drift-analysis

name	drift-analysis
description	Read signals → decide which to act on, which to ignore, which to defer. Never act on a single signal in isolation; require corroboration.
allowed-tools	["Read"]

Drift Analysis

Signals live under data/signals/<skill_slug>.json in your workdir. Each is one row from skill_curation_signals:

{
  "id": 42,
  "skill_slug": "user-profile",
  "section_id": "topic-files",
  "signal_type": "structure_diff",
  "payload": { "sub_kind": "heading_add", "target": "identity/personal.md#health-log" },
  "observed_at": 1717000000000
}

Corroboration rule

A single structure_diff signal is suggestive, not actionable. To submit a proposal you must EITHER:

• cite ≥ 2 distinct signals affecting the same (skill, section), OR
• cite a single signal of weight ≥ 3 (only owner_correction qualifies).

Single-signal proposals from structure_diff alone are auto-rejected by the smoke test (signal_citations_valid). Don't try.

Triage taxonomy

For each (skill, section) with corroborated signals, classify the change shape:

• additive_only — signals all describe ADD operations (file_add, heading_add). Propose with confidence; the worst case is a revert.
• mixed — some adds, some modifications. Propose with extra rationale citing each signal's contribution.
• destructive — at least one signal cites a removal. Only propose if signals explicitly state user intent to remove (owner_correction with intent="remove"; structure_diff alone is insufficient). When in doubt, SKIP.

Skip conditions

Do nothing on (skill, section) when ANY of these hold:

• Signals contradict (one says add, another says remove the same target).
• All signals are < 24h old (structure_diff low-weight gate hasn't matured; the snapshot may be a brief experiment).
• Section is frozen (returns from /skills/<slug> as frozen=true).
• Section had a proposal applied or auto-applied within the last 7 days (cooldown).
• Section had a proposal reverted/conflict in the last 14 days.

Self-restraint

It is correct and expected to finish a run with zero proposals. The optimizer's success is measured over weeks, not per run. The owner prefers "no action needed" to a low-confidence proposal. If you're unsure, skip.

Recording your decisions

Use POST /api/skill-curation/runs/<runId>/finalize's notes field to explain skips. Examples:

• "user-profile/routing-table: 3 signals but contradictory (one heading_add for ## Health, two heading_remove for same — likely an in-progress restructure). Skipping until next cadence."
• "today/section-shape: 1 structure_diff signal — below corroboration threshold."
• "context/file-responsibilities: section frozen since 2026-04-22 owner revert. Skipping."