nc://toolaria

Toolaria: rescues oversized tool results before they flood the context, by spilling them to disk and handing the model a short preview plus a fetch handle. A Nodecode rewrite of Sahil-SS9/Toolaria by Sahil-SS9.

Get it

Tell the agent "add the toolaria cell", or run nodecode add nc://toolaria.

Set it up

Nothing is required. The section is toolaria in ~/.nodecode/config.jsonc, and every field has a working default:

{
  "toolaria": {
    "max_result_chars": 12000,     // a result longer than this is rescued
    "head_lines": 40,              // lines the preview keeps from the start
    "tail_lines": 15,              // and from the end
    "excerpt_max_chars": 8000,     // the hard cap on the preview
    "fetch_max_chars": 4000,       // the cap on one range or grep answer
    "full_fetch_max_chars": 50000, // full mode refuses a blob over this
    "store_path": "~/.nodecode/toolaria",
    "ttl_hours": 72,               // older blobs are swept at start
    "keep_out": ["rescuer_fetch", "eval"]
  }
}

keep_out names tools whose results are never spilled. eval is there by default because its output is usually the thing the operator asked to see, and the fetch tool's own answers must not be re-spilled.

Use it

Nothing to call: the cell hooks the core's :TOOL point, so every tool result longer than max_result_chars comes back as a preview plus one handle line:

[Toolaria: tool result rescued. tool=web_extract; at=2026-11-08 04:12:31; size=171,874 chars; lines=1,994; blob=5ddff67d3a1b]
Preview (first 40 / last 15 lines); this is a preview, NOT the full output:
...
Retrieve more with rescuer_fetch(id="5ddff67d3a1b", mode=...): outline | search(query) | grep(pattern) | range(start,count) | stat | full

The model then reads only the slices it needs, with rescuer_fetch:

/rescuer answers the store's line: how many blobs, how much disk, where they are, the thresholds, and what is kept out. (help :toolaria) carries the same in the model's manual.

Blobs are content-addressed - the first 12 hex characters of the text's SHA-256 - so the same result rescued twice is one blob, and a handle stays valid until the blob is swept after ttl_hours.

What the rewrite leaves out

Pass-by-reference (tla:<id> expanded inside another tool's arguments), which needs a host middleware this contract does not expose; embedding search, so search ranks lexically and the cell needs no Python or model download; the optional at-rest encryption and its key rotation; hot-blob pinning and the size-budget sweep; the chained <id>@N grammar and the expansion audit ledger. NOTICE says the same, in the original's terms.

Test

.github/test-cell.sh tools/nodecode-toolaria

Offline: every test starts the cell against a store in the temporary directory, pushes a result through the installed hook, and reads it back through every fetch mode.