EPS Dashboard

Goal

A filed proposed kind:infra task currently only auto-runs if a LIVE PM session performs a STATUS pass (research-pm.md §444 "Standing rule — infra auto-dispatch"). With no PM session in the loop, the task orphans indefinitely — incident: code-review fast-follow task #684 was filed by the #682 session and sat at proposed for ~17h until a human hand-spawned it. The always-on 10-min autonomous_session_watch.py cron only re-drives blocked + no_compute_available tasks (#642); it deliberately skips proposed. Close the gap from BOTH ends.

Required changes

1. File-time auto-dispatch (primary mechanism)

Whoever FILES a ripe, auto-dispatchable kind:infra task should IMMEDIATELY dispatch it (spawn_session.py spawn-issue --issue <N> --auto) instead of relying on PM pickup. Add a mechanism that makes this the default and hard to forget — e.g. a task.py new-adjacent helper or a thin scripts/ wrapper that files + dispatches in one call — plus the instruction wired into the routing surfaces (CLAUDE.md routing, workflow-fix-on-bug.md, research-pm.md) that "filing a ripe infra fix == dispatching it." It MUST no-op gracefully when no Happy daemon is reachable (pod-side / headless filing) — those cases fall through to #2. Design the exact mechanism in the plan (do not bake session-spawning into task.py's state-mutation core if it breaks the pod-side "never shell out to task.py" rule — prefer the orchestration layer).

2. Always-on watcher safety-net sweep (backstop)

Add a pass to autonomous_session_watch.py (the 10-min cron) that sweeps ripe proposed kind:infra tasks and dispatches them via spawn-issue --auto, catching anything filed by a context that could not self-dispatch (pods, manual task.py new, a crashed filer). It MUST reuse the SAME scope + gating the PM rule uses: the 3-concurrent-auto-dispatched- session cap, holds/predicate gating, on_hold exclusion, and skip any task that already has a live issue-mapped session (no double-dispatch).

Constraints / acceptance

Idempotent: never spawn a 2nd session for a task that already has a live one (dedup on issue-mapped sessions). This is the most important invariant.
Shared concurrency cap: respect the existing 3-session cap shared with PM auto-dispatch; don't blow past it when both paths are active.
Don't dispatch proposed tasks that aren't kind:infra, that are on_hold, or that have unmet holds/predicates.
Honor autonomous-mode conventions (EPM_AUTONOMOUS_SESSION).
The existing PM-session infra auto-dispatch rule keeps working unchanged — this ADDS paths, it does not remove the PM one.
Tests pin: (a) file-time dispatch fires on a ripe infra filing; (b) the watcher sweep dispatches an orphaned proposed infra task; (c) neither path double-dispatches a task that already has a live session; (d) the concurrency cap is respected.

Provenance

Surfaced from the #684 incident (2026-06-27): a code-review fast-follow infra task filed by the #682 session sat at proposed ~17h because no PM STATUS pass ran. User directive: "fix this ... and add something so that these infra fixes just get dispatched automatically by whoever files them."