3.9 KiB
| id | date | status | supersedes | superseded-by | affected-paths | affected-components | ||||
|---|---|---|---|---|---|---|---|---|---|---|
| 0039 | 2026-07-14 | Accepted |
|
|
0039 — Doc-hygiene deletion autonomy is tiered on evidence quality and recoverability, not file type
Context
Lifecycle rules let os-doc-hygiene:clean delete files (true git deletion in a dedicated hygiene commit; git history is the archive — ticket #35). The question (ticket #43) was what deletes silently versus what requires confirmation. Re-confirming every run would reduce the rulebook to an advisory list and strip :calibrate of its payoff; but an LLM judgment must never silently destroy a file, and untracked or dirty files have no history to recover from. /os-adr:find confirmed no existing ADR governed clean's safety tiers.
Decision
Rule-backed deletes are auto — a rule confirmed per the strong-model gate (#39) is the standing consent. The auto/confirm line is drawn on evidence quality plus recoverability: tracked + clean worktree + deterministic signal = auto; tracked-but-dirty or untracked = confirm (verified at runtime via git ls-files + dirty check, never trusting the rule); classifier-judged served_when signals are ALWAYS forced to confirm regardless of tracked status, while scanner-proven served_when_path signals may delete silently. The temporary tier defaults to retain_recent 3 + max_age_days 3, age measured from git commit time falling back to mtime for untracked entries (directory-inode mtime, no recursive walk — #48). graphify-out/** and .dochygiene/** form an explicit IGNORE surface that is never walked — the ignore surface is its own list, never inferred from .gitignore (gitignored ≠ deletable, gitignored ≠ keepable). A per-rule confirm:true escape hatch exists but is human-settable only; a model-proposed rule may only ask a human to set it.
Consequences
Routine hygiene runs unattended on the safe population while every non-recoverable or judgment-based deletion stops at a human. The auto/confirm line moves exactly where evidence quality changes, so improving a rule's evidence (see the determinism-promotion ADR) directly buys silence. 3-day/retain-3 retention is aggressive by design; the newest 3 entries per rule always survive to show current trajectory. Full tier matrix: plugins/os-doc-hygiene/lifecycle-spec.md.
Alternatives rejected
(1) Confirm everything — rejected: rulebook becomes advisory; confirm-fatigue makes the human rubber-stamp. (2) Tier by file type or a recoverable_via/regenerate class — rejected: recoverability is a property of git state at runtime, not of the rule's claim. (3) 90-day retention windows — rejected as far too slow in the age of AI. (4) mtime-only age — rejected: clone/branch-switch resets every mtime, silently putting the whole rulebook to sleep. (5) Model-settable confirm:true — rejected: used too liberally, everything drifts back to always-confirm.
Amendment (2026-07-15 — wayfinder map #49)
Tier interactions decided on map #49: (1) the "class, never path" rule-quality test is relaxed for the keep tier only — exact-path singleton keep rules are allowed (the test exists to prevent bad deletion rules; a singleton keep merely protects) and leave the rulebook only by hand-deletion, never an automated revisit: an automated revisit would be the one place a machine argues a keep back toward deletion, the exact failure mode the keep tier prevents. Instance globs stay forbidden for temporary/delete-once-served. (2) Nominations memory (ADR-0038 amendment) is memory, not deletion authority: rejections and consults never filter files and are not individually gated at persistence; rules alone authorize deletion. (3) Consults resurface in :calibrate only — :check/:clean are unchanged. Detail: plugins/os-doc-hygiene/lifecycle-spec.md §2/§8; decision tickets #51, #52, #56.