Consider privacy-preserving context input events

[caioribeiroclw-pixel]

Hi — useful wrapper. One observability gap I keep running into with Claude Code/Cursor-style agent traces is that tool/model spans answer "what ran?", but not always "what context was loaded into the session before the model call?"

For headless/debug workflows, a privacy-preserving context input event would be useful alongside tool calls, tokens, cost, and timing.

A minimal event could avoid raw prompt/context content and record only categorical + hashed evidence, e.g.:

{
  "event.name": "context.input.loaded",
  "gen_ai.conversation.id": "...",
  "context.input.kind": "agent_instructions",
  "context.input.source.path": "AGENTS.md",
  "context.input.source.bytes_hash": "sha256:...",
  "context.input.delivered.hash": "sha256:...",
  "context.input.loaded_by": "native_file_discovery|hook|mcp|generated_fallback",
  "context.input.activation": "session_start|on_demand",
  "context.input.scope": "repo|workspace|session",
  "context.input.duplicate.suppression_policy": "suppress_equal_dedupe_key_within_scope"
}

Why I think this matters:

• Claude Code now has native OTel, and wrappers like this can make headless runs debuggable.
• Skills/hooks/MCP/context files can all add instructions before a model call.
• If the same AGENTS.md/skill/rule is discovered through two paths, normal tool-call traces may not show whether it was loaded once, duplicated, transformed, or clipped.
• Logging raw prompt/context is often too sensitive, but hashes + paths + categorical fields are usually enough to debug "wrong/missing/duplicate context" failures.

I put together a small public fixture that converts a local session JSONL into OTel-like context.input.loaded events without raw prompts, raw context, tool args, memory contents, or transcript bodies:

https://github.com/caioribeiroclw-pixel/pluribus/tree/main/examples/context-input-evidence

Question: would a small, opt-in context-input event like this fit claude_telemetry's scope, or is it better left to a separate pre/post-processor that enriches traces after the run?