git.go                 #   git worktree/clone/fetch helpers (os/exec, no go-git)
   vaultkv.go             #   AppRole-authenticated Vault client + KV-v2 read/write
   authentik.go           #   Authentik REST client (outpost search, token view_key)
   seedoutpost.go         #   seed-outpost flow (Authentik token -> Vault KV)
+  seedoauth.go           #   seed-oauth flow (oauth2-proxy credential set in Vault KV)
 go.mod                   # module git.unkin.net/unkin/agent-tools
 Makefile                 # build / test / lint / completions / rpm / version-bump
 packaging/nfpm.yaml      # nfpm spec (envsubst-templated) for the RPM (all binaries)
@@ -148,12 +150,61 @@ printed. Errors are wrapped per stage (login / read denied / outpost missing /
 view_key / write denied) via the `ErrVaultDenied`, `ErrVaultNotFound` and
 `ErrOutpostNotFound` sentinels.
 
+## agentvault seed-oauth
+
+`agentvault seed-oauth --path <kv/path> --client-id <id>` makes a KV-v2 path
+hold a complete oauth2-proxy credential set, in-process:
+
+1. AppRole login (shared `approleLogin`), then a KV-v2 read via
+   `ReadKVOptional` — a 404 or a deleted version means "empty", not an error,
+   so the first seed of a path works.
+2. Desired keys are computed over the existing map: `client_id` from the flag
+   (`kept`/`created`/`updated`), `client_secret` and `cookie_secret` generated
+   from 32 `crypto/rand` bytes only when absent or when `--rotate` is set
+   (`kept`/`created`/`rotated`). `cookie_secret` is base64url so it decodes to
+   exactly the 32 bytes oauth2-proxy demands; `client_secret` is standard
+   base64.
+3. Any other key on the path is carried through unchanged (`preserved`), which
+   is why the write goes through `WriteKVAny` rather than `WriteKV`.
+4. The write is skipped entirely when nothing changed; the command then prints
+   `version: unchanged`.
+
+Only key names, per-key actions and the new KV version are printed. Errors are
+wrapped per stage (login / read denied / write denied) via `ErrVaultDenied`.
+
 ## Gotchas
 
 - `watchpr` exits 0 with no output changes on `--once` (just prints state).
-- The token cache is process-wide (`sync.Once`); tests call the unexported
-  `fetchGiteaToken` to avoid it.
+- Gitea tokens expire in ~1h, shorter than a watch: the client re-mints once on a
+  401/403 and replays the request. If the fresh token is rejected too, `watchpr`
+  exits non-zero rather than polling blind.
+- `watchpr` polls anonymously when no token can be minted (public repos work
+  fine); only a real 401/403 reaches for Vault.
+- The token cache is process-wide (mutex-guarded); `RefreshGiteaToken` replaces
+  it. Tests call the unexported `fetchGiteaToken` to avoid the cache.
 - `agentvault` never puts a secret in an error string: Vault decode failures and
-  Authentik `view_key` responses are reported without their bodies.
+  Authentik `view_key` responses are reported without their bodies, and
+  `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.
+- `agentws prune` discovers worktrees from the worktree root *and* from
+  `git worktree list` on each source checkout, merging the two so git's own
+  `locked`/`prunable` flags reach entries the directory scan already found.
+  Removing a worktree is only safe because the local branch keeps its commits, so
+  the cases with no branch to fall back on are kept: a detached HEAD carrying
+  commits on no remote, a locked checkout, or one with a sequencer operation
+  half-finished (`rebase-merge`, `MERGE_HEAD`, `CHERRY_PICK_HEAD`, …). A directory
+  whose backing repo is gone is deleted outright, but only ever inside the
+  worktree root.
 - CI "combined status" comes from `/commits/{sha}/status`; an empty head SHA
   yields an empty state without an API call.

README.md

@@ -6,7 +6,8 @@ token from Vault, so automated PRs, comments and pushes are attributed to the
 agent — not to whoever happens to run the command. Set `AGENT_LOGIN` to act as a
 different agent identity.
 
-- **`agentpr`** — create pull requests and post PR comments as the agent user.
+- **`agentpr`** — create and edit pull requests, and post PR comments as the
+  agent user.
 - **`watchpr`** — poll one or more PRs and exit when one changes in a way worth
   acting on.
 - **`agentws`** — manage per-branch git worktrees for `unkin-agent`, cloning
@@ -50,6 +51,11 @@ agentpr pr create --repo unkin/argocd-apps \
 # Comment on a PR
 agentpr pr comment --repo unkin/argocd-apps --pr 42 --body "Rebased, CI green."
 
+# Edit a PR's title and/or body; an omitted flag is left unchanged
+agentpr pr edit --repo unkin/argocd-apps --pr 42 --body "Adds the ServiceAccount ..."
+agentpr pr edit --repo unkin/argocd-apps --pr 42 --title "Add woodpecker SA"
+# prints: #<number> <html_url>
+
 agentpr --version
 agentpr --help
 ```
@@ -70,6 +76,9 @@ watchpr unkin/argocd-apps#42
 # Multiple PRs, custom interval; refs accept #N or :N
 watchpr --interval 30s unkin/argocd-apps#42 unkin/terraform-vault:98
 
+# --interval takes a duration (30s, 2m, 1h30m) or a bare number of seconds
+watchpr --interval 30 unkin/argocd-apps#42
+
 # One-shot: print current state and exit 0 (great for scripts)
 watchpr --once unkin/argocd-apps#42
 watchpr --once --json unkin/argocd-apps#42
@@ -101,6 +110,13 @@ agentws list
 agentws rm benvin/my-change
 agentws rm ~/.cache/agentws/argocd-apps__benvin-my-change --delete-branch
 
+# Classify every worktree found; dry run unless --yes is given
+agentws prune
+agentws prune --json
+agentws prune --no-fetch
+agentws prune --yes
+agentws prune --yes --keep-branches
+
 # Remove every managed worktree and prune each source repo
 agentws clean
 
@@ -108,6 +124,82 @@ agentws clean
 agentws token
 ```
 
+### prune
+
+`agentws prune` finds worktrees two ways and merges the results: the managed
+directories under the worktree root, and `git worktree list` on every source
+checkout — so hand-made worktrees, stale registrations whose directory is gone,
+and leftover directories whose source repo was deleted all show up too.
+
+It then decides, per worktree, whether its work is safely upstream:
+
+| Signal (first match wins) | Verdict |
+|---|---|
+| working tree gone (registration only) | prune the registration |
+| backing repo gone | delete the leftover directory |
+| locked by `git worktree lock` | keep |
+| rebase, merge, cherry-pick, revert or bisect in progress | keep |
+| uncommitted, staged 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 |
+| detached HEAD carrying commits on no remote | keep |
+| 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.
+
+The last row is safe only because the local branch keeps the commits, so the
+reason names the branch it is relying on. A detached HEAD has no such branch, so
+unique commits there are kept instead.
+
+Output is a table (`REPO BRANCH PATH VERDICT REASON`) with every verdict's reason
+spelled out, or `--json` for scripting. Neither form needs a terminal.
+
+| Flag | Effect |
+|---|---|
+| `--yes` | apply the plan; without it nothing is touched |
+| `--keep-branches` | remove worktrees only; verdicts print as `remove` |
+| `--no-fetch` | judge against the refs already on disk, for offline use |
+| `--json` | emit the report as JSON on stdout, notes on stderr |
+| `--include-unmanaged` | also remove worktrees outside the worktree root |
+| `--include-keep` | dangerous: also remove worktrees classified `keep`, destroying uncommitted and in-progress work |
+
+Without `--include-unmanaged` a hand-made worktree is reported and then skipped,
+naming the flag that would remove it. `--include-keep` is the only way past a
+`keep`. It leaves the branch, so committed work outlives the worktree, but
+`git worktree remove --force` discards a dirty working tree and a paused
+rebase's sequencer state without a word, and no branch was carrying those.
+
+A directory under the worktree root is deleted outright only when its git dir
+and the repo's shared `.git` are both proven absent by `stat`. When git merely
+fails to answer for a checkout, the verdict is `keep` with the error as its
+reason: an unread state is never a dead one.
+
 ### Auth / credential-helper design
 
 Gitea tokens minted from Vault are short-lived (~1h), so `agentws` never
@@ -162,6 +254,42 @@ Errors name the failing stage: AppRole login, KV read denied (policy not
 applied), outpost not found (terraform not applied), `view_key` failure, or KV
 write denied.
 
+### seed-oauth
+
+Make a Vault KV-v2 path hold a complete oauth2-proxy credential set. It is a
+read-modify-write: `client_id` is set from the flag, `client_secret` and
+`cookie_secret` are generated (32 bytes from `crypto/rand`) only when missing,
+every other key on the path is written back untouched, and nothing is written
+at all when the secret is already correct. `cookie_secret` is base64url so it
+decodes to exactly the 32 bytes oauth2-proxy requires.
+
+```bash
+agentvault seed-oauth \
+  --path kubernetes/namespace/repospawner/default/oauth-credentials \
+  --client-id 4f1c…
+```
+
+```
+path:          kv/kubernetes/namespace/repospawner/default/oauth-credentials
+keys:          client_id, client_secret, cookie_secret
+  client_id:     created
+  client_secret: kept
+  cookie_secret: created
+version:       4
+```
+
+That is the common case: the provider's `client_secret` already lives on the
+path, so only the missing keys are added. A run with nothing to do prints
+`version:       unchanged` and issues no write.
+
+Flags: `--path` and `--client-id` are required; `--kv-mount` (default `kv`) and
+`--rotate` override the rest. `--rotate` regenerates both secrets — only use it
+when the IdP provider's secret is being rotated alongside, since a rotated
+`client_secret` no longer matches the provider.
+
+Errors name the failing stage: AppRole login, KV read denied, or KV write
+denied. Only key names, actions and the KV version are printed.
+
 ## Build & package
 
 ```bash

cmd/agentpr/main.go

@@ -5,6 +5,7 @@
 //
 //	agentpr pr create --repo owner/repo --base main --head feature --title T --body B
 //	agentpr pr comment --repo owner/repo --pr 12 --body "..."
+//	agentpr pr edit --repo owner/repo --pr 12 --title T --body B
 //	agentpr whoami
 package main
 
@@ -56,9 +57,9 @@ func client() (*agent.GiteaClient, error) {
 func newPRCmd() *cobra.Command {
 	cmd := &cobra.Command{
 		Use:   "pr",
-		Short: "Create PRs and post PR comments",
+		Short: "Create and edit PRs, and post PR comments",
 	}
-	cmd.AddCommand(newPRCreateCmd(), newPRCommentCmd())
+	cmd.AddCommand(newPRCreateCmd(), newPRCommentCmd(), newPREditCmd())
 	return cmd
 }
 
@@ -143,6 +144,60 @@ func newPRCommentCmd() *cobra.Command {
 	return cmd
 }
 
+func newPREditCmd() *cobra.Command {
+	var repo, title, body string
+	var pr int
+	cmd := &cobra.Command{
+		Use:          "edit",
+		Short:        "Edit a pull request's title and/or body",
+		SilenceUsage: true,
+		RunE: func(cmd *cobra.Command, args []string) error {
+			owner, name, err := agent.ParseRepo(repo)
+			if err != nil {
+				return err
+			}
+			if pr <= 0 {
+				return fmt.Errorf("--pr must be a positive PR number")
+			}
+			// Only the flags actually given are sent: omitting --title must
+			// leave the title as it is, not blank it.
+			var opts agent.EditPROptions
+			if cmd.Flags().Changed("title") {
+				// Gitea ignores an empty title, so sending one would report
+				// success while changing nothing.
+				if title == "" {
+					return fmt.Errorf("--title cannot be empty: a title can be set but not cleared")
+				}
+				opts.Title = &title
+			}
+			if cmd.Flags().Changed("body") {
+				opts.Body = &body
+			}
+			if opts.Title == nil && opts.Body == nil {
+				return fmt.Errorf("at least one of --title or --body is required")
+			}
+			c, err := client()
+			if err != nil {
+				return err
+			}
+			updated, err := c.EditPR(owner+"/"+name, pr, opts)
+			if err != nil {
+				return err
+			}
+			fmt.Printf("#%d %s\n", updated.Number, updated.HTMLURL)
+			return nil
+		},
+	}
+	f := cmd.Flags()
+	f.StringVar(&repo, "repo", "", "Repository as owner/repo (required)")
+	f.IntVar(&pr, "pr", 0, "PR number (required)")
+	f.StringVar(&title, "title", "", "New PR title (unchanged when omitted)")
+	f.StringVar(&body, "body", "", "New PR body (unchanged when omitted)")
+	_ = cmd.MarkFlagRequired("repo")
+	_ = cmd.MarkFlagRequired("pr")
+	return cmd
+}
+
 func newWhoamiCmd() *cobra.Command {
 	return &cobra.Command{
 		Use:          "whoami",

cmd/agentpr/main_test.go

@@ -2,6 +2,7 @@ package main
 
 import (
 	"io"
+	"strings"
 	"testing"
 )
 
@@ -16,3 +17,35 @@ func TestExecuteBadRepoErrors(t *testing.T) {
 		t.Fatal("Execute() = nil, want error for a malformed --repo")
 	}
 }
+
+// `pr edit` with neither --title nor --body has nothing to send; it must fail
+// with a usage error before any Vault/Gitea call, so this stays hermetic.
+func TestPREditRequiresTitleOrBody(t *testing.T) {
+	cmd := newRootCmd()
+	cmd.SetArgs([]string{"pr", "edit", "--repo", "unkin/repo", "--pr", "7"})
+	cmd.SetOut(io.Discard)
+	cmd.SetErr(io.Discard)
+	err := cmd.Execute()
+	if err == nil {
+		t.Fatal("Execute() = nil, want an error when neither --title nor --body is given")
+	}
+	if !strings.Contains(err.Error(), "--title or --body") {
+		t.Errorf("error = %q, want it to name the missing flags", err)
+	}
+}
+
+// Gitea silently ignores an empty title, so `pr edit --title ""` would report
+// success while changing nothing; it must fail before any Vault/Gitea call.
+func TestPREditRejectsEmptyTitle(t *testing.T) {
+	cmd := newRootCmd()
+	cmd.SetArgs([]string{"pr", "edit", "--repo", "unkin/repo", "--pr", "7", "--title", ""})
+	cmd.SetOut(io.Discard)
+	cmd.SetErr(io.Discard)
+	err := cmd.Execute()
+	if err == nil {
+		t.Fatal("Execute() = nil, want an error for an empty --title")
+	}
+	if !strings.Contains(err.Error(), "--title cannot be empty") {
+		t.Errorf("error = %q, want it to reject the empty title", err)
+	}
+}

cmd/agentvault/main.go

@@ -3,11 +3,13 @@
 // with the same Vault AppRole as agentpr (role_id only, no secret_id).
 //
 //	agentvault seed-outpost --outpost <name> --dest-path <kv/path>
+//	agentvault seed-oauth --path <kv/path> --client-id <id>
 package main
 
 import (
 	"fmt"
 	"os"
+	"strings"
 
 	"git.unkin.net/unkin/agent-tools/internal/agent"
 
@@ -33,7 +35,7 @@ func newRootCmd() *cobra.Command {
 		SilenceUsage: true,
 	}
 	root.SetVersionTemplate("{{.Version}}\n")
-	root.AddCommand(newSeedOutpostCmd(), newVersionCmd())
+	root.AddCommand(newSeedOutpostCmd(), newSeedOAuthCmd(), newVersionCmd())
 	return root
 }
 
@@ -73,6 +75,48 @@ func newSeedOutpostCmd() *cobra.Command {
 	return cmd
 }
 
+func newSeedOAuthCmd() *cobra.Command {
+	opts := agent.SeedOAuthOptions{}
+	cmd := &cobra.Command{
+		Use:   "seed-oauth",
+		Short: "Seed an oauth2-proxy credential set into Vault KV",
+		Long: "Make a Vault KV-v2 path hold a complete oauth2-proxy credential set: the\n" +
+			"given client_id, plus a client_secret and a 32-byte cookie_secret that are\n" +
+			"generated only when missing (or with --rotate). Existing keys are preserved\n" +
+			"and nothing is written when the secret is already correct. Secret values are\n" +
+			"never printed or logged.",
+		SilenceUsage: true,
+		RunE: func(cmd *cobra.Command, args []string) error {
+			opts.VaultAddr = agent.VaultAddr()
+			opts.RoleID = agent.RoleID()
+			res, err := agent.SeedOAuth(opts)
+			if err != nil {
+				return err
+			}
+			out := cmd.OutOrStdout()
+			_, _ = fmt.Fprintf(out, "path:          %s/%s\n", res.KVMount, res.Path)
+			_, _ = fmt.Fprintf(out, "keys:          %s\n", strings.Join(res.KeyNames(), ", "))
+			for _, k := range res.Keys {
+				_, _ = fmt.Fprintf(out, "  %-14s %s\n", k.Name+":", k.Action)
+			}
+			if res.Changed {
+				_, _ = fmt.Fprintf(out, "version:       %d\n", res.Version)
+			} else {
+				_, _ = fmt.Fprintf(out, "version:       unchanged\n")
+			}
+			return nil
+		},
+	}
+	f := cmd.Flags()
+	f.StringVar(&opts.Path, "path", "", "KV-v2 path holding the credentials, e.g. kubernetes/namespace/repospawner/default/oauth-credentials (required)")
+	f.StringVar(&opts.ClientID, "client-id", "", "OIDC client id to store (required)")
+	f.StringVar(&opts.KVMount, "kv-mount", agent.DefaultKVMount, "KV-v2 mount holding the path")
+	f.BoolVar(&opts.Rotate, "rotate", false, "Regenerate client_secret and cookie_secret even when they already exist")
+	_ = cmd.MarkFlagRequired("path")
+	_ = cmd.MarkFlagRequired("client-id")
+	return cmd
+}
+
 func newVersionCmd() *cobra.Command {
 	return &cobra.Command{
 		Use:          "version",

cmd/agentvault/main_test.go

@@ -2,6 +2,7 @@ package main
 
 import (
 	"bytes"
+	"encoding/json"
 	"io"
 	"net/http"
 	"net/http/httptest"
@@ -78,6 +79,89 @@ func TestSeedOutpostOutputHasNoSecrets(t *testing.T) {
 	}
 }
 
+const oauthPath = "kubernetes/namespace/repospawner/default/oauth-credentials"
+
+// fakeOAuthVault serves approle login plus a KV-v2 path that already holds a
+// client_secret, and records what gets written back.
+func fakeOAuthVault(t *testing.T, existing map[string]string) (vaultURL string, written *map[string]string) {
+	t.Helper()
+	writes := map[string]string{}
+	mux := http.NewServeMux()
+	mux.HandleFunc("/v1/auth/approle/login", func(w http.ResponseWriter, r *http.Request) {
+		_, _ = io.WriteString(w, `{"auth":{"client_token":"s.vaulttoken"}}`)
+	})
+	mux.HandleFunc("/v1/kv/data/"+oauthPath, func(w http.ResponseWriter, r *http.Request) {
+		if r.Method == http.MethodPost {
+			var body struct {
+				Data map[string]string `json:"data"`
+			}
+			_ = json.NewDecoder(r.Body).Decode(&body)
+			for k, v := range body.Data {
+				writes[k] = v
+			}
+			_, _ = io.WriteString(w, `{"data":{"version":4}}`)
+			return
+		}
+		payload, _ := json.Marshal(map[string]any{"data": map[string]any{"data": existing}})
+		_, _ = w.Write(payload)
+	})
+	srv := httptest.NewServer(mux)
+	t.Cleanup(srv.Close)
+	return srv.URL, &writes
+}
+
+// The command prints key names and the KV version only — never a value.
+func TestSeedOAuthOutputHasNoSecrets(t *testing.T) {
+	const existingSecret = "existing-client-secret-value"
+	vaultURL, written := fakeOAuthVault(t, map[string]string{"client_secret": existingSecret})
+	t.Setenv("VAULT_ADDR", vaultURL)
+	t.Setenv("AGENT_APPROLE_ROLE_ID", "role-xyz")
+
+	var out bytes.Buffer
+	cmd := newRootCmd()
+	cmd.SetOut(&out)
+	cmd.SetErr(&out)
+	cmd.SetArgs([]string{"seed-oauth", "--path", oauthPath, "--client-id", "mediamark-client-id"})
+	if err := cmd.Execute(); err != nil {
+		t.Fatalf("Execute: %v", err)
+	}
+
+	got := out.String()
+	for _, want := range []string{"kv/" + oauthPath, "client_id, client_secret, cookie_secret", "client_secret: kept", "cookie_secret: created", "version:       4"} {
+		if !strings.Contains(got, want) {
+			t.Errorf("output missing %q:\n%s", want, got)
+		}
+	}
+	for key, value := range *written {
+		if key == "client_id" {
+			continue
+		}
+		if strings.Contains(got, value) {
+			t.Fatalf("output leaks the %s value:\n%s", key, got)
+		}
+	}
+	if strings.Contains(got, existingSecret) {
+		t.Fatalf("output leaks the existing client_secret:\n%s", got)
+	}
+}
+
+func TestSeedOAuthRequiresFlags(t *testing.T) {
+	for name, args := range map[string][]string{
+		"no path":      {"seed-oauth", "--client-id", "mediamark-client-id"},
+		"no client-id": {"seed-oauth", "--path", oauthPath},
+	} {
+		t.Run(name, func(t *testing.T) {
+			cmd := newRootCmd()
+			cmd.SetOut(io.Discard)
+			cmd.SetErr(io.Discard)
+			cmd.SetArgs(args)
+			if err := cmd.Execute(); err == nil {
+				t.Fatal("Execute() = nil, want a missing-required-flag error")
+			}
+		})
+	}
+}
+
 func TestSeedOutpostRequiresFlags(t *testing.T) {
 	for name, args := range map[string][]string{
 		"no outpost":   {"seed-outpost", "--dest-path", destPath},

cmd/agentws/main.go

@@ -13,6 +13,8 @@
 //	agentws new <repo> [--branch benvin/<name>] [--from <base-branch>]
 //	agentws list
 //	agentws rm <path-or-branch> [--delete-branch]
+//	agentws prune [--yes] [--keep-branches] [--no-fetch] [--json]
+//	              [--include-unmanaged] [--include-keep]
 //	agentws clean
 //	agentws token
 //	agentws credential get   # git credential-helper protocol on stdin
@@ -24,6 +26,7 @@ import (
 	"io"
 	"os"
 	"path/filepath"
+	"sort"
 	"strings"
 
 	"git.unkin.net/unkin/agent-tools/internal/agent"
@@ -54,6 +57,7 @@ func newRootCmd() *cobra.Command {
 		newNewCmd(),
 		newListCmd(),
 		newRmCmd(),
+		newPruneCmd(),
 		newCleanCmd(),
 		newTokenCmd(),
 		newCredentialCmd(),
@@ -226,7 +230,14 @@ func newListCmd() *cobra.Command {
 				return nil
 			}
 			for _, w := range managed {
-				_, _ = fmt.Fprintf(out, "%s\t%s\t%s\n", w.repo, w.branch, w.path)
+				branch := w.branch
+				switch {
+				case w.orphan:
+					branch = "(orphan)"
+				case w.inspectErr != nil:
+					branch = "(unreadable)"
+				}
+				_, _ = fmt.Fprintf(out, "%s\t%s\t%s\n", w.repo, branch, w.path)
 			}
 			return nil
 		},
@@ -239,10 +250,31 @@ type managedWt struct {
 	branch string
 	path   string
 	srcDir string
+	// managed is false for worktrees found via `git worktree list` that live
+	// outside the worktree root, i.e. somebody made them by hand.
+	managed bool
+	// detached is true when the worktree has no branch to fall back on, so its
+	// commits die with the checkout.
+	detached bool
+	// locked records git's own "do not remove me" marker.
+	locked bool
+	// missing is a registration whose working tree is gone: nothing to inspect,
+	// nothing to lose.
+	missing bool
+	// orphan is a directory under the worktree root whose backing git dir is
+	// proven gone, so no git state can be read from it ever again.
+	orphan bool
+	// inspectErr is set when git refused to answer for a checkout and the reason
+	// was not a proven-absent backing repo. The state is unknown, never removable.
+	inspectErr error
 }
 
 // managedWorktrees scans the worktree root and resolves each entry's repo and
 // branch from git so branch names are accurate (not the sanitized dir name).
+// Directories whose backing repo is proven gone are returned as orphans rather
+// than dropped, so callers can see (and clean up) the leftovers; a directory git
+// merely failed to answer for is returned with its error instead, because an
+// unread state must never be mistaken for a dead one.
 func managedWorktrees() ([]managedWt, error) {
 	wr, err := worktreeRoot()
 	if err != nil {
@@ -261,24 +293,238 @@ func managedWorktrees() ([]managedWt, error) {
 			continue
 		}
 		path := filepath.Join(wr, e.Name())
-		branch, err := agent.GitCurrentBranch(path)
-		if err != nil {
-			continue // not a git worktree; skip
+		if _, err := os.Stat(filepath.Join(path, ".git")); err != nil {
+			continue // not a worktree checkout at all
 		}
-		srcDir, err := agent.SourceRepoDir(path)
-		if err != nil {
+		branch, branchErr := agent.GitCurrentBranch(path)
+		srcDir, srcErr := agent.SourceRepoDir(path)
+		if branchErr != nil || srcErr != nil {
+			entry := managedWt{repo: repoFromDirName(e.Name()), path: path, managed: true}
+			if gone, err := backingRepoGone(path); err == nil && gone {
+				entry.orphan = true
+			} else if branchErr != nil {
+				entry.inspectErr = branchErr
+			} else {
+				entry.inspectErr = srcErr
+			}
+			out = append(out, entry)
 			continue
 		}
 		out = append(out, managedWt{
-			repo:   filepath.Base(srcDir),
-			branch: branch,
-			path:   path,
-			srcDir: srcDir,
+			repo:     filepath.Base(srcDir),
+			branch:   branch,
+			path:     path,
+			srcDir:   srcDir,
+			managed:  true,
+			detached: branch == "HEAD",
 		})
 	}
 	return out, nil
 }
 
+// backingRepoGone proves, by stat alone, that a worktree directory's backing
+// repo no longer exists: its .git file names a git dir that is absent, and the
+// repo's shared .git the git dir lived in is absent too. Only that pair licenses
+// deleting the directory. Every other outcome — an unreadable .git file, a git
+// dir still on disk, a stat that failed for any reason other than "not there",
+// or a mere lost registration in a repo that is still present — reports false,
+// so a transient or unexplained failure can never be read as "safe to delete".
+func backingRepoGone(path string) (bool, error) {
+	dot := filepath.Join(path, ".git")
+	info, err := os.Lstat(dot)
+	if err != nil {
+		return false, err
+	}
+	if info.IsDir() {
+		return false, nil // a standalone checkout, not a linked worktree
+	}
+	data, err := os.ReadFile(dot)
+	if err != nil {
+		return false, err
+	}
+	rest, ok := strings.CutPrefix(strings.TrimSpace(string(data)), "gitdir:")
+	if !ok {
+		return false, fmt.Errorf("%s: not a worktree gitdir pointer", dot)
+	}
+	gitDir := strings.TrimSpace(rest)
+	if gitDir == "" {
+		return false, fmt.Errorf("%s: empty gitdir", dot)
+	}
+	if !filepath.IsAbs(gitDir) {
+		gitDir = filepath.Join(path, gitDir)
+	}
+	// The git dir is "<repo>/.git/worktrees/<name>"; both it and the shared .git
+	// it sits in must be absent before the repo counts as gone.
+	for _, dir := range []string{gitDir, filepath.Dir(filepath.Dir(gitDir))} {
+		if _, err := os.Stat(dir); err == nil {
+			return false, nil
+		} else if !os.IsNotExist(err) {
+			return false, err
+		}
+	}
+	return true, nil
+}
+
+// repoFromDirName recovers the repo name from the "<repo>__<branch>" layout used
+// under the worktree root, for entries git can no longer answer for.
+func repoFromDirName(name string) string {
+	if repo, _, ok := strings.Cut(name, "__"); ok {
+		return repo
+	}
+	return name
+}
+
+// allWorktrees is every worktree prune should consider: the managed ones under
+// the worktree root, plus whatever `git worktree list` reports for the repos
+// they belong to and for every checkout in the source root. The second source
+// finds hand-made worktrees and stale registrations whose directory is gone, and
+// carries git's own locked/prunable flags onto the entries the first source
+// already found.
+func allWorktrees() ([]managedWt, error) {
+	managed, err := managedWorktrees()
+	if err != nil {
+		return nil, err
+	}
+	wr, err := worktreeRoot()
+	if err != nil {
+		return nil, err
+	}
+
+	out := make([]managedWt, 0, len(managed))
+	index := map[string]int{}
+	for _, w := range managed {
+		index[resolvePath(w.path)] = len(out)
+		out = append(out, w)
+	}
+
+	for _, srcDir := range sourceRepos(managed) {
+		wts, err := agent.GitWorktreeList(srcDir)
+		if err != nil {
+			continue // not a repo any more, or unreadable; managed entries still stand
+		}
+		for _, wt := range wts {
+			if wt.Bare || sameDir(wt.Path, srcDir) {
+				continue
+			}
+			_, statErr := os.Stat(wt.Path)
+			entry := managedWt{
+				repo:     filepath.Base(srcDir),
+				branch:   worktreeBranch(wt),
+				path:     wt.Path,
+				srcDir:   srcDir,
+				managed:  underRoot(wt.Path, wr),
+				detached: wt.Detached,
+				locked:   wt.Locked,
+				missing:  wt.Prunable != "" || os.IsNotExist(statErr),
+			}
+			key := resolvePath(wt.Path)
+			if i, ok := index[key]; ok {
+				// Keep the managed scan's own view, but adopt the flags only git knows.
+				out[i].locked = entry.locked
+				out[i].missing = out[i].missing || entry.missing
+				out[i].detached = out[i].detached || entry.detached
+				continue
+			}
+			index[key] = len(out)
+			out = append(out, entry)
+		}
+	}
+	sort.Slice(out, func(i, j int) bool {
+		if out[i].repo != out[j].repo {
+			return out[i].repo < out[j].repo
+		}
+		return out[i].path < out[j].path
+	})
+	return out, nil
+}
+
+// resolvePath is a path key that matches however git spells the same directory.
+func resolvePath(path string) string {
+	if p, err := filepath.EvalSymlinks(path); err == nil {
+		return p
+	}
+	return filepath.Clean(path)
+}
+
+// worktreeBranch names a worktree's branch, reporting a detached checkout as
+// "HEAD" so it reads the same as GitCurrentBranch does.
+func worktreeBranch(wt agent.Worktree) string {
+	if wt.Branch != "" {
+		return wt.Branch
+	}
+	return "HEAD"
+}
+
+// sourceRepos is every repo to enumerate worktrees from: the ones the managed
+// worktrees point back at, plus every git checkout directly under the source
+// root (so a repo with only hand-made worktrees is still covered). Each is
+// normalised to its main checkout, because a directory in the source root may
+// itself be a linked worktree — enumerating from there would report the repo's
+// real checkout as a removable worktree of itself.
+func sourceRepos(managed []managedWt) []string {
+	seen := map[string]bool{}
+	var dirs []string
+	add := func(dir string) {
+		if dir == "" {
+			return
+		}
+		if main, err := agent.SourceRepoDir(dir); err == nil {
+			dir = main
+		}
+		key := resolvePath(dir)
+		if seen[key] {
+			return
+		}
+		seen[key] = true
+		dirs = append(dirs, dir)
+	}
+	for _, w := range managed {
+		add(w.srcDir)
+	}
+	if sr, err := srcRoot(); err == nil {
+		if entries, err := os.ReadDir(sr); err == nil {
+			for _, e := range entries {
+				if !e.IsDir() {
+					continue
+				}
+				dir := filepath.Join(sr, e.Name())
+				if _, err := os.Stat(filepath.Join(dir, ".git")); err == nil {
+					add(dir)
+				}
+			}
+		}
+	}
+	sort.Strings(dirs)
+	return dirs
+}
+
+// sameDir compares two paths after resolving symlinks, because git reports
+// worktree paths fully resolved while our own paths may not be.
+func sameDir(a, b string) bool {
+	if a == b {
+		return true
+	}
+	ra, errA := filepath.EvalSymlinks(a)
+	rb, errB := filepath.EvalSymlinks(b)
+	return errA == nil && errB == nil && ra == rb
+}
+
+// underRoot reports whether path sits inside root, comparing resolved paths
+// because git hands back worktree paths with symlinks already resolved.
+func underRoot(path, root string) bool {
+	if r, err := filepath.EvalSymlinks(root); err == nil {
+		root = r
+	}
+	if p, err := filepath.EvalSymlinks(path); err == nil {
+		path = p
+	}
+	rel, err := filepath.Rel(root, path)
+	if err != nil {
+		return false
+	}
+	return rel != ".." && !strings.HasPrefix(rel, ".."+string(filepath.Separator))
+}
+
 // --- rm -------------------------------------------------------------------
 
 func newRmCmd() *cobra.Command {
@@ -294,7 +540,8 @@ 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")
@@ -309,20 +556,35 @@ func resolveWorktree(target string) (managedWt, error) {
 	}
 	abs, _ := filepath.Abs(target)
 	for _, w := range managed {
-		if w.path == target || w.path == abs || w.branch == target {
+		if w.path == target || w.path == abs || (w.branch != "" && w.branch == target) {
 			return w, nil
 		}
 	}
 	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 {
+	switch {
+	case wt.inspectErr != nil:
+		// No srcDir to act through and no idea what is in there; --include-keep
+		// must not turn that into a delete.
+		return fmt.Errorf("refusing to remove %s: git state unreadable: %w", wt.path, wt.inspectErr)
+	case wt.orphan:
+		return removeOrphanDir(out, wt)
+	case wt.missing:
+		// The working tree is already gone; only the registration is left.
+		_, _ = fmt.Fprintf(out, "pruned stale registration %s\n", wt.path)
+		return agent.GitWorktreePrune(wt.srcDir)
+	}
 	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 +596,34 @@ func removeWorktree(out io.Writer, wt managedWt, deleteBranch bool) error {
 	return agent.GitWorktreePrune(wt.srcDir)
 }
 
+// removeOrphanDir deletes a worktree directory whose backing repo is gone. git
+// cannot act on it, so this is a plain delete — confined to the worktree root so
+// a bad path can never reach a real checkout.
+func removeOrphanDir(out io.Writer, wt managedWt) error {
+	wr, err := worktreeRoot()
+	if err != nil {
+		return err
+	}
+	if !underRoot(wt.path, wr) || sameDir(wt.path, wr) {
+		return fmt.Errorf("refusing to delete %s: not inside the worktree root %s", wt.path, wr)
+	}
+	if err := os.RemoveAll(wt.path); err != nil {
+		return err
+	}
+	_, _ = fmt.Fprintf(out, "deleted orphaned worktree directory %s\n", wt.path)
+	return nil
+}
+
+// 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 +642,7 @@ 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 +1,480 @@
+package main
+
+import (
+	"encoding/json"
+	"errors"
+	"fmt"
+	"io"
+	"os"
+	"path/filepath"
+	"sort"
+	"strings"
+	"text/tabwriter"
+
+	"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
+	// unfetched explains why, so a verdict can say which it was.
+	unfetched string
+}
+
+// pruneOpts is the knob set runPrune is driven by.
+type pruneOpts struct {
+	apply            bool
+	keepBranches     bool
+	noFetch          bool
+	jsonOut          bool
+	includeKeep      bool
+	includeUnmanaged bool
+}
+
+// reportEntry is the --json shape: one object per worktree, mirroring the table.
+type reportEntry struct {
+	Repo    string `json:"repo"`
+	Branch  string `json:"branch"`
+	Path    string `json:"path"`
+	Verdict string `json:"verdict"`
+	Reason  string `json:"reason"`
+	Managed bool   `json:"managed"`
+	Applied bool   `json:"applied"`
+}
+
+func newPruneCmd() *cobra.Command {
+	var opts pruneOpts
+	cmd := &cobra.Command{
+		Use:          "prune",
+		Short:        "Classify worktrees and remove the ones whose work is safely upstream",
+		Long:         "prune inspects every worktree it can find — the managed ones under the worktree\nroot plus whatever `git worktree list` reports for the source checkouts — and\nclassifies each against git and its Gitea pull request. It reports and changes\nnothing unless --yes is given.",
+		SilenceUsage: true,
+		RunE: func(cmd *cobra.Command, args []string) error {
+			return runPrune(cmd.OutOrStdout(), cmd.ErrOrStderr(), pruneClient(), opts)
+		},
+	}
+	f := cmd.Flags()
+	f.BoolVar(&opts.apply, "yes", false, "Actually remove worktrees (default is a dry run)")
+	f.BoolVar(&opts.keepBranches, "keep-branches", false, "Never delete a local branch, whatever the classification")
+	f.BoolVar(&opts.noFetch, "no-fetch", false, "Do not fetch; judge against the refs already on disk")
+	f.BoolVar(&opts.jsonOut, "json", false, "Emit JSON instead of a table")
+	f.BoolVar(&opts.includeKeep, "include-keep", false, "Dangerous: also remove worktrees classified keep (needs --yes). Destroys uncommitted changes and paused rebase/merge state, which no branch is carrying; only the branch itself survives")
+	f.BoolVar(&opts.includeUnmanaged, "include-unmanaged", false, "Also remove worktrees that live outside the worktree root")
+	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, errOut io.Writer, prs prLister, opts pruneOpts) error {
+	// In JSON mode stdout carries the document alone, so notes go to stderr.
+	notes := out
+	if opts.jsonOut {
+		notes = errOut
+	}
+
+	worktrees, err := allWorktrees()
+	if err != nil {
+		return err
+	}
+	if len(worktrees) == 0 {
+		if opts.jsonOut {
+			_, _ = fmt.Fprintln(out, "[]")
+			return nil
+		}
+		_, _ = fmt.Fprintln(out, "no managed worktrees")
+		return nil
+	}
+
+	results := classifyAll(notes, prs, worktrees, opts)
+	if err := report(out, results, opts); err != nil {
+		return err
+	}
+	if !opts.apply {
+		if !opts.jsonOut {
+			_, _ = fmt.Fprintln(out, "dry run: nothing removed (pass --yes to apply)")
+		}
+		return nil
+	}
+	return applyPrune(notes, results, opts)
+}
+
+// classifyAll groups worktrees by source repo so each repo is fetched and its
+// PRs listed once, then classifies every worktree against that repo's state.
+func classifyAll(notes io.Writer, prs prLister, worktrees []managedWt, opts pruneOpts) []pruneResult {
+	byRepo := map[string][]managedWt{}
+	var results []pruneResult
+	for _, w := range worktrees {
+		switch {
+		case w.inspectErr != nil:
+			// Unknown is not gone: a checkout git refused to answer for keeps.
+			results = append(results, pruneResult{wt: w, verdict: verdictKeep, reason: "inspection failed: " + oneLine(w.inspectErr.Error())})
+		case w.orphan:
+			results = append(results, pruneResult{wt: w, verdict: verdictRemove, reason: "backing repo gone, no git state to read"})
+		default:
+			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)
+
+	for _, srcDir := range srcDirs {
+		ctx, err := newRepoCtx(notes, prs, srcDir, opts.noFetch)
+		if err != nil {
+			for _, w := range byRepo[srcDir] {
+				results = append(results, pruneResult{wt: w, verdict: verdictKeep, reason: "repo state unknown: " + oneLine(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: " + oneLine(err.Error())}
+			}
+			results = append(results, res)
+		}
+	}
+	sort.SliceStable(results, func(i, j int) bool {
+		if results[i].wt.repo != results[j].wt.repo {
+			return results[i].wt.repo < results[j].wt.repo
+		}
+		return results[i].wt.path < results[j].wt.path
+	})
+	return results
+}
+
+// report writes the classification as a table or as JSON.
+func report(out io.Writer, results []pruneResult, opts pruneOpts) error {
+	if opts.jsonOut {
+		entries := make([]reportEntry, 0, len(results))
+		for _, r := range results {
+			entries = append(entries, reportEntry{
+				Repo:    r.wt.repo,
+				Branch:  r.wt.branch,
+				Path:    r.wt.path,
+				Verdict: plannedVerdict(r, opts),
+				Reason:  r.reason,
+				Managed: r.wt.managed,
+				Applied: opts.apply && willRemove(r, opts),
+			})
+		}
+		enc := json.NewEncoder(out)
+		enc.SetIndent("", "  ")
+		return enc.Encode(entries)
+	}
+
+	// Padded with spaces only, so the table reads the same with or without a TTY.
+	tw := tabwriter.NewWriter(out, 0, 0, 2, ' ', 0)
+	_, _ = fmt.Fprintln(tw, "REPO\tBRANCH\tPATH\tVERDICT\tREASON")
+	for _, r := range results {
+		_, _ = fmt.Fprintf(tw, "%s\t%s\t%s\t%s\t%s\n",
+			dash(r.wt.repo), dash(r.wt.branch), abbrevHome(r.wt.path), plannedVerdict(r, opts), r.reason)
+	}
+	return tw.Flush()
+}
+
+// oneLine flattens a git error onto a single line so one row stays one row.
+func oneLine(s string) string {
+	return strings.Join(strings.Fields(s), " ")
+}
+
+func dash(s string) string {
+	if s == "" {
+		return "-"
+	}
+	return s
+}
+
+// abbrevHome shortens $HOME to ~ so paths do not dominate the table.
+func abbrevHome(path string) string {
+	home, err := os.UserHomeDir()
+	if err != nil || home == "" || !strings.HasPrefix(path, home+string(filepath.Separator)) {
+		return path
+	}
+	return "~" + path[len(home):]
+}
+
+// applyPrune performs the removals the classification authorised, reporting each
+// one and collecting failures so one bad worktree does not stop the rest.
+func applyPrune(out io.Writer, results []pruneResult, opts pruneOpts) error {
+	var errs []error
+	for _, r := range results {
+		if !willRemove(r, opts) {
+			if !r.wt.managed && r.verdict != verdictKeep {
+				_, _ = fmt.Fprintf(out, "skipped %s: outside the worktree root (pass --include-unmanaged)\n", r.wt.path)
+			}
+			continue
+		}
+		if r.verdict == verdictKeep {
+			_, _ = fmt.Fprintf(out, "warn: removing %s despite %q (--include-keep)\n", r.wt.path, r.reason)
+		}
+		deleteBranch := plannedVerdict(r, opts) == verdictRemoveBranch
+		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...)
+}
+
+// willRemove is the single gate on destruction: a keep verdict needs
+// --include-keep, and a worktree outside the worktree root needs
+// --include-unmanaged.
+func willRemove(r pruneResult, opts pruneOpts) bool {
+	if !r.wt.managed && !opts.includeUnmanaged {
+		return false
+	}
+	if r.verdict == verdictKeep {
+		return opts.includeKeep
+	}
+	return true
+}
+
+// plannedVerdict is what will actually happen, so no flag prints an action it
+// will not perform. A detached HEAD has no branch to delete, and a keep forced
+// through with --include-keep never takes its branch with it.
+func plannedVerdict(r pruneResult, opts pruneOpts) string {
+	if r.verdict != verdictRemoveBranch {
+		return r.verdict
+	}
+	if opts.keepBranches || r.wt.detached {
+		return verdictRemove
+	}
+	return verdictRemoveBranch
+}
+
+// newRepoCtx refreshes a source repo and collects the signals prune classifies
+// against. A failed or skipped fetch and an unreachable Gitea are 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, noFetch bool) (repoCtx, error) {
+	ctx := repoCtx{srcDir: srcDir, prs: map[string]agent.PullRequest{}}
+	repo := filepath.Base(srcDir)
+	if noFetch {
+		ctx.unfetched = "fetch skipped"
+		_, _ = fmt.Fprintf(out, "warn: fetch %s skipped (--no-fetch, remote state unverified)\n", repo)
+	} else if err := agent.GitFetchPrune(srcDir, "origin", credentialHelperArgs()...); err != nil {
+		ctx.unfetched = "fetch failed"
+		_, _ = 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. Removal must never destroy state that
+// exists nowhere else: a vanished working tree is the one case with nothing to
+// lose, a locked or mid-rebase checkout holds sequencer state git itself refuses
+// to discard, and a dirty checkout or a detached HEAD with unique commits holds
+// work no branch is carrying. Past those guards, provably-upstream work loses its
+// branch too, and anything unproven keeps its branch so no commit becomes
+// 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}
+	if wt.missing {
+		res.verdict, res.reason = verdictRemove, "working tree gone, stale registration only"
+		return res, nil
+	}
+	if wt.locked {
+		res.verdict, res.reason = verdictKeep, "locked"
+		return res, nil
+	}
+	op, err := agent.GitInProgressOp(wt.path)
+	if err != nil {
+		return res, err
+	}
+	if op != "" {
+		res.verdict, res.reason = verdictKeep, op+" in progress"
+		return res, nil
+	}
+	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
+	}
+
+	local, err := agent.GitCommitsNotOnRemotes(wt.path)
+	if err != nil {
+		return res, err
+	}
+	if wt.detached && local > 0 {
+		res.verdict, res.reason = verdictKeep, fmt.Sprintf("detached HEAD carrying %s on no remote", commitCount(local))
+		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, %s so %s is unverified", pr.Number, ctx.unfetched, 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, %s so %s is unverified", pr.Number, ctx.unfetched, 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"
+	}
+	if res.verdict == verdictRemove && local > 0 {
+		res.reason += fmt.Sprintf(", %s on no remote so branch %s is kept", commitCount(local), wt.branch)
+	}
+	return res, nil
+}
+
+func commitCount(n int) string {
+	if n == 1 {
+		return "1 commit"
+	}
+	return fmt.Sprintf("%d commits", n)
+}

cmd/watchpr/main.go

@@ -7,6 +7,7 @@
 //	watchpr owner/repo#12 owner/repo:15
 //	watchpr --once --json owner/repo#12
 //	watchpr --interval 30s owner/repo#12
+//	watchpr --interval 30 owner/repo#12
 package main
 
 import (
@@ -34,7 +35,7 @@ func main() {
 // tests can invoke Execute and assert the exit behaviour without spawning a
 // process.
 func newRootCmd() *cobra.Command {
-	var interval time.Duration
+	var intervalFlag string
 	var once, jsonMode bool
 
 	root := &cobra.Command{
@@ -50,6 +51,10 @@ func newRootCmd() *cobra.Command {
 			if len(args) == 0 {
 				return fmt.Errorf("no PR references given (e.g. owner/repo#12)")
 			}
+			interval, err := agent.ParseDurationFlag("interval", intervalFlag)
+			if err != nil {
+				return err
+			}
 			refs := make([]agent.PRRef, 0, len(args))
 			for _, a := range args {
 				ref, err := agent.ParsePRRef(a)
@@ -58,10 +63,7 @@ func newRootCmd() *cobra.Command {
 				}
 				refs = append(refs, ref)
 			}
-			c, err := clientFor()
-			if err != nil {
-				return err
-			}
+			c := clientFor()
 			if once {
 				return runOnce(c, refs, jsonMode)
 			}
@@ -71,7 +73,7 @@ func newRootCmd() *cobra.Command {
 	root.SetVersionTemplate("{{.Version}}\n")
 
 	f := root.Flags()
-	f.DurationVar(&interval, "interval", 60*time.Second, "Polling interval")
+	f.StringVar(&intervalFlag, "interval", "60s", "Polling interval: a duration (30s, 2m, 1h30m) or a bare number of seconds")
 	f.BoolVar(&once, "once", false, "Check once, print current state, and exit")
 	f.BoolVar(&jsonMode, "json", false, "Emit JSON")
 
@@ -84,12 +86,16 @@ func newRootCmd() *cobra.Command {
 	return root
 }
 
-func clientFor() (*agent.GiteaClient, error) {
+// clientFor builds the Gitea client. Watching public repos works anonymously,
+// so an unavailable token is a warning, not a failure; a poll that is actually
+// rejected re-mints then.
+func clientFor() *agent.GiteaClient {
 	token, err := agent.GiteaToken()
 	if err != nil {
-		return nil, err
+		fmt.Fprintf(os.Stderr, "warning: no Gitea token (%v); polling anonymously\n", err)
+		token = ""
 	}
-	return agent.NewGiteaClient(token), nil
+	return agent.NewGiteaClient(token)
 }
 
 // runOnce fetches and prints the current state of each PR, then exits 0.
@@ -131,6 +137,12 @@ func runWatch(c *agent.GiteaClient, refs []agent.PRRef, interval time.Duration,
 
 	res, err := agent.Watch(c, refs, login, ticker.C, onBaseline, onError)
 	if err != nil {
+		if agent.IsAuthError(err) {
+			return fmt.Errorf("gitea authentication failed after re-minting the token, watch aborted: %w", err)
+		}
+		if agent.IsPRGone(err) {
+			return fmt.Errorf("PR no longer visible (repo deleted, renamed, or made private), watch aborted: %w", err)
+		}
 		return err
 	}
 	report(res.Ref.String(), res.Reason, res.State, jsonMode)

cmd/watchpr/main_test.go

@@ -2,6 +2,9 @@ package main
 
 import (
 	"io"
+	"net/http"
+	"net/http/httptest"
+	"strings"
 	"testing"
 )
 
@@ -28,3 +31,77 @@ func TestExecuteNoArgsErrors(t *testing.T) {
 		t.Fatal("Execute() = nil, want error when no PR references are given")
 	}
 }
+
+// Watching a public repo with no credentials available must work: the failed
+// mint is a warning, the poll goes out unauthenticated, and the command exits 0.
+func TestOnceRunsAnonymouslyWhenNoTokenIsAvailable(t *testing.T) {
+	vault := httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
+		w.WriteHeader(http.StatusForbidden)
+	}))
+	defer vault.Close()
+
+	authHeaders := 0
+	mux := http.NewServeMux()
+	mux.HandleFunc("/api/v1/repos/unkin/repo/pulls/7", func(w http.ResponseWriter, r *http.Request) {
+		if r.Header.Get("Authorization") != "" {
+			authHeaders++
+		}
+		_, _ = io.WriteString(w, `{"number":7,"state":"open","mergeable":true,"head":{"sha":"cafebabe"}}`)
+	})
+	mux.HandleFunc("/api/v1/repos/unkin/repo/commits/cafebabe/status", func(w http.ResponseWriter, r *http.Request) {
+		_, _ = io.WriteString(w, `{"state":"success"}`)
+	})
+	mux.HandleFunc("/api/v1/repos/unkin/repo/issues/7/comments", func(w http.ResponseWriter, r *http.Request) {
+		_, _ = io.WriteString(w, `[]`)
+	})
+	gitea := httptest.NewServer(mux)
+	defer gitea.Close()
+
+	t.Setenv("VAULT_ADDR", vault.URL)
+	t.Setenv("GITEA_URL", gitea.URL)
+
+	cmd := newRootCmd()
+	cmd.SetArgs([]string{"--once", "unkin/repo#7"})
+	cmd.SetOut(io.Discard)
+	cmd.SetErr(io.Discard)
+	if err := cmd.Execute(); err != nil {
+		t.Fatalf("anonymous --once should succeed without a token: %v", err)
+	}
+	if authHeaders != 0 {
+		t.Errorf("sent %d Authorization headers, want none", authHeaders)
+	}
+}
+
+// A bare integer interval means seconds and must survive flag parsing: the
+// command should fail on the missing PR reference, not on the flag value.
+func TestExecuteBareIntervalIsSeconds(t *testing.T) {
+	cmd := newRootCmd()
+	cmd.SetArgs([]string{"--interval", "15"})
+	cmd.SetOut(io.Discard)
+	cmd.SetErr(io.Discard)
+	err := cmd.Execute()
+	if err == nil {
+		t.Fatal("Execute() = nil, want the no-references error")
+	}
+	if !strings.Contains(err.Error(), "no PR references given") {
+		t.Fatalf("Execute() error = %v, want the no-references error", err)
+	}
+}
+
+// An unparseable interval is rejected before any Vault/Gitea call, with an
+// error naming the flag and showing valid forms.
+func TestExecuteBadIntervalErrors(t *testing.T) {
+	cmd := newRootCmd()
+	cmd.SetArgs([]string{"--interval", "soon", "unkin/repo#1"})
+	cmd.SetOut(io.Discard)
+	cmd.SetErr(io.Discard)
+	err := cmd.Execute()
+	if err == nil {
+		t.Fatal("Execute() = nil, want error for an unparseable --interval")
+	}
+	for _, want := range []string{"--interval", "30s"} {
+		if !strings.Contains(err.Error(), want) {
+			t.Errorf("Execute() error %q does not mention %q", err, want)
+		}
+	}
+}

internal/agent/client_test.go

@@ -2,9 +2,12 @@ package agent
 
 import (
 	"encoding/json"
+	"errors"
+	"fmt"
 	"io"
 	"net/http"
 	"net/http/httptest"
+	"strings"
 	"testing"
 )
 
@@ -95,6 +98,80 @@ func TestCreatePRRequestBody(t *testing.T) {
 	}
 }
 
+// An edit must send only the fields it was given: Gitea overwrites whatever
+// key it receives, so an omitted --title arriving as "" would blank the title.
+func TestEditPRSendsOnlySuppliedFields(t *testing.T) {
+	title, body, empty := "new title", "new body", ""
+	tests := []struct {
+		name string
+		opts EditPROptions
+		want map[string]any
+	}{
+		{"body only", EditPROptions{Body: &body}, map[string]any{"body": "new body"}},
+		{"title only", EditPROptions{Title: &title}, map[string]any{"title": "new title"}},
+		{"both", EditPROptions{Title: &title, Body: &body}, map[string]any{"title": "new title", "body": "new body"}},
+		{"explicit empty body is sent", EditPROptions{Body: &empty}, map[string]any{"body": ""}},
+	}
+	for _, tt := range tests {
+		t.Run(tt.name, func(t *testing.T) {
+			var gotBody map[string]any
+			var gotMethod, gotPath string
+			mux := http.NewServeMux()
+			mux.HandleFunc("/api/v1/repos/unkin/repo/pulls/7", func(w http.ResponseWriter, r *http.Request) {
+				gotMethod, gotPath = r.Method, r.URL.Path
+				_ = json.NewDecoder(r.Body).Decode(&gotBody)
+				_, _ = io.WriteString(w, `{"number":7,"title":"new title","html_url":"https://git.unkin.net/unkin/repo/pulls/7"}`)
+			})
+			srv := httptest.NewServer(mux)
+			defer srv.Close()
+
+			c := &GiteaClient{BaseURL: srv.URL, Token: "gitea-abc", HTTP: srv.Client()}
+			pr, err := c.EditPR("unkin/repo", 7, tt.opts)
+			if err != nil {
+				t.Fatalf("EditPR: %v", err)
+			}
+			if gotMethod != http.MethodPatch {
+				t.Errorf("method = %s, want PATCH", gotMethod)
+			}
+			if gotPath != "/api/v1/repos/unkin/repo/pulls/7" {
+				t.Errorf("path = %q", gotPath)
+			}
+			if len(gotBody) != len(tt.want) {
+				t.Errorf("payload = %v, want exactly the supplied fields %v", gotBody, tt.want)
+			}
+			for k, v := range tt.want {
+				if gotBody[k] != v {
+					t.Errorf("payload[%q] = %v, want %v", k, gotBody[k], v)
+				}
+			}
+			if pr.Number != 7 || pr.HTMLURL == "" {
+				t.Errorf("parsed PR = %+v", pr)
+			}
+		})
+	}
+}
+
+// A non-2xx must surface the API's own message rather than a bare status.
+func TestEditPRAPIError(t *testing.T) {
+	mux := http.NewServeMux()
+	mux.HandleFunc("/api/v1/repos/unkin/repo/pulls/7", func(w http.ResponseWriter, r *http.Request) {
+		w.WriteHeader(http.StatusNotFound)
+		_, _ = io.WriteString(w, `{"message":"not found","url":"https://git.unkin.net/api/swagger","errors":null}`)
+	})
+	srv := httptest.NewServer(mux)
+	defer srv.Close()
+
+	title := "new title"
+	c := &GiteaClient{BaseURL: srv.URL, Token: "t", HTTP: srv.Client()}
+	_, err := c.EditPR("unkin/repo", 7, EditPROptions{Title: &title})
+	if err == nil {
+		t.Fatal("expected error on 404")
+	}
+	if !strings.Contains(err.Error(), `"message":"not found"`) {
+		t.Errorf("error %q should carry the API message", err)
+	}
+}
+
 func TestCreateComment(t *testing.T) {
 	var gotBody map[string]string
 	mux := http.NewServeMux()
@@ -217,6 +294,100 @@ 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) {
@@ -231,3 +402,219 @@ func TestGiteaAPIError(t *testing.T) {
 		t.Fatal("expected error on 422")
 	}
 }
+
+// expiringGitea serves the PR endpoint, rejecting every token other than
+// wantToken with a 401 exactly as Gitea does once a Vault-minted token expires.
+// It records the tokens it saw, newest last.
+func expiringGitea(t *testing.T, wantToken string, seen *[]string) *httptest.Server {
+	t.Helper()
+	mux := http.NewServeMux()
+	mux.HandleFunc("/api/v1/repos/unkin/repo/pulls/7", func(w http.ResponseWriter, r *http.Request) {
+		tok := strings.TrimPrefix(r.Header.Get("Authorization"), "token ")
+		*seen = append(*seen, tok)
+		if tok != wantToken {
+			w.WriteHeader(http.StatusUnauthorized)
+			_, _ = io.WriteString(w, `{"message":"invalid username, password or token"}`)
+			return
+		}
+		_, _ = io.WriteString(w, `{"number":7,"state":"open","mergeable":true,"head":{"sha":"cafebabe"}}`)
+	})
+	return httptest.NewServer(mux)
+}
+
+// The production failure: the token expired mid-run. The client must re-mint
+// once and replay the request with the fresh token.
+func TestExpiredTokenIsRemintedAndRetried(t *testing.T) {
+	var seen []string
+	srv := expiringGitea(t, "fresh", &seen)
+	defer srv.Close()
+
+	refreshes := 0
+	c := &GiteaClient{BaseURL: srv.URL, Token: "stale", HTTP: srv.Client(),
+		Refresh: func() (string, error) { refreshes++; return "fresh", nil }}
+
+	pr, err := c.GetPR("unkin/repo", 7)
+	if err != nil {
+		t.Fatalf("GetPR after re-mint: %v", err)
+	}
+	if pr.Number != 7 {
+		t.Errorf("PR number = %d, want 7", pr.Number)
+	}
+	if refreshes != 1 {
+		t.Errorf("refreshes = %d, want 1", refreshes)
+	}
+	if len(seen) != 2 || seen[0] != "stale" || seen[1] != "fresh" {
+		t.Errorf("tokens seen = %v, want [stale fresh]", seen)
+	}
+	if c.Token != "fresh" {
+		t.Errorf("client token = %q, want the refreshed token", c.Token)
+	}
+}
+
+// A fresh token that is also rejected is a real auth failure: report it as one
+// rather than re-minting forever.
+func TestAuthFailureSurvivesRemint(t *testing.T) {
+	var seen []string
+	srv := expiringGitea(t, "never-issued", &seen)
+	defer srv.Close()
+
+	refreshes := 0
+	c := &GiteaClient{BaseURL: srv.URL, Token: "stale", HTTP: srv.Client(),
+		Refresh: func() (string, error) { refreshes++; return "still-bad", nil }}
+
+	_, err := c.GetPR("unkin/repo", 7)
+	if err == nil {
+		t.Fatal("GetPR should fail when the fresh token is rejected too")
+	}
+	if !IsAuthError(err) {
+		t.Errorf("IsAuthError(%v) = false, want true", err)
+	}
+	if refreshes != 1 {
+		t.Errorf("refreshes = %d, want 1 (re-mint exactly once)", refreshes)
+	}
+	if len(seen) != 2 {
+		t.Errorf("requests = %d, want 2", len(seen))
+	}
+}
+
+// A refresh that itself fails must surface as an auth error, not as a silent
+// success or a bare Vault error.
+func TestRemintErrorIsReportedAsAuthFailure(t *testing.T) {
+	var seen []string
+	srv := expiringGitea(t, "fresh", &seen)
+	defer srv.Close()
+
+	c := &GiteaClient{BaseURL: srv.URL, Token: "stale", HTTP: srv.Client(),
+		Refresh: func() (string, error) { return "", errors.New("vault approle login: HTTP 503") }}
+
+	_, err := c.GetPR("unkin/repo", 7)
+	if err == nil || !IsAuthError(err) {
+		t.Fatalf("GetPR error = %v, want an auth error", err)
+	}
+	if !strings.Contains(err.Error(), "vault approle login") {
+		t.Errorf("error %q should name the re-mint failure", err)
+	}
+	if len(seen) != 1 {
+		t.Errorf("requests = %d, want 1 (no replay without a token)", len(seen))
+	}
+}
+
+// A 5xx is transient, not an auth problem: no re-mint, no retry, and the caller
+// keeps its existing retry behaviour.
+func TestServerErrorDoesNotRemint(t *testing.T) {
+	requests := 0
+	mux := http.NewServeMux()
+	mux.HandleFunc("/api/v1/repos/unkin/repo/pulls/7", func(w http.ResponseWriter, r *http.Request) {
+		requests++
+		w.WriteHeader(http.StatusBadGateway)
+	})
+	srv := httptest.NewServer(mux)
+	defer srv.Close()
+
+	refreshes := 0
+	c := &GiteaClient{BaseURL: srv.URL, Token: "t", HTTP: srv.Client(),
+		Refresh: func() (string, error) { refreshes++; return "fresh", nil }}
+
+	_, err := c.GetPR("unkin/repo", 7)
+	if err == nil {
+		t.Fatal("expected error on 502")
+	}
+	if IsAuthError(err) {
+		t.Errorf("502 must not be an auth error")
+	}
+	if refreshes != 0 || requests != 1 {
+		t.Errorf("refreshes = %d, requests = %d, want 0 and 1", refreshes, requests)
+	}
+}
+
+// The replayed request must carry the original body, not an empty one.
+func TestRemintReplaysRequestBody(t *testing.T) {
+	var bodies []CreatePROptions
+	mux := http.NewServeMux()
+	mux.HandleFunc("/api/v1/repos/unkin/repo/pulls", func(w http.ResponseWriter, r *http.Request) {
+		var body CreatePROptions
+		_ = json.NewDecoder(r.Body).Decode(&body)
+		bodies = append(bodies, body)
+		if strings.TrimPrefix(r.Header.Get("Authorization"), "token ") != "fresh" {
+			w.WriteHeader(http.StatusUnauthorized)
+			return
+		}
+		_, _ = io.WriteString(w, `{"number":7}`)
+	})
+	srv := httptest.NewServer(mux)
+	defer srv.Close()
+
+	c := &GiteaClient{BaseURL: srv.URL, Token: "stale", HTTP: srv.Client(),
+		Refresh: func() (string, error) { return "fresh", nil }}
+
+	if _, err := c.CreatePR("unkin/repo", CreatePROptions{Base: "main", Head: "feature", Title: "T", Body: "B"}); err != nil {
+		t.Fatalf("CreatePR: %v", err)
+	}
+	if len(bodies) != 2 {
+		t.Fatalf("requests = %d, want 2", len(bodies))
+	}
+	if bodies[1] != bodies[0] {
+		t.Errorf("replayed body = %+v, want %+v", bodies[1], bodies[0])
+	}
+}
+
+func TestIsAuthError(t *testing.T) {
+	tests := []struct {
+		status int
+		want   bool
+	}{
+		{http.StatusUnauthorized, true},
+		{http.StatusForbidden, true},
+		{http.StatusNotFound, false},
+		{http.StatusUnprocessableEntity, false},
+		{http.StatusBadGateway, false},
+	}
+	for _, tt := range tests {
+		err := error(&APIError{Method: "GET", Path: "/p", StatusCode: tt.status})
+		if got := IsAuthError(err); got != tt.want {
+			t.Errorf("IsAuthError(HTTP %d) = %v, want %v", tt.status, got, tt.want)
+		}
+	}
+	if IsAuthError(errors.New("dial tcp: timeout")) {
+		t.Errorf("a network error is not an auth error")
+	}
+}
+
+// Anonymous polling of a public repo is a supported mode: with no token the
+// client must send no Authorization header, and must never reach for Vault.
+func TestAnonymousPollingNeverMints(t *testing.T) {
+	authHeaders := 0
+	mux := http.NewServeMux()
+	mux.HandleFunc("/api/v1/repos/unkin/repo/pulls/7", func(w http.ResponseWriter, r *http.Request) {
+		if r.Header.Get("Authorization") != "" {
+			authHeaders++
+		}
+		_, _ = io.WriteString(w, `{"number":7,"state":"open","mergeable":true,"head":{"sha":"cafebabe"}}`)
+	})
+	mux.HandleFunc("/api/v1/repos/unkin/repo/commits/cafebabe/status", func(w http.ResponseWriter, r *http.Request) {
+		_, _ = io.WriteString(w, `{"state":"success"}`)
+	})
+	mux.HandleFunc("/api/v1/repos/unkin/repo/issues/7/comments", func(w http.ResponseWriter, r *http.Request) {
+		_, _ = io.WriteString(w, `[]`)
+	})
+	srv := httptest.NewServer(mux)
+	defer srv.Close()
+
+	refreshes := 0
+	c := &GiteaClient{BaseURL: srv.URL, HTTP: srv.Client(),
+		Refresh: func() (string, error) { refreshes++; return "", errors.New("vault unreachable") }}
+
+	st, err := FetchState(c, PRRef{Owner: "unkin", Repo: "repo", Number: 7}, "unkin-agent")
+	if err != nil {
+		t.Fatalf("anonymous FetchState: %v", err)
+	}
+	if st.State != "open" || st.CIStatus != "success" || st.HeadSHA != "cafebabe" {
+		t.Errorf("state = %+v", st)
+	}
+	if refreshes != 0 {
+		t.Errorf("refreshes = %d, want 0 (a 200 must never trigger a mint)", refreshes)
+	}
+	if authHeaders != 0 {
+		t.Errorf("sent %d Authorization headers, want none", authHeaders)
+	}
+}

internal/agent/git.go

@@ -2,10 +2,12 @@ package agent
 
 import (
 	"bytes"
+	"errors"
 	"fmt"
 	"os"
 	"os/exec"
 	"path/filepath"
+	"strconv"
 	"strings"
 )
 
@@ -24,6 +26,10 @@ type Worktree struct {
 	Branch   string // short branch name ("" when detached or bare)
 	Bare     bool
 	Detached bool
+	Locked   bool
+	// Prunable is git's own reason a registration is stale (e.g. "gitdir file
+	// points to non-existent location"); empty when the worktree is intact.
+	Prunable string
 }
 
 // runGit runs git with args, using dir as the working directory (empty = the
@@ -67,6 +73,14 @@ 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 +97,112 @@ 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
+}
+
+// GitDir returns the absolute path to the git directory backing the checkout at
+// dir (per-worktree, unlike GitCommonDir).
+func GitDir(dir string) (string, error) {
+	return runGit(dir, "rev-parse", "--path-format=absolute", "--git-dir")
+}
+
+// inProgressMarkers maps a sentinel inside the git dir to the operation it means
+// is half-finished. Such a checkout holds state that lives nowhere else.
+var inProgressMarkers = []struct{ path, op string }{
+	{"rebase-merge", "rebase"},
+	{"rebase-apply", "rebase"},
+	{"MERGE_HEAD", "merge"},
+	{"CHERRY_PICK_HEAD", "cherry-pick"},
+	{"REVERT_HEAD", "revert"},
+	{"BISECT_LOG", "bisect"},
+}
+
+// GitInProgressOp names the sequencer operation underway in the checkout at dir,
+// or "" when none is.
+func GitInProgressOp(dir string) (string, error) {
+	gitDir, err := GitDir(dir)
+	if err != nil {
+		return "", err
+	}
+	for _, m := range inProgressMarkers {
+		if _, err := os.Stat(filepath.Join(gitDir, m.path)); err == nil {
+			return m.op, nil
+		} else if !os.IsNotExist(err) {
+			return "", err
+		}
+	}
+	return "", nil
+}
+
+// GitCommitsNotOnRemotes counts commits reachable from HEAD that no
+// remote-tracking ref holds, i.e. work that exists only in this checkout.
+func GitCommitsNotOnRemotes(dir string) (int, error) {
+	out, err := runGit(dir, "rev-list", "--count", "HEAD", "--not", "--remotes")
+	if err != nil {
+		return 0, err
+	}
+	n, err := strconv.Atoi(strings.TrimSpace(out))
+	if err != nil {
+		return 0, fmt.Errorf("parse rev-list count %q: %w", out, err)
+	}
+	return n, 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 {
@@ -210,6 +330,18 @@ func ParseWorktreeList(out string) []Worktree {
 			if cur != nil {
 				cur.Detached = true
 			}
+		case "locked":
+			if cur != nil {
+				cur.Locked = true
+			}
+		case "prunable":
+			if cur != nil {
+				// git omits the reason when it has none, so record the flag itself.
+				cur.Prunable = val
+				if cur.Prunable == "" {
+					cur.Prunable = "prunable"
+				}
+			}
 		}
 	}
 	flush()

internal/agent/git_test.go

@@ -42,10 +42,20 @@ branch refs/heads/benvin/foo
 worktree /home/ben/.cache/agentws/repo__detached
 HEAD 3333333333333333333333333333333333333333
 detached
+
+worktree /home/ben/.cache/agentws/repo__gone
+HEAD 4444444444444444444444444444444444444444
+branch refs/heads/benvin/gone
+prunable gitdir file points to non-existent location
+
+worktree /home/ben/.cache/agentws/repo__held
+HEAD 5555555555555555555555555555555555555555
+branch refs/heads/benvin/held
+locked
 `
 	wts := ParseWorktreeList(out)
-	if len(wts) != 3 {
-		t.Fatalf("got %d worktrees, want 3: %+v", len(wts), wts)
+	if len(wts) != 5 {
+		t.Fatalf("got %d worktrees, want 5: %+v", len(wts), wts)
 	}
 	if wts[0].Branch != "main" || wts[0].Path != "/home/ben/src/prodenv/repo" {
 		t.Errorf("wt[0] = %+v", wts[0])
@@ -56,6 +66,104 @@ detached
 	if !wts[2].Detached || wts[2].Branch != "" {
 		t.Errorf("wt[2] = %+v, want detached with empty branch", wts[2])
 	}
+	if wts[3].Prunable != "gitdir file points to non-existent location" {
+		t.Errorf("wt[3].Prunable = %q", wts[3].Prunable)
+	}
+	if !wts[4].Locked || wts[4].Prunable != "" {
+		t.Errorf("wt[4] = %+v, want locked and not prunable", wts[4])
+	}
+}
+
+// A bare "prunable" with no reason still has to read as prunable.
+func TestParseWorktreeListPrunableWithoutReason(t *testing.T) {
+	wts := ParseWorktreeList("worktree /tmp/wt\nHEAD 1111111111111111111111111111111111111111\ndetached\nprunable\n")
+	if len(wts) != 1 || wts[0].Prunable == "" {
+		t.Errorf("ParseWorktreeList = %+v, want one prunable worktree", wts)
+	}
+}
+
+func TestGitInProgressOp(t *testing.T) {
+	srcDir := newTempRepos(t)
+	gitDir, err := GitDir(srcDir)
+	if err != nil {
+		t.Fatalf("GitDir: %v", err)
+	}
+	if op, err := GitInProgressOp(srcDir); err != nil || op != "" {
+		t.Fatalf("clean checkout: op = %q, err = %v", op, err)
+	}
+
+	tests := []struct {
+		marker string
+		dir    bool
+		want   string
+	}{
+		{"MERGE_HEAD", false, "merge"},
+		{"CHERRY_PICK_HEAD", false, "cherry-pick"},
+		{"REVERT_HEAD", false, "revert"},
+		{"BISECT_LOG", false, "bisect"},
+		{"rebase-merge", true, "rebase"},
+		{"rebase-apply", true, "rebase"},
+	}
+	for _, tt := range tests {
+		path := filepath.Join(gitDir, tt.marker)
+		if tt.dir {
+			if err := os.Mkdir(path, 0o755); err != nil {
+				t.Fatal(err)
+			}
+		} else if err := os.WriteFile(path, []byte("x\n"), 0o644); err != nil {
+			t.Fatal(err)
+		}
+		op, err := GitInProgressOp(srcDir)
+		if err != nil || op != tt.want {
+			t.Errorf("%s: op = %q, err = %v; want %q", tt.marker, op, err, tt.want)
+		}
+		if err := os.RemoveAll(path); err != nil {
+			t.Fatal(err)
+		}
+	}
+}
+
+// Commits that reached a remote-tracking ref are not unique local work; commits
+// made after the push are.
+func TestGitCommitsNotOnRemotes(t *testing.T) {
+	srcDir := newTempRepos(t)
+	if n, err := GitCommitsNotOnRemotes(srcDir); err != nil || n != 0 {
+		t.Fatalf("freshly cloned main: n = %d, err = %v; want 0", n, err)
+	}
+
+	if _, err := runGit(srcDir, "checkout", "-b", "benvin/x"); err != nil {
+		t.Fatal(err)
+	}
+	commitFile(t, srcDir, "a.txt", "a\n", "local a")
+	commitFile(t, srcDir, "b.txt", "b\n", "local b")
+	if n, err := GitCommitsNotOnRemotes(srcDir); err != nil || n != 2 {
+		t.Fatalf("two unpushed commits: n = %d, err = %v; want 2", n, err)
+	}
+
+	if _, err := runGit(srcDir, "push", "origin", "benvin/x"); err != nil {
+		t.Fatal(err)
+	}
+	if n, err := GitCommitsNotOnRemotes(srcDir); err != nil || n != 0 {
+		t.Fatalf("after push: n = %d, err = %v; want 0", n, err)
+	}
+
+	commitFile(t, srcDir, "c.txt", "c\n", "local c")
+	if n, err := GitCommitsNotOnRemotes(srcDir); err != nil || n != 1 {
+		t.Fatalf("one commit past the push: n = %d, err = %v; want 1", n, err)
+	}
+}
+
+func commitFile(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)
+	}
+	if _, err := runGit(dir, "add", "."); err != nil {
+		t.Fatal(err)
+	}
+	if _, err := runGit(dir, "commit", "-m", msg); err != nil {
+		t.Fatal(err)
+	}
 }
 
 // gitSeed sets a repo-local identity so commits work without global config.
@@ -185,6 +293,134 @@ 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

@@ -24,40 +24,75 @@ func (e *APIError) Error() string {
 	return fmt.Sprintf("gitea %s %s: HTTP %d: %s", e.Method, e.Path, e.StatusCode, e.Body)
 }
 
-// isNotFound reports whether err is a Gitea 404.
-func isNotFound(err error) bool {
+// IsNotFound reports whether err is a Gitea 404. Gitea hides repositories a
+// caller may not see behind a 404 rather than a 403, so this also covers a repo
+// that was renamed, deleted, or made private.
+func IsNotFound(err error) bool {
 	var apiErr *APIError
 	return errors.As(err, &apiErr) && apiErr.StatusCode == http.StatusNotFound
 }
 
+// IsAuthError reports whether err is a Gitea 401/403: the token is expired or
+// unauthorised, which retrying the same request cannot fix.
+func IsAuthError(err error) bool {
+	var apiErr *APIError
+	return errors.As(err, &apiErr) &&
+		(apiErr.StatusCode == http.StatusUnauthorized || apiErr.StatusCode == http.StatusForbidden)
+}
+
 // GiteaClient talks to the Gitea REST API as the agent user.
 type GiteaClient struct {
 	BaseURL string
 	Token   string
 	HTTP    *http.Client
+	// Refresh mints a replacement token when the current one is rejected; Vault's
+	// Gitea tokens expire in ~1h, far short of a watchpr run.
+	Refresh func() (string, error)
 }
 
 // NewGiteaClient builds a client from the configured base URL and a Vault-minted
-// token.
+// token, re-minting from Vault when that token expires.
 func NewGiteaClient(token string) *GiteaClient {
-	return &GiteaClient{BaseURL: GiteaURL(), Token: token, HTTP: httpClient}
+	return &GiteaClient{BaseURL: GiteaURL(), Token: token, HTTP: httpClient, Refresh: RefreshGiteaToken}
 }
 
+// do sends the request and, if the token was rejected, re-mints it once and
+// replays the request with the fresh token.
 func (c *GiteaClient) do(method, path string, body any, out any) error {
-	var reader io.Reader
+	var payload []byte
 	if body != nil {
 		b, err := json.Marshal(body)
 		if err != nil {
 			return err
 		}
-		reader = bytes.NewReader(b)
+		payload = b
+	}
+	err := c.attempt(method, path, payload, out)
+	if !IsAuthError(err) || c.Refresh == nil {
+		return err
+	}
+	token, refreshErr := c.Refresh()
+	if refreshErr != nil {
+		return fmt.Errorf("%w; re-minting token: %v", err, refreshErr)
+	}
+	c.Token = token
+	return c.attempt(method, path, payload, out)
+}
+
+func (c *GiteaClient) attempt(method, path string, body []byte, out any) error {
+	var reader io.Reader
+	if body != nil {
+		reader = bytes.NewReader(body)
 	}
 	url := strings.TrimRight(c.BaseURL, "/") + path
 	req, err := http.NewRequest(method, url, reader)
 	if err != nil {
 		return err
 	}
-	req.Header.Set("Authorization", "token "+c.Token)
+	// An empty token means anonymous access, which public repos serve fine.
+	if c.Token != "" {
+		req.Header.Set("Authorization", "token "+c.Token)
+	}
 	req.Header.Set("Accept", "application/json")
 	if body != nil {
 		req.Header.Set("Content-Type", "application/json")
@@ -102,10 +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"`
@@ -121,6 +214,24 @@ func (c *GiteaClient) CreatePR(repoPath string, opts CreatePROptions) (PullReque
 	return pr, err
 }
 
+// EditPROptions are the fields an edit may change. Pointers so an unset field
+// is omitted from the payload entirely, leaving that field as it is. The two
+// fields are not symmetric: Gitea only applies a title when it is non-empty,
+// so Title can be set but never cleared and a "" title is a silent no-op,
+// while a pointer to "" Body really does blank the body.
+type EditPROptions struct {
+	Title *string `json:"title,omitempty"`
+	Body  *string `json:"body,omitempty"`
+}
+
+// EditPR updates a pull request's title and/or body
+// (PATCH /api/v1/repos/{owner}/{repo}/pulls/{index}).
+func (c *GiteaClient) EditPR(repoPath string, number int, opts EditPROptions) (PullRequest, error) {
+	var pr PullRequest
+	err := c.do(http.MethodPatch, fmt.Sprintf("/api/v1/repos/%s/pulls/%d", repoPath, number), opts, &pr)
+	return pr, err
+}
+
 // GetPR fetches a single pull request.
 func (c *GiteaClient) GetPR(repoPath string, number int) (PullRequest, error) {
 	var pr PullRequest

internal/agent/parse.go

@@ -4,6 +4,7 @@ import (
 	"fmt"
 	"strconv"
 	"strings"
+	"time"
 )
 
 // PRRef identifies a single pull request by repository and number.
@@ -48,6 +49,71 @@ func ParsePRRef(s string) (PRRef, error) {
 	return PRRef{Owner: owner, Repo: repo, Number: n}, nil
 }
 
+// ParseDurationFlag parses a duration flag value, accepting either a Go
+// duration string ("30s", "1h30m") or a bare integer read as seconds ("15").
+// flag names the flag so the error says which value was rejected.
+func ParseDurationFlag(flag, value string) (time.Duration, error) {
+	s := strings.TrimSpace(value)
+	d, err := time.ParseDuration(s)
+	if err != nil {
+		n, nerr := strconv.Atoi(s)
+		if nerr != nil {
+			return 0, fmt.Errorf("invalid --%s value %q: want a duration such as 30s, 2m or 1h30m, or a bare number of seconds such as 15", flag, value)
+		}
+		d = time.Duration(n) * time.Second
+	}
+	if d <= 0 {
+		return 0, fmt.Errorf("invalid --%s value %q: must be greater than zero", flag, value)
+	}
+	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

@@ -1,6 +1,10 @@
 package agent
 
-import "testing"
+import (
+	"strings"
+	"testing"
+	"time"
+)
 
 func TestParsePRRef(t *testing.T) {
 	tests := []struct {
@@ -77,3 +81,103 @@ func TestParseRepo(t *testing.T) {
 		}
 	}
 }
+
+func TestParseDurationFlag(t *testing.T) {
+	tests := []struct {
+		in      string
+		want    time.Duration
+		wantErr bool
+	}{
+		{"15", 15 * time.Second, false},
+		{"15s", 15 * time.Second, false},
+		{"2m", 2 * time.Minute, false},
+		{"1h30m", 90 * time.Minute, false},
+		{"500ms", 500 * time.Millisecond, false},
+		{" 45 ", 45 * time.Second, false},
+		{"0", 0, true},
+		{"0s", 0, true},
+		{"-5", 0, true},
+		{"-5s", 0, true},
+		{"15x", 0, true},
+		{"", 0, true},
+	}
+	for _, tt := range tests {
+		got, err := ParseDurationFlag("interval", tt.in)
+		if tt.wantErr {
+			if err == nil {
+				t.Errorf("ParseDurationFlag(%q): expected error, got %v", tt.in, got)
+			}
+			continue
+		}
+		if err != nil {
+			t.Errorf("ParseDurationFlag(%q): unexpected error: %v", tt.in, err)
+			continue
+		}
+		if got != tt.want {
+			t.Errorf("ParseDurationFlag(%q) = %v, want %v", tt.in, got, tt.want)
+		}
+	}
+}
+
+// The error must name the flag and show valid forms instead of surfacing
+// time.ParseDuration's "missing unit" wording.
+func TestParseDurationFlagErrorMessage(t *testing.T) {
+	_, err := ParseDurationFlag("interval", "soon")
+	if err == nil {
+		t.Fatal("ParseDurationFlag(\"soon\"): expected error")
+	}
+	for _, want := range []string{"--interval", `"soon"`, "30s", "seconds"} {
+		if !strings.Contains(err.Error(), want) {
+			t.Errorf("error %q does not mention %q", err, want)
+		}
+	}
+}
+
+// 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)
+		}
+	}
+}

internal/agent/seedoauth.go

@@ -0,0 +1,183 @@
+package agent
+
+import (
+	"crypto/rand"
+	"encoding/base64"
+	"errors"
+	"fmt"
+	"sort"
+)
+
+const (
+	// OAuthClientIDKey, OAuthClientSecretKey and OAuthCookieSecretKey are the
+	// KV fields oauth2-proxy deployments read their credentials from.
+	OAuthClientIDKey     = "client_id"
+	OAuthClientSecretKey = "client_secret"
+	OAuthCookieSecretKey = "cookie_secret"
+
+	// oauthSecretBytes is the raw length of generated secrets. oauth2-proxy
+	// rejects a cookie secret that does not decode to exactly 32 bytes.
+	oauthSecretBytes = 32
+)
+
+// Per-key outcomes reported by SeedOAuth.
+const (
+	ActionCreated   = "created"
+	ActionKept      = "kept"
+	ActionRotated   = "rotated"
+	ActionUpdated   = "updated"
+	ActionPreserved = "preserved"
+)
+
+// SeedOAuthOptions configures SeedOAuth. The CLI supplies the defaults.
+type SeedOAuthOptions struct {
+	VaultAddr string
+	RoleID    string
+	KVMount   string
+	Path      string
+	ClientID  string
+	Rotate    bool
+}
+
+// SeedOAuthKey names a key present in the secret and what happened to it.
+type SeedOAuthKey struct {
+	Name   string
+	Action string
+}
+
+// SeedOAuthResult is the non-secret summary of a seed run. Version is zero when
+// nothing changed and no write was issued.
+type SeedOAuthResult struct {
+	KVMount string
+	Path    string
+	Keys    []SeedOAuthKey
+	Version int
+	Changed bool
+}
+
+// KeyNames returns the key names present in the secret after the run.
+func (r SeedOAuthResult) KeyNames() []string {
+	names := make([]string, 0, len(r.Keys))
+	for _, k := range r.Keys {
+		names = append(names, k.Name)
+	}
+	return names
+}
+
+// SeedOAuth makes a Vault KV-v2 path hold a complete oauth2-proxy credential
+// set: client_id from the caller, plus a client_secret and cookie_secret that
+// are generated only when absent (or when Rotate is set). It is a
+// read-modify-write, so keys it does not own survive untouched, and it writes
+// nothing when the secret is already correct. Secret material never leaves this
+// function: results and errors carry only key names.
+func SeedOAuth(o SeedOAuthOptions) (SeedOAuthResult, error) {
+	var res SeedOAuthResult
+	if o.Path == "" {
+		return res, errors.New("a KV-v2 path is required")
+	}
+	if o.ClientID == "" {
+		return res, errors.New("a client id is required")
+	}
+
+	vc, err := NewVaultClient(o.VaultAddr, o.RoleID)
+	if err != nil {
+		return res, fmt.Errorf("vault approle login failed against %s (check VAULT_ADDR and AGENT_APPROLE_ROLE_ID): %w", o.VaultAddr, err)
+	}
+
+	existing, err := vc.ReadKVOptional(o.KVMount, o.Path)
+	if err != nil {
+		if errors.Is(err, ErrVaultDenied) {
+			return res, fmt.Errorf("reading %s/%s denied: the agent AppRole policy does not grant read on this path (apply the terraform-vault policy change): %w", o.KVMount, o.Path, err)
+		}
+		return res, fmt.Errorf("reading %s/%s: %w", o.KVMount, o.Path, err)
+	}
+
+	data := make(map[string]any, len(existing)+3)
+	for k, v := range existing {
+		data[k] = v
+	}
+
+	var (
+		keys    []SeedOAuthKey
+		changed bool
+	)
+
+	current, _ := existing[OAuthClientIDKey].(string)
+	switch current {
+	case o.ClientID:
+		keys = append(keys, SeedOAuthKey{OAuthClientIDKey, ActionKept})
+	case "":
+		keys = append(keys, SeedOAuthKey{OAuthClientIDKey, ActionCreated})
+		changed = true
+	default:
+		keys = append(keys, SeedOAuthKey{OAuthClientIDKey, ActionUpdated})
+		changed = true
+	}
+	data[OAuthClientIDKey] = o.ClientID
+
+	for _, gen := range []struct {
+		name string
+		enc  *base64.Encoding
+	}{
+		// oauth2-proxy accepts a standard-base64 client secret, but the cookie
+		// secret goes into a cookie and must be URL-safe.
+		{OAuthClientSecretKey, base64.StdEncoding},
+		{OAuthCookieSecretKey, base64.RawURLEncoding},
+	} {
+		current, _ := existing[gen.name].(string)
+		if current != "" && !o.Rotate {
+			keys = append(keys, SeedOAuthKey{gen.name, ActionKept})
+			continue
+		}
+		value, err := randomSecret(gen.enc)
+		if err != nil {
+			return res, fmt.Errorf("generating %s: %w", gen.name, err)
+		}
+		action := ActionCreated
+		if current != "" {
+			action = ActionRotated
+		}
+		data[gen.name] = value
+		keys = append(keys, SeedOAuthKey{gen.name, action})
+		changed = true
+	}
+
+	var others []string
+	for k := range existing {
+		switch k {
+		case OAuthClientIDKey, OAuthClientSecretKey, OAuthCookieSecretKey:
+		default:
+			others = append(others, k)
+		}
+	}
+	sort.Strings(others)
+	for _, k := range others {
+		keys = append(keys, SeedOAuthKey{k, ActionPreserved})
+	}
+
+	res = SeedOAuthResult{KVMount: o.KVMount, Path: o.Path, Keys: keys}
+	if !changed {
+		return res, nil
+	}
+
+	version, err := vc.WriteKVAny(o.KVMount, o.Path, data)
+	if err != nil {
+		if errors.Is(err, ErrVaultDenied) {
+			return SeedOAuthResult{}, fmt.Errorf("writing %s/%s denied: the agent AppRole policy does not grant create/update on this path (apply the terraform-vault policy change): %w", o.KVMount, o.Path, err)
+		}
+		return SeedOAuthResult{}, fmt.Errorf("writing %s/%s: %w", o.KVMount, o.Path, err)
+	}
+	res.Version = version
+	res.Changed = true
+	return res, nil
+}
+
+// randomSecret returns oauthSecretBytes of crypto/rand entropy in the given
+// base64 encoding.
+func randomSecret(enc *base64.Encoding) (string, error) {
+	buf := make([]byte, oauthSecretBytes)
+	if _, err := rand.Read(buf); err != nil {
+		return "", err
+	}
+	return enc.EncodeToString(buf), nil
+}

internal/agent/seedoauth_test.go

@@ -0,0 +1,421 @@
+package agent
+
+import (
+	"encoding/base64"
+	"encoding/json"
+	"io"
+	"net/http"
+	"net/http/httptest"
+	"strconv"
+	"strings"
+	"testing"
+)
+
+const (
+	oauthPath           = "kubernetes/namespace/repospawner/default/oauth-credentials"
+	oauthClientID       = "mediamark-client-id"
+	existingClientSec   = "existing-client-secret-from-authentik"
+	existingCookieSec   = "existing-cookie-secret-value-abcdefghij"
+	oauthExtraKeyValue  = "extra-key-secret-value"
+	oauthVaultClientTok = "s.vaulttoken"
+)
+
+// oauthVaultStub is a KV-v2 stand-in that actually stores what is written, so
+// read-modify-write behaviour can be asserted end to end.
+type oauthVaultStub struct {
+	data        map[string]any
+	exists      bool
+	version     int
+	readStatus  int
+	writeStatus int
+	writes      []map[string]any
+}
+
+func newOAuthVaultStub() *oauthVaultStub {
+	return &oauthVaultStub{readStatus: http.StatusOK, writeStatus: http.StatusOK}
+}
+
+// seed makes the path exist with the given fields at version 1.
+func (v *oauthVaultStub) seed(data map[string]any) *oauthVaultStub {
+	v.data = data
+	v.exists = true
+	v.version = 1
+	return v
+}
+
+func (v *oauthVaultStub) server(t *testing.T) *httptest.Server {
+	t.Helper()
+	mux := http.NewServeMux()
+	mux.HandleFunc("/v1/auth/approle/login", func(w http.ResponseWriter, r *http.Request) {
+		var body map[string]string
+		_ = json.NewDecoder(r.Body).Decode(&body)
+		if _, ok := body["secret_id"]; ok {
+			t.Errorf("secret_id must not be sent")
+		}
+		_, _ = io.WriteString(w, `{"auth":{"client_token":"`+oauthVaultClientTok+`"}}`)
+	})
+	mux.HandleFunc("/v1/kv/data/"+oauthPath, func(w http.ResponseWriter, r *http.Request) {
+		if got := r.Header.Get("X-Vault-Token"); got != oauthVaultClientTok {
+			t.Errorf("X-Vault-Token = %q, want %q", got, oauthVaultClientTok)
+		}
+		switch r.Method {
+		case http.MethodGet:
+			if v.readStatus != http.StatusOK {
+				w.WriteHeader(v.readStatus)
+				_, _ = io.WriteString(w, `{"errors":["permission denied"]}`)
+				return
+			}
+			if !v.exists {
+				w.WriteHeader(http.StatusNotFound)
+				_, _ = io.WriteString(w, `{"errors":[]}`)
+				return
+			}
+			payload, _ := json.Marshal(map[string]any{
+				"data": map[string]any{"data": v.data, "metadata": map[string]any{"version": v.version}},
+			})
+			_, _ = w.Write(payload)
+		case http.MethodPost:
+			if v.writeStatus != http.StatusOK {
+				w.WriteHeader(v.writeStatus)
+				_, _ = io.WriteString(w, `{"errors":["permission denied"]}`)
+				return
+			}
+			var body struct {
+				Data map[string]any `json:"data"`
+			}
+			_ = json.NewDecoder(r.Body).Decode(&body)
+			v.writes = append(v.writes, body.Data)
+			v.data = body.Data
+			v.exists = true
+			v.version++
+			_, _ = io.WriteString(w, `{"data":{"version":`+strconv.Itoa(v.version)+`}}`)
+		default:
+			t.Errorf("unexpected method %s", r.Method)
+		}
+	})
+	srv := httptest.NewServer(mux)
+	t.Cleanup(srv.Close)
+	return srv
+}
+
+func oauthOpts(vaultURL string) SeedOAuthOptions {
+	return SeedOAuthOptions{
+		VaultAddr: vaultURL,
+		RoleID:    "role-xyz",
+		KVMount:   DefaultKVMount,
+		Path:      oauthPath,
+		ClientID:  oauthClientID,
+	}
+}
+
+// actions flattens a result into key -> action for order-independent asserts.
+func actions(res SeedOAuthResult) map[string]string {
+	m := make(map[string]string, len(res.Keys))
+	for _, k := range res.Keys {
+		m[k.Name] = k.Action
+	}
+	return m
+}
+
+func stringField(t *testing.T, data map[string]any, key string) string {
+	t.Helper()
+	s, ok := data[key].(string)
+	if !ok {
+		t.Fatalf("written %s = %v, want a string", key, data[key])
+	}
+	return s
+}
+
+// assertDecodesTo32 fails unless the value is base64 of exactly 32 bytes, which
+// is what oauth2-proxy requires of a cookie secret.
+func assertDecodesTo32(t *testing.T, enc *base64.Encoding, value, name string) {
+	t.Helper()
+	raw, err := enc.DecodeString(value)
+	if err != nil {
+		t.Fatalf("%s is not valid base64: %v", name, err)
+	}
+	if len(raw) != oauthSecretBytes {
+		t.Errorf("%s decodes to %d bytes, want %d", name, len(raw), oauthSecretBytes)
+	}
+}
+
+func TestSeedOAuthFreshCreate(t *testing.T) {
+	v := newOAuthVaultStub()
+	res, err := SeedOAuth(oauthOpts(v.server(t).URL))
+	if err != nil {
+		t.Fatalf("SeedOAuth: %v", err)
+	}
+	if !res.Changed || res.Version != 1 {
+		t.Errorf("Changed=%v Version=%d, want a first write at version 1", res.Changed, res.Version)
+	}
+	for key, want := range map[string]string{
+		OAuthClientIDKey:     ActionCreated,
+		OAuthClientSecretKey: ActionCreated,
+		OAuthCookieSecretKey: ActionCreated,
+	} {
+		if got := actions(res)[key]; got != want {
+			t.Errorf("%s action = %q, want %q", key, got, want)
+		}
+	}
+	if len(v.writes) != 1 {
+		t.Fatalf("%d writes, want exactly 1", len(v.writes))
+	}
+	w := v.writes[0]
+	if got := stringField(t, w, OAuthClientIDKey); got != oauthClientID {
+		t.Errorf("written client_id = %q, want %q", got, oauthClientID)
+	}
+	assertDecodesTo32(t, base64.StdEncoding, stringField(t, w, OAuthClientSecretKey), OAuthClientSecretKey)
+	assertDecodesTo32(t, base64.RawURLEncoding, stringField(t, w, OAuthCookieSecretKey), OAuthCookieSecretKey)
+}
+
+// The mediamark case: a client_secret already issued by Authentik must survive
+// while the missing keys are filled in.
+func TestSeedOAuthPreservesExistingClientSecret(t *testing.T) {
+	v := newOAuthVaultStub().seed(map[string]any{OAuthClientSecretKey: existingClientSec})
+	res, err := SeedOAuth(oauthOpts(v.server(t).URL))
+	if err != nil {
+		t.Fatalf("SeedOAuth: %v", err)
+	}
+	got := actions(res)
+	for key, want := range map[string]string{
+		OAuthClientIDKey:     ActionCreated,
+		OAuthClientSecretKey: ActionKept,
+		OAuthCookieSecretKey: ActionCreated,
+	} {
+		if got[key] != want {
+			t.Errorf("%s action = %q, want %q", key, got[key], want)
+		}
+	}
+	if len(v.writes) != 1 {
+		t.Fatalf("%d writes, want exactly 1", len(v.writes))
+	}
+	if s := stringField(t, v.writes[0], OAuthClientSecretKey); s != existingClientSec {
+		t.Errorf("client_secret was replaced, want the existing value kept")
+	}
+}
+
+func TestSeedOAuthPreservesOtherKeys(t *testing.T) {
+	v := newOAuthVaultStub().seed(map[string]any{
+		OAuthClientIDKey:     oauthClientID,
+		OAuthClientSecretKey: existingClientSec,
+		"redirect_url":       "https://mediamark.unkin.net/oauth2/callback",
+		"extra":              oauthExtraKeyValue,
+	})
+	res, err := SeedOAuth(oauthOpts(v.server(t).URL))
+	if err != nil {
+		t.Fatalf("SeedOAuth: %v", err)
+	}
+	got := actions(res)
+	for _, key := range []string{"redirect_url", "extra"} {
+		if got[key] != ActionPreserved {
+			t.Errorf("%s action = %q, want %q", key, got[key], ActionPreserved)
+		}
+	}
+	if len(v.writes) != 1 {
+		t.Fatalf("%d writes, want exactly 1", len(v.writes))
+	}
+	w := v.writes[0]
+	if stringField(t, w, "extra") != oauthExtraKeyValue {
+		t.Errorf("extra key was not written back unchanged")
+	}
+	if stringField(t, w, "redirect_url") != "https://mediamark.unkin.net/oauth2/callback" {
+		t.Errorf("redirect_url was not written back unchanged")
+	}
+}
+
+func TestSeedOAuthRotateRegenerates(t *testing.T) {
+	v := newOAuthVaultStub().seed(map[string]any{
+		OAuthClientIDKey:     oauthClientID,
+		OAuthClientSecretKey: existingClientSec,
+		OAuthCookieSecretKey: existingCookieSec,
+	})
+	o := oauthOpts(v.server(t).URL)
+	o.Rotate = true
+
+	res, err := SeedOAuth(o)
+	if err != nil {
+		t.Fatalf("SeedOAuth: %v", err)
+	}
+	got := actions(res)
+	for key, want := range map[string]string{
+		OAuthClientIDKey:     ActionKept,
+		OAuthClientSecretKey: ActionRotated,
+		OAuthCookieSecretKey: ActionRotated,
+	} {
+		if got[key] != want {
+			t.Errorf("%s action = %q, want %q", key, got[key], want)
+		}
+	}
+	if len(v.writes) != 1 {
+		t.Fatalf("%d writes, want exactly 1", len(v.writes))
+	}
+	w := v.writes[0]
+	if stringField(t, w, OAuthClientSecretKey) == existingClientSec {
+		t.Errorf("client_secret unchanged under --rotate")
+	}
+	if stringField(t, w, OAuthCookieSecretKey) == existingCookieSec {
+		t.Errorf("cookie_secret unchanged under --rotate")
+	}
+	assertDecodesTo32(t, base64.RawURLEncoding, stringField(t, w, OAuthCookieSecretKey), OAuthCookieSecretKey)
+}
+
+// A complete, correct secret must not produce a new KV version.
+func TestSeedOAuthIdempotentWritesNothing(t *testing.T) {
+	v := newOAuthVaultStub()
+	url := v.server(t).URL
+	if _, err := SeedOAuth(oauthOpts(url)); err != nil {
+		t.Fatalf("first SeedOAuth: %v", err)
+	}
+	res, err := SeedOAuth(oauthOpts(url))
+	if err != nil {
+		t.Fatalf("second SeedOAuth: %v", err)
+	}
+	if res.Changed || res.Version != 0 {
+		t.Errorf("Changed=%v Version=%d, want an unchanged result", res.Changed, res.Version)
+	}
+	if len(v.writes) != 1 {
+		t.Errorf("%d writes, want the second run to write nothing", len(v.writes))
+	}
+	for _, k := range res.Keys {
+		if k.Action != ActionKept {
+			t.Errorf("%s action = %q, want %q", k.Name, k.Action, ActionKept)
+		}
+	}
+}
+
+func TestSeedOAuthClientIDUpdated(t *testing.T) {
+	v := newOAuthVaultStub().seed(map[string]any{
+		OAuthClientIDKey:     "stale-client-id",
+		OAuthClientSecretKey: existingClientSec,
+		OAuthCookieSecretKey: existingCookieSec,
+	})
+	res, err := SeedOAuth(oauthOpts(v.server(t).URL))
+	if err != nil {
+		t.Fatalf("SeedOAuth: %v", err)
+	}
+	if got := actions(res)[OAuthClientIDKey]; got != ActionUpdated {
+		t.Errorf("client_id action = %q, want %q", got, ActionUpdated)
+	}
+	if len(v.writes) != 1 || stringField(t, v.writes[0], OAuthClientIDKey) != oauthClientID {
+		t.Errorf("writes = %v, want the new client_id written", v.writes)
+	}
+}
+
+func TestSeedOAuthLoginFailure(t *testing.T) {
+	mux := http.NewServeMux()
+	mux.HandleFunc("/v1/auth/approle/login", func(w http.ResponseWriter, r *http.Request) {
+		w.WriteHeader(http.StatusBadRequest)
+		_, _ = io.WriteString(w, `{"errors":["invalid role ID"]}`)
+	})
+	vs := httptest.NewServer(mux)
+	defer vs.Close()
+
+	_, err := SeedOAuth(oauthOpts(vs.URL))
+	if err == nil {
+		t.Fatal("SeedOAuth() = nil, want an approle login error")
+	}
+	if !strings.Contains(err.Error(), "approle login failed") {
+		t.Errorf("error = %v, want it to name the approle login", err)
+	}
+}
+
+func TestSeedOAuthReadDenied(t *testing.T) {
+	v := newOAuthVaultStub()
+	v.readStatus = http.StatusForbidden
+
+	_, err := SeedOAuth(oauthOpts(v.server(t).URL))
+	if err == nil {
+		t.Fatal("SeedOAuth() = nil, want a KV read error")
+	}
+	msg := err.Error()
+	if !strings.Contains(msg, oauthPath) || !strings.Contains(msg, "policy") {
+		t.Errorf("error = %v, want it to name the path and point at the policy", err)
+	}
+	if len(v.writes) != 0 {
+		t.Errorf("wrote %v, want no write when the read is denied", v.writes)
+	}
+}
+
+func TestSeedOAuthWriteDenied(t *testing.T) {
+	v := newOAuthVaultStub()
+	v.writeStatus = http.StatusForbidden
+
+	_, err := SeedOAuth(oauthOpts(v.server(t).URL))
+	if err == nil {
+		t.Fatal("SeedOAuth() = nil, want a KV write error")
+	}
+	msg := err.Error()
+	if !strings.Contains(msg, oauthPath) || !strings.Contains(msg, "create/update") {
+		t.Errorf("error = %v, want it to name the path and the missing capability", err)
+	}
+}
+
+// A missing path is normal (first seed), not a not-found error.
+func TestSeedOAuthMissingPathIsNotAnError(t *testing.T) {
+	v := newOAuthVaultStub()
+	if _, err := SeedOAuth(oauthOpts(v.server(t).URL)); err != nil {
+		t.Fatalf("SeedOAuth on a missing path: %v", err)
+	}
+}
+
+func TestSeedOAuthRequiresPathAndClientID(t *testing.T) {
+	v := newOAuthVaultStub()
+	url := v.server(t).URL
+	for name, mutate := range map[string]func(*SeedOAuthOptions){
+		"no path":      func(o *SeedOAuthOptions) { o.Path = "" },
+		"no client id": func(o *SeedOAuthOptions) { o.ClientID = "" },
+	} {
+		t.Run(name, func(t *testing.T) {
+			o := oauthOpts(url)
+			mutate(&o)
+			if _, err := SeedOAuth(o); err == nil {
+				t.Fatal("SeedOAuth() = nil, want a required-input error")
+			}
+		})
+	}
+}
+
+// No failure path may leak stored or generated secret material.
+func TestSeedOAuthErrorsNeverLeakSecrets(t *testing.T) {
+	cases := map[string]func(*oauthVaultStub){
+		"read denied":  func(v *oauthVaultStub) { v.readStatus = http.StatusForbidden },
+		"write denied": func(v *oauthVaultStub) { v.writeStatus = http.StatusForbidden },
+		"read error":   func(v *oauthVaultStub) { v.readStatus = http.StatusInternalServerError },
+		"write error":  func(v *oauthVaultStub) { v.writeStatus = http.StatusInternalServerError },
+	}
+	for name, mutate := range cases {
+		t.Run(name, func(t *testing.T) {
+			v := newOAuthVaultStub().seed(map[string]any{
+				OAuthClientSecretKey: existingClientSec,
+				OAuthCookieSecretKey: existingCookieSec,
+				"extra":              oauthExtraKeyValue,
+			})
+			mutate(v)
+			_, err := SeedOAuth(oauthOpts(v.server(t).URL))
+			if err == nil {
+				t.Fatal("SeedOAuth() = nil, want an error")
+			}
+			for _, secret := range []string{existingClientSec, existingCookieSec, oauthExtraKeyValue} {
+				if strings.Contains(err.Error(), secret) {
+					t.Errorf("error %q leaks a secret", err)
+				}
+			}
+		})
+	}
+}
+
+// The successful result carries key names and a version, never values.
+func TestSeedOAuthResultNeverCarriesSecrets(t *testing.T) {
+	v := newOAuthVaultStub().seed(map[string]any{OAuthClientSecretKey: existingClientSec})
+	res, err := SeedOAuth(oauthOpts(v.server(t).URL))
+	if err != nil {
+		t.Fatalf("SeedOAuth: %v", err)
+	}
+	rendered := strings.Join(append(res.KeyNames(), res.Path, res.KVMount), " ")
+	written := v.writes[0]
+	for _, key := range []string{OAuthClientSecretKey, OAuthCookieSecretKey} {
+		if value := stringField(t, written, key); strings.Contains(rendered, value) {
+			t.Errorf("result leaks the %s value", key)
+		}
+	}
+}

internal/agent/token.go

@@ -86,17 +86,31 @@ func AuthentikURL() string {
 }
 
 var (
-	tokenOnce  sync.Once
-	tokenValue string
-	tokenErr   error
+	tokenMu     sync.Mutex
+	tokenMinted bool
+	tokenValue  string
+	tokenErr    error
 )
 
 // GiteaToken returns a Gitea token, minting it via Vault AppRole on first call
 // and caching it in-process for the lifetime of the command.
 func GiteaToken() (string, error) {
-	tokenOnce.Do(func() {
+	tokenMu.Lock()
+	defer tokenMu.Unlock()
+	if !tokenMinted {
 		tokenValue, tokenErr = fetchGiteaToken(VaultAddr(), RoleID(), GiteaCredsPath())
-	})
+		tokenMinted = true
+	}
+	return tokenValue, tokenErr
+}
+
+// RefreshGiteaToken mints a fresh Gitea token and replaces the cached one, for
+// callers that outlive the ~1h token TTL.
+func RefreshGiteaToken() (string, error) {
+	tokenMu.Lock()
+	defer tokenMu.Unlock()
+	tokenValue, tokenErr = fetchGiteaToken(VaultAddr(), RoleID(), GiteaCredsPath())
+	tokenMinted = true
 	return tokenValue, tokenErr
 }
 

internal/agent/vaultkv.go

@@ -106,8 +106,40 @@ func (c *VaultClient) ReadKV(mount, path string) (map[string]any, error) {
 	return out.Data.Data, nil
 }
 
+// ReadKVOptional is ReadKV but treats a missing secret as empty: a 404 or a
+// deleted version (data: null) yields an empty map, not an error. Callers that
+// read-modify-write a path that may not exist yet use this.
+func (c *VaultClient) ReadKVOptional(mount, path string) (map[string]any, error) {
+	var out struct {
+		Data struct {
+			Data map[string]any `json:"data"`
+		} `json:"data"`
+	}
+	if err := c.do(http.MethodGet, kvDataPath(mount, path), nil, &out); err != nil {
+		if errors.Is(err, ErrVaultNotFound) {
+			return map[string]any{}, nil
+		}
+		return nil, err
+	}
+	if out.Data.Data == nil {
+		return map[string]any{}, nil
+	}
+	return out.Data.Data, nil
+}
+
 // WriteKV writes a KV-v2 secret and returns the version it created.
 func (c *VaultClient) WriteKV(mount, path string, data map[string]string) (int, error) {
+	fields := make(map[string]any, len(data))
+	for k, v := range data {
+		fields[k] = v
+	}
+	return c.WriteKVAny(mount, path, fields)
+}
+
+// WriteKVAny writes a KV-v2 secret whose fields are not all strings (so a
+// read-modify-write can put back values it did not author) and returns the
+// version it created.
+func (c *VaultClient) WriteKVAny(mount, path string, data map[string]any) (int, error) {
 	var out struct {
 		Data struct {
 			Version int `json:"version"`

internal/agent/watch.go

@@ -1,6 +1,20 @@
 package agent
 
-import "time"
+import (
+	"errors"
+	"fmt"
+	"time"
+)
+
+// errPRGone marks a 404 from the PR lookup itself. A 404 from any other endpoint
+// can be a proxy or ingress blip and is left to the ordinary failure cap.
+var errPRGone = errors.New("PR no longer visible")
+
+// IsPRGone reports whether err is a 404 from the PR lookup, meaning the PR is no
+// longer visible rather than one endpoint being briefly unreachable.
+func IsPRGone(err error) bool {
+	return errors.Is(err, errPRGone)
+}
 
 // PRState is a point-in-time snapshot of the PR attributes watchpr tracks.
 type PRState struct {
@@ -20,13 +34,16 @@ type PRState struct {
 func FetchState(c *GiteaClient, ref PRRef, agentLogin string) (PRState, error) {
 	pr, err := c.GetPR(ref.RepoPath(), ref.Number)
 	if err != nil {
+		if IsNotFound(err) {
+			return PRState{}, fmt.Errorf("%w: %w", errPRGone, err)
+		}
 		return PRState{}, err
 	}
 	// A 404 here means the head commit is gone (branch deleted after a squash/
 	// rebase merge); the PR object is still authoritative, so treat CI as absent
 	// rather than discarding the merge signal and hanging the watch loop.
 	ci, err := c.CommitStatus(ref.RepoPath(), pr.Head.Sha)
-	if err != nil && !isNotFound(err) {
+	if err != nil && !IsNotFound(err) {
 		return PRState{}, err
 	}
 	comments, err := c.ListComments(ref.RepoPath(), ref.Number)
@@ -78,12 +95,22 @@ func terminalState(st PRState) (bool, string) {
 	return false, ""
 }
 
+// MaxPollFailures is how many consecutive failed polls of the same PR are
+// tolerated before Watch gives up. The abort fires on the 20th failed tick, so
+// at watchpr's default 60s interval a watch rides out ~19 minutes of failure.
+const MaxPollFailures = 20
+
 // Watch establishes a baseline for each ref, then polls on every tick until a
 // tracked PR changes meaningfully, returning the first such change. A PR that is
 // already terminal (merged/closed) at baseline is reported immediately rather
-// than polled forever. Poll errors are handed to onError and never stop the
-// loop; only a baseline fetch error aborts. onBaseline, if set, fires once after
-// all baselines are captured and before the first tick.
+// than polled forever. Transient poll errors are handed to onError and the loop
+// continues, but never blindly: a baseline fetch error, an authentication
+// failure surviving a token re-mint, a 404 from the PR lookup itself (the repo
+// is gone, renamed, or no longer visible), and MaxPollFailures consecutive
+// failures of one PR all abort, because a watcher that sees nothing must not
+// look healthy.
+// onBaseline, if set, fires once after all baselines are captured and before the
+// first tick.
 func Watch(f StateFetcher, refs []PRRef, agentLogin string, ticks <-chan time.Time, onBaseline func(), onError func(PRRef, error)) (WatchResult, error) {
 	prev := make(map[string]PRState, len(refs))
 	for _, ref := range refs {
@@ -99,16 +126,25 @@ func Watch(f StateFetcher, refs []PRRef, agentLogin string, ticks <-chan time.Ti
 	if onBaseline != nil {
 		onBaseline()
 	}
+	fails := make(map[string]int, len(refs))
 	for range ticks {
 		for _, ref := range refs {
 			key := ref.String()
 			cur, err := f.FetchState(ref, agentLogin)
 			if err != nil {
+				if IsAuthError(err) || IsPRGone(err) {
+					return WatchResult{}, fmt.Errorf("polling %s: %w", key, err)
+				}
+				fails[key]++
 				if onError != nil {
 					onError(ref, err)
 				}
+				if fails[key] >= MaxPollFailures {
+					return WatchResult{}, fmt.Errorf("polling %s: giving up after %d consecutive failures: %w", key, fails[key], err)
+				}
 				continue
 			}
+			fails[key] = 0
 			if changed, reason := MeaningfulChange(prev[key], cur); changed {
 				return WatchResult{Ref: ref, Reason: reason, State: cur}, nil
 			}

internal/agent/watch_test.go

@@ -5,6 +5,7 @@ import (
 	"fmt"
 	"net/http"
 	"net/http/httptest"
+	"strings"
 	"sync/atomic"
 	"testing"
 	"time"
@@ -360,3 +361,443 @@ func TestCountNonAgentComments(t *testing.T) {
 		t.Errorf("countNonAgentComments = %d, want 2", n)
 	}
 }
+
+// The production failure: the Vault-minted token expired mid-watch and every
+// poll 401'd, which the loop logged as a warning and polled past forever. An
+// auth error that survived the client's re-mint must end the watch with an
+// error so watchpr exits non-zero instead of watching blind.
+func TestWatchAbortsOnAuthError(t *testing.T) {
+	open := base()
+	merged := base()
+	merged.State = "closed"
+	merged.Merged = true
+	f := &fakeFetcher{
+		states: []PRState{open, open, merged},
+		errs:   []error{nil, &APIError{Method: "GET", Path: "/p", StatusCode: 401, Body: "invalid token"}, nil},
+	}
+
+	warned := 0
+	ticks := make(chan time.Time, 2)
+	ticks <- time.Now()
+	ticks <- time.Now()
+	_, err := Watch(f, []PRRef{open.Ref}, "unkin-agent", ticks, nil, func(PRRef, error) { warned++ })
+	if err == nil {
+		t.Fatal("Watch should return the auth failure, not keep polling")
+	}
+	if !IsAuthError(err) {
+		t.Errorf("Watch error = %v, want an auth error", err)
+	}
+	if warned != 0 {
+		t.Errorf("auth failure was logged as a warning %d time(s); it must abort", warned)
+	}
+	if f.calls != 2 {
+		t.Errorf("fetch calls = %d, want 2 (baseline + the failing poll)", f.calls)
+	}
+}
+
+// A 5xx keeps its retry behaviour: warn and poll on.
+func TestWatchContinuesPastServerError(t *testing.T) {
+	open := base()
+	merged := base()
+	merged.State = "closed"
+	merged.Merged = true
+	f := &fakeFetcher{
+		states: []PRState{open, open, merged},
+		errs:   []error{nil, &APIError{Method: "GET", Path: "/p", StatusCode: 502, Body: "bad gateway"}, nil},
+	}
+
+	warned := 0
+	ticks := make(chan time.Time, 2)
+	ticks <- time.Now()
+	ticks <- time.Now()
+	res, err := Watch(f, []PRRef{open.Ref}, "unkin-agent", ticks, nil, func(PRRef, error) { warned++ })
+	if err != nil {
+		t.Fatalf("Watch: %v", err)
+	}
+	if warned != 1 {
+		t.Errorf("warnings = %d, want 1", warned)
+	}
+	if res.Reason != "PR merged" {
+		t.Errorf("reason = %q, want %q", res.Reason, "PR merged")
+	}
+}
+
+// The production failure: a watched repo was renamed mid-watch, so every poll
+// 404'd (Gitea hides a repo the caller may not see rather than 403ing) and the
+// loop warned past it forever while reporting nothing. A 404 on a tracked PR
+// must end the watch with an error naming that PR.
+func TestWatchAbortsOnMidRunNotFound(t *testing.T) {
+	const sha = "deadbeefdeadbeef"
+	var polls atomic.Int32
+
+	mux := http.NewServeMux()
+	mux.HandleFunc("/api/v1/repos/unkin/repo/pulls/7", func(w http.ResponseWriter, r *http.Request) {
+		if polls.Add(1) > 1 { // repo renamed/made private after the baseline
+			w.WriteHeader(http.StatusNotFound)
+			_, _ = fmt.Fprint(w, `{"message":"Not Found"}`)
+			return
+		}
+		_, _ = fmt.Fprintf(w, `{"number":7,"state":"open","merged":false,"mergeable":true,"head":{"sha":%q}}`, sha)
+	})
+	mux.HandleFunc("/api/v1/repos/unkin/repo/commits/"+sha+"/status", func(w http.ResponseWriter, r *http.Request) {
+		_, _ = fmt.Fprint(w, `{"state":"success"}`)
+	})
+	mux.HandleFunc("/api/v1/repos/unkin/repo/issues/7/comments", func(w http.ResponseWriter, r *http.Request) {
+		_, _ = fmt.Fprint(w, `[]`)
+	})
+	srv := httptest.NewServer(mux)
+	defer srv.Close()
+
+	c := &GiteaClient{BaseURL: srv.URL, Token: "t", HTTP: srv.Client()}
+	ref := PRRef{Owner: "unkin", Repo: "repo", Number: 7}
+
+	tk := time.NewTicker(5 * time.Millisecond)
+	defer tk.Stop()
+	var warned atomic.Int32
+	done := make(chan error, 1)
+	go func() {
+		_, err := Watch(c, []PRRef{ref}, "unkin-agent", tk.C, nil,
+			func(PRRef, error) { warned.Add(1) })
+		done <- err
+	}()
+
+	select {
+	case err := <-done:
+		if err == nil {
+			t.Fatal("Watch should abort on a mid-run 404, not keep polling")
+		}
+		if !IsNotFound(err) {
+			t.Errorf("Watch error = %v, want a 404", err)
+		}
+		if !strings.Contains(err.Error(), ref.String()) {
+			t.Errorf("Watch error = %v, want it to name %s", err, ref.String())
+		}
+		if n := warned.Load(); n != 0 {
+			t.Errorf("404 was logged as a warning %d time(s); it must abort", n)
+		}
+	case <-time.After(3 * time.Second):
+		t.Fatal("Watch hung: a vanished repo was warned past instead of aborting")
+	}
+}
+
+// A 404 from a sub-resource is not proof the PR is gone: an ingress can serve
+// one during a Gitea rolling restart. Only the PR lookup itself is authoritative,
+// so a comments 404 must warn and keep polling like any other transient failure,
+// and still catch the merge that lands afterwards.
+func TestWatchSurvivesCommentsNotFound(t *testing.T) {
+	const sha = "0badc0de0badc0de"
+	var polls atomic.Int32
+
+	mux := http.NewServeMux()
+	mux.HandleFunc("/api/v1/repos/unkin/repo/pulls/7", func(w http.ResponseWriter, r *http.Request) {
+		if polls.Add(1) >= 4 {
+			_, _ = fmt.Fprintf(w, `{"number":7,"state":"closed","merged":true,"mergeable":true,"head":{"sha":%q}}`, sha)
+			return
+		}
+		_, _ = fmt.Fprintf(w, `{"number":7,"state":"open","merged":false,"mergeable":true,"head":{"sha":%q}}`, sha)
+	})
+	mux.HandleFunc("/api/v1/repos/unkin/repo/commits/"+sha+"/status", func(w http.ResponseWriter, r *http.Request) {
+		_, _ = fmt.Fprint(w, `{"state":"success"}`)
+	})
+	mux.HandleFunc("/api/v1/repos/unkin/repo/issues/7/comments", func(w http.ResponseWriter, r *http.Request) {
+		if n := polls.Load(); n == 2 || n == 3 { // proxy blip across two polls
+			w.WriteHeader(http.StatusNotFound)
+			_, _ = fmt.Fprint(w, `{"message":"Not Found"}`)
+			return
+		}
+		_, _ = fmt.Fprint(w, `[]`)
+	})
+	srv := httptest.NewServer(mux)
+	defer srv.Close()
+
+	c := &GiteaClient{BaseURL: srv.URL, Token: "t", HTTP: srv.Client()}
+	ref := PRRef{Owner: "unkin", Repo: "repo", Number: 7}
+
+	tk := time.NewTicker(5 * time.Millisecond)
+	defer tk.Stop()
+	var warned atomic.Int32
+	type outcome struct {
+		res WatchResult
+		err error
+	}
+	done := make(chan outcome, 1)
+	go func() {
+		res, err := Watch(c, []PRRef{ref}, "unkin-agent", tk.C, nil,
+			func(PRRef, error) { warned.Add(1) })
+		done <- outcome{res, err}
+	}()
+
+	select {
+	case o := <-done:
+		if o.err != nil {
+			t.Fatalf("Watch: %v (a comments 404 must not be terminal)", o.err)
+		}
+		if o.res.Reason != "PR merged" {
+			t.Errorf("reason = %q, want %q", o.res.Reason, "PR merged")
+		}
+		if n := warned.Load(); n != 2 {
+			t.Errorf("warnings = %d, want 2", n)
+		}
+	case <-time.After(3 * time.Second):
+		t.Fatal("Watch hung: a comments 404 must warn and keep polling")
+	}
+}
+
+// A comments 404 costs a poll from the same budget as any other failure: it must
+// not be free, and a permanently 404ing sub-resource must still end the watch.
+func TestWatchCommentsNotFoundCountsTowardCap(t *testing.T) {
+	const sha = "1badc0de1badc0de"
+
+	mux := http.NewServeMux()
+	mux.HandleFunc("/api/v1/repos/unkin/repo/pulls/7", func(w http.ResponseWriter, r *http.Request) {
+		_, _ = fmt.Fprintf(w, `{"number":7,"state":"open","merged":false,"mergeable":true,"head":{"sha":%q}}`, sha)
+	})
+	mux.HandleFunc("/api/v1/repos/unkin/repo/commits/"+sha+"/status", func(w http.ResponseWriter, r *http.Request) {
+		_, _ = fmt.Fprint(w, `{"state":"success"}`)
+	})
+	var comments atomic.Int32
+	mux.HandleFunc("/api/v1/repos/unkin/repo/issues/7/comments", func(w http.ResponseWriter, r *http.Request) {
+		if comments.Add(1) > 1 { // healthy at baseline, gone from the first poll on
+			w.WriteHeader(http.StatusNotFound)
+			_, _ = fmt.Fprint(w, `{"message":"Not Found"}`)
+			return
+		}
+		_, _ = fmt.Fprint(w, `[]`)
+	})
+	srv := httptest.NewServer(mux)
+	defer srv.Close()
+
+	c := &GiteaClient{BaseURL: srv.URL, Token: "t", HTTP: srv.Client()}
+	ref := PRRef{Owner: "unkin", Repo: "repo", Number: 7}
+
+	tk := time.NewTicker(time.Millisecond)
+	defer tk.Stop()
+	var warned atomic.Int32
+	done := make(chan error, 1)
+	go func() {
+		_, err := Watch(c, []PRRef{ref}, "unkin-agent", tk.C, nil,
+			func(PRRef, error) { warned.Add(1) })
+		done <- err
+	}()
+
+	select {
+	case err := <-done:
+		if err == nil {
+			t.Fatal("Watch should give up once the comments 404 stops being transient")
+		}
+		if !strings.Contains(err.Error(), "consecutive failures") {
+			t.Errorf("Watch error = %v, want it to report the failure cap", err)
+		}
+		if n := warned.Load(); n != MaxPollFailures {
+			t.Errorf("warnings = %d, want %d", n, MaxPollFailures)
+		}
+	case <-time.After(3 * time.Second):
+		t.Fatal("Watch hung: a permanently 404ing comments endpoint must hit the cap")
+	}
+}
+
+// The PR lookup is the call whose 404 means the PR is gone, so it aborts on the
+// very first occurrence rather than spending the failure budget.
+func TestWatchAbortsOnFirstPRLookupNotFound(t *testing.T) {
+	const sha = "2badc0de2badc0de"
+	var polls atomic.Int32
+
+	mux := http.NewServeMux()
+	mux.HandleFunc("/api/v1/repos/unkin/repo/pulls/7", func(w http.ResponseWriter, r *http.Request) {
+		if polls.Add(1) > 1 {
+			w.WriteHeader(http.StatusNotFound)
+			_, _ = fmt.Fprint(w, `{"message":"Not Found"}`)
+			return
+		}
+		_, _ = fmt.Fprintf(w, `{"number":7,"state":"open","merged":false,"mergeable":true,"head":{"sha":%q}}`, sha)
+	})
+	mux.HandleFunc("/api/v1/repos/unkin/repo/commits/"+sha+"/status", func(w http.ResponseWriter, r *http.Request) {
+		_, _ = fmt.Fprint(w, `{"state":"success"}`)
+	})
+	mux.HandleFunc("/api/v1/repos/unkin/repo/issues/7/comments", func(w http.ResponseWriter, r *http.Request) {
+		_, _ = fmt.Fprint(w, `[]`)
+	})
+	srv := httptest.NewServer(mux)
+	defer srv.Close()
+
+	c := &GiteaClient{BaseURL: srv.URL, Token: "t", HTTP: srv.Client()}
+	ref := PRRef{Owner: "unkin", Repo: "repo", Number: 7}
+
+	tk := time.NewTicker(5 * time.Millisecond)
+	defer tk.Stop()
+	done := make(chan error, 1)
+	go func() {
+		_, err := Watch(c, []PRRef{ref}, "unkin-agent", tk.C, nil,
+			func(PRRef, error) { t.Errorf("a PR-lookup 404 must abort, not warn") })
+		done <- err
+	}()
+
+	select {
+	case err := <-done:
+		if !IsNotFound(err) {
+			t.Fatalf("Watch error = %v, want a 404", err)
+		}
+		if n := polls.Load(); n != 2 {
+			t.Errorf("PR fetches = %d, want 2 (baseline + the 404 that aborts)", n)
+		}
+	case <-time.After(3 * time.Second):
+		t.Fatal("Watch hung: a vanished PR must abort")
+	}
+}
+
+// A 5xx blip must not kill a long watch: it warns, keeps polling, and still
+// catches the merge that lands afterwards.
+func TestWatchSurvivesTransientServerError(t *testing.T) {
+	const sha = "feedfacefeedface"
+	var polls atomic.Int32
+
+	mux := http.NewServeMux()
+	mux.HandleFunc("/api/v1/repos/unkin/repo/pulls/7", func(w http.ResponseWriter, r *http.Request) {
+		switch n := polls.Add(1); {
+		case n == 2 || n == 3: // gateway blip across two polls
+			w.WriteHeader(http.StatusBadGateway)
+			_, _ = fmt.Fprint(w, `bad gateway`)
+		case n >= 4:
+			_, _ = fmt.Fprintf(w, `{"number":7,"state":"closed","merged":true,"mergeable":true,"head":{"sha":%q}}`, sha)
+		default:
+			_, _ = fmt.Fprintf(w, `{"number":7,"state":"open","merged":false,"mergeable":true,"head":{"sha":%q}}`, sha)
+		}
+	})
+	mux.HandleFunc("/api/v1/repos/unkin/repo/commits/"+sha+"/status", func(w http.ResponseWriter, r *http.Request) {
+		_, _ = fmt.Fprint(w, `{"state":"success"}`)
+	})
+	mux.HandleFunc("/api/v1/repos/unkin/repo/issues/7/comments", func(w http.ResponseWriter, r *http.Request) {
+		_, _ = fmt.Fprint(w, `[]`)
+	})
+	srv := httptest.NewServer(mux)
+	defer srv.Close()
+
+	c := &GiteaClient{BaseURL: srv.URL, Token: "t", HTTP: srv.Client()}
+	ref := PRRef{Owner: "unkin", Repo: "repo", Number: 7}
+
+	tk := time.NewTicker(5 * time.Millisecond)
+	defer tk.Stop()
+	var warned atomic.Int32
+	type outcome struct {
+		res WatchResult
+		err error
+	}
+	done := make(chan outcome, 1)
+	go func() {
+		res, err := Watch(c, []PRRef{ref}, "unkin-agent", tk.C, nil,
+			func(PRRef, error) { warned.Add(1) })
+		done <- outcome{res, err}
+	}()
+
+	select {
+	case o := <-done:
+		if o.err != nil {
+			t.Fatalf("Watch: %v", o.err)
+		}
+		if o.res.Reason != "PR merged" {
+			t.Errorf("reason = %q, want %q", o.res.Reason, "PR merged")
+		}
+		if n := warned.Load(); n != 2 {
+			t.Errorf("warnings = %d, want 2", n)
+		}
+	case <-time.After(3 * time.Second):
+		t.Fatal("Watch hung: a transient 5xx must not stop the watch")
+	}
+}
+
+// pollFailures scripts a fetcher whose polls fail with a 502 at the given call
+// indexes (0 is the baseline); the final call returns merged.
+func pollFailures(calls int, failAt map[int]bool) *fakeFetcher {
+	open, merged := base(), base()
+	merged.State = "closed"
+	merged.Merged = true
+	f := &fakeFetcher{states: make([]PRState, calls), errs: make([]error, calls)}
+	for i := range calls {
+		f.states[i] = open
+		if failAt[i] {
+			f.errs[i] = &APIError{Method: "GET", Path: "/p", StatusCode: 502, Body: "bad gateway"}
+		}
+	}
+	f.states[calls-1] = merged
+	return f
+}
+
+// A permanently wedged endpoint (5xx forever) must eventually give up instead of
+// warning on every tick for the life of the process.
+func TestWatchAbortsAfterConsecutiveFailures(t *testing.T) {
+	failAt := map[int]bool{}
+	for i := 1; i <= MaxPollFailures; i++ {
+		failAt[i] = true
+	}
+	f := pollFailures(MaxPollFailures+1, failAt)
+
+	warned := 0
+	ticks := make(chan time.Time, MaxPollFailures)
+	for range MaxPollFailures {
+		ticks <- time.Now()
+	}
+	close(ticks)
+	_, err := Watch(f, []PRRef{base().Ref}, "unkin-agent", ticks, nil, func(PRRef, error) { warned++ })
+	if err == nil {
+		t.Fatal("Watch should give up once the failures stop being transient")
+	}
+	if !strings.Contains(err.Error(), "consecutive failures") {
+		t.Errorf("Watch error = %v, want it to report the failure cap", err)
+	}
+	if warned != MaxPollFailures {
+		t.Errorf("warnings = %d, want %d", warned, MaxPollFailures)
+	}
+	if f.calls != MaxPollFailures+1 {
+		t.Errorf("fetch calls = %d, want %d", f.calls, MaxPollFailures+1)
+	}
+}
+
+// The cap counts consecutive failures only: a single successful poll clears it,
+// so an intermittent endpoint is watched indefinitely and the merge is caught.
+func TestWatchFailureCountResetsOnSuccess(t *testing.T) {
+	const runs = MaxPollFailures - 1
+	failAt := map[int]bool{}
+	for i := 1; i <= runs; i++ { // first run of failures
+		failAt[i] = true
+	}
+	for i := runs + 2; i <= 2*runs+1; i++ { // second run, after one good poll
+		failAt[i] = true
+	}
+	f := pollFailures(2*runs+3, failAt)
+
+	ticks := make(chan time.Time, 2*runs+2)
+	for range 2*runs + 2 {
+		ticks <- time.Now()
+	}
+	close(ticks)
+	res, err := Watch(f, []PRRef{base().Ref}, "unkin-agent", ticks, nil, func(PRRef, error) {})
+	if err != nil {
+		t.Fatalf("Watch: %v (a successful poll must reset the failure count)", err)
+	}
+	if res.Reason != "PR merged" {
+		t.Errorf("reason = %q, want %q", res.Reason, "PR merged")
+	}
+}
+
+// Anonymous watching of a public repo must poll on without a credential in
+// sight: no token, no mint, no exit until something actually changes.
+func TestWatchAnonymousKeepsPolling(t *testing.T) {
+	open := base()
+	f := &fakeFetcher{states: []PRState{open}}
+
+	ticks := make(chan time.Time, 2)
+	ticks <- time.Now()
+	ticks <- time.Now()
+	close(ticks)
+	res, err := Watch(f, []PRRef{open.Ref}, "unkin-agent", ticks, nil,
+		func(_ PRRef, e error) { t.Errorf("unexpected poll error: %v", e) })
+	if err != nil {
+		t.Fatalf("Watch: %v", err)
+	}
+	if res.Reason != "" {
+		t.Errorf("reason = %q, want no change reported", res.Reason)
+	}
+	if f.calls != 3 {
+		t.Errorf("fetch calls = %d, want 3 (baseline + two polls)", f.calls)
+	}
+}