Feature idea: context coverage summary for long transcripts

[caioribeiroclw-pixel]

This is a feature idea for the generated transcript index, not a bug report.

One thing I keep wanting when reading long Claude Code transcripts is a quick way to answer:

did the context Claude loaded/read actually influence the plan or edits, or did it just pass through the transcript?

The HTML view already makes the session much easier to inspect. A small "context coverage" / "attention" summary could make it useful for debugging a common failure mode in coding-agent sessions: retrieval/search/read succeeded, but the assistant ignored the important context when planning or editing.

Possible shape, deliberately heuristic and opt-in:

• list large context-delivery events (Read, Grep, Glob, MCP/search/tool outputs, pasted docs, queued user prompts once Missing queue-operation prompts #98 is handled)
• show whether each delivered item is referenced later in assistant text, plan/todo text, tool arguments, or edited files
• flag likely gaps, e.g. read_file:docs/auth.md delivered but no later mention/citation/path reference before edits to auth code
• expose coarse counts on the index page: delivered_context, referenced_later, unreferenced_before_first_edit, edited_files_with_prior_context
• keep it non-authoritative: "attention hints", not proof; no model judgment required

A concrete UI could be a collapsible section on index.html:

Context coverage (experimental)
- 14 files/tool results delivered
- 9 referenced later in assistant text/tool args
- 3 read before edits but never referenced
- 2 edited files had no obvious prior read/search context

Why this feels useful: a transcript viewer can become more than a chronological archive. It can help reviewers spot the boundary where a coding agent had the right context available but failed to use it, which is different from a retrieval failure.

No expectation that this should block current work; filing because it seems adjacent to search/filtering and summary features (#12, #61, #69), but with a more review/debugging-oriented goal.

[Necmttn]

This would be most useful if every hint links back to evidence.

For each delivered context item, keep source kind, tool_use_id or message uuid, file/path/query, byte or line range when available, first delivered turn, first referenced turn, and first edit turn. Then the UI can show unreferenced before edit as a traceable heuristic, not a model judgment.

That makes false positives inspectable instead of mysterious.

---

Generated with ax.