OpenClaw to Hermes Migration (8/13) — Porting an AI Agent with a 52-Item Checklist

7 blocks, 77 items. Get the order wrong and you start over.

ํ•ต์‹ฌ ์š”์•ฝ

  • This post describes the 7-block / 77-item checklist structure for migrating an AI agent's configuration assets from one environment to another. The M block (17 items) stays in the source platform; the items ported to Hermes total 52 (P+O+G+H+S+I).
  • Porting order: P → O → G → H → S → I. The post explains block by block why stacking upper-layer blocks without their dependencies causes reference errors.
  • Covers the 3-tier skill classification (Always-On 5 / On-Demand 9 / Archived 45), the 30-day and 60-day inactivity thresholds, config/loader change patterns just before cutover, and residual token verification.

Overall Structure: 7 Blocks, 77 Items

The unit of work in a porting package is the block, not the individual file. A block is a bundle of files and rules that share a functional purpose. Each block depends on other blocks being present to operate correctly.

Block Code Items Owner
Meta M 17 OpenClaw manager
Persona P 8 Hermes manager
Operational O 7 Hermes manager
Governance G 18 Hermes manager
Heartbeat H 7 Hermes manager
Self-Evolution S 6 Hermes manager
Integrations I 6 Hermes manager

The M block holds metadata and archive items that remain in the source environment. The scope that must be ported and verified in the target environment is 52 items (P+O+G+H+S+I).

Why Order Matters

Blocks form a dependency graph. Upper-layer blocks reference objects defined by lower-layer blocks.

  • Deploying G (Governance) without P (Persona) leaves the agent identity that governance rules must reference undefined.
  • Deploying H (Heartbeat) without O (Operational) causes heartbeat to attempt write operations before the memory-map exists.
  • S (Self-Evolution) is only meaningful after both G and O are installed. Without a baseline to measure change deltas against, the gating logic is effectively disabled.

Installation order: P → O → G → H → S → I. Build the foundation layers first, then stack the behavioral layers on top. Reverse-order installation causes reference errors or runtime warnings.

P Block — Persona (8 Items)

The file set that defines agent identity.

P01 IDENTITY + P02 + P04 VALUES + P05 BEING → consolidated into SOUL.md - Total: 168 lines, 8,381 bytes. - In the source environment, identity files were spread across multiple locations. In this porting effort, they are merged into a single SOUL.md. The goal is not to reduce file count but to allow identity fields to be read and overwritten from a single file, reducing context-loading cost.

USER.md compression - Original: 128 lines - After compression: 1,211 characters (limit: 1,375 characters) - User profile and behavioral preferences compressed for context window efficiency. 164 characters of headroom remaining.

The most commonly missed items in the P block are VALUES and BEING. Porting only IDENTITY leaves the agent knowing who it is but not how it should reason.

O Block — Operational (7 Items)

The infrastructure layer for day-to-day operation.

  • ops board: active task state tracking file
  • linter: output quality inspection rules
  • memory-map: domain knowledge index
  • 25 lessons: accumulated lessons from the source environment

Lessons must not be copied verbatim. Items that reference source-platform-specific paths or service names become dead references in the target environment. During porting, screen for and remove platform-bound references, or rewrite them against the target environment.

G Block — Governance (18 Items)

The block with the highest item count. Includes 3-tier execution governance and a pre-commit guard.

3-tier structure:

Tier 1 (Always-On)   → auto-loaded every session, cannot be disabled
Tier 2 (On-Demand)   → loaded when conditions are met
Tier 3 (Archived)    → inactive state, manual restore only

The tiers partition "what to load always" from a cost perspective. Tier 1 rules are loaded every session regardless of cost. Tier 2 rules load only when specific keywords or context are detected. Tier 3 rules are stored but excluded from the active load set.

pre-commit guard: checks for credential leaks, personal information exposure, and prohibited patterns before commit. The guard enforces governance rules at save time rather than at execution time.

If the P block defines who the agent is, the G block defines what the agent is permitted to do. The two blocks are a pair.

H Block — Heartbeat (7 Items)

The autonomous operation loop.

  • 14-mode pool: 14 operation types available to heartbeat
  • 13 escalation triggers: criteria for actions the agent cannot process automatically

The 14-mode pool is not a simple ping mechanism. It is a collection of distinct operations — research scans, memory cleanup, daily briefings, reflection passes, and others — each with its own frequency, input conditions, and output format.

The 13 escalation triggers define the conditions under which the agent must request user intervention. The purpose of escalation rules is to establish the boundary between what the agent decides autonomously and what requires a human decision.

S Block — Self-Evolution (6 Items)

The agent self-improvement pipeline.

Three gating conditions:

  • reflect: pattern analysis after session end
  • research: judgment that external information collection is required
  • delta gating: lessons are updated only when changes exceed the threshold

Delta gating is the core of the S block. Updating lessons at every session end degrades the signal-to-noise ratio over time. Changes that do not cross the threshold are not saved. This is the storage strategy.

I Block — Integrations (6 Items)

External system connections.

Connection Status
Discord Active
Hub (Home Server) Active
GWS (Google Workspace) Active
Browser Active
Additional System A Connected (standby)
Additional System B Pending

4 Active, 1 Connected (ready but not yet activated), 1 Pending (configuration in progress). The Active / Connected / Pending distinction separates "credential exists" from "runtime calls permitted," allowing the two to be managed independently.

Skill Classification System

3-tier classification across all 77 skills:

Tier Classification Count Condition
Tier 1 Always-On 5 Auto-loaded every session
Tier 2 On-Demand 9 Conditional load
Tier 3 Archived 45 Inactive state

Inactivity thresholds: - 30 days unused → Archived candidate - 60 days unused → deactivation review

The 45 Tier 3 skills can be manually restored when needed. The files are preserved; they are only excluded from the active load set. Demoting skills by inactivity threshold is a design choice to separate "capable skills" from "frequently used skills," controlling per-session initialization cost.

Pre-Cutover State Verification

config.yaml settings at cutover:

skills:
  disabled: 45        # all Tier 3 disabled
platform_disabled:
  discord: 14         # 14 Discord-specific skills disabled separately

Independent of Tier 3 deactivation, platform-specific skills are blocked once more via a per-platform switch. Tier-based deactivation and platform-based deactivation operate on different axes and require cross-axis management.

loader.sh sandbox fix:

REAL_HOME=$(eval echo ~$USER)
HERMES_HOME="$REAL_HOME/.hermes"

In Claude Code sandbox environments, $HOME may be set to an isolated path rather than the actual user home. If the loader computes its data directory from $HOME, agent assets are created in the wrong location. Pinning REAL_HOME with eval echo ~$USER is the workaround. This pattern applies to any environment where HOME is redefined — sandbox, chroot, or container.

Residual Token Verification

Two residual checks after porting is complete:

Item Result
mir/Mir token residue 0
yongjin token residue 0

If authentication tokens from the previous environment are present in the target environment, API calls will be routed to the wrong account. Residual token verification is not functional verification — it is account isolation verification. This is the final check before cutover.

Checklist Summary

The 52-item porting sequence in order:

  1. Assemble SOUL.md (P block)
  2. Compress USER.md and verify character limit
  3. Port ops board / linter / memory-map / lessons (O block)
  4. Configure 3-tier governance + pre-commit guard (G block)
  5. Register 14-mode heartbeat pool + escalation criteria (H block)
  6. Set up reflect / research / delta gating pipeline (S block)
  7. Connect 4 Active integrations, verify status of remaining (I block)
  8. config.yaml: disabled 45 / platform_disabled.discord 14
  9. Fix loader.sh REAL_HOME
  10. Run residual token verification

Applicability and Open Questions

This checklist generalizes to any task involving moving an agent's configuration assets from one environment to another. The same structure applies not only to platform porting but also to reinstallation on the same platform, replication to multi-user environments, and rebuilding isolated sandbox environments. Block count and item count vary per agent, but block dependency order / tier-based load control / platform-axis separation / residual token verification are common patterns across all cases.

Open question: whether to express the inter-block dependency graph as an explicit DAG, or to leave it as an implicit "recommended order." An explicit DAG enables automated validation but adds overhead. An implicit order is lightweight but defers order errors to runtime. Moving to an explicit DAG appears inevitable at scale, but the inflection point remains open.

Series overview: Series index

๋Œ“๊ธ€

๋Œ“๊ธ€ ์“ฐ๊ธฐ

์ด ๋ธ”๋กœ๊ทธ์˜ ์ธ๊ธฐ ๊ฒŒ์‹œ๋ฌผ

Agent Memory Engine (2/10) — Building an AI Agent Memory System with SQLite Alone

← 1/10 SQLite Memory Engine f… ๐Ÿ“š Series Index 3/10 memcore Completed → A memory engine that runs without daemons, without dependencies, anywhere ํ•ต์‹ฌ ์š”์•ฝ This post covers how to design an AI agent memory engine (memcore) on a single-file SQLite backend. - Structure: 25 modules, ~6,464 lines, 22 DB tables - Principles: no daemon, single-file portability, no required external dependencies, host-agnostic - Core techniques: 4-layer hybrid search (topic · vector · FTS5 · LIKE), U-tag dialectic-based confidence evolution, ingestion-layer bias rejection rules What This Post Covers One approach to building AI agent memory without a server or cloud vector DB. Specifically: (1) the limitations of text-file-based memory, (2) vector DB alternatives compared and selection criteria, (3) how to layer search tiers, (4) how to evolve confidence scores over time, and (5) how to block biased memory accumulation at the ingestion stage. Design Principles: Why Single-File SQLite Text-file-based ...
Read more »

"ML Foundations (9/9) — PyTorch vs TensorFlow, and the Road to Local LLMs"

์ด๋ฏธ์ง€
← 8/9 Deep Learning Architec… ๐Ÿ“š Series Index (series end) The final part. So far we've focused on models. Now we focus on the tools that actually run them . We'll lay out the philosophical difference between PyTorch and TensorFlow, and trace how Hugging Face transformers , llama.cpp , MLX , and Ollama built the bridge to running large language models on your own machine. By the end you should have the full mental model of "download a pretrained LLM and serve it locally." 0. Learning Objectives Explain the eager-vs-graph difference between PyTorch and TensorFlow. Explain, in graph terms, how autograd automates the backward pass. Use Hugging Face transformers ' AutoModel , AutoTokenizer , and pipeline abstractions. Describe llama.cpp's GGUF quantization, INT4 inference, and CPU-first flow. Describe Apple MLX's use of unified memory and how it differs from PyTorch. Run a local LLM with Ollama and call it through an OpenAI-compatible API. ...
Read more »

"RAG Core Study (14/26) — Evaluation Sets with RAGAS & DeepEval"

์ด๋ฏธ์ง€
Series overview: Series index ๐Ÿ“š Series Index Without an evaluation set, every RAG improvement is a story, not evidence. RAG systems fail in more than one place: retrieval, context selection, answer generation, and grounding. That means "looks better" is not a metric. Part 14 explains how to build a golden dataset , how RAGAS and DeepEval fit into the workflow, and why teams must separate retrieval quality from answer quality if they want tuning decisions to hold up. 0. Prerequisites Part 13 reranking — system quality now depends on multi-stage ranking. Part 1 grounding — a correct-looking answer is not enough. Part 12 Hybrid — multiple retrieval settings need comparison on fixed data. 1. Learning Objectives Build a small but useful golden dataset for RAG. Distinguish retrieval metrics from answer metrics . Use RAGAS and DeepEval in the right roles. Avoid the common traps in judge-model-based evaluation. 2. ํ•ต์‹ฌ ์š”์•ฝ An evaluation set for RAG is ...
Read more »

"ML Foundations (8/9) — Deep Learning Architectures: CNN, RNN, Attention"

์ด๋ฏธ์ง€
← 7/9 Deep Learning Training ๐Ÿ“š Series Index 9/9 PyTorch vs TensorFlow → What did we keep adding on top of an MLP? Vision went CNN, sequences went RNN/LSTM, and eventually both converged on Attention and the Transformer. Each architecture is easier to remember if you read it as what it gave up to gain something else . This part is the one-line summary of the decisive inflection points: LeNet → AlexNet → ResNet → LSTM → Attention → Transformer. 0. Learning Objectives Explain why CNN's convolution, pooling, and weight sharing match images so well. Trace what got unblocked from LeNet → AlexNet → VGG → ResNet as depth grew. State BPTT for RNNs and the vanishing/exploding gradient problem. Write the LSTM gate equations (forget, input, output) with intuition. Write the attention formula in query/key/value form and the scaled dot-product variant. State the precise reasons Transformers replaced RNNs (parallelism + long-range dependencies). 1. ํ•ต์‹ฌ ์š”์•ฝ CNN : weight sharin...
Read more »

"ML Foundations (7/9) — Deep Learning Training: Optimizers, Regularization, Initialization"

← 6/9 Neural Networks ๐Ÿ“š Series Index 8/9 Deep Learning Architec… → Building an MLP, as we did in Part 6, is not the same as getting it to train well . The right optimizer, regularization, initialization, and learning rate — when these line up, deep networks converge. When they don't, the network refuses to learn at all. This part is the catalog of those crafts, with formulas, paper citations, and code in one place. 0. Learning Objectives Compare and write the update rules for SGD, Momentum, Nesterov, and Adam. Explain how Dropout, BatchNorm, and LayerNorm work and where they belong in a model. Derive the variance formulas for Xavier and He initialization and match them to activations. Implement step, cosine, and warmup learning-rate schedules in PyTorch. Explain why gradient clipping is effectively required for RNNs and Transformers. Diagnose the most common training failures (NaN, plateau, overfitting) and apply first-line fixes. 1. ํ•ต์‹ฌ ์š”์•ฝ SGD : \(w \leftarro...
Read more »

OpenClaw to Hermes Migration (2/13) — What to Preserve, Partially Port, or Discard

← 1/13 Current Structure Inve… ๐Ÿ“š Series Index 3/13 What Moves to Hermes → A code review record classifying operational scripts into three tiers: preserve, partial-preserve, and discard. What This Post Covers An asset classification framework (preserve / partial-preserve / discard) for migrating a legacy agent system to a successor A concrete case applying the Strangler Fig + Subset Migration strategy to a memory pipeline Classification results evaluated across LOC, external dependencies, and replaceability axes A method for slimming down a bloated single file (1,310 LOC) using the migration as a natural entry point Problem Definition: Port Everything or Cut Deep? The OpenClaw inventory comprises multiple agents, multiple scheduler/daemon processes, and multiple operational scripts. When moving to the successor system, Hermes, the choice is not binary. A full port carries accumulated technical debt along with it; a minimal port discards hard-won domain logic. Code-revie...
Read more »

AI Agents I Built (5/7) — Building an Automated Blogger API Publishing System

← 4/7 My Life Organizer Agent ๐Ÿ“š Series Index 6/7 Dual → Designing a Markdown-Based Auto-Publish Pipeline with Blogger API v3 + OAuth 2.0 Key Summary Blogger API v3 + OAuth 2.0 enables a pipeline that converts Markdown files into publishable HTML posts blogger_publish.py handles the full flow — frontmatter parsing → body cleanup → HTML conversion → CSS injection → API post — in a single script For bulk publishing, quota management (delay, retry, TOC suppression) is the critical design concern Platform Selection Criteria API access is mandatory for AI agent-driven content publishing. A comparison of platforms: Platform API Cost AdSense Decision WordPress.com REST API Paid plan required Paid only Cost overhead Ghost REST API Self-hosted or paid Manual setup Ops overhead Blogger v3 API Free Native integration Adopted Blogger is free, exposes a REST API, and integrates natively with AdSense. For personal blog automation, this combination i...
Read more »