Summary (deferred — post-v1)
The write side of agent skills — create_skill / use_skill invoked mid-turn — complementing E7 (#246), which is the read/match/load side. This is a model-authors-its-own-tools surface, so it needs a safety boundary.
Framework scope vs app scope (SO-1)
- Framework provides only a validation / approval-gate contract — a seam where a proposed skill (authored by the model at runtime) is validated and gated before it becomes loadable/executable.
- The authoring UX and the persistence of authored skills are app-owned (SO-1 keeps skill management out of the framework). The framework does not ship the
create_skill tool implementation or its storage.
Why gated
A model writing and then immediately using a new tool is a trust/safety escalation. The approval-gate contract is the framework's only responsibility here: it must be possible to require human/app approval before an authored skill is trusted.
Scope
@netscript/ai skills seam (approval-gate contract only). wave:defer (post-v1). POC reference: eis-chat #69 (lib/skills.ts + builtin-skills.ts, feat/agent-skills; safety/trust gate needed).
Part of #238 · Refs #246 (E7) · POC eis-chat#69
Summary (deferred — post-v1)
The write side of agent skills —
create_skill/use_skillinvoked mid-turn — complementing E7 (#246), which is the read/match/load side. This is a model-authors-its-own-tools surface, so it needs a safety boundary.Framework scope vs app scope (SO-1)
create_skilltool implementation or its storage.Why gated
A model writing and then immediately using a new tool is a trust/safety escalation. The approval-gate contract is the framework's only responsibility here: it must be possible to require human/app approval before an authored skill is trusted.
Scope
@netscript/aiskills seam (approval-gate contract only). wave:defer (post-v1). POC reference: eis-chat #69 (lib/skills.ts+builtin-skills.ts,feat/agent-skills; safety/trust gate needed).Part of #238 · Refs #246 (E7) · POC eis-chat#69