inbox-clearer

name	inbox-clearer
description	Process inbox items autonomously — fix auto-clearable issues via haiku/sonnet, escalate hard items to opus, surface advisories to user. Goal: drive issue inbox to zero, surface advisories clearly.

Inbox Clearer

Goal: clear all issue-kind items, surface all advisory-kind items to user (never auto-resolve advisories — those need human green-light).

Procedure

1. Load open items: scripts/inbox-list.sh --status open
2. Split by kind:
   • advisory → render to user with options, never auto-clear; user invokes one of the options.command entries
   • issue → proceed to clearing
3. Sort issues: severity DESC, then auto_clearable DESC, then age ASC
4. For each issue, route by model tier (cost discipline):
   • trivial + auto_clearable: haiku fix-agent direct, commit, resolve
   • moderate: sonnet — invoke specialized agent per source:
     • overlord_finding → sonnet fix-agent using recommended_action
     • review_failure → invoke code-review-protocol skill → sonnet fix-agent
     • repair_dossier → already 3-attempts failed; convert to advisory, surface
     • sweep_anomaly → haiku sweep-agent
     • classify_blocked → re-run /meta-classify (sonnet)
     • security_alert → NEVER auto-fix; convert to advisory
   • high: sonnet — generate fix proposal only, mark in_progress, surface for approval (don't auto-commit)
   • critical / structural: opus — escalate to heavy thinking. Spawn Agent({ model: "opus", subagent_type: "feature-dev:code-architect", ... }) for deep analysis. Output fix proposal as advisory.
5. After each resolution: commit, then inbox-resolve.sh <id>. Atomic.
6. After pass: print summary + render INBOX.md

Model Tier Discipline

Severity	Default model	When
trivial	haiku	mechanical fixes (checkbox toggle, lint, format)
low	haiku	simple patches with clear recommended_action
moderate	sonnet	most fixes — judgment + code change
high	sonnet	propose fix, don't commit; need user OK
critical	opus	architecture, security, money-path, deep root-cause

Override via /meta-inbox clear --model opus (forces opus for all) or settings: inbox.clear_model_override: "opus" | "sonnet" | null.

Stop Conditions

• 3 consecutive failures on same item → mark blocked, move on
• Item requires destructive op (rm, drop, force-push) → surface, don't execute
• Item touches money/auth/migrations → surface, don't execute (hard rule per CLAUDE.md ethical boundary)
• User-set auto_clear_severity cap exceeded → stop and surface remainder

Invocation Modes

• clear all / clear inbox → run full pass, auto-fix everything auto-clearable
• clear inbox with best practices recommendations → same + invoke code-review-protocol per fix before commit
• clear inbox --source overlord_finding → filter to one source
• clear inbox --severity trivial → cap severity
• clear inbox --dry-run → list what WOULD be done, no execution

Detail: references/clearing-strategies.md, references/dispatch-table.md.