package handler

import (
	"bytes"
	"context"
	"encoding/json"
	"errors"
	"fmt"
	"io"
	"log/slog"
	"net/http"
	"net/url"
	"regexp"
	"sort"
	"strconv"
	"strings"
	"time"
	"unicode"

	"github.com/go-chi/chi/v5"
	"github.com/jackc/pgx/v5"
	"github.com/jackc/pgx/v5/pgtype"
	"github.com/multica-ai/multica/server/internal/channelmedia"
	"github.com/multica-ai/multica/server/internal/dispatch"
	"github.com/multica-ai/multica/server/internal/issueguard"
	"github.com/multica-ai/multica/server/internal/issuestatus"
	"github.com/multica-ai/multica/server/internal/issueworkflow"
	"github.com/multica-ai/multica/server/internal/logger"
	obsmetrics "github.com/multica-ai/multica/server/internal/metrics"
	"github.com/multica-ai/multica/server/internal/middleware"
	"github.com/multica-ai/multica/server/internal/qoderruntime"
	"github.com/multica-ai/multica/server/internal/service"
	"github.com/multica-ai/multica/server/internal/util"
	agentpkg "github.com/multica-ai/multica/server/pkg/agent"
	db "github.com/multica-ai/multica/server/pkg/db/generated"
	"github.com/multica-ai/multica/server/pkg/protocol"
)

// IssueResponse is the JSON response for an issue.
type IssueResponse struct {
	ID          string  `json:"id"`
	WorkspaceID string  `json:"workspace_id"`
	Number      int32   `json:"number"`
	Identifier  string  `json:"identifier"`
	Title       string  `json:"title"`
	Description *string `json:"description"`
	Status      string  `json:"status"`
	// StatusCategory encodes lifecycle using the legacy seven-value wire enum. It is
	// omitted when an endpoint cannot resolve a custom status, so consumers must
	// fall back to their catalog rather than treat a blank as "no category".
	StatusCategory string `json:"status_category,omitempty"`
	// StatusName is a CUSTOM status's display name, carried beside the key so a
	// consumer that only ever sees `status` is not left holding a bare handle.
	// A key derived from a non-Latin name is opaque by construction
	// (`in_review_2`), and an agent reading an issue has nothing else to match
	// against the status a human named for it.
	//
	// Always emitted, unlike StatusCategory. Empty is a MEANING here — "this is
	// a built-in, localize it from the key" — not the "this endpoint did not
	// resolve it" that an absent status_category signals. Keeping the key
	// present is also what lets TestIssueToMap_KeysMatchIssueResponse see the
	// field at all: with omitempty a built-in fixture hides it from BOTH
	// renderings, and the drift guard goes green on a payload that has drifted.
	// (MUL-6749)
	StatusName    string  `json:"status_name"`
	Priority      string  `json:"priority"`
	AssigneeType  *string `json:"assignee_type"`
	AssigneeID    *string `json:"assignee_id"`
	CreatorType   string  `json:"creator_type"`
	CreatorID     string  `json:"creator_id"`
	ParentIssueID *string `json:"parent_issue_id"`
	// DuplicateOf is the original this issue duplicates (MUL-7349), set only
	// while the mark counts: the issue is cancelled and the original still
	// exists. newStatusCategoryFiller resolves it from duplicateOfIssueID;
	// an endpoint that skips the filler emits null rather than a bare pointer.
	DuplicateOf *IssueRefResponse `json:"duplicate_of"`
	ProjectID   *string           `json:"project_id"`
	Position    float64           `json:"position"`
	// Stage groups sub-issues under the same parent into ordered barrier
	// groups (null = unstaged). See service/issue_wakeup_system.go for how a
	// closed stage wakes the parent's assignee.
	Stage     *int32  `json:"stage"`
	StartDate *string `json:"start_date"`
	DueDate   *string `json:"due_date"`
	CreatedAt string  `json:"created_at"`
	UpdatedAt string  `json:"updated_at"`
	Revision  int64   `json:"revision"`
	// LastActivityAt is the latest semantic issue activity. It stays nullable
	// while the operator-run historical backfill is incomplete.
	LastActivityAt *string `json:"last_activity_at"`
	// Metadata is the per-issue KV map (see issue_metadata.go). Always emitted
	// (empty object when unset) so frontend code can `issue.metadata[key]`
	// without nil-guarding the parent field.
	Metadata map[string]any `json:"metadata"`
	// Properties is the custom-property value bag keyed by property definition
	// UUID (see property.go). Always emitted, mirroring Metadata.
	Properties  map[string]any          `json:"properties"`
	Reactions   []IssueReactionResponse `json:"reactions,omitempty"`
	Attachments []AttachmentResponse    `json:"attachments,omitempty"`
	// Labels are bulk-attached by list/detail endpoints so the client can render
	// chips without an N+1 round-trip per row. Pointer + omitempty so paths that
	// don't load labels (e.g. UpdateIssue, batch UpdateIssues, the issue:updated
	// WS broadcast) emit no `labels` field at all — the client merge then
	// preserves whatever labels are already in cache. nil pointer = "field
	// absent, do not touch"; non-nil (incl. empty slice) = authoritative list.
	Labels *[]LabelResponse `json:"labels,omitempty"`
	// SourceContext is detail-only. List, board, search, and children responses
	// deliberately omit the potentially large immutable snapshot.
	SourceContext *sourceContextDetailResponse `json:"source_context,omitempty"`
	// duplicateOfIssueID is the raw mark, kept off the wire; see DuplicateOf.
	duplicateOfIssueID pgtype.UUID
}

// IssueRefResponse names another issue inside a response: enough to render
// a link and its status without a second request.
type IssueRefResponse struct {
	ID         string `json:"id"`
	Identifier string `json:"identifier"`
	Title      string `json:"title"`
	Status     string `json:"status"`
}

// validIssuePriorities mirrors the CHECK constraint on the issue table. Write
// handlers pre-validate it so callers get a clean 400 with the allowed values
// instead of a database CHECK violation bubbling up as a 500.
var validIssuePriorities = []string{"urgent", "high", "medium", "low", "none"}

// validIssueStatuses contains the built-in status keys. Workspace writes,
// grouping, and filters also support custom keys from the workspace catalog.
var validIssueStatuses = issuestatus.Canonical()
var validIssueStatusCategories = issuestatus.Categories()

// Status sort ranks by the CONCRETE status key, in the catalog's own display
// order. Sorting is a display question, so it resolves on status; only behavior
// decisions aggregate on the four lifecycle categories (MUL-7379).
//
// Ranking by category instead collapses the seven built-ins into four buckets —
// Backlog ties with Todo, and In Progress / In Review / Blocked all tie — which
// leaves the created_at tiebreak deciding the visible order of a list the user
// explicitly asked to sort by status.
//
// issueTableStatusOrder is the same "board order" the status GROUPING path
// already builds, so sorting and grouping now share one source of truth instead
// of drifting apart. Archived statuses are included because issues stay on them
// after archival and still have to rank somewhere real.
//
// Keeping the indexed column bare avoids the per-row issue_effective_status()
// call that would otherwise turn a bounded page into a workspace scan.
func (h *Handler) issueStatusSortExpression(
	ctx context.Context,
	workspaceID pgtype.UUID,
	addArg func(any) string,
) (string, error) {
	entries, err := h.issueStatusCatalog().ListIssueStatusEntries(ctx, db.ListIssueStatusEntriesParams{
		WorkspaceID:     workspaceID,
		IncludeArchived: true,
	})
	if err != nil {
		return "", err
	}
	return fmt.Sprintf(
		"COALESCE(array_position(%s::text[], i.status), 100000)",
		addArg(issueTableStatusOrder(entries)),
	), nil
}

// resolveIssueStatusKey checks a status against the workspace's catalog and
// returns the CANONICAL key to store. This is the application-layer replacement
// for the enum CHECK that migration 337 dropped, so every write path must route
// through it — a missed entrypoint is how an unresolvable key would reach the
// column.
//
// Returning the resolved key (rather than a bare bool) is load-bearing:
// resolution is case- and whitespace-insensitive, so `"  HUMAN_REVIEW "` and
// `"human_review"` both validate. Writing the caller's raw string back would
// then store a value the column's format constraint rejects, turning an input
// the API just accepted into a 500. Callers must persist what this returns.
func (h *Handler) resolveIssueStatusKey(w http.ResponseWriter, r *http.Request, workspaceID pgtype.UUID, status string) (string, bool) {
	key, _, ok := h.resolveIssueStatusKeyKind(w, r, workspaceID, status)
	return key, ok
}

// resolveIssueStatusKeyKind is resolveIssueStatusKey plus whether the target is
// a CUSTOM status. Callers use that to decide whether the write needs the
// shared catalog lock — see runWithIssueStatusGuard.
func (h *Handler) resolveIssueStatusKeyKind(w http.ResponseWriter, r *http.Request, workspaceID pgtype.UUID, status string) (string, bool, bool) {
	entry, err := issuestatus.Resolve(r.Context(), h.Queries, workspaceID, status)
	if err != nil {
		if errors.Is(err, issuestatus.ErrUnknownStatus) {
			// Labels, not bare keys: a derived key says nothing about what the
			// status means, so listing `in_review_2` alone leaves the caller no
			// way to find the one they were told to use. (MUL-6749)
			allowed, listErr := issuestatus.ActiveKeyLabels(r.Context(), h.Queries, workspaceID)
			if listErr != nil || len(allowed) == 0 {
				allowed = issuestatus.Canonical()
			}
			writeError(w, http.StatusBadRequest, fmt.Sprintf(
				"invalid status %q; valid values: %s", status, strings.Join(allowed, ", ")))
			return "", false, false
		}
		slog.Warn("resolve issue status failed", append(logger.RequestAttrs(r), "error", err)...)
		writeError(w, http.StatusInternalServerError, "failed to validate status")
		return "", false, false
	}
	return entry.Key, !entry.IsSystem, true
}

// errIssueStatusArchivedRace signals that the target custom status was archived
// between the request's pre-flight validation and the write itself. Callers map
// it to 409: the request was valid when it arrived, and retrying against the
// refreshed catalog is the right remedy.
var errIssueStatusArchivedRace = errors.New("issue status was archived while the write was in flight")

// assertIssueStatusStillActive is the write half of the archive race guard. It
// takes the SHARED catalog lock and RE-RESOLVES the status inside the caller's
// transaction.
//
// The re-resolve is the part that actually closes the race. Handlers resolve a
// status up front, before any transaction, to answer a bad request with a clean
// 400 — but an archive can commit between that pre-flight check and the write.
// Re-checking here, under the lock, means the status is provably active at the
// moment the row is written. ArchiveIssueStatus holds the EXCLUSIVE side around
// retirement, so the two orderings are both covered:
//
//   - archive first: it commits, this re-resolve then fails and the write is
//     rejected, so no new assignment lands on an archived status;
//   - writer first: archive blocks until this transaction commits, then its
//     issue count rejects archival of the now-occupied status.
//
// A built-in status is a no-op: it can never be archived (enforced by
// issue_status_system_not_archivable), so the common path takes no lock and
// pays nothing. (MUL-6243)
func assertIssueStatusStillActive(ctx context.Context, qtx *db.Queries, workspaceID pgtype.UUID, statusKey string) error {
	if statusKey == "" || issuestatus.IsBuiltIn(statusKey) {
		return nil
	}
	// Catalog lock before any row lock, everywhere, so the two write paths
	// cannot deadlock against each other.
	if err := qtx.LockIssueStatusCatalogShared(ctx, workspaceID); err != nil {
		return err
	}
	if _, err := issuestatus.Resolve(ctx, qtx, workspaceID, statusKey); err != nil {
		if errors.Is(err, issuestatus.ErrUnknownStatus) {
			return errIssueStatusArchivedRace
		}
		return err
	}
	return nil
}

// runWithIssueStatusGuard runs an issue write that lands on a custom status
// inside a transaction that re-verifies the status under the shared catalog
// lock (see assertIssueStatusStillActive). Request writes also carry trusted
// wakeup actor identity in transaction-local settings, including built-in targets.
func (h *Handler) runWithIssueStatusGuard(ctx context.Context, workspaceID pgtype.UUID, statusKey string, fn func(q *db.Queries) error) error {
	_, hasActor := ctx.Value(wakeupActorKey{}).(wakeupActor)
	if !hasActor && (statusKey == "" || issuestatus.IsBuiltIn(statusKey)) {
		return fn(h.Queries)
	}
	tx, err := h.beginWakeupWrite(ctx)
	if err != nil {
		return err
	}
	defer tx.Rollback(ctx)

	qtx := h.Queries.WithTx(tx)
	if err := assertIssueStatusStillActive(ctx, qtx, workspaceID, statusKey); err != nil {
		return err
	}
	if err := fn(qtx); err != nil {
		return err
	}
	return tx.Commit(ctx)
}

// updateIssueWithStatusGuard writes params under the status-archive guard.
func (h *Handler) updateIssueWithStatusGuard(ctx context.Context, workspaceID pgtype.UUID, statusKey string, params db.UpdateIssueParams) (db.Issue, error) {
	var issue db.Issue
	var cancelledWakeups []db.AgentTaskQueue
	err := h.runWithIssueStatusGuard(ctx, workspaceID, statusKey, func(q *db.Queries) error {
		if params.DuplicateOfIssueID.Valid {
			if err := lockAndCheckDuplicateMark(ctx, q, workspaceID, params.ID, params.DuplicateOfIssueID); err != nil {
				return err
			}
		}
		var innerErr error
		issue, cancelledWakeups, innerErr = updateIssueStoppingWakeups(ctx, q, params)
		return innerErr
	})
	if err != nil {
		return issue, err
	}
	h.broadcastCancelledWakeups(ctx, workspaceID, cancelledWakeups)
	return issue, nil
}

// updateIssueStoppingWakeups writes params and, when the write moves the issue
// into a done/closed status, ends its wakeups with the same queries. Callers
// run it inside the status write's transaction and broadcast the returned runs
// after commit.
func updateIssueStoppingWakeups(ctx context.Context, q *db.Queries, params db.UpdateIssueParams) (db.Issue, []db.AgentTaskQueue, error) {
	issue, err := q.UpdateIssue(ctx, params)
	if err != nil || !params.Status.Valid {
		return issue, nil, err
	}
	cancelled, err := service.StopClosedIssueWakeups(ctx, q, issue)
	if err != nil {
		return db.Issue{}, nil, fmt.Errorf("stop closed issue wakeups: %w", err)
	}
	return issue, cancelled, nil
}

func (h *Handler) broadcastCancelledWakeups(ctx context.Context, workspaceID pgtype.UUID, cancelled []db.AgentTaskQueue) {
	if len(cancelled) > 0 {
		h.TaskService.BroadcastCancelledTasks(ctx, uuidToString(workspaceID), cancelled)
	}
}

// writeIssueStatusRaceError renders errIssueStatusArchivedRace as a 409 and
// reports whether it handled the error.
func writeIssueStatusRaceError(w http.ResponseWriter, err error) bool {
	if errors.Is(err, errIssueStatusArchivedRace) {
		writeError(w, http.StatusConflict,
			"the target status was archived while this request was in flight; reload the status list and retry")
		return true
	}
	return false
}

func validateIssueEnum(w http.ResponseWriter, field, value string, allowed []string) bool {
	for _, a := range allowed {
		if value == a {
			return true
		}
	}
	writeError(w, http.StatusBadRequest, fmt.Sprintf("invalid %s %q; valid values: %s", field, value, strings.Join(allowed, ", ")))
	return false
}

// fillStatusCategories resolves status_category for responses whose status is
// CUSTOM. The pure builders below map built-ins directly and leave custom
// statuses empty, so this is the step that makes the
// field authoritative on every payload a client caches or buckets by.
//
// Uses one Resolver for the whole slice: built-in statuses cost no query, and a
// page full of custom ones costs one catalog read rather than one per row. The
// Resolver includes ARCHIVED statuses, because an issue left on an archived
// status still belongs in its category's column. (MUL-6243)
func (h *Handler) fillStatusCategories(ctx context.Context, wsID pgtype.UUID, resps []IssueResponse) {
	var originals []pgtype.UUID
	for i := range resps {
		originals = appendDuplicateOriginal(originals, resps[i].Status, resps[i].duplicateOfIssueID)
	}
	fill := h.newStatusCategoryFiller(ctx, wsID, originals...)
	for i := range resps {
		fill(&resps[i])
	}
}

// newStatusCategoryFiller returns a request-scoped filler backed by ONE
// Resolver. Reuse it across every response a request builds: the Resolver reads
// the catalog at most once, so a page of custom-status rows costs one query
// rather than one per row. Creating a filler per row would reintroduce the N+1
// this exists to avoid. (MUL-6243)
//
// originals are the duplicate originals the caller's rows point at, collected
// with appendDuplicateOriginal. A page passes them so they resolve in one read
// up front, the way labelsByIssue loads a page's labels; an original that was
// not passed still resolves, one read per distinct original. (MUL-7349)
func (h *Handler) newStatusCategoryFiller(ctx context.Context, wsID pgtype.UUID, originals ...pgtype.UUID) func(*IssueResponse) {
	resolver := issuestatus.NewResolver(wsID)
	refs := duplicateOriginals{}
	h.loadDuplicateOriginals(ctx, wsID, refs, originals)
	return func(resp *IssueResponse) {
		if resp == nil {
			return
		}
		// Every response that resolves its status also resolves its duplicate
		// mark, so the two never disagree about which endpoints carry them.
		h.fillDuplicateOf(ctx, wsID, resp, refs)
		if resp.StatusCategory != "" {
			return
		}
		resp.StatusCategory = issuestatus.WireCategory(resp.Status, resolver.Category(ctx, h.Queries, resp.Status))
		// Same Resolver, same single catalog read, so the name rides along for
		// free. Built-ins return "" and stay omitted. (MUL-6749)
		resp.StatusName = resolver.Name(ctx, h.Queries, resp.Status)
	}
}

// fillStatusCategory is the single-response form. Only for endpoints that build
// exactly ONE response; anything looping must use newStatusCategoryFiller.
func (h *Handler) fillStatusCategory(ctx context.Context, wsID pgtype.UUID, resp *IssueResponse) {
	h.newStatusCategoryFiller(ctx, wsID)(resp)
}

// duplicateOfPointer keeps a duplicate mark only while the issue is
// cancelled. Whether the original still exists is checked when the response
// is filled (fillDuplicateOf), so a pointer an older server left behind
// never surfaces. Same rule as IssueHasDuplicates in issue.sql.
func duplicateOfPointer(status string, id pgtype.UUID) pgtype.UUID {
	if status != issuestatus.Cancelled {
		return pgtype.UUID{}
	}
	return id
}

// duplicateOriginals memoises the originals a request's duplicates point at,
// by id. A nil entry is an original that no longer exists (or could not be
// read), so it is looked up once and then left off every response.
type duplicateOriginals map[pgtype.UUID]*db.ListIssueRefsInWorkspaceRow

// appendDuplicateOriginal adds the original a row points at, if the row
// carries a live duplicate mark.
func appendDuplicateOriginal(originals []pgtype.UUID, status string, id pgtype.UUID) []pgtype.UUID {
	if pointer := duplicateOfPointer(status, id); pointer.Valid {
		return append(originals, pointer)
	}
	return originals
}

// loadDuplicateOriginals reads every id not yet in refs in one query.
func (h *Handler) loadDuplicateOriginals(ctx context.Context, wsID pgtype.UUID, refs duplicateOriginals, ids []pgtype.UUID) {
	var missing []pgtype.UUID
	for _, id := range ids {
		if _, seen := refs[id]; !seen {
			refs[id] = nil
			missing = append(missing, id)
		}
	}
	if len(missing) == 0 {
		return
	}
	rows, err := h.Queries.ListIssueRefsInWorkspace(ctx, db.ListIssueRefsInWorkspaceParams{
		WorkspaceID: wsID,
		Ids:         missing,
	})
	if err != nil {
		slog.Warn("resolve duplicate originals failed", "error", err, "count", len(missing))
		return
	}
	for i := range rows {
		refs[rows[i].ID] = &rows[i]
	}
}

// fillDuplicateOf resolves a duplicate mark to its original (MUL-7349),
// through refs so a page's originals cost one read (see
// newStatusCategoryFiller). A missing original leaves DuplicateOf unset.
func (h *Handler) fillDuplicateOf(ctx context.Context, wsID pgtype.UUID, resp *IssueResponse, refs duplicateOriginals) {
	id := resp.duplicateOfIssueID
	if !id.Valid {
		return
	}
	h.loadDuplicateOriginals(ctx, wsID, refs, []pgtype.UUID{id})
	row := refs[id]
	if row == nil {
		return
	}
	// The response already rendered its own identifier with the workspace
	// prefix; the original shares it.
	prefix := strings.TrimSuffix(resp.Identifier, "-"+strconv.Itoa(int(resp.Number)))
	resp.DuplicateOf = &IssueRefResponse{
		ID:         uuidToString(row.ID),
		Identifier: prefix + "-" + strconv.Itoa(int(row.Number)),
		Title:      row.Title,
		Status:     row.Status,
	}
}

func issueToResponse(i db.Issue, issuePrefix string) IssueResponse {
	identifier := issuePrefix + "-" + strconv.Itoa(int(i.Number))
	// Built-ins map to public categories without a catalog lookup. A custom
	// status is filled by endpoints that resolve the workspace catalog.
	statusCategory := ""
	if issuestatus.IsBuiltIn(i.Status) {
		statusCategory = i.Status
	}
	return IssueResponse{
		ID:                 uuidToString(i.ID),
		WorkspaceID:        uuidToString(i.WorkspaceID),
		Number:             i.Number,
		Identifier:         identifier,
		Title:              i.Title,
		Description:        textToPtr(i.Description),
		Status:             i.Status,
		StatusCategory:     statusCategory,
		Priority:           i.Priority,
		AssigneeType:       textToPtr(i.AssigneeType),
		AssigneeID:         uuidToPtr(i.AssigneeID),
		CreatorType:        i.CreatorType,
		CreatorID:          uuidToString(i.CreatorID),
		ParentIssueID:      uuidToPtr(i.ParentIssueID),
		duplicateOfIssueID: duplicateOfPointer(i.Status, i.DuplicateOfIssueID),
		ProjectID:          uuidToPtr(i.ProjectID),
		Position:           i.Position,
		Stage:              int4ToPtr(i.Stage),
		StartDate:          dateToPtr(i.StartDate),
		DueDate:            dateToPtr(i.DueDate),
		CreatedAt:          timestampToString(i.CreatedAt),
		UpdatedAt:          timestampToString(i.UpdatedAt),
		Revision:           i.Revision,
		LastActivityAt:     timestampToNanoPtr(i.LastActivityAt),
		Metadata:           parseIssueMetadata(i.Metadata),
		Properties:         parseIssueProperties(i.Properties),
	}
}

// issueListRowToResponse converts a list-query row (no description) to an IssueResponse.
func issueListRowToResponse(i db.ListIssuesRow, issuePrefix string) IssueResponse {
	// Same pure built-in resolution as issueToResponse. (MUL-6243)
	statusCategory := ""
	if issuestatus.IsBuiltIn(i.Status) {
		statusCategory = i.Status
	}
	identifier := issuePrefix + "-" + strconv.Itoa(int(i.Number))
	return IssueResponse{
		ID:                 uuidToString(i.ID),
		WorkspaceID:        uuidToString(i.WorkspaceID),
		Number:             i.Number,
		Identifier:         identifier,
		Title:              i.Title,
		Description:        textToPtr(i.Description),
		Status:             i.Status,
		StatusCategory:     statusCategory,
		Priority:           i.Priority,
		AssigneeType:       textToPtr(i.AssigneeType),
		AssigneeID:         uuidToPtr(i.AssigneeID),
		CreatorType:        i.CreatorType,
		CreatorID:          uuidToString(i.CreatorID),
		ParentIssueID:      uuidToPtr(i.ParentIssueID),
		duplicateOfIssueID: duplicateOfPointer(i.Status, i.DuplicateOfIssueID),
		ProjectID:          uuidToPtr(i.ProjectID),
		Position:           i.Position,
		Stage:              int4ToPtr(i.Stage),
		StartDate:          dateToPtr(i.StartDate),
		DueDate:            dateToPtr(i.DueDate),
		CreatedAt:          timestampToString(i.CreatedAt),
		UpdatedAt:          timestampToString(i.UpdatedAt),
		Revision:           i.Revision,
		LastActivityAt:     timestampToNanoPtr(i.LastActivityAt),
		Metadata:           parseIssueMetadata(i.Metadata),
		Properties:         parseIssueProperties(i.Properties),
	}
}

// labelsByIssue bulk-loads labels for the given issue IDs and returns a map
// keyed by issue UUID string. On error or empty input, returns an empty map —
// label rendering is non-critical and we'd rather serve issues without labels
// than fail the whole list call.
func (h *Handler) labelsByIssue(ctx context.Context, wsUUID pgtype.UUID, issueIDs []pgtype.UUID) map[string][]LabelResponse {
	out := map[string][]LabelResponse{}
	if len(issueIDs) == 0 {
		return out
	}
	rows, err := h.Queries.ListLabelsForIssues(ctx, db.ListLabelsForIssuesParams{
		IssueIds:    issueIDs,
		WorkspaceID: wsUUID,
	})
	if err != nil {
		slog.Warn("ListLabelsForIssues failed", "error", err)
		return out
	}
	for _, r := range rows {
		issueID := uuidToString(r.IssueID)
		out[issueID] = append(out[issueID], LabelResponse{
			ID:           uuidToString(r.ID),
			WorkspaceID:  uuidToString(r.WorkspaceID),
			ResourceType: r.ResourceType,
			Name:         r.Name,
			Description:  r.Description,
			Color:        r.Color,
			CreatedAt:    timestampToString(r.CreatedAt),
			UpdatedAt:    timestampToString(r.UpdatedAt),
		})
	}
	return out
}

func openIssueRowToResponse(i db.ListOpenIssuesRow, issuePrefix string) IssueResponse {
	// Same pure built-in resolution as issueToResponse. (MUL-6243)
	statusCategory := ""
	if issuestatus.IsBuiltIn(i.Status) {
		statusCategory = i.Status
	}
	identifier := issuePrefix + "-" + strconv.Itoa(int(i.Number))
	return IssueResponse{
		ID:                 uuidToString(i.ID),
		WorkspaceID:        uuidToString(i.WorkspaceID),
		Number:             i.Number,
		Identifier:         identifier,
		Title:              i.Title,
		Description:        textToPtr(i.Description),
		Status:             i.Status,
		StatusCategory:     statusCategory,
		Priority:           i.Priority,
		AssigneeType:       textToPtr(i.AssigneeType),
		AssigneeID:         uuidToPtr(i.AssigneeID),
		CreatorType:        i.CreatorType,
		CreatorID:          uuidToString(i.CreatorID),
		ParentIssueID:      uuidToPtr(i.ParentIssueID),
		duplicateOfIssueID: duplicateOfPointer(i.Status, i.DuplicateOfIssueID),
		ProjectID:          uuidToPtr(i.ProjectID),
		Position:           i.Position,
		Stage:              int4ToPtr(i.Stage),
		StartDate:          dateToPtr(i.StartDate),
		DueDate:            dateToPtr(i.DueDate),
		CreatedAt:          timestampToString(i.CreatedAt),
		UpdatedAt:          timestampToString(i.UpdatedAt),
		Revision:           i.Revision,
		LastActivityAt:     timestampToNanoPtr(i.LastActivityAt),
		Metadata:           parseIssueMetadata(i.Metadata),
		Properties:         parseIssueProperties(i.Properties),
	}
}

type IssueAssigneeGroupResponse struct {
	ID           string          `json:"id"`
	AssigneeType *string         `json:"assignee_type"`
	AssigneeID   *string         `json:"assignee_id"`
	Issues       []IssueResponse `json:"issues"`
	Total        int64           `json:"total"`
}

type GroupedIssuesResponse struct {
	Groups []IssueAssigneeGroupResponse `json:"groups"`
}

type groupedIssueRow struct {
	db.ListIssuesRow
	GroupTotal int64
}

func assigneeGroupID(assigneeType pgtype.Text, assigneeID pgtype.UUID) string {
	if assigneeType.Valid && assigneeID.Valid {
		return "assignee:" + assigneeType.String + ":" + uuidToString(assigneeID)
	}
	return "assignee:unassigned"
}

// SearchIssueResponse extends IssueResponse with search metadata.
type SearchIssueResponse struct {
	IssueResponse
	MatchSource               string  `json:"match_source"`
	MatchedSnippet            *string `json:"matched_snippet,omitempty"`
	MatchedDescriptionSnippet *string `json:"matched_description_snippet,omitempty"`
	MatchedCommentSnippet     *string `json:"matched_comment_snippet,omitempty"`
}

// extractSnippet extracts a snippet of text around the first occurrence of query.
// Returns up to ~120 runes centered on the match. Uses rune-based slicing to
// avoid splitting multi-byte UTF-8 characters (important for CJK content).
// For multi-word queries, tries phrase match first; if not found, locates the
// earliest occurring individual term and centers the snippet around it.
func extractSnippet(content, query string) string {
	runes := []rune(content)
	lowerRunes := []rune(strings.ToLower(content))
	queryRunes := []rune(strings.ToLower(query))

	idx := findRuneSubstring(lowerRunes, queryRunes)

	// If phrase not found, try individual terms for multi-word queries.
	matchLen := len(queryRunes)
	if idx < 0 {
		terms := strings.Fields(strings.ToLower(query))
		if len(terms) > 1 {
			earliest := -1
			earliestLen := 0
			for _, term := range terms {
				termRunes := []rune(term)
				pos := findRuneSubstring(lowerRunes, termRunes)
				if pos >= 0 && (earliest < 0 || pos < earliest) {
					earliest = pos
					earliestLen = len(termRunes)
				}
			}
			if earliest >= 0 {
				idx = earliest
				matchLen = earliestLen
			}
		}
	}

	if idx < 0 {
		if len(runes) > 120 {
			return string(runes[:120]) + "..."
		}
		return content
	}
	start := idx - 40
	if start < 0 {
		start = 0
	}
	end := idx + matchLen + 80
	if end > len(runes) {
		end = len(runes)
	}
	snippet := string(runes[start:end])
	if start > 0 {
		snippet = "..." + snippet
	}
	if end < len(runes) {
		snippet = snippet + "..."
	}
	return snippet
}

// findRuneSubstring returns the index of needle in haystack, or -1 if not found.
func findRuneSubstring(haystack, needle []rune) int {
	if len(needle) == 0 || len(haystack) < len(needle) {
		return -1
	}
	for i := 0; i <= len(haystack)-len(needle); i++ {
		match := true
		for j := range needle {
			if haystack[i+j] != needle[j] {
				match = false
				break
			}
		}
		if match {
			return i
		}
	}
	return -1
}

// descriptionContains checks if the description text contains the search phrase or all terms.
func descriptionContains(desc pgtype.Text, phrase string, terms []string) bool {
	if !desc.Valid || desc.String == "" {
		return false
	}
	lower := strings.ToLower(desc.String)
	if strings.Contains(lower, strings.ToLower(phrase)) {
		return true
	}
	if len(terms) > 1 {
		for _, t := range terms {
			if !strings.Contains(lower, strings.ToLower(t)) {
				return false
			}
		}
		return true
	}
	return false
}

// escapeLike escapes LIKE special characters (%, _, \) in user input.
func escapeLike(s string) string {
	s = strings.ReplaceAll(s, `\`, `\\`)
	s = strings.ReplaceAll(s, `%`, `\%`)
	s = strings.ReplaceAll(s, `_`, `\_`)
	return s
}

// splitSearchTerms splits a query into individual search terms, filtering empty strings.
func splitSearchTerms(q string) []string {
	fields := strings.FieldsFunc(q, func(r rune) bool {
		return unicode.IsSpace(r)
	})
	terms := make([]string, 0, len(fields))
	for _, f := range fields {
		if f != "" {
			terms = append(terms, f)
		}
	}
	return terms
}

// identifierNumberRe matches letter-leading prefixes like "MUL-123" or "V2-12".
var identifierNumberRe = regexp.MustCompile(`(?i)^[a-z][a-z0-9]*-(\d+)$`)

// parseQueryNumber extracts an issue number from the query if it looks like
// an identifier (e.g. "MUL-123" or "V2-12") or a bare number (e.g. "123").
// Numbers outside PostgreSQL int4 range remain ordinary text search terms.
func parseQueryNumber(q string) (int, bool) {
	q = strings.TrimSpace(q)
	// Check for identifier pattern like "MUL-123"
	if m := identifierNumberRe.FindStringSubmatch(q); m != nil {
		if n, err := strconv.ParseInt(m[1], 10, 32); err == nil && n > 0 {
			return int(n), true
		}
	}
	// Check for bare number
	if n, err := strconv.ParseInt(q, 10, 32); err == nil && n > 0 {
		return int(n), true
	}
	return 0, false
}

// searchResult holds a raw row from the dynamic search query.
type searchResult struct {
	issue                 db.Issue
	matchSource           string
	matchedCommentContent string
}

// buildSearchQuery builds a two-stage, workspace-scoped candidate pipeline for issue search.
// Search patterns are lowercased and escaped in Go so every flag uses the same
// case-insensitive LIKE semantics as the legacy query. The pipeline deliberately
// trades the title, description, and comment content GIN fast paths for one
// predictable pass over each relation within the selected workspace.
func buildSearchQuery(phrase string, terms []string, queryNum int, hasNum bool, includeClosed bool, terminalStatusKeys []string) (string, []any) {
	// Lowercase in Go so SQL only needs LOWER() on the column side.
	phrase = strings.ToLower(phrase)
	for i, term := range terms {
		terms[i] = strings.ToLower(term)
	}

	args := []any{}
	nextArg := func(value any) string {
		args = append(args, value)
		return fmt.Sprintf("$%d", len(args))
	}

	escapedPhrase := escapeLike(phrase)
	phraseParam := nextArg(escapedPhrase)                     // $1: exact title
	phraseContainsParam := nextArg("%" + escapedPhrase + "%") // $2: contains
	phraseStartsWithParam := nextArg(escapedPhrase + "%")     // $3: starts with
	wsParam := nextArg(nil)                                   // $4: workspace_id, filled by caller

	var termContainsParams []string
	if len(terms) > 1 {
		for _, term := range terms {
			termContainsParams = append(termContainsParams, nextArg("%"+escapeLike(term)+"%"))
		}
	}

	numParam := ""
	if hasNum {
		numParam = nextArg(queryNum)
	}

	terminalStatusesParam := ""
	if !includeClosed {
		// Negate only known terminal keys so an unknown legacy key remains
		// searchable instead of disappearing from the default result set.
		terminalStatusesParam = nextArg(terminalStatusKeys)
	}

	limitParam := nextArg(nil)
	offsetParam := nextArg(nil)

	// Stage one scans this workspace's issues once and retains only the narrow
	// flags and sort fields needed to choose a page. Do not force this CTE to be
	// MATERIALIZED: production EXPLAIN showed 28-68% lower execution time after
	// removing that fence. Full issue rows are hydrated after LIMIT/OFFSET below.
	const (
		loweredIssueTitle       = "lowered_issue_title.lowered"
		loweredIssueDescription = "lowered_issue_description.lowered"
	)
	titlePhrasePredicate := fmt.Sprintf("%s LIKE %s", loweredIssueTitle, phraseContainsParam)
	titleTermPredicates := make([]string, 0, len(termContainsParams))
	for _, termParam := range termContainsParams {
		titleTermPredicates = append(titleTermPredicates, fmt.Sprintf("%s LIKE %s", loweredIssueTitle, termParam))
	}
	titleMatchParts := []string{titlePhrasePredicate}
	if len(termContainsParams) > 1 {
		titleMatchParts = append(titleMatchParts, "("+strings.Join(titleTermPredicates, " AND ")+")")
	}
	titleMatchExpr := "(" + strings.Join(titleMatchParts, " OR ") + ")"

	issueFlagColumns := []string{
		"i.id AS issue_id",
		"i.status",
		"i.updated_at",
		fmt.Sprintf("%s = %s AS title_exact", loweredIssueTitle, phraseParam),
		fmt.Sprintf("%s LIKE %s AS title_starts_with", loweredIssueTitle, phraseStartsWithParam),
		fmt.Sprintf("%s AS title_phrase", titlePhrasePredicate),
		fmt.Sprintf("COALESCE(%s LIKE %s, FALSE) AS description_phrase", loweredIssueDescription, phraseContainsParam),
	}
	if hasNum {
		issueFlagColumns = append(issueFlagColumns, fmt.Sprintf("i.number = %s AS number_exact", numParam))
	}
	for index, termParam := range termContainsParams {
		issueFlagColumns = append(issueFlagColumns,
			fmt.Sprintf("%s AS title_term_%d", titleTermPredicates[index], index),
			fmt.Sprintf("COALESCE(%s LIKE %s, FALSE) AS description_term_%d", loweredIssueDescription, termParam, index),
		)
	}

	issueWhere := "i.workspace_id = " + wsParam
	if terminalStatusesParam != "" {
		issueWhere += fmt.Sprintf(" AND NOT (i.status = ANY(%s::text[]))", terminalStatusesParam)
	}
	// PostgreSQL otherwise inlines scalar LATERAL subqueries and recomputes the
	// LOWER expressions for every flag. The OFFSET 0 fences cache the normalized
	// title and description per issue row without materializing the whole CTE.
	// Once the title contains the phrase or every search term, it has already
	// satisfied eligibility and won both relevance-rank and match-source
	// precedence, so the LEFT JOIN can skip normalizing the description and its
	// flags safely fall back to FALSE. Correctness requires every title rank and
	// match-source branch below to remain ahead of its description counterpart.
	// Future indexable text predicates must stay outside these projection-only
	// fences so expression indexes can still match them.
	issueMatchesCTE := fmt.Sprintf(`issue_matches AS (
		SELECT %s
		FROM issue i
		CROSS JOIN LATERAL (
			SELECT LOWER(i.title) AS lowered
			OFFSET 0
		) lowered_issue_title
		LEFT JOIN LATERAL (
			SELECT LOWER(COALESCE(i.description, '')) AS lowered
			WHERE NOT %s
			OFFSET 0
		) lowered_issue_description ON TRUE
		WHERE %s
	)`, strings.Join(issueFlagColumns, ",\n\t\t\t"), titleMatchExpr, issueWhere)

	// Comments are also scanned once, workspace-first. This intentionally avoids
	// the legacy planner choice between global content GIN postings and repeated
	// correlated/hashed subplans (MUL-4059); idx_comment_workspace bounds the
	// candidate scan instead. Aggregation retains only per-issue flags plus the
	// latest matching comment ID, and content is fetched by primary key after the
	// final page is known. Per-term BOOL_OR flags keep the legacy eligibility rule
	// where terms may be spread across comments, while comment_all_terms keeps
	// ranking/snippet tied to one comment.
	const loweredCommentContent = "lowered_comment.lowered"
	commentFlagColumns := []string{
		"c.issue_id",
		fmt.Sprintf("BOOL_OR(%s LIKE %s) AS comment_phrase", loweredCommentContent, phraseContainsParam),
	}
	commentCandidateFlags := []string{"aggregated_comments.comment_phrase"}
	commentTerms := make([]string, 0, len(termContainsParams))
	for index, termParam := range termContainsParams {
		alias := fmt.Sprintf("comment_term_%d", index)
		commentFlagColumns = append(commentFlagColumns,
			fmt.Sprintf("BOOL_OR(%s LIKE %s) AS %s", loweredCommentContent, termParam, alias),
		)
		commentCandidateFlags = append(commentCandidateFlags, "aggregated_comments."+alias)
		commentTerms = append(commentTerms, fmt.Sprintf("%s LIKE %s", loweredCommentContent, termParam))
	}

	commentSnippetPredicate := fmt.Sprintf("%s LIKE %s", loweredCommentContent, phraseContainsParam)
	if len(commentTerms) > 1 {
		commentAllTerms := "(" + strings.Join(commentTerms, " AND ") + ")"
		commentFlagColumns = append(commentFlagColumns,
			fmt.Sprintf("BOOL_OR(%s) AS comment_all_terms", commentAllTerms),
		)
		commentSnippetPredicate += " OR " + commentAllTerms
	}
	// Keep the ordered aggregate in the measured single comment pass. Replacing
	// it with DISTINCT ON/window ranking changes that production-tested plan;
	// looking the ID up later would repeat text predicates after pagination.
	// The aggregate stores matching UUIDs per issue (not content), and the ID
	// tie-break makes equal created_at values deterministic.
	commentFlagColumns = append(commentFlagColumns, fmt.Sprintf(
		"(ARRAY_AGG(c.id ORDER BY c.created_at DESC, c.id DESC) FILTER (WHERE %s))[1] AS snippet_comment_id",
		commentSnippetPredicate,
	))

	// PostgreSQL otherwise inlines this scalar LATERAL subquery and recomputes
	// LOWER(c.content) for every flag. OFFSET 0 is the intentional planner fence
	// that keeps the long comment body lowercased once per row. Future indexable
	// text predicates must stay outside this projection-only fence so an index on
	// LOWER(c.content) can still match them.
	commentMatchesCTE := fmt.Sprintf(`comment_matches AS MATERIALIZED (
		SELECT *
		FROM (
			SELECT %s
			FROM comment c
			CROSS JOIN LATERAL (
				SELECT LOWER(c.content) AS lowered
				OFFSET 0
			) lowered_comment
			WHERE c.workspace_id = %s
			GROUP BY c.issue_id
		) aggregated_comments
		WHERE %s
	)`,
		strings.Join(commentFlagColumns, ",\n\t\t\t\t"),
		wsParam,
		strings.Join(commentCandidateFlags, " OR "),
	)

	// Stage two combines the two narrow sources, applies the legacy eligibility
	// and ranking rules once, and materializes only the requested page.
	eligibleParts := []string{
		"im.title_phrase",
		"im.description_phrase",
		"COALESCE(cm.comment_phrase, FALSE)",
	}
	if len(termContainsParams) > 1 {
		var allTerms []string
		for index := range termContainsParams {
			allTerms = append(allTerms, fmt.Sprintf(
				"(im.title_term_%[1]d OR im.description_term_%[1]d OR COALESCE(cm.comment_term_%[1]d, FALSE))",
				index,
			))
		}
		eligibleParts = append(eligibleParts, "("+strings.Join(allTerms, " AND ")+")")
	}
	if hasNum {
		eligibleParts = append(eligibleParts, "im.number_exact")
	}
	eligibleExpr := "(" + strings.Join(eligibleParts, " OR ") + ")"

	rankCases := []string{}
	if hasNum {
		rankCases = append(rankCases, "WHEN im.number_exact THEN 0")
	}
	rankCases = append(rankCases,
		"WHEN im.title_exact THEN 1",
		"WHEN im.title_starts_with THEN 2",
		"WHEN im.title_phrase THEN 3",
	)
	if len(termContainsParams) > 1 {
		var titleTerms []string
		for index := range termContainsParams {
			titleTerms = append(titleTerms, fmt.Sprintf("im.title_term_%d", index))
		}
		rankCases = append(rankCases, "WHEN ("+strings.Join(titleTerms, " AND ")+") THEN 4")
	}
	rankCases = append(rankCases, "WHEN im.description_phrase THEN 5")
	if len(termContainsParams) > 1 {
		var descriptionTerms []string
		for index := range termContainsParams {
			descriptionTerms = append(descriptionTerms, fmt.Sprintf("im.description_term_%d", index))
		}
		rankCases = append(rankCases, "WHEN ("+strings.Join(descriptionTerms, " AND ")+") THEN 6")
	}
	rankCases = append(rankCases, "WHEN COALESCE(cm.comment_phrase, FALSE) THEN 7")
	if len(termContainsParams) > 1 {
		rankCases = append(rankCases, "WHEN COALESCE(cm.comment_all_terms, FALSE) THEN 8")
	}
	rankExpr := "CASE " + strings.Join(rankCases, " ") + " ELSE 9 END"

	// title_exact deliberately preserves the legacy escapeLike quirk: a title
	// containing _, %, or \\ is still searchable, but the escaped phrase does
	// not compare equal and therefore is not treated as a cancelled direct hit.
	directHitParts := []string{"im.title_exact"}
	if hasNum {
		directHitParts = append(directHitParts, "im.number_exact")
	}
	// Cancelled issues sort behind every live match unless an exact title or
	// identifier shows that the user is targeting that specific issue.
	cancelledRank := fmt.Sprintf(
		"CASE WHEN im.status = 'cancelled' AND NOT (%s) THEN 1 ELSE 0 END",
		strings.Join(directHitParts, " OR "),
	)
	statusRank := `CASE im.status
		WHEN 'in_progress' THEN 0
		WHEN 'in_review' THEN 1
		WHEN 'todo' THEN 2
		WHEN 'blocked' THEN 3
		WHEN 'backlog' THEN 4
		WHEN 'done' THEN 5
		WHEN 'cancelled' THEN 6
		ELSE 7
	END`

	matchSourceParts := []string{"WHEN im.title_phrase THEN 'title'"}
	if len(termContainsParams) > 1 {
		var titleTerms []string
		for index := range termContainsParams {
			titleTerms = append(titleTerms, fmt.Sprintf("im.title_term_%d", index))
		}
		matchSourceParts = append(matchSourceParts, "WHEN ("+strings.Join(titleTerms, " AND ")+") THEN 'title'")
	}
	matchSourceParts = append(matchSourceParts, "WHEN im.description_phrase THEN 'description'")
	if len(termContainsParams) > 1 {
		var descriptionTerms []string
		for index := range termContainsParams {
			descriptionTerms = append(descriptionTerms, fmt.Sprintf("im.description_term_%d", index))
		}
		matchSourceParts = append(matchSourceParts, "WHEN ("+strings.Join(descriptionTerms, " AND ")+") THEN 'description'")
	}
	matchSourceExpr := "CASE " + strings.Join(matchSourceParts, " ") + " ELSE 'comment' END"

	rankedCandidatesCTE := fmt.Sprintf(`ranked_candidates AS (
		SELECT im.issue_id, im.updated_at, cm.snippet_comment_id,
			%s AS cancelled_rank,
			%s AS relevance_rank,
			%s AS status_rank,
			%s AS match_source
		FROM issue_matches im
		LEFT JOIN comment_matches cm ON cm.issue_id = im.issue_id
		WHERE %s
	)`, cancelledRank, rankExpr, statusRank, matchSourceExpr, eligibleExpr)

	pageCandidatesCTE := fmt.Sprintf(`page_candidates AS MATERIALIZED (
		SELECT issue_id, updated_at, snippet_comment_id, cancelled_rank, relevance_rank, status_rank, match_source
		FROM ranked_candidates
		ORDER BY cancelled_rank, relevance_rank, status_rank, updated_at DESC, issue_id ASC
		LIMIT %s OFFSET %s
	)`, limitParam, offsetParam)

	query := fmt.Sprintf(`WITH %s,
	%s,
	%s,
	%s
	SELECT i.id, i.workspace_id, i.title, i.description, i.status, i.priority,
		i.assignee_type, i.assignee_id, i.creator_type, i.creator_id,
		i.parent_issue_id, i.acceptance_criteria, i.context_refs, i.position,
		i.start_date, i.due_date, i.created_at, i.updated_at, i.last_activity_at, i.number, i.project_id,
		i.revision, i.duplicate_of_issue_id,
		pc.match_source,
		COALESCE(c.content, '') AS matched_comment_content
	FROM page_candidates pc
	JOIN issue i ON i.id = pc.issue_id AND i.workspace_id = %s
	LEFT JOIN comment c ON c.id = pc.snippet_comment_id AND c.workspace_id = %s
	ORDER BY pc.cancelled_rank, pc.relevance_rank, pc.status_rank, pc.updated_at DESC, pc.issue_id ASC`,
		issueMatchesCTE,
		commentMatchesCTE,
		rankedCandidatesCTE,
		pageCandidatesCTE,
		wsParam,
		wsParam,
	)

	return query, args
}

func (h *Handler) SearchIssues(w http.ResponseWriter, r *http.Request) {
	ctx := r.Context()
	workspaceID := h.resolveWorkspaceID(r)

	q := r.URL.Query().Get("q")
	if q == "" {
		writeError(w, http.StatusBadRequest, "q parameter is required")
		return
	}

	limit := 20
	offset := 0
	if l := r.URL.Query().Get("limit"); l != "" {
		if v, err := strconv.Atoi(l); err == nil && v > 0 {
			limit = v
		}
	}
	if limit > 50 {
		limit = 50
	}
	if o := r.URL.Query().Get("offset"); o != "" {
		if v, err := strconv.Atoi(o); err == nil && v >= 0 {
			offset = v
		}
	}

	includeClosed := r.URL.Query().Get("include_closed") == "true"

	wsUUID, ok := parseUUIDOrBadRequest(w, workspaceID, "workspace_id")
	if !ok {
		return
	}
	terms := splitSearchTerms(q)
	queryNum, hasNum := parseQueryNumber(q)
	var terminalStatusKeys []string
	if !includeClosed {
		resolvedKeys, err := h.terminalIssueStatusKeys(ctx, wsUUID)
		if err != nil {
			if writeSearchCancellation(w, r, err) {
				return
			}
			slog.Warn("expand terminal status categories failed", append(logger.RequestAttrs(r), "error", err)...)
			writeError(w, http.StatusInternalServerError, "failed to resolve status categories")
			return
		}
		terminalStatusKeys = resolvedKeys
	}

	sqlQuery, args := buildSearchQuery(q, terms, queryNum, hasNum, includeClosed, terminalStatusKeys)
	// Fill placeholder args: $4 = workspace_id, last two = limit, offset
	args[3] = wsUUID
	args[len(args)-2] = limit
	args[len(args)-1] = offset

	var results []searchResult
	err := runSearchQuery(ctx, h.TxStarter, sqlQuery, args, func(rows pgx.Rows) error {
		for rows.Next() {
			var sr searchResult
			if err := rows.Scan(
				&sr.issue.ID,
				&sr.issue.WorkspaceID,
				&sr.issue.Title,
				&sr.issue.Description,
				&sr.issue.Status,
				&sr.issue.Priority,
				&sr.issue.AssigneeType,
				&sr.issue.AssigneeID,
				&sr.issue.CreatorType,
				&sr.issue.CreatorID,
				&sr.issue.ParentIssueID,
				&sr.issue.AcceptanceCriteria,
				&sr.issue.ContextRefs,
				&sr.issue.Position,
				&sr.issue.StartDate,
				&sr.issue.DueDate,
				&sr.issue.CreatedAt,
				&sr.issue.UpdatedAt,
				&sr.issue.LastActivityAt,
				&sr.issue.Number,
				&sr.issue.ProjectID,
				&sr.issue.Revision,
				&sr.issue.DuplicateOfIssueID,
				&sr.matchSource,
				&sr.matchedCommentContent,
			); err != nil {
				return fmt.Errorf("scan: %w", err)
			}
			results = append(results, sr)
		}
		return rows.Err()
	})
	if err != nil {
		if writeSearchCancellation(w, r, err) {
			return
		}
		// Statement-timeout surfaces as SQLSTATE 57014. Return a 503
		// so the frontend can distinguish a timeout ("try a more
		// specific query") from a generic 500. This is the fail-fast
		// path when GIN search indexes are absent or the database is
		// overloaded; see runSearchQuery header for context.
		if isSearchStatementTimeout(err) {
			slog.Warn("search issues timed out",
				"workspace_id", workspaceID,
				"query", q,
				"timeout", searchStatementTimeout)
			writeError(w, http.StatusServiceUnavailable, "search timed out; please refine your query or try again")
			return
		}
		slog.Warn("search issues failed", "error", err, "workspace_id", workspaceID, "query", q)
		writeError(w, http.StatusInternalServerError, "failed to search issues")
		return
	}
	if writeSearchCancellation(w, r, ctx.Err()) {
		return
	}

	prefix := h.getIssuePrefix(ctx, wsUUID)
	var originals []pgtype.UUID
	for _, sr := range results {
		originals = appendDuplicateOriginal(originals, sr.issue.Status, sr.issue.DuplicateOfIssueID)
	}
	fillSearch := h.newStatusCategoryFiller(ctx, wsUUID, originals...)
	resp := make([]SearchIssueResponse, len(results))
	for i, sr := range results {
		sir := SearchIssueResponse{
			IssueResponse: issueToResponse(sr.issue, prefix),
			MatchSource:   sr.matchSource,
		}
		fillSearch(&sir.IssueResponse)
		// Always populate comment snippet when a matching comment exists
		if sr.matchedCommentContent != "" {
			snippet := extractSnippet(sr.matchedCommentContent, q)
			sir.MatchedCommentSnippet = &snippet
			// Keep backward compat: also set MatchedSnippet for comment-source matches
			if sr.matchSource == "comment" {
				sir.MatchedSnippet = &snippet
			}
		}
		// Populate description snippet when description matches
		if sr.matchSource == "description" || descriptionContains(sr.issue.Description, q, terms) {
			if sr.issue.Description.Valid && sr.issue.Description.String != "" {
				snippet := extractSnippet(sr.issue.Description.String, q)
				sir.MatchedDescriptionSnippet = &snippet
			}
		}
		resp[i] = sir
	}

	if writeSearchCancellation(w, r, ctx.Err()) {
		return
	}
	writeJSON(w, http.StatusOK, map[string]any{
		"issues": resp,
	})
}

// QueryIssues is the POST twin of ListIssues for filter sets too large for a
// GET request line — the table's agents-working facet can carry hundreds of
// issue ids, and common reverse proxies cap request lines around 8 KB. The
// body is a flat JSON object with EXACTLY the same keys and string encodings
// as ListIssues' query parameters; the handler rebuilds the query string and
// delegates, so the two transports cannot drift.
func (h *Handler) QueryIssues(w http.ResponseWriter, r *http.Request) {
	var params map[string]string
	if err := json.NewDecoder(io.LimitReader(r.Body, 1<<20)).Decode(&params); err != nil {
		writeError(w, http.StatusBadRequest, "invalid request body")
		return
	}
	values := make(url.Values, len(params))
	for key, value := range params {
		values.Set(key, value)
	}
	r.URL.RawQuery = values.Encode()
	h.ListIssues(w, r)
}

func (h *Handler) ListIssues(w http.ResponseWriter, r *http.Request) {
	ctx := r.Context()

	workspaceID := h.resolveWorkspaceID(r)
	wsUUID, ok := parseUUIDOrBadRequest(w, workspaceID, "workspace_id")
	if !ok {
		return
	}

	// Parse optional filter params. Malformed UUIDs in filters return 400 —
	// silently coercing them to a zero UUID would mask a client bug and let
	// the query return an empty result set (or worse, match a NULL row).
	var priorityFilter pgtype.Text
	if p := r.URL.Query().Get("priority"); p != "" {
		priorityFilter = pgtype.Text{String: p, Valid: true}
	}
	var assigneeFilter pgtype.UUID
	if a := r.URL.Query().Get("assignee_id"); a != "" {
		id, ok := parseUUIDOrBadRequest(w, a, "assignee_id")
		if !ok {
			return
		}
		assigneeFilter = id
	}
	var assigneeIdsFilter []pgtype.UUID
	if ids := r.URL.Query().Get("assignee_ids"); ids != "" {
		for _, raw := range strings.Split(ids, ",") {
			if s := strings.TrimSpace(raw); s != "" {
				id, ok := parseUUIDOrBadRequest(w, s, "assignee_ids")
				if !ok {
					return
				}
				assigneeIdsFilter = append(assigneeIdsFilter, id)
			}
		}
	}
	var creatorFilter pgtype.UUID
	if c := r.URL.Query().Get("creator_id"); c != "" {
		id, ok := parseUUIDOrBadRequest(w, c, "creator_id")
		if !ok {
			return
		}
		creatorFilter = id
	}
	var projectFilter pgtype.UUID
	if p := r.URL.Query().Get("project_id"); p != "" {
		id, ok := parseUUIDOrBadRequest(w, p, "project_id")
		if !ok {
			return
		}
		projectFilter = id
	}
	// involves_user_id widens the assignee filter to surface issues where the
	// user is the indirect assignee (their owned agent, or a squad they belong
	// to / lead / have an agent inside). Direct member-assignment is excluded
	// by design — that is the meaning of `assignee_id` (tab 1), and tab 3 must
	// be disjoint from tab 1.
	var involvesUserFilter pgtype.UUID
	if u := r.URL.Query().Get("involves_user_id"); u != "" {
		id, ok := parseUUIDOrBadRequest(w, u, "involves_user_id")
		if !ok {
			return
		}
		involvesUserFilter = id
	}

	metadataFilter, ok := parseMetadataFilterParam(w, r.URL.Query().Get("metadata"))
	if !ok {
		return
	}
	propertiesFilter, ok := parsePropertiesFilterParam(w, r.URL.Query().Get("properties"))
	if !ok {
		return
	}
	dateFilter, ok := parseIssueDateFilter(w, r.URL.Query())
	if !ok {
		return
	}

	// open_only=true returns all non-done/cancelled issues (no limit).
	if r.URL.Query().Get("open_only") == "true" {
		// Serialize the parsed AND-of-ORs groups into the single jsonb param
		// the static query unrolls (see properties_filter in ListOpenIssues).
		var openPropertiesFilter []byte
		if len(propertiesFilter) > 0 {
			marshaled, marshalErr := json.Marshal(propertiesFilter)
			if marshalErr != nil {
				writeError(w, http.StatusInternalServerError, "failed to list issues")
				return
			}
			openPropertiesFilter = marshaled
		}
		terminalStatusKeys, err := h.terminalIssueStatusKeys(ctx, wsUUID)
		if err != nil {
			writeError(w, http.StatusInternalServerError, "failed to resolve status categories")
			return
		}
		issues, err := h.Queries.ListOpenIssues(ctx, db.ListOpenIssuesParams{
			WorkspaceID:        wsUUID,
			TerminalStatusKeys: terminalStatusKeys,
			Priority:           priorityFilter,
			AssigneeID:         assigneeFilter,
			AssigneeIds:        assigneeIdsFilter,
			CreatorID:          creatorFilter,
			ProjectID:          projectFilter,
			InvolvesUserID:     involvesUserFilter,
			MetadataFilter:     metadataFilter,
			PropertiesFilter:   openPropertiesFilter,
		})
		if err != nil {
			writeError(w, http.StatusInternalServerError, "failed to list issues")
			return
		}

		prefix := h.getIssuePrefix(ctx, wsUUID)
		ids := make([]pgtype.UUID, len(issues))
		var originals []pgtype.UUID
		for i, issue := range issues {
			ids[i] = issue.ID
			originals = appendDuplicateOriginal(originals, issue.Status, issue.DuplicateOfIssueID)
		}
		labelsMap := h.labelsByIssue(ctx, wsUUID, ids)
		fillOpen := h.newStatusCategoryFiller(ctx, wsUUID, originals...)
		resp := make([]IssueResponse, len(issues))
		for i, issue := range issues {
			resp[i] = openIssueRowToResponse(issue, prefix)
			fillOpen(&resp[i])
			labels := labelsMap[resp[i].ID]
			if labels == nil {
				labels = []LabelResponse{}
			}
			resp[i].Labels = &labels
		}

		writeJSON(w, http.StatusOK, map[string]any{
			"issues": resp,
			"total":  len(resp),
		})
		return
	}

	limit := 100
	offset := 0
	if l := r.URL.Query().Get("limit"); l != "" {
		if v, err := strconv.Atoi(l); err == nil && v > 0 {
			limit = v
		}
	}
	if limit > 100 {
		limit = 100
	}
	if o := r.URL.Query().Get("offset"); o != "" {
		if v, err := strconv.Atoi(o); err == nil && v >= 0 {
			offset = v
		}
	}

	statusesFilter := splitCommaParam(r.URL.Query().Get("statuses"))
	if len(statusesFilter) == 0 {
		statusesFilter = splitCommaParam(r.URL.Query().Get("status"))
	}
	// status_category filters by lifecycle phase rather than by exact key, so
	// one board column can hold all concrete and custom statuses in that phase.
	// Without this the board would need one column — and one request — per
	// status. (MUL-6243, MUL-7240)
	statusCategoriesFilter := splitCommaParam(r.URL.Query().Get("status_categories"))
	if len(statusCategoriesFilter) == 0 {
		statusCategoriesFilter = splitCommaParam(r.URL.Query().Get("status_category"))
	}
	prioritiesFilter := splitCommaParam(r.URL.Query().Get("priorities"))
	if len(prioritiesFilter) == 0 {
		prioritiesFilter = splitCommaParam(r.URL.Query().Get("priority"))
	}

	// assignee_types narrows the list to issues assigned to the given actor
	// kinds (member / agent / squad). Mirrors the same param on
	// ListGroupedIssues so the workspace Members/Agents tabs can filter
	// server-side instead of post-filtering loaded pages on the client.
	assigneeTypesFilter := splitCommaParam(r.URL.Query().Get("assignee_types"))
	for _, assigneeType := range assigneeTypesFilter {
		if !isIssueActorType(assigneeType) {
			writeError(w, http.StatusBadRequest, "invalid assignee_types")
			return
		}
	}

	// scheduled=true restricts the result to issues that have at least one of
	// start_date / due_date set. Used by the Project Gantt view, which only
	// renders schedulable rows and shouldn't pay for the full project list.
	var scheduledFilter pgtype.Bool
	if r.URL.Query().Get("scheduled") == "true" {
		scheduledFilter = pgtype.Bool{Bool: true, Valid: true}
	}

	// Parse sort and direction params for dynamic ORDER BY.
	// Manual sort (position) is always ASC — direction is ignored because
	// the user defines order through drag-and-drop, reversing it has no
	// product meaning.
	sortCol := "position"
	sortIsExpr := false
	sortIsProperty := false
	sortByStatus := false
	if s := r.URL.Query().Get("sort"); s != "" {
		switch s {
		case "position", "title", "created_at", "updated_at", "start_date", "due_date":
			sortCol = s
		case "last_activity":
			sortCol = "last_activity_at"
		case "status":
			sortCol = "status"
			sortIsExpr = true
			sortByStatus = true
		case "priority":
			sortCol = "CASE i.priority WHEN 'urgent' THEN 0 WHEN 'high' THEN 1 WHEN 'medium' THEN 2 WHEN 'low' THEN 3 ELSE 4 END"
			sortIsExpr = true
		default:
			// property:<definitionId> sorts by the custom-property value
			// (typed expression); unknown/archived definitions degrade to
			// position order instead of erroring stale clients.
			expr, handled, sortErr := h.propertySortExpr(r, workspaceID, s)
			if !handled {
				writeError(w, http.StatusBadRequest, "invalid sort value")
				return
			}
			if sortErr != nil {
				if sortErr.Error() == "invalid sort value" || sortErr.Error() == "invalid workspace id" {
					writeError(w, http.StatusBadRequest, sortErr.Error())
					return
				}
				slog.Warn("propertySortExpr failed", append(logger.RequestAttrs(r), "error", sortErr)...)
				writeError(w, http.StatusInternalServerError, "failed to resolve sort")
				return
			}
			if expr != "" {
				sortCol = expr
				sortIsExpr = true
				sortIsProperty = true
			}
		}
	}
	sortDir := "ASC"
	if sortCol == "last_activity_at" {
		sortDir = "DESC"
	}
	if sortCol != "position" {
		if d := r.URL.Query().Get("direction"); d != "" {
			switch strings.ToLower(d) {
			case "asc":
				sortDir = "ASC"
			case "desc":
				sortDir = "DESC"
			default:
				writeError(w, http.StatusBadRequest, "invalid direction value")
				return
			}
		}
	}

	// Build dynamic SQL — same approach as ListGroupedIssues.
	where := []string{"i.workspace_id = $1"}
	args := []any{wsUUID}
	addArg := func(v any) string {
		args = append(args, v)
		return "$" + strconv.Itoa(len(args))
	}
	if len(statusCategoriesFilter) > 0 {
		// Expanded to concrete status keys rather than filtered through
		// issue_effective_status(): wrapping the column in a function makes the
		// (workspace_id, status) index unusable and turns a two-page index read
		// into a full workspace scan. (MUL-6243)
		keys, err := issuestatus.ExpandCategories(r.Context(), h.Queries, wsUUID, statusCategoriesFilter)
		if err != nil {
			slog.Warn("expand status categories failed", append(logger.RequestAttrs(r), "error", err)...)
			writeError(w, http.StatusInternalServerError, "failed to resolve status categories")
			return
		}
		where = append(where, fmt.Sprintf("i.status = ANY(%s::text[])", addArg(keys)))
	}
	if len(statusesFilter) > 0 {
		where = append(where, fmt.Sprintf("i.status = ANY(%s::text[])", addArg(statusesFilter)))
	}
	if len(prioritiesFilter) > 0 {
		where = append(where, fmt.Sprintf("i.priority = ANY(%s::text[])", addArg(prioritiesFilter)))
	}
	if assigneeFilter.Valid {
		where = append(where, fmt.Sprintf("i.assignee_id = %s::uuid", addArg(assigneeFilter)))
	}
	if len(assigneeIdsFilter) > 0 {
		where = append(where, fmt.Sprintf("i.assignee_id = ANY(%s::uuid[])", addArg(assigneeIdsFilter)))
	}
	if len(assigneeTypesFilter) > 0 {
		where = append(where, fmt.Sprintf("i.assignee_type = ANY(%s::text[])", addArg(assigneeTypesFilter)))
	}
	if creatorFilter.Valid {
		where = append(where, fmt.Sprintf("i.creator_id = %s::uuid", addArg(creatorFilter)))
	}
	if projectFilter.Valid {
		where = append(where, fmt.Sprintf("i.project_id = %s::uuid", addArg(projectFilter)))
	}

	// Table facets must be part of the server window. Applying them after
	// LIMIT/OFFSET hides matches that live on later pages and makes `total`
	// disagree with the rows the user sees/exports.
	assigneeFilters, ok := parseActorFilterList(w, r.URL.Query().Get("assignee_filters"), "assignee_filters")
	if !ok {
		return
	}
	includeNoAssignee := r.URL.Query().Get("include_no_assignee") == "true"
	if len(assigneeFilters) > 0 || includeNoAssignee {
		ors := make([]string, 0, len(assigneeFilters)+1)
		for _, filter := range assigneeFilters {
			ors = append(ors, fmt.Sprintf(
				"(i.assignee_type = %s::text AND i.assignee_id = %s::uuid)",
				addArg(filter.actorType),
				addArg(filter.actorID),
			))
		}
		if includeNoAssignee {
			ors = append(ors, "(i.assignee_type IS NULL AND i.assignee_id IS NULL)")
		}
		where = append(where, "("+strings.Join(ors, " OR ")+")")
	}

	creatorFilters, ok := parseActorFilterList(w, r.URL.Query().Get("creator_filters"), "creator_filters")
	if !ok {
		return
	}
	if len(creatorFilters) > 0 {
		ors := make([]string, 0, len(creatorFilters))
		for _, filter := range creatorFilters {
			ors = append(ors, fmt.Sprintf(
				"(i.creator_type = %s::text AND i.creator_id = %s::uuid)",
				addArg(filter.actorType),
				addArg(filter.actorID),
			))
		}
		where = append(where, "("+strings.Join(ors, " OR ")+")")
	}

	projectIDs, ok := parseUUIDParamList(w, r.URL.Query().Get("project_ids"), "project_ids")
	if !ok {
		return
	}
	includeNoProject := r.URL.Query().Get("include_no_project") == "true"
	if len(projectIDs) > 0 || includeNoProject {
		ors := make([]string, 0, 2)
		if len(projectIDs) > 0 {
			ors = append(ors, fmt.Sprintf("i.project_id = ANY(%s::uuid[])", addArg(projectIDs)))
		}
		if includeNoProject {
			ors = append(ors, "i.project_id IS NULL")
		}
		where = append(where, "("+strings.Join(ors, " OR ")+")")
	}

	labelIDs, ok := parseUUIDParamList(w, r.URL.Query().Get("label_ids"), "label_ids")
	if !ok {
		return
	}
	if len(labelIDs) > 0 {
		where = append(where, fmt.Sprintf(
			"EXISTS (SELECT 1 FROM issue_to_label itl WHERE itl.issue_id = i.id AND itl.label_id = ANY(%s::uuid[]))",
			addArg(labelIDs),
		))
	}
	// ids restricts the window to an explicit id set (the table's
	// agents-working facet sends the live running-issue ids). Presence with an
	// EMPTY list is meaningful — it must yield an empty window, not degrade to
	// the unrestricted one, so gate on Has() rather than the parsed length.
	if r.URL.Query().Has("ids") {
		idsFilter, ok := parseUUIDParamList(w, r.URL.Query().Get("ids"), "ids")
		if !ok {
			return
		}
		if idsFilter == nil {
			idsFilter = []pgtype.UUID{}
		}
		where = append(where, fmt.Sprintf("i.id = ANY(%s::uuid[])", addArg(idsFilter)))
	}
	if r.URL.Query().Get("top_level_only") == "true" {
		where = append(where, "i.parent_issue_id IS NULL")
	}
	where = appendIssueTableSearchFilter(where, addArg, r.URL.Query().Get("q"))
	if scheduledFilter.Valid {
		where = append(where, "(i.start_date IS NOT NULL OR i.due_date IS NOT NULL)")
	}
	if metadataFilter != nil {
		where = append(where, fmt.Sprintf("i.metadata @> %s::jsonb", addArg(string(metadataFilter))))
	}
	if propertiesFilter != nil {
		where = append(where, propertiesFilterPredicate(propertiesFilter, addArg))
	}
	where = appendIssueDateFilter(where, addArg, dateFilter)
	if involvesUserFilter.Valid {
		ref := addArg(involvesUserFilter)
		where = append(where, fmt.Sprintf(`(
    (i.assignee_type = 'agent' AND i.assignee_id IN (
       SELECT a.id FROM agent a
        WHERE a.workspace_id = $1
          AND a.owner_id     = %[1]s::uuid
    ))
    OR (i.assignee_type = 'squad' AND i.assignee_id IN (
       SELECT sm.squad_id
         FROM squad_member sm
         JOIN squad s ON s.id = sm.squad_id
        WHERE s.workspace_id = $1
          AND sm.member_type = 'member'
          AND sm.member_id   = %[1]s::uuid
       UNION
       SELECT s.id
         FROM squad s
         JOIN agent a ON a.id = s.leader_id
        WHERE s.workspace_id = $1
          AND a.workspace_id = $1
          AND a.owner_id     = %[1]s::uuid
       UNION
       SELECT sm.squad_id
         FROM squad_member sm
         JOIN squad s ON s.id = sm.squad_id
         JOIN agent a ON a.id = sm.member_id
        WHERE s.workspace_id = $1
          AND sm.member_type = 'agent'
          AND a.workspace_id = $1
          AND a.owner_id     = %[1]s::uuid
    ))
)`, ref))
	}

	whereSql := strings.Join(where, " AND ")

	// Build ORDER BY clause.
	// Count queries use only filter parameters. Sort-only CASE parameters must
	// be appended afterwards or COUNT receives unused/untyped bind positions.
	filterArgCount := len(args)
	if sortByStatus {
		var err error
		sortCol, err = h.issueStatusSortExpression(r.Context(), wsUUID, addArg)
		if err != nil {
			slog.Warn("resolve status sort failed", append(logger.RequestAttrs(r), "error", err)...)
			writeError(w, http.StatusInternalServerError, "failed to resolve sort")
			return
		}
	}
	orderBy := sortCol
	if !sortIsExpr {
		orderBy = "i." + sortCol
	}
	orderBy += " " + sortDir
	if sortCol == "start_date" || sortCol == "due_date" || sortCol == "last_activity_at" || sortIsProperty {
		// Property values are sparse: issues without one sort last in both
		// directions (mirrors the client comparator).
		orderBy += " NULLS LAST"
	}
	// created_at alone is not unique (bulk imports share timestamps); without
	// a unique final key the database may reorder ties between two
	// LIMIT/OFFSET requests, duplicating or dropping rows at page boundaries.
	if sortCol == "last_activity_at" {
		orderBy += ", i.id DESC"
	} else {
		orderBy += ", i.created_at DESC, i.id DESC"
	}

	offsetRef := addArg(int64(offset))
	limitRef := addArg(int64(limit))

	query := fmt.Sprintf(`SELECT i.id, i.workspace_id, i.title, i.description, i.status, i.priority,
       i.assignee_type, i.assignee_id, i.creator_type, i.creator_id,
       i.parent_issue_id, i.position, i.start_date, i.due_date, i.created_at, i.updated_at, i.last_activity_at, i.number, i.project_id, i.metadata, i.stage, i.properties,
	   i.revision, i.duplicate_of_issue_id
FROM issue i
WHERE %s
ORDER BY %s
LIMIT %s OFFSET %s`, whereSql, orderBy, limitRef, offsetRef)

	rows, err := h.DB.Query(ctx, query, args...)
	if err != nil {
		slog.Warn("ListIssues query failed", "error", err)
		writeError(w, http.StatusInternalServerError, "failed to list issues")
		return
	}
	defer rows.Close()

	var issues []db.ListIssuesRow
	for rows.Next() {
		var row db.ListIssuesRow
		if err := rows.Scan(
			&row.ID,
			&row.WorkspaceID,
			&row.Title,
			&row.Description,
			&row.Status,
			&row.Priority,
			&row.AssigneeType,
			&row.AssigneeID,
			&row.CreatorType,
			&row.CreatorID,
			&row.ParentIssueID,
			&row.Position,
			&row.StartDate,
			&row.DueDate,
			&row.CreatedAt,
			&row.UpdatedAt,
			&row.LastActivityAt,
			&row.Number,
			&row.ProjectID,
			&row.Metadata,
			&row.Stage,
			&row.Properties,
			&row.Revision,
			&row.DuplicateOfIssueID,
		); err != nil {
			slog.Warn("ListIssues scan failed", "error", err)
			writeError(w, http.StatusInternalServerError, "failed to list issues")
			return
		}
		issues = append(issues, row)
	}
	if err := rows.Err(); err != nil {
		slog.Warn("ListIssues rows failed", "error", err)
		writeError(w, http.StatusInternalServerError, "failed to list issues")
		return
	}

	// Get the true total count for pagination awareness.
	countQuery := fmt.Sprintf(`SELECT COUNT(*) FROM issue i WHERE %s`, whereSql)
	// Count query uses the same args minus the OFFSET and LIMIT params (last two added).
	countArgs := args[:filterArgCount]
	var total int64
	if err := h.DB.QueryRow(ctx, countQuery, countArgs...).Scan(&total); err != nil {
		slog.Warn("ListIssues count failed", "error", err)
		writeError(w, http.StatusInternalServerError, "failed to count issues")
		return
	}

	prefix := h.getIssuePrefix(ctx, wsUUID)
	ids := make([]pgtype.UUID, len(issues))
	for i, issue := range issues {
		ids[i] = issue.ID
	}
	labelsMap := h.labelsByIssue(ctx, wsUUID, ids)
	resp := make([]IssueResponse, len(issues))
	for i, issue := range issues {
		resp[i] = issueListRowToResponse(issue, prefix)
		labels := labelsMap[resp[i].ID]
		if labels == nil {
			labels = []LabelResponse{}
		}
		resp[i].Labels = &labels
	}
	h.fillStatusCategories(ctx, wsUUID, resp)

	writeJSON(w, http.StatusOK, map[string]any{
		"issues": resp,
		"total":  total,
	})
}

type issueActorFilter struct {
	actorType string
	actorID   pgtype.UUID
}

type issueDateFilter struct {
	column string
	start  time.Time
	end    time.Time
}

func parseIssueDateFilter(w http.ResponseWriter, values url.Values) (*issueDateFilter, bool) {
	field := strings.TrimSpace(values.Get("date_field"))
	startRaw := strings.TrimSpace(values.Get("date_start"))
	endRaw := strings.TrimSpace(values.Get("date_end"))
	if field == "" && startRaw == "" && endRaw == "" {
		return nil, true
	}
	if field == "" || startRaw == "" || endRaw == "" {
		writeError(w, http.StatusBadRequest, "date_field, date_start, and date_end are required together")
		return nil, false
	}

	column := ""
	switch field {
	case "created_at":
		column = "created_at"
	case "updated_at":
		column = "updated_at"
	default:
		writeError(w, http.StatusBadRequest, "invalid date_field")
		return nil, false
	}

	start, err := time.Parse(time.RFC3339Nano, startRaw)
	if err != nil {
		writeError(w, http.StatusBadRequest, "invalid date_start")
		return nil, false
	}
	end, err := time.Parse(time.RFC3339Nano, endRaw)
	if err != nil {
		writeError(w, http.StatusBadRequest, "invalid date_end")
		return nil, false
	}
	if !start.Before(end) {
		writeError(w, http.StatusBadRequest, "date_start must be before date_end")
		return nil, false
	}

	return &issueDateFilter{column: column, start: start, end: end}, true
}

func appendIssueDateFilter(where []string, addArg func(any) string, filter *issueDateFilter) []string {
	if filter == nil {
		return where
	}
	startRef := addArg(filter.start)
	endRef := addArg(filter.end)
	return append(where, fmt.Sprintf(
		"i.%s >= %s AND i.%s < %s",
		filter.column,
		startRef,
		filter.column,
		endRef,
	))
}

// appendIssueTableSearchFilter adds a quick identity search to the ordinary
// ListIssues window. Unlike the ranked global search endpoint, this predicate
// preserves the table's active filters, explicit sort, total, and pagination.
// Every word must appear in the title; a complete identifier (or bare issue
// number) also matches the immutable numeric issue number.
func appendIssueTableSearchFilter(where []string, addArg func(any) string, raw string) []string {
	query := strings.TrimSpace(raw)
	if query == "" {
		return where
	}

	words := splitSearchTerms(strings.ToLower(query))
	ors := make([]string, 0, 2)
	if len(words) > 0 {
		titleMatches := make([]string, 0, len(words))
		for _, word := range words {
			pattern := "%" + escapeLike(word) + "%"
			titleMatches = append(titleMatches, fmt.Sprintf("LOWER(i.title) LIKE %s", addArg(pattern)))
		}
		ors = append(ors, "("+strings.Join(titleMatches, " AND ")+")")
	}
	if number, ok := parseQueryNumber(query); ok {
		ors = append(ors, fmt.Sprintf("i.number = %s", addArg(number)))
	}
	if len(ors) == 0 {
		return where
	}
	return append(where, "("+strings.Join(ors, " OR ")+")")
}

func splitCommaParam(raw string) []string {
	if raw == "" {
		return nil
	}
	parts := strings.Split(raw, ",")
	out := make([]string, 0, len(parts))
	for _, part := range parts {
		if trimmed := strings.TrimSpace(part); trimmed != "" {
			out = append(out, trimmed)
		}
	}
	return out
}

func isIssueActorType(s string) bool {
	return s == "member" || s == "agent" || s == "squad"
}

func parseUUIDParamList(w http.ResponseWriter, raw, fieldName string) ([]pgtype.UUID, bool) {
	parts := splitCommaParam(raw)
	if len(parts) == 0 {
		return nil, true
	}
	ids := make([]pgtype.UUID, 0, len(parts))
	for _, part := range parts {
		id, ok := parseUUIDOrBadRequest(w, part, fieldName)
		if !ok {
			return nil, false
		}
		ids = append(ids, id)
	}
	return ids, true
}

func parseActorFilterList(w http.ResponseWriter, raw, fieldName string) ([]issueActorFilter, bool) {
	parts := splitCommaParam(raw)
	if len(parts) == 0 {
		return nil, true
	}
	filters := make([]issueActorFilter, 0, len(parts))
	for _, part := range parts {
		pieces := strings.SplitN(part, ":", 2)
		if len(pieces) != 2 || !isIssueActorType(pieces[0]) || strings.TrimSpace(pieces[1]) == "" {
			writeError(w, http.StatusBadRequest, "invalid "+fieldName)
			return nil, false
		}
		id, ok := parseUUIDOrBadRequest(w, strings.TrimSpace(pieces[1]), fieldName)
		if !ok {
			return nil, false
		}
		filters = append(filters, issueActorFilter{
			actorType: pieces[0],
			actorID:   id,
		})
	}
	return filters, true
}

func (h *Handler) ListGroupedIssues(w http.ResponseWriter, r *http.Request) {
	ctx := r.Context()
	if h.DB == nil {
		writeError(w, http.StatusInternalServerError, "database is unavailable")
		return
	}

	groupBy := r.URL.Query().Get("group_by")
	if groupBy == "" {
		groupBy = "assignee"
	}
	if groupBy != "assignee" {
		writeError(w, http.StatusBadRequest, "unsupported group_by")
		return
	}

	workspaceID := h.resolveWorkspaceID(r)
	wsUUID, ok := parseUUIDOrBadRequest(w, workspaceID, "workspace_id")
	if !ok {
		return
	}

	limit := 50
	offset := 0
	if l := r.URL.Query().Get("limit"); l != "" {
		if v, err := strconv.Atoi(l); err == nil && v > 0 {
			limit = v
		}
	}
	if limit > 100 {
		limit = 100
	}
	if o := r.URL.Query().Get("offset"); o != "" {
		if v, err := strconv.Atoi(o); err == nil && v > 0 {
			offset = v
		}
	}

	where := []string{"i.workspace_id = $1"}
	args := []any{wsUUID}
	addArg := func(v any) string {
		args = append(args, v)
		return "$" + strconv.Itoa(len(args))
	}

	statuses := splitCommaParam(r.URL.Query().Get("statuses"))
	if len(statuses) == 0 {
		statuses = splitCommaParam(r.URL.Query().Get("status"))
	}
	if len(statuses) > 0 {
		where = append(where, fmt.Sprintf("i.status = ANY(%s::text[])", addArg(statuses)))
	}
	// See ListIssues: category filtering is what lets the board keep a fixed
	// column count as a workspace adds custom statuses. (MUL-6243)
	statusCategories := splitCommaParam(r.URL.Query().Get("status_categories"))
	if len(statusCategories) == 0 {
		statusCategories = splitCommaParam(r.URL.Query().Get("status_category"))
	}
	if len(statusCategories) > 0 {
		// See ListIssues: expanded to keys so the (workspace_id, status) index
		// still drives the scan. (MUL-6243)
		keys, err := issuestatus.ExpandCategories(r.Context(), h.Queries, wsUUID, statusCategories)
		if err != nil {
			slog.Warn("expand status categories failed", append(logger.RequestAttrs(r), "error", err)...)
			writeError(w, http.StatusInternalServerError, "failed to resolve status categories")
			return
		}
		where = append(where, fmt.Sprintf("i.status = ANY(%s::text[])", addArg(keys)))
	}

	priorities := splitCommaParam(r.URL.Query().Get("priorities"))
	if len(priorities) == 0 {
		priorities = splitCommaParam(r.URL.Query().Get("priority"))
	}
	if len(priorities) > 0 {
		where = append(where, fmt.Sprintf("i.priority = ANY(%s::text[])", addArg(priorities)))
	}

	assigneeTypes := splitCommaParam(r.URL.Query().Get("assignee_types"))
	if len(assigneeTypes) > 0 {
		for _, assigneeType := range assigneeTypes {
			if !isIssueActorType(assigneeType) {
				writeError(w, http.StatusBadRequest, "invalid assignee_types")
				return
			}
		}
		where = append(where, fmt.Sprintf("i.assignee_type = ANY(%s::text[])", addArg(assigneeTypes)))
	}

	if raw := r.URL.Query().Get("assignee_id"); raw != "" {
		id, ok := parseUUIDOrBadRequest(w, raw, "assignee_id")
		if !ok {
			return
		}
		where = append(where, fmt.Sprintf("i.assignee_id = %s::uuid", addArg(id)))
	}
	if raw := r.URL.Query().Get("assignee_ids"); raw != "" {
		ids, ok := parseUUIDParamList(w, raw, "assignee_ids")
		if !ok {
			return
		}
		if len(ids) > 0 {
			where = append(where, fmt.Sprintf("i.assignee_id = ANY(%s::uuid[])", addArg(ids)))
		}
	}
	if raw := r.URL.Query().Get("creator_id"); raw != "" {
		id, ok := parseUUIDOrBadRequest(w, raw, "creator_id")
		if !ok {
			return
		}
		where = append(where, fmt.Sprintf("i.creator_id = %s::uuid", addArg(id)))
	}
	if raw := r.URL.Query().Get("project_id"); raw != "" {
		id, ok := parseUUIDOrBadRequest(w, raw, "project_id")
		if !ok {
			return
		}
		where = append(where, fmt.Sprintf("i.project_id = %s::uuid", addArg(id)))
	}
	if filter, ok := parseMetadataFilterParam(w, r.URL.Query().Get("metadata")); !ok {
		return
	} else if filter != nil {
		where = append(where, fmt.Sprintf("i.metadata @> %s::jsonb", addArg(string(filter))))
	}
	if filter, ok := parsePropertiesFilterParam(w, r.URL.Query().Get("properties")); !ok {
		return
	} else if filter != nil {
		where = append(where, propertiesFilterPredicate(filter, addArg))
	}
	// Mirror the involves_user_id 4-branch UNION from sqlc's ListIssues /
	// ListOpenIssues / CountIssues. ListGroupedIssues is a hand-written dynamic
	// SQL builder that does not share parameters with sqlc, so the fragment is
	// re-implemented here in lock-step. Member-direct assignment is excluded by
	// design: that semantics belongs to tab 1 (`assignee_id`), and tab 3 must
	// stay disjoint from tab 1.
	if raw := r.URL.Query().Get("involves_user_id"); raw != "" {
		id, ok := parseUUIDOrBadRequest(w, raw, "involves_user_id")
		if !ok {
			return
		}
		ref := addArg(id)
		where = append(where, fmt.Sprintf(`(
    (i.assignee_type = 'agent' AND i.assignee_id IN (
       SELECT a.id FROM agent a
        WHERE a.workspace_id = $1
          AND a.owner_id     = %[1]s::uuid
    ))
    OR (i.assignee_type = 'squad' AND i.assignee_id IN (
       SELECT sm.squad_id
         FROM squad_member sm
         JOIN squad s ON s.id = sm.squad_id
        WHERE s.workspace_id = $1
          AND sm.member_type = 'member'
          AND sm.member_id   = %[1]s::uuid
       UNION
       SELECT s.id
         FROM squad s
         JOIN agent a ON a.id = s.leader_id
        WHERE s.workspace_id = $1
          AND a.workspace_id = $1
          AND a.owner_id     = %[1]s::uuid
       UNION
       SELECT sm.squad_id
         FROM squad_member sm
         JOIN squad s ON s.id = sm.squad_id
         JOIN agent a ON a.id = sm.member_id
        WHERE s.workspace_id = $1
          AND sm.member_type = 'agent'
          AND a.workspace_id = $1
          AND a.owner_id     = %[1]s::uuid
    ))
)`, ref))
	}

	assigneeFilters, ok := parseActorFilterList(w, r.URL.Query().Get("assignee_filters"), "assignee_filters")
	if !ok {
		return
	}
	includeNoAssignee := r.URL.Query().Get("include_no_assignee") == "true"
	if len(assigneeFilters) > 0 || includeNoAssignee {
		ors := make([]string, 0, len(assigneeFilters)+1)
		for _, filter := range assigneeFilters {
			ors = append(ors, fmt.Sprintf(
				"(i.assignee_type = %s::text AND i.assignee_id = %s::uuid)",
				addArg(filter.actorType),
				addArg(filter.actorID),
			))
		}
		if includeNoAssignee {
			ors = append(ors, "(i.assignee_type IS NULL AND i.assignee_id IS NULL)")
		}
		where = append(where, "("+strings.Join(ors, " OR ")+")")
	}

	creatorFilters, ok := parseActorFilterList(w, r.URL.Query().Get("creator_filters"), "creator_filters")
	if !ok {
		return
	}
	if len(creatorFilters) > 0 {
		ors := make([]string, 0, len(creatorFilters))
		for _, filter := range creatorFilters {
			ors = append(ors, fmt.Sprintf(
				"(i.creator_type = %s::text AND i.creator_id = %s::uuid)",
				addArg(filter.actorType),
				addArg(filter.actorID),
			))
		}
		where = append(where, "("+strings.Join(ors, " OR ")+")")
	}

	projectIDs, ok := parseUUIDParamList(w, r.URL.Query().Get("project_ids"), "project_ids")
	if !ok {
		return
	}
	includeNoProject := r.URL.Query().Get("include_no_project") == "true"
	if len(projectIDs) > 0 || includeNoProject {
		ors := make([]string, 0, 2)
		if len(projectIDs) > 0 {
			ors = append(ors, fmt.Sprintf("i.project_id = ANY(%s::uuid[])", addArg(projectIDs)))
		}
		if includeNoProject {
			ors = append(ors, "i.project_id IS NULL")
		}
		where = append(where, "("+strings.Join(ors, " OR ")+")")
	}

	labelIDs, ok := parseUUIDParamList(w, r.URL.Query().Get("label_ids"), "label_ids")
	if !ok {
		return
	}
	if len(labelIDs) > 0 {
		where = append(where, fmt.Sprintf(
			"EXISTS (SELECT 1 FROM issue_to_label itl WHERE itl.issue_id = i.id AND itl.label_id = ANY(%s::uuid[]))",
			addArg(labelIDs),
		))
	}

	dateFilter, ok := parseIssueDateFilter(w, r.URL.Query())
	if !ok {
		return
	}
	where = appendIssueDateFilter(where, addArg, dateFilter)

	if groupAssigneeType := r.URL.Query().Get("group_assignee_type"); groupAssigneeType != "" {
		if groupAssigneeType == "none" {
			where = append(where, "(i.assignee_type IS NULL AND i.assignee_id IS NULL)")
		} else {
			if !isIssueActorType(groupAssigneeType) {
				writeError(w, http.StatusBadRequest, "invalid group_assignee_type")
				return
			}
			rawID := r.URL.Query().Get("group_assignee_id")
			if rawID == "" {
				writeError(w, http.StatusBadRequest, "invalid group_assignee_id")
				return
			}
			assigneeID, ok := parseUUIDOrBadRequest(w, rawID, "group_assignee_id")
			if !ok {
				return
			}
			where = append(where, fmt.Sprintf(
				"(i.assignee_type = %s::text AND i.assignee_id = %s::uuid)",
				addArg(groupAssigneeType),
				addArg(assigneeID),
			))
		}
	}

	sortCol := "position"
	sortIsExpr := false
	sortIsProperty := false
	if s := r.URL.Query().Get("sort"); s != "" {
		switch s {
		case "position", "title", "created_at", "updated_at", "start_date", "due_date":
			sortCol = s
		case "last_activity":
			sortCol = "last_activity_at"
		case "status":
			var err error
			sortCol, err = h.issueStatusSortExpression(r.Context(), wsUUID, addArg)
			if err != nil {
				slog.Warn("resolve grouped status sort failed", append(logger.RequestAttrs(r), "error", err)...)
				writeError(w, http.StatusInternalServerError, "failed to resolve sort")
				return
			}
			sortIsExpr = true
		case "priority":
			sortCol = "CASE i.priority WHEN 'urgent' THEN 0 WHEN 'high' THEN 1 WHEN 'medium' THEN 2 WHEN 'low' THEN 3 ELSE 4 END"
			sortIsExpr = true
		default:
			// property:<definitionId> sorts by the custom-property value
			// (typed expression); unknown/archived definitions degrade to
			// position order instead of erroring stale clients.
			expr, handled, sortErr := h.propertySortExpr(r, workspaceID, s)
			if !handled {
				writeError(w, http.StatusBadRequest, "invalid sort value")
				return
			}
			if sortErr != nil {
				if sortErr.Error() == "invalid sort value" || sortErr.Error() == "invalid workspace id" {
					writeError(w, http.StatusBadRequest, sortErr.Error())
					return
				}
				slog.Warn("propertySortExpr failed", append(logger.RequestAttrs(r), "error", sortErr)...)
				writeError(w, http.StatusInternalServerError, "failed to resolve sort")
				return
			}
			if expr != "" {
				sortCol = expr
				sortIsExpr = true
				sortIsProperty = true
			}
		}
	}
	sortDir := "ASC"
	if sortCol == "last_activity_at" {
		sortDir = "DESC"
	}
	if sortCol != "position" {
		if d := r.URL.Query().Get("direction"); d != "" {
			switch strings.ToLower(d) {
			case "asc":
				sortDir = "ASC"
			case "desc":
				sortDir = "DESC"
			default:
				writeError(w, http.StatusBadRequest, "invalid direction value")
				return
			}
		}
	}

	intraGroupOrder := sortCol
	if !sortIsExpr {
		intraGroupOrder = "i." + sortCol
	}
	intraGroupOrder += " " + sortDir
	if sortCol == "start_date" || sortCol == "due_date" || sortCol == "last_activity_at" || sortIsProperty {
		intraGroupOrder += " NULLS LAST"
	}
	// Unique final key — see ListIssues: created_at ties would otherwise make
	// ROW_NUMBER() unstable across per-group offset pages.
	if sortCol == "last_activity_at" {
		intraGroupOrder += ", i.id DESC"
	} else {
		intraGroupOrder += ", i.created_at DESC, i.id DESC"
	}

	offsetRef := addArg(int64(offset))
	limitRef := addArg(int64(limit))
	query := fmt.Sprintf(`
WITH ranked AS (
	SELECT
		i.id, i.workspace_id, i.title, i.description, i.status, i.priority,
		i.assignee_type, i.assignee_id, i.creator_type, i.creator_id,
		i.parent_issue_id, i.position, i.start_date, i.due_date, i.created_at, i.updated_at, i.last_activity_at,
		i.number, i.project_id, i.metadata, i.stage, i.properties, i.revision, i.duplicate_of_issue_id,
		COUNT(*) OVER (PARTITION BY i.assignee_type, i.assignee_id) AS group_total,
		ROW_NUMBER() OVER (
			PARTITION BY i.assignee_type, i.assignee_id
			ORDER BY %s
		) AS rn
	FROM issue i
	WHERE %s
)
SELECT
	id, workspace_id, title, description, status, priority,
	assignee_type, assignee_id, creator_type, creator_id,
	parent_issue_id, position, start_date, due_date, created_at, updated_at, last_activity_at,
	number, project_id, metadata, stage, properties, revision, duplicate_of_issue_id, group_total
FROM ranked
WHERE rn > %s AND rn <= %s + %s
ORDER BY
	CASE assignee_type
		WHEN 'member' THEN 0
		WHEN 'agent' THEN 1
		WHEN 'squad' THEN 2
		ELSE 3
	END,
	assignee_type NULLS LAST,
	assignee_id NULLS LAST,
	rn`, intraGroupOrder, strings.Join(where, " AND "), offsetRef, offsetRef, limitRef)

	rows, err := h.DB.Query(ctx, query, args...)
	if err != nil {
		slog.Warn("ListGroupedIssues query failed", "error", err)
		writeError(w, http.StatusInternalServerError, "failed to list grouped issues")
		return
	}
	defer rows.Close()

	groupedRows := []groupedIssueRow{}
	for rows.Next() {
		var row groupedIssueRow
		if err := rows.Scan(
			&row.ID,
			&row.WorkspaceID,
			&row.Title,
			&row.Description,
			&row.Status,
			&row.Priority,
			&row.AssigneeType,
			&row.AssigneeID,
			&row.CreatorType,
			&row.CreatorID,
			&row.ParentIssueID,
			&row.Position,
			&row.StartDate,
			&row.DueDate,
			&row.CreatedAt,
			&row.UpdatedAt,
			&row.LastActivityAt,
			&row.Number,
			&row.ProjectID,
			&row.Metadata,
			&row.Stage,
			&row.Properties,
			&row.Revision,
			&row.DuplicateOfIssueID,
			&row.GroupTotal,
		); err != nil {
			slog.Warn("ListGroupedIssues scan failed", "error", err)
			writeError(w, http.StatusInternalServerError, "failed to list grouped issues")
			return
		}
		groupedRows = append(groupedRows, row)
	}
	if err := rows.Err(); err != nil {
		slog.Warn("ListGroupedIssues rows failed", "error", err)
		writeError(w, http.StatusInternalServerError, "failed to list grouped issues")
		return
	}

	ids := make([]pgtype.UUID, len(groupedRows))
	var originals []pgtype.UUID
	for i, row := range groupedRows {
		ids[i] = row.ID
		originals = appendDuplicateOriginal(originals, row.Status, row.DuplicateOfIssueID)
	}
	labelsMap := h.labelsByIssue(ctx, wsUUID, ids)
	prefix := h.getIssuePrefix(ctx, wsUUID)
	// One Resolver for the whole page — a per-row filler would query the
	// catalog once per custom-status row. (MUL-6243)
	fillGrouped := h.newStatusCategoryFiller(ctx, wsUUID, originals...)

	groups := []IssueAssigneeGroupResponse{}
	groupIndex := map[string]int{}
	for _, row := range groupedRows {
		groupID := assigneeGroupID(row.AssigneeType, row.AssigneeID)
		idx, exists := groupIndex[groupID]
		if !exists {
			idx = len(groups)
			groupIndex[groupID] = idx
			groups = append(groups, IssueAssigneeGroupResponse{
				ID:           groupID,
				AssigneeType: textToPtr(row.AssigneeType),
				AssigneeID:   uuidToPtr(row.AssigneeID),
				Issues:       []IssueResponse{},
				Total:        row.GroupTotal,
			})
		}

		issue := issueListRowToResponse(row.ListIssuesRow, prefix)
		fillGrouped(&issue)
		labels := labelsMap[issue.ID]
		if labels == nil {
			labels = []LabelResponse{}
		}
		issue.Labels = &labels
		groups[idx].Issues = append(groups[idx].Issues, issue)
	}

	writeJSON(w, http.StatusOK, GroupedIssuesResponse{Groups: groups})
}

func (h *Handler) GetIssue(w http.ResponseWriter, r *http.Request) {
	id := chi.URLParam(r, "id")
	issue, ok := h.loadIssueForUser(w, r, id)
	if !ok {
		return
	}
	prefix := h.getIssuePrefix(r.Context(), issue.WorkspaceID)
	resp := issueToResponse(issue, prefix)
	h.fillStatusCategory(r.Context(), issue.WorkspaceID, &resp)
	detailLabels := h.labelsByIssue(r.Context(), issue.WorkspaceID, []pgtype.UUID{issue.ID})[uuidToString(issue.ID)]
	if detailLabels == nil {
		detailLabels = []LabelResponse{}
	}
	resp.Labels = &detailLabels

	// Fetch issue reactions.
	reactions, err := h.Queries.ListIssueReactions(r.Context(), issue.ID)
	if err == nil && len(reactions) > 0 {
		resp.Reactions = make([]IssueReactionResponse, len(reactions))
		for i, rx := range reactions {
			resp.Reactions[i] = issueReactionToResponse(rx)
		}
	}

	// Fetch issue-level attachments.
	attachments, err := h.Queries.ListAttachmentsByIssue(r.Context(), db.ListAttachmentsByIssueParams{
		IssueID:     issue.ID,
		WorkspaceID: issue.WorkspaceID,
	})
	if err == nil && len(attachments) > 0 {
		mode := attachmentURLModeFromRequest(r)
		resp.Attachments = make([]AttachmentResponse, len(attachments))
		for i, a := range attachments {
			resp.Attachments[i] = h.attachmentToResponse(a, mode)
		}
	}

	if sourceContext, err := h.issueSourceContextDetail(r.Context(), issue); err == nil {
		resp.SourceContext = sourceContext
	} else if !errors.Is(err, pgx.ErrNoRows) {
		slog.Warn("load issue source context failed", append(logger.RequestAttrs(r), "issue_id", uuidToString(issue.ID), "error", err)...)
		// The frozen context is part of the issue's execution contract, not an
		// optional decoration. Returning a successful detail response without it
		// would let agents run with silently incomplete instructions.
		writeError(w, http.StatusInternalServerError, "failed to load issue source context")
		return
	}

	writeJSON(w, http.StatusOK, resp)
}

func (h *Handler) ListChildIssues(w http.ResponseWriter, r *http.Request) {
	id := chi.URLParam(r, "id")
	issue, ok := h.loadIssueForUser(w, r, id)
	if !ok {
		return
	}
	children, err := h.Queries.ListChildIssues(r.Context(), issue.ID)
	if err != nil {
		writeError(w, http.StatusInternalServerError, "failed to list child issues")
		return
	}
	prefix := h.getIssuePrefix(r.Context(), issue.WorkspaceID)
	ids := make([]pgtype.UUID, len(children))
	for i, child := range children {
		ids[i] = child.ID
	}
	labelsMap := h.labelsByIssue(r.Context(), issue.WorkspaceID, ids)
	// Sub-issue progress is computed from these rows (the CLI's `issue children`
	// stage counts, among others), so they carry the resolved category — a
	// custom completed status must count as completed. One Resolver for the whole list:
	// built-in statuses still cost no query, and a list full of custom ones
	// costs one catalog read rather than one per row.
	statusResolver := issuestatus.NewResolver(issue.WorkspaceID)
	resp := make([]IssueResponse, len(children))
	for i, child := range children {
		resp[i] = issueToResponse(child, prefix)
		resp[i].StatusCategory = issuestatus.WireCategory(child.Status, statusResolver.Category(r.Context(), h.Queries, child.Status))
		labels := labelsMap[resp[i].ID]
		if labels == nil {
			labels = []LabelResponse{}
		}
		resp[i].Labels = &labels
	}
	writeJSON(w, http.StatusOK, map[string]any{
		"issues": resp,
	})
}

// Cap on the number of parents we'll fan-out children for in one request.
// Swimlane's visible-lane count is naturally bounded by what fits on screen
// (typically <= 50), but cap explicitly so a malicious caller can't ANY()
// across the whole workspace's issue set in a single round trip.
const listChildrenByParentsLimit = 200

// ListChildrenByParents returns the union of children for the
// provided parent ids. Replaces the N-call fan-out Swimlane would otherwise
// have to make on mount (one /issues/:id/children per visible parent lane).
//
// Workspace scope is enforced at the query level — any parent_id that doesn't
// belong to the caller's workspace simply yields zero children, so callers
// can't probe parents across workspace boundaries.
func (h *Handler) ListChildrenByParents(w http.ResponseWriter, r *http.Request) {
	workspaceID := h.resolveWorkspaceID(r)
	wsUUID, ok := parseUUIDOrBadRequest(w, workspaceID, "workspace_id")
	if !ok {
		return
	}

	raw := r.URL.Query().Get("parent_ids")
	if raw == "" {
		// Empty input is a no-op response (not an error) — simplifies the
		// client which calls this unconditionally on Swimlane mount even
		// when there are zero visible parent lanes.
		writeJSON(w, http.StatusOK, map[string]any{"issues": []IssueResponse{}})
		return
	}

	parts := strings.Split(raw, ",")
	if len(parts) > listChildrenByParentsLimit {
		writeError(w, http.StatusBadRequest, "too many parent_ids")
		return
	}
	parentIDs := make([]pgtype.UUID, 0, len(parts))
	for _, s := range parts {
		s = strings.TrimSpace(s)
		if s == "" {
			continue
		}
		id, ok := parseUUIDOrBadRequest(w, s, "parent_ids")
		if !ok {
			return
		}
		parentIDs = append(parentIDs, id)
	}
	if len(parentIDs) == 0 {
		writeJSON(w, http.StatusOK, map[string]any{"issues": []IssueResponse{}})
		return
	}

	children, err := h.Queries.ListChildrenByParents(r.Context(), db.ListChildrenByParentsParams{
		WorkspaceID: wsUUID,
		ParentIds:   parentIDs,
	})
	if err != nil {
		writeError(w, http.StatusInternalServerError, "failed to list child issues")
		return
	}
	prefix := h.getIssuePrefix(r.Context(), wsUUID)
	ids := make([]pgtype.UUID, len(children))
	for i, child := range children {
		ids[i] = child.ID
	}
	labelsMap := h.labelsByIssue(r.Context(), wsUUID, ids)
	// Sub-issue progress is computed from these rows (the CLI's `issue children`
	// stage counts, among others), so they carry the resolved category — a
	// custom completed status must count as completed. One Resolver for the whole list:
	// built-in statuses still cost no query, and a list full of custom ones
	// costs one catalog read rather than one per row.
	statusResolver := issuestatus.NewResolver(wsUUID)
	resp := make([]IssueResponse, len(children))
	for i, child := range children {
		resp[i] = issueToResponse(child, prefix)
		resp[i].StatusCategory = issuestatus.WireCategory(child.Status, statusResolver.Category(r.Context(), h.Queries, child.Status))
		labels := labelsMap[resp[i].ID]
		if labels == nil {
			labels = []LabelResponse{}
		}
		resp[i].Labels = &labels
	}
	writeJSON(w, http.StatusOK, map[string]any{
		"issues": resp,
	})
}

func (h *Handler) ChildIssueProgress(w http.ResponseWriter, r *http.Request) {
	wsID := h.resolveWorkspaceID(r)
	wsUUID, ok := parseUUIDOrBadRequest(w, wsID, "workspace_id")
	if !ok {
		return
	}

	terminalStatusKeys, err := h.terminalIssueStatusKeys(r.Context(), wsUUID)
	if err != nil {
		writeError(w, http.StatusInternalServerError, "failed to resolve status categories")
		return
	}
	rows, err := h.Queries.ChildIssueProgress(r.Context(), db.ChildIssueProgressParams{
		WorkspaceID:        wsUUID,
		TerminalStatusKeys: terminalStatusKeys,
	})
	if err != nil {
		writeError(w, http.StatusInternalServerError, "failed to get child issue progress")
		return
	}

	type progressEntry struct {
		ParentIssueID string `json:"parent_issue_id"`
		Total         int64  `json:"total"`
		Done          int64  `json:"done"`
	}
	resp := make([]progressEntry, len(rows))
	for i, row := range rows {
		resp[i] = progressEntry{
			ParentIssueID: uuidToString(row.ParentIssueID),
			Total:         row.Total,
			Done:          row.Done,
		}
	}
	writeJSON(w, http.StatusOK, map[string]any{
		"progress": resp,
	})
}

// QuickCreateIssueRequest is the body for POST /api/issues/quick-create. The
// user picks an actor (agent or squad) in the modal and types one line of
// natural language; the server validates the actor's reachability up front,
// queues a quick-create task, and returns 202 immediately. The agent
// translates the prompt into a `multica issue create` invocation in the
// background; success and failure both surface as inbox notifications to
// the requester.
//
// Exactly one of AgentID / SquadID is required. When SquadID is set, the
// task is enqueued against the squad's leader agent and the leader receives
// the same Operating Protocol briefing it would for an issue assigned to
// the squad, so it can choose to delegate to a squad member as usual.
//
// ProjectID is optional and lets the modal target a specific project so
// the agent's `multica issue create` invocation passes `--project <uuid>`
// instead of letting it default. The frontend remembers the user's last
// pick per workspace, so frequent users skip retyping "in project X".
//
// ParentIssueID is optional and is set by the "Add sub issue" entry point
// when the modal is opened from an existing issue. The agent passes it
// through as `--parent <uuid>` so the new issue is filed as a sub-issue,
// keeping the sub-issue intent of the entry point regardless of whether
// the user submits via manual or agent mode.
type QuickCreateIssueRequest struct {
	AgentID       string   `json:"agent_id,omitempty"`
	SquadID       string   `json:"squad_id,omitempty"`
	Prompt        string   `json:"prompt"`
	Priority      string   `json:"priority,omitempty"`
	DueDate       string   `json:"due_date,omitempty"`
	ProjectID     string   `json:"project_id,omitempty"`
	ParentIssueID string   `json:"parent_issue_id,omitempty"`
	AttachmentIDs []string `json:"attachment_ids,omitempty"`
}

// QuickCreateIssueResponse echoes the queued task id so the frontend can
// correlate the eventual inbox item, even though completion is fully async.
type QuickCreateIssueResponse struct {
	TaskID string `json:"task_id"`
}

func (h *Handler) QuickCreateIssue(w http.ResponseWriter, r *http.Request) {
	var req QuickCreateIssueRequest
	if err := json.NewDecoder(r.Body).Decode(&req); err != nil {
		writeError(w, http.StatusBadRequest, "invalid request body")
		return
	}
	prompt := strings.TrimSpace(req.Prompt)
	if prompt == "" {
		writeError(w, http.StatusBadRequest, "prompt is required")
		return
	}
	priority := strings.ToLower(strings.TrimSpace(req.Priority))
	if priority != "" && priority != "urgent" && priority != "high" && priority != "medium" && priority != "low" {
		writeError(w, http.StatusBadRequest, "priority must be one of: urgent, high, medium, low")
		return
	}
	dueDate := strings.TrimSpace(req.DueDate)
	if dueDate != "" {
		parsed, err := util.ParseCalendarDate(dueDate)
		if err != nil {
			writeError(w, http.StatusBadRequest, "invalid due_date format, expected YYYY-MM-DD")
			return
		}
		dueDate = parsed.Time.Format("2006-01-02")
	}

	hasAgent := strings.TrimSpace(req.AgentID) != ""
	hasSquad := strings.TrimSpace(req.SquadID) != ""
	if hasAgent == hasSquad {
		writeError(w, http.StatusBadRequest, "exactly one of agent_id or squad_id is required")
		return
	}

	workspaceID := h.resolveWorkspaceID(r)
	wsUUID, ok := parseUUIDOrBadRequest(w, workspaceID, "workspace_id")
	if !ok {
		return
	}

	requesterID, ok := requireUserID(w, r)
	if !ok {
		return
	}
	requesterUUID, ok := parseUUIDOrBadRequest(w, requesterID, "requester_id")
	if !ok {
		return
	}

	// Resolve the actor to the agent that will actually run the task. For
	// agent picks that's the agent itself; for squad picks it's the squad's
	// leader agent. The leader receives a squad-leader briefing on dispatch
	// (see daemon.go), matching the behavior of an issue assigned to the
	// squad — picking a squad here is functionally "ask the squad leader to
	// create this issue, on behalf of the squad".
	var agentUUID pgtype.UUID
	var squadUUID pgtype.UUID
	if hasSquad {
		var ok bool
		squadUUID, ok = parseUUIDOrBadRequest(w, req.SquadID, "squad_id")
		if !ok {
			return
		}
		squad, err := h.Queries.GetSquadInWorkspace(r.Context(), db.GetSquadInWorkspaceParams{
			ID:          squadUUID,
			WorkspaceID: wsUUID,
		})
		if err != nil {
			writeError(w, http.StatusNotFound, "squad not found")
			return
		}
		if squad.ArchivedAt.Valid {
			writeError(w, http.StatusBadRequest, "squad is archived")
			return
		}
		agentUUID = squad.LeaderID
	} else {
		var ok bool
		agentUUID, ok = parseUUIDOrBadRequest(w, req.AgentID, "agent_id")
		if !ok {
			return
		}
	}

	// Reuse the same workspace-membership / archived / private-agent
	// ownership rules as `validateAssigneePair` so a user can't POST a
	// private agent_id they shouldn't be able to dispatch (the frontend
	// filters them out, but the handler is the trust boundary). Squad
	// picks reach this with the resolved leader agent; the same rules
	// apply — a private leader behind a squad the user can't reach
	// should still be rejected.
	if status, msg := h.validateAssigneePair(
		r.Context(), r, workspaceID,
		pgtype.Text{String: "agent", Valid: true},
		agentUUID,
	); status != 0 {
		writeError(w, status, msg)
		return
	}

	// Re-load the agent for the runtime liveness check below. Safe by
	// construction: validateAssigneePair just confirmed it exists in this
	// workspace and the caller has visibility.
	agent, err := h.Queries.GetAgentInWorkspace(r.Context(), db.GetAgentInWorkspaceParams{
		ID:          agentUUID,
		WorkspaceID: wsUUID,
	})
	if err != nil {
		writeError(w, http.StatusNotFound, "agent not found")
		return
	}
	// Quick-create needs the agent to run NOW, so any non-ready verdict refuses
	// — but with the verdict's own code, so "CLI cannot run" no longer arrives
	// as "runtime is offline" and sends the user to reconnect a machine that is
	// already connected (MUL-6164).
	if verdict, err := service.AgentReadiness(r.Context(), h.runtimeLookup(obsmetrics.RuntimeLookupSourceIssue), agent); err != nil {
		writeError(w, http.StatusInternalServerError, "failed to check agent runtime")
		return
	} else if !verdict.Ready() {
		writeAgentUnavailable(w, verdict.Detail, verdict.Reason)
		return
	}

	// Daemon CLI version gate. The agent-side prompt + create-flow rely on
	// behaviors introduced in MinQuickCreateCLIVersion (URL attachment
	// handling, quick-create attachment binding, no-retry on partial failure).
	// Older daemons either double-create issues on partial CLI failures, drop
	// attachment bindings, or mishandle pasted screenshot URLs; fail closed
	// before enqueuing rather than surface the breakage as an inbox failure
	// twenty seconds later. Dev-built
	// daemons (git-describe shape) are exempted inside CheckMinCLIVersion
	// so `make daemon` works without weakening staging or production.
	if status, payload := h.checkQuickCreateDaemonVersion(r.Context(), obsmetrics.RuntimeLookupSourceIssue, agent.RuntimeID); status != 0 {
		writeJSON(w, status, payload)
		return
	}
	if priority != "" || dueDate != "" {
		if status, payload := h.checkQuickCreateDaemonVersionAtLeast(
			r.Context(), obsmetrics.RuntimeLookupSourceIssue, agent.RuntimeID, agentpkg.MinQuickCreateFieldsCLIVersion,
		); status != 0 {
			writeJSON(w, status, payload)
			return
		}
	}

	attachmentIDs, ok := parseUUIDSliceOrBadRequest(w, req.AttachmentIDs, "attachment_ids")
	if !ok {
		return
	}

	// Optional project_id — validate it belongs to the same workspace before
	// pinning the task to it. The handler is the trust boundary; the frontend
	// already only shows projects from the active workspace, but we re-check
	// here so a forged request can't smuggle a foreign project ID through.
	var projectUUID pgtype.UUID
	if strings.TrimSpace(req.ProjectID) != "" {
		pid, ok := parseUUIDOrBadRequest(w, req.ProjectID, "project_id")
		if !ok {
			return
		}
		if _, err := h.Queries.GetProjectInWorkspace(r.Context(), db.GetProjectInWorkspaceParams{
			ID:          pid,
			WorkspaceID: wsUUID,
		}); err != nil {
			writeError(w, http.StatusBadRequest, "project not found")
			return
		}
		projectUUID = pid
	}

	// Optional parent_issue_id — validate same-workspace membership just like
	// the regular CreateIssue path. Frontend seeds this from the "Add sub
	// issue" entry, but the handler re-checks so a forged request can't
	// smuggle a foreign parent UUID through.
	var parentIssueUUID pgtype.UUID
	if strings.TrimSpace(req.ParentIssueID) != "" {
		pid, ok := parseUUIDOrBadRequest(w, req.ParentIssueID, "parent_issue_id")
		if !ok {
			return
		}
		parent, err := h.Queries.GetIssueInWorkspace(r.Context(), db.GetIssueInWorkspaceParams{
			ID:          pid,
			WorkspaceID: wsUUID,
		})
		if err != nil || !parent.ID.Valid {
			writeError(w, http.StatusBadRequest, "parent issue not found in this workspace")
			return
		}
		parentIssueUUID = pid
	}

	task, err := h.TaskService.EnqueueQuickCreateTask(r.Context(), wsUUID, requesterUUID, agentUUID, squadUUID, prompt, priority, dueDate, projectUUID, parentIssueUUID, attachmentIDs)
	if err != nil {
		if writeIssueLimitReached(w, err) {
			return
		}
		slog.Warn("quick-create enqueue failed", append(logger.RequestAttrs(r), "error", err)...)
		writeError(w, http.StatusInternalServerError, "failed to enqueue quick-create task")
		return
	}

	writeJSON(w, http.StatusAccepted, QuickCreateIssueResponse{TaskID: uuidToString(task.ID)})
}

// writeAgentUnavailable returns 422 with a stable error code so the modal
// can show a "switch agent" hint without parsing the human-readable reason.
// writeAgentUnavailable refuses a trigger whose agent cannot run. `code` stays
// agent_unavailable for installed clients; reason_code carries the machine-
// readable distinction they need to phrase the fix — agent_runtime_required
// ("bind a runtime") is not runtime_offline ("reconnect the machine").
func writeAgentUnavailable(w http.ResponseWriter, reason string, reasonCode dispatch.ReasonCode) {
	w.Header().Set("Content-Type", "application/json")
	w.WriteHeader(http.StatusUnprocessableEntity)
	json.NewEncoder(w).Encode(map[string]any{
		"code":        "agent_unavailable",
		"reason":      reason,
		"reason_code": reasonCode,
	})
}

// isRuntimeOnline returns true when the given runtime is currently
// reachable (status == "online"). Quick-create rejects submissions whose
// agent's runtime is offline so the user gets immediate feedback in the
// modal instead of an inbox failure twenty seconds later.
func (h *Handler) isRuntimeOnline(ctx context.Context, runtimeID pgtype.UUID) bool {
	rt, err := h.getAgentRuntime(ctx, obsmetrics.RuntimeLookupSourceIssue, runtimeID)
	if err != nil {
		return false
	}
	return rt.Status == "online"
}

// checkQuickCreateDaemonVersion enforces MinQuickCreateCLIVersion against the
// CLI version the daemon reported at registration time (stored on the runtime
// row's metadata.cli_version). Returns (0, nil) when the version is
// acceptable, otherwise (status, payload) ready to hand to writeJSON.
//
// Failure shape is stable so the modal can branch on the `code` field and
// surface a "needs upgrade" hint that points at the specific runtime:
//
//	422 {
//	  "code": "daemon_version_unsupported",
//	  "current_version": "0.2.18" | "",
//	  "min_version":     "0.2.21",
//	  "runtime_id":      "<uuid>"
//	}
func (h *Handler) checkQuickCreateDaemonVersion(ctx context.Context, source string, runtimeID pgtype.UUID) (int, map[string]any) {
	return h.checkQuickCreateDaemonVersionAtLeast(ctx, source, runtimeID, agentpkg.MinQuickCreateCLIVersion)
}

func (h *Handler) checkQuickCreateDaemonVersionAtLeast(ctx context.Context, source string, runtimeID pgtype.UUID, minimum string) (int, map[string]any) {
	rt, err := h.getAgentRuntime(ctx, source, runtimeID)
	if err != nil {
		// Runtime row vanished between the online check and here — treat
		// as unavailable rather than wedging the request on a 500.
		return http.StatusUnprocessableEntity, map[string]any{
			"code":   "agent_unavailable",
			"reason": "agent's runtime is no longer registered",
		}
	}
	if rt.Provider == qoderruntime.Provider {
		return http.StatusUnprocessableEntity, map[string]any{
			"code":   "agent_unavailable",
			"reason": "Qoder Cloud Agents supports existing issue runs only; create the issue manually and assign it to this agent",
		}
	}
	current := readRuntimeCLIVersion(rt.Metadata)
	switch err := agentpkg.CheckMinCLIVersionFor(current, minimum); {
	case err == nil:
		return 0, nil
	case errors.Is(err, agentpkg.ErrCLIVersionMissing), errors.Is(err, agentpkg.ErrCLIVersionTooOld):
		return http.StatusUnprocessableEntity, map[string]any{
			"code":            "daemon_version_unsupported",
			"current_version": current,
			"min_version":     minimum,
			"runtime_id":      uuidToString(runtimeID),
		}
	default:
		// Defensive fall-through: unknown error from the version check is
		// also fail-closed, since the gate exists precisely because we
		// can't trust older daemons with this flow.
		return http.StatusUnprocessableEntity, map[string]any{
			"code":            "daemon_version_unsupported",
			"current_version": current,
			"min_version":     minimum,
			"runtime_id":      uuidToString(runtimeID),
		}
	}
}

// readRuntimeCLIVersion pulls metadata.cli_version off a runtime row. The
// metadata column is JSONB on the wire; the daemon stores the multica CLI
// version under that key during registration (see DaemonRegister).
func readRuntimeCLIVersion(metadata []byte) string {
	if len(metadata) == 0 {
		return ""
	}
	var m map[string]any
	if err := json.Unmarshal(metadata, &m); err != nil {
		return ""
	}
	if v, ok := m["cli_version"].(string); ok {
		return v
	}
	return ""
}

type CreateIssueRequest struct {
	Title         string   `json:"title"`
	Description   *string  `json:"description"`
	Status        string   `json:"status"`
	Priority      string   `json:"priority"`
	AssigneeType  *string  `json:"assignee_type"`
	AssigneeID    *string  `json:"assignee_id"`
	ParentIssueID *string  `json:"parent_issue_id"`
	ProjectID     *string  `json:"project_id"`
	Stage         *int32   `json:"stage,omitempty"`
	StartDate     *string  `json:"start_date"`
	DueDate       *string  `json:"due_date"`
	AttachmentIDs []string `json:"attachment_ids,omitempty"`
	// LabelIDs are issue-scoped labels to attach to the new issue in the same
	// transaction as the create. Unknown or non-issue ids are rejected with
	// 400 (service.ErrIssueLabelNotFound) rather than silently dropped.
	LabelIDs []string `json:"label_ids,omitempty"`
	// Properties is an ID-keyed bag whose values use the same typed wire shape
	// as Issue.properties and the standalone property PUT endpoint.
	Properties map[string]json.RawMessage `json:"properties,omitempty"`
	// OriginType / OriginID stamp the new issue with its provenance so
	// platform-internal flows can deterministically locate it later. Only
	// trusted callers should set these — currently the daemon CLI passes
	// them through for quick-create tasks (origin_type=quick_create,
	// origin_id=agent_task_queue.id).
	OriginType *string `json:"origin_type,omitempty"`
	OriginID   *string `json:"origin_id,omitempty"`

	AllowDuplicate bool `json:"allow_duplicate,omitempty"`
}

// UnmarshalJSON rejects duplicate object members before Go's ordinary map
// decoding can silently apply last-wins semantics. This is especially
// important for properties: two values for one definition must reject the
// whole create, never choose one implicitly.
func (r *CreateIssueRequest) UnmarshalJSON(data []byte) error {
	if err := rejectDuplicateJSONKeys(data); err != nil {
		return err
	}
	type createIssueRequestAlias CreateIssueRequest
	var decoded createIssueRequestAlias
	if err := json.Unmarshal(data, &decoded); err != nil {
		return err
	}
	*r = CreateIssueRequest(decoded)
	return nil
}

func rejectDuplicateJSONKeys(data []byte) error {
	decoder := json.NewDecoder(bytes.NewReader(data))
	decoder.UseNumber()
	if err := rejectDuplicateJSONValue(decoder); err != nil {
		return err
	}
	if _, err := decoder.Token(); !errors.Is(err, io.EOF) {
		if err == nil {
			return errors.New("request body must contain one JSON value")
		}
		return err
	}
	return nil
}

func rejectDuplicateJSONValue(decoder *json.Decoder) error {
	token, err := decoder.Token()
	if err != nil {
		return err
	}
	delimiter, ok := token.(json.Delim)
	if !ok {
		return nil
	}
	switch delimiter {
	case '{':
		seen := make(map[string]struct{})
		for decoder.More() {
			keyToken, err := decoder.Token()
			if err != nil {
				return err
			}
			key, ok := keyToken.(string)
			if !ok {
				return errors.New("object key must be a string")
			}
			if _, duplicate := seen[key]; duplicate {
				return fmt.Errorf("duplicate JSON key %q", key)
			}
			seen[key] = struct{}{}
			if err := rejectDuplicateJSONValue(decoder); err != nil {
				return err
			}
		}
		_, err = decoder.Token()
		return err
	case '[':
		for decoder.More() {
			if err := rejectDuplicateJSONValue(decoder); err != nil {
				return err
			}
		}
		_, err = decoder.Token()
		return err
	default:
		return errors.New("invalid JSON delimiter")
	}
}

func parseIssueCreateProperties(w http.ResponseWriter, values map[string]json.RawMessage) (map[pgtype.UUID]json.RawMessage, bool) {
	if len(values) == 0 {
		return nil, true
	}
	parsed := make(map[pgtype.UUID]json.RawMessage, len(values))
	for propertyID, value := range values {
		id, err := util.ParseUUID(propertyID)
		if err != nil {
			writeJSON(w, http.StatusBadRequest, map[string]any{
				"code": "invalid_issue_property", "property_id": propertyID,
				"error": "property id must be a UUID",
			})
			return nil, false
		}
		if _, duplicate := parsed[id]; duplicate {
			writeJSON(w, http.StatusBadRequest, map[string]any{
				"code": "invalid_issue_property", "property_id": propertyID,
				"error": "property was provided more than once",
			})
			return nil, false
		}
		parsed[id] = value
	}
	return parsed, true
}

func writeIssueCreatePropertyError(w http.ResponseWriter, err error) bool {
	var propertyErr *service.IssuePropertyValidationError
	if errors.As(err, &propertyErr) {
		writeJSON(w, http.StatusBadRequest, map[string]any{
			"code": "invalid_issue_property", "property_id": propertyErr.PropertyID,
			"error": propertyErr.Message,
		})
		return true
	}
	if errors.Is(err, service.ErrIssuePropertiesTooLarge) {
		writeJSON(w, http.StatusBadRequest, map[string]any{
			"code": "issue_properties_too_large", "error": err.Error(),
		})
		return true
	}
	return false
}

func duplicateIssueMessage(issue IssueResponse) string {
	return issueguard.DuplicateMessage(issue.Identifier, issue.Title, issue.Status)
}

func (h *Handler) CreateIssue(w http.ResponseWriter, r *http.Request) {
	r = h.withWakeupActor(r)
	var req CreateIssueRequest
	if err := json.NewDecoder(r.Body).Decode(&req); err != nil {
		writeError(w, http.StatusBadRequest, "invalid request body")
		return
	}

	if req.Title == "" {
		writeError(w, http.StatusBadRequest, "title is required")
		return
	}

	workspaceID := h.resolveWorkspaceID(r)
	wsUUID, ok := parseUUIDOrBadRequest(w, workspaceID, "workspace_id")
	if !ok {
		return
	}

	// Get creator from context (set by auth middleware)
	creatorID, ok := requireUserID(w, r)
	if !ok {
		return
	}

	status := req.Status
	if status == "" {
		status = "todo"
	}
	priority := req.Priority
	if priority == "" {
		priority = "none"
	}
	status, ok = h.resolveIssueStatusKey(w, r, wsUUID, status)
	if !ok {
		return
	}
	if !validateIssueEnum(w, "priority", priority, validIssuePriorities) {
		return
	}
	if req.Stage != nil && *req.Stage < 1 {
		writeError(w, http.StatusBadRequest, "stage must be >= 1")
		return
	}

	var assigneeType pgtype.Text
	var assigneeID pgtype.UUID
	if req.AssigneeType != nil {
		assigneeType = pgtype.Text{String: *req.AssigneeType, Valid: true}
	}
	if req.AssigneeID != nil {
		id, ok := parseUUIDOrBadRequest(w, *req.AssigneeID, "assignee_id")
		if !ok {
			return
		}
		assigneeID = id
	}

	var parentIssueID pgtype.UUID
	var projectID pgtype.UUID
	if req.ParentIssueID != nil {
		id, ok := parseUUIDOrBadRequest(w, *req.ParentIssueID, "parent_issue_id")
		if !ok {
			return
		}
		parentIssueID = id
		// The parent is loaded only to reject a cross-workspace or missing one
		// BEFORE the assignee gate runs, so the caller gets 400 "parent issue not
		// found" rather than a 403 that leaks nothing about which input was wrong.
		// The row itself is no longer needed: the assignee gate keys on the actor's
		// originator, not on a scope bound to the parent (MUL-6951).
		if assigneeType.Valid && (assigneeType.String == "agent" || assigneeType.String == "squad") {
			parent, err := h.Queries.GetIssueInWorkspace(r.Context(), db.GetIssueInWorkspaceParams{
				ID:          parentIssueID,
				WorkspaceID: wsUUID,
			})
			if err != nil || !parent.ID.Valid {
				writeError(w, http.StatusBadRequest, "parent issue not found in this workspace")
				return
			}
		}
	}

	if status, msg := h.validateAssigneePair(r.Context(), r, workspaceID, assigneeType, assigneeID); status != 0 {
		writeError(w, status, msg)
		return
	}

	if req.ProjectID != nil {
		id, ok := parseUUIDOrBadRequest(w, *req.ProjectID, "project_id")
		if !ok {
			return
		}
		projectID = id
	}

	// Project workflow (MUL-7420). A new issue in a project on a named
	// workflow must start on a status the workflow lists, and a column the
	// project hands off gives the new issue to its handler unless the caller
	// chose an assignee. A sub-issue without its own project inherits the
	// parent's, as IssueService.Create does.
	workflowProjectID := projectID
	if !workflowProjectID.Valid && parentIssueID.Valid {
		if parent, err := h.Queries.GetIssueInWorkspace(r.Context(), db.GetIssueInWorkspaceParams{
			ID: parentIssueID, WorkspaceID: wsUUID,
		}); err == nil {
			workflowProjectID = parent.ProjectID
		}
	}
	createWorkflow, createWorkflowProject, err := issueworkflow.ForProject(r.Context(), h.Queries, wsUUID, workflowProjectID)
	if err != nil {
		slog.Error("create issue: resolve project workflow", append(logger.RequestAttrs(r), "error", err)...)
		writeError(w, http.StatusInternalServerError, "failed to resolve project workflow")
		return
	}
	if createWorkflow != nil {
		if err := issueworkflow.CheckStatus(createWorkflow, status); err != nil {
			writeIssueWorkflowError(w, err)
			return
		}
		if req.AssigneeType == nil && req.AssigneeID == nil {
			if step, ok := createWorkflow.Step(status); ok && step.HandsOff() {
				handoffCreatorType, handoffCreatorID := h.resolveActor(r, creatorID, workspaceID)
				if t, id, ok := issueworkflow.ResolveHandler(step, createWorkflowProject, handoffCreatorType, parseUUID(handoffCreatorID)); ok {
					assigneeType = pgtype.Text{String: t, Valid: true}
					assigneeID = id
					if status, msg := h.validateAssigneePair(r.Context(), r, workspaceID, assigneeType, assigneeID); status != 0 {
						writeError(w, status, msg)
						return
					}
				}
			}
		}
	}
	// Project existence and the final parent boundary check are enforced inside
	// IssueService.Create atomically with the create. The handler preloads a
	// supplied parent only because the assignee gate must bind any autopilot
	// authority fallback to that server-verified issue before admission.

	attachmentIDs, ok := parseUUIDSliceOrBadRequest(w, req.AttachmentIDs, "attachment_ids")
	if !ok {
		return
	}

	labelIDs, ok := parseUUIDSliceOrBadRequest(w, req.LabelIDs, "label_ids")
	if !ok {
		return
	}
	properties, ok := parseIssueCreateProperties(w, req.Properties)
	if !ok {
		return
	}

	var startDate pgtype.Date
	if req.StartDate != nil && *req.StartDate != "" {
		d, err := util.ParseCalendarDate(*req.StartDate)
		if err != nil {
			writeError(w, http.StatusBadRequest, "invalid start_date format, expected YYYY-MM-DD")
			return
		}
		startDate = d
	}

	var dueDate pgtype.Date
	if req.DueDate != nil && *req.DueDate != "" {
		d, err := util.ParseCalendarDate(*req.DueDate)
		if err != nil {
			writeError(w, http.StatusBadRequest, "invalid due_date format, expected YYYY-MM-DD")
			return
		}
		dueDate = d
	}

	// Determine creator identity: agent (via X-Agent-ID header) or member.
	creatorType, actualCreatorID := h.resolveActor(r, creatorID, workspaceID)

	// Optional origin stamping (quick-create / autopilot). Only the
	// allowed origin types are accepted; anything else is rejected so a
	// rogue caller can't mint arbitrary origin labels. Both fields must
	// be provided together.
	var originType pgtype.Text
	var originID pgtype.UUID
	if req.OriginType != nil || req.OriginID != nil {
		if req.OriginType == nil || req.OriginID == nil {
			writeError(w, http.StatusBadRequest, "origin_type and origin_id must be provided together")
			return
		}
		switch *req.OriginType {
		case "quick_create":
			// Allowed — daemon CLI passes this through from a quick-create task.
		default:
			writeError(w, http.StatusBadRequest, "unsupported origin_type")
			return
		}
		oid, ok := parseUUIDOrBadRequest(w, *req.OriginID, "origin_id")
		if !ok {
			return
		}
		originType = pgtype.Text{String: *req.OriginType, Valid: true}
		originID = oid
	} else if creatorType == "agent" {
		// MUL-4305: an agent creating an issue via the ordinary create path
		// carries no explicit origin, which historically left the new issue
		// unattributed. Any run later derived from it (agent assignment,
		// squad-leader trigger) then lost the top-of-chain human originator,
		// so A2A @-mentions from those runs failed the canInvokeAgent gate
		// against private agents. Stamp the acting task as the issue's origin
		// so resolveOriginatorForIssueTask can inherit its originator — the
		// same trick CreateComment uses with comment.source_task_id (MUL-4015).
		//
		// The task id is taken from the SERVER-trusted X-Task-ID: the auth
		// middleware deletes whatever the client sent and re-stamps
		// X-Agent-ID / X-Task-ID only from a validated mat_ token (MUL-3428), so
		// a member-forged pair never reaches here — it is gone before
		// resolveActor runs, and the request resolves to creatorType=="member".
		// We still re-check the task belongs to the acting agent before trusting
		// it.
		if taskIDHeader := r.Header.Get("X-Task-ID"); taskIDHeader != "" {
			if taskUUID, perr := util.ParseUUID(taskIDHeader); perr == nil {
				if task, terr := h.Queries.GetAgentTask(r.Context(), taskUUID); terr == nil && uuidToString(task.AgentID) == actualCreatorID {
					originType = pgtype.Text{String: "agent_create", Valid: true}
					originID = taskUUID
				}
			}
		}
	}

	// Prefix is workspace-level; pre-compute once so both the broadcast
	// payload builder and the HTTP response share the same value.
	prefix := h.getIssuePrefix(r.Context(), wsUUID)

	// One filler for this create, shared by the broadcast payload and the HTTP
	// response below, so a custom-status create reads the catalog once per
	// request rather than once per payload. (MUL-6243)
	fillCreated := h.newStatusCategoryFiller(r.Context(), wsUUID)

	// Analytics agent ID: assignee agent when the issue is being assigned
	// to an agent, otherwise the creator agent for agent-authored issues.
	// Resolved here (not in the service) because creator identity is HTTP-side.
	analyticsAgentID := ""
	if assigneeType.Valid && assigneeType.String == "agent" {
		analyticsAgentID = uuidToString(assigneeID)
	}
	if creatorType == "agent" && analyticsAgentID == "" {
		analyticsAgentID = actualCreatorID
	}

	attachmentMode := attachmentURLModeFromRequest(r)
	buildAttachmentResponses := func(atts []db.Attachment) []AttachmentResponse {
		if len(atts) == 0 {
			return nil
		}
		out := make([]AttachmentResponse, len(atts))
		for i, a := range atts {
			out[i] = h.attachmentToResponse(a, attachmentMode)
		}
		return out
	}

	res, err := h.IssueService.Create(r.Context(), service.IssueCreateParams{
		WorkspaceID:    wsUUID,
		Title:          req.Title,
		Description:    ptrToText(req.Description),
		Status:         status,
		Priority:       priority,
		AssigneeType:   assigneeType,
		AssigneeID:     assigneeID,
		CreatorType:    creatorType,
		CreatorID:      parseUUID(actualCreatorID),
		ParentIssueID:  parentIssueID,
		ProjectID:      projectID,
		StartDate:      startDate,
		DueDate:        dueDate,
		OriginType:     originType,
		OriginID:       originID,
		Stage:          ptrToInt4(req.Stage),
		AttachmentIDs:  attachmentIDs,
		LabelIDs:       labelIDs,
		Properties:     properties,
		AllowDuplicate: req.AllowDuplicate,
	}, service.IssueCreateOpts{
		ActorID:          actualCreatorID,
		AnalyticsAgentID: analyticsAgentID,
		Platform:         func() string { p, _, _ := middleware.ClientMetadataFromContext(r.Context()); return p }(),
		BroadcastPayload: func(issue db.Issue, atts []db.Attachment, labels []db.IssueLabel) map[string]any {
			payload := issueToResponse(issue, prefix)
			// The event other tabs receive must carry the category too — filling
			// only the HTTP response below is too late for them, and a create
			// they cannot bucket forces a full refetch. Shares one filler with
			// the HTTP response so a custom-status create reads the catalog once
			// per request, not once per payload. (MUL-6243)
			fillCreated(&payload)
			payload.Attachments = buildAttachmentResponses(atts)
			// Carry the authoritative label snapshot so every online client
			// renders the new issue already labeled. Non-nil (even empty)
			// pointer = authoritative list; the old flow's separate
			// issue_labels:changed broadcast is gone.
			labelResponses := labelsToResponse(labels)
			payload.Labels = &labelResponses
			return map[string]any{"issue": payload}
		},
	})

	if errors.Is(err, service.ErrActiveDuplicate) {
		dup := *res.DuplicateIssue
		existing := issueToResponse(dup, h.getIssuePrefix(r.Context(), dup.WorkspaceID))
		h.fillStatusCategory(r.Context(), dup.WorkspaceID, &existing)
		writeJSON(w, http.StatusConflict, map[string]any{
			"code":  "active_duplicate_issue",
			"error": duplicateIssueMessage(existing),
			"issue": existing,
		})
		return
	}
	if errors.Is(err, service.ErrParentIssueNotFound) {
		writeError(w, http.StatusBadRequest, "parent issue not found in this workspace")
		return
	}
	if errors.Is(err, service.ErrProjectNotFound) {
		writeError(w, http.StatusBadRequest, "project not found in this workspace")
		return
	}
	if errors.Is(err, service.ErrIssueLabelNotFound) {
		writeError(w, http.StatusBadRequest, "one or more labels not found in this workspace")
		return
	}
	if errors.Is(err, service.ErrIssueStatusUnavailable) {
		writeError(w, http.StatusConflict,
			"the target status was archived while this request was in flight; reload the status list and retry")
		return
	}
	if writeIssueCreatePropertyError(w, err) {
		return
	}
	if writeIssueLimitReached(w, err) {
		return
	}
	if err != nil {
		slog.Warn("create issue failed", append(logger.RequestAttrs(r), "error", err, "workspace_id", workspaceID)...)
		writeError(w, http.StatusInternalServerError, "failed to create issue")
		return
	}

	issue := res.Issue
	slog.Info("issue created", append(logger.RequestAttrs(r), "issue_id", uuidToString(issue.ID), "title", issue.Title, "status", issue.Status, "workspace_id", workspaceID)...)

	resp := issueToResponse(issue, prefix)
	fillCreated(&resp)
	resp.Attachments = buildAttachmentResponses(res.Attachments)
	// Echo the authoritative labels attached in the create transaction. Always
	// non-nil (empty slice when none) so a newer client can tell the backend
	// understood label_ids and skip its legacy post-create attach fallback.
	labelResponses := labelsToResponse(res.Labels)
	resp.Labels = &labelResponses
	writeJSON(w, http.StatusCreated, resp)
}

type UpdateIssueRequest struct {
	ExpectedRevision *int64  `json:"expected_revision,omitempty"`
	Title            *string `json:"title"`
	// TitleBase is the title adopted by the editor before producing Title. It
	// protects title edits without coupling them to unrelated issue mutations.
	TitleBase   *string `json:"title_base,omitempty"`
	Description *string `json:"description"`
	// DescriptionBase is the authoritative Markdown the editor had adopted
	// before producing Description. It lets the server preserve channel media
	// that landed asynchronously after that base without making media already
	// present in the base impossible for the user to delete. Older clients omit
	// it and receive conservative channel-media preservation.
	DescriptionBase *string  `json:"description_base,omitempty"`
	Status          *string  `json:"status"`
	Priority        *string  `json:"priority"`
	AssigneeType    *string  `json:"assignee_type"`
	AssigneeID      *string  `json:"assignee_id"`
	Position        *float64 `json:"position"`
	StartDate       *string  `json:"start_date"`
	DueDate         *string  `json:"due_date"`
	ParentIssueID   *string  `json:"parent_issue_id"`
	ProjectID       *string  `json:"project_id"`
	Stage           *int32   `json:"stage"`
	// AttachmentIDs lets the description editor bind newly uploaded files to
	// this issue so they surface in `GET /api/issues/:id/attachments` and the
	// editor's preview Eye keeps working past a refresh. Existing bindings
	// are idempotent — re-sending the same id is a no-op.
	AttachmentIDs []string `json:"attachment_ids"`
	// SuppressRun, when true, applies the assignee/status change as usual but
	// skips starting the agent run this write would otherwise trigger
	// ("暂时不启动" — MUL-3375). It is not an undo: the change takes effect and
	// the issue can be run later via manual run/rerun. Optional; omitted or
	// false keeps today's behavior. Mirrors comment suppress_agent_ids.
	SuppressRun bool `json:"suppress_run,omitempty"`
	// HandoffNote is retained at the API boundary for installed clients that
	// predate the handoff UI removal. It is consumed only when this write starts
	// a run and is never stored on the issue itself.
	HandoffNote string `json:"handoff_note,omitempty"`
	// StopPreviousAssigneeRuns, when a status change hands the issue off to a
	// project workflow step's handler, cancels the active runs of the agent or
	// squad it was taken from. Ignored when the write does not hand off.
	// (MUL-7420)
	StopPreviousAssigneeRuns bool `json:"stop_previous_assignee_runs,omitempty"`
	// EnterWorkflowStep makes a project move enter the destination workflow's
	// step the issue lands on, handing it off like a status change would. A
	// project move without it is administrative. Batch updates ignore it.
	// (MUL-7420)
	EnterWorkflowStep bool `json:"enter_workflow_step,omitempty"`
	// DuplicateOfIssueID marks this issue as a duplicate of another issue in
	// the same workspace (MUL-7349). A duplicate is an ordinary cancelled issue
	// that remembers its original, so this also sets status to cancelled.
	// Any later status change away from cancelled removes the mark; there is
	// no explicit null. Write-only: read the relation back from
	// GET /api/issues/{id}/duplicates. Not accepted by batch updates.
	DuplicateOfIssueID *string `json:"duplicate_of_issue_id,omitempty"`
}

func mergeIssueChannelMediaDescription(current, incoming string, base *string, attachments []db.Attachment) string {
	currentIDs := channelmedia.MarkedIDs(current)
	if len(currentIDs) == 0 {
		return incoming
	}

	baseIDs := map[string]bool{}
	if base != nil {
		for _, id := range channelmedia.MarkedIDs(*base) {
			baseIDs[id] = true
		}
	}
	attachmentsByID := make(map[string]db.Attachment, len(attachments))
	for _, attachment := range attachments {
		attachmentsByID[uuidToString(attachment.ID)] = attachment
	}

	merged := incoming
	for _, id := range currentIDs {
		attachment, exists := attachmentsByID[id]
		if !exists {
			// A deleted attachment must not be resurrected from stale Markdown.
			continue
		}
		downloadPath := channelmedia.DownloadPath(id)
		hasLink := strings.Contains(merged, downloadPath)
		knownToEditor := base != nil && baseIDs[id]
		if knownToEditor && !hasLink {
			// The editor adopted this media and then removed its link: preserve
			// the user's explicit deletion rather than treating it as a race.
			continue
		}
		if !hasLink {
			merged = channelmedia.Append(merged, channelmedia.Block(
				id,
				attachment.Filename,
				strings.HasPrefix(attachment.ContentType, "image/"),
			))
			continue
		}
		// Tiptap may omit HTML comments when serializing an otherwise intact
		// image. Restore provenance without duplicating the visible link.
		if !channelmedia.HasMarker(merged, id) {
			merged = channelmedia.Append(merged, channelmedia.Marker(id))
		}
	}
	return merged
}

func refreshUntouchedNullableIssueParams(params *db.UpdateIssueParams, current db.Issue, rawFields map[string]json.RawMessage, workflowAssignee bool) {
	_, assigneeTypeTouched := rawFields["assignee_type"]
	_, assigneeIDTouched := rawFields["assignee_id"]
	// Assignee type and id form one validated value. If either half was
	// supplied, retain the pre-validation counterpart in params rather than
	// combining the supplied half with a concurrently-written counterpart that
	// has never been validated with it. A workflow-resolved assignee is also
	// an intended write, even though the caller sent no assignee fields.
	if !assigneeTypeTouched && !assigneeIDTouched && !workflowAssignee {
		params.AssigneeType = current.AssigneeType
		params.AssigneeID = current.AssigneeID
	}
	if _, touched := rawFields["start_date"]; !touched {
		params.StartDate = current.StartDate
	}
	if _, touched := rawFields["due_date"]; !touched {
		params.DueDate = current.DueDate
	}
	if _, touched := rawFields["parent_issue_id"]; !touched {
		params.ParentIssueID = current.ParentIssueID
	}
	if _, touched := rawFields["project_id"]; !touched {
		params.ProjectID = current.ProjectID
	}
	if _, touched := rawFields["stage"]; !touched {
		params.Stage = current.Stage
	}
}

var errIssueFieldConflict = errors.New("issue text field conflict")

func (h *Handler) updateIssueAtomically(ctx context.Context, workspaceID pgtype.UUID, params db.UpdateIssueParams, rawFields map[string]json.RawMessage, titleBase, descriptionBase *string, attachmentIDs []pgtype.UUID, statusKey string, workflowAssignee bool) (db.Issue, db.Issue, bool, error) {
	if h.TxStarter == nil {
		return db.Issue{}, db.Issue{}, false, errors.New("atomic issue update requires transaction starter")
	}
	tx, err := h.beginWakeupWrite(ctx)
	if err != nil {
		return db.Issue{}, db.Issue{}, false, fmt.Errorf("begin atomic issue update: %w", err)
	}
	defer tx.Rollback(ctx)

	qtx := h.Queries.WithTx(tx)
	// This path opens its own transaction, so it carries the archive-race guard
	// itself rather than going through runWithIssueStatusGuard. The catalog lock
	// must precede both attachment and issue row locks everywhere. (MUL-6243)
	if err := assertIssueStatusStillActive(ctx, qtx, workspaceID, statusKey); err != nil {
		return db.Issue{}, db.Issue{}, false, err
	}
	if len(attachmentIDs) > 0 {
		if _, err := qtx.LockAttachmentsForIssueLink(ctx, db.LockAttachmentsForIssueLinkParams{
			WorkspaceID:   workspaceID,
			AttachmentIds: attachmentIDs,
		}); err != nil {
			return db.Issue{}, db.Issue{}, false, fmt.Errorf("lock issue attachments: %w", err)
		}
	}
	current, err := qtx.LockIssueForDescriptionUpdate(ctx, db.LockIssueForDescriptionUpdateParams{
		ID:          params.ID,
		WorkspaceID: workspaceID,
	})
	if err != nil {
		return db.Issue{}, db.Issue{}, false, fmt.Errorf("lock issue for update: %w", err)
	}

	if params.Title.Valid && titleBase != nil && current.Title != *titleBase && current.Title != params.Title.String {
		return db.Issue{}, current, false, errIssueFieldConflict
	}

	if params.Description.Valid {
		attachments, listErr := qtx.ListAttachmentsByIssue(ctx, db.ListAttachmentsByIssueParams{
			IssueID:     current.ID,
			WorkspaceID: current.WorkspaceID,
		})
		if listErr != nil {
			return db.Issue{}, current, false, fmt.Errorf("list issue attachments for description merge: %w", listErr)
		}
		currentDescription := ""
		if current.Description.Valid {
			currentDescription = current.Description.String
		}
		incomingDescription := params.Description.String
		// No baseline REJECTION here, deliberately: the description editor
		// autosaves on a debounce, and its base could not be kept in step with
		// what the server had already accepted — a save whose own echo landed
		// while the editor was dirty, or any stored description that was not
		// byte-identical to its own trimmed form, reported a conflict with no
		// second writer present and then wedged the editor for the session
		// (MUL-6971). The guard also never covered the writers most likely to
		// race a human here — mobile and the CLI/agent path send no base at
		// all — so it mostly rejected the user's own autosave.
		//
		// `descriptionBase` stays in the request: it is ALSO the merge metadata
		// below, which is what lets a user delete channel media the editor had
		// adopted instead of having it restored on every save. Description
		// writes are last-write-wins; concurrent edits are recorded by the
		// `description_updated` activity.
		params.Description = pgtype.Text{
			String: mergeIssueChannelMediaDescription(currentDescription, incomingDescription, descriptionBase, attachments),
			Valid:  true,
		}
	}
	refreshUntouchedNullableIssueParams(&params, current, rawFields, workflowAssignee)

	issue, cancelledWakeups, err := updateIssueStoppingWakeups(ctx, qtx, params)
	if err != nil {
		return db.Issue{}, current, false, fmt.Errorf("update locked issue: %w", err)
	}

	attachmentsChanged := false
	if len(attachmentIDs) > 0 {
		linked, linkErr := qtx.LinkAttachmentsToIssue(ctx, db.LinkAttachmentsToIssueParams{
			IssueID:       issue.ID,
			WorkspaceID:   issue.WorkspaceID,
			AttachmentIds: attachmentIDs,
			BumpRevision:  issue.Revision == current.Revision,
		})
		if linkErr != nil {
			return db.Issue{}, current, false, fmt.Errorf("link issue attachments: %w", linkErr)
		}
		attachmentsChanged = linked.LinkedCount > 0
		if linked.IssueRevision > 0 {
			issue, err = qtx.GetIssueInWorkspace(ctx, db.GetIssueInWorkspaceParams{
				ID:          issue.ID,
				WorkspaceID: issue.WorkspaceID,
			})
			if err != nil {
				return db.Issue{}, current, false, fmt.Errorf("reload issue after attachment link: %w", err)
			}
		}
	}
	if err := tx.Commit(ctx); err != nil {
		return db.Issue{}, current, false, fmt.Errorf("commit atomic issue update: %w", err)
	}
	h.broadcastCancelledWakeups(ctx, workspaceID, cancelledWakeups)
	return issue, current, attachmentsChanged, nil
}

func (h *Handler) UpdateIssue(w http.ResponseWriter, r *http.Request) {
	r = h.withWakeupActor(r)
	id := chi.URLParam(r, "id")
	prevIssue, ok := h.loadIssueForUser(w, r, id)
	if !ok {
		return
	}
	userID := requestUserID(r)
	workspaceID := uuidToString(prevIssue.WorkspaceID)

	// Read body as raw bytes so we can detect which fields were explicitly sent.
	bodyBytes, err := io.ReadAll(r.Body)
	if err != nil {
		writeError(w, http.StatusBadRequest, "failed to read request body")
		return
	}

	var req UpdateIssueRequest
	if err := json.Unmarshal(bodyBytes, &req); err != nil {
		writeError(w, http.StatusBadRequest, "invalid request body")
		return
	}

	// Track which fields were explicitly present in JSON (even if null)
	var rawFields map[string]json.RawMessage
	json.Unmarshal(bodyBytes, &rawFields)

	if prevIssue.TriageState.Valid {
		if field := triageLockedField(rawFields); field != "" {
			writeIssueInTriage(w, field)
			return
		}
	}

	duplicateOfID, ok := parseDuplicateMark(w, &req, rawFields, prevIssue.ID)
	if !ok {
		return
	}

	// Pre-fill nullable fields (bare sqlc.narg) with current values
	params := db.UpdateIssueParams{
		SourceTaskID:  h.wakeupSourceTaskID(r),
		ID:            prevIssue.ID,
		AssigneeType:  prevIssue.AssigneeType,
		AssigneeID:    prevIssue.AssigneeID,
		StartDate:     prevIssue.StartDate,
		DueDate:       prevIssue.DueDate,
		ParentIssueID: prevIssue.ParentIssueID,
		ProjectID:     prevIssue.ProjectID,
		Stage:         prevIssue.Stage,
		// Checked against the locked rows in updateIssueWithStatusGuard.
		DuplicateOfIssueID: duplicateOfID,
	}
	if req.ExpectedRevision != nil {
		if *req.ExpectedRevision < 1 {
			writeError(w, http.StatusBadRequest, "expected_revision must be a positive integer")
			return
		}
		if prevIssue.Revision != *req.ExpectedRevision {
			writeRevisionConflict(w, "issue", prevIssue.ID, *req.ExpectedRevision, prevIssue.Revision)
			return
		}
		params.ExpectedRevision = pgtype.Int8{Int64: *req.ExpectedRevision, Valid: true}
	}

	// COALESCE fields — only set when explicitly provided
	if req.Title != nil {
		params.Title = pgtype.Text{String: *req.Title, Valid: true}
	}
	if req.Description != nil {
		params.Description = pgtype.Text{String: *req.Description, Valid: true}
	}
	// statusKeyForGuard is the resolved key when this request sets a status, and
	// empty otherwise. Empty means "this write does not touch status", which the
	// guard treats as nothing to protect.
	statusKeyForGuard := ""
	if req.Status != nil {
		statusKey, _, ok := h.resolveIssueStatusKeyKind(w, r, prevIssue.WorkspaceID, *req.Status)
		if !ok {
			return
		}
		statusKeyForGuard = statusKey
		params.Status = pgtype.Text{String: statusKey, Valid: true}
	}
	if req.Priority != nil {
		if !validateIssueEnum(w, "priority", *req.Priority, validIssuePriorities) {
			return
		}
		params.Priority = pgtype.Text{String: *req.Priority, Valid: true}
	}
	if req.Position != nil {
		params.Position = pgtype.Float8{Float64: *req.Position, Valid: true}
	}
	// Nullable fields — only override when explicitly present in JSON
	if _, ok := rawFields["assignee_type"]; ok {
		if req.AssigneeType != nil {
			params.AssigneeType = pgtype.Text{String: *req.AssigneeType, Valid: true}
		} else {
			params.AssigneeType = pgtype.Text{Valid: false} // explicit null = unassign
		}
	}
	if _, ok := rawFields["assignee_id"]; ok {
		if req.AssigneeID != nil {
			id, ok := parseUUIDOrBadRequest(w, *req.AssigneeID, "assignee_id")
			if !ok {
				return
			}
			params.AssigneeID = id
		} else {
			params.AssigneeID = pgtype.UUID{Valid: false} // explicit null = unassign
		}
	}
	if _, ok := rawFields["start_date"]; ok {
		if req.StartDate != nil && *req.StartDate != "" {
			d, err := util.ParseCalendarDate(*req.StartDate)
			if err != nil {
				writeError(w, http.StatusBadRequest, "invalid start_date format, expected YYYY-MM-DD")
				return
			}
			params.StartDate = d
		} else {
			params.StartDate = pgtype.Date{Valid: false} // explicit null = clear date
		}
	}
	if _, ok := rawFields["due_date"]; ok {
		if req.DueDate != nil && *req.DueDate != "" {
			d, err := util.ParseCalendarDate(*req.DueDate)
			if err != nil {
				writeError(w, http.StatusBadRequest, "invalid due_date format, expected YYYY-MM-DD")
				return
			}
			params.DueDate = d
		} else {
			params.DueDate = pgtype.Date{Valid: false} // explicit null = clear date
		}
	}
	if _, ok := rawFields["parent_issue_id"]; ok {
		if req.ParentIssueID != nil {
			newParentID, ok := parseUUIDOrBadRequest(w, *req.ParentIssueID, "parent_issue_id")
			if !ok {
				return
			}
			// Cannot set self as parent. Compare against prevIssue.ID (the
			// resolved entity), not the raw URL string — `id` may be an
			// identifier like "MUL-7".
			if newParentID == prevIssue.ID {
				writeError(w, http.StatusBadRequest, "an issue cannot be its own parent")
				return
			}
			// Validate parent exists in the same workspace.
			if _, err := h.Queries.GetIssueInWorkspace(r.Context(), db.GetIssueInWorkspaceParams{
				ID:          newParentID,
				WorkspaceID: prevIssue.WorkspaceID,
			}); err != nil {
				writeError(w, http.StatusBadRequest, "parent issue not found in this workspace")
				return
			}
			// Cycle detection: walk up from the new parent to ensure we don't reach this issue.
			cursor := newParentID
			for depth := 0; depth < 10; depth++ {
				ancestor, err := h.Queries.GetIssue(r.Context(), cursor)
				if err != nil || !ancestor.ParentIssueID.Valid {
					break
				}
				if ancestor.ParentIssueID == prevIssue.ID {
					writeError(w, http.StatusBadRequest, "circular parent relationship detected")
					return
				}
				cursor = ancestor.ParentIssueID
			}
			params.ParentIssueID = newParentID
		} else {
			params.ParentIssueID = pgtype.UUID{Valid: false} // explicit null = remove parent
		}
	}
	if _, ok := rawFields["project_id"]; ok {
		if req.ProjectID != nil {
			projectUUID, ok := parseUUIDOrBadRequest(w, *req.ProjectID, "project_id")
			if !ok {
				return
			}
			if _, err := h.Queries.GetProjectInWorkspace(r.Context(), db.GetProjectInWorkspaceParams{
				ID:          projectUUID,
				WorkspaceID: prevIssue.WorkspaceID,
			}); err != nil {
				if !isNotFound(err) {
					slog.Error("update issue: validate project scope",
						append(logger.RequestAttrs(r), "project_id", uuidToString(projectUUID), "error", err)...)
					writeError(w, http.StatusInternalServerError, "failed to validate project")
					return
				}
				writeError(w, http.StatusBadRequest, "project not found in this workspace")
				return
			}
			params.ProjectID = projectUUID
		} else {
			params.ProjectID = pgtype.UUID{Valid: false}
		}
	}
	if _, ok := rawFields["stage"]; ok {
		if req.Stage != nil {
			if *req.Stage < 1 {
				writeError(w, http.StatusBadRequest, "stage must be >= 1")
				return
			}
			params.Stage = pgtype.Int4{Int32: *req.Stage, Valid: true}
		} else {
			params.Stage = pgtype.Int4{Valid: false} // explicit null = unstage
		}
	}

	// Validate the resulting (assignee_type, assignee_id) pair when the caller
	// touches either field. Existing data on the issue is left alone if the
	// caller is not changing it.
	//
	// The scope is THIS issue: an unattributed autopilot run that verifiably owns
	// the work on it may point it at a private agent, exactly as it may when
	// creating a child under it. Before MUL-6691 this passed nil, so the reported
	// flow — create DRA-109 unassigned, then assign it — was refused even though
	// the identical lineage was accepted on the create path.
	_, touchedType := rawFields["assignee_type"]
	_, touchedID := rawFields["assignee_id"]

	// Project workflow (MUL-7420): keep the status inside the project's
	// workflow, and hand the issue off when it enters a step with a handler.
	// A handoff assignee passes the same gate as a hand-picked one.
	// An agent's run may only move the issue from the step it works on.
	callerType, _ := h.resolveActor(r, userID, workspaceID)
	callerRun := h.agentRunOnIssue(r, callerType, prevIssue)
	handoff, err := h.applyIssueWorkflow(r.Context(), prevIssue, &params, req.Status != nil, touchedType || touchedID, req.EnterWorkflowStep, callerRun)
	if err != nil {
		if writeIssueWorkflowError(w, err) {
			return
		}
		slog.Error("update issue: resolve project workflow", append(logger.RequestAttrs(r), "error", err)...)
		writeError(w, http.StatusInternalServerError, "failed to resolve project workflow")
		return
	}
	if params.Status.Valid && params.Status.String != statusKeyForGuard {
		// A project move remapped the status; guard the key actually written.
		statusKeyForGuard = params.Status.String
	}
	if touchedType || touchedID || handoff != nil {
		if status, msg := h.validateAssigneePair(r.Context(), r, workspaceID, params.AssigneeType, params.AssigneeID); status != 0 {
			writeError(w, status, msg)
			return
		}
	}

	attachmentIDs, ok := parseUUIDSliceOrBadRequest(w, req.AttachmentIDs, "attachment_ids")
	if !ok {
		return
	}

	// The agent this write hands the issue to gets the change as its own run,
	// not also through its event wakeups.
	writeCtx := withWakeupHandedTo(r.Context(), h.handedToAgent(r.Context(), prevIssue, params, handoff != nil))
	var issue db.Issue
	attachmentsChanged := false
	if req.Description != nil || req.TitleBase != nil || req.DescriptionBase != nil || len(attachmentIDs) > 0 {
		var lockedPrev db.Issue
		issue, lockedPrev, attachmentsChanged, err = h.updateIssueAtomically(
			writeCtx, prevIssue.WorkspaceID, params, rawFields, req.TitleBase, req.DescriptionBase, attachmentIDs, statusKeyForGuard, handoff != nil,
		)
		if lockedPrev.ID.Valid {
			prevIssue = lockedPrev
		}
	} else {
		issue, err = h.updateIssueWithStatusGuard(writeCtx, prevIssue.WorkspaceID, statusKeyForGuard, params)
	}
	if err != nil {
		if writeIssueStatusRaceError(w, err) {
			return
		}
		if writeDuplicateMarkError(w, err) {
			return
		}
		if errors.Is(err, errIssueFieldConflict) {
			writeEditConflict(w, "issue", prevIssue.ID)
			return
		}
		if errors.Is(err, pgx.ErrNoRows) && req.ExpectedRevision != nil {
			current, reloadErr := h.Queries.GetIssueInWorkspace(r.Context(), db.GetIssueInWorkspaceParams{ID: prevIssue.ID, WorkspaceID: prevIssue.WorkspaceID})
			if reloadErr == nil {
				writeRevisionConflict(w, "issue", current.ID, *req.ExpectedRevision, current.Revision)
				return
			}
		}
		slog.Warn("update issue failed", append(logger.RequestAttrs(r), "error", err, "issue_id", id, "workspace_id", workspaceID)...)
		writeError(w, http.StatusInternalServerError, "failed to update issue")
		return
	}

	// Determine actor identity: agent (via X-Agent-ID header) or member.
	actorType, actorID := h.resolveActor(r, userID, workspaceID)
	h.advanceRunStep(r.Context(), callerRun, prevIssue, issue)

	prefix := h.getIssuePrefix(r.Context(), issue.WorkspaceID)
	resp := issueToResponse(issue, prefix)
	slog.Info("issue updated", append(logger.RequestAttrs(r), "issue_id", id, "workspace_id", workspaceID)...)

	h.fillStatusCategory(r.Context(), issue.WorkspaceID, &resp)
	assigneeChanged := (req.AssigneeType != nil || req.AssigneeID != nil || handoff != nil) &&
		(prevIssue.AssigneeType.String != issue.AssigneeType.String || uuidToString(prevIssue.AssigneeID) != uuidToString(issue.AssigneeID))
	statusChanged := (req.Status != nil || params.Status.Valid) && prevIssue.Status != issue.Status
	priorityChanged := req.Priority != nil && prevIssue.Priority != issue.Priority
	// project_changed gates the client's per-project issue-list refetch the way
	// status/assignee flags gate theirs. Without it the client must diff
	// project_id against its own cache, which breaks once an optimistic local
	// move has overwritten the cached value (MUL-3669 / #4548).
	projectChanged := req.ProjectID != nil && uuidToString(prevIssue.ProjectID) != uuidToString(issue.ProjectID)
	descriptionChanged := req.Description != nil && textToPtr(prevIssue.Description) != resp.Description
	titleChanged := req.Title != nil && prevIssue.Title != issue.Title
	prevStartDate := dateToPtr(prevIssue.StartDate)
	startDateChanged := prevStartDate != resp.StartDate && (prevStartDate == nil) != (resp.StartDate == nil) ||
		(prevStartDate != nil && resp.StartDate != nil && *prevStartDate != *resp.StartDate)
	prevDueDate := dateToPtr(prevIssue.DueDate)
	dueDateChanged := prevDueDate != resp.DueDate && (prevDueDate == nil) != (resp.DueDate == nil) ||
		(prevDueDate != nil && resp.DueDate != nil && *prevDueDate != *resp.DueDate)

	// A workflow handoff (MUL-7420) is decided before the event goes out so
	// the timeline can record it as one entry: the status change, who the
	// issue went to, whether their run started, and the brief it reads. The
	// run gets that brief from its claim; its handoff note stays the
	// caller's own.
	handoffNote := req.HandoffNote
	trigger, willRun := h.IssueService.WillEnqueueRun(r.Context(),
		service.IssueTriggerInput{
			Issue:              issue,
			PrevStatus:         prevIssue.Status,
			AssigneeChanged:    assigneeChanged || handoff != nil,
			StatusChanged:      statusChanged,
			EnteredHandoffStep: handoff != nil,
		},
		h.issueTriggerWriteProbe(r, actorType, actorID, issue),
	)
	willRun = willRun && !req.SuppressRun

	updatedPayload := map[string]any{
		"issue":               resp,
		"assignee_changed":    assigneeChanged,
		"status_changed":      statusChanged,
		"priority_changed":    priorityChanged,
		"project_changed":     projectChanged,
		"start_date_changed":  startDateChanged,
		"due_date_changed":    dueDateChanged,
		"description_changed": descriptionChanged,
		"title_changed":       titleChanged,
		"prev_title":          prevIssue.Title,
		"prev_assignee_type":  textToPtr(prevIssue.AssigneeType),
		"prev_assignee_id":    uuidToPtr(prevIssue.AssigneeID),
		"prev_status":         prevIssue.Status,
		"prev_priority":       prevIssue.Priority,
		"prev_start_date":     prevStartDate,
		"prev_due_date":       prevDueDate,
		"prev_description":    textToPtr(prevIssue.Description),
		"creator_type":        prevIssue.CreatorType,
		"creator_id":          uuidToString(prevIssue.CreatorID),
		// Both ends of a mark change ride here: the activity log records it
		// and clients refresh the two issues' relations from it.
		"duplicate_of_issue_id":      liveDuplicateMark(issue.Status, issue.DuplicateOfIssueID),
		"prev_duplicate_of_issue_id": liveDuplicateMark(prevIssue.Status, prevIssue.DuplicateOfIssueID),
	}
	if handoff != nil {
		updatedPayload["workflow_handoff"] = workflowHandoffPayload(issue, handoff, h.workflowHandoffBrief(r.Context(), issue, handoff), willRun)
	}
	h.publish(protocol.EventIssueUpdated, workspaceID, actorType, actorID, updatedPayload)
	if attachmentsChanged {
		// The full owner snapshot must be admitted before an auxiliary event at
		// the same revision. Otherwise clients advance only the revision here and
		// reject issue:updated as non-increasing, stranding the old issue fields.
		h.publish(protocol.EventIssueAttachmentsChanged, workspaceID, actorType, actorID, map[string]any{
			"issue_id":       uuidToString(issue.ID),
			"issue_revision": issue.Revision,
		})
	}

	// Reconcile the task queue. Whether this write starts an agent run — and
	// for whom (agent assignee or squad leader) — is decided by the single
	// WillEnqueueRun predicate, shared verbatim with the preview endpoint so
	// the two never drift (MUL-3375).
	//
	// A reassignment intentionally does NOT cancel existing tasks on the issue
	// (#4963 / MUL-4113). The previous "cancel every active task on the issue"
	// was too coarse: it silently dropped unrelated in-flight work (a
	// mention-triggered run for another agent, a squad task) with no requeue,
	// and it self-cancelled a run that reassigned the issue from inside itself.
	// Ownership handoff no longer implies interruption; the new assignee's run,
	// if any, is enqueued by WillEnqueueRun below and runs alongside whatever
	// was already in flight. No status change — not even → cancelled — cancels
	// active tasks: a user clicking "cancel" on an issue has no expectation that
	// it stops in-flight agent runs, so that implicit coupling is gone
	// (MUL-4465). Deleting an issue still cancels its tasks (see DeleteIssue),
	// because the tasks' owning issue ceases to exist.
	//
	// A workflow handoff counts as a fresh assignment even when the step's
	// handler already owns the issue: entering the step is the handler's cue
	// to work on it, with the step's brief as the handoff note. (MUL-7420)
	if handoff != nil && req.StopPreviousAssigneeRuns {
		h.stopPreviousAssigneeRuns(r.Context(), prevIssue, issue, actorType, actorID)
	}
	if willRun {
		h.dispatchIssueRun(r.Context(), issue, trigger, actorType, actorID, handoffNote)
	}

	// Sub-issue rules on the parent (the child_done system rule and sub-issue
	// conditions): the write recorded the change; process it now. A failure
	// is retried by the scheduler sweep.
	if statusChanged || prevIssue.ParentIssueID != issue.ParentIssueID || prevIssue.Stage != issue.Stage {
		h.processChildEvents(r.Context(), issue.ParentIssueID, prevIssue.ParentIssueID)
	}

	writeJSON(w, http.StatusOK, resp)
}

// validateAssigneePair verifies the (assignee_type, assignee_id) pair refers
// to an existing entity in the workspace. For agent and squad assignees it
// also rejects archived targets and runs the INVOKE gate — canInvokeAgent, not
// the softer canAccessPrivateAgent view gate: assigning an issue produces a
// run, so it must clear the same predicate as chat / @-mention (MUL-3963).
// That means owner-only for a private agent, with NO workspace-admin bypass
// and NO unconditional agent-to-agent bypass — an agent caller (X-Agent-ID) is
// judged by the top-of-chain human originator like everywhere else.
// An autopilot run needs no special case here: since MUL-6951 a scheduled run
// carries its trigger owner's originator, so it is judged by exactly the same
// predicate as that human acting directly.
//
// Returns (statusCode, errorMessage). statusCode == 0 means the pair is valid;
// callers should treat any non-zero status as a rejection and surface it back
// to the client.
func (h *Handler) validateAssigneePair(ctx context.Context, r *http.Request, workspaceID string, assigneeType pgtype.Text, assigneeID pgtype.UUID) (int, string) {
	// Both unset → unassigned issue, valid.
	if !assigneeType.Valid && !assigneeID.Valid {
		return 0, ""
	}
	// Exactly one of type/id provided → callers must always pair them.
	if assigneeType.Valid != assigneeID.Valid {
		return http.StatusBadRequest, "assignee_type and assignee_id must be provided together"
	}
	wsUUID, err := util.ParseUUID(workspaceID)
	if err != nil {
		return http.StatusBadRequest, "invalid workspace_id"
	}
	switch assigneeType.String {
	case "member":
		if _, err := h.Queries.GetMemberByUserAndWorkspace(ctx, db.GetMemberByUserAndWorkspaceParams{
			UserID:      assigneeID,
			WorkspaceID: wsUUID,
		}); err != nil {
			return http.StatusBadRequest, "assignee_id does not refer to a member of this workspace"
		}
		return 0, ""
	case "agent":
		agent, err := h.Queries.GetAgentInWorkspace(ctx, db.GetAgentInWorkspaceParams{
			ID:          assigneeID,
			WorkspaceID: wsUUID,
		})
		if err != nil {
			return http.StatusBadRequest, "assignee_id does not refer to an agent of this workspace"
		}
		if agent.ArchivedAt.Valid {
			return http.StatusBadRequest, "cannot assign to archived agent"
		}
		actorType, actorID := h.resolveActor(r, requestUserID(r), workspaceID)
		effectiveInvoker := h.invokeOriginatorFromRequest(r, actorType, actorID)
		if !h.canInvokeAgent(ctx, agent, actorType, actorID, effectiveInvoker, workspaceID) {
			// Names the missing permission, not the target's configuration: the
			// old "private agent" wording both disclosed the agent's permission
			// mode and was simply wrong for a `public_to` agent scoped to
			// specific people. This is NOT full enumeration-safety — the
			// not-in-workspace branch above still answers 400 where this
			// answers 403, so existence remains observable; the guarantee here
			// is only that the reason no longer names the target's permission
			// mode (MUL-6380 / GH #7180).
			return http.StatusForbidden, "you do not have permission to assign work to this agent"
		}
		return 0, ""
	case "squad":
		squad, err := h.Queries.GetSquadInWorkspace(ctx, db.GetSquadInWorkspaceParams{
			ID:          assigneeID,
			WorkspaceID: wsUUID,
		})
		if err != nil {
			return http.StatusBadRequest, "assignee_id does not refer to a squad in this workspace"
		}
		if squad.ArchivedAt.Valid {
			return http.StatusBadRequest, "cannot assign to an archived squad"
		}
		leader, err := h.Queries.GetAgent(ctx, squad.LeaderID)
		if err != nil || leader.ArchivedAt.Valid {
			return http.StatusBadRequest, "squad leader is archived; cannot assign to this squad"
		}
		actorType, actorID := h.resolveActor(r, requestUserID(r), workspaceID)
		effectiveInvoker := h.invokeOriginatorFromRequest(r, actorType, actorID)
		if !h.canInvokeAgent(ctx, leader, actorType, actorID, effectiveInvoker, workspaceID) {
			// Same wording rule as the agent branch above; "this squad"
			// avoids disclosing the leader agent's permission mode.
			return http.StatusForbidden, "you do not have permission to assign work to this squad"
		}
		return 0, ""
	default:
		return http.StatusBadRequest, "assignee_type must be 'member', 'agent', or 'squad'"
	}
}

// shouldEnqueueAgentTask returns true when an issue creation or assignment
// should trigger the assigned agent. Backlog issues are skipped — backlog
// acts as a parking lot where issues can be pre-assigned without immediately
// triggering execution. Moving out of backlog is handled separately in
// UpdateIssue.
func (h *Handler) shouldEnqueueAgentTask(ctx context.Context, issue db.Issue) bool {
	// Only the fixed backlog key parks work; custom unstarted statuses do not.
	// An issue in Triage is not parked but refused: it produces no run from any
	// entry point until it is accepted (MUL-7189 §2.3).
	if issue.TriageState.Valid || issuestatus.Effective(ctx, h.Queries, issue.WorkspaceID, issue.Status) == "backlog" {
		return false
	}
	return h.isAgentAssigneeReady(ctx, issue)
}

// shouldEnqueueAssigneeFallback returns true when comment routing can fall back
// to the issue's assigned agent. Fires for any status — comments are
// conversational and can happen at any stage, including after completion
// (e.g. follow-up questions on a done issue).
//
// Mirrors the private-agent gate that resolveMentionedAgentCommentTriggers applies on the
// @mention path: once an owner/admin assigns a private agent to an issue, the
// agent's UUID is "welded" onto the issue and remains visible to every member
// who can view it. Without this check any of those members could dispatch a new
// task to the private agent simply by commenting (#3300).
func (h *Handler) shouldEnqueueAssigneeFallback(ctx context.Context, issue db.Issue, actorType, actorID string, opts commentTriggerComputeOptions) bool {
	_, hasPending, ok := h.assigneeFallbackAgent(ctx, issue, actorType, actorID, opts)
	return ok && !hasPending
}

func (h *Handler) assigneeFallbackAgent(ctx context.Context, issue db.Issue, actorType, actorID string, opts commentTriggerComputeOptions) (db.Agent, bool, bool) {
	if !issue.AssigneeType.Valid || issue.AssigneeType.String != "agent" || !issue.AssigneeID.Valid {
		return db.Agent{}, false, false
	}
	agent, err := h.Queries.GetAgent(ctx, issue.AssigneeID)
	if err != nil || !agent.RuntimeID.Valid || agent.ArchivedAt.Valid {
		return db.Agent{}, false, false
	}
	if !h.canInvokeAgent(ctx, agent, actorType, actorID, opts.OriginatorUserID, uuidToString(issue.WorkspaceID)) {
		return db.Agent{}, false, false
	}
	// Coalescing queue: pending is still a valid route target, but callers
	// that actually enqueue tasks use this flag to avoid piling on duplicates.
	hasPending, err := h.hasPendingTaskForIssueAndAgent(ctx, issue.ID, issue.AssigneeID, opts)
	if err != nil {
		return db.Agent{}, false, false
	}
	return agent, hasPending, true
}

// isAgentRunningOnIssue reports whether the calling agent's current task
// (identified by X-Task-ID) is running for the exact issue being promoted.
// That is the only true self-loop on backlog→active: the agent flipping
// the same issue its own task is executing for would immediately re-enqueue
// itself, complete the run, flip again, and so on.
//
// Same-agent cross-issue handoff (Agent A finishing a task on issue I1 then
// promoting issue I2 — even when I2 is also assigned to A) is NOT a loop
// and must fire; that is the documented serial sub-task chain. Member
// actors never match.
//
// X-Task-ID is guaranteed to be present and consistent when actorType is
// "agent": resolveActor demotes the actor to "member" otherwise (handler.go
// resolveActor). We still recheck defensively — a future caller could pass
// agent identity through a different path.
func (h *Handler) isAgentRunningOnIssue(r *http.Request, actorType string, issue db.Issue) bool {
	if actorType != "agent" {
		return false
	}
	taskIDStr := r.Header.Get("X-Task-ID")
	if taskIDStr == "" {
		return false
	}
	taskUUID, err := util.ParseUUID(taskIDStr)
	if err != nil {
		return false
	}
	task, err := h.Queries.GetAgentTask(r.Context(), taskUUID)
	if err != nil {
		return false
	}
	if !task.IssueID.Valid {
		return false
	}
	return uuidToString(task.IssueID) == uuidToString(issue.ID)
}

// isAgentAssigneeReady checks if an issue is assigned to an active agent
// with a valid runtime.
func (h *Handler) isAgentAssigneeReady(ctx context.Context, issue db.Issue) bool {
	if !issue.AssigneeType.Valid || issue.AssigneeType.String != "agent" || !issue.AssigneeID.Valid {
		return false
	}

	agent, err := h.Queries.GetAgent(ctx, issue.AssigneeID)
	if err != nil {
		return false
	}
	// The shared verdict, not a local re-check (service.AgentReadiness). Only a
	// BLOCKED verdict stops the enqueue: an offline machine still queues,
	// because that work runs when the machine comes back.
	verdict, err := service.AgentReadiness(ctx, h.runtimeLookup(obsmetrics.RuntimeLookupSourceIssue), agent)
	if err != nil || !verdict.Blocked() {
		return err == nil
	}
	// Assignment has no response the assigner reads for this outcome, so a
	// refusal that needs human repair leaves the explanation on the issue
	// (MUL-6164). An unbound agent keeps its silent skip: the agent list
	// already shows it has no runtime, and nothing about it is new here.
	if service.RuntimeBlockedNeedsNotice(verdict.Reason) {
		h.noteRuntimeUnusable(ctx, issue, agent, verdict)
	}
	return false
}

func (h *Handler) DeleteIssue(w http.ResponseWriter, r *http.Request) {
	r = h.withWakeupActor(r)
	id := chi.URLParam(r, "id")
	issue, ok := h.loadIssueForUser(w, r, id)
	if !ok {
		return
	}

	h.TaskService.CancelTasksForIssue(r.Context(), issue.ID)
	// Fail any linked autopilot runs before delete (ON DELETE SET NULL clears issue_id).
	_ = h.AutopilotService.FailAutopilotRunsByIssue(r.Context(), issue.ID)

	deleteResult, err := h.deleteIssueAndCollectAttachmentURLs(r.Context(), issue, nil)
	if err != nil {
		writeError(w, http.StatusInternalServerError, "failed to delete issue")
		return
	}

	h.deleteS3Objects(r.Context(), deleteResult.AttachmentURLs)
	userID := requestUserID(r)
	actorType, actorID := h.resolveActor(r, userID, uuidToString(issue.WorkspaceID))
	// Always emit the resolved UUID — frontend caches key by UUID, so an
	// identifier-style payload ("MUL-123") would leave stale entries on
	// other clients after an identifier-path delete.
	resolvedID := uuidToString(issue.ID)
	h.publish(protocol.EventIssueDeleted, uuidToString(issue.WorkspaceID), actorType, actorID, map[string]any{"issue_id": resolvedID})
	h.publishIssueSnapshots(r.Context(), deleteResult.DetachedChildren, actorType, actorID)
	h.publishClearedDuplicates(r.Context(), deleteResult.ClearedDuplicates, actorType, actorID)
	slog.Info("issue deleted", append(logger.RequestAttrs(r), "issue_id", resolvedID, "workspace_id", uuidToString(issue.WorkspaceID))...)
	w.WriteHeader(http.StatusNoContent)
}

// deleteIssueAndCollectAttachmentURLs serializes issue deletion with channel
// media binding. The delete-side FOR UPDATE conflicts with the binder's
// FOR KEY SHARE, and URL collection happens only after that lock is held:
// bind-first means the new URL is collected; delete-first means the bind rolls
// back without consuming its durable object intent.
type issueDeleteResult struct {
	AttachmentURLs   []string
	DetachedChildren []db.Issue
	// ClearedDuplicates lost their duplicate mark because their original was
	// deleted. They stay cancelled.
	ClearedDuplicates []clearedDuplicate
}

// clearedDuplicate pairs a duplicate whose mark a delete cleared with the
// original that delete removed, so the broadcast can still name it.
type clearedDuplicate struct {
	Issue    db.Issue
	Original db.Issue
}

func (h *Handler) deleteIssueAndCollectAttachmentURLs(ctx context.Context, issue db.Issue, excludedIssueIDs []pgtype.UUID) (issueDeleteResult, error) {
	return h.deleteIssuesAndCollectAttachmentURLs(ctx, []db.Issue{issue}, excludedIssueIDs)
}

func (h *Handler) deleteIssuesAndCollectAttachmentURLs(ctx context.Context, issues []db.Issue, excludedIssueIDs []pgtype.UUID) (issueDeleteResult, error) {
	sort.Slice(issues, func(i, j int) bool {
		return uuidToString(issues[i].ID) < uuidToString(issues[j].ID)
	})
	tx, err := h.beginWakeupWrite(ctx)
	if err != nil {
		return issueDeleteResult{}, fmt.Errorf("begin issue delete: %w", err)
	}
	defer tx.Rollback(ctx)
	qtx := h.Queries.WithTx(tx)

	result := issueDeleteResult{}
	for _, issue := range issues {
		if _, err := qtx.LockIssueForDelete(ctx, db.LockIssueForDeleteParams{
			ID: issue.ID, WorkspaceID: issue.WorkspaceID,
		}); err != nil {
			return issueDeleteResult{}, fmt.Errorf("lock issue for delete: %w", err)
		}
		detached, err := qtx.DetachDirectChildIssues(ctx, db.DetachDirectChildIssuesParams{
			WorkspaceID: issue.WorkspaceID, ParentIssueID: issue.ID, ExcludedIssueIds: excludedIssueIDs,
		})
		if err != nil {
			return issueDeleteResult{}, fmt.Errorf("detach child issues: %w", err)
		}
		result.DetachedChildren = append(result.DetachedChildren, detached...)
		cleared, err := qtx.ClearIssueDuplicatesOf(ctx, db.ClearIssueDuplicatesOfParams{
			WorkspaceID: issue.WorkspaceID, IssueID: issue.ID, ExcludedIssueIds: excludedIssueIDs,
		})
		if err != nil {
			return issueDeleteResult{}, fmt.Errorf("clear duplicate marks: %w", err)
		}
		for _, duplicate := range cleared {
			result.ClearedDuplicates = append(result.ClearedDuplicates, clearedDuplicate{Issue: duplicate, Original: issue})
		}
		attachmentURLs, err := qtx.ListAttachmentURLsByIssueOrComments(ctx, issue.ID)
		if err != nil {
			return issueDeleteResult{}, fmt.Errorf("list issue attachment URLs: %w", err)
		}
		result.AttachmentURLs = append(result.AttachmentURLs, attachmentURLs...)
		if sourceContext, contextErr := qtx.GetIssueSourceContextByIssue(ctx, db.GetIssueSourceContextByIssueParams{WorkspaceID: issue.WorkspaceID, IssueID: issue.ID}); contextErr == nil {
			contextAttachments, listErr := qtx.ListAttachmentsBySourceContext(ctx, db.ListAttachmentsBySourceContextParams{
				WorkspaceID: issue.WorkspaceID, SourceContextID: sourceContext.ID,
			})
			if listErr != nil {
				return issueDeleteResult{}, fmt.Errorf("list source context attachments: %w", listErr)
			}
			// The storage interface's legacy bulk delete has no result channel. Keep
			// durable object intents before removing DB ownership so the periodic
			// reconciler retries any ambiguous or failed best-effort delete.
			if h.Storage != nil {
				for _, clone := range contextAttachments {
					storageKey := h.Storage.KeyFromURL(clone.Url)
					if storageKey == "" {
						continue
					}
					if _, intentErr := qtx.RecordSourceContextDeletionObjectIntent(ctx, db.RecordSourceContextDeletionObjectIntentParams{
						StorageKey: storageKey, WorkspaceID: issue.WorkspaceID, SourceContextID: sourceContext.ID,
						AttachmentID: clone.ID, ObjectUrl: clone.Url,
					}); intentErr != nil {
						return issueDeleteResult{}, fmt.Errorf("record source context delete intent: %w", intentErr)
					}
				}
			}
			clones, cloneErr := qtx.DeleteAttachmentsBySourceContext(ctx, db.DeleteAttachmentsBySourceContextParams{WorkspaceID: issue.WorkspaceID, SourceContextID: sourceContext.ID})
			if cloneErr != nil {
				return issueDeleteResult{}, fmt.Errorf("delete source context attachments: %w", cloneErr)
			}
			for _, clone := range clones {
				result.AttachmentURLs = append(result.AttachmentURLs, clone.Url)
			}
			if _, contextErr = qtx.DeleteIssueSourceContextByIssue(ctx, db.DeleteIssueSourceContextByIssueParams{WorkspaceID: issue.WorkspaceID, IssueID: issue.ID}); contextErr != nil {
				return issueDeleteResult{}, fmt.Errorf("delete issue source context: %w", contextErr)
			}
		} else if !errors.Is(contextErr, pgx.ErrNoRows) {
			return issueDeleteResult{}, fmt.Errorf("load issue source context for delete: %w", contextErr)
		}
		if err := qtx.DeleteIssue(ctx, db.DeleteIssueParams{ID: issue.ID, WorkspaceID: issue.WorkspaceID}); err != nil {
			return issueDeleteResult{}, fmt.Errorf("delete issue: %w", err)
		}
	}
	if err := tx.Commit(ctx); err != nil {
		return issueDeleteResult{}, fmt.Errorf("commit issue delete: %w", err)
	}
	return result, nil
}

// publishIssueSnapshots broadcasts issue:updated for issues a delete rewrote:
// detached children and duplicates whose original was deleted.
func (h *Handler) publishIssueSnapshots(ctx context.Context, issues []db.Issue, actorType, actorID string) {
	for _, issue := range issues {
		response := issueToResponse(issue, h.getIssuePrefix(ctx, issue.WorkspaceID))
		h.fillStatusCategory(ctx, issue.WorkspaceID, &response)
		h.publish(protocol.EventIssueUpdated, uuidToString(issue.WorkspaceID), actorType, actorID, map[string]any{"issue": response})
	}
}

// publishClearedDuplicates broadcasts issue:updated for duplicates whose
// original a delete removed. The pointer went with the original, so the
// payload names both ends the way UpdateIssue does, plus the identifier the
// activity log can no longer look up (MUL-7349).
func (h *Handler) publishClearedDuplicates(ctx context.Context, cleared []clearedDuplicate, actorType, actorID string) {
	for _, c := range cleared {
		prefix := h.getIssuePrefix(ctx, c.Issue.WorkspaceID)
		response := issueToResponse(c.Issue, prefix)
		h.fillStatusCategory(ctx, c.Issue.WorkspaceID, &response)
		h.publish(protocol.EventIssueUpdated, uuidToString(c.Issue.WorkspaceID), actorType, actorID, map[string]any{
			"issue":                        response,
			"duplicate_of_issue_id":        liveDuplicateMark(c.Issue.Status, c.Issue.DuplicateOfIssueID),
			"prev_duplicate_of_issue_id":   liveDuplicateMark(c.Issue.Status, c.Original.ID),
			"prev_duplicate_of_identifier": prefix + "-" + strconv.Itoa(int(c.Original.Number)),
		})
	}
}

// Batch operations

type BatchUpdateIssuesRequest struct {
	IssueIDs []string           `json:"issue_ids"`
	Updates  UpdateIssueRequest `json:"updates"`
}

func (h *Handler) BatchUpdateIssues(w http.ResponseWriter, r *http.Request) {
	r = h.withWakeupActor(r)
	bodyBytes, err := io.ReadAll(r.Body)
	if err != nil {
		writeError(w, http.StatusBadRequest, "failed to read request body")
		return
	}

	var req BatchUpdateIssuesRequest
	if err := json.Unmarshal(bodyBytes, &req); err != nil {
		writeError(w, http.StatusBadRequest, "invalid request body")
		return
	}

	if len(req.IssueIDs) == 0 {
		writeError(w, http.StatusBadRequest, "issue_ids is required")
		return
	}

	userID, ok := requireUserID(w, r)
	if !ok {
		return
	}

	// Detect which fields in "updates" were explicitly set (including null).
	var rawTop map[string]json.RawMessage
	json.Unmarshal(bodyBytes, &rawTop)
	var rawUpdates map[string]json.RawMessage
	if raw, exists := rawTop["updates"]; exists {
		json.Unmarshal(raw, &rawUpdates)
	}
	if _, ok := rawUpdates["duplicate_of_issue_id"]; ok {
		writeError(w, http.StatusBadRequest, "duplicate_of_issue_id is not supported in batch updates")
		return
	}

	// Short-circuit when no mutation field is present in `updates`. Without
	// this, the loop below runs N no-op UPDATEs (every if-guard skips, every
	// COALESCE preserves the existing value) and reports `{"updated": N}` —
	// the response cheerfully claims success while nothing changed. Most
	// real-world cases that hit this path are caller mistakes (status placed
	// at the top level, "update" misspelled as singular). Telling the truth
	// here — `{"updated": 0}` — keeps the wire shape stable while making the
	// count match reality. See multica-ai/multica#1660.
	hasMutation := req.Updates.Title != nil ||
		req.Updates.Description != nil ||
		req.Updates.Status != nil ||
		req.Updates.Priority != nil ||
		req.Updates.Position != nil
	if !hasMutation {
		for _, k := range []string{"assignee_type", "assignee_id", "start_date", "due_date", "parent_issue_id", "project_id", "stage"} {
			if _, ok := rawUpdates[k]; ok {
				hasMutation = true
				break
			}
		}
	}
	if !hasMutation {
		writeJSON(w, http.StatusOK, map[string]any{"updated": 0})
		return
	}
	if req.Updates.Priority != nil {
		if !validateIssueEnum(w, "priority", *req.Updates.Priority, validIssuePriorities) {
			return
		}
	}

	workspaceID := h.resolveWorkspaceID(r)
	wsUUID, ok := parseUUIDOrBadRequest(w, workspaceID, "workspace_id")
	if !ok {
		return
	}
	// Status is validated against this workspace's catalog, so it has to wait
	// for wsUUID above. One check for the whole batch — every issue in it
	// shares the workspace — and a rejection rather than a silent skip, so a
	// bad status cannot report `{"updated": N}`. (MUL-6243)
	batchStatusKey := ""
	if req.Updates.Status != nil {
		batchStatusKey, _, ok = h.resolveIssueStatusKeyKind(w, r, wsUUID, *req.Updates.Status)
		if !ok {
			return
		}
	}
	if !h.validateBatchTriageLocks(w, r, wsUUID, req.IssueIDs, rawUpdates) {
		return
	}
	// The batch shares one project_id, so it is checked once here rather than
	// per issue, and rejected instead of skipped like the per-item guards in
	// the loop: a foreign project invalidates the whole request.
	batchProjectID := pgtype.UUID{Valid: false}
	if _, ok := rawUpdates["project_id"]; ok && req.Updates.ProjectID != nil {
		projectUUID, ok := parseUUIDOrBadRequest(w, *req.Updates.ProjectID, "project_id")
		if !ok {
			return
		}
		if _, err := h.Queries.GetProjectInWorkspace(r.Context(), db.GetProjectInWorkspaceParams{
			ID:          projectUUID,
			WorkspaceID: wsUUID,
		}); err != nil {
			if !isNotFound(err) {
				slog.Error("batch update issues: validate project scope",
					append(logger.RequestAttrs(r), "project_id", uuidToString(projectUUID), "error", err)...)
				writeError(w, http.StatusInternalServerError, "failed to validate project")
				return
			}
			writeError(w, http.StatusBadRequest, "project not found in this workspace")
			return
		}
		batchProjectID = projectUUID
	}

	updated := 0
	// One Resolver for the whole batch — a per-issue filler would query the
	// catalog once per custom-status row. (MUL-6243)
	fillBatch := h.newStatusCategoryFiller(r.Context(), wsUUID)
	// Parents whose sub-issues changed this batch. Their rules are evaluated
	// once after the loop against the final state (MUL-4155), not per child
	// against a mid-batch snapshot.
	var changedParents []pgtype.UUID
	for _, issueID := range req.IssueIDs {
		issueUUID, err := util.ParseUUID(issueID)
		if err != nil {
			continue
		}
		prevIssue, err := h.Queries.GetIssueInWorkspace(r.Context(), db.GetIssueInWorkspaceParams{
			ID:          issueUUID,
			WorkspaceID: wsUUID,
		})
		if err != nil {
			continue
		}

		params := db.UpdateIssueParams{
			SourceTaskID:  h.wakeupSourceTaskID(r),
			ID:            prevIssue.ID,
			AssigneeType:  prevIssue.AssigneeType,
			AssigneeID:    prevIssue.AssigneeID,
			StartDate:     prevIssue.StartDate,
			DueDate:       prevIssue.DueDate,
			ParentIssueID: prevIssue.ParentIssueID,
			ProjectID:     prevIssue.ProjectID,
			Stage:         prevIssue.Stage,
		}

		if req.Updates.Title != nil {
			params.Title = pgtype.Text{String: *req.Updates.Title, Valid: true}
		}
		if req.Updates.Description != nil {
			params.Description = pgtype.Text{String: *req.Updates.Description, Valid: true}
		}
		if req.Updates.Status != nil {
			params.Status = pgtype.Text{String: batchStatusKey, Valid: true}
		}
		if req.Updates.Priority != nil {
			params.Priority = pgtype.Text{String: *req.Updates.Priority, Valid: true}
		}
		if req.Updates.Position != nil {
			params.Position = pgtype.Float8{Float64: *req.Updates.Position, Valid: true}
		}
		if _, ok := rawUpdates["assignee_type"]; ok {
			if req.Updates.AssigneeType != nil {
				params.AssigneeType = pgtype.Text{String: *req.Updates.AssigneeType, Valid: true}
			} else {
				params.AssigneeType = pgtype.Text{Valid: false}
			}
		}
		if _, ok := rawUpdates["assignee_id"]; ok {
			if req.Updates.AssigneeID != nil {
				assigneeUUID, err := util.ParseUUID(*req.Updates.AssigneeID)
				if err != nil {
					continue
				}
				params.AssigneeID = assigneeUUID
			} else {
				params.AssigneeID = pgtype.UUID{Valid: false}
			}
		}
		if _, ok := rawUpdates["start_date"]; ok {
			if req.Updates.StartDate != nil && *req.Updates.StartDate != "" {
				d, err := util.ParseCalendarDate(*req.Updates.StartDate)
				if err != nil {
					continue
				}
				params.StartDate = d
			} else {
				params.StartDate = pgtype.Date{Valid: false}
			}
		}
		if _, ok := rawUpdates["due_date"]; ok {
			if req.Updates.DueDate != nil && *req.Updates.DueDate != "" {
				d, err := util.ParseCalendarDate(*req.Updates.DueDate)
				if err != nil {
					continue
				}
				params.DueDate = d
			} else {
				params.DueDate = pgtype.Date{Valid: false}
			}
		}

		if _, ok := rawUpdates["parent_issue_id"]; ok {
			if req.Updates.ParentIssueID != nil {
				newParentID, err := util.ParseUUID(*req.Updates.ParentIssueID)
				if err != nil {
					continue
				}
				// Cannot set self as parent.
				if newParentID == prevIssue.ID {
					continue
				}
				// Validate parent exists in the same workspace.
				if _, err := h.Queries.GetIssueInWorkspace(r.Context(), db.GetIssueInWorkspaceParams{
					ID:          newParentID,
					WorkspaceID: prevIssue.WorkspaceID,
				}); err != nil {
					continue
				}
				// Cycle detection: walk up from the new parent to ensure we don't reach this issue.
				cycleDetected := false
				cursor := newParentID
				for depth := 0; depth < 10; depth++ {
					ancestor, err := h.Queries.GetIssue(r.Context(), cursor)
					if err != nil || !ancestor.ParentIssueID.Valid {
						break
					}
					if ancestor.ParentIssueID == prevIssue.ID {
						cycleDetected = true
						break
					}
					cursor = ancestor.ParentIssueID
				}
				if cycleDetected {
					continue
				}
				params.ParentIssueID = newParentID
			} else {
				params.ParentIssueID = pgtype.UUID{Valid: false}
			}
		}
		if _, ok := rawUpdates["project_id"]; ok {
			// Resolved before the loop; an explicit null stays invalid and clears.
			params.ProjectID = batchProjectID
		}
		if _, ok := rawUpdates["stage"]; ok {
			if req.Updates.Stage != nil {
				if *req.Updates.Stage < 1 {
					continue
				}
				params.Stage = pgtype.Int4{Int32: *req.Updates.Stage, Valid: true}
			} else {
				params.Stage = pgtype.Int4{Valid: false} // explicit null = unstage
			}
		}

		// Validate the resulting assignee pair when this batch update touches
		// either assignee field. Skip the issue silently on failure.
		//
		// Scoped PER ISSUE (prevIssue is this iteration's row), so one bound
		// issue in the batch can never lend its authority to the others: an
		// unbound entry simply fails the check and is skipped. This IS a real
		// agent-reachable authorization point — a task token authenticates as its
		// bound workspace member, so requireUserID above is satisfied and
		// resolveActor still classifies the caller as an agent (MUL-6691).
		_, batchTouchedType := rawUpdates["assignee_type"]
		_, batchTouchedID := rawUpdates["assignee_id"]
		// Project workflow (MUL-7420), per issue: an issue whose workflow does
		// not list the batch status is skipped like the other per-item guards;
		// entering a handoff step reassigns this issue to the step's handler.
		callerType, _ := h.resolveActor(r, userID, workspaceID)
		callerRun := h.agentRunOnIssue(r, callerType, prevIssue)
		handoff, err := h.applyIssueWorkflow(r.Context(), prevIssue, &params, req.Updates.Status != nil, batchTouchedType || batchTouchedID, false, callerRun)
		if err != nil {
			continue
		}
		issueStatusGuard := batchStatusKey
		if params.Status.Valid {
			issueStatusGuard = params.Status.String
		}
		if batchTouchedType || batchTouchedID || handoff != nil {
			if status, _ := h.validateAssigneePair(r.Context(), r, workspaceID, params.AssigneeType, params.AssigneeID); status != 0 {
				continue
			}
		}

		writeCtx := withWakeupHandedTo(r.Context(), h.handedToAgent(r.Context(), prevIssue, params, handoff != nil))
		var issue db.Issue
		if req.Updates.Description != nil {
			// One batch-level base cannot describe multiple issue documents.
			// Preserve every marked channel-media block conservatively, matching
			// legacy single-update clients that omit description_base.
			var lockedPrev db.Issue
			issue, lockedPrev, _, err = h.updateIssueAtomically(
				writeCtx, prevIssue.WorkspaceID, params, rawUpdates, nil, nil, nil, issueStatusGuard, handoff != nil,
			)
			if err == nil {
				prevIssue = lockedPrev
			}
		} else {
			issue, err = h.updateIssueWithStatusGuard(writeCtx, wsUUID, issueStatusGuard, params)
		}
		if err != nil {
			// The archive race is a property of the batch's shared target
			// status, not of one issue, so every remaining item would fail the
			// same way. Abort with 409 instead of reporting a partial update.
			if writeIssueStatusRaceError(w, err) {
				return
			}
			slog.Warn("batch update issue failed", "issue_id", issueID, "error", err)
			continue
		}

		prefix := h.getIssuePrefix(r.Context(), issue.WorkspaceID)
		resp := issueToResponse(issue, prefix)
		actorType, actorID := h.resolveActor(r, userID, workspaceID)
		h.advanceRunStep(r.Context(), callerRun, prevIssue, issue)

		fillBatch(&resp)
		assigneeChanged := (req.Updates.AssigneeType != nil || req.Updates.AssigneeID != nil || handoff != nil) &&
			(prevIssue.AssigneeType.String != issue.AssigneeType.String || uuidToString(prevIssue.AssigneeID) != uuidToString(issue.AssigneeID))
		statusChanged := (req.Updates.Status != nil || params.Status.Valid) && prevIssue.Status != issue.Status
		priorityChanged := req.Updates.Priority != nil && prevIssue.Priority != issue.Priority
		projectChanged := req.Updates.ProjectID != nil && uuidToString(prevIssue.ProjectID) != uuidToString(issue.ProjectID)

		batchHandoffNote := req.Updates.HandoffNote
		trigger, willRun := h.IssueService.WillEnqueueRun(r.Context(),
			service.IssueTriggerInput{
				Issue:              issue,
				PrevStatus:         prevIssue.Status,
				AssigneeChanged:    assigneeChanged || handoff != nil,
				StatusChanged:      statusChanged,
				EnteredHandoffStep: handoff != nil,
			},
			h.issueTriggerWriteProbe(r, actorType, actorID, issue),
		)
		willRun = willRun && !req.Updates.SuppressRun
		batchPayload := map[string]any{
			"issue":                      resp,
			"assignee_changed":           assigneeChanged,
			"status_changed":             statusChanged,
			"priority_changed":           priorityChanged,
			"project_changed":            projectChanged,
			"prev_status":                prevIssue.Status,
			"duplicate_of_issue_id":      liveDuplicateMark(issue.Status, issue.DuplicateOfIssueID),
			"prev_duplicate_of_issue_id": liveDuplicateMark(prevIssue.Status, prevIssue.DuplicateOfIssueID),
		}
		if handoff != nil {
			batchPayload["workflow_handoff"] = workflowHandoffPayload(issue, handoff, h.workflowHandoffBrief(r.Context(), issue, handoff), willRun)
		}
		h.publish(protocol.EventIssueUpdated, workspaceID, actorType, actorID, batchPayload)

		// Reassignment does not cancel existing tasks (#4963 / MUL-4113) —
		// mirrors UpdateIssue. See that handler for the rationale.
		//
		// Same single predicate as UpdateIssue — batch must not grow its own
		// copy of the enqueue rule (the historical source of four-entry-point
		// drift, MUL-3375). suppress_run applies batch-wide.
		if handoff != nil && req.Updates.StopPreviousAssigneeRuns {
			h.stopPreviousAssigneeRuns(r.Context(), prevIssue, issue, actorType, actorID)
		}
		if willRun {
			h.dispatchIssueRun(r.Context(), issue, trigger, actorType, actorID, batchHandoffNote)
		}

		// No status change — not even → cancelled — cancels active tasks here,
		// mirroring UpdateIssue (MUL-4465). See that handler for the rationale.

		if statusChanged || prevIssue.ParentIssueID != issue.ParentIssueID || prevIssue.Stage != issue.Stage {
			changedParents = append(changedParents, issue.ParentIssueID, prevIssue.ParentIssueID)
		}

		updated++
	}

	// Each write recorded its sub-issue change; processing claims them per
	// parent, so each parent is evaluated once against the batch's final
	// state, independent of issue_ids order (MUL-4155). A failure is retried
	// by the sweep.
	h.processChildEvents(r.Context(), changedParents...)

	slog.Info("batch update issues", append(logger.RequestAttrs(r), "count", updated)...)
	writeJSON(w, http.StatusOK, map[string]any{"updated": updated})
}

type BatchDeleteIssuesRequest struct {
	IssueIDs []string `json:"issue_ids"`
}

func (h *Handler) BatchDeleteIssues(w http.ResponseWriter, r *http.Request) {
	var req BatchDeleteIssuesRequest
	if err := json.NewDecoder(r.Body).Decode(&req); err != nil {
		writeError(w, http.StatusBadRequest, "invalid request body")
		return
	}

	if len(req.IssueIDs) == 0 {
		writeError(w, http.StatusBadRequest, "issue_ids is required")
		return
	}

	userID, ok := requireUserID(w, r)
	if !ok {
		return
	}

	workspaceID := h.resolveWorkspaceID(r)
	wsUUID, ok := parseUUIDOrBadRequest(w, workspaceID, "workspace_id")
	if !ok {
		return
	}
	issues := make([]db.Issue, 0, len(req.IssueIDs))
	excludedIDs := make([]pgtype.UUID, 0, len(req.IssueIDs))
	seenIssueIDs := make(map[pgtype.UUID]struct{}, len(req.IssueIDs))
	for _, issueID := range req.IssueIDs {
		issueUUID, err := util.ParseUUID(issueID)
		if err != nil {
			continue
		}
		if _, duplicate := seenIssueIDs[issueUUID]; duplicate {
			continue
		}
		issue, err := h.Queries.GetIssueInWorkspace(r.Context(), db.GetIssueInWorkspaceParams{
			ID:          issueUUID,
			WorkspaceID: wsUUID,
		})
		if err != nil {
			continue
		}

		seenIssueIDs[issueUUID] = struct{}{}
		issues = append(issues, issue)
		excludedIDs = append(excludedIDs, issue.ID)
		h.TaskService.CancelTasksForIssue(r.Context(), issue.ID)
		_ = h.AutopilotService.FailAutopilotRunsByIssue(r.Context(), issue.ID)
	}
	deleteResult, err := h.deleteIssuesAndCollectAttachmentURLs(r.Context(), issues, excludedIDs)
	if err != nil {
		slog.Warn("batch delete issues failed", "error", err)
		writeError(w, http.StatusInternalServerError, "failed to delete issues")
		return
	}
	h.deleteS3Objects(r.Context(), deleteResult.AttachmentURLs)
	actorType, actorID := h.resolveActor(r, userID, workspaceID)
	for _, issue := range issues {
		h.publish(protocol.EventIssueDeleted, workspaceID, actorType, actorID, map[string]any{"issue_id": uuidToString(issue.ID)})
	}
	h.publishIssueSnapshots(r.Context(), deleteResult.DetachedChildren, actorType, actorID)
	h.publishClearedDuplicates(r.Context(), deleteResult.ClearedDuplicates, actorType, actorID)
	deleted := len(issues)

	slog.Info("batch delete issues", append(logger.RequestAttrs(r), "count", deleted)...)
	writeJSON(w, http.StatusOK, map[string]any{"deleted": deleted})
}