From aurora
Nightly self-improvement loop. Reads the day's `learnings/<date>.jsonl`, clusters them by skill/agent, identifies recurring patterns (≥ 3 occurrences across ≥ 2 projects), and opens a GitHub PR against `skills/` (or `agents/`) with a proposed update — e.g., a new fingerprint refinement, an updated SKILL.md guidance, a new helper script. The PR is HITL-gated via `aurora-promote` (kind: skill_compost_pr) and never auto-merged. The mechanism that makes the swarm get smarter with use.
How this skill is triggered — by the user, by Claude, or both
Slash command
/aurora:aurora-compostThe summary Claude sees in its skill listing — used to decide when to auto-load this skill
Every agent writes one-line learnings as it works. Most are mundane. Some recur. The ones that recur, across projects, with consistent rationale — those are worth turning into a skill update. Compost is what turns the day's accumulated learnings into a versioned, reviewed change to the swarm itself.
Every agent writes one-line learnings as it works. Most are mundane. Some recur. The ones that recur, across projects, with consistent rationale — those are worth turning into a skill update. Compost is what turns the day's accumulated learnings into a versioned, reviewed change to the swarm itself.
aurora compost --since N --dry-run to preview proposed PRs${AURORA_HOME}/learnings/*.jsonl — every agent's appended learnings, one JSON object per line:
{
"ts": "2026-05-09T03:42:11Z",
"agent": "surgeon",
"skill": "aurora-fingerprint",
"project_id": "CAND-2026-05-09-aabbccdd",
"kind": "fingerprint-resolution",
"summary": "selector-broken/wnd-aaname-mismatch resolved by re-walking parent ariaName chain",
"context": { "...": "..." }
}
The fingerprint SQLite at ${AURORA_HOME}/fingerprints.db — for cross-referencing cluster sizes
Org memory at .aurora/org/
For each composted pattern, one GitHub PR against the swarm's own repo. PR title format:
Skill update: <skill-name>: <one-line rationale>
PR body:
# Compost-step skill update
## Pattern observed
<rationale: what recurred, in what context, why an update is warranted>
## Evidence
- 12 occurrences across 4 projects in last 30 days
- Cluster `selector-broken/wnd-aaname-mismatch`
- All 12 resolved by Surgeon with the same remediation; 0 regressions
## Proposed change
- `skills/aurora-fingerprint/scripts/cluster.py` — add a refinement detector for the new pattern
- `skills/aurora-fingerprint/SKILL.md` — document the refinement in the canonical taxonomy table
## Risk
- Low — additive; no existing behavior modified
## Verification
- [ ] aurora-policy validate (passes)
- [ ] tests/aurora-fingerprint pass with the new fixture
- [ ] HITL approval via Action Center
After opening the PR, this skill calls aurora-promote with kind: skill_compost_pr to gate it. The PR is never auto-merged — even when CI is green and the change is trivially additive. The whole point of the gate is bounded autonomy.
A learning becomes a compost candidate when ALL hold:
summary text (TF-IDF threshold)kind == fingerprint-resolution, the proposed remediation worked the last 3 times in a row (no regressions)If only some of those hold, append the learning to a "watching" list and revisit next compost. Don't propose half-baked PRs.
| Pattern type | Target |
|---|---|
| New fingerprint refinement | skills/aurora-fingerprint/scripts/cluster.py (derive_refinement) |
| New canonical kind | skills/aurora-fingerprint/SKILL.md taxonomy table |
| Updated PDD anti-pattern | skills/aurora-pdd/templates/ambiguity-rubric.md |
| New REFramework lint rule | .claude/rules/aurora-conventions.md |
| Vendor-specific selector quirk | .aurora/org/vendor-selector-quirks.md (org memory, not a skill) |
| New gate template | skills/aurora-promote/templates/<kind>.json |
| New agent role (rare) | agents/<name>.md — always HITL, always with explicit Conductor schedule |
lib/aurora/*.py core code. Code changes are PRs from humans (or surgeon-style fixes), not from compost. Compost only updates skills, agent definitions, rules, and org memory.A one-line summary plus the PR list:
aurora-compost: 2026-05-09 — 1 PR opened (aurora-fingerprint refinement), 2 watching, 0 deferred
- https://github.com/aurora-demo-org/uipath-for-coding-agents/pull/47
Guides creation, editing, and verification of skills for AI coding agents using test-driven development with subagent scenarios. Use when authoring or debugging skills.
npx claudepluginhub mlbrilliance/uipath-for-coding-agents --plugin aurora