Read the agentic stack across Model, Harness, Coordination & Lineage, and Environment & Security.
Concentric growth rings encode cross-layer telemetry over the selected window. 48 radial clock sectors map Model token density, Harness execution rhythms, Coordination queue handoffs, and Environment security invariants.
Synthesized executive intelligence across model caching, multi-harness dispatch, and autonomous coordination.
Prompt prefix alignment in Claude & Codex harnesses kept uncached input at 1.54M tokens direct.
Context pressure remained nominal; tool invocations converged with zero unredacted body capture.
Lineage walk verified 0 cycles across 18 sender/receiver actor pairs under continuous watchdog assurance.
Analyze fleet execution dynamically across harnesses, architectural layers, and chronological time buckets.
| Harness / Tool | Status | Calls / Messages | Observed Tokens | Token Share | Action |
|---|
| Architectural Layer | Recorded Facts | Volume Share | Active Anomalies | Primary Insight Focus | Action |
|---|
| Date / Time Bucket | Model Vol | Harness Vol | Coord Vol | Env Vol | Total Facts | Cross-Layer Distribution |
|---|
Observe cross-layer causality: Model → Harness → Coordination → Environment across real agent execution spans.
Every agent action decomposed across the 5 critical execution layers: 10 Behavioral Tool Families (ToolArchetype), Built-in Harness Primitives, Skills (skill__*), MCP Servers (mcp__*), and Lifecycle Hooks (HookEvent).
read_file)
Core L1
edit_file)
Atomic
write_file)
Artifact
grep_search)
Discovery
find_by_name)
Discovery
run_command)
Sandbox
invoke_subagent)
Delegation
Provides local, private contextual memory, profile manifests, and semantic search across owner identity without unapproved publication.
holoself_context, holoself_search, holoself_status, holoself_proposal_create
Provides real-time Cypher graph queries, node upserts, edge lineages, and relationship knowledge across the agent topology.
cypher_query, graph_status, graph_upsert_node, graph_upsert_edge
Message and capability telemetry across harnesses and models. Counts by role, direction, and modality. Message bodies are never captured. I02 accounts tokens. I08 accounts settled action failure. All-time counts — window_from/window_to label the snapshot but do not filter.
Token and cost attribution across models, tasks, and actors. Unknown cost is never treated as zero.
Direct parent-child session hierarchy with explicit lineage tracking. Missing parents remain unresolved without false roots. Direct parent-child session counts from v_subagent_lineage; depth, cycle, and unresolved-parent verdicts come from canonical insight I09 (missing parents remain unresolved and are never imputed as roots). All-time counts — window_from/window_to label the snapshot but do not filter.
Sender and receiver actor pairs across queue handoffs and return loops. The counts are queue flow, not token attribution.
Observed actor time split into active execution, idle-with-work, container hold, and leaf stall, plus session yield. Partitioned by actor and session across execution bands.
Closed-vocabulary observations across policy, grant, approval, attempt, effective authority, and effect. A grant is not effective authority. Command text is never stored. All-time counts — window_from/window_to label the snapshot but do not filter.
CPU deltas and max sampled RSS per unit, process scope, and process attribution. I14 environment.resources is CPU deltas and max sampled RSS. I18 coverage.source counts sources. Max RSS per unit, process scope, and process attribution. The SQL view does not recompute CPU or coverage. All-time counts — window_from/window_to label the snapshot but do not filter.
Verification evidence for recovery, vortex loops, cache bleed, and independent peer review.
Distance from the cohort of completed sessions. A trip is an anomaly signal, not an inferred root cause. A trip is not a cause. All-time counts — window_from/window_to label the snapshot but do not filter.
Envelope shape distributions, plus keys added, removed, and retyped. Short samples report insufficient evidence. L1 distance of envelope shape distributions, plus keys added, removed, and retyped. A short sample is insufficient_sample. All-time counts — window_from/window_to label the snapshot but do not filter.
Downstream sessions on evidenced edges only. The count reflects dependency exposure, not inherited authority. All-time counts — window_from/window_to label the snapshot but do not filter.
Simulated fast-tier cost from an operator-pinned table. It is an analytical simulation, never an active runtime mutation. It is not an effect and not a routing change. All-time counts — window_from/window_to label the snapshot but do not filter.
Offline and connected reader for actors, sessions, work lineage, model accounting, and security posture.
All Layers stay addressable. Counts roll through Model Layer, Harness Layer, Coordination & Lineage, Environment & Security, and Coordination & Topology.
usage_facts —
message_facts — · activity_facts — · tool_facts — · skill_facts —
queue_facts — · session_edges —
security_facts —
At most seven headline questions. The compiled strip below is the I01–I25 catalog index, not this headline list.
Where do direct model tokens go when cache is kept separate?
Which observed model cost is known, and which estimate stays unknown?
Which harness actions failed after they settled?
Where does work lineage stop, cycle, or leave a parent unresolved?
Which placements account for CPU time and sampled memory?
Which downstream sessions does an evidenced grant reach?
Is a trajectory trip a cause, or only a distance from the cohort?
Catalog index I01–I25. L0 is status, finding, practical meaning, scope, and missingness. This strip does not recompute measures.
Count distinct evidenced call IDs. Usage-only rows are separate. No admitted call facts → not_observable. The presenter does not recompute this measure.
Status: Computed when usage rows exist; Refused on an empty population or a failed store read. Formula: disjoint sums of input_uncached, cache_write, cache_read, and output per model. Reasoning stays inside output and is not added again. cache_read_ratio_of_sums and cache_read_mean_of_ratios are separate measures.
Inclusive sum over the reachable observation union. Nonadditive. Incomplete nonnegative lineage is a lower bound. Ambiguous duplicates refuse the canonical total.
Status: Computed, or Partial when observed or estimated USD is unknown. Formula: observed_usd and estimated_usd stay on separate measures and are never added. Unknown is not zero. The pricing table ships empty, so estimates stay unknown.
Observed bindings that differ from contemporaneous desired bindings, or from paired observed bindings. Unknown values are excluded and counted. Missing paired population → not_observable.
Distinct loaded and invoked capability IDs, and the invoked∩loaded / loaded rate. Absent loaded evidence refuses the rate. Actions still count.
Evidenced context-used / context-limit at invocation. Prompt tokens do not substitute. A missing limit refuses the ratio.
Status: Computed, or Partial when unsettled actions remain. Formula: failed divided by settled terminal actions (failed IS NOT NULL). Unsettled actions (failed IS NULL) are counted separately and are never treated as success.
Status: Computed, or Partial when a parent is unresolved or a cycle stops the walk. Formula: explicit session_edges only. Relations include spawn, delegate, handoff, and return. Roots, depth, and diamonds are measured. Cycles are flagged. Unresolved parents are counted and never invented as roots.
Wall union, summed attempt-seconds, and summed/wall. Half-open same-clock segments. A censored contribution is partial. wall=0 refuses the concurrency ratio.
Shared usage, weighted beneficiaries, and residual. Rational weights are ≥ 0 and sum ≤ 1. Absent policy leaves the shared bucket.
Sum or median of ready-to-start elapsed intervals, and a separate wall union. Missing ready, start, or cross-clock uncertainty refuses the affected intervals.
Distinct incarnations and placement intervals by locus, plus same-label rebuild count. An unknown incarnation gets an unresolved count.
Status: Computed when identities and samples exist; Refused on an empty population. Formula: CPU deltas and max sampled RSS. Per root_key and harness, sum CPU deltas on readable non-decreasing samples and take the max sampled RSS per rss_unit, with launched versus discovered attribution. RSS units are never mixed.
Normalized successful and failed syscall counts by operation and resource class. An unmatched call has an unknown result. No trace → not_observable.
Paired declared / configured / tested dimension comparisons. A lone declaration is insufficient. Namespace equality is configuration, not a universal verification.
Comparator counts of consistent, mismatch, and insufficient by an explicit rule. No applicable policy → NOT_ASSESSED, never a zero violation count.
Status: Computed for an unwindowed scope; Refused when a from or to bound is set. Formula: read, projected, unprojected-by-reason, and refused counts per source. A missing source is not a zero.
Currently accepted eligible work / rule-eligible evaluated work, with attempt success kept separate. No acceptance rule → not_observable.
Matched-population cost / accepted work, and a separate mean of per-work ratios. Zero accepted is null, not infinity. Cost stays on I04.
Stratified difference of same-version measures. Match clock basis, grain, accounting, units, and coverage. Biased or unmatched selection → noncomparable.
Distance from the cohort centroid of completed sessions. A trip is not a cause. L1 assumptions block that reading. The SQL trigger only marks the session dirty.
L1 distance of envelope_shape key-path and variant_type distributions, plus keys added, removed, and retyped. A short sample is reason insufficient_sample and badge insufficient_evidence.
Downstream sessions on status = evidenced edges only. The count is dependency exposure, not inherited or effective authority. PageRank is not an input.
Simulated fast-tier cost and context pressure from an operator-pinned tier table. It is not an effect and not a routing change. An unpinned cohort or tier profile refuses the simulation and keeps observed I02.
Recorded window: —
Total rows: —
Total rows: —
Harnesses: —
| SQL View Token | Layer | Canonical Insight | Description |
|---|---|---|---|
v_message_mix | Model · Harness | I02 / I08 | Message and capability telemetry by role and modality |
v_mission_token_burn | Model | I02 / I04 | Token and cost attribution across models and actors |
v_subagent_lineage | Coordination | I09 | Parent-child session edges, depth, cycle detection |
v_handoff_rework_analysis | Coordination | M2 / Handoff | Sender and receiver actor pairs across queue returns |
v_seat_idle_ranking | Harness | Actor Yield | Active execution, idle-with-work, container hold, leaf stall |
v_security_posture | Environment | I16 / I17 | Closed-vocabulary observations across policy and authority |
v_process_resources | Environment | I14 / I18 | CPU deltas and max sampled RSS per harness and root key |
v_executive_governance | Coordination | Scorecard | Verification evidence for recovery, cache bleed, peer review |
v_trajectory_vortex | Harness | I22 | Distance from cohort centroid of completed sessions |
v_schema_drift | Harness | I23 | L1 distance of envelope shape key-paths and variants |
v_authority_blast_radius | Environment | I24 | Downstream sessions on evidenced dependency edges |
v_counterfactual_routing | Model | I25 | Simulated fast-tier cost from operator-pinned tier table |