nodecode-tameru
Tameru (貯める, to save up): big tool results, kept small, nothing lost — the lines that matter and every distinct error, with the exact original one tool call away. A Nodecode rewrite of 0xWhiteMage/hermes-tameru-plugin by 0xWhiteMage.
A test run, a build log, a directory listing: most of a long tool result is
noise, and it rides every later request at the model's expense. This cell
shortens such a result on its way into the context — the first lines, the
last lines, and the first line of every distinct error template, with one
marker line where a run of the middle went — and keeps the exact original, so
nothing is lost behind the shortening: tameru_expand hands any of it back,
whole, by line range, or by a regex.
| Surface | What it does |
|---|---|
| the hook | A tool result of min_chars characters or more comes back shortened, its first line ending in a [tameru: kept … · ref=…] header. A result that repeats one template forty times is one exemplar and its count ([×40] …). A result holding a probable credential is shortened but not kept. |
tameru_expand |
Reads the original back by the ref its header names: all of it, only start_line..end_line, or only the lines matching grep, up to max_chars. It says when a ref is unknown — the store is memory only, so a restart forgets it. |
/tameru |
What shortening has saved and which originals are held; /tameru expand REF prints one here; /tameru reset forgets both. |
Everything is local and deterministic — no model call, no network, no child process, nothing written to disk.
Get it
Tell the agent "add the tameru cell", or run nodecode add nc://tameru.
Set it up
Nothing is required. The section tameru in the config tunes it; every
member may be left out.
| Member | Default | What it does |
|---|---|---|
min_chars |
4000 | the smallest result worth shortening |
head_lines |
8 | lines kept from the start |
tail_lines |
8 | lines kept from the end |
max_line_chars |
300 | the longest one kept line may be, folded in its middle |
error_lines |
40 | how many distinct error lines are kept |
store_limit |
16 | how many originals tameru_expand can read back |
skip_tools |
— | more tool names whose results are never shortened |
Use it
It works while the model works: a big result arrives short, and the model
may call tameru_expand when it needs a line the shortening left out. A
result is only rewritten when the reduction saves at least a fifth of its
characters; a short one, a result of tameru_expand's own, and one from a
tool on the keep list come back exactly as they were.
What of the original it does not do: the original is a context engine that subclasses another agent's compressor and prunes results already in the history, on a token budget scaled to the model's window, with an archive lookup after a restart; Nodecode has no compressor to subclass, so this cell acts at the tool-result seam instead, keeps the original in memory rather than in a session archive, and has no summariser pass. See NOTICE.
MIT-licensed. See LICENSE.
