When autonomous agents enter unproductive retry loops, repeatedly fail at tool execution, or terminate unexpectedly, conventional monitoring often leaves you with logs but no clear explanation.
Last Signal reconstructs execution traces, detects failure patterns using deterministic rules, and generates evidence-grounded digital autopsies, entirely offline.
llama3.2 (offline): Upstream cluster outage; agent lacked circuit-breaker or local Chroma fallback.
Step through the full execution stream, audit verified state-machine rules, and inspect the generated incident autopsy.
Traditional monitoring relies on bloated cloud dashboards and fuzzy LLM prompts. Last Signal provides an offline-first forensic engine built on four immutable pillars.
Verified state-machine rules identify failure patterns without treating LLM guesses as facts. Rules audit exact tool failure counts, temporal crash deltas, and retry loops directly from raw JSONL traces.
The forensic core operates locally without requiring external cloud AI services. Proprietary agent prompts, tool parameters, and internal memory vectors never leave your host workstation or VPC.
Optional Ollama-powered interpretation provides contextual explanations and recovery suggestions using local open-weight models (default: llama3.2). AI reasoning is strictly bounded by deterministic evidence.
Generate structured Markdown post-mortems containing chronological evidence, verified finding tables, and remediation checklists. Formatted specifically for pull requests and incident retrospectives.
Non-invasive process supervision that works with any Python script or agent framework without requiring SDK code refactors.
Launches the agent as a supervised child process using subprocess.Popen(shell=False). Background worker threads stream unbuffered stdout and stderr, capturing JSONL events and fatal tracebacks without modifying agent code.
Auto-selects the latest failed trace from .lastsignal/traces/ and evaluates it against all 4 state-machine failure rules. Validates failure frequencies, retry loops, and temporal proximity to abnormal termination.
Synthesizes deterministic findings, chronological event logs, and optional local Ollama root-cause hypotheses into a publication-ready Markdown post-mortem document directly in reports/.
At least two distinct operational failures associated with a single tool in one execution session.
Sequential retries executed without an intervening successful operation before process exit.
A tool failure occurs as the immediate precursor event prior to abnormal process termination.
Unhandled fatal execution failure without a recorded cleanup or recovery checkpoint.
Last Signal is open source. The Python CLI establishes the forensic foundation; now we are actively inviting TypeScript and JavaScript developers to help design and implement the native TypeScript agent supervisor for Node.js, Bun, and Deno environments.