Merge pull request 'watchpr: start the conflict window at the arm' (#21) from benvin/watchpr-arm-window-start into main

Reviewed-on: #21

AGENTS.md

@@ -8,10 +8,11 @@ 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 and edit pull requests and issues, and post comments on
-  either, as `unkin-agent` (fixes the "tea posts as Ben" attribution problem).
-  Subcommands: `pr create`, `pr comment`, `pr edit`, `issue create`,
-  `issue comment`, `issue edit`, `whoami`.
+- **`agentpr`** — create and edit pull requests and issues, close and reopen
+  issues, and post comments on either, as `unkin-agent` (fixes the "tea posts
+  as Ben" attribution problem). Subcommands: `pr create`, `pr comment`,
+  `pr edit`, `issue create`, `issue comment`, `issue edit`, `issue close`,
+  `issue reopen`, `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
@@ -29,7 +30,7 @@ parsing, watch-state comparison, git worktree helpers).
 ## Structure
 
 ```
-cmd/agentpr/main.go      # agentpr CLI (pr + issue create/comment/edit, whoami)
+cmd/agentpr/main.go      # agentpr CLI (pr + issue create/comment/edit, issue close/reopen, whoami)
 cmd/watchpr/main.go      # watchpr CLI (poll + meaningful-change exit)
 cmd/agentws/main.go      # agentws CLI (new / list / rm / clean / token / credential)
 cmd/agentws/prune.go     # agentws prune (classify worktrees, remove the safe ones)
@@ -37,7 +38,7 @@ 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/issue create/edit/get, comments, status, whoami)
+  gitea.go               #   Gitea REST client (PR/issue create/edit/get, issue state, 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)
@@ -131,7 +132,8 @@ make test           # go test -v -race ./...
 `internal/agent` covers PR-ref parsing, the `MeaningfulChange` table (benign vs
 alerting transitions), request-body construction, and the Vault+Gitea client
 against `httptest` servers (fake AppRole login + gitea creds + PR/issue create
-+ edit / comment / whoami / status). No live Vault/Gitea access is required for tests.
++ edit / close / reopen / comment / whoami / status). No live Vault/Gitea access
+is required for tests.
 
 ## agentvault seed-outpost
 
@@ -218,8 +220,45 @@ wrapped per stage (login / read denied / write denied) via `ErrVaultDenied`.
   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.
+- `watchpr` alerts on changes, not on conditions: the baseline snapshot is a real
+  read, so a PR that is already conflicted or already CI-failing when watching
+  starts is polled on rather than reported. That suppression is printed with the
+  baseline, since a silent watcher and a watcher with nothing to say look alike.
+  The baseline goes to stderr in both modes — a JSON record under `--json` —
+  because automation is what most needs it and stdout carries the result.
+- Mergeability is the one `watchpr` rule needing a run of observations
+  (`prWatch`). Gitea 1.26 sends `mergeable` as a plain bool — always present,
+  never null — and sends `false` both for a real conflict and while it recomputes
+  the merge base after a push, which is exactly when an agent starts a watch. So
+  `false` counts only once this watch has seen a merge computation start (the PR
+  was mergeable, or `head.sha`/`base.sha` moved), and then only once the run of
+  `false`s has spanned `conflictWindow` (2m); a mergeable poll, an unknown one or
+  a failed poll all break the run. The debounce is wall-clock, measured from the
+  tick that fired each poll, because what it outlasts is the recompute and
+  `--interval` spans seconds to hours. The window starts at the later of the
+  run's first observation and the arm, which is what keeps arming and the run
+  independent: movement arms without resetting the run — `base.sha` is the base
+  branch tip as of the response, so it moves for every open PR whenever main does
+  and a reset there could never complete on a busy repo — while an arming poll
+  still cannot confirm a run it played no part in. Re-arming an already-armed
+  watch is a no-op, so a base moving under every poll advances the window once
+  and never again. Both times are pointers because the zero `time.Time` is a
+  legal clock value and cannot also mean "no run". `Mergeability` stays tri-state
+  for what the bool cannot carry: an absent/null flag from some other Gitea, and
+  a state no successful poll ever filled in. Head and base SHAs are read only to
+  arm — a push never alerts. The residue: a conflict landed by the push just
+  before the watch began is unreported while the commits stay put, since Gitea's
+  payload has no field separating it from a check in flight (`merge_base` is the
+  true merge base and does not move on recheck). The commits rarely stay put —
+  a merge to the base branch moves `base.sha` under every open PR — but that
+  move only arms; the two minutes it then has to outlast are the recompute the
+  move started, not the falses before it.
 - CI "combined status" comes from `/commits/{sha}/status`; an empty head SHA
   yields an empty state without an API call.
 - Gitea backs every PR with an issue of the same number and serves comments from
   `/issues/{n}/comments`, so `agentpr pr comment` and `agentpr issue comment`
   are one implementation under two flag names (`--pr` / `--issue`).
+- `issue close`/`issue reopen` read the issue before the PATCH: Gitea answers a
+  no-op state change with 200, so without the read an already-closed issue would
+  report success. There is no `pr close`: closing a pull request is a human's
+  call, not an agent's.

README.md

@@ -6,8 +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 and edit pull requests and issues, and post comments on
-  either, as the agent user.
+- **`agentpr`** — create and edit pull requests and issues, close and reopen
+  issues, and post comments on either, 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
@@ -68,6 +68,12 @@ agentpr issue comment --repo unkin/argocd-apps --issue 43 --body "Fixed in #44."
 agentpr issue edit --repo unkin/argocd-apps --issue 43 --body "The pipeline fails with ..."
 # prints: #<number> <html_url>
 
+# Close or reopen an issue; an issue already in that state is an error, not a
+# silent success
+agentpr issue close --repo unkin/argocd-apps --issue 43
+agentpr issue reopen --repo unkin/argocd-apps --issue 43
+# prints: #<number> <state> <html_url>
+
 agentpr --version
 agentpr --help
 ```
@@ -79,7 +85,38 @@ Non-zero exit on any API error.
 Poll PRs and exit (reporting what changed) when a tracked PR **merges/closes**,
 gets a **new comment from someone other than the agent**, its **CI fails**
 (failure/error), or it **loses mergeability** (a conflict appears). Benign
-transitions — CI `pending`→`success`, the agent's own comments — are ignored.
+transitions — CI `pending`→`success`, the agent's own comments and pushes — are
+ignored, and so is any condition the PR was already in at the baseline: a PR
+that is already conflicted or already failing when watching starts keeps being
+watched, and the baseline line — written to stderr, as a JSON record under
+`--json` — says which conditions it is staying silent about.
+
+Gitea reports `mergeable: false` both for a real conflict and while it
+recomputes the merge base after a push, so a conflict is only reported once the
+watch has seen a merge computation start (the PR was mergeable, or its head or
+base commit moved) **and** the non-mergeable polls have **then** run unbroken for
+two minutes. The window runs from whichever came later, so the poll that starts
+the merge computation never confirms a run of falses that predates it. The
+debounce is a duration, not a poll count, because what it has to outlast is
+Gitea's recompute and `--interval` ranges from seconds to hours.
+
+So a conflict introduced by the push immediately before the watch started is not
+reported while the commits stay put: nothing in Gitea's payload separates it from
+a merge check still in flight, and the baseline line is the only notice of it.
+That is a narrower gap than it looks, because `base.sha` is the base branch's
+tip — it moves for every open PR whenever anything merges to the base branch, so
+the commits rarely stay put for long. The first such move arms the rule, and is
+itself silent: the window then runs from that arming poll rather than from the
+falses that predate it, so what is eventually reported is a conflict that
+outlasted the merge computation the move started.
+
+Arming is a one-way latch. Moves after that one — and every move seen by a watch
+that was already armed at the baseline — neither re-arm nor restart the window,
+because a restart on every `base.sha` move could never complete on a busy base
+branch. That is the cost side of the same trade: on a busy base branch a
+conflict can be confirmed while the newest merge recompute is less than two
+minutes old. The window guarantees that the run of falses outlasted *a* merge
+computation this watch saw start, not that it outlasted the most recent one.
 
 ```bash
 # Watch until something meaningful happens (default interval 60s)

cmd/agentpr/main.go

@@ -9,6 +9,8 @@
 //	agentpr issue create --repo owner/repo --title T --body B
 //	agentpr issue comment --repo owner/repo --issue 12 --body "..."
 //	agentpr issue edit --repo owner/repo --issue 12 --title T --body B
+//	agentpr issue close --repo owner/repo --issue 12
+//	agentpr issue reopen --repo owner/repo --issue 12
 //	agentpr whoami
 package main
 
@@ -69,9 +71,15 @@ func newPRCmd() *cobra.Command {
 func newIssueCmd() *cobra.Command {
 	cmd := &cobra.Command{
 		Use:   "issue",
-		Short: "File and edit issues, and post issue comments",
+		Short: "File, edit, close and reopen issues, and post issue comments",
 	}
-	cmd.AddCommand(newIssueCreateCmd(), newCommentCmd("issue", "issue", "Post a comment on an issue"), newIssueEditCmd())
+	cmd.AddCommand(
+		newIssueCreateCmd(),
+		newCommentCmd("issue", "issue", "Post a comment on an issue"),
+		newIssueEditCmd(),
+		newIssueStateCmd("close", "Close an issue", agent.IssueStateClosed),
+		newIssueStateCmd("reopen", "Reopen a closed issue", agent.IssueStateOpen),
+	)
 	return cmd
 }
 
@@ -301,6 +309,44 @@ func newIssueEditCmd() *cobra.Command {
 	return cmd
 }
 
+// newIssueStateCmd builds `issue close` and `issue reopen`, which differ only
+// in the state they ask for. An issue already in that state is an error, not a
+// silent success: Gitea answers the PATCH with 200 either way.
+func newIssueStateCmd(use, short, state string) *cobra.Command {
+	var repo string
+	var issue int
+	cmd := &cobra.Command{
+		Use:          use,
+		Short:        short,
+		SilenceUsage: true,
+		RunE: func(cmd *cobra.Command, args []string) error {
+			owner, name, err := agent.ParseRepo(repo)
+			if err != nil {
+				return err
+			}
+			if issue <= 0 {
+				return fmt.Errorf("--issue must be a positive issue number")
+			}
+			c, err := client()
+			if err != nil {
+				return err
+			}
+			updated, err := c.SetIssueState(owner+"/"+name, issue, state)
+			if err != nil {
+				return err
+			}
+			fmt.Printf("#%d %s %s\n", updated.Number, updated.State, updated.HTMLURL)
+			return nil
+		},
+	}
+	f := cmd.Flags()
+	f.StringVar(&repo, "repo", "", "Repository as owner/repo (required)")
+	f.IntVar(&issue, "issue", 0, "Issue number (required)")
+	_ = cmd.MarkFlagRequired("repo")
+	_ = cmd.MarkFlagRequired("issue")
+	return cmd
+}
+
 func newWhoamiCmd() *cobra.Command {
 	return &cobra.Command{
 		Use:          "whoami",

cmd/agentpr/main_test.go

@@ -149,3 +149,56 @@ func TestCommentCommandsStayInStep(t *testing.T) {
 		t.Error("issue comment must take --issue and only --issue")
 	}
 }
+
+// Both state commands need a repo and an issue number; cobra and the RunE
+// guard must reject their absence before anything reaches for a token.
+func TestIssueStateCommandsRequireFlags(t *testing.T) {
+	for _, verb := range []string{"close", "reopen"} {
+		t.Run(verb+" without --issue", func(t *testing.T) {
+			err := execute("issue", verb, "--repo", "unkin/repo")
+			if err == nil {
+				t.Fatal("Execute() = nil, want an error when --issue is missing")
+			}
+			if !strings.Contains(err.Error(), "issue") {
+				t.Errorf("error = %q, want it to name the missing --issue flag", err)
+			}
+		})
+		t.Run(verb+" without --repo", func(t *testing.T) {
+			err := execute("issue", verb, "--issue", "12")
+			if err == nil {
+				t.Fatal("Execute() = nil, want an error when --repo is missing")
+			}
+			if !strings.Contains(err.Error(), "repo") {
+				t.Errorf("error = %q, want it to name the missing --repo flag", err)
+			}
+		})
+		t.Run(verb+" with a zero --issue", func(t *testing.T) {
+			err := execute("issue", verb, "--repo", "unkin/repo", "--issue", "0")
+			if err == nil {
+				t.Fatal("Execute() = nil, want an error for a non-positive issue number")
+			}
+			if !strings.Contains(err.Error(), "--issue must be a positive") {
+				t.Errorf("error = %q, want it to reject the issue number", err)
+			}
+		})
+		t.Run(verb+" with a malformed --repo", func(t *testing.T) {
+			if err := execute("issue", verb, "--repo", "not-a-repo", "--issue", "12"); err == nil {
+				t.Fatal("Execute() = nil, want error for a malformed --repo")
+			}
+		})
+	}
+}
+
+// Closing and reopening are issue-only: a PR is closed by a human, so the pr
+// group must not grow these verbs by accident.
+func TestPRHasNoStateCommands(t *testing.T) {
+	for _, verb := range []string{"close", "reopen"} {
+		if cmd, _, err := newRootCmd().Find([]string{"pr", verb}); err == nil && cmd.Name() == verb {
+			t.Errorf("pr %s exists; closing a PR is not agentpr's to do", verb)
+		}
+		cmd, _, err := newRootCmd().Find([]string{"issue", verb})
+		if err != nil || cmd.Name() != verb {
+			t.Fatalf("issue %s not found: %v", verb, err)
+		}
+	}
+}

cmd/watchpr/main.go

@@ -1,8 +1,19 @@
 // Command watchpr polls one or more Gitea pull requests and exits when a
 // tracked PR changes in a way worth alerting on: it merges or closes, gets a
 // new comment from someone other than the agent, its CI fails, or it loses
-// mergeability. Benign transitions (CI pending→success, the agent's own
-// comments) are ignored.
+// mergeability after the baseline. Benign transitions (CI pending→success, the
+// agent's own pushes and comments) are ignored, and so is any condition that was
+// already true at the baseline -- which watchpr prints on stderr (as a JSON
+// record under --json), so a run started against a conflicted or CI-red PR says
+// so. A conflict introduced by the push immediately before the watch started is
+// not reported while the commits stay put, because nothing in Gitea's payload
+// separates it from a merge check still in flight; the baseline line is the only
+// notice of it. The first later move of the head or base arms the rule without
+// alerting, and the window then runs from that arm rather than from the falses
+// that predate it. Moves after that neither re-arm nor restart the window -- a
+// restart on every base move could never complete on a busy base branch -- so a
+// conflict can be confirmed while the newest merge recompute is less than two
+// minutes old.
 //
 //	watchpr owner/repo#12 owner/repo:15
 //	watchpr --once --json owner/repo#12
@@ -13,7 +24,9 @@ package main
 import (
 	"encoding/json"
 	"fmt"
+	"io"
 	"os"
+	"strings"
 	"time"
 
 	"git.unkin.net/unkin/agent-tools/internal/agent"
@@ -43,7 +56,8 @@ func newRootCmd() *cobra.Command {
 		Short: "Poll Gitea PRs and exit when one changes meaningfully.",
 		Long: "watchpr polls each PR every --interval and exits (reporting what changed)\n" +
 			"when a PR merges/closes, gets a new non-agent comment, its CI fails, or it\n" +
-			"loses mergeability. Accepts refs as owner/repo#N or owner/repo:N.",
+			"loses mergeability after the baseline. Conditions already true at the\n" +
+			"baseline are printed, not alerted on. Refs take owner/repo#N or owner/repo:N.",
 		Version:      version,
 		Args:         cobra.ArbitraryArgs,
 		SilenceUsage: true,
@@ -63,7 +77,7 @@ func newRootCmd() *cobra.Command {
 				}
 				refs = append(refs, ref)
 			}
-			c := clientFor()
+			c := clientFor(jsonMode)
 			if once {
 				return runOnce(c, refs, jsonMode)
 			}
@@ -89,15 +103,35 @@ func newRootCmd() *cobra.Command {
 // 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 {
+func clientFor(jsonMode bool) *agent.GiteaClient {
 	token, err := agent.GiteaToken()
 	if err != nil {
-		fmt.Fprintf(os.Stderr, "warning: no Gitea token (%v); polling anonymously\n", err)
+		warn(os.Stderr, jsonMode, "no Gitea token (%v); polling anonymously", err)
 		token = ""
 	}
 	return agent.NewGiteaClient(token)
 }
 
+// warnRecord is the --json form of a warning. Under --json every notice watchpr
+// itself writes to stderr -- warnings and the baseline -- is an NDJSON record,
+// so a caller parsing them line by line never has to guess which shape a line
+// is. A terminal failure is the one exception: SilenceErrors stays off, so
+// cobra prints it as a plain "Error: ..." line and the exit status is non-zero.
+type warnRecord struct {
+	Warning string `json:"warning"`
+}
+
+// warn writes a non-fatal notice to stderr in whichever form the caller asked
+// for.
+func warn(w io.Writer, jsonMode bool, format string, args ...any) {
+	msg := fmt.Sprintf(format, args...)
+	if jsonMode {
+		_ = json.NewEncoder(w).Encode(warnRecord{Warning: msg})
+		return
+	}
+	_, _ = fmt.Fprintf(w, "warning: %s\n", msg)
+}
+
 // runOnce fetches and prints the current state of each PR, then exits 0.
 func runOnce(c *agent.GiteaClient, refs []agent.PRRef, jsonMode bool) error {
 	login := agent.AgentLogin()
@@ -126,13 +160,11 @@ func runWatch(c *agent.GiteaClient, refs []agent.PRRef, interval time.Duration,
 	ticker := time.NewTicker(interval)
 	defer ticker.Stop()
 
-	onBaseline := func() {
-		if !jsonMode {
-			fmt.Fprintf(os.Stderr, "watching %d PR(s) every %s; baseline established\n", len(refs), interval)
-		}
+	onBaseline := func(states []agent.PRState) {
+		emitBaselines(os.Stderr, states, interval, jsonMode)
 	}
 	onError := func(ref agent.PRRef, err error) {
-		fmt.Fprintf(os.Stderr, "warning: polling %s: %v\n", ref.String(), err)
+		warn(os.Stderr, jsonMode, "polling %s: %v", ref.String(), err)
 	}
 
 	res, err := agent.Watch(c, refs, login, ticker.C, onBaseline, onError)
@@ -143,6 +175,32 @@ func runWatch(c *agent.GiteaClient, refs []agent.PRRef, interval time.Duration,
 	return nil
 }
 
+// baselineRecord is the --json form of baselineLine. Both go to stderr, leaving
+// the stdout contract a single result record: a caller automating watchpr is
+// precisely the one who needs to be told the watch started against a PR that is
+// already conflicted, since no alert will ever follow for it.
+type baselineRecord struct {
+	Baseline   bool          `json:"baseline"`
+	Suppressed string        `json:"suppressed,omitempty"`
+	State      agent.PRState `json:"state"`
+}
+
+// emitBaselines writes the state each watch started from, in whichever form the
+// caller asked for.
+func emitBaselines(w io.Writer, states []agent.PRState, interval time.Duration, jsonMode bool) {
+	if jsonMode {
+		enc := json.NewEncoder(w)
+		for _, st := range states {
+			_ = enc.Encode(baselineRecord{Baseline: true, Suppressed: suppressedAtBaseline(st), State: st})
+		}
+		return
+	}
+	_, _ = fmt.Fprintf(w, "watching %d PR(s) every %s; baseline established\n", len(states), interval)
+	for _, st := range states {
+		_, _ = fmt.Fprintln(w, baselineLine(st))
+	}
+}
+
 // describeFailure names the cause of a terminal failure so a watcher that stops
 // says why. An anonymous rejection, a permission boundary and a token that
 // outlived its Vault lease are three different problems and only the last is
@@ -176,8 +234,39 @@ func report(key, reason string, st agent.PRState, jsonMode bool) {
 }
 
 func printState(st agent.PRState) {
-	fmt.Printf("%s state=%s merged=%t mergeable=%t ci=%s head=%s comments(non-agent)=%d\n",
-		st.Ref.String(), st.State, st.Merged, st.Mergeable, ciOrNone(st.CIStatus), shortSHA(st.HeadSHA), st.NonAgentComments)
+	fmt.Println(stateLine(st))
+}
+
+func stateLine(st agent.PRState) string {
+	return fmt.Sprintf("%s state=%s merged=%t mergeable=%s ci=%s head=%s base=%s comments(non-agent)=%d",
+		st.Ref.String(), st.State, st.Merged, st.Mergeable, ciOrNone(st.CIStatus),
+		shortSHA(st.HeadSHA), shortSHA(st.BaseSHA), st.NonAgentComments)
+}
+
+// baselineLine describes the state a watch started from, naming the conditions
+// it will deliberately stay silent about: watchpr alerts on changes, so a PR
+// that is already conflicted or already CI-red produces no alert for either, and
+// that suppression has to be visible to whoever started the watch.
+func baselineLine(st agent.PRState) string {
+	line := "  baseline " + stateLine(st)
+	if s := suppressedAtBaseline(st); s != "" {
+		return line + " -- " + s + "; not alerting on a pre-existing condition"
+	}
+	return line
+}
+
+func suppressedAtBaseline(st agent.PRState) string {
+	var conds []string
+	switch st.Mergeable {
+	case agent.MergeNo:
+		conds = append(conds, "already non-mergeable (a real conflict, or Gitea still recomputing)")
+	case agent.MergeUnknown:
+		conds = append(conds, "mergeability unknown (the conflict rule is disarmed until Gitea reports mergeable or the commits move)")
+	}
+	if st.CIStatus == "failure" || st.CIStatus == "error" {
+		conds = append(conds, "CI already "+st.CIStatus)
+	}
+	return strings.Join(conds, ", ")
 }
 
 func ciOrNone(s string) string {

cmd/watchpr/main_test.go

@@ -1,6 +1,8 @@
 package main
 
 import (
+	"bytes"
+	"encoding/json"
 	"errors"
 	"fmt"
 	"io"
@@ -8,6 +10,7 @@ import (
 	"net/http/httptest"
 	"strings"
 	"testing"
+	"time"
 
 	"git.unkin.net/unkin/agent-tools/internal/agent"
 )
@@ -182,3 +185,129 @@ func TestDescribeFailureNamesTheCause(t *testing.T) {
 		})
 	}
 }
+
+// A watch started against an already conflicted or CI-red PR sends no alert for
+// either, so the baseline line has to say that is what it is doing: before this,
+// such a run printed nothing in either direction.
+func TestBaselineLineNamesSuppressedConditions(t *testing.T) {
+	st := agent.PRState{
+		Ref:       agent.PRRef{Owner: "unkin", Repo: "repo", Number: 7},
+		State:     "open",
+		Mergeable: agent.MergeNo,
+		CIStatus:  "failure",
+		HeadSHA:   "cafebabecafebabe",
+		BaseSHA:   "deadbeefdeadbeef",
+	}
+	got := baselineLine(st)
+	for _, want := range []string{
+		"unkin/repo#7", "mergeable=false", "ci=failure", "head=cafebabe", "base=deadbeef",
+		"already non-mergeable", "CI already failure", "not alerting",
+	} {
+		if !strings.Contains(got, want) {
+			t.Errorf("baselineLine = %q, want it to mention %q", got, want)
+		}
+	}
+
+	clean := st
+	clean.Mergeable = agent.MergeYes
+	clean.CIStatus = "success"
+	if got := baselineLine(clean); strings.Contains(got, "not alerting") {
+		t.Errorf("baselineLine = %q, want no suppression note for a clean baseline", got)
+	}
+}
+
+// --json is the mode automation uses, and automation is exactly who needs to be
+// told the watch started against an already-conflicted PR -- the one case that
+// will never produce an alert. It must therefore be emitted in JSON mode too,
+// on stderr, where it cannot corrupt the result record on stdout.
+func TestBaselineIsEmittedInJSONMode(t *testing.T) {
+	st := agent.PRState{
+		Ref:       agent.PRRef{Owner: "unkin", Repo: "repo", Number: 7},
+		State:     "open",
+		Mergeable: agent.MergeNo,
+		CIStatus:  "failure",
+		HeadSHA:   "cafebabecafebabe",
+	}
+	var buf bytes.Buffer
+	emitBaselines(&buf, []agent.PRState{st}, 30*time.Second, true)
+
+	var got baselineRecord
+	if err := json.Unmarshal(buf.Bytes(), &got); err != nil {
+		t.Fatalf("baseline is not a JSON record (%v); got %q", err, buf.String())
+	}
+	if !got.Baseline {
+		t.Error("record does not mark itself as the baseline")
+	}
+	if got.State.Ref != st.Ref || got.State.Mergeable != agent.MergeNo {
+		t.Errorf("record state = %+v, want the conflicted snapshot the watch started from", got.State)
+	}
+	for _, want := range []string{"already non-mergeable", "CI already failure"} {
+		if !strings.Contains(got.Suppressed, want) {
+			t.Errorf("suppressed = %q, want it to mention %q", got.Suppressed, want)
+		}
+	}
+
+	buf.Reset()
+	clean := st
+	clean.Mergeable = agent.MergeYes
+	clean.CIStatus = "success"
+	emitBaselines(&buf, []agent.PRState{clean}, 30*time.Second, true)
+	if strings.Contains(buf.String(), "suppressed") {
+		t.Errorf("clean baseline = %q, want no suppression field", buf.String())
+	}
+}
+
+// Under --json stderr is the baseline and warning channel, so every notice
+// watchpr writes there has to be one shape: a caller parsing it line by line
+// must never meet a bare `warning:` line between two NDJSON records. Cobra's
+// terminal `Error: ...` line is not covered here -- SilenceErrors stays off, so
+// it is plain text on stderr alongside a non-zero exit.
+func TestJSONModeWarningsAndBaselineAreRecords(t *testing.T) {
+	st := agent.PRState{
+		Ref:       agent.PRRef{Owner: "unkin", Repo: "repo", Number: 7},
+		State:     "open",
+		Mergeable: agent.MergeNo,
+	}
+	var buf bytes.Buffer
+	warn(&buf, true, "no Gitea token (%v); polling anonymously", errors.New("vault refused"))
+	emitBaselines(&buf, []agent.PRState{st}, 30*time.Second, true)
+	warn(&buf, true, "polling %s: %v", st.Ref.String(), errors.New("HTTP 502"))
+
+	lines := strings.Split(strings.TrimSpace(buf.String()), "\n")
+	if len(lines) != 3 {
+		t.Fatalf("stderr = %q, want 3 records", buf.String())
+	}
+	for _, line := range lines {
+		var rec map[string]any
+		if err := json.Unmarshal([]byte(line), &rec); err != nil {
+			t.Errorf("stderr line %q is not a JSON record: %v", line, err)
+		}
+	}
+	var first warnRecord
+	if err := json.Unmarshal([]byte(lines[0]), &first); err != nil || !strings.Contains(first.Warning, "vault refused") {
+		t.Errorf("first record = %q, want the token warning", lines[0])
+	}
+
+	buf.Reset()
+	warn(&buf, false, "polling %s: %v", st.Ref.String(), errors.New("HTTP 502"))
+	if got := buf.String(); !strings.HasPrefix(got, "warning: ") {
+		t.Errorf("plain-mode warning = %q, want the warning: prefix", got)
+	}
+}
+
+// An unknown mergeability disarms the conflict rule exactly as a non-mergeable
+// baseline does, so the baseline has to name it too.
+func TestBaselineNamesAnUnknownMergeability(t *testing.T) {
+	st := agent.PRState{
+		Ref:       agent.PRRef{Owner: "unkin", Repo: "repo", Number: 7},
+		State:     "open",
+		Mergeable: agent.MergeUnknown,
+		CIStatus:  "success",
+	}
+	if got := suppressedAtBaseline(st); !strings.Contains(got, "mergeability unknown") {
+		t.Errorf("suppressed = %q, want it to name the unknown mergeability", got)
+	}
+	if got := baselineLine(st); !strings.Contains(got, "not alerting") {
+		t.Errorf("baselineLine = %q, want the suppression note", got)
+	}
+}

internal/agent/client_test.go

@@ -946,3 +946,144 @@ func TestEditIssueSendsOnlySuppliedFields(t *testing.T) {
 		})
 	}
 }
+
+// Closing and reopening differ only in the state sent; both must read the
+// issue first and then PATCH the issue endpoint with that state alone.
+func TestSetIssueStateRequest(t *testing.T) {
+	tests := []struct {
+		name    string
+		current string
+		state   string
+	}{
+		{"close an open issue", IssueStateOpen, IssueStateClosed},
+		{"reopen a closed issue", IssueStateClosed, IssueStateOpen},
+	}
+	for _, tt := range tests {
+		t.Run(tt.name, func(t *testing.T) {
+			var methods []string
+			var gotBody map[string]any
+			var gotPath string
+			mux := http.NewServeMux()
+			mux.HandleFunc("/api/v1/repos/unkin/repo/issues/12", func(w http.ResponseWriter, r *http.Request) {
+				methods = append(methods, r.Method)
+				gotPath = r.URL.Path
+				if r.Method == http.MethodGet {
+					_, _ = io.WriteString(w, `{"number":12,"state":"`+tt.current+`","title":"T","html_url":"https://git.unkin.net/unkin/repo/issues/12"}`)
+					return
+				}
+				_ = json.NewDecoder(r.Body).Decode(&gotBody)
+				_, _ = io.WriteString(w, `{"number":12,"state":"`+tt.state+`","title":"T","html_url":"https://git.unkin.net/unkin/repo/issues/12"}`)
+			})
+			srv := httptest.NewServer(mux)
+			defer srv.Close()
+
+			c := &GiteaClient{BaseURL: srv.URL, Token: "gitea-abc", HTTP: srv.Client()}
+			issue, err := c.SetIssueState("unkin/repo", 12, tt.state)
+			if err != nil {
+				t.Fatalf("SetIssueState: %v", err)
+			}
+			if len(methods) != 2 || methods[0] != http.MethodGet || methods[1] != http.MethodPatch {
+				t.Errorf("requests = %v, want a GET then a PATCH", methods)
+			}
+			if gotPath != "/api/v1/repos/unkin/repo/issues/12" {
+				t.Errorf("path = %q", gotPath)
+			}
+			if len(gotBody) != 1 || gotBody["state"] != tt.state {
+				t.Errorf("payload = %v, want only {\"state\":%q}", gotBody, tt.state)
+			}
+			if issue.State != tt.state || issue.Number != 12 {
+				t.Errorf("parsed issue = %+v", issue)
+			}
+		})
+	}
+}
+
+// Gitea answers a no-op state change with 200, so an issue already in the
+// requested state must fail rather than report a change that never happened —
+// and no PATCH may be sent.
+func TestSetIssueStateAlreadyInState(t *testing.T) {
+	patches := 0
+	mux := http.NewServeMux()
+	mux.HandleFunc("/api/v1/repos/unkin/repo/issues/12", func(w http.ResponseWriter, r *http.Request) {
+		if r.Method == http.MethodPatch {
+			patches++
+		}
+		_, _ = io.WriteString(w, `{"number":12,"state":"closed","title":"T","html_url":"https://git.unkin.net/unkin/repo/issues/12"}`)
+	})
+	srv := httptest.NewServer(mux)
+	defer srv.Close()
+
+	c := &GiteaClient{BaseURL: srv.URL, Token: "t", HTTP: srv.Client()}
+	_, err := c.SetIssueState("unkin/repo", 12, IssueStateClosed)
+	if err == nil {
+		t.Fatal("expected an error closing an already-closed issue")
+	}
+	if !errors.Is(err, ErrIssueStateUnchanged) {
+		t.Errorf("errors.Is(%v, ErrIssueStateUnchanged) = false", err)
+	}
+	if !strings.Contains(err.Error(), "unkin/repo#12") {
+		t.Errorf("error %q should name the issue", err)
+	}
+	if patches != 0 {
+		t.Errorf("PATCH requests = %d, want 0", patches)
+	}
+}
+
+// A non-2xx on either leg must surface the API's own message, not a bare
+// status, and must not be mistaken for a successful change.
+func TestSetIssueStateAPIError(t *testing.T) {
+	tests := []struct {
+		name   string
+		failOn string
+		status int
+	}{
+		{"read fails", http.MethodGet, http.StatusNotFound},
+		{"write fails", http.MethodPatch, http.StatusForbidden},
+	}
+	for _, tt := range tests {
+		t.Run(tt.name, func(t *testing.T) {
+			mux := http.NewServeMux()
+			mux.HandleFunc("/api/v1/repos/unkin/repo/issues/12", func(w http.ResponseWriter, r *http.Request) {
+				if r.Method == tt.failOn {
+					w.WriteHeader(tt.status)
+					_, _ = io.WriteString(w, `{"message":"no dice","url":"https://git.unkin.net/api/swagger","errors":null}`)
+					return
+				}
+				_, _ = io.WriteString(w, `{"number":12,"state":"open","title":"T","html_url":"https://git.unkin.net/unkin/repo/issues/12"}`)
+			})
+			srv := httptest.NewServer(mux)
+			defer srv.Close()
+
+			c := &GiteaClient{BaseURL: srv.URL, Token: "t", HTTP: srv.Client()}
+			_, err := c.SetIssueState("unkin/repo", 12, IssueStateClosed)
+			if err == nil {
+				t.Fatalf("expected an error when %s returns %d", tt.failOn, tt.status)
+			}
+			if !strings.Contains(err.Error(), `"message":"no dice"`) {
+				t.Errorf("error %q should carry the API message", err)
+			}
+		})
+	}
+}
+
+// Only Gitea's two states are accepted, and a bad one is rejected before any
+// request goes out.
+func TestSetIssueStateRejectsUnknownState(t *testing.T) {
+	requests := 0
+	srv := httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
+		requests++
+	}))
+	defer srv.Close()
+
+	c := &GiteaClient{BaseURL: srv.URL, Token: "t", HTTP: srv.Client()}
+	_, err := c.SetIssueState("unkin/repo", 12, "merged")
+	if err == nil {
+		t.Fatal("expected an error for an unknown state")
+	}
+	if !strings.Contains(err.Error(), `invalid issue state "merged"`) {
+		t.Errorf("error = %q, want it to name the invalid state", err)
+	}
+	if requests != 0 {
+		t.Errorf("requests = %d, want 0", requests)
+	}
+}

internal/agent/gitea.go

@@ -189,17 +189,22 @@ func (c *GiteaClient) Whoami() (User, error) {
 
 // PullRequest is the subset of Gitea's PR object we track.
 type PullRequest struct {
-	Number    int    `json:"number"`
-	State     string `json:"state"`
-	Title     string `json:"title"`
-	Merged    bool   `json:"merged"`
-	Mergeable bool   `json:"mergeable"`
-	HTMLURL   string `json:"html_url"`
+	Number    int          `json:"number"`
+	State     string       `json:"state"`
+	Title     string       `json:"title"`
+	Merged    bool         `json:"merged"`
+	Mergeable Mergeability `json:"mergeable"`
+	HTMLURL   string       `json:"html_url"`
 	Head      struct {
 		Sha   string `json:"sha"`
 		Ref   string `json:"ref"`
 		Label string `json:"label"`
 	} `json:"head"`
+	// Base.Sha is the base branch tip as of this response, not the merge base
+	// recorded when the PR was opened.
+	Base struct {
+		Sha string `json:"sha"`
+	} `json:"base"`
 }
 
 // prPageSize is the per-page limit for the pulls listing; maxPRPages caps how
@@ -329,6 +334,46 @@ func (c *GiteaClient) EditIssue(repoPath string, number int, opts EditOptions) (
 	return issue, err
 }
 
+// Issue states Gitea accepts on a state change. Gitea has no third state: an
+// issue is open or closed.
+const (
+	IssueStateOpen   = "open"
+	IssueStateClosed = "closed"
+)
+
+// ErrIssueStateUnchanged reports a state change asked for the state the issue
+// is already in. Gitea answers such a PATCH with 200 and changes nothing, so
+// without this check closing an already-closed issue would look like it worked.
+var ErrIssueStateUnchanged = errors.New("issue is already in that state")
+
+// GetIssue fetches a single issue
+// (GET /api/v1/repos/{owner}/{repo}/issues/{index}).
+func (c *GiteaClient) GetIssue(repoPath string, number int) (Issue, error) {
+	var issue Issue
+	err := c.do(http.MethodGet, fmt.Sprintf("/api/v1/repos/%s/issues/%d", repoPath, number), nil, &issue)
+	return issue, err
+}
+
+// SetIssueState closes or reopens an issue. It reads the issue first so an
+// issue already in the requested state fails with ErrIssueStateUnchanged
+// instead of reporting a change that never happened.
+func (c *GiteaClient) SetIssueState(repoPath string, number int, state string) (Issue, error) {
+	if state != IssueStateOpen && state != IssueStateClosed {
+		return Issue{}, fmt.Errorf("invalid issue state %q: want %q or %q", state, IssueStateOpen, IssueStateClosed)
+	}
+	current, err := c.GetIssue(repoPath, number)
+	if err != nil {
+		return Issue{}, err
+	}
+	if current.State == state {
+		return current, fmt.Errorf("%s#%d: %w (%s)", repoPath, number, ErrIssueStateUnchanged, state)
+	}
+	var issue Issue
+	payload := map[string]string{"state": state}
+	err = c.do(http.MethodPatch, fmt.Sprintf("/api/v1/repos/%s/issues/%d", repoPath, number), payload, &issue)
+	return issue, err
+}
+
 // Comment is the subset of an issue comment we track.
 type Comment struct {
 	ID   int64  `json:"id"`

internal/agent/watch.go

@@ -3,6 +3,7 @@ package agent
 import (
 	"errors"
 	"fmt"
+	"strings"
 	"time"
 )
 
@@ -16,17 +17,65 @@ func IsPRGone(err error) bool {
 	return errors.Is(err, errPRGone)
 }
 
+// Mergeability is Gitea's mergeable flag. Gitea 1.26 always sends a plain bool,
+// and sends false both for a real conflict and while it recomputes the merge
+// base after a push, so false on its own decides nothing (prWatch resolves it).
+// Unknown covers what the bool cannot carry: an absent or null flag from another
+// Gitea build, and a snapshot no successful poll ever filled in.
+type Mergeability int
+
+const (
+	MergeUnknown Mergeability = iota
+	MergeYes
+	MergeNo
+)
+
+func (m Mergeability) String() string {
+	switch m {
+	case MergeYes:
+		return "true"
+	case MergeNo:
+		return "false"
+	}
+	return "unknown"
+}
+
+func (m Mergeability) MarshalJSON() ([]byte, error) {
+	switch m {
+	case MergeYes:
+		return []byte("true"), nil
+	case MergeNo:
+		return []byte("false"), nil
+	}
+	return []byte("null"), nil
+}
+
+func (m *Mergeability) UnmarshalJSON(b []byte) error {
+	switch strings.TrimSpace(string(b)) {
+	case "true":
+		*m = MergeYes
+	case "false":
+		*m = MergeNo
+	case "null":
+		*m = MergeUnknown
+	default:
+		return fmt.Errorf("mergeable: unexpected value %s", b)
+	}
+	return nil
+}
+
 // PRState is a point-in-time snapshot of the PR attributes watchpr tracks.
 type PRState struct {
-	Ref              PRRef  `json:"ref"`
-	State            string `json:"state"` // open / closed
-	Merged           bool   `json:"merged"`
-	HeadSHA          string `json:"head_sha"`
-	Mergeable        bool   `json:"mergeable"`
-	CIStatus         string `json:"ci_status"` // success / pending / failure / error / ""
-	NonAgentComments int    `json:"non_agent_comments"`
-	Title            string `json:"title"`
-	URL              string `json:"url"`
+	Ref              PRRef        `json:"ref"`
+	State            string       `json:"state"` // open / closed
+	Merged           bool         `json:"merged"`
+	HeadSHA          string       `json:"head_sha"`
+	BaseSHA          string       `json:"base_sha"`
+	Mergeable        Mergeability `json:"mergeable"`
+	CIStatus         string       `json:"ci_status"` // success / pending / failure / error / ""
+	NonAgentComments int          `json:"non_agent_comments"`
+	Title            string       `json:"title"`
+	URL              string       `json:"url"`
 }
 
 // FetchState builds a PRState for the given ref. agentLogin's comments are
@@ -55,6 +104,7 @@ func FetchState(c *GiteaClient, ref PRRef, agentLogin string) (PRState, error) {
 		State:            pr.State,
 		Merged:           pr.Merged,
 		HeadSHA:          pr.Head.Sha,
+		BaseSHA:          pr.Base.Sha,
 		Mergeable:        pr.Mergeable,
 		CIStatus:         ci,
 		NonAgentComments: countNonAgentComments(comments, agentLogin),
@@ -95,6 +145,121 @@ func terminalState(st PRState) (bool, string) {
 	return false, ""
 }
 
+// conflictWindow is how long an unbroken run of non-mergeable polls must span
+// before it is reported as a conflict. What the run has to outlast is Gitea's
+// merge recompute, which is a duration, so counting polls cannot express it:
+// --interval varies from seconds to hours between callers. Recomputes have been
+// observed finishing in ~5-20s with no bound on the tail, so this is set well
+// clear of that; the cost of it being too long is only a later conflict alert,
+// while too short is a false one.
+const conflictWindow = 2 * time.Minute
+
+// prWatch tracks one PR across polls, because mergeability needs more memory
+// than the previous snapshot. Gitea reports mergeable=false while it recomputes
+// the merge base after a push, so a false is only trusted once this watch has
+// seen a merge computation start: the PR was mergeable at some point, or its
+// head or base SHA moved. A bare false inherited from the baseline says nothing
+// -- it is equally a conflict the operator is already waiting on and a recompute
+// in flight -- so it arms nothing. Arming and the run of falses are independent:
+// movement only arms and never resets the run, because a base branch that moves
+// under the PR on every push to main would otherwise restart it forever. The
+// window is instead measured from the later of the run's start and the arm, so
+// an arming poll confirms nothing it has not itself outlasted.
+type prWatch struct {
+	prev  PRState
+	armed bool
+	// armedAt is when the rule armed mid-watch; nil while disarmed, and nil when
+	// the baseline armed it, since then there is no transition to measure from.
+	armedAt *time.Time
+	// conflictSince is when the current unbroken run of non-mergeable polls
+	// began; nil when no run is in progress.
+	conflictSince *time.Time
+}
+
+func newPRWatch(baseline PRState) *prWatch {
+	return &prWatch{prev: baseline, armed: baseline.Mergeable == MergeYes}
+}
+
+// arm records the disarmed->armed transition and when it happened. Re-arming is
+// a no-op, so a base branch moving under every poll advances nothing.
+func (w *prWatch) arm(now time.Time) {
+	if w.armed {
+		return
+	}
+	w.armed = true
+	at := now
+	w.armedAt = &at
+}
+
+// mergeInputsChanged reports whether the commits Gitea merges have moved, which
+// starts a fresh merge computation whose result is attributable to this watch.
+func mergeInputsChanged(prev, cur PRState) bool {
+	return cur.HeadSHA != prev.HeadSHA || cur.BaseSHA != prev.BaseSHA
+}
+
+// track folds one snapshot's mergeability into the run of observations. The
+// observations that confirm a conflict must be adjacent, so anything but
+// another non-mergeable one breaks the run.
+func (w *prWatch) track(st PRState, now time.Time) {
+	switch st.Mergeable {
+	case MergeNo:
+		if w.conflictSince == nil {
+			at := now
+			w.conflictSince = &at
+		}
+	case MergeYes:
+		w.arm(now)
+		w.conflictSince = nil
+	default:
+		w.conflictSince = nil
+	}
+}
+
+// missed records a poll that never produced a snapshot; the run of adjacent
+// non-mergeable observations does not survive the gap.
+func (w *prWatch) missed() {
+	w.conflictSince = nil
+}
+
+// windowStart is the later of the run's first observation and the arm, so the
+// window always covers observations this watch can attribute a merge
+// computation to. A baseline arm records no time and leaves the run governing.
+func (w *prWatch) windowStart() *time.Time {
+	if w.conflictSince == nil {
+		return nil
+	}
+	if w.armedAt != nil && w.armedAt.After(*w.conflictSince) {
+		return w.armedAt
+	}
+	return w.conflictSince
+}
+
+// confirmed reports whether the run of non-mergeable observations has spanned
+// the recompute window since it started counting, so neither a lone
+// non-mergeable poll nor the poll that armed the rule confirms anything.
+func (w *prWatch) confirmed(now time.Time) bool {
+	start := w.windowStart()
+	return start != nil && now.Sub(*start) >= conflictWindow
+}
+
+// observe folds in the newest snapshot, taken at now, and reports whether the
+// watch should end.
+func (w *prWatch) observe(cur PRState, now time.Time) (bool, string) {
+	changed, reason := MeaningfulChange(w.prev, cur)
+	if mergeInputsChanged(w.prev, cur) {
+		w.arm(now)
+	}
+	w.prev = cur
+	w.track(cur, now)
+	if changed {
+		return true, reason
+	}
+	if w.armed && w.confirmed(now) && cur.State == "open" {
+		return true, "PR lost mergeability (conflict)"
+	}
+	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.
@@ -109,10 +274,11 @@ const MaxPollFailures = 20
 // 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))
+// onBaseline, if set, receives every captured baseline once, before the first
+// tick, so a caller can show what state the watch started from.
+func Watch(f StateFetcher, refs []PRRef, agentLogin string, ticks <-chan time.Time, onBaseline func([]PRState), onError func(PRRef, error)) (WatchResult, error) {
+	watches := make(map[string]*prWatch, len(refs))
+	baselines := make([]PRState, 0, len(refs))
 	for _, ref := range refs {
 		st, err := f.FetchState(ref, agentLogin)
 		if err != nil {
@@ -121,13 +287,16 @@ func Watch(f StateFetcher, refs []PRRef, agentLogin string, ticks <-chan time.Ti
 		if terminal, reason := terminalState(st); terminal {
 			return WatchResult{Ref: ref, Reason: reason, State: st}, nil
 		}
-		prev[ref.String()] = st
+		watches[ref.String()] = newPRWatch(st)
+		baselines = append(baselines, st)
 	}
 	if onBaseline != nil {
-		onBaseline()
+		onBaseline(baselines)
 	}
 	fails := make(map[string]int, len(refs))
-	for range ticks {
+	// The tick carries the time it fired, which is the clock the conflict
+	// window is measured on.
+	for now := range ticks {
 		for _, ref := range refs {
 			key := ref.String()
 			cur, err := f.FetchState(ref, agentLogin)
@@ -136,6 +305,7 @@ func Watch(f StateFetcher, refs []PRRef, agentLogin string, ticks <-chan time.Ti
 					return WatchResult{}, fmt.Errorf("polling %s: %w", key, err)
 				}
 				fails[key]++
+				watches[key].missed()
 				if onError != nil {
 					onError(ref, err)
 				}
@@ -145,10 +315,9 @@ func Watch(f StateFetcher, refs []PRRef, agentLogin string, ticks <-chan time.Ti
 				continue
 			}
 			fails[key] = 0
-			if changed, reason := MeaningfulChange(prev[key], cur); changed {
+			if changed, reason := watches[key].observe(cur, now); changed {
 				return WatchResult{Ref: ref, Reason: reason, State: cur}, nil
 			}
-			prev[key] = cur
 		}
 	}
 	return WatchResult{}, nil
@@ -172,15 +341,15 @@ func isFailedCI(state string) bool {
 
 // MeaningfulChange compares a previous state to the current one and reports
 // whether a change warrants alerting the operator, with a human-readable
-// reason. Benign transitions (CI pending→success, the agent's own comments, an
-// unchanged snapshot) return false.
+// reason. Benign transitions (CI pending→success, the agent's own comments, a
+// new head or base commit, an unchanged snapshot) return false. Mergeability is
+// not decided here: it takes a whole run of observations, which prWatch keeps.
 //
 // Alerting conditions:
 //   - the PR merged
 //   - the PR closed without merging
 //   - a new comment from someone other than the agent
 //   - CI transitioned into failure/error
-//   - the PR lost mergeability (a conflict appeared) for two consecutive polls
 func MeaningfulChange(prev, cur PRState) (bool, string) {
 	if !prev.Merged && cur.Merged {
 		return true, "PR merged"
@@ -195,12 +364,5 @@ func MeaningfulChange(prev, cur PRState) (bool, string) {
 	if isFailedCI(cur.CIStatus) && !isFailedCI(prev.CIStatus) {
 		return true, "CI failed (" + cur.CIStatus + ")"
 	}
-	// Gitea computes mergeability asynchronously, so a PR can briefly report
-	// mergeable=false right after a push. Require the loss to persist across two
-	// consecutive polls (both prev and cur false, still open) before treating it
-	// as a real conflict; a single false poll is debounced.
-	if !prev.Mergeable && !cur.Mergeable && cur.State == "open" {
-		return true, "PR lost mergeability (conflict)"
-	}
 	return false, ""
 }

internal/agent/watch_test.go

@@ -1,6 +1,7 @@
 package agent
 
 import (
+	"encoding/json"
 	"errors"
 	"fmt"
 	"net/http"
@@ -17,7 +18,8 @@ func base() PRState {
 		State:            "open",
 		Merged:           false,
 		HeadSHA:          "abc123",
-		Mergeable:        true,
+		BaseSHA:          "base000",
+		Mergeable:        MergeYes,
 		CIStatus:         "pending",
 		NonAgentComments: 0,
 	}
@@ -66,25 +68,11 @@ func TestMeaningfulChange(t *testing.T) {
 			wantChange: true,
 		},
 		{
-			// A single mergeable=false poll is debounced: Gitea often reports
-			// this transiently right after a push.
-			name:       "mergeable true to false for one poll is benign",
-			mutate:     func(s *PRState) { s.Mergeable = false },
-			wantChange: false,
-		},
-		{
-			// mergeable=false persisting into a second consecutive poll is a
-			// real conflict and alerts.
-			name:       "mergeable false persisting a second poll alerts",
-			mutatePrev: func(s *PRState) { s.Mergeable = false },
-			mutate:     func(s *PRState) { s.Mergeable = false },
-			wantChange: true,
-		},
-		{
-			// mergeable recovered (false then true) must not alert.
-			name:       "mergeable recovered false to true is benign",
-			mutatePrev: func(s *PRState) { s.Mergeable = false },
-			mutate:     func(s *PRState) {},
+			// Mergeability takes a run of observations, so no pair of snapshots
+			// decides it here; prWatch owns that rule.
+			name:       "mergeable false pair alone is not a pairwise change",
+			mutatePrev: func(s *PRState) { s.Mergeable = MergeNo },
+			mutate:     func(s *PRState) { s.Mergeable = MergeNo },
 			wantChange: false,
 		},
 		{
@@ -222,16 +210,16 @@ func TestWatchDetectsMergeAfterBaseline(t *testing.T) {
 	merged.Merged = true
 	f := &fakeFetcher{states: []PRState{open, merged}} // baseline open, then merged
 
-	baselineFired := false
+	var baselines []PRState
 	ticks := make(chan time.Time, 1)
 	ticks <- time.Now()
 	res, err := Watch(f, []PRRef{open.Ref}, "unkin-agent",
-		ticks, func() { baselineFired = true }, nil)
+		ticks, func(sts []PRState) { baselines = sts }, nil)
 	if err != nil {
 		t.Fatalf("Watch: %v", err)
 	}
-	if !baselineFired {
-		t.Errorf("onBaseline should fire for an open baseline")
+	if len(baselines) != 1 || baselines[0].Ref != open.Ref {
+		t.Errorf("onBaseline received %v, want the one open baseline", baselines)
 	}
 	if res.Reason != "PR merged" {
 		t.Errorf("reason = %q, want %q", res.Reason, "PR merged")
@@ -801,3 +789,603 @@ func TestWatchAnonymousKeepsPolling(t *testing.T) {
 		t.Errorf("fetch calls = %d, want 3 (baseline + two polls)", f.calls)
 	}
 }
+
+// drainableTicks returns a channel holding n ticks and already closed, so Watch
+// polls exactly n times and then returns instead of blocking. The ticks are
+// spaced far past conflictWindow, so a run of non-mergeable polls confirms on
+// its second observation; spacedTicks drives the intervals where it must not.
+func drainableTicks(n int) <-chan time.Time {
+	return spacedTicks(n, 10*time.Minute)
+}
+
+// spacedTicks is drainableTicks with the poll interval named. The tick carries
+// the time Watch measures the conflict window on, which makes every wall-clock
+// assertion in these tests exact and instant.
+func spacedTicks(n int, interval time.Duration) <-chan time.Time {
+	start := time.Date(2026, 9, 24, 12, 0, 0, 0, time.UTC)
+	ticks := make(chan time.Time, n)
+	for i := 1; i <= n; i++ {
+		ticks <- start.Add(time.Duration(i) * interval)
+	}
+	close(ticks)
+	return ticks
+}
+
+// The production bug: a PR that was already conflicted (and already CI-failing)
+// when watching began must not be reported as having just changed. That state is
+// what the watcher is waiting to see resolved, so the loop keeps polling.
+func TestWatchIgnoresBaselineConflictAndFailure(t *testing.T) {
+	stuck := base()
+	stuck.Mergeable = MergeNo
+	stuck.CIStatus = "failure"
+	f := &fakeFetcher{states: []PRState{stuck}}
+
+	res, err := Watch(f, []PRRef{stuck.Ref}, "unkin-agent", drainableTicks(5), nil, nil)
+	if err != nil {
+		t.Fatalf("Watch: %v", err)
+	}
+	if res.Reason != "" {
+		t.Fatalf("Watch ended with %q; a conflict/failure predating the watch is not a change", res.Reason)
+	}
+	if f.calls != 6 {
+		t.Errorf("fetch calls = %d, want 6 (baseline + 5 polls)", f.calls)
+	}
+}
+
+// Mergeability lost after the baseline still alerts, on the second consecutive
+// conflicted poll.
+func TestWatchDetectsConflictAfterBaseline(t *testing.T) {
+	ok := base()
+	conflicted := base()
+	conflicted.Mergeable = MergeNo
+	f := &fakeFetcher{states: []PRState{ok, conflicted, conflicted}}
+
+	res, err := Watch(f, []PRRef{ok.Ref}, "unkin-agent", drainableTicks(3), nil, nil)
+	if err != nil {
+		t.Fatalf("Watch: %v", err)
+	}
+	if res.Reason != "PR lost mergeability (conflict)" {
+		t.Errorf("reason = %q, want the mergeability loss", res.Reason)
+	}
+	if f.calls != 3 {
+		t.Errorf("fetch calls = %d, want 3 (baseline + the two conflicted polls)", f.calls)
+	}
+}
+
+// CI that goes green→red during the watch alerts.
+func TestWatchDetectsCIFailureAfterBaseline(t *testing.T) {
+	ok := base()
+	ok.CIStatus = "pending"
+	green := base()
+	green.CIStatus = "success"
+	red := base()
+	red.CIStatus = "failure"
+	f := &fakeFetcher{states: []PRState{ok, green, red}}
+
+	res, err := Watch(f, []PRRef{ok.Ref}, "unkin-agent", drainableTicks(3), nil, nil)
+	if err != nil {
+		t.Fatalf("Watch: %v", err)
+	}
+	if res.Reason != "CI failed (failure)" {
+		t.Errorf("reason = %q, want the CI failure (pending→success must pass silently)", res.Reason)
+	}
+	if f.calls != 3 {
+		t.Errorf("fetch calls = %d, want 3 (the success poll must not end the watch)", f.calls)
+	}
+}
+
+// The agent's own pushes and comments must not end a watch; a comment from
+// anyone else must.
+func TestWatchIgnoresAgentActivity(t *testing.T) {
+	start := base()
+	pushed := base()
+	pushed.HeadSHA = "def456" // the agent pushed a fix; non-agent comments unchanged
+	commented := pushed
+	commented.NonAgentComments = 1
+	f := &fakeFetcher{states: []PRState{start, pushed, pushed, commented}}
+
+	res, err := Watch(f, []PRRef{start.Ref}, "unkin-agent", drainableTicks(3), nil, nil)
+	if err != nil {
+		t.Fatalf("Watch: %v", err)
+	}
+	if res.Reason != "new comment from a non-agent user" {
+		t.Errorf("reason = %q, want the non-agent comment", res.Reason)
+	}
+	if f.calls != 4 {
+		t.Errorf("fetch calls = %d, want 4 (the agent's push and comment must not end the watch)", f.calls)
+	}
+}
+
+func TestPRWatchMergeability(t *testing.T) {
+	snap := func(m Mergeability) PRState {
+		s := base()
+		s.Mergeable = m
+		return s
+	}
+	tests := []struct {
+		name     string
+		baseline Mergeability
+		polls    []Mergeability
+		wantPoll int // 1-based poll that ends the watch; 0 for none
+	}{
+		{"conflicted before the watch never alerts", MergeNo, []Mergeability{MergeNo, MergeNo, MergeNo}, 0},
+		{"loss after a mergeable baseline alerts on the second poll", MergeYes, []Mergeability{MergeNo, MergeNo}, 2},
+		{"a lone conflicted poll is debounced", MergeYes, []Mergeability{MergeNo, MergeYes, MergeNo}, 0},
+		{"unknown at baseline does not arm the rule", MergeUnknown, []Mergeability{MergeNo, MergeNo, MergeNo}, 0},
+		{"unknown at baseline then a real loss alerts", MergeUnknown, []Mergeability{MergeYes, MergeNo, MergeNo}, 3},
+		{"a baseline conflict resolved then lost again alerts", MergeNo, []Mergeability{MergeYes, MergeNo, MergeNo}, 3},
+		{"unknown between conflicted polls breaks the run", MergeYes, []Mergeability{MergeNo, MergeUnknown, MergeNo}, 0},
+		{"the run restarts after an unknown", MergeYes, []Mergeability{MergeNo, MergeUnknown, MergeNo, MergeNo}, 4},
+		{"unknown alone is never a conflict", MergeYes, []Mergeability{MergeUnknown, MergeUnknown, MergeUnknown}, 0},
+	}
+	for _, tt := range tests {
+		t.Run(tt.name, func(t *testing.T) {
+			w := newPRWatch(snap(tt.baseline))
+			start := time.Date(2026, 9, 24, 12, 0, 0, 0, time.UTC)
+			got := 0
+			for i, m := range tt.polls {
+				changed, reason := w.observe(snap(m), start.Add(time.Duration(i+1)*10*time.Minute))
+				if !changed {
+					continue
+				}
+				if reason != "PR lost mergeability (conflict)" {
+					t.Fatalf("poll %d ended the watch with %q, want a mergeability loss", i+1, reason)
+				}
+				got = i + 1
+				break
+			}
+			if got != tt.wantPoll {
+				t.Errorf("alerted on poll %d, want %d", got, tt.wantPoll)
+			}
+		})
+	}
+}
+
+// Gitea 1.26 always sends mergeable as a plain bool, so unknown is reserved for
+// what that bool cannot carry: an absent or null flag must decode as unknown
+// rather than as a conflict, and unknown must encode back as null.
+func TestMergeabilityDecoding(t *testing.T) {
+	tests := map[string]Mergeability{
+		`{"number":7}`:                   MergeUnknown,
+		`{"number":7,"mergeable":null}`:  MergeUnknown,
+		`{"number":7,"mergeable":true}`:  MergeYes,
+		`{"number":7,"mergeable":false}`: MergeNo,
+	}
+	for body, want := range tests {
+		var pr PullRequest
+		if err := json.Unmarshal([]byte(body), &pr); err != nil {
+			t.Fatalf("Unmarshal(%s): %v", body, err)
+		}
+		if pr.Mergeable != want {
+			t.Errorf("Unmarshal(%s) mergeable = %v, want %v", body, pr.Mergeable, want)
+		}
+	}
+	out, err := json.Marshal(PRState{Mergeable: MergeUnknown})
+	if err != nil {
+		t.Fatalf("Marshal: %v", err)
+	}
+	if !strings.Contains(string(out), `"mergeable":null`) {
+		t.Errorf("unknown mergeability encoded as %s, want null", out)
+	}
+}
+
+// The inverse of the bug this PR fixed: demanding a mergeable observation to arm
+// the rule left a PR that was non-mergeable at baseline silent forever, even
+// after a push gave Gitea a fresh merge computation to answer for. A head that
+// moves during the watch arms the rule, so the conflict the new head keeps
+// reporting is attributable to this watch and is reported.
+func TestWatchAlertsWhenAPushArmsABaselineConflict(t *testing.T) {
+	baseline := base()
+	baseline.Mergeable = MergeNo // still recomputing; the push is not visible yet
+
+	pushed := base()
+	pushed.Mergeable = MergeNo
+	pushed.HeadSHA = "def456"
+
+	f := &fakeFetcher{states: []PRState{baseline, pushed, pushed, pushed}}
+	res, err := Watch(f, []PRRef{baseline.Ref}, "unkin-agent", drainableTicks(3), nil, nil)
+	if err != nil {
+		t.Fatalf("Watch: %v", err)
+	}
+	if res.Reason != "PR lost mergeability (conflict)" {
+		t.Fatalf("reason = %q, want the mergeability loss; a conflict must not be unreportable because the baseline caught the recompute", res.Reason)
+	}
+	if f.calls != 3 {
+		t.Errorf("fetch calls = %d, want 3 (baseline + the two conflicted polls on the new head)", f.calls)
+	}
+}
+
+// A base-branch move restarts the same merge computation, so a conflict that
+// only becomes visible once someone merges into main alerts too.
+func TestWatchAlertsWhenBaseMovedIntoAConflict(t *testing.T) {
+	baseline := base()
+	baseline.Mergeable = MergeNo
+
+	moved := base()
+	moved.Mergeable = MergeNo
+	moved.BaseSHA = "base111"
+
+	f := &fakeFetcher{states: []PRState{baseline, moved, moved}}
+	res, err := Watch(f, []PRRef{baseline.Ref}, "unkin-agent", drainableTicks(3), nil, nil)
+	if err != nil {
+		t.Fatalf("Watch: %v", err)
+	}
+	if res.Reason != "PR lost mergeability (conflict)" {
+		t.Errorf("reason = %q, want the mergeability loss", res.Reason)
+	}
+}
+
+// The fix this PR exists for: a conflict predating the watch, on a head and base
+// that never move, is the condition the operator is already waiting on and must
+// stay silent however long the watch runs.
+func TestWatchNeverAlertsOnAStableBaselineConflict(t *testing.T) {
+	stuck := base()
+	stuck.Mergeable = MergeNo
+	f := &fakeFetcher{states: []PRState{stuck}}
+
+	res, err := Watch(f, []PRRef{stuck.Ref}, "unkin-agent", drainableTicks(100), nil, nil)
+	if err != nil {
+		t.Fatalf("Watch: %v", err)
+	}
+	if res.Reason != "" {
+		t.Fatalf("Watch ended with %q after 100 unchanged polls; a conflict that predates the watch is not a change", res.Reason)
+	}
+}
+
+// Head and base movement arm the conflict rule but are not themselves alerts: a
+// push, or a base that moves under the PR, must not end a watch.
+func TestWatchArmingMovementDoesNotAlert(t *testing.T) {
+	start := base()
+	pushed := base()
+	pushed.HeadSHA = "def456"
+	rebased := base()
+	rebased.HeadSHA = "789abc"
+	rebased.BaseSHA = "base111"
+	f := &fakeFetcher{states: []PRState{start, pushed, rebased}}
+
+	res, err := Watch(f, []PRRef{start.Ref}, "unkin-agent", drainableTicks(5), nil, nil)
+	if err != nil {
+		t.Fatalf("Watch: %v", err)
+	}
+	if res.Reason != "" {
+		t.Fatalf("Watch ended with %q; a new head or base is not a change worth alerting on", res.Reason)
+	}
+}
+
+// The polls that confirm a conflict must be adjacent. A failed poll produces no
+// snapshot, so the run cannot span it and two non-adjacent falses do not fire.
+func TestWatchConflictRunDoesNotSpanAFailedPoll(t *testing.T) {
+	ok := base()
+	conflicted := base()
+	conflicted.Mergeable = MergeNo
+	f := &fakeFetcher{
+		states: []PRState{ok, conflicted, conflicted, conflicted, conflicted},
+		errs:   []error{nil, nil, errors.New("HTTP 502"), nil, nil},
+	}
+
+	res, err := Watch(f, []PRRef{ok.Ref}, "unkin-agent", drainableTicks(4), nil, func(PRRef, error) {})
+	if err != nil {
+		t.Fatalf("Watch: %v", err)
+	}
+	if res.Reason != "PR lost mergeability (conflict)" {
+		t.Fatalf("reason = %q, want the mergeability loss on the two adjacent polls", res.Reason)
+	}
+	if f.calls != 5 {
+		t.Errorf("fetch calls = %d, want 5: the falses either side of the failed poll are not a run", f.calls)
+	}
+}
+
+// perRefFetcher scripts a separate sequence per ref, so a multi-ref watch can be
+// driven with each PR doing something different.
+type perRefFetcher struct {
+	states map[string][]PRState
+	calls  map[string]int
+}
+
+func (f *perRefFetcher) FetchState(ref PRRef, _ string) (PRState, error) {
+	key := ref.String()
+	seq := f.states[key]
+	i := f.calls[key]
+	if i >= len(seq) {
+		i = len(seq) - 1
+	}
+	f.calls[key]++
+	return seq[i], nil
+}
+
+// Each ref keeps its own run of observations: one PR's mergeability, pushes and
+// resolutions must neither arm nor disarm another's conflict rule.
+func TestWatchTracksRefsIndependently(t *testing.T) {
+	refA := PRRef{Owner: "unkin", Repo: "repo", Number: 1}
+	refB := PRRef{Owner: "unkin", Repo: "repo", Number: 2}
+	snap := func(ref PRRef, m Mergeability, head string) PRState {
+		s := base()
+		s.Ref = ref
+		s.Mergeable = m
+		s.HeadSHA = head
+		return s
+	}
+
+	tests := []struct {
+		name string
+		a, b []PRState
+		want string // reason, "" for no alert
+	}{
+		{
+			name: "a mergeable neighbour does not arm a baseline conflict",
+			a:    []PRState{snap(refA, MergeYes, "aaa")},
+			b:    []PRState{snap(refB, MergeNo, "bbb")},
+			want: "",
+		},
+		{
+			name: "a neighbour's push does not arm a baseline conflict",
+			a:    []PRState{snap(refA, MergeYes, "aaa"), snap(refA, MergeYes, "aa2"), snap(refA, MergeYes, "aa3")},
+			b:    []PRState{snap(refB, MergeNo, "bbb")},
+			want: "",
+		},
+		{
+			name: "a neighbour's mergeable polls do not clear another's run",
+			a:    []PRState{snap(refA, MergeYes, "aaa")},
+			b:    []PRState{snap(refB, MergeYes, "bbb"), snap(refB, MergeNo, "bbb"), snap(refB, MergeNo, "bbb")},
+			want: "PR lost mergeability (conflict)",
+		},
+	}
+	for _, tt := range tests {
+		t.Run(tt.name, func(t *testing.T) {
+			f := &perRefFetcher{
+				states: map[string][]PRState{refA.String(): tt.a, refB.String(): tt.b},
+				calls:  map[string]int{},
+			}
+			res, err := Watch(f, []PRRef{refA, refB}, "unkin-agent", drainableTicks(6), nil, nil)
+			if err != nil {
+				t.Fatalf("Watch: %v", err)
+			}
+			if res.Reason != tt.want {
+				t.Fatalf("reason = %q, want %q", res.Reason, tt.want)
+			}
+			if tt.want != "" && res.Ref != refB {
+				t.Errorf("alert names %s, want %s", res.Ref, refB)
+			}
+		})
+	}
+}
+
+// A watch that will deliberately stay silent about a pre-existing conflict has
+// to hand its caller the baseline it is staying silent about.
+func TestWatchReportsBaselineStates(t *testing.T) {
+	stuck := base()
+	stuck.Mergeable = MergeNo
+	stuck.CIStatus = "failure"
+	f := &fakeFetcher{states: []PRState{stuck}}
+
+	var got []PRState
+	if _, err := Watch(f, []PRRef{stuck.Ref}, "unkin-agent", drainableTicks(1),
+		func(sts []PRState) { got = sts }, nil); err != nil {
+		t.Fatalf("Watch: %v", err)
+	}
+	if len(got) != 1 {
+		t.Fatalf("onBaseline received %d state(s), want 1", len(got))
+	}
+	if got[0].Mergeable != MergeNo || got[0].CIStatus != "failure" {
+		t.Errorf("baseline = %+v, want the conflicted, CI-red snapshot the watch started from", got[0])
+	}
+}
+
+// The residual gap, pinned so it is a decision rather than an accident: the
+// agent pushes, the push conflicts, watchpr starts inside Gitea's recompute and
+// nothing moves again. Every poll answers false and none of them is
+// attributable to this watch, so no alert is ever sent -- the baseline line is
+// the only notice. Gitea's payload carries no field that separates this from a
+// merge check still running.
+func TestWatchNeverAlertsOnAConflictLandedByThePushBeforeTheWatch(t *testing.T) {
+	pushed := base()
+	pushed.Mergeable = MergeNo
+	pushed.HeadSHA = "def456"
+	f := &fakeFetcher{states: []PRState{pushed}}
+
+	res, err := Watch(f, []PRRef{pushed.Ref}, "unkin-agent", drainableTicks(50), nil, nil)
+	if err != nil {
+		t.Fatalf("Watch: %v", err)
+	}
+	if res.Reason != "" {
+		t.Fatalf("Watch ended with %q; nothing distinguishes this conflict from a merge check in flight, so it must stay silent", res.Reason)
+	}
+}
+
+// The whole mergeability rule as a sequence table, driven through Watch at the
+// intervals that decide it. The debounce is wall-clock, so the same sequence of
+// snapshots must alert or stay silent according to how far apart the polls are.
+func TestWatchConflictSequences(t *testing.T) {
+	snap := func(m Mergeability, head, bse string) PRState {
+		st := base()
+		st.Mergeable = m
+		st.HeadSHA = head
+		st.BaseSHA = bse
+		return st
+	}
+	repeat := func(st PRState, n int) []PRState {
+		out := make([]PRState, n)
+		for i := range out {
+			out[i] = st
+		}
+		return out
+	}
+	movingBase := func(m Mergeability, n int) []PRState {
+		out := make([]PRState, n)
+		for i := range out {
+			out[i] = snap(m, "h1", fmt.Sprintf("b%d", i))
+		}
+		return out
+	}
+	const (
+		quick   = 5 * time.Second  // well inside conflictWindow
+		normal  = 60 * time.Second // watchpr's default
+		relaxed = 10 * time.Minute // past conflictWindow in a single gap
+	)
+	const conflict = "PR lost mergeability (conflict)"
+
+	yes := snap(MergeYes, "h1", "b1")
+	no := snap(MergeNo, "h1", "b1")
+	pushedNo := snap(MergeNo, "h2", "b1")
+	movedNo := snap(MergeNo, "h1", "b2")
+
+	tests := []struct {
+		name     string
+		interval time.Duration
+		baseline PRState
+		polls    []PRState
+		errs     []error
+		want     string
+	}{
+		{
+			name:     "a push then the recompute's falses is not a conflict",
+			interval: normal, baseline: yes, polls: repeat(pushedNo, 2), want: "",
+		},
+		{
+			name:     "a fast poller rides out the whole recompute after a push",
+			interval: quick, baseline: yes, polls: repeat(pushedNo, 20), want: "",
+		},
+		{
+			name:     "a base move then the recompute's falses is not a conflict",
+			interval: normal, baseline: no, polls: repeat(movedNo, 2), want: "",
+		},
+		{
+			name:     "a base moving under every poll still confirms a conflict",
+			interval: normal, baseline: yes, polls: movingBase(MergeNo, 20), want: conflict,
+		},
+		{
+			name:     "a base move does not confirm a run it took no part in",
+			interval: normal, baseline: no, polls: []PRState{no, no, no, movedNo}, want: "",
+		},
+		{
+			name:     "a push does not confirm a run it took no part in",
+			interval: normal, baseline: no, polls: []PRState{no, no, no, pushedNo}, want: "",
+		},
+		{
+			name:     "an armed run confirms a window after the arm, not before",
+			interval: normal, baseline: no,
+			polls: []PRState{no, no, no, movedNo, movedNo, movedNo}, want: conflict,
+		},
+		{
+			name:     "a conflict that predates the watch stays silent forever",
+			interval: relaxed, baseline: no, polls: repeat(no, 50), want: "",
+		},
+		{
+			name:     "a sustained loss after a mergeable baseline alerts",
+			interval: normal, baseline: yes, polls: repeat(no, 3), want: conflict,
+		},
+		{
+			name:     "a mergeable poll breaks the run however long it ran",
+			interval: relaxed, baseline: yes, polls: []PRState{no, yes, no}, want: "",
+		},
+		{
+			name:     "an unknown poll breaks the run however long it ran",
+			interval: relaxed, baseline: yes, polls: []PRState{no, snap(MergeUnknown, "h1", "b1"), no}, want: "",
+		},
+		{
+			name:     "a baseline conflict that clears and returns alerts",
+			interval: relaxed, baseline: no, polls: []PRState{yes, no, no}, want: conflict,
+		},
+		{
+			name:     "a failed poll breaks the run",
+			interval: relaxed, baseline: yes, polls: []PRState{no, no, no},
+			errs: []error{nil, nil, errors.New("HTTP 502"), nil}, want: "",
+		},
+		{
+			name:     "the run restarts after a failed poll and still confirms",
+			interval: relaxed, baseline: yes, polls: []PRState{no, no, no, no},
+			errs: []error{nil, nil, errors.New("HTTP 502"), nil, nil}, want: conflict,
+		},
+	}
+	for _, tt := range tests {
+		t.Run(tt.name, func(t *testing.T) {
+			f := &fakeFetcher{states: append([]PRState{tt.baseline}, tt.polls...), errs: tt.errs}
+			res, err := Watch(f, []PRRef{tt.baseline.Ref}, "unkin-agent",
+				spacedTicks(len(tt.polls), tt.interval), nil, func(PRRef, error) {})
+			if err != nil {
+				t.Fatalf("Watch: %v", err)
+			}
+			if res.Reason != tt.want {
+				t.Fatalf("reason = %q, want %q", res.Reason, tt.want)
+			}
+		})
+	}
+}
+
+// Two refs watched together keep separate runs and separate clocks: a PR being
+// pushed to must neither delay nor trigger the conflict its neighbour is really
+// in, and the alert must name the conflicted one.
+func TestWatchConflictIsolatedFromANeighboursPushes(t *testing.T) {
+	conflicted := PRRef{Owner: "unkin", Repo: "repo", Number: 1}
+	pushing := PRRef{Owner: "unkin", Repo: "repo", Number: 2}
+	snap := func(ref PRRef, m Mergeability, head string) PRState {
+		st := base()
+		st.Ref = ref
+		st.Mergeable = m
+		st.HeadSHA = head
+		return st
+	}
+	f := &perRefFetcher{
+		states: map[string][]PRState{
+			conflicted.String(): {snap(conflicted, MergeYes, "a1"), snap(conflicted, MergeNo, "a1")},
+			pushing.String():    {snap(pushing, MergeYes, "b1"), snap(pushing, MergeYes, "b2"), snap(pushing, MergeYes, "b3")},
+		},
+		calls: map[string]int{},
+	}
+
+	res, err := Watch(f, []PRRef{conflicted, pushing}, "unkin-agent", drainableTicks(4), nil, nil)
+	if err != nil {
+		t.Fatalf("Watch: %v", err)
+	}
+	if res.Reason != "PR lost mergeability (conflict)" {
+		t.Fatalf("reason = %q, want the conflict on %s", res.Reason, conflicted)
+	}
+	if res.Ref != conflicted {
+		t.Errorf("alert names %s, want %s", res.Ref, conflicted)
+	}
+}
+
+// The window is measured from the arm, not from the run the arming poll walked
+// in on: a base branch that moves under a long-conflicted PR arms the rule and
+// then has to outlast the recompute itself before anything is reported.
+func TestWatchConflictWindowStartsAtTheArm(t *testing.T) {
+	stuck := base()
+	stuck.Mergeable = MergeNo
+	moved := stuck
+	moved.BaseSHA = "base111"
+
+	f := &fakeFetcher{states: []PRState{stuck, stuck, stuck, stuck, moved, moved, moved}}
+	res, err := Watch(f, []PRRef{stuck.Ref}, "unkin-agent", spacedTicks(6, time.Minute), nil, nil)
+	if err != nil {
+		t.Fatalf("Watch: %v", err)
+	}
+	if res.Reason != "PR lost mergeability (conflict)" {
+		t.Fatalf("reason = %q, want the mergeability loss two minutes after the base moved", res.Reason)
+	}
+	if f.calls != 7 {
+		t.Errorf("fetch calls = %d, want 7: the window runs from the arming poll, not from the run it inherited", f.calls)
+	}
+}
+
+// The zero time is a legal clock value, so it must not double as the "no run in
+// progress" sentinel: a caller whose ticks start at the zero time gets the same
+// debounce as any other.
+func TestWatchConflictWindowCountsFromTheZeroTime(t *testing.T) {
+	ok := base()
+	conflicted := base()
+	conflicted.Mergeable = MergeNo
+
+	ticks := make(chan time.Time, 2)
+	ticks <- time.Time{}
+	ticks <- time.Time{}.Add(10 * time.Minute)
+	close(ticks)
+
+	f := &fakeFetcher{states: []PRState{ok, conflicted, conflicted}}
+	res, err := Watch(f, []PRRef{ok.Ref}, "unkin-agent", ticks, nil, nil)
+	if err != nil {
+		t.Fatalf("Watch: %v", err)
+	}
+	if res.Reason != "PR lost mergeability (conflict)" {
+		t.Fatalf("reason = %q, want the mergeability loss; a run starting at the zero time still counts", res.Reason)
+	}
+}