| name | agent-libos-skill-navigation |
| description | Discover an applicable Agent Skill, activate its exact snapshot, read declared resources from that loaded snapshot, and unload it when its guidance or tools are no longer needed. Use when the required Skill is not already loaded or when a registered Skill must be found; never use Skill lifecycle operations to obtain Capability authority or bypass a denied tool. |
| allowed-tools | discover_skills activate_skill read_skill_resource unload_skill |
Navigate Skills
First inspect Loaded skills and visible schemas. If the exact Skill is already usable, do not rediscover or reactivate it: activation and unload are durable, audited, non-idempotent lifecycle mutations. Select the smallest direct match; a similar name is not an interchangeable implementation.
Tool guide
discover_skills
Use {"text":"git pull request","limit":8} only when the exact ID is uncertain or a registered Skill may apply.
- The result contains
skills, catalog_scope, and has_more. Applicable built-ins are a complete prefix and do not count against limit; text and limit bound only registered entries.
- Model-facing discovery merges applicable built-ins with registered entries; it does not scan directories for unregistered packages.
catalog_scope="builtin_only" means registry visibility was absent, not that no other Skill exists.
- Registered discovery needs
read on the configured skill:* registry resource. A limited grant can be consumed, so use a useful first-call limit. Missing authority silently narrows scope to built-ins.
- There is no cursor. If
has_more=true, refine text or raise limit within the Host maximum. If a later call changes to builtin_only, report the authority change instead of concluding that prior registered entries disappeared.
- Built-in summaries include
active and available_tools; registered summaries and the metadata-only prompt catalog omit active. Use Loaded skills as the primary state check.
- Discovery neither loads instructions nor grants tool, Capability, provider, or approval authority.
activate_skill
Pass exactly one returned or prompt-provided ID, for example {"skill_id":"agent-libos-jsonrpc"}.
A built-in activates only when the image already owns every declared static tool under the exact IDs. It atomically projects all of them into model visibility; it does not change the full tool table, add JIT, grant Capability, or need current-process skill:<id> authority.
A registered activation requires execute skill:<id> and can add static/JIT bindings. Resume an ASK decision rather than duplicating it. Reactivation atomically replaces its prior snapshot/bindings and retires superseded JIT.
Both persist an immutable snapshot/provenance plus audit/event evidence. authority_changed=false means Capability did not expand, not that no state changed.
Built-in success includes activation_kind="builtin_projection", authority_changed=false, maps/hashes, and empty jit_tool_ids. Registered success includes identity, maps, and hashes but omits those two fields. New guidance/schemas appear next turn.
Direct projection exposes each registered JIT schema. Multiplexed images expose only run_jit_tool; use the loaded contract and {"tool_name":"exact-name","arguments":{...}}. Never guess a hidden schema.
read_skill_resource
Read a package-relative resource after activation and before unload, for example {"skill_id":"example-skill","path":"references/protocol.md","max_bytes":8192}.
- Use an exact path from loaded instructions or
resources metadata. Absolute paths, .., and inventions fail; this reads neither workspace files nor the live registry package.
- Reads use the immutable loaded snapshot, unaffected by later registry replacement, and need no second
skill:<id> read grant.
max_bytes rejects rather than truncates. Omit it for the Host default or choose at least disclosed size_bytes; do not probe repeatedly.- Inspect
kind, size_bytes, and sha256. For kind="text", use content; for kind="base64", use content_base64. Exactly one payload form is useful.
- Built-in tool-Skill packages contain only
SKILL.md, so bundled resources normally belong to registered Skills.
unload_skill
Call {"skill_id":"example-skill"} only after its guidance/resources are no longer needed.
Built-in unload needs no current-process skill:<id> right. Registered unload separately requires execute skill:<id>. It removes that snapshot/bindings, retires JIT, and restores overlapping/base bindings. It never revokes Capability or reverses external effects.
The result includes activation_kind, removed_tools, and authority_changed=false. Empty removed_tools is normal for a built-in because image-owned full-table bindings remain.
Recommended workflow
- Check
Loaded skills and visible schemas. If the exact Skill is already usable, stop navigation and perform the task.
- Activate an exact catalog built-in directly; otherwise discover once with focused text and a sufficient limit.
- Select by exact ID, description, tools, source type, and package hash—not name resemblance.
- Activate once. Match returned
skill_id and any discovered package_sha256.
- Validate the tool maps. For a built-in, require
jit_tool_ids == {} and set(tool_names) == keys(tool_ids), aligned with available_tools. For a registered Skill, require set(tool_names) == keys(tool_ids) union keys(jit_tool_ids).
- Next turn, confirm its
Loaded skills instructions and read each schema before use. Read only required declared resources.
- Use the Skill to finish the user task. Keep it loaded across related steps; unload only when reduced prompt/tool exposure is useful and no later step depends on it.
- Recheck after transitions. Same-image fork/restore can preserve snapshots; fresh-child spawn or exec uses target-image defaults and may drop projections.
Failure and recovery
- Already loaded: do not reactivate; it rewrites lifecycle evidence and can replace registered JIT.
- Unknown ID or absent registered entry: refine discovery only when
has_more=true; otherwise report that no visible registered match exists. Do not scan paths or invent an ID through these tools.
- Unsupported built-in membership, tool-ID mismatch, invalid provenance, or hash mismatch: stop. Do not request partial activation or substitute an unverified global tool.
- Discovery/activation denial: request only exact catalog
read or skill:<id> execute; never duplicate pending ASK.
- Unknown activation settlement: inspect the next turn's
Loaded skills; for built-ins, a later discovery active value is also usable. Do not blindly repeat a non-idempotent activation.
- Missing/oversized resource: compare loaded metadata and correct the path or ceiling once; no filesystem bypass.
- Multiplexed JIT without a documented name/argument contract: stop and report the incomplete Skill package rather than guessing.
- Unknown unload settlement: verify loaded prompt state on the next turn. Schema presence alone is insufficient because the base image or another Skill may own the same binding.
- Process transition: reassess current loaded state and exact schemas; never rely only on what was loaded in the parent or pre-exec process.
Completion evidence
Navigation is complete when one exact Skill is loaded, identity/maps are consistent, guidance/schemas are visible, and required resources came from that snapshot with hashes recorded. Domain completion still follows that Skill.
After unload, require the exact Skill to be absent from the next Loaded skills section. For a built-in, discovery may additionally show active=false; registered summaries have no such field. Report removed_tools as full-table deletions only, not as a claim that every model schema or prior external effect disappeared.
