Alibi

We check whether your agent has an alibi for what it did.

Alibi finds where a long AI-agent run went wrong. It borrows the playbook of automotive driver-assistance systems: treat the run as a time series, filter it forward, detect the fault, then smooth backward to the moment it began. The only sensor is Jev, TypeSafe AI's System One model, which answers with calibrated probabilities instead of text.

It is built for long traces, where reading everything at once breaks down: on real coding failures of 59K to 120K tokens, a single whole-trace read found the root-cause step in 0% of cases; Alibi found it in 16% and 7.8%, and points to the right neighbourhood (within 3 steps) in 21% and 17%.

How it works

flowchart LR
  A[Trace] --> B[10K-token chapters]
  B --> C[Forward filter:<br/>Jev rates each chapter]
  C -->|memory card| C
  C --> D[CUSUM alarm<br/>on health]
  D --> E[Look back from the alarm:<br/>chapter + per-step evidence]
  E --> F[Top 3 steps to read,<br/>earliest strong suspect first]

What you get

A short trace is refused, without spending a call:

$ alibi diagnose examples/sample_trace.json
Trace is 219 tokens, under the 50,000-token threshold: a single direct read is
enough; Alibi adds value on long traces.

A long one comes back as three steps to read, in order. This is a real run from the locked TrajErrBench set (a Claude Opus coding agent failing on a qutebrowser issue), replayed from its recorded result:

$ alibi diagnose trace.json
Read these 3 steps first, in order.
80,778 tokens, 10 chapters, alarm at chapter 6; 17 Jev calls, 35 s, $0.0096
1. step 74 (assistant), chapter 6, score 0.277
   'Now I see the full picture. The test on line 458 expects
    `str(proc.outcome) == 'Testprocess crashed.'` for SIGSEGV...'
2. step 76 (assistant), chapter 6, score 0.202
   '## Phase 5: FIX ANALYSIS\n\nNow I have a clear understanding. Let me
    implement the changes to `guiprocess.py`...'
3. step 78 (assistant), chapter 6, score 0.178
   'Now let me implement all the changes:\nTool calls:\nstr_replace_editor(...'

Step 74 is the labelled root cause: the agent reads the test wrong and every later edit builds on that reading. Being right at rank 1 happens on 16% of these traces; the honest claim is that three steps out of 118 is a much smaller haystack.

Results

Locked test sets, each run once against a pass bar written down beforehand:

Against published methods on the LongRCA leaderboard (all on DeepSeek-V4-Flash; exact root step, same 128 SWE-bench Pro failures):

Alibi ties ECHO (Fisher p = 0.66) and trails RCTA (p < 0.000001), which traces each suspect back to the handoff instruction between agents.

When to use it, and when not to

• Use it for traces of tens of thousands of tokens or more (default gate: 50K tokens).
• Do not use it for short traces: a single direct read by any capable model is enough, and Alibi says so without spending a call.
• Treat the output as "read these 3 steps first", not as a verdict.

Speed and cost

• Each chapter is one Jev call, and chapters are judged in parallel on the way back. A 38K-token trace is 4.6 chapters and 15 seconds of judge time; a 120K-token trace is 17 chapters and about 48 seconds. No reasoning tokens are generated.
• Cost scales with trace length: about $0.10 per million trace tokens at Jev's price of $0.042 per million input tokens (roughly $0.01 for a 100K-token trace).

Install

Claude Code:

/plugin marketplace add ahmedezz26/alibi
/plugin install alibi@alibi

Set TYPESAFE_API_KEY in your environment. Then ask Claude Code: "Why did my last session go wrong?"

Command line (the PyPI distribution is agent-alibi; the import package is alibi):

uv tool install agent-alibi
ALIBI_JUDGE_BACKEND=typesafe ALIBI_ALLOW_PAID_MODELS=1 TYPESAFE_API_KEY=... \
  alibi diagnose path/to/trace.json

Privacy

Traces above the length gate are sent to TypeSafe's API; below it, nothing leaves your machine. There is no telemetry. Claude Code session parsing is best effort: the transcript format is internal to Claude Code and may change. Session transcripts often contain source code and secrets, so read SECURITY.md before diagnosing one.

Contributing

See CONTRIBUTING.md. Plumbing, adapters, bug fixes and docs are welcome as ordinary pull requests; changes to the estimation method need a pre-registered measurement, for the reason the research log makes obvious.

Background

Alibi started as an experiment by an ADAS engineer: can the tracking filters used in cars work on AI-agent traces? The research log with every pre-registered test, including the ones that failed, is in docs/research-log.md.

License

MIT