feat: +ae context primitive - Agentic Entrepreneurship as first-class tool - #55872
Open
MYaelMendez wants to merge 51 commits into
Open
feat: +ae context primitive - Agentic Entrepreneurship as first-class tool#55872MYaelMendez wants to merge 51 commits into
MYaelMendez wants to merge 51 commits into
Conversation
…tool Adds ae_context tool to Hermes Agent core: - +æ activate: loads agentic entrepreneurship context bundle (language, stack, blueprints, NAICS, revenue proof, mission) - -æ deactivate: unloads context - search: live NAICS API classification (with static fallback) - blueprints: list available organizational blueprints - revenue: show real 33.65 revenue proof - mission: #opensourceware NousResearch#250 challenge - install: skill install command + public artifacts Prompt transforms: >_ → >_æ| when activated. This is the first blueprint-level tool — above skills, below core. Blueprints (#) are organizational patterns. Skills do tasks. $ executes within economic boundary (requires confirmation). $OS = economic infrastructure for agentic entrepreneurship. #startabusiness +æ
alt-glitch
added
invalid
…nd standalone computer_use primitive
…surfaces - Kill fakeInference stub; wire remoteUse.hermesNative to real WebLLM/WebGPU - Add remoteUse.factory: agentic CSS/HTML factory (spec -> local Qwen3-0.6B -> live preview -> write) - Add remoteUse.computerUse: bounded local desktop control surface - Local Python bridge (bridge.py + lib_gate.py) with gated app execution, scaffold app - Register factory/computerUse commands in package.json + subscriptions - Add hermes-native skill + llm-bench-rig design notes - Scoped .gitignore for build artifacts
- Full workbench chrome + editor + terminal + git decoration skin - Palette: void NousResearch#50505, gold #D4AF37, neon-mint #00ffa3, ink #e6e6e6 - 184 workbench colors, 19 tokenColor rules, semantic token colors - Registered as theme contribution in vscode-remote-use extension
- Render local runtime as an operator surface, not a CLI - 6 surfaces (Hermes Native, Factory, Computer Use, Media, DAO, home://) - Each INVOKE routes to remoteUse.* command via VS Code host postMessage - Mesh bus log + live clock; gold-on-void skin
- templates/surfaces/index.html: v4 dashboard hub mounting 8 surfaces - tabs (Surfaces/Inference/Factory/Computer/Media/Secret/Boot) + sidebar cards - file:// + remoteUse.* command dispatch, dock terminal, bridge probe - remoteUse.surfaces command; remoteUse.htmlSurface now opens the hub - gold-on-void design system; no parallel standalone pages
- victus-capabilities.html: probed machine map (CPU/GPU/CUDA/disk/toolchain/rust/surfaces) + client WebGPU/WebGL probe - hub: Machine tab + capability surface entry - extension: file:// -> asWebviewUri conversion so surfaces render inside VS Code webview (raw file:// is blocked)
…now present (verified) - victus-capabilities.html: docker is the only remaining gap; cl/link + wasm-pack 0.15 marked installed - victus-capability-map.md: rewrite from verified reality (BuildTools 2022 installed, not on PATH; wasm-pack 0.15.0; CUDA 13.3) - subagents completed: wasm-pack 0.15.0 installed, repo secret sweep CLEAN (1170 files), capability doc written
- nvidia-tiles.html: 6-tile sovereign GPU dashboard (>_n identity, green NVIDIA theme): GPU smi telemetry, compute path, processes, bench, routes, envelope; host-bridge wired (remoteUse.nvidia) - extension: remoteUse.nvidia command w/ nvidia-smi bridge (smi + procs + torch probe), Computation tab + surface entry in hub - pioneers: Hermes has no CUDA/NVIDIA skill; this is the first
First CUDA/NVIDIA skill for Hermes (none existed). Formalizes >_n: operator grammar: - SKILL.md: >_n: NemoClaw pattern, nvidia-smi host bridge, webview file:// pitfall, WebGPU fallback, verification checklist - references/host-bridge.md: full remoteUse.nvidia handler + message contract - templates/nvidia-tiles.html: shipped green-NVIDIA 6-tile surface - related_skills: sovereign-vscode-surface-dev, hermes-surface-development, hermes-operator-grammar, optimizing-attention-flash
…ile IR - description now scopes nvcc + CUDA Tile IR (not just telemetry) - Workflow 4: compile/run a real nvcc kernel; surface Bench tile should call it - Resources: link NVIDIA/cuda-tile as the compiler backend under the surface
…actor environments/host_loop.py — the sovereign local-loop primitive: planner -> executor -> auditor -> manifest (bounded, deterministic, offline). Local counterpart to environments/agent_loop.py (OpenAI-spec). Brain is a local model socket (ollama default); hands are local processes. Air-gappable. templates/surfaces/gpu_agent.py — GPUAgent now subclasses LocalHostLoop (single source of truth for the loop). CUDA tools unchanged: probe_gpu / compile_kernel (nvcc via vcvars64.bat) / run_kernel (exec on RTX). Writes a run manifest proving offline probe/compile/run. test_gpu_agent.py — covers the base loop (mock planner) + GPU hands + full offline loop (serial-only). 33 passed, 1 skipped (ollama) in repo gate.
environments/rust_wasm_tools.py — Rust/WASM tool-set for LocalHostLoop:
rust_build_wasm : cargo build --target wasm32-unknown-unknown (or wasm-pack)
rust_run_wasm : execute via wasmtime if present; else honest 'no_runtime'
capability report (Victus has no wasmtime on PATH)
Found + fixed a real toolchain trap: a standalone Rust install at
'C:\Program Files\Rust stable MSVC 1.96' shadows rustup on PATH, so wasm32
builds fail with 'can't find crate for std' despite the target being installed.
_rustup_env() prepends rustup's bin to PATH so cargo->rustc resolve to the
rustup shim (1.97.0 after force reinstall). _wasm_std_present() guards cleanly.
Fixture: environments/testdata/wasm_hello (minimal cdylib) — verifies the
build path produces a real .wasm (383 bytes). 4 wasm tests pass.
…tLoop
WasmAgent(LocalHostLoop): offline ollama brain + Rust/WASM hands
- rust_build_wasm : cargo build --target wasm32-unknown-unknown (real .wasm)
- rust_run_wasm : wasmtime if present; else honest 'no_runtime' report
- per-agent SYSTEM prompt lists the exact tool names + crate path so the
planner emits valid TOOL: calls (model was inventing 'cargo new' before)
- platform-portable: paths resolved via rustup, not hardcoded to Victus
Hands renamed _hand_build/_hand_run to avoid shadowing LocalHostLoop.run().
Test: 3 passed (build real .wasm, run degrades, full loop drives build).
…col (stdlib-only) environments/gpu_mcp.py — zero-dependency MCP server over stdio (JSON-RPC initialize/tools/list/tools/call, hand-rolled with stdlib so the sovereign stack stays air-gappable; no MCP SDK install needed). Tools (imported from the existing offline-verified hands, no logic dup): CUDA : probe_gpu, compile_kernel (nvcc+vcvars), run_kernel Rust : rust_build_wasm, rust_run_wasm (honest no_runtime if wasmtime absent) Any MCP client (Claude Desktop, Cursor, gemini-cli) can now drive the local RTX 3050 + wasm toolchain as tools. Verified by piping JSON-RPC over stdin: initialize/tools/list + live probe_gpu/compile_kernel/rust_build_wasm calls. test_gpu_mcp.py: 5 passed (real stdio JSON-RPC round-trip).
…GPU-MCP
+æ://cc home:// is now a real conductor scheme (longest-prefix outranks +æ://).
It routes to environments/gpu_mcp.py — the protocol-native control surface for
the Victus node: local CUDA (probe/compile/run) + Rust/WASM hands over MCP.
- hermes_cli/conductor.py: _cc_dispatch + register('+æ://cc', ...) -> mcp surface
(address mcp://gpu-mcp, launch 'python environments/gpu_mcp.py', node home://)
- gpu_mcp.py header documents canonical scheme +æ://cc home://
- test_conductor.py: +æ://cc routes to gpu-mcp; longer-prefix beats bare +æ://
Verified: _dispatch('+æ://cc home://') -> mcp://gpu-mcp surface. 13 conductor
tests pass; full gate 58 passed / 1 skipped.
The +æ^glocal offline brain now comes up automatically when vscode-remote-use
activates in your local VS Code:
- ensureBrain(): pings :11434; if down, spawns 'ollama serve' in a hidden
terminal so qwen2.5-coder:3b / gemma4 (on-disk weights) come online — no
network needed at generate time.
- remoteUse.brain command: check/raise the local brain on demand.
- activationEvents + command contribution added.
Agents (gpu_agent, wasm_agent, gpu_mcp) hit localhost:11434 only, so the
sovereign loop is air-gappable once weights are cached.
Compiled + reinstalled VSIX (211.84KB); out/extension.js verified to contain
ensureBrain() call + 'ollama serve' logic. tsc clean.
…MCP server
The extension is now the installable 'Hermes Agent' VS Code plugin (not
'vscode-remote-use'):
- displayName 'Hermes Agent' + keywords (hermes, cuda, wasm, mcp, offline,
sovereign, gpu, nvidia, glocal, ae) so a search for 'hermes' resolves.
- contributes.mcp.servers.gpu-mcp -> python ./gpu_mcp/gpu_mcp.py (stdio).
Fills the 'MCP Servers' slot: any agent client (Agentic.js, Claude Desktop,
Cursor) gains probe_gpu/compile_kernel/run_kernel/rust_build_wasm/
rust_run_wasm as local tools. This is the +æ://cc home:// control surface
exposed through the standard plugin channel.
- self-contained gpu_mcp/ bundle (host_loop + rust_wasm_tools + gpu_agent +
gpu_mcp) ships inside the extension so the MCP server runs standalone, no
hermes-fork repo required. Canonical source stays in repo; bundle = shipping copy.
- portable vcvars: gpu_agent._find_vcvars() searches common VS layouts +
HERMES_MSVC_VCVARS env instead of a hardcoded Victus path -> worldwide
x86_64-pc-windows compatible (PR-prep for PR NousResearch#2).
Verified: bundled gpu_mcp runs standalone (5 tools, live RTX 3050 probe);
VSIX 234KB reinstalled as hermes-agent@0.2.0 with mcp server in manifest.
Repo tests: 30 passed / 1 skipped.
… name Names the whole primitive: ae://glocal-agent (the canonical scheme for the +ae^glocal sovereign-local agent = offline brain + local CUDA/Rust-WASM hands, exposed as GPU-MCP control surface). Alias of +ae://cc home:// under one name. - conductor: register ae://glocal-agent -> gpu-mcp surface (mcp://gpu-mcp) - gpu_mcp.py header documents ae://glocal-agent as canonical scheme - test: ae://glocal-agent home:// -> mcp://gpu-mcp; short form -> home:// Verified: dispatch resolves; 20 conductor+MCP tests pass.
The viewport (local HTML/CSS/Rust-WASM surface) is the mandate; VS Code is
its host runtime, not the mandate itself.
- conductor: vscode:// relabeled viewport_host (v=viewport, compute=local)
- conductor: new viewport:// scheme = the mandate (local HTML/CSS/WASM
viewport control plane); alias of vscode:// as host
- SPEC.md / CAPABILITIES.md: declare v = viewport, not Visual Studio
- tests: viewport:// is the mandate; vscode:// is its host
vscode:// stays (Microsoft protocol, immutable); our conductor framing owns
the meaning. Verified: viewport:// + vscode:// both resolve v=viewport;
16 conductor tests pass.
… Agent viewport viewport://hermes-agent is the concrete instance of the viewport mandate: the ae://glocal-agent primitive (offline ollama brain + GPU-MCP control surface) rendered as the local gold-on-void HTML/CSS/WASM viewport (templates/surfaces/index.html). Other viewport:// nodes stay generic. Verified: resolves node=hermes-agent, agent=ae://glocal-agent, control_surface=mcp://gpu-mcp, brain=ollama://localhost:11434; 17 conductor tests pass.
…l RTX proof templates/startabusiness.html — a gold-on-void boot-sequence page (Orbitron + JetBrains Mono, neural mesh canvas, scanlines) for would-be agentic entrepreneurs. Follows the #startabusiness required sections (topbar, command block, 8-step execution, Stripe block, agent roster, boot log, stats, CTA, seal) but grounds the COMPUTE step in the ACTUAL stack we built this session: - ae://glocal-agent (canonical scheme for the sovereign local agent) - GPU-MCP (mcp://gpu-mcp) — CUDA probe/compile/run + Rust->wasm32 - viewport://hermes-agent — the rendered gold-on-void control surface - offline ollama brain (qwen2.5-coder, on-disk weights, ZERO network) - nvcc-matmul executed on the GeForce RTX silicon The thesis: a consumer GeForce RTX IS the business — no cloud bill, no API rent. CTA links to Doola Wyoming DAO LLC formation (affiliate). Neutral glyph placeholder (never defaults to ae, which is Yael's). Verified: well-formed HTML, all 9 required sections + Doola CTA present, real RTX proof wired into boot log + execution sequence.
…niverse
templates/sim.html — a one-file, zero-runtime sovereign agent simulation that
runs the ae://glocal-agent loop VISUALLY. This is what solo teams never got
from the Omniverse: a working simulation, not a spec to stand up a mesh first.
- defaults to SIM MODE: deterministic local animation of brain -> probe_gpu
-> compile_kernel (nvcc) -> run_kernel (silicon) -> rust_build_wasm ->
manifest, on a canvas-rendered RTX die with live VRAM/temp/util bars.
- upgrades to LIVE MODE automatically when ollama is detected at :11434
(honest SIM/LIVE badge; no fake data in live mode).
- exposes the 5 GPU-MCP tools as the mcp://gpu-mcp surface, with the
#startabusiness CTA as the exit ramp (own the runtime).
- no :5173 host, no bridge, no QR mesh — opens by double-click. The full
Omniverse mesh stays optional, not the on-ramp.
Verified: well-formed HTML, only cosmetic external deps (fonts + Doola),
all loop steps + 5 tools + SIM/LIVE toggle present.
…vider, reachy/vscode surfaces, landingpage, tests Consolidates 17 modified + 100 untracked source files: - hermes_cli/conductor.py: expanded scheme routing - hermes_cli/runtime_provider.py: runtime/provider resolution - hermes_cli/main.py: CLI entry growth - apps/reachy (app/viewport), vscode-remote-use extension - landingpage (index/script/style), templates/ae.html - tests: test_conductor, runtime_provider_resolution - blueprints, mech, qc64_basic, secret_source_bridge, neuromitosis Excludes vendored third-party trees (tools/vlc, llm-bench-rig, media, movies) via updated .gitignore.
- Promote first-party GPU-MCP into its own public repo (MYaelMendez/gpu-mcp) and consume it as a git submodule; delete the two stale in-tree copies (environments/gpu_mcp.py and the diverged vscode-remote-use/gpu_mcp bundle) that had drifted from canonical. - conductor: route +ae://cc, ae://glocal-agent, computer:// launches to 'python -m gpu_mcp' (module form) instead of the deleted script path. - vscode-remote-use/package.json: MCP server now launches via 'python -m gpu_mcp' (installed from the submodule) rather than a bundled file. - New scheme +ae://glocal cloud computer: the hybrid sovereign compute surface. HANDS are ALWAYS local (gpu-mcp, sovereign CUDA/Rust-WASM). BRAIN defaults local; cloud (Nous Portal) is opt-in only via an explicit token (portal/nous/ 2nd 'cloud'). Brain/hands separation — global when you opt in, local by default. - tests: rewire assertions + 2 new hybrid-surface tests (local-default, opt-in portal). Full conductor suite: 24 passed.
…heme Previous commit 7936edc captured only the submodule gitlink; the working changes (conductor scheme, package.json, deleted stale copies, test updates) were unstaged and missed. Restating the real change set: - conductor: +ae://glocal cloud computer hybrid surface; +ae://cc, ae://glocal-agent, computer:// now launch 'python -m gpu_mcp'. - delete stale environments/gpu_mcp.py + vscode-remote-use/gpu_mcp bundle. - vscode-remote-use/package.json: MCP via 'python -m gpu_mcp'. - tests: rewire + 2 hybrid-surface tests (24 passing).
Adds a QR-based peer-pairing primitive so two sovereign nodes (e.g. Victus + Legion on the same wifi) can link without cloud IDs or LAN port scanning: - +ae://mesh offer <name> -> emits a QR PNG encoding a peer manifest (ae://peer?host=<name>&mesh=pc://mesh/<name>/local&port=<lan>&token=<ephemeral>&via=wifi). New node emits; sovereign node scans + accepts. No secrets in the code. - +ae://mesh accept <pay> -> registers pc://mesh/<name>/local as a live, addressable route. Explicit opt-in only; bad payload rejected gracefully. - Peers persist locally (mesh_peers.json, gitignored) — no cloud. - Reuses existing +ae://qrcode image writer; registers per-name route (pc://mesh/<name>/) so it never shadows pc://mesh/victus/... nodes. - test: full offer->accept->live-route round trip + bad-payload rejection. Conductor suite: 25 passed.
Adds the fleet registry primitive to the chassis, backing the github-pages fleet.html bootstrap surface. Local-default, opt-in, no cloud. - +ae://fleet offer <name> [--models m1,m2] emit capability QR (ae://fleet manifest) - +ae://fleet join <manifest> --node <n> --models <csv> register device into MoD - +ae://fleet list show current fleet (MoD) - _FLEET_REGISTRY persisted locally (fleet_registry.json), mirrors _PEER_REGISTRY Sets up the next layer: +ae://route (MoD x MoA x MoM dispatch).
Prior rewire (8f897c2) deleted environments/gpu_mcp.py and moved the server into the gpu-mcp submodule (runs as 'python -m gpu_mcp'), but the test still referenced the deleted script path and invoked it as a file — breaking the relative import (ImportError -> no stdout -> out[1] IndexError). Point the test at the submodule and invoke it as a module. 5 gpu-mcp tests now pass; 25 conductor tests unaffected. Total 30.
Local-only media pipe: ffmpeg encodes a source (testsrc | gpu_mcp viz) on the RTX via h264_nvenc and listens on tcp://localhost:17344?listen; VLC connects as client (vlc://play tcp://localhost:17344). No cloud, no rent. - _cuda_vlc_dispatch (play <src>|status|stop) + process lifecycle (lock-guarded) - registers +ae://cuda-vlc in SchemeDispatcher - test_cuda_vlc_surface_routing (hermetic status/stop; live path ad-hoc) - templates/cuda-vlc.html agentic GUI (drives the scheme, NVENC telemetry) Verified end-to-end: ffmpeg(NVENC)+VLC both run, VLC connected to the tcp stream, stop tears down. 31 passed.
- agents/vps_node.py: stdlib HTTP node holding ae.core records (signed), answers rtx3050:// compute via gpu-mcp's real surface (compile+run kernel), serves the did:web PDS endpoint. --selftest covers sign/verify+store+route. - conductor: +ae://vps scheme routes status|record|route to the node. - fixed urlparse netloc-vs-path action bug; dispatch drives GPUAgent directly. Verified live: +ae://vps route rtx://rtx ran matmul on RTX 3050 (526ms). #hermiphicationisinevitable
Closes the GLOCAL secret handoff gap. github.io hosts the bridge SURFACE; the SECRET lives only on Victus (C:\ae\secrets\secrets.json, git-ignored). This scheme reads that local file and feeds agents at runtime — never uploads, never echoes raw values to stdout, never touches cloud. +æ://secrets status -> file present? count? (no values) +æ://secrets list -> keys+kinds+masked preview +æ://secrets get <K> -> resolve (masked in stdout; value carried for agent only) +æ://secrets path -> where the local source lives Uses shared agents/secret_source. Verified 9/9 ad-hoc (no raw leak, get resolves). Replaces the broken 'upload secrets to cloud' handoff. #hermiphicationisinevitable
Polish per user: 'make glocal secrets a skill, let's polish it'. - sources: list candidate paths + which exist (debug a failed handoff) - docstring documents status|list|get|sources|path Verified 9/9 ad-hoc (digital_ocean resolves locally, masked; sources flags real file). #hermiphicationisinevitable
teknium1
reviewed
Jul 15, 2026
teknium1
left a comment
teknium1
left a comment
Contributor
There was a problem hiding this comment.
Thanks for the entrepreneurship-context proposal. The current implementation needs substantial re-scoping before it can be evaluated as a Hermes contribution.
Problems
tools/ae_context_tool.py:95-122stores activation in module-global_AE_ACTIVE; this is shared by all sessions in a gateway process rather than being bound to one conversation.tools/ae_context_tool.py:103-126only toggles that flag and returns JSON, while the schema attools/ae_context_tool.py:220-228claims prompt transformation and enforced execution boundaries. No prompt-builder or approval integration is present in this implementation.toolsets.py:67-68adds this niche feature to_HERMES_CORE_TOOLS. Currenttoolsets.py:430-568propagates core tools across default surfaces, andtools/tool_search.py:10-11makes them non-deferred.crates/hermes-conductor/target/CACHEDIR.TAG:1and relatedtarget/paths are generated build artifacts.
Suggested changes
- Salvage the entrepreneurship material as an opt-in skill or external context-engine/plugin, using the session lifecycle seam in
agent/agent_init.py:1755-1969. - Split unrelated changes and remove generated artifacts; then add session-isolation and user-visible behavior tests.
Automated hermes-sweeper review.
|
|
||
| # ── Activation / deactivation ────────────────────────────────── | ||
|
|
||
| _AE_ACTIVE = False |
Contributor
There was a problem hiding this comment.
_AE_ACTIVE is process-global, so gateway sessions share it. Activation/deactivation must be session-bound; otherwise one user's call changes another user's reported context state.
| "+æ — Agentic Entrepreneurship context primitive. " | ||
| "Load agentic entrepreneurship context (+æ), search NAICS business classifications (live API), " | ||
| "discover blueprints, view revenue proof, and check mission status. " | ||
| "When activated, the prompt becomes >_æ| and # (blueprint) / $ (execution) boundaries are enforced. " |
Contributor
There was a problem hiding this comment.
The implementation only toggles _AE_ACTIVE and returns a tool result; it does not modify the cached system prompt or install an execution/approval guard. Please do not claim this boundary is enforced until it is wired through a session-safe supported mechanism.
| @@ -64,6 +64,8 @@ | |||
| "honcho_context", "honcho_profile", "honcho_search", "honcho_conclude", | |||
| # Home Assistant smart home control (gated on HASS_TOKEN via check_fn) | |||
| "ha_list_entities", "ha_get_state", "ha_list_services", "ha_call_service", | |||
| # +æ — Agentic Entrepreneurship context primitive | |||
| "ae_context", | |||
Contributor
There was a problem hiding this comment.
Adding this to _HERMES_CORE_TOOLS exposes the schema on every default CLI, cron, and messaging session, and core tools cannot be deferred. This niche content needs an opt-in skill, plugin, context engine, or MCP surface instead.
| @@ -0,0 +1,3 @@ | |||
| Signature: 8a477f597d28d172789f06886806bc55 | |||
Contributor
There was a problem hiding this comment.
This is Cargo-generated cache metadata. Remove target/ build outputs from the PR; source and reproducible build configuration belong in version control, not local artifacts.
teknium1
added
sweeper:risk-session-state
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
3 participants
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
+ae - Agentic Entrepreneurship Context Primitive
Adds
+aeas a first-class context tool in Hermes Agent. Not a skill, not a plugin - a context layer that loads agentic entrepreneurship context and transforms the prompt from>_to>_ae|.Actions
Files
Verification
Proposal
https://myaelmendez.github.io/PRIMITIVE.md
Built for Hermes Agent Accelerated Business Hackathon (Nous x NVIDIA x Stripe).
$OS = economic infrastructure for agentic entrepreneurship.
#startabusiness +ae