nodecode-recovery-peer
Recovery Peer: one bounded second opinion from another model when a tool call keeps failing or stalls. A Nodecode rewrite of rezaaibox-ui/nexus-recovery-peer by rezaaibox-ui.
The agent that is stuck is the last thing that will notice. When the same call fails the same way -- or returns the same output -- three times in a row, this cell asks a model on a provider of your own for at most two different next steps and how to check each, and appends them to the result the agent is about to read. The agent still decides, and nothing the peer says counts as verified until it is run.
| The stall | What the cell does |
|---|---|
| The same call fails the same way, 3 times | asks once, on the third result |
| The same read returns the same output, 3 times | asks once, on the third result |
| Anything else | passes the result through untouched |
It never replays the call, never substitutes a model for the session's own, and never asks twice for one stall. A loop costs one small call, not one per step.
Get it
Tell the agent "add the recovery-peer cell", or run:
nodecode add nc://recovery-peer
Set it up
It needs one setting, peer: the provider and model of the second opinion, as
provider/model -- a model you also have, on a provider other than the one you
run on. In /settings, or in the recovery-peer section of
~/.nodecode/config.jsonc (config.example.jsonc shows the whole section):
"recovery-peer": { "peer": "some-lane/some-model" }
Without it the section refuses, and /cells and the notices block say which
key is missing. other_provider_only (on by default) keeps the promise that
it is really a second opinion: a peer on the provider you are already using is
not asked.
Use it
Once it is on, there is nothing to call: every tool call of every session
passes through the cell's :TOOL hook.
| Verb | What it answers |
|---|---|
/recovery-peer |
on or off, the peer, this session's streak, the last word |
/recovery-peer on / off |
start or stop watching tool results |
/recovery-peer ask |
ask the peer now, about the last stall |
/recovery-peer reset |
forget every streak and the last word |
recovery_peer tool |
the model asks by hand: task, failure, tool |
(help :recovery-peer) |
the same, as prose |
The peer sees the task (the operator's newest ask, when the log can say), the
call as eval(form) and how it ended, each clipped to clip characters, with
the obvious credentials replaced by [redacted]. Redaction is best-effort,
not a security boundary: do not name a peer whose provider should not see this
conversation's error text.
Test it
.github/test-cell.sh tools/nodecode-recovery-peer
The tests are offline: the peer is bound to scripted answers, so "the third identical failure is the one that asks", "one stall asks once" and "a peer that raises answers a refusal instead of failing the call" are facts a test reads rather than a provider's mood.
