Capture generic slash-command skill invocations

[dipree]

Problem

Prompt-based skill/prompt invocations were only captured for one fixed slash-command shape. Agents that invoke skills via ordinary slash commands (Codex, Cursor, OpenCode, Gemini, Factory) recorded no skill events, even though a leading /<command> is exactly such an invocation.

Solution

Record a prompt_invocation skill event for any leading /<command> in a submitted prompt.

This is safe because a slash command only reaches a recorded prompt if it submitted a model turn:

• Runtime/UI-only commands (/mcp, /model, /help, /clear, …) are handled in the agent REPL, never submit a turn, and never fire the turn-start hook — so they're excluded for free, no denylist needed.
• Pasted filesystem paths (/Users/..., /tmp/...) are the only non-command class and are rejected via a filesystem-root denylist. A bare root-name command (/dev) stays a command; only a root followed by a path segment (/dev/null) is rejected.

Details:

• Only the command token is stored — never prompt arguments (privacy).
• Pi's /skill:<name> is normalized to the bare name so it dedupes against Pi's native event.
• Built-in vs custom is just a label; both are injected prompts and recorded.
• Wiring is unchanged — handleLifecycleTurnStart already appends for every agent that exposes event.Prompt.

Behavior note: Claude Code sessions now also get prompt_invocation events for slash commands, alongside tool_invocation events from its native Skill tool (different event types, not deduped).

Testing

• go test ./cmd/entire/cli/agent ./cmd/entire/cli; mise run fmt + golangci-lint clean.
• Table tests cover built-ins, custom commands, :/path namespacing, arg-stripping, the Pi form, bare root-name commands, and path/non-command rejection.

🤖 Generated with Claude Code

[Copilot]

Pull request overview

This PR expands prompt-based skill invocation capture from the Pi-specific /skill:<name> shape to a generic “leading slash command” shape (/<command>), enabling Codex/Cursor/OpenCode/Gemini/Factory-style prompt skills to be persisted as prompt_invocation skill events (storing only the command token for privacy).

Changes:

• Add a generic prompt slash-command parser and event appender (prompt_slash_command), including filesystem-path rejection and Pi /skill: normalization for dedupe.
• Update TurnStart lifecycle handling to append the generic prompt-derived skill event only when the agent matches the session owner (or ownership is unset).
• Add unit tests for parsing variants and lifecycle tests for TurnStart recording + forwarded-hook dedupe.

Reviewed changes

Copilot reviewed 5 out of 5 changed files in this pull request and generated 1 comment.

Show a summary per file

File	Description
cmd/entire/cli/lifecycle.go	Appends a generic prompt-derived skill event on TurnStart after session ownership is known.
cmd/entire/cli/lifecycle_test.go	Adds lifecycle coverage for recording the prompt-derived event and avoiding duplicates from forwarded hooks.
cmd/entire/cli/agent/skill_events.go	Introduces the prompt_slash_command signal constant.
cmd/entire/cli/agent/skill_events_prompt.go	Implements parsing/normalization/denylist logic and dedupe for prompt slash-command skill events.
cmd/entire/cli/agent/skill_events_prompt_test.go	Adds table-driven tests for parser variants and adapter-vs-generic dedupe behavior.