SOTA Agent Harness Comparative Analysis & v3.4 Architecture (Vox vs. Cursor, Antigravity & Claude Code)

This document synthesizes an empirical trajectory post-mortem and architectural benchmark comparing the Vox Axis agent harness (vox-orchestrator, vox-orchestrator-mcp, vox-gui) against October 2026 frontier harnesses (Cursor, Antigravity, and Claude Code). It specifies and documents Milestone v3.4 (Phases 178–182), delivering prefix-locked two-tier tool selection (FROZEN_PREFIX_CORE_TOOLS), AST symbol-slice extraction and signature-preserving skeleton folding, <2ms pre-edit cross-crate Cargo.toml and repo-policy preflight guards, streaming delta.tool_calls read prefetch, workspace test canonicalization, and FIFO multi-call test doubles.

1. Empirical Trajectory Post-Mortem (v3.3 Execution Audit)

Analyzing a full multi-crate milestone implementation (Phases 173–177) exposed five concrete classes of latency and token waste common to Cursor, Antigravity, and Claude Code:

Failure / Friction Class	What Happened in Standard Agent Loops	Measured Cost	Vox v3.4 Deterministic Elimination
1. Unverified Cross-Crate Imports & Signatures	Edits introduced use vox_telemetry::... in vox-gui (E0432) and guessed struct/method signatures across crate boundaries (E0599, E0609), discovered only after a cold cargo test build.	~70s wall-clock, ~4,200 tokens	preflight_check_write_guards (ast_slice_and_preflight.rs) checks use vox_* imports against target Cargo.toml in <1ms inside PreToolUse.
2. Per-Turn BM25 Tool Prefix Reshuffling	Re-sorting top-20 tool schemas on every turn (select_tools_for_turn_ranked) mutates the leading tools JSON array in the request prefix, invalidating provider KV caches.	3,000–6,000 uncached prefix tokens/turn	select_prefix_locked_tools_for_turn freezes FROZEN_PREFIX_CORE_TOOLS (12 core tools) in strictly invariant order at indices 0..12, appending spoke-ranked tools after the frozen prefix.
3. Line-Window File Reads & Mid-Line Character Truncation	Inspecting 1,000–2,200 LoC files required multiple 200-line reads (~15,000 tokens), and cap_and_summarize_tool_output sliced mid-function.	~15,000 input tokens	extract_symbol_slice extracts exact fn/struct/impl/interface blocks with top-of-file imports; fold_code_to_ast_skeleton preserves all signatures while eliding function bodies.
4. Workspace Test Runner CLI Traps	cargo test -p vox-gui --lib failed on a binary-only crate; pnpm test -- --run <file> ran all 336 Vitest suites (22.5s) instead of 2 (0.78s).	29x slower Vitest run (22.5s vs 0.78s)	canonicalize_workspace_command (lifecycle_hooks.rs) rewrites -p vox-gui --lib to --bin vox-gui and test -- --run <file> to exec vitest run <file>.
5. Post-Hoc Repo Guard & God-Object Drift	noVersionedModelIds.test.ts and toestub arch/god_object (>500 non-blank LoC) only ran after implementation completed, requiring a second refactor turn.	3 extra edit+test turns (~45s)	Inline write-time preflight (preflight_check_write_guards) blocks hardcoded versioned model literals / banned crypto imports and warns/flags >450/>500 non-blank LoC immediately.

---

2. Head-to-Head Matrix: Vox v3.4 vs. Cursor, Antigravity & Claude Code

Capability Axis	Cursor (Oct 2026)	Antigravity (Oct 2026)	Claude Code (Oct 2026)	Vox Axis Harness (v3.4)
Prompt KV-Cache Stability	Static tool set per mode	Fixed tool declarations	Prompt-cached system + deferred ToolSearch	Prefix-Locked Two-Tier (FROZEN_PREFIX_CORE_TOOLS + Spoke Tail + Deferred Catalog Hint)
Code Read Granularity	File / line-range chunks + semantic search	800-line read_file windows	Line-range Read + Grep	AST Symbol-Slice (extract_symbol_slice) + Signature-Preserving Skeleton Folding (fold_code_to_ast_skeleton)
Pre-Edit Verification	Post-edit shadow workspace LSP	Post-edit run_command / cargo	Post-edit PreToolUse / PostToolUse shell hooks	In-Memory <2ms Pre-Edit Crate-Edge (Cargo.toml), arch/god_object LoC, noVersionedModelIds & Crypto Guard
Streaming Tool Execution	Speculative fast-apply on single file	Batch tool execution after turn stream	Sequential/batch after tool_use block	SSE delta.tool_calls Incremental Assembly + Read-Only Prefetch Readiness (try_parse_streaming_read_prefetch)
Model Routing & Learning	Static cloud model picker + Auto	Static model inheritance (inherit/pro/flash)	Manual /model + subagent tier	Empirical Bayes CommunityIntelligenceBundleV1 + Online Bradley-Terry Arena + Local 5-Spoke MENS Fast-Apply

---

3. Milestone v3.4 Implementation Architecture (Phases 178–182)

3.1 Phase 178: Prefix-Locked Two-Tier Tool Selection (CACHE-PREFIX-01..04)

• Module: crates/vox-orchestrator-mcp/src/chat_tools/chat/agent_loop/ast_slice_and_preflight.rs
• Mechanism:
  • FROZEN_PREFIX_CORE_TOOLS defines the 12 highest-frequency workspace tools in a fixed canonical sequence.
  • select_prefix_locked_tools_for_turn guarantees that indices 0..frozen_prefix_count are identical across turns regardless of user prompt wording, while the remaining slots up to max_tools (20) are populated from the spoke-conditioned BM25 ranking (select_tools_for_turn_ranked).
  • Overflow tools (rank 20..28) are summarized into PrefixLockedToolSelection::deferred_catalog_hint ([DEFERRED MCP TOOLS AVAILABLE ON DEMAND: ...]) and recorded in turn telemetry (frozen_prefix_tools_count, deferred_tool_catalog_hint).

3.2 Phase 179: AST Symbol-Slice Read & Skeleton Folding (AST-SLICE-01..04)

• Module: crates/vox-orchestrator-mcp/src/chat_tools/chat/agent_loop/ast_slice_and_preflight.rs
• Mechanism:
  • extract_symbol_slice(source, symbol_name) scans Rust, TypeScript/TSX, and .vox source for fn, struct, enum, trait, impl, interface, type, const, or function declarations, includes preceding attributes/doc-comments (#[...], ///, @...) and top-of-file use/import headers, and extracts the balanced brace/statement span with 1-based line numbers.
  • fold_code_to_ast_skeleton(source, max_chars) preserves all top-level and member signatures (pub fn, pub struct, export interface, use, import) while folding implementation bodies into /* ... N body line(s) elided ... */.

3.3 Phase 180: Pre-Edit Cross-Crate Grounding & Write-Time Repo Guards (PREFLIGHT-GROUND-01..04)

• Modules: ast_slice_and_preflight.rs & lifecycle_hooks.rs
• Mechanism:
  • preflight_check_write_guards(path, content, cargo_toml_content) inspects write payloads in <2ms:
    1. Detects use vox_<crate>::... imports missing from the target crate’s Cargo.toml (unlinked_crates).
    2. Computes non-blank LoC (non_blank_lines), setting god_object_warning at 450..=500 lines and god_object_exceeded above 500 lines.
    3. Flags banned direct crypto imports (chacha20poly1305, ed25519_dalek, x25519_dalek, sha2, sha3, md5) outside crates/vox-crypto/ and hardcoded versioned model IDs (claude-sonnet-4.6, etc.) in .ts/.tsx files.
  • Wired directly into evaluate_pre_tool_use_hook so write tools block banned literals/imports immediately and inject LoC/dependency advisories into additional_context.

3.4 Phase 181: Speculative Streaming Tool Prefetch (STREAM-EXEC-01..04)

• Module: crates/vox-orchestrator-mcp/src/chat_tools/chat/agent_loop/salvage_and_stream.rs
• Mechanism:
  • try_parse_streaming_read_prefetch(tool_name, arguments_raw) checks in-flight delta.tool_calls accumulators as chunks arrive and validates read-only (is_read_only_memoizable) JSON arguments as soon as their outer {...} closes, enabling zero-fallback single-pass SSE tool turns.

3.5 Phase 182: Workspace Test Canonicalizer, FIFO FakeBackend & MENS Fast-Apply (FAST-APPLY-01..04)

• Modules: lifecycle_hooks.rs & crates/vox-gui/src/remote/calls.rs
• Mechanism:
  • canonicalize_workspace_command rewrites cargo fmt --all → vox run scripts/fmt.vox -- --all, -p vox-gui --lib → -p vox-gui --bin vox-gui, and test -- --run <file> → exec vitest run <file>.
  • resolve_mens_fast_apply_target selects the lowest-latency local 5-spoke MENS model (SelectionAxes::LATENCY_FIRST, complexity = 3, prefer_local = true) for fast code-edit application.
  • FakeBackend in crates/vox-gui/src/remote/calls.rs uses VecDeque<Result<Value, String>> per tool/GET key so multi-call tools (such as 2-arm chat_arena_compare) dequeue distinct responses in FIFO order while retaining the final entry for single-response callers.