OpenClaw to Hermes Migration (10/13) — Applying a Strangler Fig Variant in Single-User Environments

What This Post Covers

This post documents a migration technique for coexisting a legacy system and a new system in a single-user (one operator) environment. The canonical Strangler Fig assumes facade routing and is designed for environments with high volumes of statistically uniform external traffic. In a single-user environment where input is unevenly distributed and traffic splitting carries little statistical meaning, a variant that removes the facade and retains only "comparison baseline preservation" is effective. This post covers how that variant works, its connection to Subset Migration, actual measured values, and confirmed limitations.

Three core arguments: - The essence of Strangler Fig is not incremental replacement but "preservation of prior state." - The sequencing in Subset Migration must be driven by "measurability." - A comparison baseline serves two distinct roles: anomaly diagnosis and improvement magnitude verification.

Two systems are used as examples: OpenClaw, the stable system, and Hermes, the new system absorbing it. The documented incident is a token runaway in Hermes v0.8.0 (5 inputs → 111 internal tool calls → 6.5M cumulative tokens, repetition factor 42.79×). This regression event serves as the reference point for validating the variant technique.

Canonical Strangler Fig and the Single-User Variant

The canonical Strangler Fig works as follows. A new system is placed alongside the legacy system, and a facade receives all external requests. The facade routes some requests to the legacy system and some to the new system according to defined rules. Over time, the new system handles an increasing share, and the legacy system withers. Finally, the legacy system is decommissioned.

This pattern operates in environments with many external users and statistically uniform traffic distribution. In a single-user environment, facade routing loses its meaning because input distribution is not uniform. On some days, ten requests may go to one side and zero to the other. Under these conditions, facade-distributed results cannot serve as a statistically valid comparison baseline.

The key change in the variant is removing the facade and replacing it with a shared measurement layer applied simultaneously to both systems.

Given the same input, both systems are measured against identical formulas to check whether they produce equivalent output. Rather than routing requests, the systems share a measurement layer.

In practice, this means the OpenClaw metrics (repetition factor, schema re-sends, delegation, crash, daily cost, cache hit) are measured in Hermes using the same columns and the same formulas. As long as this comparison baseline is active, both systems can be evaluated against a common standard.

The implication of this variant: The essence of Strangler Fig is not "incremental replacement" but "preservation of prior state." Replacing everything at once destroys preservation. Without preservation, comparison cannot be established. Without comparison, the basis for judging improvement or regression disappears.

Subset Migration — Measurability as the Primary Ordering Criterion

Subset Migration is the pattern of migrating systems not all at once, but starting with the most measurable subset. Once a migrated subset operates correctly, the next subset is migrated.

The critical decision in applying this pattern is determining "what is measurable." Migrating an unmeasurable component first means state changes immediately after migration cannot be observed.

The priority order applied:

  1. Memory engine — SQLite-based storage. Input/output schema is well-defined; read/write counts are measurable.
  2. Measurement scripts — Code computing token counts, repetition factors, and cache hit rates. This module must be migrated first so that all subsequent migration steps can be measured.
  3. Environment variables — Model selection, routing options, and other explicitly declared values.
  4. Credentials — API keys, OAuth tokens. These are static values; migration risk is low.

Components deferred to lower priority were high-variability areas: the tool set and skill triggers. Tools are added and removed frequently; skill trigger keywords change. Migrating high-variability components first means "the effect of migration" and "the effect of tool changes" cannot be separated. Migrating at a point where the measurement baseline is shifting makes regression root-cause analysis impossible.

Determining this ordering required evaluating 34 decision items. When the dependencies between items are mapped, it becomes self-evident that measurement scripts must follow the memory engine. Without measurement, the effects of all subsequent migration work remain unobservable.

Migration Phases and Regression Behavior

The full flow breaks into five phases.

Phase 1 — First attempt. Memory engine and measurement scripts are migrated to Hermes first, followed by environment variables and credentials. Hermes v0.8.0 then begins receiving inputs. OpenClaw is frozen and maintained as a preserved baseline.

Phase 2 — Runaway observed. 5 inputs → 111 tool calls → 6.5M tokens, repetition factor 42.79×. Hermes v0.8.0 over-repeated internal tool calls per input unit and re-sent the same schema multiple times, causing cumulative token growth.

Phase 3 — Rollback to stable system. Input is redirected back to OpenClaw. This was only possible because the stable system remained alive in a frozen state. If the new system had been deployed as the sole system, no rollback target would have existed, forcing continued operation in a runaway state or a full work stoppage.

Phase 4 — Redesign. Design defects in the initial implementation are isolated. The 34 decision items are re-examined, and the direct cause candidates for the runaway (conditions triggering repetition factor spikes, schema re-send origination points) are isolated separately. The version produced by this process is the "post-redesign" version.

Phase 5 — Retry, under observation. Hermes is restarted on the post-redesign version and baseline is measured: 23.80×. Approximately 45% reduction from v0.8.0's 42.79×. However, this is a single baseline data point, and a regression was detected in subsequent operational-day measurements (next section). Phase 5 is therefore "under verification," not "successful."

The structural implication of this flow: the migration did not complete in a single pass. All five phases — first attempt → runaway → rollback → redesign → retry — were only possible because OpenClaw remained alive as the stable system.

Two Roles of the Comparison Baseline

The baseline operated in two distinct modes.

First, anomaly diagnosis. The absolute value of 42.79× alone does not allow severity to be assessed. Placing it alongside the same-formula operational data from OpenClaw reveals the deviation from the mean. Because OpenClaw records were preserved, the diagnosis "this is an anomaly and indicates a system design problem" was possible.

Second, improvement magnitude verification. The retry-phase baseline of 23.80× and v0.8.0's 42.79× are results from the same formula, so comparison is valid: approximately 45% reduction. If the v0.8.0 measurement logs had been discarded at the time of the runaway, 23.80 would become an isolated figure with no reference point. Whether it represents improvement or deterioration would be indeterminate.

Methodological Limitations

The Strangler Fig variant removes facade routing, so it does not run identical inputs through both systems simultaneously for comparison. OpenClaw is frozen; only Hermes receives new inputs. Therefore, the comparison between 23.80 and 42.79 is strictly "applying the same formula to input distributions from different time periods." This is a structural limitation — the input sample is not controlled — and controlled re-measurement will be necessary in a later phase.

First Operational Day and Regression Detection

First operational day KPIs:

  • Repetition: 13.82×
  • Delegation: 1.0
  • Crash: 0
  • Daily cost: $0
  • Cache hit: 25.64%

Repetition measured below the baseline of 23.80×, crash 0, delegation 1.0 (sub-agent delegated as intended), daily cost $0 (local model-centric). Up to this point, results were moving in the intended direction.

Second operational day regression:

  • Repetition: 16.22× (up from 13.82 the previous day)
  • Delegation: 0.0 (down from 1.0 the previous day)
  • Crash: 1 (up from 0 the previous day)

Three indicators degraded simultaneously. Root cause is being isolated. This regression is recorded because the purpose of building a comparison baseline is precisely to capture moments like this. Recording only improving days nullifies the baseline's function.

Currently in the Phase 0 observation window. The data set is 1 baseline point + 2 operational-day points. This sample size is insufficient to distinguish a trend from transient oscillation. Until the observation window closes, the 23.80× baseline and subsequent operational KPIs are treated as provisional observations only.

Summary — Applicable Scope and Open Questions

Applying the Strangler Fig and Subset Migration variants in a single-user environment confirmed the following principles at the technique level:

  • Preserve a measurement baseline instead of a facade. Secure both a rollback target and a comparison standard simultaneously.
  • Migrate measurable subsets first. High-variability components are lower priority. Migration effects and tool-change effects must be separable for regression root-cause analysis to be possible.
  • Migration does not complete in a single pass. Design must assume multi-phase iteration; the critical prerequisite for that assumption is the survival of the stable system.

Applicable scope: environments where external traffic is not statistically uniform (internal tooling, personal automation, single-team internal systems); environments where both systems can share identical metrics; environments where inputs can be redirected back to the stable system on regression.

Two open questions remain: 1. To what extent can a baseline comparison with uncontrolled input samples be used as valid evidence? 2. Can the regression observed in Phase 0 (Repetition 16.22, Delegation 0.0, Crash 1) be attributed to a single root cause, or is it multi-factorial?

Both are under investigation. Until verification results are available, both the 23.80× baseline and all subsequent operational KPIs are treated as provisional observations only.

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 »