Skip to content
Unverified: This page has not yet been anchored to automated code tests.

Axis Harness River: Autonomous 3-Tier Cascade, Live Alluvial Telemetry & Ludus Unification

Axis Harness River: Autonomous 3-Tier Cascade, Live Alluvial Telemetry & Ludus Unification

Section titled “Axis Harness River: Autonomous 3-Tier Cascade, Live Alluvial Telemetry & Ludus Unification”

Defines the Axis Harness River visual and data architecture for vox-gui, replacing the static HarnessRedirect empty state with a live 4-zone telemetry HUD: a top throughput waveform and 3-tier model cascade strip (Interceptor, Graph Fan-Out, Causal Verifier), a Canvas 2D alluvial particle river with a central BLAKE3 receipt chokepoint and sweeping phase laser, an 8-slice SVG agent trajectory braid, and a deterministic 4-stage verification retention funnel unified with LudusSandbox gamification rewards.

1. Context & Visual AI Audit of the Reference Post

Section titled “1. Context & Visual AI Audit of the Reference Post”

On 2026-10-13, @marfinxx published a viral demonstration (status/2109992692167970904, quoting their article status/2081445614957850624 “Master Agent Architecture: Unifying Harness, Loop, and Graph Engineering”) depicting a 3-tier autonomous engineering harness (“Interceptor $\to$ Parallel Fan-Out $\to$ Causal Red-Team Verifier”) alongside a monochrome brutalist telemetry HUD (“the live river”).

Visual AI inspection of the 25.1-second primary video (720×872 portrait HUD) and the companion article video (1280×720) reveals four vertically stacked telemetry regions:

  1. Zone A — Top Telemetry & Waveform Strip (~12% height):
    • Left: Audio-oscilloscope bar waveform encoding token throughput (THROUGHPUT TOK/S), CACHE HIT %, SAVINGS $, and ACTIVE AGENTS.
    • Center/Right: 3-Tier Model Assignment Readout (TIER 1: INTERCEPT [<18ms / -88.4% TOKENS], TIER 2: GRAPH FAN-OUT [PARALLEL SUB-AGENTS], TIER 3: CAUSAL RED-TEAM [ADVERSARIAL VERIFY]) and 5 stage-completion bars (HARNESS INTERCEPT, EVIDENCE LOOP, GRAPH FAN-OUT, CAUSAL RED-TEAM, SANDBOX PR GATE).
  2. Zone B — Center “Live River” Alluvial Particle Bundle + Laser Scanner (~48% height):
    • An alluvial/Sankey particle stream rendered on 2D Canvas: hundreds of high-contrast cubic-Bézier strands originate from top source slots (workspace files / sub-agent branches), converge through a luminous central Singularity (the Interceptor & Verification Chokepoint), and fan out to bottom verified output targets.
    • A horizontal sweeping laser scanner bar glides vertically across the river with an inverted high-contrast callout pill naming the active execution layer (e.g., PHASE 1: HARNESS INTERCEPT [BLAKE3_HASH_CACHE], LEVEL 1: EVIDENCE TEST LOOP [CARGO_NEXTEST_COLD], PHASE 2: GRAPH FAN-OUT [PARALLEL_BRANCHES], LEVEL 3: CAUSAL RED-TEAM [ SOCRATIC_DOUBT ]).
  3. Zone C — Agent Role Trajectory Braid (~18% height):
    • Labeled AGENT ROLE TRAJECTORY FLOW — LAST 8 SLICES: a multi-strand sinusoidal bump chart where each active agent’s horizontal strand weaves vertically across 8 time slices as it transitions between roles (INTERCEPT, FAN-OUT, VERIFY, IDLE).
  4. Zone D — Deterministic Evidence Footer (~22% height, 3 columns):
    • Left (FAILURE / DOUBT ATTRIBUTION LEDGER): Horizontal bars breaking down intercepted or rejected turns (STALE_FILE_HASH, COMPILE_REGRESSION, ADVERSARIAL_EDGE_CASE, FABRICATED_RECEIPT).
    • Center (VERIFICATION RETENTION FUNNEL): 4-stage inverted trapezoid funnel showing candidate patch survival (100% CANDIDATE TURNS $\to$ AST / LINT GATE $\to$ COLD TEST SUITE $\to$ ADVERSARIAL PASS).
    • Right (TELEMETRY DRIFT + ZERO-CASCADE SLO): Monospace pass/fail badges for BLAKE3 tool receipts, lock contention, and session spend vs. budget.

1.2 Make-vs-Buy Assessment (External Repo vs. Vox Native)

Section titled “1.2 Make-vs-Buy Assessment (External Repo vs. Vox Native)”

Inspection of the companion article (status/2081445614957850624) confirms there is no external open-source repository to import. The article’s code listings are ~90 lines of illustrative Python (hashlib.sha256, subprocess.run(["pytest"]), asyncio.gather) paired with a synthetic Canvas animation.

By contrast, Vox already owns production-grade Rust backends that exceed every mechanism described in the post:

Capability Claimed in PostExisting Vox Production SSOTGap to Close in Axis GUI
Tier 1 Interceptor (SHA-256 file cache, token filtering)crates/vox-orchestrator/src/tool_receipt.rs (ToolReceiptLedger, BLAKE3 call_args_hash + result_hash, HMAC tag, ReceiptValidationResult) & crates/vox-orchestrator/src/cache_predictor.rs (CachePredictor, CacheSignal, overlap_ratio)Surface BLAKE3 cache hits, token thrift %, and receipt verification state in the top strip and river chokepoint.
3-Tier Model Routing (Economy $\to$ Standard $\to$ Strong)crates/vox-orchestrator/src/tier_cascade.rs (RoutingTier::{Economy, Standard, Strong}, CompositeSignal) + DriveConsole.tsx (ClutchId, RiskId) + ModelPickerSearch.tsxMap ClutchId (free/efficiency/balanced/genius) + RiskId (high/moderate/low) + RailRouting into an explicit 3-tier INTERCEPT / FAN-OUT / VERIFIER cascade header with click-to-open model/routing controls.
Causal Loop & Adversarial Verificationvox-oratio Socratic Doubt (suspect_task, finding_validated, task_doubted) + vox-Verify (compilation_succeeded, test_passed, contract_satisfied) in crates/vox-orchestrator/src/events.rsFeed live AgentEventFrame / TurnEventDto streams into the 4-stage VerificationFunnelFooter and sweeping laser scanner callout.
Dockable Harness Panel in ChatChatSurface.tsx already registers 'harness' in OPT_IN_PANEL_IDS (HarnessDockPanel), but mounts HarnessRedirect.tsx, a static empty state!Upgrade HarnessRedirect / HarnessRiverPanel so Panels ▾ → Harness in Chat and Workspace → Harness render the live responsive Axis Harness River.

Decision: Do not import external code. Build the Axis Harness River natively in crates/vox-gui/ui/src/components/surfaces/Harness/ on pure Canvas 2D + SVG + React 19, wired directly to Vox’s existing Tauri events, RailRouting, and DriveConsole state.


2. Audit of Pre-Existing Gamification & Visualizer Surfaces

Section titled “2. Audit of Pre-Existing Gamification & Visualizer Surfaces”

2.1 Why LudusSandbox (Vox Urbs) Falls Short

Section titled “2.1 Why LudusSandbox (Vox Urbs) Falls Short”

Our existing citybuilder visualizer in crates/vox-gui/ui/src/components/gamify/LudusSandbox.tsx and crates/vox-gui/ui/src/components/gamify/urbs/worldRenderer.ts has four structural weaknesses compared to the Harness River:

  1. Poll-bound, not event-reactive (20s interval): LudusSandbox polls workspace_town_scan and fetchHarnessSnapshot every 20_000 ms. It does not react to vox://agent-events (token_streamed, tool_completed, task_doubted), making it feel static while agents are actively coding.
  2. Low-contrast emoji aesthetics: Agents are rendered as HTML 🧍 emoji divs floating over brown/grey isometric rectangles (#1c1917), which lacks the high-contrast kinetic clarity of an engineering HUD.
  3. Disconnected from Model Tiers & Token Thrift: AQVAE (the aqueduct overlay) is hardcoded to lit: false (LudusSandbox.tsx line 183), and neither BLAKE3 cache savings nor the 3-tier model cascade (Economy / Standard / Strong) is visible or rewarded.
  4. Split-brain standalone app: apps/experimental/visualizer/ is an orphaned v0.2.0 standalone Vite app with hardcoded mock tasks.

2.2 Unification, Upgrade, and Retirement Strategy

Section titled “2.2 Unification, Upgrade, and Retirement Strategy”
  • Upgrade Harness (HarnessRiverPanel + HarnessRedirect):
    • Replace the dead empty-state in HarnessRedirect with HarnessRiverPanel while preserving the onFocusComposer action button ("Focus composer") so deep links and existing tests remain green.
    • Support two responsive layouts via dockview width observation (HARNESS_FULL_WIDTH_PX = 420):
      • Compact Dock Mode (< 420px, docked beside Chat transcript): Vertical 3-Tier Cascade header + Live Alluvial River Canvas with sweeping phase laser + compact 4-stage Verification Funnel strip.
      • Wide / Full Surface Mode (>= 420px, wide dock or Workspace → Harness / Gamify): All 4 zones expanded (Top Waveform & 3-Tier Strip, Center Alluvial River + Singularity + Laser Scanner, 8-Slice Agent Role Trajectory Braid, and 3-Column Evidence Footer).
  • Unify with GamifyView & Dashboard (River | Urbs Mode Switcher + Token Thrift Yield):
    • In GamifyView.tsx and Dashboard.tsx, add a simulation mode toggle ("Urbs Map" vs. "Live River") alongside a Ludus Thrift & Verification Yield badge (+XP / ◆ Crystals derived from BLAKE3 cache hits and cold-verified turns). Keep "Urbs Map" (<LudusSandbox />) as the default initial tab in GamifyView so GamifyView.test.tsx’s synchronous <canvas> query continues to pass without regression, while allowing one-click switching to "Live River".
  • Retire apps/experimental/visualizer:

flowchart LR
subgraph Controls["Loquela & Model Selector"]
CMP["ChatModelPicker<br/>(Auto / Pinned Family)"]
DC["DriveConsole<br/>(ClutchId + RiskId)"]
RR["RailRouting<br/>(Decision Preview + Alternatives)"]
end
subgraph Stream["Live Orchestrator Telemetry"]
AE["vox://agent-events<br/>(task_started, token_streamed,<br/>task_doubted, task_completed)"]
TE["ChatMessage.events<br/>(TurnEventDto: tool_called,<br/>verify_passed, routing_decision)"]
end
subgraph Store["Pure Correlation Layer (riverMath.ts)"]
TC["deriveTierCascade()<br/>Tier 1 Intercept · Tier 2 Fan-Out · Tier 3 Verifier"]
RF["deriveRiverSnapshot()<br/>Throughput · BLAKE3 Thrift % · 4-Stage Funnel · 8-Slice Braid"]
end
subgraph UI["HarnessRiverPanel (Dockable + Full Surface)"]
Z1["Zone A: WaveformTierStrip"]
Z2["Zone B: RiverCanvas (Alluvial Béziers + Singularity + Laser)"]
Z3["Zone C: AgentTrajectoryBraid (8-Slice SVG)"]
Z4["Zone D: VerificationFunnelFooter (3-Col Evidence + Ludus Yield)"]
end
CMP --> TC
DC --> TC
RR --> TC
AE --> RF
TE --> RF
TC --> Z1
TC --> Z2
RF --> Z1
RF --> Z2
RF --> Z3
RF --> Z4

3.1 Deterministic 3-Tier Cascade Mapping (deriveTierCascade)

Section titled “3.1 Deterministic 3-Tier Cascade Mapping (deriveTierCascade)”

The 3 tiers in the Harness River HUD always reflect the user’s live DriveConsole (ClutchId + RiskId), ChatModelPicker override, and RailRouting:

Tier SlotRole in HarnessAuto Mapping from ClutchId / RiskId / RailRouting
Tier 1: INTERCEPTBLAKE3 receipt cache (ToolReceiptLedger), AST hash check, token triageAlways maps to Economy tier (Flash-Lite / Haiku / Local MENS) or routing.alternatives[last] (<18ms target).
Tier 2: GRAPH FAN-OUTParallel branch execution across active orchestrator tasksMaps to chatModelOverride if pinned; else routing.model (Auto → <label>); scales branch concurrency display with ClutchId (free $\to$ economy, efficiency/balanced $\to$ standard, genius $\to$ frontier).
Tier 3: CAUSAL VERIFIERSocratic doubt (vox-oratio), adversarial test gate (vox-Verify)Driven by RiskId: low risk (strict) elevates Verifier to Strong / Opus / Argon + Socratic Doubt; moderate runs deterministic compile + test gate; high risk runs fast receipt-only verification.

Clicking any tier pill in WaveformTierStrip invokes onOpenRouting?.() (opening the inline Routing drawer in ChatSurface) or onFocusComposer?.() (focusing the Loquela model/mode bar).


4. Planning for Antigravity IDE + Gemini 3.8 Flash / Gemini 4 Argon

Section titled “4. Planning for Antigravity IDE + Gemini 3.8 Flash / Gemini 4 Argon”

Per gemini-3-5-flash-antigravity-limitations-2026-06-18.md and antigravity-handoff-and-skill-gaps-2026-06-18.md, the implementation plan in 2026-10-13-axis-harness-river.md enforces six structural guardrails:

  1. Explicit Model Tier Routing (FLASH-3.8 vs. ARGON-4):
    • Gemini 3.8 Flash ([MODEL: FLASH-3.8]) executes bounded, single-component React/SVG tasks and mechanical wiring where all types and props are pre-specified (Tasks 3, 4, 5, 7, 8).
    • Gemini 4 Argon ([MODEL: ARGON-4]) is assigned to tasks requiring multi-invariant state correlation, pure cubic-Bézier + DPR Canvas math, or multi-component responsive assembly (Tasks 1, 2, 6), and serves as the two-strike escalation target for any Flash task that hits 2 consecutive test failures.
  2. Atomic Green Tasks: Every task creates or updates its unit test first, implements the exact module, and runs timeout 180s pnpm --dir crates/vox-gui/ui exec vitest run <test-file> to green before finishing.
  3. Verify-Before-Use (Step 0): Every task starts with an rg -n command confirming the exact line numbers and signatures of target files before any edit.
  4. Zero Placeholders: Every task embeds complete, copy-pasteable TypeScript/React code with zero TODOs or elided blocks.
  5. Strict File-Write Disjointness: Wave 1 tasks (Tasks 1–5) write to completely disjoint new files under crates/vox-gui/ui/src/components/surfaces/Harness/ and can run in parallel ([PARALLEL-SAFE]). Wave 2 tasks (Tasks 6–8) wire existing surfaces (HarnessRedirect.tsx, ChatSurface.tsx, GamifyView.tsx, Dashboard.tsx) and run strictly sequentially ([SEQUENTIAL]).
  6. Two-Strike Circuit Breaker: If any command fails twice with the same error under Gemini 3.8 Flash, the agent halts immediately with DRIVE: STOPPED and escalates the task to Gemini 4 Argon.