anti-ceremony
Guard outcome work against process overhead. Triggers: RPI pre-Plan guard; explicit "full anti-ceremony audit" requests.
git clone --depth 1 https://github.com/boshu2/agentops /tmp/anti-ceremony && cp -r /tmp/anti-ceremony/skills-codex/anti-ceremony ~/.claude/skills/anti-ceremonySKILL.md
# Anti-Ceremony Keep work tied to the caller-visible outcome and the proof still needed for it. The default guard is a single, artifact-free judgment before RPI Plan. ## Modes | Trigger | Mode | |---|---| | RPI pre-Plan invocation or ordinary use | Quick guard | | Explicit request for a full anti-ceremony audit | Full honesty audit | Never promote the quick guard into the full audit without the explicit trigger. ## Quick guard 1. Freeze the caller-visible outcome in one short statement. 2. Park proposed process work that has no concrete consumer, gated subject or release decision, observed defect, or retirement condition. 3. Name the implementation or independent proof still missing. 4. Name the condition that ends this traversal. 5. Return `STOP` when the traversal would only create control artifacts or the frozen outcome is already sufficiently proved; otherwise return `CONTINUE`. Perform the guard exactly once in memory. Write no worksheet, ledger, capture, dashboard, prompt packet, or other hidden artifact. Parking is a field in the response, not a tracker or delivery mutation. Return this shape: ```yaml decision: CONTINUE | STOP reason: <exactly one sentence> frozen_outcome: <nonempty string> parked_process_work: [<work item>, ...] remaining_proof: [<proof item>, ...] stop_condition: <nonempty string> ``` The two lists contain strings and may be empty. Done when every field is present and the decision explains whether one RPI traversal earns its cost. ## Full honesty audit Run this mode only when the caller explicitly requests it. Examine each named control artifact or ceremony step for its consumer, gated subject or release decision, observed defect, and retirement condition. Return concise keep, park, or drop findings in the response; create no audit artifact unless the caller separately requests one. ## Boundaries The guard makes one admission judgment. It does not retry, repair, replan, dispatch core phases, operate Git or a tracker, deliver, release, or select a next action. On malformed or missing inputs, return `STOP` with the missing field named in the one-sentence reason and stop.
Switch a caller-selected coding-agent Triggers: "switch account", "rotate coding-agent account".
Use when operating ACFS flywheel health checks, init, and agent loop tooling from ~/acfs/bin/acfs.
Use Agent Mail as an optional messaging and Triggers: "coordinate writers", "reserve files".
Operate explicit orchestrator, implementer Triggers: "agent-native factory", "role-shaped agent panes", "persistent workers".
Run AGY headlessly via scheduled ticks or `agy -p`, capture agentapi JSONL evidence, and validate automated AGY loops or event streams.
|-
Use an explicitly selected AGY runtime for Triggers: "agy", "antigravity", "AGY evidence".
|-