EPS Dashboard

Overview / Motivation

Auto-filed via the workflow-fix-on-bug protocol from a workflow gap found during a chat-mode session-cleanup diagnostic (2026-06-27). The idle-unmapped reaper in autonomous_session_watch.py is effectively INERT for the exact session class it was built to reap: detached, unmapped, genuinely-idle tmux-launched Claude sessions (the 2026-06-12 incident where 25 such sessions held ~23 GB RSS).

Goal

Make the idle-unmapped pass actually reap a detached, unmapped, genuinely-idle tmux-launched Claude session — closing the gap where every such session is kept forever via the tty=True / idle=? fail-toward-keep paths — WITHOUT ever reaping a detached session that is still doing real work.

Workflow gap

Bug observed: In the live 2026-06-27 17:34 and 17:43 watcher passes, every detached, unmapped tmux Claude session logged mapped=False tty=True idle=? missed=0->0 action=clear. The idle-unmapped pass reaped zero of them — it is inert for the tmux-launched unmapped class.
Why it is a workflow gap: The pass (CLAUDE.md § Crons; this file §9036-9138) exists specifically to reap "live-but-idle inner Claude + unmapped" sessions that the reconcile/zombie passes structurally exclude. The detached-tmux refinement (_detached_tmux_pane_ttys + _is_live_user_tty, ~L8987-9056) was added to handle exactly the "detached tmux pane holds a controlling pty" case — but in production it is not firing for these sessions, so the reaper never advances past decide_idle_unmapped's has_tty -> ("clear", 0) / idle_age_s is None -> ("skip", missed) guards.
Live root-cause evidence (gathered this session):
- For a still-live detached session (pid 2730037, session cmqvqzomyw6ngyc0utvuiklhd, tmux claude-0626-193535), the wrapper's controlling tty is /dev/pts/9 (fd 0/1/2).
- The watcher's own detached-pane query — tmux list-panes -a -F '#{pane_tty}\t#{session_attached}' — DOES return /dev/pts/9 ... 0 (attached=0), i.e. /dev/pts/9 IS in the detached set.
- Therefore _is_live_user_tty(2730037, detached_set) SHOULD return False (detached pane → not a live-user tty) → has_tty=False → the decision should fall through to the idle check. Yet the watcher logged tty=True. So in the cron execution, _is_live_user_tty returns True when a same-host live check says False.
- Leading hypotheses for the implementer to confirm: (a) the cron-time tmux list-panes -a subprocess returns empty / errors under the minimal cron env (fail-soft → empty detached set → every tty-bearing wrapper kept — the refinement never active in production); (b) the sessions were transiently attached at pass time; (c) a pid-resolution mismatch (the pid passed to _is_live_user_tty is not the pty-holding wrapper). Reproduce by running the watcher's exact tmux subprocess under the cron environment and by logging the resolved detached set + per-session controlling-tty path.
- Secondary gap: even if has_tty were correctly False, these sessions log idle=? (transcript unresolvable via _transcript_idle_age_s / _resolve_transcript_via_happy_log for manually-tmux-launched, non-daemon-spawned sessions) → the idle_age_s is None -> ("skip", missed) fail-toward-keep path keeps them indefinitely. A genuinely-idle detached session with an unresolvable happy-log transcript is never reapable.
Confidence (emitter): medium — the symptom (pass reaps nothing for this class) is certain and reproducible from the logs; the exact internal cause (a vs b vs c, plus the idle=? secondary gap) needs the planner/implementer to confirm before choosing the fix.

CRITICAL constraint — do NOT reap working detached sessions

The fail-toward-keep conservatism is load-bearing and must be preserved. At diagnosis time, of six detached unmapped tmux sessions:

ONE was actively running an experimenter agent while detached (claude-0626-192206, ↓ 177.6k, "Waiting for 1 background agent to finish").
FOUR had a typed-but-unsent user command queued in the input box (promote it useful, promote 685 useful, ping me when #691 finishes, check progress).
Only ONE was genuinely done (empty input, no running agent) — it was closed by hand this session.

So any reap predicate MUST additionally exclude a detached session that (i) has a running descendant agent (experimenter / codex companion / bg workload python) or (ii) owns a running pod-<N>. The fix tightens ONLY the genuinely-dead case; a false positive here kills a live experiment. This is precisely why it routes through the full /issue planner + adversarial-critic + code-review pipeline rather than a quick edit.

Proposed change (candidate diff sketch — refine in planning)

Diagnose why _is_live_user_tty returns True in the cron pass for a confirmed-detached pane (instrument the detached set + per-session controlling-tty path; verify the cron-env tmux query is non-empty).
If the cron-env tmux query is the cause, make the detached-pane resolution robust to the cron environment (explicit socket / $HOME) and/or log when the detached set comes back empty so a silently-inert refinement is visible.
For the idle=? secondary gap: when a session is confirmed detached + unmapped + has NO running descendant agent + owns NO running pod, allow a corroborating idleness fallback (e.g. tmux session_activity filtered to real turns, or the wrapper process activity) instead of an unconditional keep — but only as an ADDITIONAL gate layered on the no-running-work checks above, never replacing fail-toward-keep on its own.
Add/extend tests/test_autonomous_session_watch.py coverage: detached-pane → reap-eligible; detached-pane-with-running-experimenter → keep; attached-pane → keep; idle=? + running work → keep.

Scope / surfaces

Primary target: scripts/autonomous_session_watch.py (decide_idle_unmapped, _is_live_user_tty, _detached_tmux_pane_ttys, _transcript_idle_age_s, and the idle-unmapped pass call site ~L9338-9382).
Tests: tests/test_autonomous_session_watch.py.
Grep the workflow surface for sibling references before editing.

Constraints / invariants

Workflow-surface only — never experiment code, configs/, or tasks/.
scripts/workflow_lint.py --check-asks passes; ruff on touched files passes; CLAUDE.md § Crons / .claude/rules/background-automation.md stay consistent with the new behavior.
Preserve fail-soft: any tmux/transcript query failure must keep the conservative keep-all behavior; the fix may only ever reap MORE in a CONFIRMED-dead case, never reap an attached / uncertain / working one.
This session runs under EPM_WORKFLOW_FIX_SESSION=1 and carries a workflow_fix_target: Provenance line — it MUST NOT auto-route any of its own subagents' workflow-fix candidates (recursion guard).

Provenance

workflow_fix_target: scripts/autonomous_session_watch.py
fingerprint: 296170f25def

target_file: scripts/autonomous_session_watch.py bug_observed: Every detached, unmapped tmux Claude session logs tty=True idle=? action=clear in the live watcher passes, so the idle-unmapped pass reaps nothing for the tmux-launched class it was built for. why_workflow_gap: The detached-tmux refinement (_detached_tmux_pane_ttys/_is_live_user_tty) and the idle=? fail-toward-keep guard together keep every tmux-launched unmapped session forever, defeating the pass that exists to reap exactly this class (the 2026-06-12 23 GB incident). proposed_change: Make the idle-unmapped reaper actually reap detached, unmapped, genuinely-idle tmux-launched Claude sessions, gated to exclude any session running a descendant agent or owning a running pod. diff_sketch: |

in the idle-unmapped pass: confirm _is_live_user_tty correctly returns False for a detached pane under the cron env (instrument detached set + controlling-tty path)
add a corroborating idleness fallback for idle=? sessions ONLY when detached + no running descendant agent + no running pod

tests: detached -> reap; detached+running-experimenter -> keep; attached -> keep confidence: medium related_task: chat-mode session-cleanup diagnostic 2026-06-27