メニュー
Interrupt controller converting polled to proactive agent execution. Per-runtime strategies, configurable thresholds, Deacon heartbeat supervision. Fights LLM passivity bias.
デフォルトでは、最初にソースを確認する Prompt が選択されています。直接コマンドに切り替えるか、ローカルコピーをダウンロードすることもできます。
インストールを決める前に、SKILL.md と SkillsMP に表示されている付属ファイルをお読みください。
SOC 職業分類に基づく
Codex または Claude でインストール この Prompt をコピーして Codex、Claude、または他のアシスタントに貼り付けると、Skill ページを確認してインストールできます。
直接コマンドでは確認用 Prompt が省略されます。実行前にソースを確認してください。
npx skills add https://github.com/Tibsfox/gsd-skill-creator --skill gupp-propulsionコマンドは1行のまま表示されます。コピー前に横へスクロールして全体を確認してください。
ローカルで確認しますか？SkillsMP が現在取得できるファイルをダウンロードできます。
SKILL.md を表示中
Use the moment a request is underspecified and you are about to act on an ASSUMED goal. Enumerate the candidate goals the request could mean, project the action or artifact each one implies, and check whether they diverge: if two or more plausible goals lead to materially different actions or artifacts, ask exactly ONE targeted clarifying question; otherwise proceed on the most-likely goal and state the assumption in one line. Clarification is an evidence-producing action, not a delay — but one question, never an interrogation. Distinct from gsd-spec-phase (GSD-phase-bound, emits a heavy SPEC.md) and intent-router (fetch strategy, not goal disambiguation). Backed by agent goal-state inference (arxiv 2606.16813v1). Triggers on acting under an assumed goal when the request admits more than one materially different reading.
Use at write time to vector/embedding memory — Grove content-addressed insertion, chroma/pgvector upserts, memory-consolidation promoting session traces to MEMORY.md, or embedding externally-ingested content. Scores each candidate record against a fixed panel of sentinel queries and quarantines any record that would become the nearest neighbor of too many unrelated queries — a hub — whether from adversarial poisoning or accidental over-generality. This is the memory-record-side sibling of skill-injection-guardian (file-side) and the write-side complement of memory-use-warrant (read-side). Quarantine, never silently drop; a human reviews. Backed by the admission-time hubness gate (arxiv 2606.19692v1). Triggers on inserting into vector memory, consolidating memory, and embedding stranger content.
Run this appropriateness check the moment you are about to integrate a retrieved long-term memory — a Grove content-addressed hit, a chroma/pgvector neighbour, a memory-consolidation digest, or a MEMORY.md line — into a response, especially anything touching private origins, Fox Companies IP, credentials, or Center Camp / consent-governed content. It answers a question intent-router never asks: not WHAT to fetch or HOW, but WHETHER a correctly-retrieved item should reach output. Relevance is not appropriateness — a perfect similarity match can still be a boundary violation. Default is FAIL-CLOSED: if in-context authorization is uncertain, the memory may inform behaviour but must not be surfaced. Backed by RBI-Eval (arxiv 2606.06055v1). Triggers on surfacing recalled sensitive memory into a response.
| name | gupp-propulsion |
| description | Interrupt controller converting polled to proactive agent execution. Per-runtime strategies, configurable thresholds, Deacon heartbeat supervision. Fights LLM passivity bias. |
| format | "2025-10-02T00:00:00.000Z" |
| version | 1.0.0 |
| status | ACTIVE |
| updated | "2026-04-25T00:00:00.000Z" |
| triggers | ["agent has work assigned to its hook and is not yet executing","stall detection identifies an agent with hooked work that has gone idle","new agent is spawned and needs GUPP enforcement injected into its context","runtime-specific GUPP strategy selection is needed","heartbeat supervision needs to be configured for active agents"] |
| references_subdir | true |
| word_budget | 800 |
The Gas Town Universal Propulsion Principle: If there is work on your hook, YOU MUST RUN IT.
GUPP is the interrupt controller of the Gastown chipset. In a real processor, the interrupt controller converts external events (keyboard press, network packet, timer tick) into CPU interrupts that force immediate handling. GUPP does the same for AI coding agents: it converts the polled model (agent checks periodically, waits for user input) into an interrupt-driven model (work appears on hook, agent executes immediately).
RLHF-trained LLM assistants have a strong bias toward waiting for the user — asking clarifying questions, seeking confirmation, pausing for feedback. This is appropriate in interactive chat but catastrophic for autonomous multi-agent orchestration, where every moment of idle waiting stalls the pipeline. Gastown discovered this empirically: agents spawned with a work item would introduce themselves, summarize the task, and wait, asking "Shall I proceed?" when the answer was already on their hook. GUPP overrides this trained passivity with an explicit, non-negotiable execution mandate: if you have hooked work, you do not wait, do not ask for confirmation, do not pause for feedback. You begin immediately. This is physics, not politeness.
This skill activates when:
GUPP enforcement adapts to each runtime via the Runtime HAL. The HAL exposes getGUPPStrategy() which returns one of:
hook_injection — Claude Code (SessionStart hook, highest fidelity)startup_fallback — Codex (tmux gt prime, medium fidelity)polling — Gemini / Cursor / unknown (prompt preamble + state polling)A Deacon heartbeat loop runs inside the witness-observer at the configured nudge_interval, detecting agents that received hooks but stalled. After restart_threshold, the watchdog requests a session restart. Per-runtime strategy tables, configurable thresholds, the full heartbeat protocol, and the Claude Code HAL note are in references/runtime-strategies.md.
These safety rules MUST remain in effect at all times — they are not advisory.
GUPP enforcement is ADVISORY within GSD's structured workflow. The GSD orchestrator's phase gates, verification steps, and checkpoint protocols take precedence over GUPP urgency. If a GSD checkpoint requires human verification, GUPP does not override it. If a phase gate blocks progress pending approval, GUPP does not force execution past the gate.
The hierarchy is:
GSD orchestrator (highest authority)
-> Phase gates and checkpoints
-> GUPP propulsion (execution enforcement)
-> Agent autonomy (lowest -- must obey all above)
GUPP operates between the orchestrator's structural gates. Within a phase, within an approved plan, within a committed task — GUPP demands immediate execution. But it never overrides the orchestrator's decision about what to execute or when a phase is complete.
The watchdog restart mechanism has strict limits to prevent infinite restart loops:
| Limit | Value |
|---|---|
| Max restart threshold | 1800s (30 min, hard cap) |
| Max restarts per bead | 3 |
| Restart cooldown | 60s minimum between consecutive restarts |
After 3 restarts for the same work item, escalation to human is mandatory.
After 3 restarts for the same bead without resolution, GUPP stops attempting automated recovery and escalates to a human operator via durable mail to the mayor. The mayor surfaces this through whatever channel the human is monitoring. GUPP does not attempt a fourth restart.
GUPP enforcement never destroys work in progress. Restarts preserve the agent's branch and commits. Nudges are non-destructive signals. Escalations are informational. At no point does GUPP delete files, reset branches, or discard uncommitted changes.
Reads runtime capabilities from runtime-hal (getGUPPStrategy(), thresholds, restart support). Targets polecat-worker (the primary enforcement target). Runs inside witness-observer (Deacon heartbeat). Sends nudges via nudge-sync and escalations via mail-async. Reads agent activity timestamps from beads-state and hook status from hook-persistence. mayor-coordinator receives escalations. sling-dispatch sets hooks that activate GUPP enforcement.
Implementation detail moved to:
references/runtime-strategies.md — per-runtime enforcement tables (Claude Code, Codex, Gemini, Cursor), threshold reference, Deacon heartbeat pattern.references/boundaries.md — full Safety Boundaries detail (advisory-in-GSD hierarchy, restart limits, escalation, no-data-destruction).references/metrics-and-learning.md — observable metrics (response time, nudge effectiveness, strategy success, stall frequency) and the learning feedback loop.references/gastown-origin.md — the GUPP principle from Gastown, how it fights model training bias.heartbeat.md — Deacon heartbeat supervision pattern and integration points.