ai-discover

Read-only discovery step of the auto-improvement loop. One bounded agent turn per cycle reads the focus files (Read/Grep/Glob only, no shell, no subagents) and replies with a JSON array of testable su

SKILL.md

GitHub

ai-discover — the discovery step of the auto-improvement loop

This skill drives Phase A (discovery) of an auto-improvement cycle. It is discovery only: it never applies a change, never runs the keep-or-revert A/B, and never decides whether a finding is kept or drafted as a pull request. Those are the spine's deterministic Python gate / keeper / pipeline, which no model can argue past. That separation is the point — the measurement is the product.

How discovery actually runs

The spine runs discovery as ONE bounded agent turn per cycle (spine/agent_discovery.py discover_surfaces_via_agent), not a fan-out. Do NOT spawn subagents: the turn must end with the JSON array itself, and children spawned from an unattended runner orphan and hang the run (see the SessionAgentRunner notes in spine/agent_runner.py).

  • Tools are Read, Grep and Glob only. There is no shell and no edit: discovery is read-only by capability, so you cannot run a profiler or write a test here.
  • Read the priority slice you are given, stop after 2-3 solid findings or the read budget, and reply with ONLY a JSON array (at most the stated limit) of {file, line, symbol, rule, message, hypothesis} — rule is AGENT, or COVERAGE for a coverage gap. Reply [] when nothing is testable. The final reply IS the output; there is no separate artifact file.
  • The same surface shape feeds both tracks. On the bug track the profile names the reproducing test path and author_bug_fix later writes the RED test and the fix in an isolated worktree; on the perf track author_perf_fix writes the edit. The spine dedups by (kind, basename::symbol) locus, then gates (RED -> GREEN -> STAYGREEN for bugs, serial pinned A/B for perf).

Reporting a candidate honestly

State the expected win as a hypothesis, not a result. The spine measures it; if your estimate was wrong the candidate is reverted and that is a normal, useful outcome. A guess presented as a measurement is the failure mode this whole app exists to prevent.

What this skill never does

  • Never edits the ruler, the measurement harness, the tests-of-record, or anything outside the active target profile's edit allowlist. Those paths are mechanically rejected, so an edit there wastes the whole cycle.
  • Never publishes or merges a pull request. Survivors are drafted as GitHub draft PRs by the spine, and a human publishes them.
  • Never fabricates a measured number. A fabricated win is the worst possible reward-hack, because it corrupts the record the loop reasons from.