venus
VENUS-209
Hub verbosity backstop — curb agent-to-agent chatter (Elazar 2a)
Done normal
plpm-llmmsgsrv-cc
Complements fleet CLAUDE.md agent-to-agent payload-only norm. Hub-side signal to surface/curb agent→agent reasoning-narration + social ack. Design-first: detection is the hard part (narration vs dense payload both long → length is weak). Reuse v_cli_drift_7d + daily coaching report + CDA-DRIFT marker. Human-facing msgs EXEMPT. Measurement view (agent-to-agent volume+length dist per sender) lands BEFORE active coaching; coach outliers only. coder-llmmsgsrv proposing detection approach + view spec. Ref mem:working-signal-plain-language.
Questions
No questions.
Activity
-
coder proposal (measured on copy DB, 7d a2a: 3628 msgs/3.22M ch/avg888): 3 of 4 detectors DEAD — length is information-free (avg len FLAT across all density buckets), kind is a proxy (69% dm), narration-lexicon 2/1532 hits. Only referent-density (digits+/_=`( per 1000ch) separates prose from payload — BUT it false-positives on reasoned escalation/disagreement (the top-2 lowest-density long msgs of the week were a blocked-deletion escalation + a bin-venus directive-contradiction flag). Active nag would suppress escalations. RECOMMEND measurement-only. Building v_a2a_verbosity_7d + weekly tail report (read-only, no hub code). Active-coaching decision → Elazar (1a measurement-only my rec / 1b nag-anyway). Independent bug found: expects_reply NEVER SET fleet-wide → filed separately as the nag-safe structural path.
-
View v_a2a_verbosity_7d SHIPPED at c194757 (v2.34.11): llmmsg-srv dfde7fc→c194757 (view appended to [migrate-analytics-views] DROP/CREATE block, read-only/lazy/zero-behavior-change) + sh.git abe1acb (llmmsg-kpi-report.sh 1.8→1.9, tail section LIMIT 15 with do-not-nag caveat, rides existing daily report, degrades to '(view missing)' pre-bounce). Cols: sender,msgs,chars,avg_len,max_len,long_msgs,prose_msgs,prose_chars,pct_of_hub_bytes. GATE = audit RE-CLEAR of c194757 (comment rewrite in review) → PM bounce. PM RULING: human exemption is ROW-level (shipped) not logical-send — agent-addressed fan-out copies count as a2a volume even when Elazar CC'd; logical-send under-counts. Not escalated (low-stakes, reversible). Junk noop-head-check commit in c194757 bumped 2.34.10→11, hub.mjs byte-identical to comment-fix (shared-tree gate covers).
-
Shipped + verified 2026-07-12 (hub v2.34.13, head c57b694). v_a2a_verbosity_7d live — read-only/lazy/no send-path touch; smoke-test returns per-sender baseline (top by bytes: pm-venus 449msg/328k/72prose, coder-llmmsgsrv 170/304k/34 [avg 1787], pm-llmmsgsrv 268/240k/41). sh.git llmmsg-kpi-report.sh v1.9 (abe1acb) prints tail w/ do-not-nag caveat on existing daily report. Human exemption ROW-level per PM ruling (both sender+recipient NOT LIKE elazar%). ACTIVE NAG NOT built — coder evidence: only working detector (referent-density) false-positives on reasoned escalation/disagreement; measurement-only recommended. Elazar owes 1a(measurement-only,rec)/1b(nag-anyway). Ref mem:working-signal-plain-language.
task
5w ago by wi-cli-venus
5w ago
2026-07-12 07:16