Export privacy-safe context receipts from traces

[caioribeiroclw-pixel]

claude-tap already captures the most valuable layer for debugging coding agents: the real request/response traffic, diffs, tool calls, token usage, and portable JSONL/HTML artifacts.

One thing I think would be useful on top of that is an optional privacy-safe context receipt export: a derived artifact that proves which context sources were present/transformed/deduped/truncated without including the raw prompts, source code, memory contents, or transcript bodies.

Why this might fit claude-tap:

• the README positions claude-tap around inspecting exact context and sharing one portable artifact
• raw traces are excellent locally, but harder to share in issues/PRs because they may contain system prompts, repo code, secrets, memory results, or private transcript text
• a redacted receipt would let users say “this run loaded these context inputs / these chunks / this tool schema set” with hashes, counts, paths/URIs, and categorical metadata only

A minimal event/record could look like:

{
  "event": "context.input.loaded",
  "conversation_id": "...",
  "kind": "system_prompt|context_file|skill|tool_schema|mcp_result|session_history",
  "source": {
    "path": "AGENTS.md",
    "uri": null,
    "bytes_hash": "sha256:..."
  },
  "delivered": {
    "hash": "sha256:...",
    "estimated_chars": 1200,
    "truncated": false,
    "full_render_status": "available|unavailable_not_materialized"
  },
  "loaded_by": "claude-tap|client|mcp|hook|unknown",
  "activation": "upfront|tool_result|dynamic",
  "duplicate": {
    "dedupe_key": "...",
    "suppression_policy": "suppress_equal_dedupe_key_within_scope|keep_distinct"
  }
}

Non-goals / safety boundary:

• no raw prompt text
• no raw source code chunks
• no raw MCP memory contents
• no tool arguments/results unless the user explicitly opts into existing raw trace behavior
• no hosted telemetry requirement; local artifact only is enough

This would complement, not replace, the current raw trace viewer: raw trace for local debugging, receipt export for safer sharing and audit.

I have a small public fixture exploring the shape here, in case useful as prior art: https://github.com/caioribeiroclw-pixel/pluribus/tree/main/examples/context-input-evidence

If this is too far from claude-tap’s scope, that is useful signal too. In that case the cleaner split may be: claude-tap continues to record raw local traces, and an external post-processor converts JSONL traces into redacted context receipts.