unloop v0.2.0: ArgShield Path & Argument Hardening, Stdio Session Recorder & Semantic Anti-Oscillation Watchdog

Highlights of unloop v0.2.0 (Rank 5 Sprint)

• ArgShield Tool Argument Hardening:
  • Enforces strict path traversal boundaries via \os.path.commonpath.
  • Blocks cross-platform DOS/Windows reserved device names (\CON, \PRN, \AUX, \NUL, \COM1-9, \LPT1-9).
  • Tokenizes and strips flag injection attempts (--flag=value, -flag).
  • Redacts high-entropy API keys and passwords before wire execution.
• UN-01 Transparent Stdio / JSON-RPC Session Recorder:
  • Intercepts wire-level agent tool streams into high-performance SQLite in WAL mode.
  • Length-prefixed RFC 8785 canonical JSON-RPC hash chaining for audit-grade tamper resistance.
• UN-02 Semantic Anti-Oscillation Watchdog:
  • Tracks sliding-window tool executions to detect repetitive thrashing, alternating cycles, and state loops.
  • Trips circuit-breakers dynamically before runaway token or budgetary exhaustion.
• Deterministic 6-Slot Compaction:
  • OpenClaw/Hermes constraint verification and memory preservation across agent handoffs.

unloop v0.1.0: The Time-Travel Debugger & Anti-Oscillation Watchdog for AI Agents

unloop v0.1.0: The Time-Travel Debugger & Anti-Oscillation Watchdog for AI Agents

unloop is an open-source, interactive time-travel debugger and trajectory flight-recorder for autonomous AI agents.

It breaks multi-turn agents out of infinite critique/retry loops, lets engineers rewind cognitive turns like a video player, inspect state diffs, fork alternate execution branches, and capture reproducible single-file .unloop CI replay cassettes.

---

⚡ 60-Second Quickstart

pip install unloop

Wrap Any CLI Agent Without Code Changes

Flight-record agents like Claude Code, Copilot, or Aider with automatic loop protection:

unloop wrap -- claude

Drop-In Python Function Tracing

from unloop import trace, Session

@trace
def my_agent_step(prompt: str, session: Session = None):
    # Every cognitive turn, tool call, and state delta is persisted
    return run_turn(prompt)

Interactive Terminal Debugger (TUI)

Launch the full-screen terminal inspector:

unloop tui session.unloop

• Navigate cognitive turns with arrow keys.
• Inspect JSON state deltas and tool argument hashes.
• Rewind, mutate corrupt context, and fork alternative reasoning trajectories.

---

🛡️ Core Capabilities & Architectural Invariants

• Deterministic Trajectory Persistence (SQLite WAL):
  • Append-only event store with sub-2ms write tax.
  • Encrypted at rest via optional AES-256 and PII redaction rules.
• OscillationWatchdog:
  • Detects agent death-spirals (repetitive tool calls, ping-pong state cycles, oscillatory self-critique loops) before token budgets are exhausted.
  • Freezes execution or drops into an interactive breakpoint when loop threshold is exceeded.
• State Deltas & Dynamic Forking:
  • Calculates semantic state diffs between arbitrary cognitive turns.
  • Supports branching: rewind to turn $N$, modify prompt or tool output, and execute an alternate path without starting over.
• CI Replay Cassettes:
  • Package multi-turn failures into portable .unloop files for deterministic regression tests and bug reports.

---

🧪 Quality Gates & Verification Suite

• Unit & Integration Suite: 42/42 tests passing in 4.73s.
• Code Coverage: 94.56% coverage (exceeding strict 90% threshold).
• Multi-Platform Verification: Tested across Ubuntu, macOS, and Windows on Python 3.10–3.14.
• Packaging: Built wheel and sdist distributions attached.

---

Full Changelog: https://github.com/invarcore/unloop/commits/v0.1.0