Pull-based session retros via a reviewed graph edge

The first design for session retros was push-based: a Claude Code SubagentStop hook would fire ax retro emit after every turn, inserting a retro row keyed by session. That has two problems.

First, Stop hooks run on the agent's hot path. Even a fast hook introduces visible turn latency, and any failure inside it blocks the agent from returning to the user. Second, the hook can only emit a deterministic heuristic retro (counts of turns, top failing tool, etc.) because it cannot pause to think; the retros are cheap but shallow.

We adopt a pull-based variant instead. ax retro emit still exists for the deterministic path, but the primary surface is now ax retro pending + ax retro brief + a retro-reviewer subagent dispatched by the /retro skill. A session is "pending retro" iff it has no outbound reviewed graph edge AND is finished (either ended_at is set, or its last turn is older than an idle threshold). The skill drains the backlog on demand, in parallel, with the user's idle Opus quota - a "quota arbitrage" pattern: weekly windows that would otherwise expire unused get spent on harness improvement instead.

reviewed is a typed RELATION (session -> retro) with a UNIQUE(in, out) index, not a foreign-key column on retro. We already have retro.session as a record reference, so the edge is technically duplicate state. We accept the duplication for three reasons. (1) The existing convention - spawned, produced, invoked, corrected_by

• is that every session-relationship is a RELATION; reviewed matches the rest of the graph. (2) Graph traversals stay symmetric (session->reviewed->retro->proposal reads cleanly in both directions). (3) The hot query for ax retro pending is WHERE count(->reviewed) = 0, which is cheaper and clearer than a NOT IN (SELECT session FROM retro) subquery and parallels how we ask "no spawned children" or "no produced commits" elsewhere.

The pending query is two-pass and deliberately avoids the per-session count(turn) subquery that blew up the v0 prototype (48s for a LIMIT 5). Pass one selects sessions with ended_at in the window; pass two selects sessions without ended_at whose started_at is older than the idle threshold (proxy for "user closed the tab"). Subagent sessions (source = 'claude-subagent') are excluded by default because their retros belong to the parent session's review. Turn count and last-turn timestamp are fetched lazily inside ax retro brief, not here.

Trade-off: the subagent-driven path is per-session more expensive than a single Stop hook firing, so we'd lose on cost if we ran it on every session unconditionally. That's the point of pulling. The user (or a weekly cron) chooses when to spend the quota; sessions accumulate in the pending queue until then. The deterministic ax retro emit heuristic remains as a fallback for Stop-hook users who want the cheap signal regardless.