vivekchand/clawmetry

Automated harness-observability audit found a gap for openclaw. Harness exposes: OpenClaw 2026.9.3's safer-update pipeline rehearses core/plugin changes in isolated candidate state before activation, tracks eligible 2026.9.2 migrations, and can recover abandoned update records without stopping a healthy Gateway (#136997, #138839, #141109, #141175, #141562). ClawMetry misses: No function in the adapter reads any update/migration record (candidate state, activation status, abandoned-update recovery outcome). _openclaw_doctor_findings() surfaces general doctor diagnostics but nothing update-pipeline specific, so an agent stuck in a half-activated or abandoned update state reports no signal to ClawMetry's health surfaces. Where (harness): CHANGELOG.md 2026.9.3 highlights Severity: medium Closest capability: new Filed by scripts/harness/audit.py. Fingerprint: hgap-17edab7b0b (used to dedupe — keep it in the body).

Automated harness-observability audit found a gap for openclaw. Harness exposes: Models settings now let a user manage multiple connected provider accounts and set supported account priority (i.e. which credential/account is preferred per provider) directly in the harness. ClawMetry misses: The adapter's provider/model surface only covers ClawRouter (_clawrouter_detect, reading ~/.openclaw/clawrouter/config.json) and NemoClaw sandbox inference config (_sandbox_inference_configs). Neither reads the new connected-accounts/priority list, so a misconfigured or deprioritized account (a real cause of auth failures/cost surprises) is invisible. Where (harness): CHANGELOG.md 2026.9.3 highlights (#132451) Severity: medium Closest capability: new Filed by scripts/harness/audit.py. Fingerprint: hgap-b5710e3e2e (used to dedupe — keep it in the body).

What the roadmap promises PicoClaw, NanoClaw, plus a lightweight SDK for CrewAI, LangGraph, AutoGen, and any OpenTelemetry-instrumented agent. Listed under the Planned section as of 2026-07-01. Cross-repo scope This work belongs in vivekchand/clawmetry (OSS). The OTLP receiver already accepts any OpenTelemetry-instrumented agent; the gap is first-class adapters and _CHANNEL_DIRS entries for the named runtimes. What this means in concrete terms ClawMetry currently has adapters for OpenClaw, NemoClaw, and Claude Code. PicoClaw and NanoClaw are smaller/embedded OpenClaw variants. CrewAI/LangGraph/AutoGen need thin SDK wrappers that emit OTLP to ClawMetry's /v1/traces receiver — which already exists. The deliverable is: (1) add PicoClaw + NanoClaw to _CHANNEL_DIRS in clawmetry/sync.py, (2) publish a clawmetry-sdk package with a one-line CrewAI/LangGraph/AutoGen integration that emits OTLP. Where in the codebase Likely affects (scoped to vivekchand/clawmetry): clawmetry/sync.py — _CH

Context From the TokPinch comparison (blog). Today our VelocityBreaker is token-based (max_tokens/window). TokPinch adds two things we lack: Cost spiral: trigger on >$2 spent in a 5-minute window (dollar, not tokens — clearer for users). Exponential-backoff pause: on loop/spiral, pause the session 5→30 min escalating, not just reject the single request. What to build Extend VelocityBreaker (or add CostSpiralBreaker) with a $ threshold: cost_spiral: { enabled, window_seconds: 300, max_usd: 2.0 } using ProxyDB.get_spending(since_ts) (already exists, ~L409). Auto-backoff pause: when loop/spiral/rate fires, write a HITL-style pause with an escalating TTL (5,10,20,30 min) keyed by session, so subsequent requests 429 session_paused until the cool-off elapses — auto-resume after. Reuse the _HITL_DIR/pause_<session> mechanism (~L78-85) but with an expiry timestamp + escalation counter. Acceptance $2 in 5 min → 429 cost_spiral; session auto-paused 5 min; repeat offence escalates

Context From the TokPinch comparison (blog). ClawMetry's LoopDetector (clawmetry/proxy.py ~L778) catches identical request hashes (max_similar in window_seconds). The VelocityBreaker (~L858) caps tokens/window. Gap: Neither catches a content-agnostic request-RATE spike. TokPinch's "rapid fire" rule blocks when one session exceeds >20 requests/minute regardless of content — a runaway that varies its payload slips past our identical-hash detector and may stay under the token cap. What to build A RateBreaker alongside VelocityBreaker: Config: rate_breaker: { enabled, window_seconds: 60, max_requests: 20 }. ProxyDB.get_recent_request_count(session_id, window) (mirror get_recent_token_count). On breach → 429 rate_exceeded (same envelope shape as velocity_exceeded) + emit a rate_exceeded event for alerts/badge/device. Acceptance 21 requests in 60s from one session → 429 rate_exceeded; the event lands in DuckDB and fires the alert pipeline. Default off; opt-in. Grounding: clawmetry/