Agent-Casuality

Causal debugging for branching multi-agent systems — plus the OpenCode capture plugin.

Agent-Casuality records what your agents did as a causal event graph, then explains failures with grounded evidence instead of guesses.

This README is identical on both registries. The full system needs both packages:

• Python (agent-casuality on PyPI) — SDK, CLI, ingest receiver (casuality-opencode-ingest), and Visual DAG Explorer (casuality-explorer). Installing only this gives you the offline demo but no live OpenCode capture.
• JavaScript (@psionic-labs/opencode-agent-casuality on npm) — OpenCode plugin (telemetry capture only). Installing only this gives you events with nowhere to persist or view them.

Source: https://github.com/Psionic-labs/Agent-Casuality

Install · 30-second demo · Single-run capture · Explorer · OpenCode · Docs

Why it exists

Logs and traces show what happened. When several agents branch, merge, and share state, that isn't enough to tell what actually caused a bad outcome.

Agent-Casuality adds three things ordinary observability lacks:

Agent execution → event capture → causal graph → slice / provenance / interaction → grounded diagnosis

Install — you need BOTH packages

Published as agent-casuality==0.1.4 (PyPI) and @psionic-labs/opencode-agent-casuality==0.1.4 (npm). Requires Python 3.11+ and uv for source checkouts; PyPI/npm consumers only need pip/opencode.

# 1. Python side (SDK + receiver + explorer + CLI)
pip install agent-casuality
# or, inside a uv project:
uv add agent-casuality

# 2. JavaScript side (OpenCode capture plugin — telemetry only)
opencode plugin add @psionic-labs/opencode-agent-casuality

npm-only installers: you still need the PyPI package — the Visual DAG Explorer and the casuality-opencode-ingest receiver ship only in agent-casuality, not in this npm package.

PyPI-only installers: you still need the npm package for live OpenCode capture — pip install agent-casuality alone cannot observe an OpenCode session without the plugin.

Optional — AI explanation extra. Only needed for LLM-grounded explain / Generate AI analysis (needs httpx + OPENROUTER_API_KEY):

# pip consumer
pip install "agent-casuality[explain]"
# uv project
uv add "agent-casuality[explain]"
# source checkout
uv sync --extra explain

Try it in 30 seconds (no server, key, or network)

uv run python examples/customer_approval.py

It captures two agent results and a wrong approval, then explains the failure:

Outcome: approved
Diagnosis: The terminal event is marked 'failure' because the decision returned
'approved' with customer_status='eligible' and risk_score=0.2.
Evidence:
- Minimal tested chain: customer_status + risk_score -> decision -> terminal failure
- Canonical baselines were 'ineligible' and 0.8
- Interaction between customer_status and risk_score: 1.0
Limitations: the evidence does not establish why the upstream tools produced these values.

Single-run capture: ingest → opencode → explorer

The minimal end-to-end loop. One isolated SQLite file per run keeps traces separate (.casuality/single-run.db below is just an example — any path works) and port 8797 avoids clashing with the default 8766 explorer.

Terminal 1 — start the receiver:

# PyPI install
casuality-opencode-ingest --db .casuality/single-run.db
# Source checkout
uv run casuality-opencode-ingest --db .casuality/single-run.db

Terminal 2 — run one instrumented agent task (plugin posts automatically):

opencode run "Run the existing tests before editing. Diagnose and fix the bug in discount.py so both tests pass. Do not modify the tests. Rerun the tests and summarize the change."

Optional — enable AI explanations for this project:

uv add "agent-casuality[explain]"
# pip equivalent: pip install "agent-casuality[explain]"

Terminal 3 — inspect the trace in the Visual DAG Explorer:

# PyPI install
casuality-explorer --db .casuality/single-run.db --port 8797
# Source checkout
uv run python -m explorer.server --db .casuality/single-run.db --port 8797

Then open http://127.0.0.1:8797. Each opencode run appends to the same --db file; use a fresh --db path per task for one-trace-per-file (any filename works).

See it: Visual DAG Explorer

Open any capture as an interactive causal graph — no API key needed. The Explorer ships inside the agent-casuality PyPI wheel (no clone needed):

# PyPI install
pip install agent-casuality
casuality-explorer
casuality-explorer --db .casuality/opencode.db --port 8766

# Source checkout
uv run python -m explorer.server
uv run python -m explorer.server --db .casuality/opencode.db --port 8766
# Single-run variant from the section above:
uv run python -m explorer.server --db .casuality/single-run.db --port 8797

Then open http://127.0.0.1:8766 (or :8797 for the single-run variant). The bundled demo trace loads: a failing test, the fix edit, and the passing retry, with the failure path highlighted.

Freshly captured shared_state_contamination run (240 events, 6 causal edges) — the checker's failure node selected, its causal chain in the inspector.

Bundled demo trace in the Timeline view — per-agent lanes (run, bash, edit), the retry's tool_result ("app tests pass") selected as the failure target, with the "Why did it fail?" strip and event inspector alongside.

What you get:

• Graph + Timeline views of the declared causal parents — every edge was recorded at capture time, none inferred.
• Event inspector — click any node for its command, result, parents, and children.
• Analysis tabs (chain, provenance, interaction, evidence, metrics) stay informative even with no ground truth, built from capture facts.
• Your own runs: point it at a live capture — casuality-explorer --db .casuality/opencode.db (source checkout: uv run python -m explorer.server --db .casuality/opencode.db).
• Optional AI interpretation: install the explain extra (pip install "agent-casuality[explain]" / uv add "agent-casuality[explain]"), set OPENROUTER_API_KEY (and optionally OPENROUTER_MODEL, default nvidia/nemotron-3-ultra-550b-a55b:free), then click Generate AI analysis in the "Why did it fail?" strip. The model text lives in its own labeled block — never mixed with the evidence.

Full acceptance procedure: TEST.md §29.

Use it with your agents

import casuality

casuality.init(".casuality/events.db")

@casuality.agent(role="risk_checker")
def check_risk(customer): ...

@casuality.merge_decision(...)
def approve(customer_status, risk_score): ...

See examples/customer_approval.py for a complete runnable example.

Works with OpenCode

Capture live coding-agent sessions. Telemetry is fail-open: it never interrupts your agent. Requires both packages (receiver from PyPI, plugin from npm).

# Terminal 1: receiver (PyPI install; source checkout: uv run casuality-opencode-ingest ...)
casuality-opencode-ingest --db .casuality/opencode.db
# Terminal 2: agent (plugin posts automatically)
opencode run "Read RESEARCH.md, implement sort_items in task.py as recommended, run the test script."
# Terminal 3: Visual DAG
casuality-explorer --db .casuality/opencode.db

Full consumer flow: install agent-casuality (PyPI) → install the OpenCode plugin (opencode plugin add @psionic-labs/opencode-agent-casuality, npm) → start the receiver → run OpenCode → launch casuality-explorer → inspect the recorded trace. No repository clone is needed.

Representative committed result (stale_research, OpenCode 1.19.0): completed (exit 0), 331 events, capture coverage 1.0, diagnosis scored (see benchmark/results/opencode/opencode.md). No video artifact ships with this release. Setup guide →

What the plugin captures

The npm plugin forwards OpenCode's global event stream plus explicit hooks (tool.execute.before/after, permission.ask, command.execute.before, chat.message, chat.params). The Python receiver maps these onto canonical events, registers file writes with Resource-Version machinery, and redacts secrets before SQLite persistence. It is telemetry capture only — the Explorer and CLI live in the PyPI distribution.

Benchmarks

Offline, provider-neutral suite (uv run casuality-benchmark ...) plus a real OpenCode coding-agent benchmark. Full results in docs/benchmarks.md and benchmark/results/.

• Five deterministic scenarios: 5/5 pass.
• Experiment 2 (shared-state recovery): 100/100.
• Experiment 3 (semantic-port vs raw deletion): 20/20 semantic-port successes; raw-deletion failure rate met target.
• Experiment 1 (offline): passes; external Fastino run is provider-blocked/deferred (no structured-output capability).
• Real OpenCode benchmark (5 runs / 5 tasks, OpenCode 1.19.0): capture micro/macro-average 1.0; cause recall mean 0.3 (structural-slice ancestry under explicit-evidence edges only); joint ancestry 0.0 — genuine causal interaction is unsupported from live traces.

Terminology is precise: live-trace branch results are reported as joint ancestry (not causal interaction), and minimal slices as a required-cause preservation proxy (not causal minimality).

Documentation