Pine/deepseek-harness
1
0
Fork 0
Code Issues Pull Requests Actions 1 Packages Projects Releases 1 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
packages/skill/tool-skill/README.md
+15 -1
View File
@@ -36,7 +36,7 @@ Tool execution does not add a synthetic context message. Its freshly loaded resu
#### What the model sees
If model-invocable skills exist and this exact `skill` tool is visible, the agent receives the catalog template below as a durable user-role message before the first request, with one data-dependent entry per sorted skill. Later membership, description, or visibility changes append a complete replacement using the same `<available_skills>` envelope; deleting every skill appends an empty envelope with an explicit instruction not to use older names. The template's closing sentence is the seam rule against double-loading: the host's user-explicit `skill.invoke` injects the same `renderSkillContent` output (shared from `@deepseek-ai/dsh-skill`) inline, and the catalog tells the model to follow that block instead of re-loading the skill through the tool; the replacement-catalog template carries the same sentence in both arms, including the emptied catalog.
If model-invocable skills exist and this exact `skill` tool is visible, the agent receives the catalog template below as a durable user-role message before the first request, with one data-dependent entry per sorted skill. Later membership, description, or visibility changes append a complete replacement using the same `<available_skills>` envelope; deleting every skill appends an empty envelope with an explicit instruction not to use older names. The template's closing sentence is the seam rule against double-loading: the user-explicit gesture boundary (the pre-step listener below) injects the same `renderSkillContent` output (shared from `@deepseek-ai/dsh-skill`) inline, and the catalog tells the model to follow that block instead of re-loading the skill through the tool; the replacement-catalog template carries the same sentence in both arms, including the emptied catalog.
##### Skill catalog template
@@ -145,6 +145,20 @@ Only a failing call adds these retained tokens.
Append-only; newly visible content follows the reusable request prefix and does not invalidate existing KV-cache entries.
### User-explicit invocation injection
#### What the model sees
A whitespace-bounded `/name` token anywhere in a claimed user message, naming a user-invocable skill in the workspace catalog, injects that skill's full `<skill_content>` rendering (the exact result-template shape above) as a `user`-role instructions context appended after every other injection of that step — background first, the material to act on last. Only direct user input is scanned, the check runs on the loaded definition, and unknown or user-disabled names stay ordinary prose. This is the sole entry point for `disable-model-invocation` skills, which the catalog and the `skill` tool never expose; the catalog's closing sentence tells the model to follow the injected block instead of re-loading it.
#### Token effect
Each gesture adds one rendered skill body to that turn as injected context — the same size as the tool result for the same skill, paid deterministically at the user's request instead of at the model's discretion. Repeated gestures for one skill within one step inject once.
#### KV Cache effect
Append-only; the injection lands after the reusable request prefix inside the step's message batch and does not invalidate existing KV-cache entries.
## Known Limitations and Deferred Work
- **The catalog omits `whenToUse`, source, and provider metadata** — routing is based only on name and a capped description; `whenToUse` remains provider metadata and is not rendered by the loaded wrapper either.