Vox-Populi Extraction Follow-Up Plan (2026)

Vox-Populi Extraction Follow-Up Plan

For agentic workers: REQUIRED SUB-SKILL: Use superpowers:subagent-driven-development (recommended) or superpowers:executing-plans to implement this plan task-by-task. Steps use checkbox (- [ ]) syntax for tracking.

Parent spec: plugin-system-redesign-2026.md

Predecessor work: SP1–SP8 implementation plans + the SP3 sub-batches A–D, NVML extraction, and mesh scaffold all landed on the claude/infallible-lalande-baf300 branch (~70 commits). This follow-up plan picks up where they stopped.

Status update (2026-05-08): Browser, oratio, and transport extractions all landed end-to-end in this session. Verified slim-core: chromiumoxide, symphonia, rubato, candle-core, burn, wgpu, nvml-wrapper are all OUT of cargo tree -p vox-cli for default builds. populi-transport is now opt-in (vox-orchestrator’s default no longer pulls axum’s mesh router, JWT, dashmap, blake3, ed25519-dalek, turso). vox-skills retirement is partial: parser/bundle/manifest types moved to vox-plugin-host::skill_*; SkillRegistry (DB-backed install/search/hydrate) remains in vox-skills as it has different semantics. Remaining work: vox-populi mens/tensor candle module cleanup (still feature-gated, kept for the in-tree merge CLI’s adapter-schema bridge), full SkillRegistry unification.

Status update (2026-05-07): Completed candle/NVML extraction from vox-populi mens: deleted candle_qlora_merge.rs, adapter_schema_v3.rs, hardware/nvml.rs. Migrated vox-mens schola merge-qlora CLI to cached_code_plugin("mens-candle-cuda") dispatch. Inline serde-only v2/v3 schema in CLI (no candle deps). Burn/wgpu files (burn_stack.rs, lora/) remain — they are mens-gpu feature-gated and not candle (Unit 3 scope). All 3 CI gates pass; cargo tree -p vox-cli shows no candle/burn/wgpu/nvml.

Honest accounting of what landed vs. what’s deferred

The foundation is sound: ABI-versioned vox-plugin-host loader works end-to-end (proven by 11 host integration tests + the candle-cuda gating spike), seven extension-point traits exist, the SSOT catalog is enforced by build-time + runtime validators, and vox plugin / vox bundle CLI commands work for install / remove / build / verify. But several “extraction complete” claims overstated reality. This plan corrects the record and scopes what remains.

What’s actually in plugins

Capability	Plugin crate	Status
Mens/Candle training loop	vox-plugin-mens-candle-cuda	✅ run_full_training real (3,128 LOC moved)
NVML hardware probe	vox-plugin-nvml-probe	✅ Full extraction
Compiler skill	vox-plugin-skill-compiler	✅ Full migration
7 other agent skills	vox-plugin-skill-{git,memory,orchestrator,rag,testing,testing-validate,v0}	✅ Full migration

What’s scaffolded only (trait + plugin shell, zero code moved)

Capability	Plugin crate	LOC still in source crate
Tensor backend (Burn/wgpu)	vox-plugin-tensor-burn-wgpu	3,054 LOC in vox-tensor/src/
Audio capture (Whisper)	vox-plugin-oratio	2,275 LOC in vox-oratio/src/backends/
Audio capture (mic)	vox-plugin-oratio-mic	(subset of vox-oratio)
Cloud sync	vox-plugin-cloud	mens-cloud feature in vox-populi
Script execution	vox-plugin-script-execution	(vox-eval, vox-exec-grammar)
Mesh transport	vox-plugin-populi-mesh (composite)	4,990 LOC in vox-populi/src/transport/

What’s partially extracted

• vox-populi mens residual: Candle files deleted in this session (2026-05-07): candle_qlora_merge.rs, adapter_schema_v3.rs removed from vox-populi/src/mens/tensor/; nvml.rs removed from vox-populi/src/mens/hardware/. vox-mens schola merge-qlora CLI now dispatches via cached_code_plugin("mens-candle-cuda"). Remaining in mens/tensor/: Burn/wgpu files (burn_stack.rs, lora/, optim.rs, train.rs, tensor.rs) are feature-gated by mens-gpu and are NOT candle — they belong to the Burn/wgpu training path (Unit 3 scope). backend_candle_qlora.rs remains but is 100% plugin dispatch (no direct candle dep).

What’s not a plugin yet but probably should be

• vox-browser: chromiumoxide CDP wrapper, pulled into vox-cli unconditionally (no feature gate). No plugin scaffold exists.
• vox-dashboard: Axum SPA host with embedded assets.
• vox-publisher: social/news publishing adapters.
• vox-scientia-*: scholarly ingestion + publishing pipeline.
• vox-gamify: gamification.
• vox-audio-ingress: standalone binary that unconditionally pulls stt-candle from vox-oratio.

Slim-core diagnostic

Default cargo build -p vox-cli produces a 66 MB debug binary without ML deps. The feature-gating discipline works: candle/burn/wgpu/qlora do NOT compile in the default build. But:

• vox-cli enables vox-populi/transport unconditionally, pulling axum, tower-http, jwt, ed25519-dalek, blake3, dashmap, turso into every CLI build.
• vox-audio-ingress enables vox-oratio/stt-candle unconditionally, pulling the full Candle Whisper stack.
• vox-cli depends on vox-browser unconditionally (no feature gate).

True slim-core requires extracting these into plugins so even consumers’ default features don’t pull them in.

Five extraction units

Each unit is independently shippable. Sequence them or parallelize across worktrees.

Unit 1: vox-populi mens/tensor residual extraction (~7,852 LOC)

Scope: Move the remaining candle-using files from vox-populi/src/mens/tensor/ into vox-plugin-mens-candle-cuda:

• candle_model_qwen.rs (764 LOC) — Qwen3.5/Qwen2 transformer block, attention, MLP, RoPE, KV cache
• candle_inference_serve.rs (494 LOC) — inference server
• candle_qlora_merge.rs (389 LOC) — adapter merge logic
• burn_stack.rs (128 LOC) — Burn tensor stack
• Supporting files in mens/tensor/lora/, mens/tensor/ce_*, etc.

Approach: Add MlBackend::run_inference(model_handle, prompt_json) -> RString and MlBackend::merge_adapter(base_path, adapter_path, dest_path) -> () methods. Bump ABI 6→7. Move the implementations into the plugin behind those methods. Update vox-populi’s mens module to dispatch via plugin host (similar to SP3 sub-batch D for training).

Acceptance: vox-populi/src/mens/tensor/ no longer contains direct candle-core imports. The plugin’s integration tests cover load_model + run_inference + merge_adapter end-to-end.

Risk: Inference server is HTTP-bound (probably axum) — the plugin would need to spin up its own server or the host would need a serve_model callback. Decide architecture before code-moving.

Unit 2: vox-populi transport extraction (~4,990 LOC)

Scope: Move vox-populi/src/transport/ into vox-plugin-populi-mesh (currently a stub). Files: mod.rs, auth.rs, router.rs, mesh_replay.rs, result_attestation.rs, handlers/ (764 LOC), store/ (1,153 LOC). Plus http_client.rs, http_lifecycle.rs, node_registry.rs (~1,105 LOC).

Approach: The MeshDriver trait already exists from SP7 with method shapes (start_transport, stop_transport, dispatch, node_join, list_nodes). Real implementations replace the stub RErr("not yet implemented") returns. The plugin pulls the heavy deps (axum, tower-http, jwt, dashmap, turso). vox-cli drops vox-populi/transport from its default features.

Acceptance: vox-populi/src/transport/ deleted. vox-cli builds without axum in the dep graph (verify with cargo tree). Existing populi integration tests pass via the plugin path.

Risk: Heavy. The transport module is tightly coupled to vox-populi’s NodeRegistry, store backends (Turso schema), and JWT secret resolution (vox-secrets). Extraction may require the plugin to depend on vox-secrets + vox-db too — that’s accepted (the same pattern as mens-candle-cuda’s vox-tensor + vox-corpus deps).

Unit 3: vox-tensor extraction (~3,054 LOC)

Scope: Move all of vox-tensor/src/ into vox-plugin-tensor-burn-wgpu. Files: data.rs, lora.rs, optim.rs, vox_nn.rs, train.rs, plus tensor/{activations,ctor,elemwise,slice_reduce,cat_reshape}.rs.

Approach: TensorBackend trait exists from SP7. Real implementations replace stubs. vox-tensor either becomes a thin types-only crate or is deleted entirely.

Acceptance: vox-tensor is empty or deleted. Burn+wgpu no longer compile in default builds of any consumer.

Risk: vox-tensor is consumed by vox-populi mens (Unit 1’s territory) and possibly other crates. Sequence after Unit 1 so the consumers exist as plugin-host dispatches before vox-tensor’s body moves.

Unit 4: vox-oratio Whisper extraction (~2,275 LOC)

Scope: Move vox-oratio/src/backends/{candle_engine,candle_whisper,logit_processors,audio_io,multilingual}.rs into vox-plugin-oratio.

Approach: AudioCapture trait exists from SP7. Add a sibling trait or method group for speech-to-text (the current trait covers capture, not transcription). Possibly: split into AudioCapture (mic / device) and SpeechToText (Whisper). Bump ABI.

Acceptance: vox-oratio/src/backends/ no longer has candle-* deps. vox-audio-ingress switches to plugin dispatch and drops its stt-candle feature.

Risk: Audio pipelines are state-heavy (sample rates, frame buffers, streaming). The cdylib boundary needs to support streaming reads — RVec<u8> chunked reads work, but design carefully.

Unit 5: vox-browser extraction (~thin wrapper + chromiumoxide indirect deps)

Scope: Wrap vox-browser’s public API in a new vox-plugin-browser cdylib + a new BrowserAutomation extension-point trait (currently no scaffold).

Approach: Define BrowserAutomation trait in vox-plugin-api/src/extensions/browser_automation.rs. Bump ABI. Move chromiumoxide-using code into the plugin. Make vox-browser either a thin types-only crate or delete it.

Acceptance: vox-cli no longer transitively depends on chromiumoxide. CDP automation works via plugin dispatch.

Risk: Lower than 1-4 — vox-browser is smaller and has fewer consumers.

Sequencing recommendation

Unit 1 (mens residual)   ←── Unit 3 (tensor) depends on Unit 1's consumer migration
                              Unit 2 (transport) is independent — can run in parallel
Unit 4 (oratio whisper)  ←── Independent of mesh/tensor
Unit 5 (browser)         ←── Independent

Total estimated LOC to move: ~18,000+ across 5 units. Total estimated effort: 3–5 focused sessions (Unit 2 is the biggest).

Cross-cutting items not unit-specific

These accompany or follow Unit 1–5:

CC-1: Retire vox-skills shim entirely

ARS shim relocation — DONE (2026-05-03). Created crates/vox-openclaw-runtime as the ARS/OpenClaw facade (successor to the retired three-letter ARS crate name). All external consumers (vox-cli, vox-actor-runtime, vox-orchestrator) migrated to canonical OpenClaw runtime import paths. Physical move of the 12 ars_shim/ files out of vox-skills is deferred — ars_shim/mod.rs re-exports crate::parser::parse_skill_md, crate::SkillRegistry, and crate::install_builtins from the vox-skills crate root, creating a circular dependency if extracted. Those types need to move into vox-plugin-host first.

Full vox-skills deletion still deferred. Remaining blocker: parser.rs, registry.rs, plugin.rs consumed by vox-orchestrator’s plugin_skills_bridge — requires those types to be replicated or moved into vox-plugin-host before vox-skills can be deleted.

CC-2: Adopt AgentSkills open standard at SKILL.md level

The interop reviewer found that agentskills.io is the de facto universal format with 34+ adopters (Claude Code, Codex, Gemini CLI, Copilot, Cursor, JetBrains, etc.). Vox’s format is ~60% compatible already.

Concrete steps (small, ship as one commit):

• Add name field to each skill plugin’s SKILL.md frontmatter (alias from existing id; satisfy the spec’s lowercase-hyphen + match-dirname constraint).
• Move Vox extensions into a metadata.vox-* block when exporting.
• Document the convention in docs/src/reference/plugin-manifest.md.

CC-3: OpenClaw publish (export) path

vox-skills/src/ars_shim/openclaw.rs implements import. Add export: OpenClawClient::publish_skill() POSTs to /v1/skills. Unlocks publishing to ClawHub (44k+ community skills marketplace).

CC-4: AgentSkills export shim

vox skill export --format agentskills <id> flattens Plugin.toml + SKILL.md into a spec-compliant directory. Makes Vox skills publishable to any AgentSkills-compatible host.

CC-5: Honest deferred-status notes in existing plans

Already done in this branch — SP3, SP7, populi-mesh source comments, and parent spec all carry status notes acknowledging deferred work.

Acceptance for the whole follow-up

When all 5 units + CC-1 land:

• cargo build -p vox-cli default features: zero candle/burn/wgpu/chromiumoxide/axum/jwt in the dep graph
• vox-tensor, vox-skills deleted from workspace
• vox-populi shrinks to mesh-types + lifecycle (~few hundred LOC)
• All 7 SP7 + 1 SP3-residual + 1 mesh + 1 browser extension points have working plugin implementations
• CC-2 lands (AgentSkills compliance) → Vox skills work in 34+ external tools
• CC-3 + CC-4 land → bidirectional skill exchange with ClawHub and any AgentSkills-compatible host

The result: a genuinely slim Vox core + first-class membership in the cross-vendor agent-skill ecosystem.