mergify · Sep 30, 2026 · Sep 30, 2026 · Oct 11, 2026 · Oct 11, 2026
diff --git a/README.md b/README.md
@@ -201,6 +201,7 @@ Dashboard: **Agent Traces** (`agent-traces`); Tempo span-metrics carry `llm_mode
 | OpenClaw gateway | `diagnostics.otel` (unchanged) | `openclaw-gateway` | `transform/phoenix_openclaw` (agent id + channel from the system prompt's `Runtime:` line; drops the duplicate `openclaw.content.*` copy; **privacy:** agent `mira` spans lose all message content in Phoenix, tagged `privacy.content_dropped=mira`) | `diagnostics.otel.traces=false` |
 | Digital twin (Cloudflare Worker, public ask endpoint) | hand-rolled OTLP/JSON tracer in the Worker (`src/telemetry.js`), exported after the response over Cloudflare Tunnel + Access to the collector's **`otlp/twin`** receiver (container port 4320, not host-published, `bearertokenauth/twin`, token `secrets/twin_otel_token` from `.env` `TWIN_OTEL_BEARER_TOKEN`; `filter/twin` drops any service not named `*-digital-twin`). Dashboard `digital-twin-traces` | `digital-twin` | none (the Worker sets OpenInference itself) | Worker var `TRACING="off"`; Phoenix drops bot rejects (`method_not_allowed`, `forbidden_origin`) via `tail_sampling/twin_phoenix`, Tempo keeps them |
 | SpecStride | `lib/ralph_otel_spans.py`: prompt on the attempt/critic span (`input.value` <= `SPECSTRIDE_OTEL_PROMPT_MAX`, `specstride.prompt.{path,sha256,bytes}`), critic reply as `output.value`; exports `TRACEPARENT` so the claude/dsh child nests inside the run | `specstride` (children too) | `transform/phoenix_kind` | `SPECSTRIDE_OTEL_TRACES=false` |
+| agentic-netops intent tier (kind cluster, 4 LangGraph agents) | each agent's `ioa-observe-sdk` POSTs OTLP/HTTP **once** to the tier-owned collector (`agent-otel-collector`, ns `agentic-netops-agents`), which fans out to ClickHouse, feature-001 and this collector via an `otlphttp/phoenix` exporter (`AOS_OTLP_HTTP_ENDPOINT`, default `http://172.17.0.1:4318` from inside kind). Spans are OpenInference-NATIVE (ioa-observe-sdk is a rebranded Traceloop SDK): `ChatLiteLLM.chat` spans carry `gen_ai.prompt/completion` content + `gen_ai.usage.*`/`llm.usage.*` tokens; decorator spans are `<name>.agent` with `agent.name`; graph runs are `LangGraph.workflow` + `*.task`; A2A round-trips propagate W3C context so a request is one trace | `intent-supervisor`, `intent-mapper`, `intent-allocator`, `intent-deployer` (one per `service.name`) | `transform/phoenix_agentic_netops` (LLM kind on `ChatLiteLLM.chat` spans — content + tokens already native via `gen_ai.*`; the literal `llm.model_name=unknown` from the oauth2/ shim is fixed from `ioa_observe.association.properties.ls_model_name`; AGENT kind on `<name>.agent` + `LangGraph.workflow` spans; A2A round-trips arrive as single traces filed under the supervisor's project) | remove `otlphttp/phoenix` from the traces pipeline exporters in `deploy/agents/telemetry.yaml` (agentic-netops) and re-apply |
 
 Nesting: Claude Code honours an inbound `TRACEPARENT` env in `-p`/SDK sessions (`parent.source=env`
 on `claude_code.interaction`); dsh honours it and forwards `traceparent` to LiteLLM, which honours
@@ -209,7 +210,13 @@ caller's project; Tempo keeps each service name. Known limits: Claude Code traci
 smoke test fails when `claude_code.llm_request` spans stop mapping to LLM); detailed content
 (`new_context`, `response.model_output`, `tool_input`) exists only in `-p`/SDK sessions, and the
 system prompt is a 500-char preview + hash on the span (full text once per hash in Loki,
-event `system_prompt`); OpenClaw strips session ids by design, so its traces have no `session.id`.
+event `system_prompt`); OpenClaw strips session ids by design, so its traces have no `session.id`;
+the agentic-netops intent tier's source-side redaction (agent-tier policy T333-T335 in
+`agents/common/llm.py`) is currently inert — the SDK's context scoping keeps those
+`agentic-netops.*` attributes off exported spans — so its Phoenix spans carry FULL
+prompt/completion content (consistent with this stack's content-capture-on policy); the
+literal `llm.model_name=unknown` is corrected to the resolved model by
+`transform/phoenix_agentic_netops`.
 
 ## Docs
 

diff --git a/bin/gen-dashboards.py b/bin/gen-dashboards.py
@@ -278,42 +278,6 @@ def write(d, folder="."):
 ]
 write(dashboard("OpenClaw Agents — Cost, Latency & Traces", "agents", ag, ["agents", "llm"]), "ai-agents")
 
-# ---------------- Claude Code (by model + agent/subagent) ----------------
-# claude_code.* OTel metrics -> Prometheus. token_usage labels: model, type
-# (input/output/cacheRead/cacheCreation), query_source (main/subagent), session_id.
-# Detailed per-request EVENTS -> Loki (logs panel below). Metrics are session-driven.
-_id = itertools.count(1)
-# Claude Code's OTLP *metrics* are unreliable for real usage: short sessions exit before the 10s
-# metric flush, and the counters are ephemeral (they vanish ~5min after a session ends). The per-request
-# `api_request` LOG events are exported immediately and PERSIST in Loki — so the consumption panels are
-# driven from Loki. Fields on api_request: cost_usd, input_tokens, output_tokens, cache_read_tokens,
-# cache_creation_tokens, model, duration_ms. Subagents come from subagent_completed (agent_type, model,
-# total_tokens). All totals below are over the dashboard time range ($__range) — true persistent history.
-CC = '{service_name="claude-code"} | event_name=`api_request`'
-cc = [
-    lstat("Cost (range)", f'sum(sum_over_time({CC} | unwrap cost_usd [$__range]))', "currencyUSD", 6, 4, 0, 0),
-    lstat("Output tokens", f'sum(sum_over_time({CC} | unwrap output_tokens [$__range]))', "short", 6, 4, 6, 0),
-    lstat("Cache-read tokens", f'sum(sum_over_time({CC} | unwrap cache_read_tokens [$__range]))', "short", 6, 4, 12, 0),
-    lstat("Requests", f'sum(count_over_time({CC} [$__range]))', "short", 6, 4, 18, 0),
-    lts("Cost by MODEL", [(f'sum by (model)(sum_over_time({CC} | unwrap cost_usd [$__interval]))', "{{model}}")], "currencyUSD", 12, 8, 0, 4, stack=True),
-    lts("Tokens by TYPE (in / out / cacheRead / cacheCreation)", [
-        (f'sum(sum_over_time({CC} | unwrap input_tokens [$__interval]))', "input"),
-        (f'sum(sum_over_time({CC} | unwrap output_tokens [$__interval]))', "output"),
-        (f'sum(sum_over_time({CC} | unwrap cache_read_tokens [$__interval]))', "cacheRead"),
-        (f'sum(sum_over_time({CC} | unwrap cache_creation_tokens [$__interval]))', "cacheCreation")], "short", 12, 8, 12, 4, stack=True),
-    lts("Output tokens by MODEL", [(f'sum by (model)(sum_over_time({CC} | unwrap output_tokens [$__interval]))', "{{model}}")], "short", 12, 8, 0, 12, stack=True),
-    lts("Requests by MODEL", [(f'sum by (model)(count_over_time({CC} [$__interval]))', "{{model}}")], "short", 12, 8, 12, 12, stack=True),
-    lbar("Cache-read tokens by MODEL (range)", f'sum by (model)(sum_over_time({CC} | unwrap cache_read_tokens [$__range]))', "short", 12, 8, 0, 20),
-    lts("Request latency p95 (by model)", [(f'quantile_over_time(0.95, {CC} | unwrap duration_ms [$__interval]) by (model)', "{{model}}")], "ms", 12, 8, 12, 20),
-    # ----- Named subagents (from Loki subagent_completed events: agent_type, model, total_tokens) -----
-    lbar("Subagent runs by type", 'sum by (agent_type) (count_over_time({service_name="claude-code"} | event_name=`subagent_completed` [$__range]))', "short", 12, 8, 0, 28),
-    lbar("Subagent tokens by type", 'sum by (agent_type) (sum_over_time({service_name="claude-code"} | event_name=`subagent_completed` | unwrap total_tokens [$__range]))', "short", 12, 8, 12, 28),
-    lts("Subagent tokens over time (by type)", [('sum by (agent_type) (sum_over_time({service_name="claude-code"} | event_name=`subagent_completed` | unwrap total_tokens [$__interval]))', "{{agent_type}}")], "short", 12, 8, 0, 36, stack=True),
-    lts("Subagent activity (by model)", [('sum by (model) (count_over_time({service_name="claude-code"} | event_name=`subagent_completed` [$__interval]))', "{{model}}")], "short", 12, 8, 12, 36, stack=True),
-    logs("Recent Claude Code events (Loki — request/tool/subagent: model, tokens, agent_type)", '{service_name="claude-code"}', 24, 11, 0, 44),
-]
-write(dashboard("Claude Code", "claude-code", cc, ["ai", "claude-code"]), "ai-agents")
-
 # ---------------- LLM Cost & Consumption (unified Claude Code + LiteLLM/OpenClaw) ----------------
 _id = itertools.count(1)
 # Two cost sources: Claude Code (persistent Loki api_request events) + OpenClaw/LiteLLM (persistent
@@ -474,83 +438,6 @@ def write(d, folder="."):
 write(dashboard("SpecStride Runs & Traces", "specstride-traces", spec, ["ai", "specstride", "traces", "ralph"], refresh="30s",
                 templating=[dict(var_query("task", 'label_values({service_name="ralph"}, task)', ds=LOKI), allValue=".*")]), "ai-agents")
 
-# ---------------- Agent Mix (12.1 frontier-fading tree) ----------------
-# The fading dashboard. ONE event stream (job="agentops", from agent-pack `agentops emit`) feeds
-# every panel here; the SAME events also drive the kanban board (two sinks, one stream — they can
-# never disagree). Loki stream labels (low-cardinality): job="agentops", event, task. Everything
-# else is logfmt in the line: model, agent, verdict, rung, attempts, tokens_in, tokens_out,
-# wall_ms, dial_state, degraded, reason, contract_id, card, adhoc, task_class.
-#   Events: dispatch (task handed to a leaf), verdict (pass|fail|escalated), escalation (kicked up
-#   the ladder), adhoc (bebop ask — a question, no card). LOCAL model = qwen* ; FRONTIER = the rest
-#   (claude-*, gpt-*). GPU-degraded-by-reason is the ONE Prometheus panel: metric
-#   agentops_gpu_degraded{reason=...} from the S3/H1 watchdog — reason="cpu-fallback" (H1) shows as
-#   its own series so a silent CPU-fallback is visually distinct from a device-health degrade (S3).
-_id = itertools.count(1)
-AOVERD = '{job="agentops", event="verdict", task=~"$task"}'
-AODISP = '{job="agentops", event="dispatch", task=~"$task"}'
-AOESC  = '{job="agentops", event="escalation", task=~"$task"}'
-LOCAL_RE = 'qwen.*'                       # local worker; everything else is frontier
-MNOTE = (
-    "**Frontier-fading agent tree (12.1).** opus plans/verifies, the local qwen worker executes, and "
-    "trust is earned per task-class. Every panel is the persistent Loki event stream "
-    "(`job=\"agentops\"`, one event per dispatch/verdict/escalation) over the dashboard range — the "
-    "**same** events that drive the kanban board. **Local** = `qwen*`; **Frontier** = `claude-*` / "
-    "`gpt-*`. The goal of fading is the two numbers up top moving in opposite directions: local task "
-    "share **up**, frontier tokens (cost proxy) **down**, without local-first pass-rate dropping below "
-    "the promotion bar (0.85). GPU-degraded is split **by reason** so a silent CPU-fallback (H1, "
-    "`reason=cpu-fallback`) reads distinctly from a device-health degrade (S3)."
-)
-mix = [
-    text("About this dashboard", MNOTE, 24, 4, 0, 0),
-    # ----- headline stats over the range -----
-    lstat("Tasks completed", f'sum(count_over_time({AOVERD} [$__range]))', "short", 6, 4, 0, 4),
-    lstat("Local task share",
-          f'sum(count_over_time({AOVERD} | logfmt | model=~`{LOCAL_RE}` [$__range])) '
-          f'/ sum(count_over_time({AOVERD} | logfmt [$__range]))',
-          "percentunit", 6, 4, 6, 4,
-          thresholds=[{"color": "red", "value": None}, {"color": "orange", "value": 0.5}, {"color": "green", "value": 0.8}]),
-    lstat("Escalation rate",
-          f'sum(count_over_time({AOESC} [$__range])) '
-          f'/ sum(count_over_time({AODISP} [$__range]))',
-          "percentunit", 6, 4, 12, 4,
-          thresholds=[{"color": "green", "value": None}, {"color": "orange", "value": 0.2}, {"color": "red", "value": 0.4}]),
-    stat("GPU degraded now", 'max(agentops_gpu_degraded) or vector(0)', "short", 6, 4, 18, 4,
-         thresholds=[{"color": "green", "value": None}, {"color": "red", "value": 1}]),
-    # ----- frontier share: tokens + task counts, local vs frontier -----
-    lts("Tokens: local vs frontier (cost proxy)", [
-        (f'sum(sum_over_time({AOVERD} | logfmt | model=~`{LOCAL_RE}` | unwrap tokens_out [$__interval]))', "local (qwen)"),
-        (f'sum(sum_over_time({AOVERD} | logfmt | model!~`{LOCAL_RE}` | unwrap tokens_out [$__interval]))', "frontier (claude/gpt)")],
-        "short", 12, 8, 0, 8, stack=True),
-    lts("Task mix: local vs frontier (count)", [
-        (f'sum(count_over_time({AOVERD} | logfmt | model=~`{LOCAL_RE}` [$__interval]))', "local (qwen)"),
-        (f'sum(count_over_time({AOVERD} | logfmt | model!~`{LOCAL_RE}` [$__interval]))', "frontier (claude/gpt)")],
-        "short", 12, 8, 12, 8, stack=True),
-    # ----- the fading metrics, BY task-class -----
-    lbar("Local-first pass rate by class (promote ≥ 0.85)",
-         f'sum by (task)(count_over_time({AOVERD} | logfmt | model=~`{LOCAL_RE}` | verdict=`pass` | attempts=`1` [$__range])) '
-         f'/ sum by (task)(count_over_time({AOVERD} | logfmt | model=~`{LOCAL_RE}` [$__range]))',
-         "percentunit", 12, 8, 0, 16, legend="{{task}}"),
-    lbar("Escalation rate by class",
-         f'sum by (task)(count_over_time({AOESC} [$__range])) '
-         f'/ sum by (task)(count_over_time({AODISP} [$__range]))',
-         "percentunit", 12, 8, 12, 16, legend="{{task}}"),
-    # ----- dial position + cost proxy breakdown -----
-    lbar("Tasks by dial position (fading state)",
-         f'sum by (dial_state)(count_over_time({AOVERD} | logfmt [$__range]))',
-         "short", 12, 8, 0, 24, legend="{{dial_state}}"),
-    lbar("Frontier tokens by model (range) — cost proxy",
-         f'sum by (model)(sum_over_time({AOVERD} | logfmt | model!~`{LOCAL_RE}` | unwrap tokens_out [$__range]))',
-         "short", 12, 8, 12, 24, legend="{{model}}"),
-    # ----- GPU degraded windows, SPLIT BY reason (H1: cpu-fallback distinct from S3 device-health) -----
-    ts("GPU degraded windows (by reason) — cpu-fallback = H1, others = S3 device-health",
-       [('agentops_gpu_degraded', "{{reason}}")], "short", 24, 7, 0, 32),
-    # ----- raw event stream -----
-    logs("Recent agent-pack events (dispatch / verdict / escalation / adhoc)",
-         '{job="agentops", task=~"$task"}', 24, 11, 0, 39),
-]
-write(dashboard("Agent Mix — Frontier Fading (12.1)", "agent-mix", mix, ["ai", "agents", "12.1", "fading"],
-                templating=[var_query("task", 'label_values({job="agentops"}, task)', ds=LOKI)]), "ai-agents")
-
 # ---------------- Containers ----------------
 _id = itertools.count(1)
 cn = [