multica/server/pkg/db/queries/agent.sql

-- name: ListAgents :many
SELECT * FROM agent
WHERE workspace_id = $1 AND archived_at IS NULL AND kind = 'user'
ORDER BY created_at ASC;

-- name: ListAllAgents :many
SELECT * FROM agent
WHERE workspace_id = $1 AND kind = 'user'
ORDER BY created_at ASC;

-- name: ListAllAgentsAnyKind :many
-- Every agent row in the workspace, INCLUDING `kind = 'system'` execution
-- carriers (agent builder sessions) that ListAllAgents deliberately hides.
--
-- Only for surfaces that aggregate over raw `agent_task_queue` / task_usage
-- rows, which carry no kind filter of their own: a system carrier runs real
-- tasks and books real usage, so a per-agent rollup returns it whether or not
-- any list endpoint will ever name it. Those surfaces need the full population
-- to decide what to hide (MUL-5409). Do NOT use this to build a user-facing
-- agent list — `kind = 'system'` must stay out of every picker and assignee
-- surface.
SELECT * FROM agent
WHERE workspace_id = $1
ORDER BY created_at ASC;

-- name: GetAgent :one
SELECT * FROM agent
WHERE id = $1;

-- name: GetAgentForUpdate :one
-- Serializes read-modify-write updates to disabled_runtime_skills so two
-- concurrent per-skill toggles cannot overwrite each other.
SELECT * FROM agent
WHERE id = $1
FOR UPDATE;

-- name: GetAgentInWorkspace :one
SELECT * FROM agent
WHERE id = $1 AND workspace_id = $2 AND kind = 'user';

-- name: LockAgentForAutopilotAssignment :one
-- Serializes creating, retargeting, or resuming an active Autopilot with
-- Runtime teardown. Teardown takes FOR UPDATE on this same Agent row before it
-- clears runtime_id and pauses matching Autopilots. FOR SHARE also stabilizes
-- runtime_id and archived_at against concurrent ordinary Agent updates. The
-- shared lock makes the
-- two outcomes exhaustive:
--   * assignment commits first, then teardown sees and pauses it; or
--   * teardown commits first, then assignment re-reads runtime_id=NULL and
--     rejects the active Autopilot.
SELECT * FROM agent
WHERE id = $1 AND workspace_id = $2 AND kind = 'user'
FOR SHARE;

-- name: CreateAgent :one
INSERT INTO agent (
    workspace_id, name, description, avatar_url, runtime_mode,
    runtime_config, runtime_id, visibility, max_concurrent_tasks, owner_id,
    instructions, custom_env, custom_args, mcp_config, model, thinking_level,
    service_tier, conversation_starters,
    composio_toolkit_allowlist, permission_mode
) VALUES (
    $1, $2, $3, $4, $5,
    $6, $7, $8, $9, $10,
    $11, $12, $13, $14, $15, $16,
    $17, COALESCE(sqlc.narg('conversation_starters')::jsonb, '[]'::jsonb),
    sqlc.narg('composio_toolkit_allowlist')::text[],
    COALESCE(sqlc.narg('permission_mode'), 'private')
)
RETURNING *;

-- name: CreateAgentBuilder :one
-- One hidden builder agent per creation session. Keeping the execution carrier
-- session-scoped freezes its model/runtime configuration when multiple builder
-- flows are open concurrently, while `kind = 'system'` keeps it out of normal
-- agent lists and assignment surfaces.
INSERT INTO agent (
    workspace_id, name, description, runtime_mode, runtime_config, runtime_id,
    visibility, permission_mode, max_concurrent_tasks, owner_id, instructions,
    custom_env, custom_args, model, kind, system_key
) VALUES (
    @workspace_id, @name, '', @runtime_mode, '{}'::jsonb, @runtime_id,
    'private', 'private', 1, @owner_id, @instructions,
    '{}'::jsonb, '[]'::jsonb, sqlc.narg('model'), 'system', @system_key
)
RETURNING *;

-- name: DeleteSystemAgentByID :exec
-- Builder sessions own their hidden execution agent. Deleting the session
-- removes that carrier and its task rows; the kind guard prevents this cleanup
-- path from ever deleting a user-authored agent.
DELETE FROM agent
WHERE id = $1 AND kind = 'system' AND system_key LIKE 'agent_builder:%';

-- name: RebindAgentBuilderRuntime :one
-- Re-points a builder carrier at another runtime mid-conversation. The carrier
-- is what SendDirectChatMessage reads to stamp a chat task's runtime_id, so this
-- UPDATE is the only thing that actually moves subsequent replies; the live-draft
-- picker alone never did. model is reset wholesale because model ids are
-- per-runtime — the new runtime resolves its own default instead of inheriting an
-- id it may not serve. The kind/system_key guard mirrors DeleteSystemAgentByID so
-- this path can never touch a user-authored agent.
--
-- Callers MUST hold LockChatSessionForRuntimeBind on the owning chat_session for
-- the whole transaction, otherwise a concurrent send can stamp a task with the
-- pre-switch runtime after the pending-task check has already passed (MUL-5163).
--
-- chat_session.runtime_id is deliberately left untouched: the daemon only resumes
-- a stored provider session when chat_session.runtime_id matches the claiming
-- task's runtime (see the chat claim path), so leaving the old pointer in place is
-- exactly what makes the new runtime start a fresh provider session.
UPDATE agent
SET runtime_id = @runtime_id,
    runtime_mode = @runtime_mode,
    model = sqlc.narg('model'),
    updated_at = now()
WHERE id = @id AND kind = 'system' AND system_key LIKE 'agent_builder:%'
RETURNING *;

-- name: UpdateAgent :one
-- composio_toolkit_allowlist is set wholesale: the API layer is responsible
-- for normalising the request payload to either (a) the new slug list — sent
-- here verbatim — or (b) an empty array to explicitly disable Composio.
-- Distinguish "field omitted" (preserve) from "explicit clear" via
-- ClearAgentComposioToolkitAllowlist below, mirroring the
-- thinking_level / mcp_config two-query pattern: COALESCE can't restore NULL.
UPDATE agent SET
    name = COALESCE(sqlc.narg('name'), name),
    description = COALESCE(sqlc.narg('description'), description),
    avatar_url = COALESCE(sqlc.narg('avatar_url'), avatar_url),
    runtime_config = COALESCE(sqlc.narg('runtime_config'), runtime_config),
    runtime_mode = COALESCE(sqlc.narg('runtime_mode'), runtime_mode),
    runtime_id = COALESCE(sqlc.narg('runtime_id'), runtime_id),
    visibility = COALESCE(sqlc.narg('visibility'), visibility),
    permission_mode = COALESCE(sqlc.narg('permission_mode'), permission_mode),
    status = COALESCE(sqlc.narg('status'), status),
    max_concurrent_tasks = COALESCE(sqlc.narg('max_concurrent_tasks'), max_concurrent_tasks),
    instructions = COALESCE(sqlc.narg('instructions'), instructions),
    custom_env = COALESCE(sqlc.narg('custom_env'), custom_env),
    custom_args = COALESCE(sqlc.narg('custom_args'), custom_args),
    mcp_config = COALESCE(sqlc.narg('mcp_config'), mcp_config),
    model = COALESCE(sqlc.narg('model'), model),
    thinking_level = COALESCE(sqlc.narg('thinking_level'), thinking_level),
    service_tier = COALESCE(sqlc.narg('service_tier'), service_tier),
    conversation_starters = COALESCE(sqlc.narg('conversation_starters'), conversation_starters),
    composio_toolkit_allowlist = COALESCE(sqlc.narg('composio_toolkit_allowlist')::text[], composio_toolkit_allowlist),
    updated_at = now()
WHERE id = $1
RETURNING *;

-- name: ClearAgentComposioToolkitAllowlist :one
-- Explicit NULL-clear for composio_toolkit_allowlist. The COALESCE-based
-- UpdateAgent cannot set the column back to NULL — sending an empty array
-- through there would persist `{}` (still a non-NULL, equivalent to "no
-- toolkits" but distinct from "field never configured"). The API uses this
-- dedicated query when the agent owner removes every toolkit; subsequent
-- dispatch decisions treat NULL identically to `{}` (both -> no overlay).
UPDATE agent SET composio_toolkit_allowlist = NULL, updated_at = now()
WHERE id = $1
RETURNING *;

-- name: ClearAgentThinkingLevel :one
-- Explicit NULL-clear for thinking_level. COALESCE-based UpdateAgent cannot
-- set the column back to NULL, so the API layer routes "user picked Default"
-- through this dedicated query.
UPDATE agent SET thinking_level = NULL, updated_at = now()
WHERE id = $1
RETURNING *;

-- name: ClearAgentServiceTier :one
-- Explicit NULL-clear for service_tier. COALESCE-based UpdateAgent cannot
-- set the column back to NULL, so the API routes "Runtime default" here.
UPDATE agent SET service_tier = NULL, updated_at = now()
WHERE id = $1
RETURNING *;

-- name: ClearAgentMcpConfig :one
UPDATE agent SET mcp_config = NULL, updated_at = now()
WHERE id = $1
RETURNING *;

-- name: UpdateAgentCustomEnv :one
-- Replaces an agent's custom_env map wholesale. Used by the dedicated
-- env-management endpoint (POST/PUT /api/agents/{id}/env), which is the
-- only post-creation write path for env. UpdateAgent has been stripped
-- of custom_env handling so all env mutations flow through here and the
-- handler's audit-log + **** sentinel guard.
UPDATE agent
SET custom_env = $2, updated_at = now()
WHERE id = $1
RETURNING *;

-- name: UpdateAgentDisabledRuntimeSkills :one
UPDATE agent
SET disabled_runtime_skills = $2, updated_at = now()
WHERE id = $1
RETURNING *;

-- name: ArchiveAgent :one
UPDATE agent SET archived_at = now(), archived_by = $2, updated_at = now()
WHERE id = $1
RETURNING *;

-- name: ArchiveAgentsByRuntime :many
-- Bulk-archives every active agent bound to any runtime in the given set.
-- Used when revoking a leaving member's runtimes so agents pinned to those
-- runtimes can no longer be assigned new work. Returns the affected rows so
-- the caller can broadcast agent:archived per agent.
--
-- System agents are exempt: they belong to the workspace rather than to the
-- member who happened to create them, and the workspace's entry point runs
-- through one. Archiving Mika because a colleague left would take the default
-- agent away from everyone. Its runtime does go offline with the departure, so
-- it needs rebinding — but it stays visible and recoverable instead of
-- vanishing.
UPDATE agent
SET archived_at = now(), archived_by = @archived_by, updated_at = now()
WHERE runtime_id = ANY(@runtime_ids::uuid[]) AND archived_at IS NULL
  AND (system_key IS NULL OR system_key = '')
RETURNING *;

-- name: ArchiveAgentsByIDs :many
-- Narrow archive that only touches the explicit ID list. Used by the
-- cascade-delete endpoint so the user's expected_active_agent_ids list
-- is the authoritative bound on what gets archived: any agent that
-- appeared on the runtime after the user opened the dialog is filtered
-- out here so it can't be silently archived even in the (vanishingly
-- rare) case where a row-level race slips past the runtime FOR UPDATE
-- lock. Returns the affected rows so the caller can broadcast
-- agent:archived per agent.
UPDATE agent
SET archived_at = now(), archived_by = @archived_by, updated_at = now()
WHERE id = ANY(@agent_ids::uuid[]) AND archived_at IS NULL
RETURNING *;

-- name: ListActiveAgentsByRuntime :many
-- Returns every non-archived agent bound to a runtime. Backs the cascade
-- delete dialog: when DELETE /api/runtimes/:id refuses with
-- runtime_has_active_agents, the response carries this list so the front-end
-- can render exactly the agents that will be unbound if the user confirms,
-- and so the cascade endpoint's expected_active_agent_ids check has a stable
-- snapshot to compare against. Ordered by name for a deterministic display.
SELECT * FROM agent
WHERE runtime_id = $1 AND archived_at IS NULL AND kind = 'user'
ORDER BY name ASC;

-- name: ListActiveAgentsByRuntimeForUpdate :many
-- FOR UPDATE variant used inside the cascade-delete transaction. Locks
-- each currently-active agent row so a concurrent archive/move of one
-- of those rows blocks until our transaction commits. Pair with
-- LockAgentRuntime, which holds the runtime row exclusively to also
-- block FK-validated INSERTs / runtime_id updates that would otherwise
-- add a new agent to the runtime mid-cascade. Together they guarantee
-- that the set we compared against expected_active_agent_ids is stable.
SELECT * FROM agent
WHERE runtime_id = $1 AND archived_at IS NULL AND kind = 'user'
ORDER BY name ASC
FOR UPDATE;

-- name: ListUserAgentsByRuntimeForUpdate :many
-- Locks active AND archived user agents before a runtime teardown. Locking only
-- the active snapshot leaves a restore race: an archived row can become active
-- after confirmation and then be silently unbound. Runtime/profile delete
-- locks the parent runtime first (blocking new FK references), then calls this
-- in stable ID order (blocking restore/archive/move on existing rows).
SELECT * FROM agent
WHERE runtime_id = $1 AND kind = 'user'
ORDER BY id
FOR UPDATE;

-- name: RestoreAgent :one
UPDATE agent SET archived_at = NULL, archived_by = NULL, updated_at = now()
WHERE id = $1
RETURNING *;

-- name: ListAgentTasks :many
SELECT * FROM agent_task_queue
WHERE agent_id = $1
ORDER BY created_at DESC;

-- name: CreateAgentTask :one
-- Fenced against workspace teardown: lock_task_owner_rows (migration 284)
-- locks the owners' workspace rows in the writer's own transaction and returns
-- false once they are gone, so this statement writes no row instead of stranding
-- a task in a workspace that has just been deleted (MUL-5999).
-- head_sha stamps the commit under review into the task's context JSONB so the
-- reviewer-loop dedup (HasPendingTaskForIssueAndAgent) can tell a pending run
-- against an OLD head apart from a fresh request against a NEW head (TEN-356).
-- Empty/absent head_sha leaves context NULL, preserving pre-TEN-356 behavior for
-- issues with no linked PR. Issue-linked tasks never hit quick-create context
-- parsing (parseQuickCreateContext short-circuits on IssueID.Valid), so this
-- key rides harmlessly alongside.
-- id is minted by the application as a UUIDv7 (pkg/dbid) so consecutive
-- enqueues cluster in a narrow contiguous primary-key range instead of
-- scattering across the B-tree. On a table with existing v4 ids, that range
-- is not necessarily the tree's right edge.
-- COALESCE keeps the column's gen_random_uuid() default reachable, so a caller
-- that passes no id still inserts — it just gets a random v4, exactly as before.
-- The same pattern is used by every INSERT listed in pkg/dbid's write table.
INSERT INTO agent_task_queue (
    agent_id, runtime_id, issue_id, status, priority, trigger_comment_id,
    coalesced_comment_ids, trigger_summary, force_fresh_session, is_leader_task,
    squad_id, context, originator_user_id, accountable_user_id, runtime_mcp_overlay, runtime_connected_apps,
    originator_source, delegated_from_task_id, rule_version_id, rerun_of_task_id, trigger_evidence_kind, trigger_evidence_ref_id,
    id
)
SELECT
    $1, $2, $3, 'queued', $4, sqlc.narg(trigger_comment_id),
    COALESCE(sqlc.narg(coalesced_comment_ids)::uuid[], '{}'),
    sqlc.narg(trigger_summary),
    COALESCE(sqlc.narg('force_fresh_session')::boolean, FALSE),
    COALESCE(sqlc.narg('is_leader_task')::boolean, FALSE),
    sqlc.narg(squad_id),
    CASE
        WHEN COALESCE(sqlc.narg('head_sha')::text, '') <> ''
        THEN jsonb_build_object('head_sha', sqlc.narg('head_sha')::text)
        ELSE NULL
    END,
    sqlc.narg(originator_user_id),
    sqlc.narg(accountable_user_id),
    sqlc.narg(runtime_mcp_overlay),
    sqlc.narg(runtime_connected_apps),
    sqlc.narg(originator_source),
    sqlc.narg(delegated_from_task_id),
    sqlc.narg(rule_version_id),
    sqlc.narg(rerun_of_task_id),
    sqlc.narg(trigger_evidence_kind),
    sqlc.narg(trigger_evidence_ref_id),
    COALESCE(sqlc.narg('id')::uuid, gen_random_uuid())
WHERE lock_task_owner_rows($1, $3, $2)
RETURNING *;

-- name: CreateDeferredChannelIssueTask :one
-- Fenced against workspace teardown: lock_task_owner_rows (migration 284)
-- locks the owners' workspace rows in the writer's own transaction and returns
-- false once they are gone, so this statement writes no row instead of stranding
-- a task in a workspace that has just been deleted (MUL-5999).
-- Channel /issue media resolves after issue creation. Persist the assigned
-- issue task up front for crash safety, but keep it inert until attachment
-- binding settles or the fire_at fallback is promoted by the normal sweeper.
INSERT INTO agent_task_queue (
    agent_id, runtime_id, issue_id, status, priority, trigger_comment_id,
    coalesced_comment_ids, trigger_summary, force_fresh_session, is_leader_task,
    squad_id, context, originator_user_id, accountable_user_id, runtime_mcp_overlay, runtime_connected_apps,
    originator_source, delegated_from_task_id, rule_version_id, rerun_of_task_id,
    trigger_evidence_kind, trigger_evidence_ref_id, fire_at,
    id
)
SELECT
    $1, $2, $3, 'deferred', $4, sqlc.narg(trigger_comment_id),
    COALESCE(sqlc.narg(coalesced_comment_ids)::uuid[], '{}'),
    sqlc.narg(trigger_summary),
    COALESCE(sqlc.narg('force_fresh_session')::boolean, FALSE),
    COALESCE(sqlc.narg('is_leader_task')::boolean, FALSE),
    sqlc.narg(squad_id),
    jsonb_strip_nulls(jsonb_build_object(
        'head_sha', NULLIF(COALESCE(sqlc.narg('head_sha')::text, ''), ''),
        'channel_issue_media_pending', TRUE
    )),
    sqlc.narg(originator_user_id),
    sqlc.narg(accountable_user_id),
    sqlc.narg(runtime_mcp_overlay),
    sqlc.narg(runtime_connected_apps),
    sqlc.narg(originator_source),
    sqlc.narg(delegated_from_task_id),
    sqlc.narg(rule_version_id),
    sqlc.narg(rerun_of_task_id),
    sqlc.narg(trigger_evidence_kind),
    sqlc.narg(trigger_evidence_ref_id),
    @fire_at,
    COALESCE(sqlc.narg('id')::uuid, gen_random_uuid())
WHERE lock_task_owner_rows($1, $3, $2)
RETURNING *;

-- name: PromoteDeferredChannelIssueTask :one
-- Early promotion is idempotent at the service layer: a task already promoted
-- by the fire_at sweeper no longer matches and is treated as settled.
UPDATE agent_task_queue
SET status = 'queued', fire_at = NULL
WHERE id = $1 AND issue_id IS NOT NULL AND status = 'deferred'
RETURNING *;

-- name: SetDeferredChannelIssueTaskRuntimeOverlay :execrows
-- The issue and its inert media-gated task commit atomically before the optional
-- external overlay is built. Only hydrate the untouched original plan: a comment
-- merge may already have replaced the trigger, attribution, and matching overlay.
UPDATE agent_task_queue
SET runtime_mcp_overlay = sqlc.narg(runtime_mcp_overlay),
    runtime_connected_apps = sqlc.narg(runtime_connected_apps)
WHERE id = @id
  AND status = 'deferred'
  AND context->>'channel_issue_media_pending' = 'true'
  AND trigger_comment_id IS NULL
  AND originator_user_id IS NOT DISTINCT FROM sqlc.narg(expected_originator_user_id)::uuid;

-- name: CreateQuickCreateTask :one
-- Fenced against workspace teardown: lock_task_owner_rows (migration 284)
-- locks the owners' workspace rows in the writer's own transaction and returns
-- false once they are gone, so this statement writes no row instead of stranding
-- a task in a workspace that has just been deleted (MUL-5999).
-- Quick-create tasks have no issue / chat / autopilot link; the entire job
-- description (prompt, requester, workspace) lives in context JSONB. The
-- daemon detects this variant via context.type == "quick_create".
-- The requester who opened the quick-create modal is a direct_human originator
-- and accountable; attribution provenance is stamped so this path is not a
-- NULL-source enqueue bypass (MUL-4302 §2).
INSERT INTO agent_task_queue (
    agent_id, runtime_id, issue_id, status, priority, context, originator_user_id,
    accountable_user_id, runtime_mcp_overlay, runtime_connected_apps,
    originator_source, trigger_evidence_kind, trigger_evidence_ref_id,
    id
)
SELECT
    $1, $2, NULL, 'queued', $3, $4,
    sqlc.narg(originator_user_id),
    sqlc.narg(accountable_user_id),
    sqlc.narg(runtime_mcp_overlay),
    sqlc.narg(runtime_connected_apps),
    sqlc.narg(originator_source),
    sqlc.narg(trigger_evidence_kind),
    sqlc.narg(trigger_evidence_ref_id),
    COALESCE(sqlc.narg('id')::uuid, gen_random_uuid())
WHERE lock_task_owner_rows($1, NULL, $2)
RETURNING *;

-- name: CreateDeferredAgentTask :one
-- Fenced against workspace teardown: lock_task_owner_rows (migration 284)
-- locks the owners' workspace rows in the writer's own transaction and returns
-- false once they are gone, so this statement writes no row instead of stranding
-- a task in a workspace that has just been deleted (MUL-5999).
-- Deferred tasks are inert until PromoteDueDeferredTasksForRuntime flips them
-- to queued. Used for comment-routing escalation: a thread-owner primary task
-- gets a delayed assignee fallback without waking both agents at t=0.
-- Attribution is resolved and stamped at creation (not at promotion), from the
-- same trigger comment as the primary task, so the fallback assignee's run
-- carries a non-NULL source and evidence rather than bypassing attribution
-- (MUL-4302 §2).
INSERT INTO agent_task_queue (
    agent_id, runtime_id, issue_id, status, priority, trigger_comment_id,
    trigger_summary, is_leader_task, squad_id, escalation_for_task_id, fire_at,
    originator_user_id, accountable_user_id, originator_source,
    delegated_from_task_id, trigger_evidence_kind, trigger_evidence_ref_id,
    id
)
SELECT
    @agent_id, @runtime_id, @issue_id, 'deferred', @priority,
    sqlc.narg(trigger_comment_id),
    sqlc.narg(trigger_summary),
    COALESCE(sqlc.narg('is_leader_task')::boolean, FALSE),
    sqlc.narg(squad_id),
    @escalation_for_task_id,
    @fire_at,
    sqlc.narg(originator_user_id),
    sqlc.narg(accountable_user_id),
    sqlc.narg(originator_source),
    sqlc.narg(delegated_from_task_id),
    sqlc.narg(trigger_evidence_kind),
    sqlc.narg(trigger_evidence_ref_id),
    COALESCE(sqlc.narg('id')::uuid, gen_random_uuid())
WHERE lock_task_owner_rows($1, $3, $2)
RETURNING *;

-- name: LinkTaskToIssue :exec
-- Fenced against workspace teardown: lock_task_owner_rows (migration 284)
-- locks the owners' workspace rows in the writer's own transaction and returns
-- false once they are gone, so this statement writes no row instead of stranding
-- a task in a workspace that has just been deleted (MUL-5999).
-- Attaches the issue a quick-create task produced back to the task row, once
-- the agent has finished and the issue exists. Guarded by `issue_id IS NULL`
-- so this never overwrites an issue id that was set at task creation (only
-- quick-create tasks land here unset). Fixes the activity row staying on
-- "Creating issue" forever after completion.
UPDATE agent_task_queue
SET issue_id = $2
WHERE id = $1 AND issue_id IS NULL
  AND lock_task_owner_rows(NULL, $2, NULL);

-- name: CreateRetryTask :one
-- Fenced against workspace teardown: lock_task_owner_rows (migration 284)
-- locks the owners' workspace rows in the writer's own transaction and returns
-- false once they are gone, so this statement writes no row instead of stranding
-- a task in a workspace that has just been deleted (MUL-5999).
--
-- Fenced against slot contention too: ON CONFLICT DO NOTHING yields the single
-- queued/dispatched slot idx_one_pending_task_per_issue_agent_v2 allows per
-- (issue, agent) rather than raising 23505. A manual rerun may now be enqueued
-- behind a still-running task, and it can commit at any point — including
-- between a caller's "is a successor already pending?" check and this insert,
-- since that check takes no lock. Raising here would abort the caller's
-- transaction, and on the FailTask path that transaction also carries the
-- parent's failed status, so the failure would roll back and leave the task
-- stuck in 'running'. Yielding instead makes the policy explicit: whoever
-- already holds the slot keeps it, and a deliberate human rerun therefore wins
-- over an automatic retry.
--
-- Both fences surface the same way — zero rows, i.e. pgx.ErrNoRows from this
-- :one query. Callers must treat that as "no retry was created" and still commit
-- their own work, NOT as a failure.
-- Clones a parent task into a fresh queued attempt. Carries forward the
-- agent's resume context (session_id/work_dir) so the child can continue
-- the conversation when the backend supports it. Resume-unsafe failures are
-- retried as fresh sessions so the child does not inherit a stuck agent
-- conversation. Keep the CASE WHEN predicates in sync with
-- resumeUnsafeFailureReason and the resume lookup blacklists. attempt is
-- incremented; max_attempts, trigger_comment_id, coalesced_comment_ids,
-- is_leader_task, and squad_id are inherited so the retried task receives the
-- parent's complete planned comment batch and keeps the same squad-role
-- provenance. delivered_comment_ids intentionally stays at its '{}' default:
-- the child must earn its own delivery receipt at claim time.
--
-- originator_user_id is inherited so the Composio overlay decision sees the
-- same top-of-chain human across the retry: the user behind the original
-- run has not changed. The Composio overlay follows the agent's invocation
-- permission and uses the agent owner's connection (MUL-3963); originator is
-- carried for A2A/audit, not as an originator == agent.owner_id gate.
-- A system retry is NOT a new attribution event (MUL-4302 §5): it inherits the
-- parent's accountable human, source label, delegation lineage, rule version,
-- and trigger evidence UNCHANGED, and records retry_of_task_id = p.id so retry
-- and manual rerun stay separable in reporting. parent_task_id keeps its
-- existing meaning for the retry/resume machinery; retry_of_task_id is the
-- attribution-facing lineage column.
--
-- chat_input_task_id is inherited straight from the parent so the whole retry
-- chain keeps consuming the ORIGINAL root input batch (MUL-4351): the root
-- direct task set it to its own id, every descendant copies that value, and a
-- claim always reads the same user messages. A plain copy (not
-- COALESCE(parent.chat_input_task_id, parent.id)) is deliberate: legacy/channel
-- parents carry NULL and must stay NULL so their retries keep the trailing
-- selector — promoting a pre-migration NULL row to the task-owned path on retry
-- would risk replaying untagged history during a rolling deploy.
--
-- Chat retries are queued at GREATEST(priority, 3) so a transiently-failed
-- earlier turn is re-claimed ahead of any fresh chat task (priority 2) the user
-- queued while the failing turn was still running — the retry continues the
-- older turn first. Combined with creating the retry inside FailTask's
-- transaction, this leaves no window for a newer input task to jump ahead.
--
-- fire_at arms a backoff before the retry: when non-NULL the child is inserted
-- as 'deferred' with that fire_at and stays inert until the existing
-- PromoteDueDeferredTasksForRuntime sweeper (run promote-first on every claim
-- poll) flips it to 'queued'. Used for runtime_offline so the child waits for a
-- healthy runtime, and for provider_network's final attempt so it waits ~5s
-- instead of firing back-to-back with the immediate retry (MUL-4910). NULL
-- keeps the historical behaviour: an immediately-claimable 'queued' child.
--
-- max_attempts overrides the inherited budget when non-NULL (NULL inherits
-- p.max_attempts unchanged). Callers persist the reason-aware effective ceiling
-- here so the row stays self-consistent — e.g. provider_network's chain records
-- attempt=3, max_attempts=3 rather than leaking attempt=3, max_attempts=2 to the
-- task API (MUL-4910). The Go retryAttemptCeiling already refuses to raise a
-- disabled (max_attempts<=1) task, so this only ever widens, never revives.
INSERT INTO agent_task_queue (
    agent_id, runtime_id, issue_id, chat_session_id, autopilot_run_id,
    status, priority, trigger_comment_id, coalesced_comment_ids, trigger_summary, context,
    session_id, work_dir,
    attempt, max_attempts, parent_task_id, force_fresh_session, is_leader_task,
    squad_id, originator_user_id, accountable_user_id, runtime_mcp_overlay, runtime_connected_apps,
    originator_source, delegated_from_task_id, rule_version_id,
    trigger_evidence_kind, trigger_evidence_ref_id, retry_of_task_id,
    chat_input_task_id, fire_at,
    channel_context_revision, id
)
SELECT
    p.agent_id, p.runtime_id, p.issue_id, p.chat_session_id, p.autopilot_run_id,
    CASE WHEN sqlc.narg(fire_at)::timestamptz IS NOT NULL THEN 'deferred' ELSE 'queued' END,
    CASE WHEN p.chat_session_id IS NOT NULL THEN GREATEST(p.priority, 3) ELSE p.priority END,
    p.trigger_comment_id, p.coalesced_comment_ids, p.trigger_summary, p.context,
    CASE WHEN p.failure_reason IS NOT DISTINCT FROM 'codex_semantic_inactivity' THEN NULL ELSE p.session_id END,
    CASE WHEN p.failure_reason IS NOT DISTINCT FROM 'codex_semantic_inactivity' THEN NULL ELSE p.work_dir END,
    p.attempt + 1, COALESCE(sqlc.narg(max_attempts)::int, p.max_attempts), p.id,
    p.failure_reason IS NOT DISTINCT FROM 'codex_semantic_inactivity',
    p.is_leader_task,
    p.squad_id,
    p.originator_user_id,
    p.accountable_user_id,
    sqlc.narg(runtime_mcp_overlay),
    sqlc.narg(runtime_connected_apps),
    p.originator_source, p.delegated_from_task_id, p.rule_version_id,
    p.trigger_evidence_kind, p.trigger_evidence_ref_id, p.id,
    p.chat_input_task_id, sqlc.narg(fire_at),
    p.channel_context_revision,
    -- Named new_task_id, not id: $1 above is the PARENT task's id.
    COALESCE(sqlc.narg('new_task_id')::uuid, gen_random_uuid())
FROM agent_task_queue p
WHERE p.id = $1
  AND lock_task_owner_rows(p.agent_id, p.issue_id, p.runtime_id)
ON CONFLICT (issue_id, agent_id) WHERE status IN ('queued', 'dispatched')
       OR (status = 'deferred' AND context->>'channel_issue_media_pending' = 'true')
DO NOTHING
RETURNING *;

-- name: CreateManualQuickCreateRetryTask :one
-- A human retry of an issue-less quick-create is a new direct_human run, not
-- an automatic retry. It preserves the immutable quick-create context JSON
-- (including source_context_id), but deliberately starts a fresh session and
-- records rerun_of_task_id so attribution/reporting can distinguish the human
-- action from CreateRetryTask's retry_of_task_id lineage.
--
-- The caller locks and transfers the pending issue_source_context in the same
-- transaction. lock_task_owner_rows provides the usual workspace-teardown
-- fence; a deleted workspace therefore yields pgx.ErrNoRows and no task.
INSERT INTO agent_task_queue (
    agent_id, runtime_id, status, priority, context,
    force_fresh_session, is_leader_task, squad_id,
    originator_user_id, accountable_user_id,
    runtime_mcp_overlay, runtime_connected_apps,
    originator_source, rerun_of_task_id, id
)
SELECT
    p.agent_id, p.runtime_id, 'queued', p.priority, p.context,
    TRUE, p.is_leader_task, p.squad_id,
    sqlc.arg(actor_user_id), sqlc.arg(actor_user_id),
    sqlc.narg(runtime_mcp_overlay), sqlc.narg(runtime_connected_apps),
    'direct_human', p.id, sqlc.arg(new_task_id)
FROM agent_task_queue p
WHERE p.id = sqlc.arg(source_task_id)
  AND p.status = 'failed'
  AND p.issue_id IS NULL
  AND p.chat_session_id IS NULL
  AND p.autopilot_run_id IS NULL
  AND lock_task_owner_rows(p.agent_id, p.issue_id, p.runtime_id)
RETURNING *;

-- name: DeleteUnstartedQuickCreateRetryTask :execrows
-- FailTask creates an automatic retry in the same transaction as the parent
-- terminal write. An issue-less source-context retry has no pending-slot
-- unique key, so a competing retry or an already-attached context can make
-- the subsequent attach-authority transfer lose after this row was inserted.
-- Remove only that still-uncommitted child and let the parent's failure commit.
DELETE FROM agent_task_queue
WHERE id = sqlc.arg(task_id)
  AND status IN ('queued', 'deferred')
  AND issue_id IS NULL
  AND chat_session_id IS NULL
  AND autopilot_run_id IS NULL;

-- name: CancelAgentTasksByIssue :many
-- Cancels every active task on the issue and returns the affected rows so the
-- caller can reconcile each agent's status and broadcast task:cancelled events
-- (#1587). Prior :exec form silently dropped that info, leaving agents stuck at
-- status="working" with no self-correction. Only issue-deletion cleanup calls
-- this now; a status flip to cancelled/done no longer does (MUL-4465).
UPDATE agent_task_queue
SET status = 'cancelled', completed_at = now(), prepare_lease_expires_at = NULL
WHERE issue_id = $1 AND status IN ('queued', 'dispatched', 'running', 'waiting_local_directory', 'deferred')
RETURNING *;

-- name: CancelPendingTasksByIssueAndAgent :many
-- Cancels the not-yet-started tasks for a single (issue, agent) pair, so the
-- manual rerun flow can enqueue a replacement without colliding with
-- idx_one_pending_task_per_issue_agent_v2 and without dropping the rerun's own
-- attribution onto a row somebody else created.
--
-- 'running' and 'waiting_local_directory' are deliberately NOT cancelled: an
-- agent is executing in them. Neither status appears in that unique index, so a
-- fresh queued row can be inserted alongside one, and ClaimAgentTask's
-- per-(issue, agent) serialization holds the new row until the active run
-- reaches a terminal state. Cancelling them made a manual rerun kill the pass
-- the agent was still working on; interrupting an in-flight run is CancelTask's
-- job, not rerun's.
--
-- Everything that has NOT begun executing is still cleared, including deferred
-- escalations, so rerun keeps its prior "replace the pending plan" behaviour for
-- those rows.
UPDATE agent_task_queue
SET status = 'cancelled', completed_at = now(), prepare_lease_expires_at = NULL
WHERE issue_id = $1 AND agent_id = $2
  AND status IN ('queued', 'dispatched', 'deferred')
RETURNING *;

-- name: CancelAgentTasksByAgent :many
-- Bulk-cancel every active (queued/dispatched/running) task for an agent.
-- Returns the affected rows so callers can broadcast task:cancelled events.
-- Mirrors the shape of CancelAgentTasksByIssue / CancelPendingTasksByIssueAndAgent
-- (also :many + RETURNING + completed_at) so the three sibling cancel paths
-- behave consistently.
UPDATE agent_task_queue
SET status = 'cancelled', completed_at = now(), prepare_lease_expires_at = NULL
WHERE agent_id = $1 AND status IN ('queued', 'dispatched', 'running', 'waiting_local_directory', 'deferred')
RETURNING *;

-- name: CancelAgentTasksByTriggerComment :many
-- Cancels active tasks whose planned batch contains the edited/deleted comment.
-- The body may already have been embedded as either the primary trigger or a
-- coalesced input; cancellation prevents an agent from acting on a stale or
-- deleted version. Must run before deletion clears trigger_comment_id.
UPDATE agent_task_queue
SET status = 'cancelled', completed_at = now(), prepare_lease_expires_at = NULL
WHERE (trigger_comment_id = $1 OR $1 = ANY(coalesced_comment_ids))
  AND status IN ('queued', 'dispatched', 'running', 'waiting_local_directory', 'deferred')
RETURNING *;

-- name: CancelAgentTasksByChatSession :many
-- Cancels active tasks belonging to a chat session. Called from
-- DeleteChatSession so the daemon doesn't keep running work whose result
-- has nowhere to land. Must run BEFORE the chat_session row is deleted —
-- the FK ON DELETE SET NULL would otherwise nullify chat_session_id and we
-- could no longer reach those tasks.
UPDATE agent_task_queue
SET status = 'cancelled', completed_at = now(), prepare_lease_expires_at = NULL
WHERE chat_session_id = $1 AND status IN ('queued', 'dispatched', 'running', 'waiting_local_directory', 'deferred')
RETURNING *;

-- name: GetAgentTask :one
SELECT * FROM agent_task_queue
WHERE id = $1;

-- name: GetAgentTaskForDelegatedFailureUpdate :one
-- Serializes the idempotent delegated-failure recovery signal for one failed
-- task. FailTask and the stale-task sweepers can converge on the same row; the
-- lock ensures they cannot create two recovery comments/tasks for it.
SELECT * FROM agent_task_queue
WHERE id = $1
FOR UPDATE;

-- name: HasRetryTaskForParent :one
-- Defense-in-depth for the recovery path: a delegated failure with any retry
-- child is still intermediate and must not wake its coordinator.
SELECT count(*) > 0
FROM agent_task_queue
WHERE parent_task_id = $1
  AND status <> 'cancelled';

-- name: GetAgentTaskInWorkspace :one
-- Loads a task only when its owning agent lives in the given workspace.
-- agent_id is NOT NULL on every task row (and ON DELETE CASCADE, so the agent
-- always exists), which makes this the universal tenant guard for
-- user-initiated cancellation — independent of which optional source FK
-- (issue / chat_session / autopilot_run) happens to be set. It is what lets
-- run_only autopilot tasks and quick_create tasks (whose issue does not exist
-- yet) be cancelled at all, instead of 404-ing on a missing source FK.
SELECT atq.* FROM agent_task_queue atq
JOIN agent a ON a.id = atq.agent_id
WHERE atq.id = $1 AND a.workspace_id = $2;

-- name: ClaimAgentTask :one
-- Claims the next queued task for an agent on one healthy runtime, enforcing
-- per-(issue, agent) serialization:
-- a task is only claimable when no other task for the same issue AND same agent is
-- already dispatched or running. This allows different agents to work on the same
-- issue in parallel while preventing a single agent from running duplicate tasks.
-- Chat tasks (issue_id IS NULL) use chat_session_id for serialization instead.
-- Quick-create tasks have no issue / chat / autopilot link, so they serialize on
-- "any other quick-create-shaped task" (all four FKs NULL) for the same agent —
-- otherwise a user mashing the create button could fire concurrent quick-creates
-- whose completion lookup would race over "most recent issue by this agent".
UPDATE agent_task_queue
SET status = 'dispatched',
    dispatched_at = now(),
    prepare_lease_expires_at = now() + make_interval(secs => @prepare_lease_secs::double precision)
WHERE id = (
    SELECT atq.id FROM agent_task_queue atq
    WHERE atq.agent_id = @agent_id
      AND atq.runtime_id = @runtime_id
      AND atq.status = 'queued'
      AND EXISTS (
          SELECT 1
          FROM agent a
          JOIN agent_runtime r ON r.id = atq.runtime_id
          WHERE a.id = atq.agent_id
            -- A task's persisted runtime is not authority after an agent rebind.
            AND a.runtime_id = atq.runtime_id
            -- Private runtimes only execute their owner's agents. Ownerless
            -- runtime/agent rows remain claimable only so the handler can
            -- settle them explicitly before daemon delivery; filtering them
            -- here would leave every task silently queued until the TTL.
            -- Public runtimes remain shareable across agent owners.
            AND (
                r.visibility = 'public'
                OR (
                    r.visibility = 'private'
                    AND (
                        r.owner_id IS NULL
                        OR a.owner_id IS NULL
                        OR r.owner_id = a.owner_id
                    )
                )
            )
            AND r.status = 'online'
            AND COALESCE(r.last_seen_at, r.updated_at) >=
                now() - make_interval(secs => @runtime_stale_secs::double precision)
      )
      AND NOT EXISTS (
          SELECT 1 FROM agent_task_queue active
          WHERE active.agent_id = atq.agent_id
            AND active.status IN ('dispatched', 'running', 'waiting_local_directory')
            AND (
              (atq.issue_id IS NOT NULL AND active.issue_id = atq.issue_id)
              OR (atq.chat_session_id IS NOT NULL AND active.chat_session_id = atq.chat_session_id)
              OR (
                atq.issue_id IS NULL
                AND atq.chat_session_id IS NULL
                AND atq.autopilot_run_id IS NULL
                AND active.issue_id IS NULL
                AND active.chat_session_id IS NULL
                AND active.autopilot_run_id IS NULL
              )
            )
      )
    ORDER BY atq.priority DESC, atq.created_at ASC, atq.id ASC
    LIMIT 1
    FOR UPDATE SKIP LOCKED
)
RETURNING *;

-- name: SetTaskDeliveredCommentIDs :one
-- Replace (rather than append to) the delivery receipt for this claim. A stale
-- dispatched task may be reclaimed by a daemon with different capabilities,
-- so only the ids embedded in the newest response count as delivered. The CAS
-- keeps a stale handler from writing after execution starts, and the subset
-- guard prevents acknowledging an id outside the task's enqueue-time plan.
UPDATE agent_task_queue
SET delivered_comment_ids = @delivered_comment_ids::uuid[]
WHERE id = @task_id
  AND runtime_id = @runtime_id
  AND status = 'dispatched'
  AND started_at IS NULL
  AND dispatched_at = @dispatched_at
  AND trigger_comment_id IS NOT DISTINCT FROM sqlc.narg(expected_trigger_comment_id)::uuid
  AND NOT EXISTS (
      SELECT 1
      FROM unnest(@delivered_comment_ids::uuid[]) AS delivered(id)
      WHERE delivered.id IS NULL
         OR (
             delivered.id IS DISTINCT FROM trigger_comment_id
             AND NOT (delivered.id = ANY(coalesced_comment_ids))
         )
  )
RETURNING delivered_comment_ids;

-- name: RequeueAgentTaskAfterClaimFailure :one
-- Claim finalization (task token + optional comment receipt) failed before any
-- response bytes were written. Return only that exact claim generation to the
-- queue so another poll can retry immediately instead of waiting for stale
-- dispatch recovery. The dispatched_at CAS prevents an old handler from
-- rolling back a newer reclaim.
UPDATE agent_task_queue
SET status = 'queued',
    dispatched_at = NULL,
    prepare_lease_expires_at = NULL,
    delivered_comment_ids = '{}'
WHERE id = @task_id
  AND runtime_id = @runtime_id
  AND status = 'dispatched'
  AND started_at IS NULL
  AND dispatched_at = @dispatched_at
RETURNING *;

-- name: ReclaimStaleDispatchedTaskForRuntime :one
-- Re-delivers a task whose previous claim likely succeeded server-side but
-- whose response never reached the daemon. The task is still in `dispatched`
-- with no `started_at`, so the daemon has not acknowledged it via StartTask.
-- Refresh dispatched_at so the server-side dispatch timeout measures from the
-- recovered delivery attempt.
UPDATE agent_task_queue
SET dispatched_at = now(),
    prepare_lease_expires_at = now() + make_interval(secs => @prepare_lease_secs::double precision)
WHERE id = (
    SELECT atq.id FROM agent_task_queue atq
    WHERE atq.runtime_id = $1
      AND atq.status = 'dispatched'
      AND atq.started_at IS NULL
      AND atq.dispatched_at < now() - make_interval(secs => @claim_recovery_secs::double precision)
      AND (atq.prepare_lease_expires_at IS NULL OR atq.prepare_lease_expires_at < now())
      AND EXISTS (
          -- Keep this authorization fence in sync with ClaimAgentTask.
          SELECT 1
          FROM agent a
          JOIN agent_runtime r ON r.id = atq.runtime_id
          WHERE a.id = atq.agent_id
            AND a.runtime_id = atq.runtime_id
            AND (
                r.visibility = 'public'
                OR (
                    r.visibility = 'private'
                    AND (
                        r.owner_id IS NULL
                        OR a.owner_id IS NULL
                        OR r.owner_id = a.owner_id
                    )
                )
            )
            AND r.status = 'online'
            AND COALESCE(r.last_seen_at, r.updated_at) >=
                now() - make_interval(secs => @runtime_stale_secs::double precision)
      )
    ORDER BY atq.priority DESC, atq.dispatched_at ASC
    LIMIT 1
    FOR UPDATE SKIP LOCKED
)
RETURNING *;

-- name: ReclaimStaleDispatchedTasksForRuntimes :many
-- Batch variant of ReclaimStaleDispatchedTaskForRuntime (MUL-4257): re-delivers
-- up to @max_tasks tasks across the whole runtime set in one round trip, so a
-- machine-level batch claim recovers lost-response dispatches for every runtime
-- it hosts without one query per runtime. Same eligibility as the singular
-- query (dispatched, never started, past the recovery window, expired/absent
-- prepare lease) and the same dispatched_at refresh; only the runtime filter
-- (= ANY) and the LIMIT (max_tasks instead of 1) differ.
UPDATE agent_task_queue
SET dispatched_at = now(),
    prepare_lease_expires_at = now() + make_interval(secs => @prepare_lease_secs::double precision)
WHERE id IN (
    SELECT atq.id FROM agent_task_queue atq
    WHERE atq.runtime_id = ANY(@runtime_ids::uuid[])
      AND atq.status = 'dispatched'
      AND atq.started_at IS NULL
      AND atq.dispatched_at < now() - make_interval(secs => @claim_recovery_secs::double precision)
      AND (atq.prepare_lease_expires_at IS NULL OR atq.prepare_lease_expires_at < now())
      AND EXISTS (
          -- Keep this authorization fence in sync with ClaimAgentTask.
          SELECT 1
          FROM agent a
          JOIN agent_runtime r ON r.id = atq.runtime_id
          WHERE a.id = atq.agent_id
            AND a.runtime_id = atq.runtime_id
            AND (
                r.visibility = 'public'
                OR (
                    r.visibility = 'private'
                    AND (
                        r.owner_id IS NULL
                        OR a.owner_id IS NULL
                        OR r.owner_id = a.owner_id
                    )
                )
            )
            AND r.status = 'online'
            AND COALESCE(r.last_seen_at, r.updated_at) >=
                now() - make_interval(secs => @runtime_stale_secs::double precision)
      )
    ORDER BY atq.priority DESC, atq.dispatched_at ASC
    LIMIT @max_tasks::int
    FOR UPDATE SKIP LOCKED
)
RETURNING *;

-- name: ExtendAgentTaskPrepareLease :one
-- Keeps a dispatched task protected while the daemon resolves/cache/materializes
-- startup inputs before StartTask. Once the daemon stops extending this short
-- lease, the stale-dispatched reclaim path can recover the task without waiting
-- for a long global recovery window.
UPDATE agent_task_queue
SET prepare_lease_expires_at = now() + make_interval(secs => @lease_secs::double precision)
WHERE id = $1
  AND runtime_id = $2
  AND status IN ('dispatched', 'waiting_local_directory')
  AND started_at IS NULL
RETURNING *;

-- name: StartAgentTask :one
-- Transitions a task to running. Accepts either 'dispatched' (the normal
-- claim → run flow) or 'waiting_local_directory' (the daemon held the row in
-- a wait state while another task owned the local_directory path lock; once
-- the lock was acquired the daemon flips here). wait_reason is cleared on
-- the transition so a future read can't conflate "currently waiting" with
-- "previously waited".
UPDATE agent_task_queue
SET status = 'running',
    started_at = now(),
    wait_reason = NULL,
    prepare_lease_expires_at = NULL
WHERE id = $1 AND status IN ('dispatched', 'waiting_local_directory')
RETURNING *;

-- name: MarkAgentTaskWaitingLocalDirectory :one
-- Transitions a freshly-dispatched task into 'waiting_local_directory' while
-- the daemon waits for another in-flight task to release the path lock on a
-- project_resource of type local_directory. wait_reason carries a short
-- human-readable hint (typically the contested path) that the UI surfaces
-- alongside the status.
--
-- The CHECK only allows the transition from 'dispatched' so a daemon can't
-- mark an already-running or terminal task as waiting; the StartAgentTask
-- mutation handles the reverse transition once the lock is acquired.
UPDATE agent_task_queue
SET status = 'waiting_local_directory',
    wait_reason = $2,
    prepare_lease_expires_at = now() + make_interval(secs => @prepare_lease_secs::double precision)
WHERE id = $1 AND status = 'dispatched'
RETURNING *;

-- name: CompleteAgentTask :one
-- session_rollout_missing (MUL-5305): when true the daemon withheld this task's
-- Codex session because its rollout was never written to the store. Forcing
-- session_id NULL and flagging the row happen in THIS terminal transaction so an
-- auto-retry created and woken by the same commit can never observe the bad
-- pointer or a missing gap flag.
--
-- retired_session_id (GH #6066): a session this run abandoned as unresumable.
-- Recorded even on the completed path, because that is exactly the case the
-- resume lookups could not see — a fresh-session retry that SUCCEEDS still has
-- to retire the transcript it retried away from, or an older completed row
-- pointing at the same id resurrects it on the next run.
UPDATE agent_task_queue
SET status = 'completed', completed_at = now(), result = $2,
    session_id = CASE WHEN sqlc.arg('session_rollout_missing') THEN NULL ELSE $3 END,
    work_dir = $4,
    durable_work_dir = COALESCE(sqlc.narg('durable_work_dir'), durable_work_dir),
    branch_name = COALESCE(sqlc.narg('branch_name'), branch_name),
    session_rollout_missing = sqlc.arg('session_rollout_missing'),
    retired_session_id = COALESCE(sqlc.narg('retired_session_id'), retired_session_id),
    prepare_lease_expires_at = NULL
WHERE id = $1 AND status = 'running'
RETURNING *;

-- name: GetLastTaskSession :one
-- Returns the session_id and work_dir from the most recent task for a given
-- (agent_id, issue_id) pair, used for session resumption on the auto-retry
-- path. We accept 'completed', 'failed' AND 'cancelled' tasks: a failed task
-- may have established a real agent session before crashing (orphaned by a
-- daemon restart, runtime offline, or sweeper timeout), and the daemon pins
-- the resume pointer mid-flight via UpdateAgentTaskSession. Without this,
-- an auto-retry of a mid-run failure would silently start a fresh
-- conversation and lose the in-flight context — exactly what MUL-1128's B
-- branch is meant to fix.
--
-- A cancelled task is in exactly the same position, and excluding it was the
-- issue-side half of GH #6340: the pinned session is real (the provider emitted
-- it), the user stopped the run rather than the provider rejecting it, and
-- cancellation records no failure_reason/error for the poison filters below to
-- match on. So a stop-then-comment-again sequence keeps its context instead of
-- silently starting cold. A user who wants a clean slate has manual rerun,
-- which never takes this path (see below).
--
-- Manual rerun (TaskService.RerunIssue) does NOT take this path. The claim
-- handler branches on rerun_of_task_id FIRST and resolves the session/workdir
-- from that exact source task (so a parallel task on the same issue can't be
-- resumed by mistake), reusing the source workdir and resuming its session only
-- when the source failure is resume-safe. The rerun row still carries
-- force_fresh_session=true purely as a rollback-safe signal: an OLD claim
-- handler that predates the rerun_of_task_id branch falls back to this query,
-- and force_fresh_session=true makes it start clean instead of resuming the
-- wrong execution (MUL-4869).
--
-- Tasks that ended in a known "poisoned" terminal state are also excluded
-- here so even auto-retry does not inherit the bad session. The daemon
-- classifies these failures (iteration_limit, agent_fallback_message,
-- api_invalid_request, codex_semantic_inactivity, agent_error.context_overflow,
-- codex_resume_oversized)
-- when it detects either an agent fallback marker in the output, an upstream
-- API 400 that means the conversation history itself is unprocessable
-- (oversized image, malformed base64, etc.), a Codex semantic inactivity
-- timeout whose recorded session may replay the same stuck state, a context
-- window overflow that would immediately overflow again on resume, or a Codex
-- thread/resume response too large to read back (MUL-5722). Keep this
-- list in sync with resumeUnsafeFailureReason and GetLastChatTaskSession.
--
-- The error-text ILIKE clause is defense-in-depth for the api_invalid_request
-- shape: a legacy row tagged 'agent_error' (pre-MUL-1921), a deploy-window
-- row that the old code wrote between migration and rollout, or a future
-- error format that escapes the daemon classifier all still get filtered
-- here as long as the canonical Anthropic 400 marker is present in the
-- error text. Migration 079 backfills the failure_reason column itself,
-- so observability stays accurate; this clause guarantees session resume
-- never picks up a bad session even when failure_reason hasn't caught up.
--
-- Selection is per-session, not per-row: we first reduce to the single most
-- recent terminal row FOR EACH session_id (DISTINCT ON), then apply the
-- resume-unsafe filter. This closes a poisoning wormhole (GH #5975): when the
-- SAME session_id has both an older 'completed' row (the turn that first baked
-- in the bad history) and a newer poisoned 'failed' row, a plain row-level
-- filter would drop the poisoned row and happily fall back to the older
-- completed row — which points at the exact same dead session. Judging by each
-- session's LATEST terminal state instead means:
--   - a newer poisoned row invalidates the whole session, older completed rows
--     included;
--   - a different, healthy session (distinct id) is still eligible as a
--     fallback;
--   - if that same id later terminates cleanly again, the newer completed row
--     restores its resumability.
--
-- The oversized-image ILIKE (dimension phrase AND image.source.base64.data) is
-- the GH #5975 analogue of the api_invalid_request defense-in-depth above: a
-- Kiro/ACP resume rejected for an oversized historical image is classified
-- 'api_invalid_request' at write time, but this text clause also blocks a
-- legacy/unclassified row (or a new-server + old-daemon deploy window) from
-- resuming the poisoned session. Both markers are required so the clause stays
-- exactly as narrow as classifyPoisonedError and the Kiro detector — an
-- unrelated error that only mentions image dimensions is NOT excluded.
--
-- MUL-5722 needed a different shape, and the reason is worth stating because
-- the first attempt got it wrong: a row-level guard like the ones above CANNOT
-- work for an overflowed resume. That failure happens before the turn starts,
-- so the backend has no session id to report and the row lands with session_id
-- NULL — which the latest_per_session CTE drops before any filter runs. The
-- only row naming the oversized thread is the OLDER completed one, which looks
-- perfectly healthy. Hence resume_overflow_at below: block by TIME, which needs
-- nothing from the failed row but its existence. See the clause for details.
--
-- The final pair of regexes is the provider-agnostic version of the same
-- guard, and it matters most for self-hosted installs: daemons upgrade on
-- their own cadence, so a host still running a daemon without
-- taskfailure.UnresumableHistory reports an empty-message rejection as
-- agent_error.unknown and would otherwise keep resuming the transcript the
-- provider just refused (GH #6066). Both regexes must match — an emptiness
-- complaint AND a locator pointing into the message history — mirroring
-- emptyContentRe and historyMessageLocatorRe in pkg/taskfailure/resume.go.
-- Keep the three in sync.
--
-- retired_sessions is the explicit half of the same rule (GH #6066). The
-- per-session latest state above can only judge sessions that some row still
-- POINTS at, so it cannot see a session a run deliberately abandoned: a fresh
-- retry that succeeded reports its own new id (or none), leaving the poisoned
-- id visible on nothing but an older completed row, which then looks perfectly
-- healthy. retired_session_id records the abandonment itself, so one row
-- retiring a session removes it from every later lookup no matter how many
-- clean rows still reference it.
WITH retired_sessions AS (
    SELECT DISTINCT r.retired_session_id AS session_id
    FROM agent_task_queue r
    WHERE r.agent_id = $1 AND r.issue_id = $2
      AND r.retired_session_id IS NOT NULL
), resume_overflow_at AS (
    SELECT MAX(COALESCE(t.completed_at, t.started_at, t.dispatched_at, t.created_at)) AS at
    FROM agent_task_queue t
    WHERE t.agent_id = $1 AND t.issue_id = $2
      AND t.status = 'failed'
      AND (
        COALESCE(t.failure_reason, '') = 'codex_resume_oversized'
        OR (COALESCE(t.error, '') ILIKE '%thread/resume failed%' AND COALESCE(t.error, '') ILIKE '%token too long%')
      )
), latest_per_session AS (
    SELECT DISTINCT ON (t.session_id)
        t.session_id, t.work_dir, t.runtime_id, t.status, t.failure_reason, t.error,
        COALESCE(t.completed_at, t.started_at, t.dispatched_at, t.created_at) AS terminal_at
    FROM agent_task_queue t
    WHERE t.agent_id = $1 AND t.issue_id = $2
      AND t.session_id IS NOT NULL
      AND t.status IN ('completed', 'failed', 'cancelled')
    ORDER BY t.session_id, COALESCE(t.completed_at, t.started_at, t.dispatched_at, t.created_at) DESC
)
SELECT session_id, work_dir, runtime_id FROM latest_per_session
WHERE session_id NOT IN (SELECT session_id FROM retired_sessions)
  AND (
    status IN ('completed', 'cancelled')
    OR (
      status = 'failed'
      AND COALESCE(failure_reason, '') NOT IN ('iteration_limit', 'agent_fallback_message', 'api_invalid_request', 'codex_semantic_inactivity', 'agent_error.context_overflow', 'codex_resume_oversized')
      AND NOT (COALESCE(error, '') ILIKE '%400%' AND COALESCE(error, '') ILIKE '%invalid_request_error%')
      AND NOT (COALESCE(error, '') ILIKE '%image dimensions exceed max allowed size%' AND COALESCE(error, '') ILIKE '%image.source.base64.data%')
      -- A provider credential-resolution failure ("Could not resolve
      -- authentication method...") is deterministic on resume: the missing
      -- api_key / auth_token / header is baked into the session's provider
      -- state, so replaying it reproduces the same provider error forever. It
      -- is classified agent_error.unknown (resume-safe), so this text guard is
      -- the only thing that keeps a wedged issue from resuming the same dead
      -- session on its next trigger — there is no daemon upgrade to wait for.
      -- Keep in sync with ResumeUnsafeFailure and GetLastChatTaskSession.
      -- The phrase itself lives in taskfailure.AuthMethodUnresolved, which the
      -- daemon's in-turn fresh-session retry reads (GH #6777). This guard stays
      -- because it is the only protection for rows an older daemon wrote.
      AND NOT (COALESCE(error, '') ILIKE '%could not resolve authentication method%')
      AND NOT (COALESCE(error, '') ~* 'must not be empty|must be non-?empty|must have non-?empty|non-?empty content|cannot be empty|should not be empty'
               AND COALESCE(error, '') ~* 'role[^a-z0-9]{0,2}assistant|assistant message|message at position|messages\.[0-9]|messages\[[0-9]')
    )
  )
  -- MUL-5722: a resume that overflowed the reader names no session, so it can
  -- only be excluded by time, not by matching the failed row. Drop every
  -- session whose last terminal activity predates the newest such failure: one
  -- of them IS the oversized thread, and the row that would tell us which is
  -- exactly the row that could not be written. Anything that terminated after
  -- the overflow is the fresh thread that replaced it, so this un-blocks itself
  -- as soon as one succeeds.
  AND (
    (SELECT at FROM resume_overflow_at) IS NULL
    OR terminal_at > (SELECT at FROM resume_overflow_at)
  )
ORDER BY terminal_at DESC
LIMIT 1;

-- name: GetLatestTaskRolloutMissing :one
-- Reports whether the most recent terminal task for (agent_id, issue_id)
-- withheld its Codex session because the rollout was missing (MUL-5305). When
-- true, GetLastTaskSession fell back to an older session, so the next run must
-- disclose that the most recent turn's context could not be carried over. Any
-- later task that records a real session resets this to FALSE by being the new
-- most-recent row, so the disclosure fires once and then clears.
SELECT COALESCE(session_rollout_missing, FALSE) FROM agent_task_queue
WHERE agent_id = $1 AND issue_id = $2
  AND status IN ('completed', 'failed')
  AND started_at IS NOT NULL
ORDER BY COALESCE(completed_at, started_at, dispatched_at, created_at) DESC
LIMIT 1;

-- name: GetLatestChatTaskRolloutMissing :one
-- Chat-session counterpart of GetLatestTaskRolloutMissing (MUL-5305): reports
-- whether the most recent terminal task on this chat session withheld its Codex
-- session because the rollout was missing. When true the next chat claim resumed
-- an older session (or none), so it must disclose the continuity gap.
SELECT COALESCE(session_rollout_missing, FALSE) FROM agent_task_queue
WHERE chat_session_id = sqlc.arg('chat_session_id')
  AND (
    sqlc.narg('channel_context_revision')::bigint IS NULL
    OR COALESCE(channel_context_revision, 1) = sqlc.narg('channel_context_revision')::bigint
  )
  AND status IN ('completed', 'failed')
  AND started_at IS NOT NULL
ORDER BY COALESCE(completed_at, started_at, dispatched_at, created_at) DESC
LIMIT 1;

-- name: GetLastTaskStartedAtForIssueAndAgent :one
-- Returns the started_at of the most recent prior task for this (agent, issue)
-- pair, used as the "since" anchor for counting comments that arrived since the
-- agent's last run. Any terminal state counts as "a run happened". Tasks with
-- no started_at (never dispatched / the just-claimed current task) are excluded,
-- so this never returns the current claim's own row. MUST use started_at, never
-- completed_at: a long run would otherwise miss comments posted while it ran.
SELECT started_at FROM agent_task_queue
WHERE agent_id = $1 AND issue_id = $2 AND started_at IS NOT NULL
ORDER BY started_at DESC
LIMIT 1;

-- name: FailAgentTask :one
-- Marks a task as failed. session_id and work_dir are merged via COALESCE so
-- if the agent already established a real session before failing (e.g. it
-- crashed mid-conversation, was cancelled, or hit a tool error) the resume
-- pointer is preserved on the task row. The next chat task can then fall
-- back to GetLastChatTaskSession and continue the conversation instead of
-- silently starting over.
--
-- failure_reason is a coarse classifier consumed by the auto-retry path;
-- 'agent_error' is the safe default when the daemon doesn't supply one.
--
-- session_rollout_missing (MUL-5305): when true the daemon withheld this task's
-- Codex session (its rollout was missing). Force session_id NULL — overriding
-- the COALESCE that would otherwise preserve a stale mid-flight pin — and flag
-- the row, in the SAME transaction that creates and wakes the auto-retry, so the
-- retry can never claim the bad pointer or miss the continuity gap.
UPDATE agent_task_queue
SET status = 'failed',
    completed_at = now(),
    error = $2,
    failure_reason = COALESCE(sqlc.narg('failure_reason'), 'agent_error'),
    session_id = CASE WHEN sqlc.arg('session_rollout_missing') THEN NULL ELSE COALESCE(sqlc.narg('session_id'), session_id) END,
    work_dir = COALESCE(sqlc.narg('work_dir'), work_dir),
    durable_work_dir = COALESCE(sqlc.narg('durable_work_dir'), durable_work_dir),
    branch_name = COALESCE(sqlc.narg('branch_name'), branch_name),
    session_rollout_missing = sqlc.arg('session_rollout_missing'),
    retired_session_id = COALESCE(sqlc.narg('retired_session_id'), retired_session_id),
    prepare_lease_expires_at = NULL
WHERE id = $1 AND status IN ('dispatched', 'running', 'waiting_local_directory')
RETURNING *;

-- name: UpdateAgentTaskSession :exec
-- Pins the resume pointer mid-flight so a daemon crash leaves a usable
-- session_id/work_dir on the task row. No-op if the task is no longer
-- in dispatched/running. waiting_local_directory tasks have no session yet
-- so this query intentionally skips them.
--
-- A row the user just cancelled accepts the pin too, but only while its
-- session slot is still empty (GH #6340). The pin is asynchronous — for Codex
-- it waits for the rollout to reach the store — so a cancel landing in that
-- window used to drop the session id for good, and the cancelled turn's
-- context with it. Filling an EMPTY slot on the run's own row is exactly what
-- the mid-flight pin is for; the `session_id IS NULL` guard is what keeps this
-- from being an overwrite, and completed/failed rows stay untouchable so a
-- straggler goroutine can never contradict a terminal report.
UPDATE agent_task_queue
SET session_id = COALESCE(sqlc.narg('session_id'), session_id),
    work_dir  = COALESCE(sqlc.narg('work_dir'), work_dir)
WHERE id = $1
  AND (
    status IN ('dispatched', 'running')
    OR (status = 'cancelled' AND session_id IS NULL)
  );

-- name: RecoverOrphanedTasksForRuntime :many
-- Called by the daemon at startup. Atomically fails any dispatched/running/
-- waiting_local_directory task that the prior incarnation of this runtime
-- owned but did not finalize. Returns the failed rows so callers can hand
-- them to the auto-retry path. waiting_local_directory rows are included
-- because the daemon holding the path lock is the same process that just
-- died — without us, the row would sit waiting forever.
UPDATE agent_task_queue
SET status = 'failed',
    completed_at = now(),
    error = 'daemon restarted while task was in flight',
    failure_reason = 'runtime_recovery',
    wait_reason = NULL,
    prepare_lease_expires_at = NULL
WHERE runtime_id = $1 AND status IN ('dispatched', 'running', 'waiting_local_directory')
RETURNING *;

-- name: FailStaleTasks :many
-- Fails tasks stuck in dispatched/running beyond the given thresholds.
--
-- Each branch pairs a wall-clock deadline with a task-appropriate liveness
-- signal, so the sweeper only kills tasks whose owning daemon is no longer
-- proving it is alive:
--
--   * Dispatched: `prepare_lease_expires_at` is refreshed every 15s by the
--     daemon between claim and StartTask (see startTaskPrepareLeaseExtender).
--     A live lease excludes the row. An expired lease may fail immediately on
--     a healthy runtime, but a disconnected runtime gets the same reconnect
--     grace as a running task.
--
--   * Running: no per-task lease is renewed once StartTask fires, so we key
--     off the daemon-wide heartbeat instead — `agent_runtime.last_seen_at`,
--     which the daemon bumps every ~15s while it is up. A running task is not
--     killed until that heartbeat has been absent for the full reconnect
--     grace, even when its own wall-clock timeout elapsed earlier. This keeps
--     healthy multi-hour work alive through a network partition.
--
-- The daemon-dead case is recovered immediately when that daemon restarts via
-- RecoverOrphanedTasksForRuntime. Until then, this query and
-- FailTasksForOfflineRuntimes share the same bounded reconnect grace.
--
-- runtime_id IS NULL: a running row with no runtime is by definition not
-- proving liveness, so the wall clock is allowed to fire — same shape as
-- the legacy pure-wall-clock behavior for that (rare / historical) case.
--
-- waiting_local_directory rows are intentionally excluded: the daemon owns
-- the wait (with its own ctx-driven timeout) and a legitimate queue ahead
-- of this task can exceed the dispatch / running timeouts without being
-- "stuck". If the daemon dies, RecoverOrphanedTasksForRuntime reclaims
-- those rows at restart.
UPDATE agent_task_queue
SET status = 'failed', completed_at = now(), error = 'task timed out',
    failure_reason = 'timeout',
    prepare_lease_expires_at = NULL
WHERE (
    status = 'dispatched'
    AND dispatched_at < now() - make_interval(secs => @dispatch_timeout_secs::double precision)
    AND (prepare_lease_expires_at IS NULL OR prepare_lease_expires_at < now())
    AND (
      runtime_id IS NULL
      OR NOT EXISTS (
        SELECT 1 FROM agent_runtime r
        WHERE r.id = agent_task_queue.runtime_id
      )
      OR EXISTS (
        SELECT 1 FROM agent_runtime r
        WHERE r.id = agent_task_queue.runtime_id
          AND (
            (
              r.status = 'online'
              AND COALESCE(r.last_seen_at, r.updated_at) >=
                  now() - make_interval(secs => @runtime_stale_secs::double precision)
            )
            OR COALESCE(r.last_seen_at, r.updated_at) <
               now() - make_interval(secs => @runtime_reconnect_grace_secs::double precision)
          )
      )
    )
  )
   OR (
    status = 'running'
    AND started_at < now() - make_interval(secs => @running_timeout_secs::double precision)
    AND (
      runtime_id IS NULL
      OR NOT EXISTS (
        SELECT 1 FROM agent_runtime r
        WHERE r.id = agent_task_queue.runtime_id
          AND COALESCE(r.last_seen_at, r.updated_at) >=
              now() - make_interval(secs => @runtime_reconnect_grace_secs::double precision)
      )
    )
  )
RETURNING *;

-- name: ExpireStaleQueuedTasks :many
-- Fails queued tasks whose runtime can no longer prove it is alive.
--
-- This used to be a pure wall clock: queued for longer than a TTL (default 2h)
-- meant failed. That conflated "nobody is coming for this task" with "the
-- queue ahead of it is long", and the second one is not a failure. MUL-6558
-- was exactly that — a self-hosted runtime with low task concurrency held its
-- own queue past 2h and healthy work died as queued_expired. The TTL knob
-- added then (MULTICA_TASK_QUEUED_TTL) only moved the cliff; it did not stop
-- a busy runtime from eventually crossing it.
--
-- So the question is not "how long has this waited" but "is anything still
-- able to pick it up". A runtime that keeps heartbeating is busy, not dead,
-- and its backlog must be allowed to drain however long that takes. The
-- liveness signal is the same one FailTasksForOfflineRuntimes uses for
-- dispatched/running rows, so a daemon going down now retires its queued and
-- its in-flight work on one clock instead of two.
--
-- The row must ALSO have been queued for a full grace of its own. Enqueue binds
-- a task to agent.runtime_id without checking that the runtime is up, so
-- runtime liveness alone would fail a task the instant it is assigned to a
-- runtime that has been offline a while — a laptop closed overnight would turn
-- "assign this issue" into a failure inside one 30s sweep tick instead of
-- waiting for the machine to come back. Requiring the task's own age keeps the
-- promise the name makes: a queued task gets one full reconnect grace before it
-- is given up on, counted from when it started waiting. It also bounds the
-- scan, which would otherwise re-evaluate the runtime subquery against every
-- queued row on every tick.
--
-- Heartbeat age is read directly rather than gated on runtime.status='online',
-- so a row stuck at 'online' with a long-dead heartbeat still releases its
-- queue.
--
-- The runtime_id IS NULL / missing-runtime arms are fail-closed defence, not
-- live paths: the schema already excludes both. runtime_id was NOT NULL from
-- migration 004 until 251 replaced it with CHECK (runtime_id IS NOT NULL OR
-- completed_at IS NOT NULL), which every insert and update is checked against
-- (NOT VALID only skips the backfill scan), so no queued row can be unbound.
-- A dangling reference is impossible because agent_task_queue_runtime_id_fkey
-- (migration 004) is ON DELETE CASCADE: deleting a runtime removes the rows
-- referencing it rather than orphaning them, and the normal delete path
-- (migration 251, MUL-5559) explicitly unbinds history tasks first in the same
-- transaction. They are kept so a future schema change cannot silently strand
-- rows that have no liveness signal at all.
--
-- A retry created by runtime_offline is exempt: it deliberately waits for that
-- runtime to reconnect, and FailExpiredRuntimeReconnectRetries owns its exit.
--
-- Concurrency safety: the daemon's claim path may race with this sweeper to
-- transition the same row out of 'queued'. We protect against that two
-- ways:
--   1. The CTE selects victims with FOR UPDATE SKIP LOCKED so a row that is
--      currently being claimed (or otherwise locked) is skipped — no lock
--      contention with the dispatch path, and we won't queue up behind it.
--   2. The outer UPDATE re-checks status='queued' AND the liveness predicate at
--      apply time. If a daemon claimed the row between selection and update
--      (e.g. lock released after the claim transaction commits), the row is
--      already 'dispatched'/'running' and the WHERE clause filters it out
--      so we cannot clobber an in-flight task.
-- Capped via LIMIT inside the CTE so a single sweep tick cannot monopolise
-- the DB when the backlog is large — the sweeper drains the rest on
-- subsequent ticks.
WITH victims AS (
    SELECT id FROM agent_task_queue
    WHERE status = 'queued'
      AND created_at < now() - make_interval(secs => @reconnect_grace_secs::double precision)
      AND (
          runtime_id IS NULL
          OR NOT EXISTS (
              SELECT 1 FROM agent_runtime r WHERE r.id = agent_task_queue.runtime_id
          )
          OR EXISTS (
              SELECT 1 FROM agent_runtime r
              WHERE r.id = agent_task_queue.runtime_id
                AND COALESCE(r.last_seen_at, r.updated_at) <
                    now() - make_interval(secs => @reconnect_grace_secs::double precision)
          )
      )
      AND NOT EXISTS (
          SELECT 1 FROM agent_task_queue retry_parent
          WHERE retry_parent.id = agent_task_queue.parent_task_id
            AND retry_parent.failure_reason = 'runtime_offline'
      )
    ORDER BY created_at ASC
    LIMIT @max_per_tick::int
    FOR UPDATE SKIP LOCKED
)
UPDATE agent_task_queue t
SET status = 'failed',
    completed_at = now(),
    error = 'runtime unavailable while task was queued',
    failure_reason = 'queued_expired',
    prepare_lease_expires_at = NULL
FROM victims v
WHERE t.id = v.id
  AND t.status = 'queued'
  AND t.created_at < now() - make_interval(secs => @reconnect_grace_secs::double precision)
  AND (
      t.runtime_id IS NULL
      OR NOT EXISTS (SELECT 1 FROM agent_runtime r WHERE r.id = t.runtime_id)
      OR EXISTS (
          SELECT 1 FROM agent_runtime r
          WHERE r.id = t.runtime_id
            AND COALESCE(r.last_seen_at, r.updated_at) <
                now() - make_interval(secs => @reconnect_grace_secs::double precision)
      )
  )
  AND NOT EXISTS (
      SELECT 1 FROM agent_task_queue retry_parent
      WHERE retry_parent.id = t.parent_task_id
        AND retry_parent.failure_reason = 'runtime_offline'
  )
RETURNING t.*;

-- name: FailExpiredRuntimeReconnectRetries :many
-- A runtime_offline retry waits in deferred until its runtime returns healthy.
-- Give that recovery state a bounded exit: after one full reconnect grace,
-- fail it with a non-retryable reason so issues, agents, and runtime GC can
-- converge. A runtime that is healthy when this query runs wins the race and
-- remains deferred for the next daemon poll to promote. The parent join is the
-- lineage discriminator; provider backoff and other deferred task types are
-- intentionally excluded.
WITH victims AS (
    SELECT retry.id
    FROM agent_task_queue retry
    JOIN agent_task_queue parent ON parent.id = retry.parent_task_id
    WHERE retry.status = 'deferred'
      AND retry.fire_at < now() - make_interval(secs => @reconnect_grace_secs::double precision)
      AND parent.failure_reason = 'runtime_offline'
      AND NOT EXISTS (
          SELECT 1 FROM agent_runtime runtime
          WHERE runtime.id = retry.runtime_id
            AND runtime.status = 'online'
            AND COALESCE(runtime.last_seen_at, runtime.updated_at) >=
                now() - make_interval(secs => @runtime_stale_secs::double precision)
      )
    ORDER BY retry.fire_at, retry.created_at
    LIMIT @max_per_tick::int
    FOR UPDATE OF retry SKIP LOCKED
)
UPDATE agent_task_queue AS retry
SET status = 'failed',
    completed_at = now(),
    error = 'runtime did not reconnect within the configured grace period',
    failure_reason = 'runtime_reconnect_timeout',
    wait_reason = NULL,
    prepare_lease_expires_at = NULL
FROM victims
WHERE retry.id = victims.id
  AND retry.status = 'deferred'
  AND retry.fire_at < now() - make_interval(secs => @reconnect_grace_secs::double precision)
  AND EXISTS (
      SELECT 1 FROM agent_task_queue parent
      WHERE parent.id = retry.parent_task_id
        AND parent.failure_reason = 'runtime_offline'
  )
  AND NOT EXISTS (
      SELECT 1 FROM agent_runtime runtime
      WHERE runtime.id = retry.runtime_id
        AND runtime.status = 'online'
        AND COALESCE(runtime.last_seen_at, runtime.updated_at) >=
            now() - make_interval(secs => @runtime_stale_secs::double precision)
  )
RETURNING retry.*;

-- name: CancelAgentTask :one
-- Automatic cancellation without an explicit persisted failure reason. Unlike
-- CancelAgentTaskByUser, this deliberately leaves recovery inputs replayable.
UPDATE agent_task_queue
SET status = 'cancelled', completed_at = now(), prepare_lease_expires_at = NULL
WHERE id = $1 AND status IN ('queued', 'dispatched', 'running', 'waiting_local_directory', 'deferred')
RETURNING *;

-- name: CancelAgentTaskByUser :one
-- An explicit user cancellation is a terminal acknowledgement for any
-- delegated-failure recovery signal planned into this task. Server-initiated
-- cancellations keep using CancelAgentTask / CancelAgentTaskWithReason so a
-- stale plan, claim failure, or other automatic repair can still be replayed.
UPDATE agent_task_queue AS task
SET status = 'cancelled',
    completed_at = now(),
    prepare_lease_expires_at = NULL,
    delivered_comment_ids = CASE
      -- Chat and ordinary issue tasks almost never carry a delegated-failure
      -- recovery signal. Keep their high-frequency user-cancel path to a
      -- no-join update; only validate task lineage after the cheap comment
      -- shape probe finds a possible recovery signal.
      WHEN task.trigger_comment_id IS NULL
       AND COALESCE(cardinality(task.coalesced_comment_ids), 0) = 0
        THEN task.delivered_comment_ids
      WHEN NOT EXISTS (
        SELECT 1
        FROM comment recovery_signal
        WHERE (
            recovery_signal.id = task.trigger_comment_id
            OR recovery_signal.id = ANY(task.coalesced_comment_ids)
        )
          AND recovery_signal.author_type = 'system'
          AND recovery_signal.type = 'progress_update'
          AND recovery_signal.source_task_id IS NOT NULL
      ) THEN task.delivered_comment_ids
      ELSE (
        SELECT COALESCE(array_agg(DISTINCT receipt.id), '{}')::uuid[]
        FROM unnest(array_cat(
            task.delivered_comment_ids,
            ARRAY(
                SELECT recovery.id
                FROM comment recovery
                JOIN agent_task_queue failed ON failed.id = recovery.source_task_id
                JOIN agent_task_queue source ON source.id = failed.delegated_from_task_id
                WHERE (
                    recovery.id = task.trigger_comment_id
                    OR recovery.id = ANY(task.coalesced_comment_ids)
                )
                  AND recovery.author_type = 'system'
                  AND recovery.type = 'progress_update'
                  AND recovery.source_task_id IS NOT NULL
                  AND failed.status = 'failed'
                  AND failed.delegated_from_task_id IS NOT NULL
                  AND failed.autopilot_run_id IS NULL
                  AND failed.trigger_evidence_kind IS DISTINCT FROM 'delegated_failure'
                  AND source.autopilot_run_id IS NULL
                  AND source.issue_id = task.issue_id
                  AND source.agent_id = task.agent_id
                  AND recovery.issue_id = source.issue_id
            )
        )) AS receipt(id)
      )
    END
WHERE task.id = $1
  AND task.status IN ('queued', 'dispatched', 'running', 'waiting_local_directory', 'deferred')
RETURNING task.*;

-- name: SetAgentTaskBranchName :exec
-- Records the delivered branch on a CANCELLED task. Needed because the daemon
-- finalizes its worktree (committing whatever the agent produced) BEFORE it
-- learns the task was cancelled, so the branch exists but the cancel path has
-- no result payload to carry it. Without this the branch is real and
-- completely undiscoverable from the UI.
--
-- Never overwrites a name already recorded by a complete/fail callback, and
-- never touches any OTHER terminal state: the daemon acks on every terminal
-- status it observes, so a late or replayed ack from a stale run could
-- otherwise write its branch into a completed/failed row whose own callback —
-- the authoritative channel for those states — recorded nothing. status is
-- stable once terminal, so this CAS cannot race a legitimate write.
UPDATE agent_task_queue
SET branch_name = COALESCE(branch_name, sqlc.arg('branch_name'))
WHERE id = sqlc.arg('id') AND status = 'cancelled';

-- name: SetAgentTaskDurableWorkDir :exec
-- Records the durable replacement for a disposable worktree on a CANCELLED
-- task. The daemon only sends this after Finalize confirms the task worktree
-- is gone. As with branch_name, the status CAS rejects stale acknowledgements
-- and COALESCE makes replays idempotent.
UPDATE agent_task_queue
SET durable_work_dir = COALESCE(durable_work_dir, sqlc.arg('durable_work_dir'))
WHERE id = sqlc.arg('id') AND status = 'cancelled';

-- name: SetAgentTaskErrorIfEmpty :exec
-- Companion to SetAgentTaskBranchName for the cancel-ack path. A cancelled
-- worktree task whose Finalize ABORTED has no branch to deliver — the error
-- text carrying the preserved-worktree path is the only pointer to the agent's
-- work, and the cancel flow discarded the rest of the result. A user-initiated
-- cancel leaves error NULL, so filling it here never overwrites a reason
-- recorded by a fail callback or the claim gate; the status CAS keeps a late
-- ack from stamping an error onto a completed/failed row (see
-- SetAgentTaskBranchName above).
UPDATE agent_task_queue
SET error = sqlc.arg('error'),
    failure_reason = COALESCE(failure_reason, sqlc.arg('failure_reason'))
WHERE id = sqlc.arg('id') AND (error IS NULL OR error = '') AND status = 'cancelled';

-- name: CancelAgentTaskWithReason :one
-- Cancels a task AND records why, for cancellations the user did not ask for.
--
-- CancelAgentTaskByUser leaves error/failure_reason NULL because the user knows
-- why. A server-initiated refusal is the opposite: without a persisted reason
-- the run surfaces as an unexplained "cancelled", and the only trace is a 4xx
-- in a daemon log the user never sees. Retrying cannot help either (the task is
-- refused for a durable reason), so this terminal state carries its explanation.
UPDATE agent_task_queue
SET status = 'cancelled',
    completed_at = now(),
    error = sqlc.arg('error'),
    failure_reason = sqlc.arg('failure_reason'),
    prepare_lease_expires_at = NULL
WHERE id = sqlc.arg('id') AND status IN ('queued', 'dispatched', 'running', 'waiting_local_directory', 'deferred')
RETURNING *;

-- name: CancelQueuedAgentTask :one
-- Queue editing is a compare-and-set: never cancel a task that the daemon
-- promoted between the user's click and this statement.
UPDATE agent_task_queue
SET status = 'cancelled', completed_at = now(), prepare_lease_expires_at = NULL
WHERE id = sqlc.arg('id')
  AND chat_session_id = sqlc.arg('chat_session_id')
  AND status = 'queued'
RETURNING *;

-- name: CancelQueuedAgentTasksForSession :many
-- Clear only positional follow-ups. The first visible task is current even
-- when it has not been claimed yet, so an all-queued session must preserve its
-- priority/FIFO head. Keep this selector in lockstep with the visible-head
-- invariant documented above ListChatMessages in chat.sql.
WITH head AS MATERIALIZED (
  SELECT candidate.id
  FROM agent_task_queue AS candidate
  WHERE candidate.chat_session_id = $1
    AND candidate.status IN ('queued', 'dispatched', 'running', 'waiting_local_directory', 'deferred')
    AND candidate.regenerate_quick_actions_for IS NULL
  ORDER BY
    CASE
      WHEN candidate.status IN ('dispatched', 'running', 'waiting_local_directory') THEN 0
      WHEN candidate.status = 'deferred' THEN 1
      ELSE 2
    END,
    candidate.priority DESC,
    candidate.created_at ASC,
    candidate.id ASC
  LIMIT 1
)
UPDATE agent_task_queue AS queued
SET status = 'cancelled', completed_at = now(), prepare_lease_expires_at = NULL
WHERE queued.chat_session_id = $1
  AND queued.status = 'queued'
  AND queued.id IS DISTINCT FROM (SELECT id FROM head)
RETURNING queued.*;

-- name: MarkChatFinalizeDeferred :one
-- Arms the deferred chat-finalize marker for a cancelled chat task whose
-- empty-transcript judgment must wait for the daemon's flush ack (#5219).
UPDATE agent_task_queue
SET chat_finalize_deferred_at = now()
WHERE id = $1
RETURNING *;

-- name: ClaimChatFinalizeDeferred :one
-- Atomically claims the deferred marker so the daemon ack and the sweeper
-- cannot both finalize the same task (double-"Stopped." guard).
UPDATE agent_task_queue
SET chat_finalize_deferred_at = NULL
WHERE id = $1 AND chat_finalize_deferred_at IS NOT NULL
RETURNING *;

-- name: ListChatFinalizeDeferredExpired :many
-- Deferred chat finalizations whose grace period elapsed without a daemon
-- ack (dead or partitioned daemon). Batch-capped like the other sweeper
-- queries so one tick can't monopolise the DB.
SELECT * FROM agent_task_queue
WHERE chat_finalize_deferred_at IS NOT NULL
  AND chat_finalize_deferred_at < now() - make_interval(secs => @grace_secs::double precision)
ORDER BY chat_finalize_deferred_at
LIMIT @max_per_tick::int;

-- name: CountRunningTasks :one
-- waiting_local_directory remains capacity-bearing until resume admission has
-- an atomic reservation/CAS gate. Consequently a waiter-only agent is reported
-- idle by RefreshAgentStatusFromTasks but still cannot claim additional work;
-- removing it here alone could exceed max_concurrent_tasks when it resumes.
SELECT count(*) FROM agent_task_queue
WHERE agent_id = $1 AND status IN ('dispatched', 'running', 'waiting_local_directory');

-- name: GetAgentForClaimUpdate :one
SELECT * FROM agent
WHERE id = $1
FOR UPDATE;

-- name: HasActiveTaskForIssue :one
-- Returns true if there is any queued, dispatched, waiting_local_directory,
-- or running task for the issue.
SELECT count(*) > 0 AS has_active FROM agent_task_queue
WHERE issue_id = $1 AND status IN ('queued', 'dispatched', 'running', 'waiting_local_directory');

-- name: HasPendingTaskForIssue :one
-- Returns true if there is a queued or dispatched (but not yet running) task for the issue.
-- Used by the coalescing queue: allow enqueue when a task is running (so
-- the agent picks up new comments on the next cycle) but skip if a pending
-- task already exists (natural dedup).
SELECT count(*) > 0 AS has_pending FROM agent_task_queue
WHERE issue_id = $1 AND status IN ('queued', 'dispatched');

-- name: HasPendingTaskForIssueAndAgent :one
-- Returns true if a specific agent already has a queued or dispatched task
-- for the given issue, or the explicitly-marked deferred task whose channel
-- media must bind before the issue agent runs. Used by @mention trigger dedup.
--
-- head_sha keys the dedup on the commit under review (TEN-356): when a caller
-- passes a non-empty head_sha, a pending task only dedups if it was stamped
-- with the SAME head_sha at enqueue time. If HEAD advanced since the pending
-- task's run began (its context head_sha differs, or predates the stamp and is
-- NULL), the dedup MISSES and a fresh review enqueues against the new HEAD.
-- When head_sha is empty/NULL (issue has no linked PR) the check falls back to
-- the pre-TEN-356 (issue_id, agent_id) key so non-PR issues keep coalescing.
SELECT count(*) > 0 AS has_pending FROM agent_task_queue
WHERE issue_id = $1 AND agent_id = $2
  AND (
    status IN ('queued', 'dispatched')
    OR (status = 'deferred' AND context->>'channel_issue_media_pending' = 'true')
  )
  AND (
    COALESCE(sqlc.narg('head_sha')::text, '') = ''
    OR context->>'head_sha' = sqlc.narg('head_sha')::text
  );

-- name: HasPendingTaskForIssueAndAgentExcludingTriggerComment :one
-- Same as HasPendingTaskForIssueAndAgent, but ignores tasks triggered by the
-- current comment being edited. Edit preview needs this because save cancels
-- that comment's old queued/dispatched tasks before re-computing triggers.
-- Carries the same head_sha dedup key as HasPendingTaskForIssueAndAgent (TEN-356).
SELECT count(*) > 0 AS has_pending FROM agent_task_queue
WHERE issue_id = @issue_id
  AND agent_id = @agent_id
  AND (
    status IN ('queued', 'dispatched')
    OR (status = 'deferred' AND context->>'channel_issue_media_pending' = 'true')
  )
  AND trigger_comment_id IS DISTINCT FROM @exclude_trigger_comment_id::uuid
  AND (
    COALESCE(sqlc.narg('head_sha')::text, '') = ''
    OR context->>'head_sha' = sqlc.narg('head_sha')::text
  );

-- name: MergeCommentIntoPendingTask :one
-- MUL-4195: fold a newly-arrived comment into an existing task for (issue,
-- agent) that has NOT yet been claimed, instead of letting the
-- HasPendingTaskForIssueAndAgent dedup silently DROP it. The task's prior
-- trigger_comment_id becomes a coalesced ("also cover") comment and
-- @new_trigger_comment_id becomes the new trigger, so the injected prompt shows
-- the latest deliberate instruction while the single run is still told to
-- address every folded comment.
--
-- Target is restricted to a pre-claim task on purpose (MUL-4195 review rounds
-- 2–4). This merge is reached when HasPendingTaskForIssueAndAgent matched a
-- 'queued'/'dispatched' task, or the channel-media deferred task described
-- below. 'dispatched' is deliberately NOT a
-- target: a dispatched / waiting_local_directory / running task has already had
-- its claim response built. Folding afterward would add a planned id that is
-- absent from that response's delivered_comment_ids receipt; completion
-- reconciliation handles it instead. Ordinary 'deferred' rows remain excluded:
-- assignee-fallback escalations have their own fire_at lifecycle and may coexist
-- with a queued primary. The one exception is a task explicitly marked
-- channel_issue_media_pending. It is the issue's sole assigned-agent task and
-- has not been claimable yet, so a new comment must merge into it instead of
-- creating a competing queued task. coalesced_comment_ids remains the pre-claim
-- plan; the claim path records the actual embedded subset in
-- delivered_comment_ids.
--
-- Recompute-on-merge (MUL-4195 review must-fix #1): originator_user_id,
-- runtime_mcp_overlay and runtime_connected_apps are re-stamped to the NEW
-- trigger comment's originator (computed by the caller). Earlier this only
-- repointed the trigger and kept the old originator's overlay/attribution, so a
-- run answering user B's comment could execute under user A's connected-app
-- capabilities and audit identity. Re-stamping means the single coalescing run
-- carries the latest deliberate instruction's originator and the matching
-- overlay — no cross-user capability bleed, no stale attribution. This also
-- removes the previous originator gate + fresh-enqueue fallback, which could not
-- create a second task anyway (the idx_one_pending_task_per_issue_agent unique
-- index allows only one queued/dispatched task per (issue, agent)) and therefore
-- silently dropped the mismatched-originator comment.
--
-- Returns pgx.ErrNoRows when no queued task exists (it was claimed/started
-- between the dedup check and this call, or the only task is already
-- dispatched/running). The caller must NOT blindly enqueue a fresh task in that
-- case — a dispatched sibling would trip the unique index — it defers to
-- completion reconciliation unless no active task exists at all.
UPDATE agent_task_queue
SET coalesced_comment_ids = (
        SELECT COALESCE(array_agg(DISTINCT e), '{}')
        FROM unnest(array_append(coalesced_comment_ids, trigger_comment_id)) AS e
        WHERE e IS NOT NULL AND e <> @new_trigger_comment_id::uuid
    ),
    trigger_comment_id = @new_trigger_comment_id::uuid,
    trigger_summary = COALESCE(sqlc.narg('new_trigger_summary'), trigger_summary),
    -- Re-attribution is ATOMIC (MUL-4302): folding a newly-arrived comment moves the
    -- WHOLE attribution snapshot to that comment's human — person columns, source
    -- label, delegation lineage, rule version, and evidence — computed by the caller
    -- as one attribution.Result. Re-stamping only the person columns would leave a
    -- run showing B accountable while still pointing at A's stale source / evidence /
    -- level. accountable comes from the resolved Result (finalizeAttribution already
    -- guaranteed originator ⟹ accountable == originator; the cross-column CHECK backs it).
    originator_user_id = sqlc.narg('new_originator_user_id')::uuid,
    accountable_user_id = sqlc.narg('new_accountable_user_id')::uuid,
    originator_source = sqlc.narg('new_originator_source'),
    delegated_from_task_id = sqlc.narg('new_delegated_from_task_id')::uuid,
    rule_version_id = sqlc.narg('new_rule_version_id')::uuid,
    trigger_evidence_kind = sqlc.narg('new_trigger_evidence_kind'),
    trigger_evidence_ref_id = sqlc.narg('new_trigger_evidence_ref_id')::uuid,
    runtime_mcp_overlay = sqlc.narg('new_runtime_mcp_overlay'),
    runtime_connected_apps = sqlc.narg('new_runtime_connected_apps')
WHERE id = (
    SELECT t.id FROM agent_task_queue t
    WHERE t.issue_id = @issue_id
      AND t.agent_id = @agent_id
      AND (
          t.status = 'queued'
          OR (t.status = 'deferred' AND t.context->>'channel_issue_media_pending' = 'true')
      )
      -- Head-scoped (TEN-356, #5914): never fold across HEADs. The physical
      -- unique index is only (issue_id, agent_id), so an insert-race loser can
      -- collide with a pending task stamped for a DIFFERENT head_sha; merging
      -- into it would give a new-HEAD comment old-HEAD review coverage. Empty/
      -- absent head_sha (no linked PR) matches any task, preserving coalescing.
      AND (
          COALESCE(sqlc.narg('head_sha')::text, '') = ''
          OR t.context->>'head_sha' = sqlc.narg('head_sha')::text
      )
    ORDER BY t.created_at DESC
    LIMIT 1
)
RETURNING id, coalesced_comment_ids;

-- name: RegisterPlannedCommentForActiveTask :one
-- #5914: durably register a comment that lost an enqueue race as a PLANNED
-- (undelivered) input on the same-(issue, agent) ACTIVE task whose queued row
-- MergeCommentIntoPendingTask could no longer target (it was claimed →
-- dispatched/running). The losing comment can predate that task's created_at,
-- so completion reconciliation's `created_at > since` window cannot see it;
-- appending it to coalesced_comment_ids (the planned set) WITHOUT touching
-- delivered_comment_ids makes reconcileCommentsOnCompletion replay it as a
-- single bounded follow-up (planned-but-not-delivered ⇒ follow-up).
--
-- Head-scoped (TEN-356): only a task stamped with the SAME head_sha is a target,
-- so a new-HEAD comment is never attached to an old-HEAD run. Returns
-- pgx.ErrNoRows when no same-head active task exists (different HEAD, or the task
-- just terminated) so the caller falls back to the active-task decision.
--
-- 'queued' is DELIBERATELY EXCLUDED (#5914, Elon round 3): a not-yet-claimed
-- task has no claim receipt, so a comment merged into it WILL be recorded as
-- delivered at claim time and never earns a completion follow-up under its own
-- attribution. A queued target must therefore go through the ATOMIC
-- MergeCommentIntoPendingTask (which re-stamps trigger/originator/accountable/
-- overlay), never a bare planned append — otherwise a second member's comment
-- could execute under the first member's identity/connected-apps (MUL-4302).
-- Only claim-receipt statuses (already-built delivered set) are safe planned-id
-- targets.
UPDATE agent_task_queue
SET coalesced_comment_ids = (
        SELECT COALESCE(array_agg(DISTINCT e), '{}')
        FROM unnest(array_append(coalesced_comment_ids, @comment_id::uuid)) AS e
        WHERE e IS NOT NULL
    )
WHERE id = (
    SELECT t.id FROM agent_task_queue t
    WHERE t.issue_id = @issue_id
      AND t.agent_id = @agent_id
      AND t.status IN ('dispatched', 'running', 'waiting_local_directory')
      AND (
          COALESCE(sqlc.narg('head_sha')::text, '') = ''
          OR t.context->>'head_sha' = sqlc.narg('head_sha')::text
      )
    ORDER BY t.created_at DESC
    LIMIT 1
)
RETURNING id, coalesced_comment_ids;

-- name: MergeDelegatedFailureCommentIntoPendingTask :one
-- A delegated failure is a platform-owned input, not a new human instruction:
-- fold it into the coordinator's pre-claim task without replacing that task's
-- attribution snapshot. The recovery comment is the newest input, so it becomes
-- the trigger and the prior trigger joins the coalesced plan. This preserves the
-- claim/prompt invariant that trigger_comment_id names the newest comment while
-- retaining the pending task's original human authority and connected apps.
UPDATE agent_task_queue
SET coalesced_comment_ids = (
        SELECT COALESCE(array_agg(DISTINCT e), '{}')
        FROM unnest(array_append(coalesced_comment_ids, trigger_comment_id)) AS e
        WHERE e IS NOT NULL AND e <> @comment_id::uuid
    ),
    trigger_comment_id = @comment_id::uuid,
    trigger_summary = sqlc.narg('trigger_summary')
WHERE id = (
    SELECT t.id FROM agent_task_queue t
    WHERE t.issue_id = @issue_id
      AND t.agent_id = @agent_id
      AND (
          t.status = 'queued'
          OR (t.status = 'deferred' AND t.context->>'channel_issue_media_pending' = 'true')
      )
      AND t.trigger_comment_id IS DISTINCT FROM @comment_id::uuid
      AND NOT (@comment_id::uuid = ANY(t.coalesced_comment_ids))
    ORDER BY t.created_at DESC
    LIMIT 1
)
RETURNING *;

-- name: HasTaskCoveringDelegatedFailureComment :one
-- Durable idempotency check for a recovery comment. The completion reconciler
-- excludes its own just-completed task when replaying a planned-but-undelivered
-- signal. A planned comment is covered only while its task can still execute;
-- after a task becomes terminal, delivered_comment_ids is the sole durable
-- receipt. This prevents cancelled/failed pre-delivery tasks from swallowing
-- the recovery obligation while still avoiding a loop after real delivery.
SELECT count(*) > 0 AS covered
FROM agent_task_queue
WHERE issue_id = @issue_id
  AND agent_id = @agent_id
  AND (
      @comment_id::uuid = ANY(delivered_comment_ids)
      OR (
          id IS DISTINCT FROM sqlc.narg('exclude_task_id')::uuid
          AND (
              status IN ('queued', 'dispatched', 'running', 'waiting_local_directory')
              OR (status = 'deferred' AND context->>'channel_issue_media_pending' = 'true')
          )
          AND (trigger_comment_id = @comment_id::uuid OR @comment_id::uuid = ANY(coalesced_comment_ids))
      )
  );

-- name: CountDelegatedFailureRecoveryTasks :one
-- Counts dedicated coordinator wakeups for one failed delegated task. Merged
-- recovery signals do not create a new row and therefore do not consume an
-- automatic attempt until they need a standalone successor.
SELECT count(*)
FROM agent_task_queue
WHERE trigger_evidence_kind = 'delegated_failure'
  AND trigger_evidence_ref_id = @failed_task_id;

-- name: AcknowledgeExhaustedDelegatedFailureRecovery :one
-- Once the bounded automatic attempts are exhausted, record a terminal
-- acknowledgement on the newest recovery task. The outbox treats this receipt
-- as settled, while the service writes a separate visible system comment that
-- tells the user why no further task will be generated.
UPDATE agent_task_queue AS acknowledged
SET delivered_comment_ids = (
    SELECT COALESCE(array_agg(DISTINCT receipt.id), '{}')::uuid[]
    FROM unnest(array_append(acknowledged.delivered_comment_ids, @comment_id::uuid)) AS receipt(id)
)
WHERE acknowledged.id = (
    SELECT attempt.id
    FROM agent_task_queue attempt
    WHERE attempt.trigger_evidence_kind = 'delegated_failure'
      AND attempt.trigger_evidence_ref_id = @failed_task_id
    ORDER BY attempt.created_at DESC, attempt.id DESC
    LIMIT 1
    FOR UPDATE
)
  AND (
      SELECT count(*)
      FROM agent_task_queue attempt_count
      WHERE attempt_count.trigger_evidence_kind = 'delegated_failure'
        AND attempt_count.trigger_evidence_ref_id = @failed_task_id
  ) >= @max_attempts::int
RETURNING acknowledged.*;

-- name: SettleDelegatedFailureRecoveriesForTask :execrows
-- Retire every delegated-failure recovery comment this task actually received.
--
-- Callers MUST run this inside the same transaction that makes the task
-- terminal, and only there. While a task is still dispatched its receipt is
-- REPLACED rather than appended (SetTaskDeliveredCommentIDs), because a reclaim
-- by a differently-capable daemon must be able to drop an id it will not
-- deliver. Settling at receipt-write time would freeze that legitimate
-- transient window into a permanently lost recovery. Once the task is terminal
-- the receipt is final, which is exactly the guarantee this marker records.
--
-- A comment that was only planned into the task and never delivered is absent
-- from delivered_comment_ids, so an automatic cancellation correctly leaves it
-- pending for the sweeper to replay.
--
-- The terminal-status predicate is the guarantee itself, not a caller
-- convention. Settling a dispatched or running task's receipt would freeze the
-- reclaim window into a permanently lost recovery, and this marker is
-- monotonic — there is no later pass that could undo it. A caller that passes a
-- non-terminal task therefore updates nothing rather than silently destroying
-- the obligation.
UPDATE comment recovery
SET recovery_settled_at = now()
FROM agent_task_queue task
WHERE task.id = @task_id
  AND task.status IN ('completed', 'failed', 'cancelled')
  AND recovery.id = ANY(task.delivered_comment_ids)
  AND recovery.author_type = 'system'
  AND recovery.type = 'progress_update'
  AND recovery.source_task_id IS NOT NULL
  AND recovery.recovery_settled_at IS NULL;

-- name: SettleDelegatedFailureRecoveryComment :execrows
-- Retire one recovery comment by id, for the terminal outcomes that are not a
-- task reaching a terminal status: exhaustion of the bounded automatic attempts
-- writes its receipt onto an attempt row that may still be running, and the
-- user-visible explanation comment is what closes the obligation. Callers run
-- this in the same transaction that records that outcome.
UPDATE comment
SET recovery_settled_at = now()
WHERE id = @comment_id
  AND author_type = 'system'
  AND type = 'progress_update'
  AND source_task_id IS NOT NULL
  AND recovery_settled_at IS NULL;

-- name: ListPendingDelegatedFailureRecoveries :many
-- Durable outbox scan for platform recovery comments that are not yet owned by
-- an executable task and have no terminal delivery receipt. Starting from the
-- explicit recovery signal avoids retroactively waking unrelated historical
-- delegated failures. A bounded runtime sweeper replays these comments after a
-- transient dispatch error or process restart.
--
-- recovery_settled_at is what keeps this scan from growing with history. Every
-- other exclusion below is reversible — an issue can leave 'done', a cancelled
-- retry can be superseded — so none of them can be frozen into the index
-- predicate. A delivery receipt on a terminal task cannot be taken back, so
-- that one condition is recorded as durable state instead of being re-proven
-- through four joins and two NOT EXISTS subqueries on every tick. The predicate
-- of idx_comment_delegated_failure_unsettled matches the first four conditions,
-- so LIMIT now bounds the rows CHECKED and not just the rows RETURNED.
SELECT recovery.*
FROM comment recovery
JOIN agent_task_queue failed ON failed.id = recovery.source_task_id
JOIN agent_task_queue source ON source.id = failed.delegated_from_task_id
JOIN issue source_issue ON source_issue.id = source.issue_id
JOIN agent source_agent ON source_agent.id = source.agent_id
LEFT JOIN issue_status source_status
  ON source_status.workspace_id = source_issue.workspace_id
 AND source_status.key = source_issue.status
WHERE recovery.author_type = 'system'
  AND recovery.type = 'progress_update'
  AND recovery.source_task_id IS NOT NULL
  AND recovery.recovery_settled_at IS NULL
  AND recovery.issue_id = source_issue.id
  AND recovery.workspace_id = source_issue.workspace_id
  AND failed.status = 'failed'
  AND failed.delegated_from_task_id IS NOT NULL
  AND failed.autopilot_run_id IS NULL
  AND failed.trigger_evidence_kind IS DISTINCT FROM 'delegated_failure'
  AND source.autopilot_run_id IS NULL
  AND source.issue_id IS NOT NULL
  AND source.agent_id <> failed.agent_id
  AND COALESCE(source_status.category, source_issue.status) NOT IN ('done', 'cancelled', 'backlog')
  AND source_agent.archived_at IS NULL
  AND source_agent.runtime_id IS NOT NULL
  AND source_agent.workspace_id = source_issue.workspace_id
  AND NOT EXISTS (
      SELECT 1
      FROM agent_task_queue retry
      WHERE retry.parent_task_id = failed.id
        AND retry.status <> 'cancelled'
  )
  AND NOT EXISTS (
      SELECT 1
      FROM agent_task_queue covering
      WHERE covering.issue_id = source_issue.id
        AND covering.agent_id = source.agent_id
        AND (
            recovery.id = ANY(covering.delivered_comment_ids)
            OR (
                (
                    covering.status IN ('queued', 'dispatched', 'running', 'waiting_local_directory')
                    OR (
                        covering.status = 'deferred'
                        AND covering.context->>'channel_issue_media_pending' = 'true'
                    )
                )
                AND (
                    covering.trigger_comment_id = recovery.id
                    OR recovery.id = ANY(covering.coalesced_comment_ids)
                )
            )
        )
  )
ORDER BY recovery.created_at ASC, recovery.id ASC
LIMIT @max_per_tick;

-- name: HasActiveTaskForIssueAndAgent :one
-- MUL-4195: true when the (issue, agent) pair has any non-terminal task in a
-- state whose completion will run completion reconciliation — queued,
-- dispatched, running, waiting_local_directory, or the explicitly-marked
-- channel-media deferred state. Used by the comment enqueue
-- path: when a merge into a pre-claim task fails (the task is already
-- dispatched/running, or a mismatched pre-claim task exists), a fresh queued
-- INSERT would collide with idx_one_pending_task_per_issue_agent AND would risk
-- a duplicate run. Instead the caller relies on that active task's completion
-- reconcile to schedule the guaranteed follow-up, and only enqueues fresh when
-- NO active task exists.
SELECT count(*) > 0 AS has_active FROM agent_task_queue
WHERE issue_id = $1 AND agent_id = $2
  AND (
    status IN ('queued', 'dispatched', 'running', 'waiting_local_directory')
    OR (status = 'deferred' AND context->>'channel_issue_media_pending' = 'true')
  );

-- name: GetLatestTaskRoleForIssueAndAgent :one
-- Returns the role markers from the agent's most recent task on this issue.
-- Used by the squad-leader self-trigger guard to tell apart leader tasks,
-- same-squad worker tasks, and generic agent tasks such as direct mentions or
-- thread-parent replies.
SELECT is_leader_task, squad_id FROM agent_task_queue
WHERE issue_id = $1 AND agent_id = $2
ORDER BY created_at DESC
LIMIT 1;

-- name: ListPendingTasksByRuntime :many
SELECT * FROM agent_task_queue
WHERE runtime_id = $1 AND status IN ('queued', 'dispatched')
ORDER BY priority DESC, created_at ASC;

-- name: ListQueuedClaimCandidatesByRuntime :many
-- Returns rows the runtime is authorized to attempt to claim. Status is restricted to
-- 'queued' (in contrast to ListPendingTasksByRuntime which also includes
-- 'dispatched') because dispatched rows are by definition already owned
-- and cannot be re-claimed — including them in the candidate list pads
-- the result with rows that always lose the per-(issue, agent) race in
-- ClaimAgentTask, wasting CPU and a SELECT every poll cycle when the
-- runtime is busy on a long-running task. Backed by the partial index
-- idx_agent_task_queue_claim_candidates so the warm path is cheap.
SELECT atq.* FROM agent_task_queue atq
WHERE atq.runtime_id = $1
  AND atq.status = 'queued'
  AND EXISTS (
      -- Keep this authorization fence in sync with ClaimAgentTask.
      SELECT 1
      FROM agent a
      JOIN agent_runtime r ON r.id = atq.runtime_id
      WHERE a.id = atq.agent_id
        AND a.runtime_id = atq.runtime_id
        AND (
            r.visibility = 'public'
            OR (
                r.visibility = 'private'
                AND (
                    r.owner_id IS NULL
                    OR a.owner_id IS NULL
                    OR r.owner_id = a.owner_id
                )
            )
        )
  )
ORDER BY atq.priority DESC, atq.created_at ASC;

-- name: CancelSupersededDeferredRetriesForRuntimes :many
-- Cancels deferred auto-retry rows that a newer active task has already
-- superseded, so one rerun click still means exactly one more run.
--
-- The narrow race: a manual rerun clears the pending slot, a concurrent FailTask
-- commits a deferred retry (runtime_offline / provider_network's final attempt
-- arm fire_at), and the rerun's own enqueue then succeeds because 'deferred' is
-- outside idx_one_pending_task_per_issue_agent_v2. Both rows now exist. Merely
-- refusing to promote the retry is not enough: once the rerun stops occupying the
-- slot the retry promotes and runs a SECOND time, duplicating the agent's
-- comments, side effects and cost with nothing to signal it. That contradicts
-- both "a manual rerun replaces the pending plan" and FailTask's own "a runnable
-- successor already exists, so no retry is needed".
--
-- Scope is deliberately tight:
--   * retry_of_task_id IS NOT NULL — only auto-retry clones, never a fresh run.
--   * escalation_for_task_id IS NULL — assignee-fallback escalations own their
--     fire_at lifecycle and are SUPPOSED to coexist with an active primary.
--   * channel-media pending rows are excluded for the same reason: that deferred
--     row is the issue's own task waiting on media, not a superseded retry.
--   * issue_id IS NOT NULL — chat / quick-create tasks have no slot semantics.
--
-- 'running' and 'waiting_local_directory' count as superseding: the duplicate
-- fires precisely when the rerun has STARTED, which is when the row would
-- otherwise no longer look blocked.
UPDATE agent_task_queue r
SET status = 'cancelled', completed_at = now(), prepare_lease_expires_at = NULL
WHERE r.runtime_id = ANY(@runtime_ids::uuid[])
  AND r.status = 'deferred'
  AND r.issue_id IS NOT NULL
  AND r.retry_of_task_id IS NOT NULL
  AND r.escalation_for_task_id IS NULL
  AND COALESCE(r.context->>'channel_issue_media_pending', '') <> 'true'
  AND EXISTS (
    SELECT 1 FROM agent_task_queue successor
    WHERE successor.issue_id = r.issue_id
      AND successor.agent_id = r.agent_id
      AND successor.id <> r.id
      AND successor.status IN ('queued', 'dispatched', 'running', 'waiting_local_directory')
  )
RETURNING *;

-- name: PromoteDueDeferredTasksForRuntime :many
-- Promotion is fenced against the single queued/dispatched slot
-- idx_one_pending_task_per_issue_agent_v2 allows per (issue, agent). A deferred
-- row is NOT covered by that index, so it can legitimately coexist with a queued
-- one — a manual rerun enqueued behind a running task, plus the deferred retry
-- that task's failure armed (runtime_offline, provider_network's final attempt).
-- Flipping such a row to 'queued' unconditionally violates the index, and because
-- the claim loop promotes before it claims, that error blocked every claim on the
-- runtime — including the rerun the operator was waiting for.
--
-- Two fences: skip a row whose (issue, agent) slot is already occupied, and
-- promote at most ONE row per (issue, agent) so a single statement cannot collide
-- with itself. A skipped row stays deferred with fire_at in the past and is
-- promoted by a later tick once the slot frees, so nothing is lost — the human's
-- rerun simply goes first. Chat / quick-create rows (issue_id NULL) are outside
-- the index and bypass both fences.
WITH due AS (
    SELECT t.id,
           t.issue_id,
           row_number() OVER (
               PARTITION BY t.issue_id, t.agent_id
               ORDER BY t.priority DESC, t.created_at ASC, t.id
           ) AS rn
    FROM agent_task_queue t
    WHERE t.runtime_id = @runtime_id
      AND t.status = 'deferred'
      AND t.fire_at <= now()
      AND EXISTS (
        SELECT 1 FROM agent_runtime r
        WHERE r.id = t.runtime_id
          AND r.status = 'online'
          AND COALESCE(r.last_seen_at, r.updated_at) >=
              now() - make_interval(secs => @runtime_stale_secs::double precision)
      )
      AND NOT EXISTS (
        SELECT 1 FROM agent_task_queue occupant
        WHERE occupant.issue_id = t.issue_id
          AND occupant.agent_id = t.agent_id
          AND occupant.id <> t.id
          AND (
            occupant.status IN ('queued', 'dispatched')
            OR (occupant.status = 'deferred' AND occupant.context->>'channel_issue_media_pending' = 'true')
          )
      )
)
UPDATE agent_task_queue
SET status = 'queued'
WHERE id IN (SELECT id FROM due WHERE issue_id IS NULL OR rn = 1)
RETURNING *;

-- name: ListQueuedClaimCandidatesByRuntimes :many
-- Batch variant of ListQueuedClaimCandidatesByRuntime (MUL-4257): returns
-- queued claim candidates across every runtime_id in the input set in ONE round
-- trip, so a daemon can list candidates for all of its runtimes with a single
-- query instead of one per runtime. Ordering matches the singular query
-- (priority, then FIFO) so the batch claim loop keeps the same fairness. The
-- runtime_id filter is served by the partial index
-- idx_agent_task_queue_claim_candidates; the cross-runtime ORDER BY still needs
-- a sort step (each runtime's slice is index-ordered, but merging several
-- runtimes' rows into one priority/FIFO order is not). The per-machine
-- candidate set is small, so this is cheap in practice.
SELECT atq.* FROM agent_task_queue atq
WHERE atq.runtime_id = ANY(@runtime_ids::uuid[])
  AND atq.status = 'queued'
  AND EXISTS (
      -- Keep this authorization fence in sync with ClaimAgentTask.
      SELECT 1
      FROM agent a
      JOIN agent_runtime r ON r.id = atq.runtime_id
      WHERE a.id = atq.agent_id
        AND a.runtime_id = atq.runtime_id
        AND (
            r.visibility = 'public'
            OR (
                r.visibility = 'private'
                AND (
                    r.owner_id IS NULL
                    OR a.owner_id IS NULL
                    OR r.owner_id = a.owner_id
                )
            )
        )
  )
ORDER BY atq.priority DESC, atq.created_at ASC;

-- name: PromoteDueDeferredTasksForRuntimes :many
-- Batch variant of PromoteDueDeferredTasksForRuntime (MUL-4257): promotes all
-- due deferred tasks across the runtime set in one UPDATE. Carries the same two
-- fences as the singular query; see its comment for why.
WITH due AS (
    SELECT t.id,
           t.issue_id,
           row_number() OVER (
               PARTITION BY t.issue_id, t.agent_id
               ORDER BY t.priority DESC, t.created_at ASC, t.id
           ) AS rn
    FROM agent_task_queue t
    WHERE t.runtime_id = ANY(@runtime_ids::uuid[])
      AND t.status = 'deferred'
      AND t.fire_at <= now()
      AND EXISTS (
        SELECT 1 FROM agent_runtime r
        WHERE r.id = t.runtime_id
          AND r.status = 'online'
          AND COALESCE(r.last_seen_at, r.updated_at) >=
              now() - make_interval(secs => @runtime_stale_secs::double precision)
      )
      AND NOT EXISTS (
        SELECT 1 FROM agent_task_queue occupant
        WHERE occupant.issue_id = t.issue_id
          AND occupant.agent_id = t.agent_id
          AND occupant.id <> t.id
          AND (
            occupant.status IN ('queued', 'dispatched')
            OR (occupant.status = 'deferred' AND occupant.context->>'channel_issue_media_pending' = 'true')
          )
      )
)
UPDATE agent_task_queue
SET status = 'queued'
WHERE id IN (SELECT id FROM due WHERE issue_id IS NULL OR rn = 1)
RETURNING *;

-- name: CancelDeferredEscalationsForTask :many
UPDATE agent_task_queue
SET status = 'cancelled', completed_at = now(), prepare_lease_expires_at = NULL
WHERE escalation_for_task_id = $1
  AND status IN ('deferred', 'queued', 'dispatched', 'waiting_local_directory')
RETURNING *;

-- name: CancelDeferredEscalationsForIssueAgent :many
WITH cancelled AS (
    UPDATE agent_task_queue fallback
    SET status = 'cancelled', completed_at = now(), prepare_lease_expires_at = NULL
    FROM agent_task_queue primary_task
    WHERE fallback.escalation_for_task_id = primary_task.id
      AND fallback.status IN ('deferred', 'queued', 'dispatched', 'waiting_local_directory')
      AND primary_task.issue_id = @issue_id
      AND primary_task.agent_id = @agent_id
    RETURNING fallback.*
)
SELECT * FROM cancelled;

-- name: ListActiveTasksByIssue :many
-- Backs the issue-detail "agent live" banner. Includes 'queued' so the
-- banner shows up the moment a task is enqueued — not only after a runtime
-- claims it. The queued window can be long when the runtime is offline or
-- busy on a prior task, and a silent UI during that window looks like the
-- platform never received the trigger.
SELECT * FROM agent_task_queue
WHERE issue_id = $1 AND status IN ('queued', 'dispatched', 'running', 'waiting_local_directory')
ORDER BY created_at DESC;

-- name: ListActiveTasksByIssueFamily :many
-- Cross-issue coordination read for parallel sub-issue work (#7768). Given a
-- family root — the target issue's parent, or the target itself when it has
-- none — return every in-flight task on the root and on all of its children,
-- so a run can see who else is already working in the family before it starts
-- overlapping work. Advisory only: nothing here gates, queues, or serialises
-- anything.
--
-- Same active set as ListActiveTasksByIssue, including 'queued': a queued
-- sibling cannot answer you yet, but it is about to touch the same code, which
-- is exactly what the caller is trying to find out. The status column tells the
-- two apart.
--
-- Issue identity is joined in because the caller renders runs from several
-- issues in one list and cannot label a row from the task alone. agent_id is
-- here for the same reason: this read spans agents, and which one is on a
-- sibling is the answer, not a detail.
--
-- Columns are named rather than embedded. This is the coordination question,
-- not the execution log: result and context are JSONB blobs, and work_dir /
-- trigger_summary / the attribution ids are all execution-log fields that a
-- caller asking "who else is here?" never reads. Selecting them would make
-- Postgres detoast and ship roughly 5x the bytes per row for nothing.
--
-- Ordered running-first so the truncation the LIMIT may impose drops the least
-- interesting rows, and bounded because a parent with hundreds of children must
-- not turn one coordination read into an unbounded scan.
SELECT
    atq.id AS task_id,
    atq.agent_id,
    atq.issue_id,
    atq.status,
    atq.created_at,
    atq.started_at,
    w.issue_prefix,
    i.number AS issue_number,
    i.title AS issue_title
FROM agent_task_queue atq
JOIN issue i ON i.id = atq.issue_id
JOIN workspace w ON w.id = i.workspace_id
WHERE i.workspace_id = @workspace_id
  AND (i.id = @root_issue_id::uuid OR i.parent_issue_id = @root_issue_id::uuid)
  AND atq.status IN ('queued', 'dispatched', 'running', 'waiting_local_directory')
ORDER BY
    CASE atq.status
        WHEN 'running' THEN 0
        WHEN 'dispatched' THEN 1
        WHEN 'waiting_local_directory' THEN 2
        ELSE 3
    END,
    atq.created_at DESC
LIMIT @row_limit;

-- name: GetWorkspaceAgentRunCounts :many
-- Total task runs per agent over the trailing 30 days, used by the Agents
-- list RUNS column. 30-day window keeps the count meaningful (a long-dormant
-- agent shouldn't show "5,420 runs from 2 years ago") and keeps the scan
-- bounded as the workspace ages.
SELECT
    atq.agent_id,
    COUNT(*)::int AS run_count
FROM agent_task_queue atq
JOIN agent a ON a.id = atq.agent_id
WHERE a.workspace_id = $1
  AND atq.created_at > now() - INTERVAL '30 days'
GROUP BY atq.agent_id;

-- name: GetWorkspaceAgentActivity30d :many
-- Returns per-agent daily activity buckets for the last 30 days. Single
-- workspace-wide read backs both surfaces:
--   - Agents list ACTIVITY column — uses only the trailing 7 buckets
--   - Agent detail "Last 30 days" panel — uses the full 30
-- 30 days contains 7 days, so one fetch + a client-side .slice(-7) wins
-- over fetching twice. Days with no completion produce no row; the
-- front-end zero-fills.
--
-- Anchored on completed_at (not created_at) because the sparkline answers
-- "what did this agent produce?" not "what was queued at it?". A task that's
-- still in flight has no completed_at and contributes nothing here — that's
-- correct: in-flight tasks are surfaced via the live presence indicator,
-- not the historical trend.
SELECT
    atq.agent_id,
    DATE_TRUNC('day', atq.completed_at)::timestamptz AS bucket,
    COUNT(*)::int AS task_count,
    COUNT(*) FILTER (WHERE atq.status = 'failed')::int AS failed_count
FROM agent_task_queue atq
JOIN agent a ON a.id = atq.agent_id
WHERE a.workspace_id = $1
  AND atq.completed_at IS NOT NULL
  AND atq.completed_at > now() - INTERVAL '30 days'
GROUP BY atq.agent_id, bucket
ORDER BY atq.agent_id, bucket;

-- name: ListWorkspaceAgentTaskSnapshot :many
-- Returns the tasks the front-end reads off one workspace-wide snapshot:
--   - All active tasks (queued / dispatched / running / waiting_local_directory)
--     — the current workload: working signal + counts (see derive-presence.ts).
--   - Each agent's most recent OUTCOME task (completed / failed) — NOT part of
--     presence since #1823; it is the "last activity" line the Squad hover card
--     renders (agent-live-peek-card.tsx). Kept in this response because shipped
--     desktop builds read it from here; see MUL-5436 for the plan to move it to
--     a dedicated lazy endpoint.
--
-- Cancelled tasks are excluded from the outcome half on purpose: cancel is a
-- procedural signal ("attempt aborted"), not an outcome. It tells us nothing
-- about whether the agent works, so it must NOT be allowed to mask a prior
-- failure. Concretely: if an agent fails and then the user cancels the queued
-- retry (or the parent issue closes and cascades cancels), the failure stays
-- the agent's last real outcome.
--
-- The outcome half is a per-agent LATERAL Top-1 rather than a workspace-wide
-- DISTINCT ON: with idx_agent_task_queue_agent_terminal_latest (migration 233)
-- each agent costs one ordered index scan + LIMIT 1, so the cost no longer
-- grows with how much terminal history the workspace has accumulated. The
-- (created_at, id) tie-break makes the pick deterministic when completed_at
-- ties or is NULL — plain completed_at DESC returned an arbitrary row there.
-- Row shape and row set are unchanged.
--
-- Both halves JOIN / scan agent because agent_task_queue has no workspace_id.
SELECT atq.* FROM agent_task_queue atq
JOIN agent a ON a.id = atq.agent_id
WHERE a.workspace_id = $1
  AND atq.status IN ('queued', 'dispatched', 'running', 'waiting_local_directory')

UNION ALL

SELECT latest.* FROM agent a
JOIN LATERAL (
  SELECT atq.*
  FROM agent_task_queue atq
  WHERE atq.agent_id = a.id
    AND atq.status IN ('completed', 'failed')
  ORDER BY atq.completed_at DESC NULLS LAST, atq.created_at DESC, atq.id DESC
  LIMIT 1
) latest ON TRUE
WHERE a.workspace_id = $1;

-- name: ListWorkspaceWorkingAgents :many
-- Workspace-level source for consumers that show currently working agents.
-- One row per visible, user-authored agent with at least one task that has
-- actually started running. work_type is optional (empty = every source);
-- source-specific reads use the same precedence as computeTaskKind:
-- chat > autopilot > issue. "issue" intentionally groups direct and
-- comment-triggered issue work. Quick-create work is present only in the
-- unfiltered projection because it has no source FK yet. mine_relation is
-- optional (empty = workspace); when set it narrows issue work to the
-- authenticated member's My Issues relation. parent_issue_id is optional
-- (NULL = workspace); when set it narrows issue work to the direct children
-- of that issue, so an issue detail's sub-issue header reads the same
-- projection as the Issues list header instead of deriving its own count.
SELECT
  a.id,
  a.name,
  a.avatar_url,
  COUNT(*)::int AS running_task_count,
  COALESCE(
    ARRAY_AGG(DISTINCT atq.issue_id ORDER BY atq.issue_id)
      FILTER (WHERE atq.issue_id IS NOT NULL),
    ARRAY[]::uuid[]
  )::uuid[] AS issue_ids
FROM agent a
JOIN agent_task_queue atq ON atq.agent_id = a.id
WHERE a.workspace_id = $1
  AND a.kind = 'user'
  AND a.archived_at IS NULL
  AND atq.status = 'running'
  AND (
    @work_type::text = ''
    OR (@work_type::text = 'chat' AND atq.chat_session_id IS NOT NULL)
    OR (
      @work_type::text = 'autopilot'
      AND atq.chat_session_id IS NULL
      AND atq.autopilot_run_id IS NOT NULL
    )
    OR (
      @work_type::text = 'issue'
      AND atq.chat_session_id IS NULL
      AND atq.autopilot_run_id IS NULL
      AND atq.issue_id IS NOT NULL
    )
  )
  AND (
    @mine_relation::text = ''
    OR EXISTS (
      SELECT 1
      FROM issue i
      WHERE i.id = atq.issue_id
        AND i.workspace_id = a.workspace_id
        AND (
          (
            @mine_relation::text IN ('assigned', 'any')
            AND i.assignee_type = 'member'
            AND i.assignee_id = @member_id::uuid
          )
          OR (
            @mine_relation::text IN ('created', 'any')
            AND i.creator_type = 'member'
            AND i.creator_id = @member_id::uuid
          )
          OR (
            @mine_relation::text IN ('involved', 'any')
            AND (
              (
                i.assignee_type = 'agent'
                AND EXISTS (
                  SELECT 1
                  FROM agent owned_agent
                  WHERE owned_agent.id = i.assignee_id
                    AND owned_agent.workspace_id = a.workspace_id
                    AND owned_agent.owner_id = @member_id::uuid
                )
              )
              OR (
                i.assignee_type = 'squad'
                AND EXISTS (
                  SELECT 1
                  FROM squad s
                  WHERE s.id = i.assignee_id
                    AND s.workspace_id = a.workspace_id
                    AND (
                      EXISTS (
                        SELECT 1
                        FROM squad_member sm
                        WHERE sm.squad_id = s.id
                          AND sm.member_type = 'member'
                          AND sm.member_id = @member_id::uuid
                      )
                      OR EXISTS (
                        SELECT 1
                        FROM agent leader
                        WHERE leader.id = s.leader_id
                          AND leader.workspace_id = a.workspace_id
                          AND leader.owner_id = @member_id::uuid
                      )
                      OR EXISTS (
                        SELECT 1
                        FROM squad_member sm
                        JOIN agent owned_member ON owned_member.id = sm.member_id
                        WHERE sm.squad_id = s.id
                          AND sm.member_type = 'agent'
                          AND owned_member.workspace_id = a.workspace_id
                          AND owned_member.owner_id = @member_id::uuid
                      )
                    )
                )
              )
            )
          )
        )
    )
  )
  AND (
    @parent_issue_id::uuid IS NULL
    OR EXISTS (
      SELECT 1
      FROM issue child
      WHERE child.id = atq.issue_id
        AND child.workspace_id = a.workspace_id
        AND child.parent_issue_id = @parent_issue_id::uuid
    )
  )
GROUP BY a.id, a.name, a.avatar_url, a.created_at
ORDER BY a.created_at ASC;

-- name: ListTasksByIssue :many
SELECT * FROM agent_task_queue
WHERE issue_id = $1
ORDER BY created_at DESC;

-- name: UpdateAgentStatus :one
UPDATE agent SET status = $2, updated_at = now()
WHERE id = $1
RETURNING *;

-- name: RefreshAgentStatusFromTasks :one
-- Persisted agent.status has no queued/resource-wait bucket. Keep dispatched
-- as working because the daemon is actively preparing that task, but do not
-- let a task parked on a local_directory mutex masquerade as executing work.
-- The status guard makes a correct status a no-op: :one returns no rows, so
-- callers neither rewrite updated_at nor publish a redundant status event.
WITH desired AS (
    SELECT CASE WHEN EXISTS (
        SELECT 1 FROM agent_task_queue q
        WHERE q.agent_id = $1 AND q.status IN ('dispatched', 'running')
    ) THEN 'working' ELSE 'idle' END AS status
)
UPDATE agent AS a
SET status = desired.status,
    updated_at = now()
FROM desired
WHERE a.id = $1 AND a.status IS DISTINCT FROM desired.status
RETURNING a.*;

-- name: GetAgentBySystemKey :one
-- Resolves a workspace's built-in agent by its stable system_key. This is the
-- identity lookup for system agents: their display name is owner-editable, so
-- nothing server-side may key off it.
SELECT * FROM agent
WHERE workspace_id = $1 AND system_key = $2 AND archived_at IS NULL
ORDER BY created_at ASC, id ASC
LIMIT 1;

-- name: CreateSystemUserAgent :one
-- Creates a product-defined agent that members can still see, chat with, and
-- assign issues to. Deliberately kind='user': kind='system' hides the row from
-- agent lists and assignment surfaces and hard deletes it with its runtime.
-- Every product-owned field is a server constant; instructions stays empty
-- because the system half ships with the binary and is layered in at claim
-- time, leaving this column free for the workspace's own notes.
--
-- CONTRACT: call only while holding the per-workspace mika advisory lock. The
-- one-per-workspace invariant rests on a check inside that lock, not on a
-- unique index — migration 172's index keys on (workspace_id, owner_id,
-- runtime_id, system_key), so two different owners or runtimes are distinct
-- tuples and would both insert.
INSERT INTO agent (
    workspace_id, name, description, avatar_url, runtime_mode, runtime_config,
    runtime_id, model, visibility, permission_mode, max_concurrent_tasks,
    owner_id, instructions, custom_env, custom_args, kind, system_key
) VALUES (
    @workspace_id, @name, @description, @avatar_url, @runtime_mode, '{}'::jsonb,
    @runtime_id, @model, @visibility, @permission_mode, @max_concurrent_tasks,
    @owner_id, '', '{}'::jsonb, '[]'::jsonb, 'user', @system_key
)
RETURNING *;