nodecode-run-debugger
Agent Run Debugger: an imported agent run's timeline, diagnosis and inert replay, offline. A Nodecode rewrite of Sahil-SS9/agent-run-debugger by Sahil-SS9.
Hand it one trace file and it tells you what the evidence supports: the order of what happened (from the recorded causal parents, not from the clock), what the source itself reports as failed, the hypotheses over it — each with its alternatives and what would falsify it — the evidence that is missing, and a replay that walks the recorded timeline and executes nothing. A completed run is never reported as a passed task, and an unknown cost stays unknown.
Get it
Tell the agent "add the run debugger cell", or run nodecode add nc://run-debugger.
Set it up
No key and no address: this cell reads files on this machine and talks to nothing. Every member of the run-debugger section has the original plugin's default, so an absent section is a working cell.
| Member | What it is |
|---|---|
max_bytes |
Hard cap on bytes read from one selected trace (default 67108864). |
max_line_bytes |
Hard cap on one JSONL line (default 8388608), so an embedded prompt snapshot cannot swallow the file. |
max_events |
Hard cap on records accepted from one trace (default 200000). |
allow_partial |
Keep the valid prefix of a truncated final line, marked PARTIAL with an import gap (default off). |
retention_days |
How long an imported run is kept under the home (default 90); only files this cell wrote are ever deleted. |
Use it
The tools the model calls, all local:
| Tool | What it does |
|---|---|
rundebug_import |
Import one or more trace files you name. A directory, an archive or a raw database is refused; nothing is extracted and nothing runs. |
rundebug_timeline |
The recorded order — causal parents first, source sequence as the tie-break — with gaps and concurrent pairs left visible. |
rundebug_explain |
What the source reports, the ranked hypotheses over it, the missing evidence, the read-only next checks, and the cost as reported. |
rundebug_replay |
The same order walked: imported command and error text is shown as data, never executed. |
rundebug_status |
The runs held, the lanes, the bounds, and where the runs are kept. |
The same verbs are one command, /rundebug import FILE, /rundebug timeline RUN, /rundebug explain RUN, /rundebug replay RUN, /rundebug report RUN, /rundebug runs, /rundebug status, and functions in the rundebug package for eval. /rundebug report writes one markdown file under the home and answers its path.
The trace is JSONL, one JSON object a line, in the original plugin's own format: a trace envelope ("schema": "hermes-run-debugger/trace", "version": 1), then attempt and event records. An event names its kind (tool_call, model_call, provider_error, interrupt, …), what it reports about itself (ok, failed, interrupted, skipped, unknown), where that came from (observed, reported, inferred, unknown), its source_sequence, and its causal_parents; a parent the trace does not carry becomes a visible gap rather than an invented edge.
