ReproSieve

ReproSieve turns one failed agent trace into a smaller, redacted, deterministic capsule that preserves the failure condition expressed by your predicate.

Important: ReproSieve preserves what the predicate recognizes. A weak or overly broad predicate can preserve the wrong failure.

Status: ReproSieve 0.1.0a4 is an experimental technical alpha. Use synthetic or disposable inputs, inspect capsules before sharing, and treat embedded predicates as arbitrary Python.

Try the complete package-owned synthetic demonstration:

python -m pip install reprosieve==0.1.0a4
reprosieve demo

It needs only the core package: no OpenAI extra, API key, agent application, or network access. It reduces the synthetic 247-event fixture, independently verifies 1-minimality, materializes retained recorded values, exports an offline predicate reproduction, and executes that export.

ReproSieve captures one supported OpenAI Agents SDK trace, removes configured secrets before persistence, replaces model and tool calls with recorded outputs, and reduces the trace against an executable failure predicate. The result is a hash-addressed capsule with an independent 1-minimality proof.

Feedback boundary: Do not submit real traces, credentials, private source, personal data, or confidential capsules. Submit synthetic/disposable reproductions or a prose description of the unsupported failure shape.

Supported path

• Python 3.11, 3.12, and 3.13.
• openai-agents>=0.18.3,<0.19; CI tests 0.18.3.
• One completed SDK trace captured through the public TracingProcessor API.
• Embedded Python predicates. ReproSieve invokes them with a direct argument vector, a clean temporary directory, provider keys removed, bounded time/output/process resources, and Python audit hooks that deny outbound network, child processes, native loading, and host-file access. These controls are defense in depth, not an OS sandbox; embedded predicates are arbitrary code.
• Deterministic recorded-output materialization and offline predicate reproduction. These paths reconstruct retained values and execute only the declared predicate; they do not rerun application or orchestration code.

Support matrix

Install capture support:

python -m pip install "reprosieve[openai]"

Workflow

The predicate script is arbitrary capsule-provided Python and is not sandboxed. It must be included in the capsule, and every command that can execute or export it requires --trust-embedded-predicate. --predicate must be the final option because everything after it is an argument vector.

reprosieve capture \
  --output failed.reprosieve \
  --workspace-root . \
  --include verify_failure.py \
  -- python app.py

reprosieve reduce failed.reprosieve \
  --output-dir reduced \
  --trust-embedded-predicate \
  --predicate python verify_failure.py

reprosieve materialize reduced/<sha256>.reprosieve \
  --output materialized.json

reprosieve reproduce-predicate reduced/<sha256>.reprosieve \
  --trust-embedded-predicate \
  --predicate python verify_failure.py

reprosieve verify-minimal reduced/<sha256>.reprosieve \
  --trust-embedded-predicate \
  --predicate python verify_failure.py

reprosieve export reduced/<sha256>.reprosieve \
  --trust-embedded-predicate \
  --output issue-repro

cd issue-repro
python reproduce.py --trust-embedded-predicate

Exported reproductions refuse to execute without that explicit flag. Use it only after inspecting and accepting the embedded Python predicate as arbitrary code running with your user account's permissions.

materialize writes recorded values as deterministic JSON. reproduce-predicate evaluates the embedded predicate in a fresh constrained directory and reports every trial. Neither command is application replay.

The pre-0.1 minimize and replay names remain warning aliases for reduce and materialize. They may be removed before 0.2.

Application replay is not part of the 0.1 CLI or promise. Experimental 0.5 library work now has one OpenAI Agents SDK adapter that re-executes an explicit application callback through injected public Model and FunctionTool interfaces. It enforces ordered exact matching and measured canaries, but remains synthetic-only and is not a 0.5 readiness claim. See the application-replay boundary.

Predicate exit codes are strict:

• 0: target failure reproduced.
• 1: target failure absent.
• 2: candidate or harness invalid.
• timeout, signal, cancellation, output overflow, unexpected exit, or missing predicate file: invalid.

Invalid is never treated as absent and is never accepted as a reduction.

Zero-setup synthetic demo

The installed package includes a synthetic 247-event mechanical fixture:

reprosieve demo

The demo accepts no external capsule, predicate, command, or executable. Therefore, unlike commands that consume user-controlled capsules, it does not require --trust-embedded-predicate. To retain the source capsule, reduced capsule and report, materialized JSON, standalone export, and summary JSON, pass a path that does not already exist:

reprosieve demo --output-dir reprosieve-demo-output

Without --output-dir, the temporary workspace is cleaned automatically. The demo never overwrites an existing path.

Reproducible proof

The repository release gate runs the same installed CLI demo through its historical source entry point:

python scripts/killer_demo.py

The release gate requires this command to finish within 20 seconds. “1-minimal” means deleting any remaining declared unit makes the failure absent or the candidate structurally invalid. It does not mean globally smallest.

Two additional copy-paste checks:

python -m scripts.verify
python -m scripts.final_release_gate

The first runs tests, lint, typing, and contract self-tests. The second runs the current exact-head security, secrets, killer-demo, and minimality checks from a clean Git state. The immutable contract-v2 release gate remains historical evidence, as documented in RELEASING.md.

Scope and comparison

ReproSieve is not a trace viewer, observability backend, VM sandbox, or general record/replay system. Existing record/replay tools preserve executions; ReproSieve adds dependency-aware reduction, redaction before its own persistence, strict tri-state predicates, and an independently checked 1-minimality claim for one recorded agent trajectory. It does not diagnose root cause, rerun application logic, or replay live model semantics.

Privacy boundary

ReproSieve redacts SDK payloads in memory before its own first write. Capture replaces the SDK default exporter unless --retain-sdk-exporter is explicitly passed. Captured target stdout and stderr are discarded because they may contain secrets. Exact canaries, bounded regexes, allow paths, deny paths, declared workspace files, and environment allowlists are available on capture.

Arbitrary personal data cannot be detected reliably. Inspect every capsule before publishing it. See the privacy contract and security review.

Development

python -m pip install -e ".[dev]"
python -m pytest
python -m ruff check .
python -m mypy
python -m build
python scripts/security_check.py
python scripts/killer_demo.py

Apache-2.0 licensed.

See SUPPORT.md for supported combinations and CONTRIBUTING.md before opening a change. Maturity levels are defined in ROADMAP.md; real-case evidence status is tracked in docs/case-studies.