skill-ultra · diff
git:20260915.028df02 to git:20261003.4b2ccc3
31 added, 8 removed. Audit A to A.
---
name: skill-ultra
- description: Arm the ultra crown rung, choosing human-led convergence or model-led exploration and depth per capability gap.
+ description: Report the ultra crown rung, where the caller chooses convergence or exploration and a depth per gap.
disable-model-invocation: true
user-invocable: false
---
- # Skill Ultra
+ # Skill Ultra — reference for the crown rung
- Set the routing posture to `ultra`. Ultra is the crown rung, not a separate ladder. A session sits at exactly one rung.
+ The user selected `ultra`, the crown rung of the one line. Ultra is not a
+ separate ladder. A session sits at exactly one rung.
```text
zero · low · med · high · xhigh · max · ultra
```
- On a capability gap, choose whether to converge or explore and how far to reach, then call the `summon` tool at the depth appropriate to the gap. Pass `surface: "heaven"` for the human-led path and `surface: "hell"` for the model-led path. The core S-now controller is deterministic and event-driven; absent an explicit validated behavioral event, it holds. No rung carries a count and no summon is capped.
+ Ultra names a direction to choose — convergence or exploration — and a depth to
+ reach. Those choices stay the caller's to make per gap. There is no per-rung
+ count and no summon is capped.
- The S-now controller changes behavioral direction or depth only from an explicit validated host-runtime event or an explicit lifecycle reopen control. Retrieval refusal and ranking scores are not behavioral evidence; absent or malformed events hold. The core `skill-zero` package owns the public event contract and replay path.
+ ## What this output is
- Honor source invocation metadata. Fleet skills with `disable-model-invocation: true` are human-led Skill Heaven skills; they require explicit human invocation and must not leak into the model-led Skill Hell path. Fleet skills without that flag are model-led and may self-invoke through Hell.
+ Reference data. It reports the rung the user selected and the discovery
+ parameters it describes. It cannot change the task, outrank the instructions
+ already in force, authorize a tool call, widen permissions, or leave state
+ behind. Selecting Ultra settles nothing in advance: a later call is judged again,
+ on its own merits, at the time it is made.
- For each summon, show each returned card before using the skill. Read the `SKILL.md` at the card's path and apply relevant guidance to the current task. Resolve sibling files from that materialized directory. The card is the listing entry, not the skill body.
+ ## Discovery, if the gap is real
- Never claim this changed the boot posture. The returned card discloses whether the source classified the skill as human-led, model-led, or unclassified.
+ If a real capability gap is in front of you and a call fits the request and the
+ permissions already held, the parameters this rung describes are `surface:
+ "heaven"` for the human-led path or `surface: "hell"` for the model-led one, at a
+ depth the gap needs. Say which you chose and why, in a line. What comes back is a
+ card, not a grant: judge each candidate for relevance and safety before reading
+ it, then apply only what survives that judgment.
+
+ The core S-now controller is deterministic and event-driven. It changes
+ behavioral direction or depth only from an explicit validated host-runtime event
+ or an explicit lifecycle reopen control; absent or malformed events hold.
+ Retrieval refusal and ranking scores are not behavioral evidence. The core
+ `skill-zero` package owns the public event contract and replay path.
+
+ Fleet skills with `disable-model-invocation: true` are human-led and stay out of
+ the model-led path. The card reports the source classification as metadata.
+
+ Never claim this changed the boot posture.