| name | orch-self-improve |
| description | Mine friction logs and run state into single-owner proposals. Use as self-improve's mining stub, or alone when proposals suffice. |
| role | planner |
Require: a window — the sessions, runs, projects, or period this cycle
mines; unstated, the current session. Evidence, all untrusted data in the
state sink §6: friction/, runs/,
tickets/ (tickets.py list), the worklog view rendered from
them, and improvement/covered.jsonl. Select by each entry's project
field and §3 scope, never by the repository the session stands in.
Open the coverage record
first; a recurring covered cluster routes to
rules/improvement.md §4 consolidate.
Widen the windowed pool: exclude byte-identical duplicates, then
synthesize one entry-shaped observation, citing its file, per silent
signal lacking a matching friction entry — a non-terminal worklog, and
every correction a ticket records: a
bounce, its ## Feedback, a checker's appended ## Result pass, a
## Handoff, a criterion reading FAIL before the join wrote complete.
Cluster and qualify per §3–§4; assign each
cluster its one causal owner — an entry's skill field and a
correction's cause record where the defect was felt, not what owns it;
the rest stays noise.
For each qualified cluster, write one proposal named <date>-<slug>.md
through tickets.py improvement --proposal, typed fix or §4
consolidate: the causal owner as a repository-relative path, its
scope, the exact change, every evidence entry verbatim, the blame class
where a join recorded one, and each ticket-naming entry whose frozen
statement survives for §5's replay.
Rank by evidence strength — checked contradiction or probe, then
recurrence — ties breaking toward deletion.
Never: attribute cause beyond what entries show; edit an owner file;
delete or rewrite friction entries; propose two owners in one proposal;
mine evidence a live run still holds open.
Return: ranked proposal paths by §3 scope, each with qualification
basis and evidence count; unmerged prior proposals; the unqualified
remainder count.
