This document records the Milestone v2.3 Before/After full-flywheel audit across all chat modifying knobs, execution tracks (sync, background, plan), interaction modes (plan, act, verify), runtime tiers (local, mesh, cloud, auto), clutch profiles (free, efficiency, balanced, genius), risk postures (high, moderate, low), OpenRouter vs Local-to-MENS model selection, strategy-aware BPE context compaction (aggressive, balanced, conservative), and the 25-problem 5-spoke + 15-problem organic search-to-compilation MENS training and evaluation harness.
| Flywheel Dimension | BEFORE (v2.2 Baseline) | AFTER (v2.3 Remediated) | Verification Artifact / Test |
|---|
| Chat Modifying Knob End-to-End Fidelity | 68.2% (15/22 knobs reached all 3 tracks; oversight_level & kb_scope dropped in buildChatTurn.ts, mode & grounding_check_enabled dropped on Sync, scope_files dropped on Plan) | 100.0% (24/24 keys wired across TS + Rust Sync, Background, and Plan tracks; 432 TS + 144 Rust matrix combinations verified) | crates/vox-gui/ui/src/lib/buildChatTurn.test.ts, crates/vox-gui/src/commands/chat_turn.rs |
| OpenRouter Selection & Local-to-MENS Regression | 2 Regressions (tier=local / clutch=free blocked VoxLocal MENS fast-path via !res.enforce_free_tier_only; clutch=free + risk=low skipped free models when effective_axes shifted to QUALITY_FIRST) | 0 Regressions (60/60 cells across 5 spokes $\times$ 4 clutches $\times$ 3 risk postures route to calibrated mens/qwen3_27b_five_spoke_v1 over uncalibrated ollama/qwen3:8b) | crates/vox-orchestrator/src/models/registry.rs, crates/vox-orchestrator-mcp/src/llm_bridge/model_route_policy/tests.rs |
| Context Management & Session Pin Hydration | Fixed 4,000-token chars / 4 tail-only window; 0% turn-0 goal anchor retention on overflow; kb_scope ignored; session pins not injected into user_prompt | Strategy-scaled cl100k_base BPE budget (Aggressive=2,400, Balanced=4,000, Conservative=8,000); 100% turn-0 anchor retention (20% head reserve); KbScopeMask 5-mode leg filter; [PINNED CONTEXT] + [EXECUTION GOVERNANCE] injection | crates/vox-orchestrator-mcp/src/chat_tools/chat/conversation.rs, crates/vox-orchestrator-mcp/src/chat_tools/chat/message.rs |
| MENS 5-Spoke & Organic Training Pipeline + Data Audit | 25/25 5-spoke suite (pass@1 = 1.00), 15/15 organic suite (100%), route_spoke_task hardcoded CostPreference::Economy | 25/25 5-spoke + 15/15 organic suite + route_spoke_task_with_policy across all (clutch, risk, prefer_local) axes + 0 malformed research_metrics rows | contracts/reports/full-flywheel-knob-matrix-before-after.v1.json |
| Knob / Field | Values Tested | BEFORE (Sync / Background / Plan) | AFTER (Sync / Background / Plan) |
|---|
execution | sync, background, plan | Routed vox_chat_message, submit_orchestrator_task, vox_plan | Preserved (100%) |
mode | plan, act, verify | Forwarded to Background only; dropped on Sync | Forwarded to Sync (sync_tool_args + [EXECUTION GOVERNANCE]) and Background (enqueue_hints.mode) |
tier | local, mesh, cloud, auto | Forwarded to Sync and Background | Forwarded to Sync and Background + fixed local VoxLocal MENS fast-path |
clutch | free, efficiency, balanced, genius | Forwarded to Sync and Background | Forwarded to Sync and Background + fixed free + low risk axis interaction |
risk | high, moderate, low | Forwarded to Sync and Background | Forwarded to Sync and Background |
oversight_level | 1, 2, 3, 4, 5 | Dropped in buildChatTurn.ts (never left TS composer) | Wired end-to-end: CHAT_TURN_KEYS $\rightarrow$ ChatTurnInput $\rightarrow$ sync_tool_args ([EXECUTION GOVERNANCE]) & SubmitTaskInput (enqueue_hints.oversight_level) |
kb_scope | auto, docs, code, memory, off | Dropped in buildChatTurn.ts and ignored in message.rs | Wired end-to-end: CHAT_TURN_KEYS $\rightarrow$ ChatTurnInput $\rightarrow$ KbScopeMask::from_kb_scope in message.rs and enqueue_hints.kb_scope |
context_files + Session Pins | Workspace paths & gui.session_context_pins.<id> | Recorded on ChatTranscriptEntry, omitted from user_prompt on Sync and omitted from scope_files on Plan | Hydrated from gui.session_context_pins.<session_id> + context_files into [PINNED CONTEXT] on Sync, file_manifest on Background, and scope_files on Plan |
grounding_check_enabled | true, false | Local post-check + Background, omitted from sync_tool_args | Forwarded in sync_tool_args (ChatMessageParams::grounding_check_enabled) + Background |
dry_run | true, false | Forwarded on wire; no prompt-level read-only guard on Sync | Forwarded on wire + injects dry_run=true (read-only simulation; do not mutate files) in [EXECUTION GOVERNANCE] |
| Clutch Profile | Risk Posture | Effective Axes (Cost, Responsiveness, Intelligence) | Free Pool Enforced | OpenRouter Tier Selected | Local / MENS Route Selected |
|---|
Free | High / Moderate | (70, 15, 15) (COST_FIRST) | true | openrouter/*:free (ModelTier::Free) | mens/qwen3_27b_five_spoke_v1 (ProviderType::VoxLocal) |
Free | Low (ModelLean::Intelligence) | (25, 15, 60) (QUALITY_FIRST) | true (allow_free_in_performance_mode = true) | Highest-prior openrouter/*:free | mens/qwen3_27b_five_spoke_v1 (ProviderType::VoxLocal) |
Efficiency | High / Moderate | (65, 20, 15) | false | ModelTier::Fast / Free workhorse | mens/qwen3_27b_five_spoke_v1 when prefer_local |
Efficiency | Low (ModelLean::Intelligence) | (25, 15, 60) | false | ModelTier::Pro workhorse | mens/qwen3_27b_five_spoke_v1 when prefer_local |
Balanced | High / Moderate | (40, 30, 30) (BALANCED) | false | ModelTier::Pro (claude-sonnet-4.6 / gemini-2.5-flash) | mens/qwen3_27b_five_spoke_v1 when prefer_local |
Balanced | Low (ModelLean::Intelligence) | (25, 20, 55) | false | ModelTier::Pro / Elite | mens/qwen3_27b_five_spoke_v1 when prefer_local |
Genius | High / Moderate / Low | (15, 15, 70) (QUALITY_FIRST) | false | ModelTier::Elite / Pro (claude-opus-4.7) | mens/qwen3_27b_five_spoke_v1 when prefer_local |
ROUTE-REG-01 (resolve.rs:494 — tier="local" / clutch="free" blocked VoxLocal MENS fast-path):
- BEFORE:
if vox_local_route_preferred && !res.enforce_free_tier_only prevented ProviderType::VoxLocal (which has is_free == true and $0 cost) from taking the vox_local_route_preferred fast-path whenever tier="local" or clutch="free" set res.enforce_free_tier_only = true.
- AFTER: Replaced
!res.enforce_free_tier_only with .filter(|m| !res.enforce_free_tier_only || m.is_free) and populated *rationale_out with "VoxLocal MENS fast-path preferred for <task> (<model_id>)".
ROUTE-REG-02 (resolve.rs:39 & registry.rs:973 — clutch=Free + risk=Low rejected free models in best_for_internal):
- BEFORE:
effective_axes(ClutchProfile::Free, RiskPosture::Low) returns (25, 15, 60) (QUALITY_FIRST), whose to_cost_preference() is CostPreference::Performance. Because SelectionIntent::allow_free_in_performance_mode defaulted to false, best_for_internal skipped all is_free == true models before enforce_free_tier_if_needed ran.
- AFTER: Set
intent.allow_free_in_performance_mode = prefer_local || matches!(clutch, Some(ClutchProfile::Free)) and added ModelRegistry::route_spoke_task_with_policy, achieving 60/60 zero-regression cells across all 5 spokes (VoxLang, RustWorkspace, StructuredShell, McpToolChain, AgentHarness).
| Metric | BEFORE (conversation.rs) | AFTER (bound_messages_by_strategy) |
|---|
| Token Estimator | Crude chars / 4 (approx_tokens) | CompactionEngine::estimate_tokens (cl100k_base BPE) |
| Strategy Budgets | Hardcoded 4,000 across all strategies | Aggressive = 2,400, Balanced = 4,000, Conservative = 8,000 (hydrated from gui.context_budget.strategy) |
| Turn-0 Goal Anchor Preservation | 0% when conversation exceeds budget (oldest-first drop) | 100% under Balanced and Conservative (reserves up to 20% of budget for turn-0 anchor) |
Knowledge Base Scope (kb_scope) | Always ran all 5 retrieval legs even when kb_scope = "off" | KbScopeMask: off skips run_retrieval_bundle; docs keeps KG+chunks; code keeps RRF+repo; memory keeps memory; auto keeps all |
- Canonical 25-Problem 5-Spoke Suite (
vox-corpus::training::spoke_roundtrip): 25/25 passed (pass@1 = 1.00, pass@k = 1.00, 5/5 per spoke across VoxLang, RustWorkspace, StructuredShell, McpToolChain, AgentHarness).
- Organic 15-Problem Search-to-Compilation Suite (
vox-corpus::training::organic_ benchmark): 15/15 passed (100% across Easy, Medium, Hard difficulty tiers and all 5 spokes).
- Curriculum Weight Floor & DPO Compiler-Diagnostic Repair Pairs: All 5 spokes satisfy
weight >= 0.10 (DEFAULT_MIN_SPOKE_WEIGHT_FLOOR) with $\sum w_i = 1.00$; extract_dpo_repair_pairs_from_outcomes deterministically pairs failed compiler/clippy diagnostics with repaired solutions.
VoxDb::audit_research_metric_shapes Data Audit: 0 malformed rows across benchmark_event, syntax_k_event, RoutingDecision, and socrates_surface (100% schema conformance).
- Before (
v2.3): Adjusting composer knobs (clutch, risk, tier, mode, execution_mode, oversight_level, kb_scope, grounding_check_enabled) only updated React state in Loquela.tsx and only took effect on the next buildChatTurn submission from that view. External MCP calls, subagent turns, vox_plan, and vox_submit_task did not inherit the live session posture.
- After (
v2.4):
- Added Tauri RPCs
get_session_chat_posture and set_session_chat_posture (crates/vox-gui/src/commands/context_budget.rs) writing to gui.chat_posture.<session_id> and gui.chat_posture.default (plus gui.compaction_strategy when set).
vox_chat_message (message.rs), vox_plan (plan.rs), and vox_submit_task (submission.rs) automatically hydrate omitted knob fields from load_persisted_session_chat_posture(state, session_id) and persist explicit overrides via persist_session_chat_posture.- Every
vox_chat_message response now emits structured compaction_audit (strategy, budget_tokens, head_anchor_preserved, total_loaded, kept_messages, dropped_messages) and posture_audit (clutch, risk, tier, mode, oversight_level, kb_scope, grounding_check_enabled, hydrated_from_session_posture).
| Keybind | Knob / Action | Semantic Shape & Coloration |
|---|
Alt+1 | Clutch $\rightarrow$ Free ($0) | Emerald capsule segment (border-emerald-500/50 bg-emerald-500/20 text-emerald-200) |
Alt+2 | Clutch $\rightarrow$ Efficient | Sky-blue capsule segment (border-sky-500/50 bg-sky-500/20 text-sky-200) |
Alt+3 | Clutch $\rightarrow$ Balanced | Amber-gold capsule segment (border-amber-500/55 bg-amber-500/20 text-amber-100) |
Alt+4 | Clutch $\rightarrow$ Genius | Violet capsule segment (border-violet-500/55 bg-violet-500/20 text-violet-100) |
Alt+C | Cycle Clutch (Free $\rightarrow$ Efficient $\rightarrow$ Balanced $\rightarrow$ Genius) | Segmented rounded-full capsule group with active glow ring |
Alt+R | Cycle Risk (High $\rightarrow$ Moderate $\rightarrow$ Low) | Rose (High), Amber (Moderate), Emerald (Low) rounded-full pill with status dot |
Alt+T | Cycle Runtime Tier (Auto $\rightarrow$ Local $\rightarrow$ Mesh $\rightarrow$ Cloud) | Neutral (Auto), Emerald (Local), Cyan (Mesh), Violet (Cloud) capsule pill |
Alt+M | Cycle Execution Track (Sync $\rightarrow$ Background $\rightarrow$ Plan) | Sky (Quick chat), Indigo (Delegate), Amber (Write a plan) capsule pill |
Alt+V | Cycle Interaction Mode (Act $\rightarrow$ Plan $\rightarrow$ Verify) | Synchronized with execution track and persisted to session posture |
Alt+O | Cycle Oversight (1: Auto .. 5: Strict) | Emerald (1-2), Amber (3: Balanced), Rose (4-5: Cautious/Strict) capsule pill |
Alt+K | Cycle @kb Scope (auto $\rightarrow$ docs $\rightarrow$ code $\rightarrow$ memory $\rightarrow$ verified $\rightarrow$ off) | Teal/Sky/Indigo/Violet/Emerald/Zinc rounded-full pill |
Alt+/ | Toggle Real-Time Keybind HUD (loquela-keybind-hud-toggle / loquela-keybind-hud-panel) | Brass capsule pill (Keys Alt+/) + interactive click-to-cycle HUD dialog |
- Live Turn Receipt & Anchor Badge (
chat-turn-receipt-badge in ChatExecutionRail.tsx): Surfaces Goal Anchor: Kept · <budget> BPE (<kept>/<total> msgs) and <clutch>/<risk> · @<kb_scope> in real time via get_session_chat_posture and vox:session-posture-change.
- Continuous 5-Spoke MENS Ratchet (
ModelRegistry::verify_full_flywheel_routing_and_spoke_ratchet): Continuously verifies 25/25 5-spoke pass@1 = 1.0 and 60/60 (clutch, risk, spoke) routing cells.
| Surface ID | Before Score | After Score (v2.3) | After Score (v2.4 Knobs + Live Badge) | Edge Sharpness $\Delta$ | Contrast $\Delta$ | PSNR (dB) | Converged |
|---|
loquela-knobs | 0.8280 | 0.9162 | 0.9222 (+0.0942) | +6.0% | +76.7% | 38.14 dB | true (0 defects) |
chat-default | 0.8502 | 0.9203 | 0.9203 (+0.0701) | +12.8% | +62.1% | 38.12 dB | true (0 defects) |
chat-dock-multi | 0.7985 | 0.9216 | 0.9216 (+0.1231) | +4.1% | +64.1% | 38.14 dB | true (0 defects) |
urbs-civilization | 0.8228 | 0.9180 | 0.9180 (+0.0952) | +24.5% | +14.5% | 38.10 dB | true (0 defects) |