bro self learn
Triggered lessons with evidence, confidence, and a memory that knows when to speak.
bro self learn stores durable lessons for the moments they matter. A lesson
is a rule plus a trigger: hook events and optional terms, commands, paths,
tools, or failed-tool evidence decide when it surfaces.
Commands
| Command | What it does |
|---|---|
bro self learn add --lesson "…" --on E… --evidence K:R… | Store a manual lesson. --on and at least one --evidence are required; both are repeatable |
bro self learn list [--json] [--source X] [--confidence X] | List lessons, optionally filtered |
bro self learn show <id> | Print one lesson as JSON |
bro self learn forget <id> | Remove a lesson |
bro self learn capture [--source drill|retro|act|mol|all] [--mol ID] [--dry-run] [--json] | Distill finished artifacts into lessons |
bro self learn probe <question> | Ask the store first; a hit prints ranked lessons and a miss records the gap |
bro self learn probe can also store an answer in its second phase with
--lesson, --on, --match-terms, --match-commands,
--match-paths, --match-tools, --match-errors, --budget, and
optional --evidence.
Manual evidence uses K:R: the supported kinds are bead, pr, session,
command, and text.
Triggers and confidence
Each --on value is one of:
session-startprompt-submitpost-tool
--on is repeatable, so a lesson can fire on several events: --on session-start --on post-tool. A comma-separated value (--on session-start,post-tool) works too. Repeats of the same event collapse
to one entry.
Match keys are conjunctive across categories and disjunctive within a
category: terms match prompt or trace text, commands match command
prefixes, paths match touched paths, tools match tool names, and
--match-errors matches a failed tool landing. --budget limits how many
times one lesson may fire in a session; the default is one.
Evidence is part of the lesson, not decoration. One unverified evidence
item is tentative; independent evidence or a lesson that held under a
real gate can make it established. proven is a recognized confidence
value, but the current confidence ladder does not assign it.
Where lessons appear
The learn connector probes at session-start, prompt-submit, and post-tool.
Each probe is fail-open and emits at most learn.maxInject lesson lines.
The per-lesson trigger budget is separate: it prevents one post-tool lesson
from nagging after every command.
{
"learn": {
"enabled": true,
"maxInject": 3,
"sources": []
}
}enabled is the kill switch. An empty sources list allows every lesson
source; a non-empty list limits injection to known sources such as
manual, probe, or capture:drill.
Learn versus bd remember
Use learn for conditioned knowledge:
when
gh pr mergeruns, sweep review debt.
Use bd remember for context-free memory:
this repo uses native TypeScript.
Memories are always-on context. Lessons wait for their trigger, so a
mistake does not become another permanent prompt paragraph. Run
bro self learn capture --dry-run before turning a finished artifact into a
lesson; it renders the proposal without writing it.