Skip to content
voxlang.orgv0.6.0SSOT

T3 Wave 4 Gap Closure & Live Pipeline Audit (2026)

Forensic code-to-specification audit comparing the T3 Wave 1–3 architecture blueprints against the live vox_chat_message, agent_loop, t3_wave3_tools, agy_worktree, webhook_intake, and Tauri IPC execution paths. Identifies ten concrete integration gaps where domain primitives were unit-tested in isolation but bypassed, stubbed, or broken on the live request path, and defines the six-phase Wave 4 (v4.8, Phases 265–270) remediation specification.

1. Executive Summary

Section titled “1. Executive Summary”

Across Milestones v4.5 (Phases 247–252), v4.6 (Phases 253–258), and v4.7 (Phases 259–264), Vox implemented the core domain primitives from the T3 research corpus (t3code, t3router, pi-t3chat, thom-chat, helioschat):

  • Typed IDs, Outbox & Compaction: VoxIdAllocatorV2, OutboxQueue, context_sanitizer, CompactionConfig.
  • Router, Consensus, Battery & Stream Guards: ConsensusOrchestrator, BatteryCostLedger, ResponseQualityAnalyzer, StreamIdleWatchdog, ToolRefusalDetector, AtomicResumableStreamBuffer.
  • Worktrees, Checkpoints, MCP & Local Sync: GitPlumbingCheckpointStore, WorkspaceLeaseMap, provision_untracked_env_files, McpOAuthPKCEFlow, CommandOutputHub, ChatTranscriptCache.
  • Agentic UX, Rules, Webhooks & Session Interop: UserRuleRegistry, resolve_user_rule_mentions, interpolate_system_prompt_variables, WebhookDedupeAndRateGate, GuestShareService, ExternalSessionImporter.
  • TOESTUB / vox-code-audit Static Detectors: 9 detectors (wire_unbounded_collection, sql_unbounded_query, v8_Map_Set_hot_path_clone, css_unscoped_transition_all, child_process_shell_true, worktree_git_index_race, git_log_unbounded_all, react_provider_inline_value, llm_stream_missing_idle_watchdog).

However, a line-by-line call-graph audit of crates/vox-orchestrator-mcp/src/chat_tools/chat/ and crates/vox-gui/ui/src/transport/t3Wave3Rpc.ts revealed 10 high-impact integration gaps (GAP-01 through GAP-10) where these modules are either broken on edge cases or bypassed on the live production execution path.

2. Forensic Gap Catalog (GAP-01 – GAP-10)

Section titled “2. Forensic Gap Catalog (GAP-01 – GAP-10)”

GAP-01: tool:<name> Fence Stream Suppression Drops Tool Calls Without Salvaging (salvage_and_stream.rs)

Section titled “GAP-01: tool:<name> Fence Stream Suppression Drops Tool Calls Without Salvaging (salvage_and_stream.rs)”
  • Location: crates/vox-orchestrator-mcp/src/chat_tools/chat/agent_loop/salvage_and_stream.rs (L211 vs L377–390).
  • Root Cause:
    1. looks_like_streamed_tool_call_envelope (L211) checks trimmed.starts_with("```tool:") and returns true, suppressing TokenStreamed events so raw tool fences are not leaked to the user transcript.
    2. When the stream finishes and native tool_calls is empty, salvage_tool_calls_from_text (L377) runs r"(?s)```(?:json|tool_call|xml)?\s*(.*?)\s*```".
    3. For a pi-t3chat-style tool:vox_read_file code fence containing {"path": "Cargo.toml"}, inner captures "tool:vox_read_file\n{\"path\": \"Cargo.toml\"}", which fails serde_json::from_str. Tier 4 (extract_balanced_json_objects) extracts {"path": "Cargo.toml"}, which lacks a "name" field (normalize_salvaged_tool_call_item returns None).
    4. Impact: Both the streamed text AND the salvaged tool call are silently dropped.
  • Remediation (T3-SALVAGE-RULES-01):
    • Add Tier 1.5 in salvage_tool_calls_from_text matching r"(?s)```tool:([A-Za-z0-9_.:-]+)\s*\n(.*?)\s*```". Extract cap[1] as name and parse cap[2] (or balanced JSON object within cap[2]) as arguments, then normalize via normalize_salvaged_tool_call_item.

GAP-02: @rule:<name> Mentions & $variable System Prompt Interpolation Bypassed on Live Chat Turn (mentions.rs, preamble_and_retrieval.rs)

Section titled “GAP-02: @rule:<name> Mentions & $variable System Prompt Interpolation Bypassed on Live Chat Turn (mentions.rs, preamble_and_retrieval.rs)”
  • Location:
  • Root Cause:
    1. resolve_user_rule_mentions (mentions.rs:36) is never called by resolve_mentions (mentions.rs:308) or prepare_turn_context_and_retrieval. Worse, MENTION_RE (@([A-Za-z0-9_.:/\\-]+)) matches @rule:rust-strict and attempts to resolve rule:rust-strict as a workspace file path, emitting [ Mention @rule:rust-strict not found in workspace ]!
    2. interpolate_system_prompt_variables (system_prompt_vars.rs:36) is never called when constructing the system prompt in build_system_prompt_with_skill or prepare_turn_context_and_retrieval.
  • Remediation (T3-SALVAGE-RULES-02..04):
    • In resolve_mentions_with_rules / resolve_mentions, strip and resolve @rule:<name> tags (plus Always-mode rules) via resolve_user_rule_mentions before running MENTION_RE file resolution, loading persisted user rules from state.db (get_user_preference("local_user", "gui.user_rules")) with default fallback.
    • In prepare_turn_context_and_retrieval (preamble_and_retrieval.rs), construct SystemPromptVarContext from state.repository (workspace_root, git branch from .git/HEAD), params.session_id, params.active_file, model_used_pre, and params.cognitive_profile, and pass system_prompt through interpolate_system_prompt_variables.

GAP-03: context_sanitizer & UpToMessageId Compaction Not Wired into conversation.rs

Section titled “GAP-03: context_sanitizer & UpToMessageId Compaction Not Wired into conversation.rs”
  • Location: crates/vox-orchestrator-mcp/src/chat_tools/chat/conversation.rs (L428–454).
  • Root Cause:
    • transcript_entry_to_llm_message (L428) pushes raw entry.content directly into ChatMessage without calling vox_orchestrator::context_sanitizer::sanitize_assistant_for_history on "assistant" turns or escape_prompt_injection_tokens on "user" / "tool" turns.
    • Closed <think>...</think> blocks and <WebSearch>...</WebSearch> blocks from earlier turns remain in context on every subsequent turn.
    • Additionally, vox_chat_compact_up_to (CompactionDirection::UpToMessageId) is not exposed in chat_tools/chat/history.rs.
  • Remediation (T3-HIST-BATTERY-01..02):
    • Wire sanitize_assistant_for_history(&entry.content) for "assistant" messages and escape_prompt_injection_tokens(&entry.content) for "user" messages inside transcript_entry_to_llm_message.
    • Add chat_compact_up_to (vox_chat_compact_up_to) to crates/vox-orchestrator-mcp/src/chat_tools/chat/history.rs and dispatch.rs so the GUI and agents can compact transcript history up to or from a specific anchor message_id.

GAP-04: BatteryCostLedger, VoxIdAllocatorV2 & Real ResponseQualityAnalyzer Scores Bypassed in message.rs & persist_and_envelope.rs

Section titled “GAP-04: BatteryCostLedger, VoxIdAllocatorV2 & Real ResponseQualityAnalyzer Scores Bypassed in message.rs & persist_and_envelope.rs”
  • Location:
  • Root Cause:
    1. BatteryCostLedger (crates/vox-orchestrator/src/budget/tracker.rs) is never instantiated on ServerState or checked/debited during chat_message.
    2. persist_and_envelope.rs:44,53 mints message IDs via format!("usr-{}", now_ts()) and format!("ast-{}", now_ts()) (which collide whenever two turns complete within the same second) instead of VoxIdAllocatorV2::derive.
    3. persist_and_envelope.rs:348 hardcodes quality_score: Some(1.0) in LlmSurfaceTelemetry instead of computing the actual ResponseQualityAnalyzer::analyze_response_quality(&params.prompt, &response_text) score.
  • Remediation (T3-HIST-BATTERY-03..04):
    • Add battery_ledger: Arc<BatteryCostLedger> and id_allocator: VoxIdAllocatorV2 to ServerState.
    • Run state.battery_ledger.check_preflight_balance("local_user", 1.0) before turn execution and state.battery_ledger.record_turn_usage("local_user", &session_id, &model_used, prompt_tokens, completion_tokens) in persist_and_build_chat_envelope.
    • Mint deterministic collision-free usr-* and ast-* IDs via state.id_allocator.derive(...).
    • Compute analyze_response_quality(&params.prompt, &response_text) in persist_and_build_chat_envelope, feed (quality.quality_score / 100.0) into LlmSurfaceTelemetry::quality_score, and include quality_score and battery_spent_bu in the response envelope’s posture_audit.

GAP-05: CommandOutputHub \r Folding, bound_snapshot_metadata DOM Shedding, Pre-Write Sparse Checkpoints, Resumable Stream Buffer & Mid-Turn Steer Draining Not Wired in agent_loop.rs

Section titled “GAP-05: CommandOutputHub \r Folding, bound_snapshot_metadata DOM Shedding, Pre-Write Sparse Checkpoints, Resumable Stream Buffer & Mid-Turn Steer Draining Not Wired in agent_loop.rs”
  • Location:
  • Root Cause:
    1. Tool outputs in agent_loop.rs:542 bypass vox_orchestrator::command_output_hub::normalize_terminal_output (\r progress bar folding) and vox_orchestrator::snapshot::bound_snapshot_metadata (20 KiB DOM snapshot shedding).
    2. Write/mutation tools (should_sparse_checkpoint(&call.name)) execute at L524 without capturing a pre-mutation sparse Git checkpoint (GitPlumbingCheckpointStore::capture_checkpoint).
    3. stream_llm_with_pref_and_events (salvage_and_stream.rs:69) emits AgentEventKind::TokenStreamed to EventBus without appending chunks to a ResumableTurnStreamRegistry (AtomicResumableStreamBuffer) on ServerState.
    4. Between tool iterations (agent_loop.rs:279), promoted steering prompts from ThreadRunQueue (ServerState::thread_run_queue) are never drained into messages.
  • Remediation (T3-LOOP-CHECKPOINT-01..04):
    • Normalize tool output strings with normalize_terminal_output and bound_snapshot_metadata before cap_and_summarize_tool_output.
    • When should_sparse_checkpoint(&call.name) is true, automatically record a pre-tool sparse checkpoint on state.checkpoint_store.
    • Hold resumable_streams: Arc<ResumableTurnStreamRegistry> and thread_run_queue: Arc<RwLock<ThreadRunQueue>> on ServerState; append streamed chunks in stream_llm_with_pref_and_events and expose vox_chat_stream_resume; drain promoted steering items (QueueDisposition::Steer) into messages at the start of each agent_loop iteration.

GAP-06: Synthetic/Stateless Handlers in t3_wave3_tools.rs

Section titled “GAP-06: Synthetic/Stateless Handlers in t3_wave3_tools.rs”
  • Location: crates/vox-orchestrator-mcp/src/t3_wave3_tools.rs (L141–231, L406–564).
  • Root Cause:
    1. handle_consensus_run (L141) generates synthetic string templates when legs is omitted instead of fanning out concurrent call_llm_with_pref calls across the requested models via futures_util::future::join_all (with graceful fallback when no upstream API key is configured).
    2. handle_context_handoff (L194) ignores state (_state) when turns is omitted instead of hydrating source_thread_id’s live conversation history via context_history_or_hydrate(state, &format!("chat_history:{source_thread_id}")) and writing the handoff brief into chat_history:{new_thread_id}.
    3. handle_guest_share_mint & handle_guest_share_moderate (L406–474) instantiate a fresh local GuestShareService::new_from_clavis() on every call instead of sharing a persistent Arc<RwLock<GuestShareService>> on ServerState and appending approved guest prompts into chat_history:{thread_id}.
    4. handle_session_import_external, handle_session_export_bundle, and handle_session_import_bundle (L477–564) ignore state:
      • import_external does not read real file_path / ~/.claude/projects, ~/.codex/sessions, or ~/.cursor/chats files when raw_payload is omitted, nor does it persist imported turns into chat_history:{session_id}.
      • export_bundle does not hydrate chat_history:{session_id} from state when messages is omitted.
      • import_bundle does not write the verified bundle’s messages into chat_history:{session_id} and state.db.
  • Remediation (T3-STATEFUL-TOOLS-01..04):
    • Wire all six handlers in t3_wave3_tools.rs to live ServerState (context_history_or_hydrate, persist_context_Async, state.guest_share_service, and concurrent join_all LLM fan-out in handle_consensus_run).

GAP-07: Missing vox_chat_regenerate & vox_chat_edit_message_in_place Backend MCP Tools (history.rs, dispatch.rs)

Section titled “GAP-07: Missing vox_chat_regenerate & vox_chat_edit_message_in_place Backend MCP Tools (history.rs, dispatch.rs)”
  • Location: crates/vox-orchestrator-mcp/src/chat_tools/chat/history.rs.
  • Root Cause:
    • t3-vox-agentic-ux-rules-webhooks-and-session-interop-spec-2026.md §1.3 specified two distinct message mutation operations:
      1. vox_chat_edit_message_in_place: mutates an assistant (or user) message’s content in-place in chat_history:{session_id} without truncating subsequent messages ([Edited] badge).
      2. vox_chat_regenerate: truncates history from the target message onward and re-triggers chat_message with the stored GenerationPostureSnapshot or an explicit model_override / reasoning_effort.
    • Currently only vox_chat_slash /retry and vox_chat_fork exist.
  • Remediation (T3-REGEN-WORKTREE-01):
    • Implement chat_edit_message_in_place and chat_regenerate in crates/vox-orchestrator-mcp/src/chat_tools/chat/history.rs and register vox_chat_edit_message_in_place, vox_chat_regenerate, vox_chat_compact_up_to, vox_chat_stream_resume, and vox_chat_queue_steer in tool_registry.rs, input_schemas.rs, and dispatch.rs.

GAP-08: DelegationWorktree::create & submit_webhook_events Internal Bypass (agy_worktree.rs, webhook_intake.rs)

Section titled “GAP-08: DelegationWorktree::create & submit_webhook_events Internal Bypass (agy_worktree.rs, webhook_intake.rs)”
  • Location:
  • Root Cause:
    1. DelegationWorktree::create spawns git worktree add without acquiring WorkspaceLeaseMap::global().lease_for(repo_root) and without calling provision_untracked_env_files(repo_root, &wt_path) (L94) after worktree creation.
    2. submit_webhook_events (webhook_intake.rs:188) validates the HMAC signature and immediately pushes events into webhook_inbox without calling WebhookDedupeAndRateGate::admit_delivery (L94) to reject replayed delivery_ids or > 60/min bursts.
  • Remediation (T3-REGEN-WORKTREE-02..03):
    • Acquire the WorkspaceLeaseMap mutex guard across git worktree add and call provision_untracked_env_files(repo_root, &wt_path) in DelegationWorktree::create.
    • Hold a WebhookDedupeAndRateGate on ServerState (or global singleton) and gate every WebhookEventItem in submit_webhook_events via admit_delivery.

GAP-09: vox_orchestrator_status Ignores compact Argument & transfer_budget.rs Uses Synthetic Fixtures

Section titled “GAP-09: vox_orchestrator_status Ignores compact Argument & transfer_budget.rs Uses Synthetic Fixtures”
  • Location:
  • Root Cause:
    • dispatch.rs:806 calls dei_tools::orchestrator_status(state).await unconditionally, ignoring args["compact"] and args["omit_idle_subagents"] instead of routing to dei_tools::orchestrator_status_compact(state, omit_idle_subagents).
    • transfer_budget.rs tests a hardcoded 5-field serde_json::json! literal instead of building a 900-row / 1.1 MiB payload and asserting that orchestrator_status_compact + cap_and_summarize_tool_output keep the wire payload under 256 KiB and 4 DB round-trips.
  • Remediation (T3-REGEN-WORKTREE-04):
    • Route vox_orchestrator_status in dispatch.rs to orchestrator_status_compact when compact == true (defaulting omit_idle_subagents to true), and upgrade crates/vox-orchestrator-mcp/tests/transfer_budget.rs to exercise the real compact projection and 1.1 MiB tool-output capping functions.

GAP-10: t3Wave3Rpc.ts unwrapMcpData Fails on Tauri invoke_mcp_tool { is_error, result } Wrapper (t3Wave3Rpc.ts)

Section titled “GAP-10: t3Wave3Rpc.ts unwrapMcpData Fails on Tauri invoke_mcp_tool { is_error, result } Wrapper (t3Wave3Rpc.ts)”
  • Location: crates/vox-gui/ui/src/transport/t3Wave3Rpc.ts (L88–104).
  • Root Cause:
    • In Tauri desktop mode, voxTransport.invokeMcpTool invokes Tauri command invoke_mcp_tool (crates/vox-gui/src/commands/mod.rs:146), which returns { is_error: false, result: { success: true, data: ... } } (or a JSON string inside result).
    • unwrapMcpData in t3Wave3Rpc.ts:96 only checks if ('data' in obj && obj.data !== undefined) return obj.data as T; at the outer object level. Because data is nested inside obj.result.data, unwrapMcpData returns the outer { is_error: false, result: ... } wrapper, which fails every caller’s field check ('winning_response' in res, 'bundle_json' in res, 'share_token' in res, etc.) and causes the GUI to fall through to hardcoded mock objects even when connected to the live Rust backend!
  • Remediation (T3-GUI-ENVELOPE-01..04):
    • Fix unwrapMcpData in t3Wave3Rpc.ts to recursively unwrap { is_error, result }, stringified JSON payloads, MCP { content: [{ type: 'text', text: '...' }] } blocks, and { success, data } envelopes (throwing when is_error === true or success === false).
    • Add typed client wrappers regenerateChatTurn, editChatMessageInPlace, compactChatUpTo, resumeChatStream, and steerChatQueue in t3Wave3Rpc.ts, wire in-place assistant message editing and posture-aware regeneration into ChatTranscript.tsx, and add Vitest coverage verifying real Tauri { is_error: false, result: { success: true, data: ... } } envelope unwrapping.

3. Milestone v4.8 Phase Mapping (Phases 265–270, 24 Requirements)

Section titled “3. Milestone v4.8 Phase Mapping (Phases 265–270, 24 Requirements)”
PhaseRequirement IDsScope
Phase 265T3-SALVAGE-RULES-01..04Fix tool:<name> fence salvage in salvage_and_stream.rs, wire @rule:<name> + Always rules into resolve_mentions (mentions.rs), and wire interpolate_system_prompt_variables into preamble_and_retrieval.rs.
Phase 266T3-HIST-BATTERY-01..04Wire sanitize_assistant_for_history & escape_prompt_injection_tokens into conversation.rs, expose vox_chat_compact_up_to, and wire live BatteryCostLedger, VoxIdAllocatorV2, and ResponseQualityAnalyzer into message.rs & persist_and_envelope.rs.
Phase 267T3-LOOP-CHECKPOINT-01..04Wire normalize_terminal_output (\r), bound_snapshot_metadata (20 KiB DOM cap), automatic pre-write sparse Git checkpoints (should_sparse_checkpoint), ResumableTurnStreamRegistry (vox_chat_stream_resume), and mid-turn ThreadRunQueue steer draining into agent_loop.rs.
Phase 268T3-STATEFUL-TOOLS-01..04Upgrade t3_wave3_tools.rs to dispatch concurrent join_all LLM calls in vox_consensus_run, hydrate/persist real chat_history:{session_id} transcripts in vox_context_handoff & vox_session_*, and persist GuestShareService on ServerState.
Phase 269T3-REGEN-WORKTREE-01..04Implement vox_chat_regenerate & vox_chat_edit_message_in_place in history.rs, wire WorkspaceLeaseMap + provision_untracked_env_files in DelegationWorktree::create, wire WebhookDedupeAndRateGate in submit_webhook_events, and wire compact in vox_orchestrator_status + transfer_budget.rs.
Phase 270T3-GUI-ENVELOPE-01..04Fix unwrapMcpData in t3Wave3Rpc.ts for nested Tauri { is_error, result: { success, data } } & MCP content blocks, wire regenerateChatTurn & editChatMessageInPlace in ChatTranscript.tsx, and verify all Rust, Vitest, TypeScript, and audit gates.