Decision Marker

Cockpit's second principle: surface every meaningful micro-decision before the agent commits to it. This solves spec.md §2 FM2 (mid-execution invisibility) for the 6 decision-worthy tools.

Behavior (handled entirely by hook)

For 6 decision-worthy tools (D3 decision):

• WebSearch — query selection
• WebFetch — URL + extraction prompt
• Task — subagent delegation
• Edit — file modification
• Write — file creation
• NotebookEdit — notebook modification

The PreToolUse hook intercepts the tool call and returns permissionDecision: "ask" with a Cockpit-formatted reason. Claude Code's built-in permission dialog displays the reason — user approves or denies.

Tools NOT in the ask list

These get silent allow (no Cockpit dialog):

• Read / Grep / Glob — exploration tools, decisions visible via thought block (F1)
• Bash — Claude Code's own permission system handles destructive ones
• AskUserQuestion / ExitPlanMode / TodoWrite — workflow tools, not decision points
• mcp__* — MCP server tools, behavior varies, v0.1 not intercepted

Why these 6, not all

Per Decision D3 (Risk-based filtering): Cockpit intervenes at "commit-direction" moments (research commit + write commit), not at exploration moments. Bash is excluded because Claude Code's existing permission system already handles destructive shell commands — Cockpit asking would create UI overlap.

Reason text format

🛂 Cockpit Decision Point — {tool_name}

Agent wants to:
  {formatted choice — query / url / file_path / etc.}

Approve this choice? (Choose 'no' to redirect agent.)

The hook extracts the relevant fields from tool_input (no agent declaration needed — <cockpit-decision> block was dropped per Q4=drop).

User actions

• Approve → tool call proceeds
• Deny → tool call blocked. v0.1 abort behavior; v0.2 will support redirect.

Behavior is hook-driven

Claude does not invoke this skill manually. It exists as documentation of Cockpit's principle. Actual behavior lives in hooks/pre_tool_use.py.