Introduction

DSH’s plugin philosophy is “everything is a plugin.” For agent developers, a common issue is: while corrections, preferences, constraints, and project facts may appear in the assistant session, this information isn’t stably accumulated, requiring background explanation to be repeated in the next session.

rxa3c/chat2skill, included in the independent community directory, addresses this problem. Its positioning is: automatically learning reusable skills and project memory from assistant sessions. It will distill corrections, preferences, constraints, and project facts into local memory and SKILL.md files, and inject them into future sessions. The following only introduces content verified from available materials; unverified items will be marked.

What is this

The name in the directory information is rxa3c/chat2skill, and the owner is rxa3c. It primarily handles reusable project context within assistant sessions, rather than merely preserving single-turn dialogue text.

The artifacts provided by the material include:

  • atomized skills: reusable skills saved in the form of SKILL.md files.
  • Local project memory: The material states that project memory is saved locally using SQLite and includes project facts.
  • PROJECT_SKILL.md: A project-level documentation file synthesized from active skills.

The material states it is best suited for Claude Code, Codex, Cursor; other agents need to support lifecycle hooks or be able to run CLI scripts.

Core Features

Post-Session Extraction

Chat2Skill analyzes corrections, preferences, constraints, and project facts in the session and generates local memory and SKILL.md files based on this.

Pre-Session Injection

In subsequent sessions, it retrieves relevant project memory and active skills, injecting relevant content into future sessions.

Candidate Skills Cycle

The feedback loop provided in the material includes:

  1. Retrieve memory and skills before the session.
  2. Use the retrieved content in subsequent sessions.
  3. The assistant works, and the user accepts, corrects, or constrains it.
  4. Extract learning signals after the session ends.
  5. Create, edit, or discard candidate skills.
  6. Validate, replay, merge, and store active skills.
  7. Rebuild PROJECT_SKILL.md and update the local profile.

LLM and Non-LLM Modes

The material states that when an LLM is available, a Proposer/Generator/Judge style flow is used; when there is no LLM, keyword detection and template-based generation are used.

Local Storage and Credentials

The local data directory is:

~/.chat2skill/

The material states this directory saves skills, profile, and history, using SQLite + markdown files. The cloud is used for stateless extraction execution, and the material claims no data is stored there. The extraction LLM call supports bringing its own API key or short-lived OAuth bearer token. The material explains that credentials are used in memory and are not persisted or logged on the server. If no available credential is found, it falls back to low-quality heuristics.

Costs and Triggers

The typical extraction cited in the material involves about 4 LLM calls:

detect
analyze
generate
judge

If replay validation is performed, up to 5 more calls are added for the history.

The session window takes the last 40 messages. Extraction is not triggered in every session, but only when a correction/constraint signal is detected.

Installation and Activation

The material does not show the complete installation command for DSH/DeepSeek Harness. Therefore, without the original DSH installation text, it is not recommended to concatenate dsh plugin add ... based on the plugin name.

The DSH-related configuration fields mentioned in the material include:

main
exports
dsh.bundle
dsh.client.inject

The entry path pointed to by main/exports is:

adapters/deepseek-harness/plugin.mjs

The material also mentions that the version in package.json is:

0.1.3

However, it does not state whether this value equals the plugin version.

The installation example for Codex given in the README is:

codex plugin marketplace add rxacc/chat2skill

Then open /plugins, select the chat2skill marketplace, and install chat2skill. It should be noted: the directory information uses rxa3c/chat2skill, while this command uses rxacc/chat2skill, and the relationship between the two is unconfirmed.

The beginning of the installation instructions for Claude Code in the material is:

claude plugin marketplace add https://github.com/rxacc/chat2s…

However, the command is truncated in the material, and the complete URL and subsequent installation steps are unconfirmed. It is not recommended to execute the incomplete command directly.

Typical Usage

Disable Response Guard

The material states the response guard default mode is strict and reads the response_guard frontmatter. If you wish to disable it, you can set:

CHAT2SKILL_RESPONSE_GUARD=false

Non-LLM Fallback

When an LLM is not available, the material claims the cycle can still use keyword detection and template-based generation. If no available credential is found, the material mentions falling back to lower-quality heuristics.

Observe Trigger Points

In daily use, focus on observing whether corrections, constraints, or stable preferences appear in the session. The material states that the extraction process only enters when a relevant signal is detected.

Suitable Scenarios and Notes

Suitable for scenarios where repeated corrections, preferences, constraints, and project facts in assistant sessions are accumulated as local memory and SKILL.md.

Notes:

  • The material states it is best suited for Claude Code, Codex, Cursor; other agents need lifecycle hooks or CLI scripts.
  • The plugin runs with the current dsh process permissions; source code and license should be checked before installation.
  • No license field or LICENSE information appeared in the material; the license status is unconfirmed.
  • The difference in installation commands between rxa3c and rxacc is unconfirmed.
  • DSH installation commands are unconfirmed; do not concatenate installation commands directly from the directory name.
  • Extraction will incur LLM call costs: typically about 4 times, with up to 5 more added for replay.

Conclusion

The value of rxa3c/chat2skill lies in converting corrections, preferences, constraints, and project facts from the session into local memory and SKILL.md, allowing subsequent sessions to continue working based on existing accumulated knowledge.

Directory page:

https://www.skillhub.cn/plugins/rxa3c/chat2skill

GitHub:

https://github.com/rxa3c/chat2skill