Coder/deepseek-harness
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

feat(tool-skill): inject user-invoked skills at the pre-step gesture boundary

Browse Source 
A whitespace-bounded /name token anywhere in a claimed user message,
naming a user-invocable skill in the workspace directory, now injects that
skill's renderSkillContent as instructions context appended after every
other injection of the step — the same agent/pre-step seam the catalog,
workspace instructions, and the runtime snapshot ride. Closed-set matching
mirrors the command registry (a miss stays plain prose), only user-source
messages are scanned, the policy check runs on the loaded definition, and
this is the sole entry point for disable-model-invocation skills. The
catalog's no-reload sentence now names the gesture boundary.
This commit is contained in:
Yichen Jiang
2026-08-08 13:14:49 +08:00
parent 7750789c8e
commit c08fa27e5c
10 changed files with 225 additions and 86 deletions
Download Patch File Download Diff File Expand all files Collapse all files

2
packages/skill/skill/README.md
View File

@@ -39,7 +39,7 @@ This package owns the `ctx.skills` interface. It does not know whether skills co
### Shared model-facing rendering
`renderSkillContent(skill)` renders one loaded skill as the canonical `<skill_content>` block (escaped `name` attribute, resource hints, verbatim body). It is the single truth for both loading paths: `dsh-tool-skill` returns it as the `skill` tool result, and the host's user-explicit `skill.invoke` injects it as a user message, so the model sees one shape regardless of who initiated the load. `escapeText` is exported beside it for consumers embedding prose in the same markup frame. The package also declares the `skill-invocation` `MessageSource` kind ({ name, args? }) that user-explicit injection stamps on its messages — transcript consumers present the invocation from this metadata instead of re-parsing the body.
`renderSkillContent(skill)` renders one loaded skill as the canonical `<skill_content>` block (escaped `name` attribute, resource hints, verbatim body). It is the single truth for both loading paths: `dsh-tool-skill` returns it as the `skill` tool result and injects it at the user-explicit gesture boundary, so the model sees one shape regardless of who initiated the load. `escapeText` is exported beside it for consumers embedding prose in the same markup frame. The package also declares the `skill-invocation` `MessageSource` kind ({ name, form: 'instructions' }) that user-explicit injection stamps on its messages — transcript consumers present the invocation from this metadata instead of re-parsing the body.
`isModelInvocable(skill)` and `isUserInvocable(skill)` read the matching positive field directly. `ctx.skills.get()` remains the trusted, policy-neutral loading primitive, so every user- or model-facing consumer must enforce the predicate that matches its surface before exposing or loading a skill.