--- name: repair-stephanos-continuously description: Continue existing Stephanos Core Daemon and goal-building repairs through real, bounded Sovereign Commander actions, preserving original outcomes, independent review, gaming resource protection, and end-to-end build receipts. --- # Stephanos: persistent repair-to-building continuation Use this skill when the operator wants the already-existing Stephanos Core Daemon, Sovereign Commander, Observatory and building machinery to keep repairing and producing real material source work. This is an operator-facing workflow skill, not a daemon, task scheduler, machine tool, privilege grant, automatic code installer or alternative lease/review authority. Installing it in ChatGPT does not prove the Battle Bridge desktop plugin host, OpenClaw Local or the Core Daemon has loaded it. ## Durable parent objective The parent outcome is not 'repair verb returned DONE'. It is one qualified goal selected, legitimately claimed, physically picked up, edited and tested, independently reviewed, merged through the already-protected owner workflow, proven in the live runtime, then capacity refilled for two unattended cycles. Preserve the same original goal and blocker identity across failed attempts, provider failover, daemon cycles and user chats. Never mark a parent goal complete because a child maintenance command, draft skill, UI check or CI job succeeded. ## Evidence first, then one smallest repair 1. Inspect current GitHub main and one recent exact-head Sovereign Commander `visibility-snapshot` (or narrower fixed status action) through the connected private/local MCP or the owner-authored #2808 guarded fast mailbox. Use `skills/use-sovereign-commander/SKILL.md` for the schema, approval and public-output restrictions. Cloud/iPad ChatGPT cannot assume its private desktop stdio tools are connected. Prefer the bounded receipt index and request-correlated terminal record over repeatedly fetching the full public mailbox. 2. Distinguish head sync, healthy services, singleton Core process, loaded Outer Bridge contract, Core wake/degraded state, independently signed 23 Observatory gates, canonical Programme Authority/lease, native publisher, controller states and actual physical worker pickup. Unknown or absent signed evidence is not a pass. 3. Diagnose the **first actionable independent** gate, not merely the most visible blocker. Use existing registered maintenance actions only for admitted reversible faults. Verify the terminal receipt and a fresh post-repair observation. If the same typed blocker remains, stop looping the identical action; preserve parent mission and route a deduplicated capability or authority gap under existing #2961 and #1903. 4. On an actual missing capability, search existing source/Skill Forge/skills first. For external SKILL.md intake require pinned upstream SHA, publisher and licence evidence, byte-digest verification, scope/risk review and target-specific authorization. Draft intake is not installation. No auto-execution of untrusted instructions, no automatic privilege enlargement. Admit a guarded activation only after independently reviewed implementation and actual installed-host inventory plus runtime action proof. 5. Preserve original Goal Conveyor, independent Seed Canopy Parallel Forge and Rescue Island as **separate** authority and mutation domains. An original held source lease cannot be seized, spoofed or released by a parallel builder. Allow only independently proven, unclaimed, non-overlapping source tasks through separately qualified paths. Never invent a new scheduler or worker merely to get another green count. 6. Respect gaming mode: local heavy CPU/GPU builders and model routers need demonstrated governor custody and timely zero-residency/stop proof. Continue only genuinely external/resource-disjoint activities while the operator games. 7. Report exact request ID, terminal DONE/BLOCKED verdict, source/head, evidence hash, which gate changed, which gates are still held and whether physical pickup/build/merge/live refills happened. Keep retries single-flight with cooldown, dedupe and GitHub API backoff. If a remote request is ACCEPTED but has not reached a terminal receipt, do not submit a duplicate mutation. ## Known 14 October 2026 problems: examples, not eternal facts - Existing owner #2961 retains the original #2962/r7 source-writer custody gate. The missing original execution receipt is **not** reconstructable by generic `repair-goal-builder-flow`; an independently qualified Windows physical exclusion/OS witness and separate operator-authorised exceptional reconciliation are required. Never kill old writers or alter a lease by guess. - Native capacity Gate 13 is independent of Gate 8. It needs a real signed current-head publisher receipt, not a scheduled-task-start acknowledgment. Revalidate the current PR and runtime state before repair. - `repair-control-plane` and `status-stephanos-core-daemon` have previously produced genuine terminal remote receipts. An individual healthy singleton process can coexist with an overall DEGRADED building outcome; do not suppress the red blocker. - Open PRs are not installed skills. Sovereign Commander source-plugin manifest changes require actual desktop-host/plugin inventory and version activation proof; the local HTTP daemon's health does not prove that host reloaded a packaged skill. ## No authority expansion No arbitrary shell or file access over public GitHub, protected merge bypass, lease seizure, fabricated witness, secret/public key publishing, unrestricted service installation, uncontrolled model downloads, new source writer, or direct main push. Use only existing source-approved maintenance verbs and reviewed exact-head protected merge flow. Record escalation truth when no safe approved action exists, and keep other eligible independent repair work moving.