Tool discovery
Harnez gives each task an immutable catalog snapshot. The catalog currently contains workspace tools and model-invocable skills. Discovery returns trusted metadata first, without putting every tool schema or skill body into the model context.
Core workspace tools are loaded when a task starts. The same discovery path supports capabilities that are not already loaded.
Discovery flow
The model-facing operations are:
| Operation | Result |
|---|---|
capabilities_list | Lists permitted tools and skills with bounded pagination |
capabilities_search | Searches permitted metadata with the lexical analyzer |
capabilities_inspect | Returns the validated contract for one capability |
tools_load | Admits a tool schema to task context and makes the tool callable |
skills_activate | Verifies and admits a skill body to task context |
A typical dynamic tool flow is:
search -> inspect -> load -> callSkills use search -> inspect -> activate instead. Loading or activation can
fail if the task's capability-context budget cannot fit the new content.
Search behavior
List and search return 20 results by default and accept a limit up to 100. Longer result sets use a cursor. Search normalizes the query, splits it into terms, and scores exact names, name terms, tags, descriptions, provider names, and capability kinds. Equal scores are ordered by canonical capability ID.
The analyzer version is returned as lexical-v1. There is no embedding lookup,
second model, or hidden semantic router. A search with no results only means
that the lexical index found no match.
Identity and authority
Discovery only includes capabilities that are marked model-discoverable and allowed by the task's grant. Tool authority follows this order:
none < discover < inspect < load < executeSkill authority ends at activate. A task can reduce its authority, but it
cannot raise the ceiling after it starts.
Every result carries a versioned reference with a catalog generation, contract
hash, and provider binding generation. Harnez checks that identity again before
loading or executing a tool. A changed contract or missing binding fails with
STALE_CAPABILITY instead of running a different implementation under an old
reference.
Skill discovery
A skill needs valid frontmatter with a name or id and a description.
Skills are model-discoverable by default. Set disable-model-invocation: true
to keep a skill available for manual use without exposing it to model
discovery. Harnez hashes the same bytes it uses to parse the manifest.
Activation reads the file again and verifies both the manifest and body hashes
before adding the body to task context.
Invalid manifests are reported as diagnostics and left out of model discovery.
See Skills for locations, frontmatter, and manual activation.
See Task runtime for task snapshots, cancellation, and capability-context lifetime.