Scope skill completion to the active agent and include marketplace-installed skills

[NightWatcher314]

Problem

HAPI's $skill completion currently appears to aggregate skills from multiple agent ecosystems (for example ~/.agents/skills, ~/.claude/skills, and ~/.codex/skills) regardless of which agent/flavor is active.

This can surface skills that the current agent cannot actually use. Example: in a Codex session, HAPI suggests $davinci-resolve-scripting because it exists under ~/.claude/skills/davinci-resolve-scripting, but the active Codex environment does not list that skill in its available skills.

This makes completion misleading: selecting a suggested skill can produce a prompt reference that the target agent will not recognize.

Expected behavior

Skill completion should be scoped to the active agent/flavor:

• Claude Code sessions should only suggest skills that Claude Code can actually load/use.
• Codex sessions should only suggest skills that Codex can actually load/use.
• Shared/project skills should still be included when they are valid for that agent.

In other words, the completion list should reflect the effective skill set of the current agent, not every skill directory HAPI knows about.

Additional request: marketplace-installed skills

Skills installed through a marketplace should also be discoverable by the corresponding agent integration:

• If a skill is installed for Codex through the marketplace, Codex sessions should include it in $ completion.
• If a skill is installed for Claude Code through the marketplace, Claude sessions should include it in $ completion.

The important point is that marketplace-installed skills should be included in the same agent-scoped resolution path, rather than being either invisible or globally mixed into every agent's suggestions.

Possible implementation direction

Add agent/flavor-aware skill root resolution, e.g.:

• Common/project roots, where applicable:
  • <cwd>/.agents/skills
  • repo ancestor .agents/skills
• Claude-specific roots:
  • ~/.claude/skills
  • marketplace roots used by Claude Code
• Codex-specific roots:
  • ~/.codex/skills
  • /etc/codex/skills
  • Codex system skills under ~/.codex/skills/.system
  • marketplace roots used by Codex
• Generic user roots:
  • ~/.agents/skills, if intended to be shared across agents

Then have /sessions/:id/skills pass the current session flavor/agent into the skill listing RPC, so the CLI lists only roots applicable to that session.

Why this matters

This avoids false-positive completions and makes $skill suggestions trustworthy: every suggested skill should be usable by the current active agent.

[github-actions]

Confirmed from the current code: skill completion is not agent/flavor-scoped today. The web route resolves the session, but calls engine.listSkills(sessionId) without passing session.metadata.flavor; the hub RPC then sends an empty listSkills request to the CLI. On the CLI side, listSkills() currently merges project/user roots for .agents, .claude, and .codex, plus /etc/codex/skills, then dedupes by skill name. That matches the false-positive behavior described here.

I did not find marketplace-specific skill roots wired into the skill completion resolver. The current resolver appears to use fixed filesystem roots only, so marketplace-installed skills would need to be added through the same flavor-aware root selection path.

Relevant code:

• hub/src/web/routes/sessions.ts:530 - /sessions/:id/skills endpoint.
• hub/src/web/routes/sessions.ts:543 - calls engine.listSkills(sessionId) without flavor/agent context.
• hub/src/sync/syncEngine.ts:891 - listSkills(sessionId) forwards only the session id.
• hub/src/sync/rpcGateway.ts:221 - sends RPC_METHODS.ListSkills with {} request payload.
• cli/src/modules/common/handlers/skills.ts:8 - CLI handler ignores request fields and calls listSkills(workingDirectory).
• cli/src/modules/common/skills.ts:24 - user roots include ~/.agents/skills, ~/.claude/skills, and ~/.codex/skills together.
• cli/src/modules/common/skills.ts:37 - project roots include .agents/skills, .claude/skills, and .codex/skills together.
• cli/src/modules/common/skills.ts:158 - resolver reads project, user, and admin roots, then merges/dedupes.
• cli/src/modules/common/skills.test.ts:61 - existing tests assert merged .agents + .claude behavior.
• cli/src/modules/common/skills.test.ts:72 - existing tests assert .agents + .codex + Codex system skill behavior.
• web/src/hooks/queries/useSkills.ts:42 - web completion data comes from api.getSkills(sessionId).

Need more info:

• Exact marketplace-installed skill directories for Claude Code and Codex, if they are not already under the agent-specific roots above.

---

HAPI Bot