[FEATURE]: Pointer-based retrieval for compacted tool results #22407

Feature hasn't been suggested before.

• I have verified this feature I'm about to request hasn't been suggested before.

Describe the enhancement you want to request

After compaction, cleared tool results are replaced with [Old tool result content cleared]. The original content persists in the database (part.state.output), but the model has no way back to it. Any question requiring a specific detail from a cleared result gets answered from whatever the summary preserved, or hallucinated.

This gets worse with cascaded compaction. Each summarization round is a lossy pass - by the second compaction, specific details from early in the session are effectively gone.

Give the model two retrieval tools for compacted content:

1. list_refs() — returns all compacted parts for the current session with IDs, titles, and sizes
2. read_ref(id) — returns the full original part.state.output for a compacted part

And emit pointer IDs where cleared content used to be:

 // Before:
 const outputText = part.state.time.compacted
? "[Old tool result content cleared]"
: part.state.output

// After:
const outputText = part.state.time.compacted
? `[ref:${part.id} — ${part.state.title}]`
: part.state.output

Implementation cost is low: one line changed, ~80 lines of new code, no new database tables or infrastructure. The content is already stored.

I ran a series of experiments testing this against summarization-only compaction. Full methodology and experiment repo: https://github.com/joshuaisaact/pointer-experiments Writeup: https://joshtuddenham.dev/blog/compaction/

Single compaction (simulating opencode's 60% pruning rate):

Cascaded compaction (x2):

Summaries of summaries lose specific details.

Pointers don't degrade because the stored content is lossless.

Related issues:
#16512, #20246, #16960 identify related symptoms (detail loss, instruction loss after compaction). Those propose better summaries or re-injecting instructions. This proposes making the original content retrievable.