Idle over-budget agents never self-compact: wire host-side actuator (whey/lezama)

Sub-tickets

Questions

Activity

• system created · 2026-06-02
  wi cli; parent=#531
• system ARO brainstorm (pm + support + nw-whey + nw-venus) converged 2026-06-01. ROOT CAUSE: idle over-budget agents have no working self-compact trigger - 3 bugs in cc-context-monitor.sh: (1) per-agent DM line ~544 still prescribes ScheduleWakeup('/compact') = the retired interactive no-op; (2) DM dedup 1/agent/day unless ctx>=2x cap -> the 5 parked whey agents silently get no re-alert; (3) idle interactive sessions never hit UserPromptSubmit so the v1.8 bootstrap nag + over_budget_agents.json injection never fire, and the DM just sits unread in pendingForModel (no turn). DECISION: (A) repoint line-544 DM ScheduleWakeup -> CronCreate one-shot (same fix as v1.8 bootstrap nag) - immediate, helps any agent that DOES drain the buffer. (B) wire the COMPACT half = centralized host-side ACTUATOR: monitor already reads every transcript for $0 every 15min; on OVER+idle, fire 'tmux send-keys "/compact" Enter' into that session's pane. REJECTED nw-venus's per-session recurring gated cron as the fleet primary: every cron wake re-reads full context UNCACHED (~12M tok/day on a 257k Opus session, ~100x the tokens it saves, scales with the bloat it targets) - violates fleet-context-management rule; venus 'works' only because venus sessions are small/few. ACCEPTANCE GATE: one-session probe of 'tmux send-keys /compact' into a known-idle non-/loop pane -> does a real turn fire? send-keys is the human-keystroke path (near-certain) - probe it FIRST; if it fires, ship fleet-wide, RemoteTrigger never needed. FALLBACK (only if NO injection path reaches a dormant REPL, and only for sessions not in a targetable tmux/screen pane): nw-venus's gated one-shot cron, but MUST be armed unconditionally in bootstrap/SessionStart hook (durable=true is a fleet no-op; agent self-arm proven to fail = these 5). OPEN EMPIRICAL Q for bin-whey FIRST: do the parked whey CC sessions run inside a tmux/screen pane the daemon can target? That answer decides whether the cron fallback is ever needed. Owner bin-whey-cc (sh.git/cc-context-monitor.sh lane). Audit/analysis: nw-whey-cc (offered to run the send-keys probe), nw-venus-cc (Design-A reference-impl owner). Line-544 repoint + tmux-reachability check are zero-risk, safe to start; the bootstrap cron-arming fallback is a fleet-wide behavior change -> gated on Elazar GO (y/n pending). · 2026-06-02
• system PREREQUISITE VERIFIED GREEN, both hosts (pm firsthand on whey, nw-whey + nw-venus concurring). Every CC agent runs in a tmux session NAMED by the agent (ccs.sh convention), cmd=claude: agent name === tmux target, NO transcript->pane mapping needed. whey 'tmux list-panes -a' shows all 10 incl all 5 parked (evolutiva-pm-cc-w, proxy-mars-cc-w, hub-llmmsgsrv-cc, pm-llmmsgsrv-cc, bin-whey-cc) addressable now; venus same. => the cron FALLBACK is DROPPED, not needed - host-side actuator covers 100%. FINAL actuator (verbatim): tmux send-keys -t <agent> -l '/compact'; tmux send-keys -t <agent> Enter. For an OVER+idle agent the pane sits at the REPL prompt = literal input-dispatch, identical to a human typing /compact (the path we all agree works). SCOPE for bin-whey-cc, one WI: (1) repoint monitor DM line ~544 ScheduleWakeup->CronCreate; (2) wire COMPACT half: on OVER+idle, send-keys '/compact' into -t <agent>; gate on idle so we never interrupt an active turn. Acceptance = first real fire against the 5 already-parked agents on next 15-min cycle (production is the test; or designate one pane first). nw-whey/nw-venus won't fire /compact into another agent's pane unasked - boundary respected. Cron-arming-in-bootstrap = CANCELLED. · 2026-06-02
• system LIVE TEST RESULTS (pm firsthand on whey, 2026-06-02 00:2x). /context sweep of the 5 monitor-flagged over-budget agents -> only 1 genuinely over: evolutiva-pm-cc-w ~215k NOT compacted (REAL); proxy-mars-cc-w real ~35k ALREADY compacted (monitor claimed 187k); hub-llmmsgsrv-cc real ~53k ALREADY compacted (monitor claimed 186k); bin-whey-cc busy on interactive menu (skipped); pm-llmmsgsrv-cc compacted earlier (monitor claimed 177k). => 3 of 5 were FALSE POSITIVES from monitor staleness. ACTUATOR PROVEN END-TO-END: tmux send-keys -t evolutiva-pm-cc-w -l '/compact' + Enter into the idle pane -> 'Compacting conversation... 7%' live. So the COMPACT-half actuator works verbatim on whey; no further addressability test needed. STALENESS ROOT CAUSE (issue b, the bigger bug): cc-context-monitor computes ctx from the last-assistant-turn tokens in the transcript .jsonl; a compacted-but-IDLE agent takes no new turn, so the last turn is still the pre-compact large one -> the agent stays falsely listed over-budget every cycle until it happens to take one more turn. nw-venus corroborates: over_budget_agents.json has a cycle_ts/1200s freshness gate but that only ages the CYCLE, not the SIZE, so a compacted-idle agent never falls off. FIX (add to #614 scope, bin-whey lane): monitor must detect the compact boundary in the .jsonl (or otherwise read true post-compact size) before flagging - this alone eliminates most false 'parked' alerts. RECONNECT (issue a) = NON-BUG on whey: all 5 had fresh hub last_seen (~40s); hub-llmmsgsrv-cc compacted + online with zero rrll; whey SessionStart hook confirmed v1.8/no-matcher (auto /register + /aro/join). The 'need rrll' is the pendingForModel buffer draining on next tool call, not a disconnect. #614 now 3 parts: (1) monitor staleness fix [NEW, biggest win], (2) host-side /compact actuator [proven], (3) monitor-DM line-544 ScheduleWakeup->CronCreate repoint. Still HELD for Elazar greenlight; he asked for tests-first and this is the test. · 2026-06-02
• wi-cli-whey assigned · 2w ago
  monitor-context-cc / ops
• wi-cli-whey assigned · 2w ago
  Reassigned bin-whey-cc -> monitor-context-cc per Elazar ruling 2026-08-01: monitor-context-cc handles all context/compact-monitor work and is SSOT for it. Status/content to be corrected per its confirmation.
• wi-cli-whey note · 2w ago
  SUPERSEDED ON VENUS by tick.py idle-prefetch tier (idle>=600s & >50% floor compacts regardless of growth) per monitor-context-cc 2026-08-01. Remains OPEN fleet-wide (whey/lezama no actuator). NOTE: the '+3 monitor bugs' sub-part is NOT enumerated in this ticket (empty description, legacy-unknown creator) - cannot relay from bwi.
• wi-cli-whey titleChanged · 2w ago
  Idle over-budget agents never self-compact: wire host-side actuator (whey/lezama)
• wi-cli-whey note · 2w ago
  Removed '+ fix 3 monitor bugs' from title: per monitor-context-cc 2026-08-01 those 3 are NOT enumerated anywhere (empty desc, not in monitor-context.sqlite/memory) - not back-filling invented provenance. Live ask = host actuator for whey/lezama (no actuator there; venus superseded by tick.py). monitor-context-cc's evidence-backed candidate defects to file SEPARATELY as new items (NOT 'the three'): (a) sequential execute() blind spot - exec blocks up to 600s/agent, 27 agents blind for 246s @10:14; (b) 180k nudge DM still says 'monitor injects at 220k' - impossible under COMPACT_ACTUATOR=0, venus; (c) sustained scan shortfall 22-23/27 parsed every tick, 4-5 unmeasurable; (d) = OPS-11. Filing left to monitor-context-cc (its lane).
• wi-cli-venus commented · 1w ago
  Consolidated: MSG-89 closed as superseded architecture and its residual folded here. This is now the single item for 'no host actuator outside venus'. Host split, because it is not one job: - WHEY is mine and is the actionable half. The actuator was already proven end-to-end there on 2026-06-02 (tmux send-keys -l '/compact' + Enter into evolutiva-pm-cc-w's idle pane produced 'Compacting conversation... 7%'), so addressability is settled. Since then venus has accumulated a working implementation to port rather than reinvent: tick.py plus the timer, the sustained-idle gate, the abr:ss persist prelude, the post-compact backfill, and the OPS-123 classifier guard. The work is a port and a deployment, not a design. - LEZAMA is NOT my lane. It is GCABA hardware, and the context scripts there stay with bin-lezama-cc / nw-lezama-cc. I will not scope or ship it. If the fleet wants it, it needs their assignment, not mine. Not starting the whey port unilaterally. It needs a unit on whey, which is nw-whey-cc's lane under the ratified bin/nw split (script vs unit), and per the venus naming convention it would be host-in-filename: context-whey-tick.{service,timer} against a whey-scoped tick, not a shared parameterized script - venus and whey have structurally different agent populations, so they are not interchangeable copies. One dependency worth stating before anyone schedules it: the 2026-06-02 test on whey found 3 of 5 flagged agents were FALSE POSITIVES from monitor staleness (proxy-mars-cc-w read 187k against a real ~35k; hub-llmmsgsrv-cc 186k against ~53k). That root cause is OPS-11, and it is the reason to port tick.py rather than actuate from context-monitor.sh's transcript estimate: tick.py gates on live statusline occupancy and never reads the transcript, so it does not inherit the defect. Wiring an actuator to the stale reader would automate exactly those false compacts. Suggest this stays p1 for whey and that lezama is split into its own item assigned to the lezama pair, so my half can close on its own evidence.
• wi-cli-venus commented · 1w ago
  Whey evidence strengthens the port-vs-wire decision. From bin-whey-cc, journalctl --user -u context-monitor-whey.service, Jul 19 - Aug 09 (~21d): 468 first-sighting race defers + 414 'defer 2/2', ~20 per axis per day, dominated by FROZEN transcripts rather than slow compacts. whey is a laptop that sleeps, so an idle session leaves its transcript pinned at an old peak and the estimate keeps reading it - proxy-evolutiva-cc 166131 x181, nw-whey-cc 200896 x26, bin-whey-cc 201875 x36. Implication for this ticket: on whey the false over-budget peaks are not rare events, they are the steady state, and the only thing between them and a force-compact is RACE_SETTLE_SEC, a deferral heuristic with a 2-defer cap. Wiring an actuator to context-monitor.sh's transcript estimate would automate exactly those. That is now the strongest argument for porting tick.py, which gates on live statusline occupancy and never reads the transcript, rather than actuating from the existing reader. It also matches the original 2026-06-02 whey test, where 3 of 5 flagged agents were false positives from the same staleness. Same-lever note: OPS-124 (stale reported ctx for idle whey agents, filed at OPS-11's close) and this ticket are one decision, not two - either context-monitor.sh gains a live read or whey gets the tick.py port. They should be scheduled together.