skill: "Unable to load skill <name>" does not hint when the name is a registered agent

[Julfried]

Summary

When a model invokes the skill tool with the name of a registered agent
instead of a skill, the only feedback is Unable to load skill <name>. The
error does not distinguish "unknown skill" from "permission denied" from "this
name belongs to an agent". Because user @-mentions of agents look similar to
skill invocations, models conflate the two registries, and the opaque error
sends the conversation into a manual debugging loop instead of a one-step
self-correction.

Environment

• opencode version: v0.0.0-beta-18743 (opencode2, V2)
• OS: Linux 7.1.12-200.fc44.x86_64 (Fedora 44, x64)
• Terminal: ghostty (TERM=xterm-256color, COLORTERM=truecolor)
• Shell: /bin/bash
• Install/channel: beta
• Active plugins: none (no plugin config, no .opencode/plugin/)

Reproduction

1. Define a subagent at ~/.config/opencode/agents/my-agent.md with
   mode: subagent in the frontmatter (any name works; no skill with the
   same name is defined).
2. In a session, ask the model to do a task "using @my-agent".
3. Observe the model call the skill tool with name: "my-agent".

Expected Behavior

The skill tool error should be actionable. Ideally:

• Unable to load skill my-agent: unknown skill. "my-agent" is a registered agent; invoke it with the subagent/task tool. when the name matches an
  agent, or minimally
• Unable to load skill my-agent: no skill with this name is available in this session — i.e. distinguish "not found" from "permission denied", and point
  at similarly named registered entities (skills and agents).

Actual Behavior

All failure modes collapse into the same opaque message:

Unable to load skill my-agent

From packages/core/src/tool/skill.ts, unableToLoad(name) is used for both
the not-found path (skills.list() miss) and mapped load/permission errors,
so the model cannot tell why the load failed or what the correct tool would
have been. In practice this required a manual config investigation
(agents directory, skills list) to work out that the mentioned name was an
agent, not a skill.

Additional Context

• The user config was correct: mode: subagent agent definitions in
  ~/.config/opencode/agents/ are the intended mechanism, and the same
  session successfully invoked a different agent via the proper tool.
• The confusion seems easy to trigger because @-mentions of agents and
  skill names appear in similar prompt positions, and both registries are
  name-based.
• Related open issue: [FEATURE]: stop using Effect.die for skill-not-found in the skill tool #40091 (skill-not-found error handling uses
  Effect.die) touches the same code path, but is about crash-style failure
  handling rather than error-message quality / agent cross-referencing.
• Frequency: reproduced 3 times across one session (each failed skill load
  recovered only via manual fallback).
• Workaround: none within the session; the model fell back to manually
  executing the agent's instructions (loading the referenced skill and
  calling the repository CLI directly), which produced correct output but
  bypassed the subagent isolation.

[holny]

I'd like to take this. In tool/plugin/skill.ts, every miss in skills.get() collapses into one opaque failure, so agent names hit the same error as unknown skills. Plan: when the id isn't a skill, resolve it against the agent registry — if it matches an agent, fail with a hint telling the model to prompt the subagent instead; keep unknown-skill and permission failures distinct. Small change plus a focused test.

[holny]

Heads-up: the PR for this issue (#46940) was accidentally closed earlier today - its head fork got deleted unintentionally, which force-closed it. Re-filed as #48159 with the same commits. Sorry for the noise.