Sahil-SS9Agent Run Debugger
Debug failed agent runs with evidence, not stories.
Agent Run Debugger is a local-first evidence debugger for failed, stalled or expensive agent runs. Import one explicitly selected trace and get a deterministic timeline, observed failure points, ranked hypotheses that cite their evidence, the gaps you still need to fill, and an inert replay of the recorded run.
Status: v1 debugger, offline-first. The end-to-end offline journey is implemented: import, timeline, diagnosis, replay, export and status. Live capture and model narration are separate, opt-in lanes.
At a glance
| Area | What it gives you |
|---|---|
| Input | Selected JSONL traces from Hermes, Claude Code or the harness-neutral v1 format |
| Timeline | Deterministic ordering from the recorded events |
| Diagnosis | Observed failure points and ranked hypotheses with evidence |
| Gaps | Missing evidence and falsification checks you can run |
| Replay | Inert playback of the recorded timeline; nothing is executed |
| Output | A private Markdown report, with redacted exports |
Why it exists
Agent failures are usually buried in long traces. Reading the raw log gives you events, but not a clear account of what failed, what is unknown, or which explanation is actually supported.
Agent Run Debugger turns one selected trace into an evidence trail:
- what happened, in order;
- where the run first diverged or failed;
- which hypotheses are supported by the recorded evidence;
- what evidence is missing;
- what to check next;
- what can be replayed safely without touching the live system.
Design principles
- Evidence first. Observed facts, hypotheses and missing evidence are kept separate.
- Inert replay. Replay re-reads the recorded timeline. It runs no subprocess, opens no network connection and touches no live database.
- Local by default. Import a trace explicitly selected by you or the model through
rundebug_import. No telemetry collector and no background capture. - Host-neutral core. The domain core does not import Hermes. A thin adapter handles the host integration.
- Explicit opt-ins. Live reads, capture, model narration and real re-execution are separate lanes with their own gates.
- Unknown stays unknown. Missing logs, unknown billing and unsupported signals are reported as gaps, not guessed or rounded to zero.
Harness support
| Format | What it takes |
|---|---|
| hermes | hermes sessions export --format jsonl — the host’s own session envelope |
| v1 | The harness-neutral trace in schemas/trace-v1.json |
| claude | Claude Code JSONL session transcripts, via a tested adapter |
hermes sessions export --format jsonl --session-id <id> ./run.jsonl
hermes run-debug import --file ./run.jsonl --format hermes
Adding another harness means adding one adapter that maps its records onto event-v1. The store, timeline, diagnosis, replay and report layers stay the same.
Surfaces
| Surface | Route |
|---|---|
| CLI | hermes run-debug import\|timeline\|explain\|replay\|export\|status |
| Slash | /rundebug <subcommand> |
| Agent tools | rundebug_import, rundebug_timeline, rundebug_explain, rundebug_replay, rundebug_status |
export is operator-CLI-only. Tool, slash and gateway routes do not expose it, so viewing a private report never becomes an implicit sharing permission.
Privacy
- Minimal metadata by default; no raw prompt or result retention.
- Redaction runs before the store, before the journal, before error logs and again before export.
- Raw input digests stay local. Exports carry redacted content digests and stable opaque references.
- The store lives under
plugin-data/agent-run-debugger/for the active profile. - Deletion only targets owned record IDs and never follows a source-path reference.
Quick start
hermes run-debug import --file ./selected-run.jsonl --format v1
hermes run-debug timeline --run <run-id>
hermes run-debug explain --run <run-id>
hermes run-debug replay --run <run-id>
hermes run-debug export --run <run-id> --out ~/private/rundebug/report.md
hermes run-debug status
Roadmap
- Native capture lane with an explicit owner gate.
- Selected-session live reads.
- Optional model narration, clearly labelled as hypothesis.
- Sandboxed experiments and comparative replays.
These are separate product lanes. v1 is useful without them: it makes the recorded evidence legible and keeps the unknown visible.
Development
uv sync --locked --extra dev
uv run python -m pytest tests/unit tests/adversarial --strict-markers -ra
uv run python -m pytest tests/integration --strict-markers -ra
uv run ruff check .
Nothing here installs into a live Hermes profile. The host test loads the plugin from a checkout into a disposable HERMES_HOME.
Licensed under Apache-2.0. See THIRD_PARTY_NOTICES.md.
Repository: https://github.com/Sahil-SS9/agent-run-debugger
README from the reviewed commit f4eefef; it updates when the author re-pins.
