total-recall-brain 3.33.0 → 3.34.0

package/plugins/skill-manager/plugin.json

@@ -0,0 +1,24 @@
+{
+  "id": "skill-manager",
+  "name": "Skill Manager",
+  "version": "1.0.0",
+  "description": "Conservative skill optimization on an explicitly selected daemon node and authorized roots.",
+  "license": "MIT",
+  "use_cases": ["skills", "automation"],
+  "cli": {
+    "command": "skill-manager",
+    "handler": "./cli.mjs",
+    "subcommands": [
+      { "name": "configure", "description": "Select a mesh node and authorized skill roots from a JSON configuration" },
+      { "name": "status", "description": "Show selection, configuration and last report" },
+      { "name": "audit", "description": "Inspect changed skill packages without applying changes" },
+      { "name": "optimize", "description": "Run configured conservative optimization on the selected node" }
+    ]
+  },
+  "tasks": [{
+    "intent": "Inspect changed skills and apply only authorized lossless optimization",
+    "schedule": "0 3 * * *",
+    "command": "optimize",
+    "placement": "selected-node"
+  }]
+}

package/scaffold/.agent/skills/decision/SKILL.md

@@ -0,0 +1,20 @@
+---
+name: decision
+description: Design and use typed decisions for routing, classification, ranking and verification; configure Jev through the existing decision capability.
+repo_scoped: false
+---
+
+# Decision
+
+Use one decision capability for closed choices, yes/no probabilities and ordered scores. Jev is a configurable engine; use a generative model when the task needs prose. Typed output still needs validation and domain evaluation.
+
+Discover the current repository's decision plugin, configuration, credential resolver and local overlay. Confirm installed commands before invoking them. Resolve model, endpoint, timeout, budget and gates from config or flags; keep credentials in the existing secret store. Public skills contain portable methods, not repository or account instances.
+
+Validate typed answers before branching. Choice/Score confidence differs from Noul's probability of yes; never default missing certainty to success. Calibrate outcome gates using representative cases. Missing config, invalid output, timeout or provider failure takes the deterministic fallback. Required instructions, scope and permissions stay enforced in code; model advice cannot override them.
+
+Read only the reference needed now:
+
+- [Integration and typed response contract](references/integration.md)
+- [Skill routing and use cases](references/use-cases.md)
+
+Keep shared state small; batch independent questions and compute arithmetic in code. Test synthetic success, abstention and failure cases before integration. Live provider evidence is separate from mocks. Record model/version, decision and measured outcome without secrets or unnecessary state. Preserve existing human approval requirements.

package/scaffold/.agent/skills/decision/references/integration.md

@@ -0,0 +1,25 @@
+# Decision integration
+
+Verified against primary documentation on 2026-10-01. Recheck alpha API contracts before implementation.
+
+## Discover and configure
+
+Use the installed decision plugin's manifest/help to discover supported commands. Its skill is the same `decision` capability; do not install an additional Jev skill. Plugin presence or documentation alone does not prove live dispatch or persistence. Validate the adapter before relying on it.
+
+Reuse the repository's credential and provider configuration. Prefer schema-validated portable defaults and repo-owned config/overlays. Use the existing layered skill contract where supported: plugin-owned `core/config.schema.json`, canonical repo-owned SSSS `skill_config`, disposable `config.json` projection. Do not invent those interfaces where no consumer exists. No personal paths/hosts, credentials or injected memory belong in shipped defaults.
+
+## Jev boundary
+
+OpenRouter documents `POST /api/alpha/decisions` for plain HTTP/OpenRouter SDK calls and `POST /api/v1/systemone` for the TypeSafe SDK format. Confirm the chosen surface's request schema; do not interchange formats blindly. Select the model in configuration and record the actual response version. The separate Jev Router selects chat models/reasoning effort; it does not implement skill selection.
+
+Use Choice for a bounded label set, Noul for probability of yes and Score for an ordered probability-weighted result. Choice and Score expose distributions and confidence; Noul does not expose a separate confidence value. Validate finite ranges, membership and required answer fields; missing confidence never becomes 1. A weighted Score is not a discrete level. Pin the actual adapter contract and fixture responses before wiring gates.
+
+## OpenAI Decisions preview
+
+OpenAI announced a Luna-powered Decisions API on 2026-09-29 for finite predefined answers from text or image context. It is in limited preview; the public endpoint, schema and this account's access are unverified. OpenRouter's `/api/alpha/decisions` currently documents Jev, and its public model catalog did not list OpenAI Decisions on 2026-10-01. Keep OpenAI as a configurable future engine in this one `decision` skill, without guessing a model slug or sending requests through the Jev endpoint. Use the deterministic path until the official contract and actual access are verified. [OpenAI announcement](https://openai.com/index/devday-2026-recap/) · [OpenRouter Jev contract](https://openrouter.ai/docs/guides/community/jev).
+
+## Evidence and fallback
+
+Tune gates with representative labelled cases, including ambiguous input, no matching option, label wording changes and incomplete state. Report false selections, missed selections, abstention, latency and cost. Typed validity does not establish semantic accuracy. Cache by task, catalog/content hashes and configuration/model version; invalidate on relevant changes. Network/configuration/parse failures preserve the deterministic path. Avoid unneeded outbound private data; log no secrets.
+
+Sources: [OpenRouter Jev guide](https://openrouter.ai/docs/guides/community/jev), [TypeSafe confidence](https://docs.typesafe.ai/confidence), [Choice](https://docs.typesafe.ai/primitives/choice), [Score](https://docs.typesafe.ai/primitives/score), [Noul](https://docs.typesafe.ai/primitives/noul).

package/scaffold/.agent/skills/decision/references/use-cases.md

@@ -0,0 +1,20 @@
+# Decision use cases
+
+Primary docs checked 2026-10-01. Recipes demonstrate designs; they are not verification of this repository's runtime.
+
+## Skill selection
+
+The official recipe uses two passes: rank the catalog and determine whether a skill is needed, then verify a shortlist with richer descriptions and brief instruction excerpts. It can return no skill. Treat the selected name as advice. Keep full manuals out of the main context until selected, and preserve required instructions independently. Ownership/scope filtering happens in code before selection.
+
+Evaluate against real and synthetic requests from the target catalog. Compare deterministic routing with assisted routing; measure wrong loads, needless loads and missed loads as well as latency/cost. The cookbook's published experiment used Jev 1.12 and was rendered 2026-07-31; its thresholds and results are not current-model guarantees. Do not copy its gates without calibration.
+
+[Skill suggestion recipe](https://docs.typesafe.ai/cookbooks/skill_suggestion)
+
+## Other applicable patterns
+
+- Intent routing and triage: choose a handler and ask independent urgency/status questions on shared state. [Intent routing](https://docs.typesafe.ai/patterns/intent-routing)
+- Large catalogs: hierarchical classification narrows categories before leaf selection. [Hierarchical classification](https://docs.typesafe.ai/cookbooks/hierarchical_classification)
+- Retrieval: score candidate passages, then filter in code; preserve source provenance. [RAG passage classification](https://docs.typesafe.ai/cookbooks/classifying_rag_passages)
+- Batch decisions: fan out independent questions about one state and compose the result in code. [Fan-out](https://docs.typesafe.ai/patterns/fan-out)
+
+Candidate matching, deduplication, citation checks and tool dispatch require their own validated decision definitions and tests. Use deterministic code for arithmetic, identifiers and access control. Periodic skill optimization can use advice to identify candidates, but it must preserve source, respect ownership, check drift and require review for semantic rewrites.

package/scaffold/.agent/skills/meta-harness/SKILL.md

@@ -1,113 +1,15 @@
 ---
 name: meta-harness
-description: Meta Harness & Agent Management Layer to orchestrate and delegate tasks across all connected IDE harnesses (Antigravity, Claude Code, Codex, Gemini, Ollama) and the computer generally.
+description: Inspect harnesses, agent processes and mesh execution when authorized.
+repo_scoped: false
 ---
 
-# Meta Harness & Agent Management Layer
+<!-- total-recall:skill-router:v1 -->
 
-The **Meta Harness** provides a unified agent management layer across all connected developer environments, CLI runtimes, and the computer generally.
+# Harness management
 
-## 🎯 Architectural Concept
+Resolve installed commands, authenticated accounts, permissions and recorded mesh access before execution. Dispatch or spawn agents only when authorized and supported by verified adapters. Read the selected CLI/process/mesh procedure; do not load every harness example.
 
-Total Recall connects to every major IDE, CLI harness, and local neural runtime:
-- **Google Antigravity (`agy`)**: Frontier multi-step reasoning, web grounding, and deep research on AI Ultra.
-- **Claude Code (`claude`)**: Direct terminal codebase refactoring, bash tool execution, and complex Unix tasks.
-- **OpenAI Codex (`codex`)**: Program synthesis, sandboxed workspace mutations, and isolated code generation.
-- **Google Gemini CLI (`gemini`)**: Fast utility completions and lightweight tool chaining.
-- **Ollama Local LLM (`ollama`)**: Zero-token-cost local neural reasoning (e.g. `gemma4:latest`) running non-interactively via standard input piping (`pipe_stdin`).
-- **IDE Surfaces**: Cursor (`.cursorrules`), Cline (`.clinerules`), VS Code (`copilot-instructions.md`), DeepSeek Harness (`dsh`).
+Track owned processes, await work, inspect real results and clean up. A catalog entry is not proof of an available integration. Use validated memory operations and preserve repository boundaries.
 
-### The Shared Brain Principle
-Because all harnesses operate over the **same portable brain**:
-- Every harness reads the same auto-compiled instruction surface (`AGENTS.md`, `GEMINI.md`, `CLAUDE.md`).
-- Every harness writes directly into the shared SSSS memory vault (`memory-vault/`).
-- Knowledge or solutions learned by one agent in one harness are immediately accessible to every other harness across the computer and mesh.
-
----
-
-## 🛠️ CLI Operations
-
-### 1. Inspect Active Harnesses
-Inspect all available developer harnesses and their local binary paths:
-```bash
-npx total-recall harness list
-```
-
-### 2. Headless Task Dispatch
-Dispatch a specific coding or research task to any external harness (locally or remotely over the mesh):
-```bash
-# Local dispatches
-npx total-recall harness dispatch claude "Inspect src/core/surface.mjs and run lint checks"
-npx total-recall harness dispatch agy "Analyze latest preprints on room temperature diamond NV registers"
-npx total-recall harness dispatch codex "Generate unit tests for src/core/context-cache.mjs"
-npx total-recall harness dispatch ollama "Summarize the core invariants of SSSS"
-
-# Cross-mesh remote dispatches
-npx total-recall harness dispatch ollama --node build-box "What is test-time compute scaling?"
-npx total-recall harness dispatch claude --node build-box "Run git status and lint tests"
-```
-
-### 3. Multi-Harness Council
-Run a task across multiple harnesses concurrently to obtain comparative solutions and consensus:
-```bash
-npx total-recall harness council "Propose architecture for decentralized research mesh"
-```
-
----
-
-## 🤖 Agent Process Management
-
-Manage background subagent tasks across the local machine and remote mesh nodes:
-
-### 1. Spawn a Background Agent
-```bash
-# Local background agent
-npx total-recall agent spawn claude "Refactor error handling in src/cli/harness.mjs" --name "Refactor Worker"
-
-# Remote background agent on a mesh node
-npx total-recall agent spawn agy --node build-box "Crawl recent preprints on quantum shuttling" --name "Quantum Scout"
-```
-
-### 2. Inspect Running Agents
-```bash
-npx total-recall agent list
-npx total-recall agent list --json
-```
-
-### 3. Stream Agent Logs
-```bash
-npx total-recall agent logs <agent-id> --tail 100
-```
-
-### 4. Terminate an Agent
-```bash
-npx total-recall agent kill <agent-id>
-```
-
----
-
-## 🌐 Headscale Mesh Diagnostics & Remote Execution
-
-The Meta Harness integrates directly with your private Headscale WireGuard mesh network:
-
-### 1. Cluster Capability Audit (`mesh doctor`)
-Concurrently audit reachability, developer runtimes (`node`, `docker`, `git`), and AI harnesses (`agy`, `claude`, `codex`, `gemini`, `ollama`) across all cluster nodes:
-```bash
-npx total-recall mesh doctor
-npx total-recall mesh doctor --json
-```
-
-### 2. Non-Interactive Remote Command Execution (`mesh exec`)
-Execute commands on remote nodes with automatic `$PATH` sanitization and structured outputs:
-```bash
-npx total-recall mesh exec build-box uname -m
-npx total-recall mesh exec cloud docker ps
-```
-
----
-
-## 💻 General Computer & OS Management
-The Meta Harness controls the computer generally:
-- **Activity Monitoring ("Follow the User")**: Tracks interaction timestamps and routes dispatches to the active device across the Headscale WireGuard mesh.
-- **Process Orchestration**: Spawns background worker loops, monitors task completion, and terminates runaway processes.
-- **Surface Sync**: Recompiles instruction surfaces across all IDE directories automatically upon graph updates.
+Before an operation, read its [task references](references/optimized/index.md), including prerequisites and constraints. If the applicable procedure is unclear, read the preserved original instead of guessing. Commands assume the skill root.

package/scaffold/.agent/skills/meta-harness/references/optimized/index.md

@@ -0,0 +1,20 @@
+# Task references
+
+Read only sections needed for the operation, including their prerequisites and constraints. Commands and plain paths assume the skill root. This index selects documentation, not permission or rule applicability.
+
+- [Overview](section-01.md)
+- [🎯 Architectural Concept](section-02.md)
+- [The Shared Brain Principle](section-03.md)
+- [🛠️ CLI Operations](section-04.md)
+- [1. Inspect Active Harnesses](section-05.md)
+- [2. Headless Task Dispatch](section-06.md)
+- [3. Multi-Harness Council](section-07.md)
+- [🤖 Agent Process Management](section-08.md)
+- [1. Spawn a Background Agent](section-09.md)
+- [2. Inspect Running Agents](section-10.md)
+- [3. Stream Agent Logs](section-11.md)
+- [4. Terminate an Agent](section-12.md)
+- [🌐 Headscale Mesh Diagnostics & Remote Execution](section-13.md)
+- [1. Cluster Capability Audit (`mesh doctor`)](section-14.md)
+- [2. Non-Interactive Remote Command Execution (`mesh exec`)](section-15.md)
+- [💻 General Computer & OS Management](section-16.md)

package/scaffold/.agent/skills/meta-harness/references/optimized/manifest.json

@@ -0,0 +1,79 @@
+{
+  "version": 1,
+  "source_sha256": "511f8c11043679b032b85c4a6ceca921d8ac17fc59be1cdb87572346e7e425b7",
+  "entrypoint_sha256": "56abd1e41c0c93292c1ea5dc02d9eed280ec0b68e51c3a34e2ab060c66ba5a2f",
+  "sections": [
+    {
+      "path": "references/optimized/section-01.md",
+      "sha256": "3d0dd724287d7de88d1bacd069f1eb2925b1fe2f128a15ee0d8baaf605f61e79"
+    },
+    {
+      "path": "references/optimized/section-02.md",
+      "sha256": "febe972174b649547733fb8c22fd6eee42b61f0834a0d18e48794c928c2381e8"
+    },
+    {
+      "path": "references/optimized/section-03.md",
+      "sha256": "d11f85b987516543e03b8ad6ac5752ed1efbfcd0a0aa9dacecaeb9c31f91501f"
+    },
+    {
+      "path": "references/optimized/section-04.md",
+      "sha256": "b0781872c9046f04a7beed3cd12e055bf82db184a8e9fb4e64ecc9cc494b7d01"
+    },
+    {
+      "path": "references/optimized/section-05.md",
+      "sha256": "4b7acbdf1f56ceb2aa9d77677e2c6ddd18407c115a67e62dbcd7d4f4ebe0c674"
+    },
+    {
+      "path": "references/optimized/section-06.md",
+      "sha256": "d5581f9390c9f1164884166b35dd1a9a59b3b23c9a03c351b1382f36282e8aeb"
+    },
+    {
+      "path": "references/optimized/section-07.md",
+      "sha256": "a430fe3f87eb1e99fe33570202a6e7e016faef343f7d89ee901bbc31e13c1868"
+    },
+    {
+      "path": "references/optimized/section-08.md",
+      "sha256": "c0ac4ceacdcfce39a72c9c4cb0ae41d466ab94d9e340c034a229e3686c14f90c"
+    },
+    {
+      "path": "references/optimized/section-09.md",
+      "sha256": "62d5a35716e489d5164e1bd6f541b92fe159a4a181ee99e1cd1acdf1dec53262"
+    },
+    {
+      "path": "references/optimized/section-10.md",
+      "sha256": "5ed4c289daae2a5c04f3d2e8d96715403ce0988a56feb89e73d7f83ff7c56473"
+    },
+    {
+      "path": "references/optimized/section-11.md",
+      "sha256": "f46d6aaabb132430bfe52e20e48ab7d4f9bb20e4548adf7b8ab2f3a122a9b382"
+    },
+    {
+      "path": "references/optimized/section-12.md",
+      "sha256": "ea3166d7ed898c766dc4f9fcbccbc35e43a0c63d65a081f4c81e95b4f0af6c0b"
+    },
+    {
+      "path": "references/optimized/section-13.md",
+      "sha256": "8f5b2410330d1773a8a801b91e6b66d471adaa9609a06720e27b5feb3e4b272c"
+    },
+    {
+      "path": "references/optimized/section-14.md",
+      "sha256": "92d093b8c5583b2eff82ffaf73cec094b4d087791427be3aa2382222db4c0605"
+    },
+    {
+      "path": "references/optimized/section-15.md",
+      "sha256": "24b6f085e0fa09dc25ef5e9d5b9b49d21adbe8c58ea7a835a38d3d9ac7272760"
+    },
+    {
+      "path": "references/optimized/section-16.md",
+      "sha256": "264af7d981ecbbdf060dd0d087eaca0bb15a17210ae4882d07fdbb45dbaddab3"
+    },
+    {
+      "path": "references/optimized/index.md",
+      "sha256": "ce64a33aed9e5987e917c83945ff7d54f14aed8a8c079142b053fddfac3e960b"
+    },
+    {
+      "path": "references/optimized/original.md",
+      "sha256": "511f8c11043679b032b85c4a6ceca921d8ac17fc59be1cdb87572346e7e425b7"
+    }
+  ]
+}

package/scaffold/.agent/skills/meta-harness/references/optimized/original.md

@@ -0,0 +1,113 @@
+---
+name: meta-harness
+description: Meta Harness & Agent Management Layer to orchestrate and delegate tasks across all connected IDE harnesses (Antigravity, Claude Code, Codex, Gemini, Ollama) and the computer generally.
+---
+
+# Meta Harness & Agent Management Layer
+
+The **Meta Harness** provides a unified agent management layer across all connected developer environments, CLI runtimes, and the computer generally.
+
+## 🎯 Architectural Concept
+
+Total Recall connects to every major IDE, CLI harness, and local neural runtime:
+- **Google Antigravity (`agy`)**: Frontier multi-step reasoning, web grounding, and deep research on AI Ultra.
+- **Claude Code (`claude`)**: Direct terminal codebase refactoring, bash tool execution, and complex Unix tasks.
+- **OpenAI Codex (`codex`)**: Program synthesis, sandboxed workspace mutations, and isolated code generation.
+- **Google Gemini CLI (`gemini`)**: Fast utility completions and lightweight tool chaining.
+- **Ollama Local LLM (`ollama`)**: Zero-token-cost local neural reasoning (e.g. `gemma4:latest`) running non-interactively via standard input piping (`pipe_stdin`).
+- **IDE Surfaces**: Cursor (`.cursorrules`), Cline (`.clinerules`), VS Code (`copilot-instructions.md`), DeepSeek Harness (`dsh`).
+
+### The Shared Brain Principle
+Because all harnesses operate over the **same portable brain**:
+- Every harness reads the same auto-compiled instruction surface (`AGENTS.md`, `GEMINI.md`, `CLAUDE.md`).
+- Every harness writes directly into the shared SSSS memory vault (`memory-vault/`).
+- Knowledge or solutions learned by one agent in one harness are immediately accessible to every other harness across the computer and mesh.
+
+---
+
+## 🛠️ CLI Operations
+
+### 1. Inspect Active Harnesses
+Inspect all available developer harnesses and their local binary paths:
+```bash
+npx total-recall harness list
+```
+
+### 2. Headless Task Dispatch
+Dispatch a specific coding or research task to any external harness (locally or remotely over the mesh):
+```bash
+# Local dispatches
+npx total-recall harness dispatch claude "Inspect src/core/surface.mjs and run lint checks"
+npx total-recall harness dispatch agy "Analyze latest preprints on room temperature diamond NV registers"
+npx total-recall harness dispatch codex "Generate unit tests for src/core/context-cache.mjs"
+npx total-recall harness dispatch ollama "Summarize the core invariants of SSSS"
+
+# Cross-mesh remote dispatches
+npx total-recall harness dispatch ollama --node build-box "What is test-time compute scaling?"
+npx total-recall harness dispatch claude --node build-box "Run git status and lint tests"
+```
+
+### 3. Multi-Harness Council
+Run a task across multiple harnesses concurrently to obtain comparative solutions and consensus:
+```bash
+npx total-recall harness council "Propose architecture for decentralized research mesh"
+```
+
+---
+
+## 🤖 Agent Process Management
+
+Manage background subagent tasks across the local machine and remote mesh nodes:
+
+### 1. Spawn a Background Agent
+```bash
+# Local background agent
+npx total-recall agent spawn claude "Refactor error handling in src/cli/harness.mjs" --name "Refactor Worker"
+
+# Remote background agent on a mesh node
+npx total-recall agent spawn agy --node build-box "Crawl recent preprints on quantum shuttling" --name "Quantum Scout"
+```
+
+### 2. Inspect Running Agents
+```bash
+npx total-recall agent list
+npx total-recall agent list --json
+```
+
+### 3. Stream Agent Logs
+```bash
+npx total-recall agent logs <agent-id> --tail 100
+```
+
+### 4. Terminate an Agent
+```bash
+npx total-recall agent kill <agent-id>
+```
+
+---
+
+## 🌐 Headscale Mesh Diagnostics & Remote Execution
+
+The Meta Harness integrates directly with your private Headscale WireGuard mesh network:
+
+### 1. Cluster Capability Audit (`mesh doctor`)
+Concurrently audit reachability, developer runtimes (`node`, `docker`, `git`), and AI harnesses (`agy`, `claude`, `codex`, `gemini`, `ollama`) across all cluster nodes:
+```bash
+npx total-recall mesh doctor
+npx total-recall mesh doctor --json
+```
+
+### 2. Non-Interactive Remote Command Execution (`mesh exec`)
+Execute commands on remote nodes with automatic `$PATH` sanitization and structured outputs:
+```bash
+npx total-recall mesh exec build-box uname -m
+npx total-recall mesh exec cloud docker ps
+```
+
+---
+
+## 💻 General Computer & OS Management
+The Meta Harness controls the computer generally:
+- **Activity Monitoring ("Follow the User")**: Tracks interaction timestamps and routes dispatches to the active device across the Headscale WireGuard mesh.
+- **Process Orchestration**: Spawns background worker loops, monitors task completion, and terminates runaway processes.
+- **Surface Sync**: Recompiles instruction surfaces across all IDE directories automatically upon graph updates.

package/scaffold/.agent/skills/meta-harness/references/optimized/section-01.md

@@ -0,0 +1,5 @@
+
+# Meta Harness & Agent Management Layer
+
+The **Meta Harness** provides a unified agent management layer across all connected developer environments, CLI runtimes, and the computer generally.
+

package/scaffold/.agent/skills/meta-harness/references/optimized/section-02.md

@@ -0,0 +1,10 @@
+## 🎯 Architectural Concept
+
+Total Recall connects to every major IDE, CLI harness, and local neural runtime:
+- **Google Antigravity (`agy`)**: Frontier multi-step reasoning, web grounding, and deep research on AI Ultra.
+- **Claude Code (`claude`)**: Direct terminal codebase refactoring, bash tool execution, and complex Unix tasks.
+- **OpenAI Codex (`codex`)**: Program synthesis, sandboxed workspace mutations, and isolated code generation.
+- **Google Gemini CLI (`gemini`)**: Fast utility completions and lightweight tool chaining.
+- **Ollama Local LLM (`ollama`)**: Zero-token-cost local neural reasoning (e.g. `gemma4:latest`) running non-interactively via standard input piping (`pipe_stdin`).
+- **IDE Surfaces**: Cursor (`.cursorrules`), Cline (`.clinerules`), VS Code (`copilot-instructions.md`), DeepSeek Harness (`dsh`).
+

package/scaffold/.agent/skills/meta-harness/references/optimized/section-03.md

@@ -0,0 +1,8 @@
+### The Shared Brain Principle
+Because all harnesses operate over the **same portable brain**:
+- Every harness reads the same auto-compiled instruction surface (`AGENTS.md`, `GEMINI.md`, `CLAUDE.md`).
+- Every harness writes directly into the shared SSSS memory vault (`memory-vault/`).
+- Knowledge or solutions learned by one agent in one harness are immediately accessible to every other harness across the computer and mesh.
+
+---
+

package/scaffold/.agent/skills/meta-harness/references/optimized/section-04.md

@@ -0,0 +1,2 @@
+## 🛠️ CLI Operations
+

package/scaffold/.agent/skills/meta-harness/references/optimized/section-05.md

@@ -0,0 +1,6 @@
+### 1. Inspect Active Harnesses
+Inspect all available developer harnesses and their local binary paths:
+```bash
+npx total-recall harness list
+```
+

package/scaffold/.agent/skills/meta-harness/references/optimized/section-06.md

@@ -0,0 +1,14 @@
+### 2. Headless Task Dispatch
+Dispatch a specific coding or research task to any external harness (locally or remotely over the mesh):
+```bash
+# Local dispatches
+npx total-recall harness dispatch claude "Inspect src/core/surface.mjs and run lint checks"
+npx total-recall harness dispatch agy "Analyze latest preprints on room temperature diamond NV registers"
+npx total-recall harness dispatch codex "Generate unit tests for src/core/context-cache.mjs"
+npx total-recall harness dispatch ollama "Summarize the core invariants of SSSS"
+
+# Cross-mesh remote dispatches
+npx total-recall harness dispatch ollama --node build-box "What is test-time compute scaling?"
+npx total-recall harness dispatch claude --node build-box "Run git status and lint tests"
+```
+

package/scaffold/.agent/skills/meta-harness/references/optimized/section-07.md

@@ -0,0 +1,8 @@
+### 3. Multi-Harness Council
+Run a task across multiple harnesses concurrently to obtain comparative solutions and consensus:
+```bash
+npx total-recall harness council "Propose architecture for decentralized research mesh"
+```
+
+---
+

package/scaffold/.agent/skills/meta-harness/references/optimized/section-08.md

@@ -0,0 +1,4 @@
+## 🤖 Agent Process Management
+
+Manage background subagent tasks across the local machine and remote mesh nodes:
+

package/scaffold/.agent/skills/meta-harness/references/optimized/section-09.md

@@ -0,0 +1,9 @@
+### 1. Spawn a Background Agent
+```bash
+# Local background agent
+npx total-recall agent spawn claude "Refactor error handling in src/cli/harness.mjs" --name "Refactor Worker"
+
+# Remote background agent on a mesh node
+npx total-recall agent spawn agy --node build-box "Crawl recent preprints on quantum shuttling" --name "Quantum Scout"
+```
+

package/scaffold/.agent/skills/meta-harness/references/optimized/section-10.md

@@ -0,0 +1,6 @@
+### 2. Inspect Running Agents
+```bash
+npx total-recall agent list
+npx total-recall agent list --json
+```
+

package/scaffold/.agent/skills/meta-harness/references/optimized/section-11.md

@@ -0,0 +1,5 @@
+### 3. Stream Agent Logs
+```bash
+npx total-recall agent logs <agent-id> --tail 100
+```
+

package/scaffold/.agent/skills/meta-harness/references/optimized/section-12.md

@@ -0,0 +1,7 @@
+### 4. Terminate an Agent
+```bash
+npx total-recall agent kill <agent-id>
+```
+
+---
+

package/scaffold/.agent/skills/meta-harness/references/optimized/section-13.md

@@ -0,0 +1,4 @@
+## 🌐 Headscale Mesh Diagnostics & Remote Execution
+
+The Meta Harness integrates directly with your private Headscale WireGuard mesh network:
+

package/scaffold/.agent/skills/meta-harness/references/optimized/section-14.md

@@ -0,0 +1,7 @@
+### 1. Cluster Capability Audit (`mesh doctor`)
+Concurrently audit reachability, developer runtimes (`node`, `docker`, `git`), and AI harnesses (`agy`, `claude`, `codex`, `gemini`, `ollama`) across all cluster nodes:
+```bash
+npx total-recall mesh doctor
+npx total-recall mesh doctor --json
+```
+

package/scaffold/.agent/skills/meta-harness/references/optimized/section-15.md

@@ -0,0 +1,9 @@
+### 2. Non-Interactive Remote Command Execution (`mesh exec`)
+Execute commands on remote nodes with automatic `$PATH` sanitization and structured outputs:
+```bash
+npx total-recall mesh exec build-box uname -m
+npx total-recall mesh exec cloud docker ps
+```
+
+---
+

package/scaffold/.agent/skills/meta-harness/references/optimized/section-16.md

@@ -0,0 +1,5 @@
+## 💻 General Computer & OS Management
+The Meta Harness controls the computer generally:
+- **Activity Monitoring ("Follow the User")**: Tracks interaction timestamps and routes dispatches to the active device across the Headscale WireGuard mesh.
+- **Process Orchestration**: Spawns background worker loops, monitors task completion, and terminates runaway processes.
+- **Surface Sync**: Recompiles instruction surfaces across all IDE directories automatically upon graph updates.