Skills from completed work
Let Chetty turn a long, successful task into a reusable procedure, and stay in control of what it saves.
Turn learning on
Open Account → Memory and switch on Skill learning. Switch it off there at any time.
An administrator sets the limits under Runtime in the AI settings, behind Edit runtime. A write budget of 0 in Maximum skill changes per exchange switches learning off for everyone, while the manual skill tools keep working.
When a conversation is reviewed
A finished exchange counts only when it made at least Minimum tool calls for a skill review successful tool calls, one by default.
Once the counted exchanges add up to Skill review every N tool calls, 100 by default, Chetty considers the completed conversation. It looks for a reusable procedure, meaning how to do something rather than facts about you or a project, and uses the same model and provider account as the conversation.
An interval that someone saved explicitly stays unchanged.
What Chetty proposes
- A new skill is an ordinary personal skill, marked as created by the agent.
- There is no approval step: it is offered exactly like a skill you saved yourself. Turn off Show to agent automatically if you only want to start it with
/skill:name. - A new skill never takes a name you can already see in your skill list, including skills that other plugins contribute.
- An existing skill is improved only when it was loaded successfully in that exchange, is still agent-created and did not change in the meantime. The model sees the whole body and replaces it as a whole.
- One review saves at most the number of changes in Maximum skill changes per exchange, one by default.
What it never touches
Skills you wrote or edited, bundled skills and instance-wide skills are never changed by a proposal.
A body longer than 8,000 characters, or one containing something that looks like a secret, is shown read-only and cannot be patched. A proposal containing something that looks like a key, token or password, or exceeding the size limit, is dropped before it is saved.
Any change you make in the manager, including the automatic switch, removes the agent mark, so from then on only you change that skill.
Limits and cost
Review usage is charged to your account, including reading the cached conversation and the fresh input for the review. Switching learning off or lowering the write budget stops further writes under the live limit.
A review gets the existing conversation and its full final answer without executing tools. It does not remove material already present in the conversation it inherits.
Skipped reviews
A review is skipped, with no model request and no skill change, when:
- the provider route cannot repeat the conversation exactly, for example a chained session or provider-hosted tools;
- capacity is unavailable or the exact source can no longer be used;
- the provider account's weekly subscription usage is at 90% or more, or unknown.
A skipped review keeps its count for a later eligible exchange, and there is no retry on a different model.
Where to see it
Account → Memory shows one line under the switch: how many reviews ran, how many skills they saved, how many proposals were rejected, how many failed or timed out, when the last one ran and how it ended. It also shows how many tool calls remain until the next review.
The daemon log has one line per review saying what it did.
The per-account counter sits in a hidden folder next to the skills and is deleted with the account. A missing counter starts at zero; an unreadable or invalid counter stops that review with an error instead of silently resetting.