migrations: 0030 research_topics + 0031 loops schema

First commit of the Research + Loops feature arc. Schema only — routes,
runtime hooks, and the UI arrive in subsequent commits.

0030 — research_topics
  Container entity with a small state machine (standby → processing →
  reviewing → publishing → published). Owns runs via a nullable
  topology_runs.research_topic_id FK, so all existing SSE/audit/gated-
  approval plumbing surfaces without changes. Many-to-many join table
  captures the role_slot the wizard assigns each agent ("lead", "critic",
  "writer") so the canvas can group avatars sensibly.

0031 — loops
  Durable recurring topology execution. graph + task_template pair with a
  triggers JSONB (any of cron / on_completion / webhook, all can be on
  simultaneously) and a repeat_policy (infinite / N iters / until). Every
  iteration writes a topology_runs row with loop_id, iteration (1-indexed),
  and parent_run_id chained back to N-1 — cross-iteration context comes
  from that hop, no extra state store needed. Webhook auth is HMAC-SHA256
  keyed by webhook_signing_key. Missed cron windows fire once and skip the
  backlog (see comment header).

Both migrations only ADD tables/columns and use ON DELETE SET NULL for the
back-refs, so they're safe to run against prod without downtime. The
existing indexes on topology_runs keep serving legacy (non-research,
non-loop) runs unchanged.

Publish approval extension (approvals.kind for the reviewing → publishing
gate) comes as a separate migration in the publish-gate commit.

migrations/0030_research.sql

@@ -0,0 +1,62 @@
+-- Research topics: user-scoped inquiry containers that group a set of agents
+-- around a shared question and drive them toward a named outcome (spec,
+-- prod_plan, roadmap, or paper). The status field is a small state machine
+-- driven by the orchestrator + a manual publish gate (approvals; see 0032).
+--
+--   standby     — created, agents assigned, not started
+--   processing  — running; each iteration writes to topology_runs
+--                 (with research_topic_id back-ref)
+--   reviewing   — orchestrator flipped it here on last run_completed;
+--                 a human approver still has to click Publish
+--   publishing  — approval landed; artifact assembly + release in flight
+--   published   — terminal
+--
+-- A topic OWNS runs (topology_runs.research_topic_id). Runs carry all the
+-- durable execution state — the topic row is just the container + status.
+
+CREATE TABLE research_topics (
+    id            UUID PRIMARY KEY,
+    workspace_id  UUID NOT NULL REFERENCES workspaces (id) ON DELETE CASCADE,
+    title         TEXT NOT NULL,
+    -- LLM-refined description from the wizard: structured markdown with the
+    -- topic prompt + refined framing + key questions + success criteria.
+    description   TEXT NOT NULL,
+    outcome_kind  TEXT NOT NULL
+        CHECK (outcome_kind IN ('spec', 'prod_plan', 'roadmap', 'paper')),
+    status        TEXT NOT NULL DEFAULT 'standby'
+        CHECK (status IN ('standby', 'processing', 'reviewing', 'publishing', 'published')),
+    created_by    UUID NOT NULL REFERENCES users (id) ON DELETE RESTRICT,
+    created_at    TIMESTAMPTZ NOT NULL DEFAULT now(),
+    updated_at    TIMESTAMPTZ NOT NULL DEFAULT now(),
+    -- Set on the reviewing→publishing transition (after the publish approval
+    -- lands). Used by list queries that surface recent publishes.
+    published_at  TIMESTAMPTZ
+);
+
+CREATE INDEX research_topics_workspace_idx
+    ON research_topics (workspace_id, updated_at DESC);
+-- Speeds up the "topics in each status" sidebar bucket count.
+CREATE INDEX research_topics_status_idx
+    ON research_topics (workspace_id, status);
+
+-- Assigned agents (many-to-many). role_slot is a free-form label the wizard
+-- captures ("lead", "critic", "writer") so the canvas can group avatars
+-- when the topic has multiple agents playing different parts.
+CREATE TABLE research_topic_agents (
+    topic_id   UUID NOT NULL REFERENCES research_topics (id) ON DELETE CASCADE,
+    agent_id   UUID NOT NULL REFERENCES agents (id) ON DELETE CASCADE,
+    role_slot  TEXT,
+    created_at TIMESTAMPTZ NOT NULL DEFAULT now(),
+    PRIMARY KEY (topic_id, agent_id)
+);
+CREATE INDEX research_topic_agents_agent_idx
+    ON research_topic_agents (agent_id);
+
+-- Back-ref from durable runs to the topic they belong to. Nullable — runs
+-- outside a research topic (chat, ad-hoc topology exec) leave this NULL and
+-- the existing indexes keep serving them.
+ALTER TABLE topology_runs
+    ADD COLUMN research_topic_id UUID REFERENCES research_topics (id) ON DELETE SET NULL;
+CREATE INDEX topology_runs_research_topic_idx
+    ON topology_runs (research_topic_id, created_at DESC)
+    WHERE research_topic_id IS NOT NULL;

migrations/0031_loops.sql

@@ -0,0 +1,71 @@
+-- Loops: durable recurring topology executions. A loop pairs a stored
+-- topology graph + task template with a trigger configuration (any of
+-- cron / on_completion / webhook) and a repeat policy (infinite / N iters /
+-- until condition). Every iteration writes a normal topology_runs row with
+-- loop_id + iteration + parent_run_id set so the existing SSE observer,
+-- audit log, and gated approvals surface transparently.
+--
+-- Cross-iteration state: iteration N+1's run driver reads parent_run_id and
+-- injects the parent's final assistant message into the first system prompt
+-- as prior-iteration context (refine-over-time by default).
+--
+-- Missed schedules: if the loop scheduler wakes and finds next_fire_at in
+-- the past (a restart or long stall), it fires ONCE and computes the next
+-- fire from the current cron expression. Missed backlog is not drained.
+
+CREATE TABLE loops (
+    id             UUID PRIMARY KEY,
+    workspace_id   UUID NOT NULL REFERENCES workspaces (id) ON DELETE CASCADE,
+    title          TEXT NOT NULL,
+    description    TEXT NOT NULL,
+    -- Same shape as topology_runs.graph: JSON-encoded agent/node connectivity
+    -- from the wizard's topology builder (or a template pick).
+    graph          JSONB NOT NULL,
+    -- Text prompt used as the task for every iteration. Templating is out of
+    -- scope for v1 — the same string runs each time (with prior-iteration
+    -- output prepended by the run driver for refine-over-time).
+    task_template  TEXT NOT NULL,
+    -- {cron?: '0 */6 * * *', on_completion?: bool, webhook_enabled?: bool}
+    -- All three can be enabled simultaneously; whichever fires first wins.
+    triggers       JSONB NOT NULL DEFAULT '{}'::JSONB,
+    -- {kind: 'infinite' | 'iters' | 'until', spec: <kind-specific>}
+    --   iters: {n: 10}                — stop after N iterations
+    --   until: {event: 'ok', within_iters: 20} — stop when N events match
+    repeat_policy  JSONB NOT NULL DEFAULT '{"kind":"infinite"}'::JSONB,
+    enabled        BOOLEAN NOT NULL DEFAULT true,
+    -- Computed by the scheduler from `triggers.cron` after each fire. NULL
+    -- when no cron trigger is set (loop runs only on webhook / on_completion
+    -- / manual).
+    next_fire_at   TIMESTAMPTZ,
+    last_run_id    UUID REFERENCES topology_runs (id) ON DELETE SET NULL,
+    -- Opaque secret ID for the webhook endpoint. HMAC-SHA256 verification of
+    -- the request body uses webhook_signing_key. Both are NULL when the
+    -- webhook trigger is disabled; both are set when it's on. Rotation:
+    -- disable trigger, re-enable to regenerate both.
+    webhook_token        TEXT UNIQUE,
+    webhook_signing_key  TEXT,
+    created_by     UUID NOT NULL REFERENCES users (id) ON DELETE RESTRICT,
+    created_at     TIMESTAMPTZ NOT NULL DEFAULT now(),
+    updated_at     TIMESTAMPTZ NOT NULL DEFAULT now()
+);
+
+CREATE INDEX loops_workspace_idx
+    ON loops (workspace_id, updated_at DESC);
+-- Scheduler tick query: SELECT ... WHERE enabled AND next_fire_at <= now().
+-- Partial index keeps the scanner tight; disabled + no-cron loops stay out.
+CREATE INDEX loops_next_fire_idx
+    ON loops (next_fire_at)
+    WHERE enabled AND next_fire_at IS NOT NULL;
+
+-- Back-refs from durable runs. loop_id groups all iterations of a loop;
+-- iteration is 1-indexed per loop; parent_run_id chains iteration N+1 back
+-- to N so the run driver can pull prior context in one hop.
+ALTER TABLE topology_runs
+    ADD COLUMN loop_id       UUID REFERENCES loops (id) ON DELETE SET NULL,
+    ADD COLUMN iteration     INT,
+    ADD COLUMN parent_run_id UUID REFERENCES topology_runs (id) ON DELETE SET NULL;
+
+-- Listing "recent iterations of loop X" hits this.
+CREATE INDEX topology_runs_loop_idx
+    ON topology_runs (loop_id, iteration DESC)
+    WHERE loop_id IS NOT NULL;