Observer Daemon — Development skill for Claude Code

observer-daemon

You can miss repeated writing-rule violations when you review agent responses by hand. Use observer-daemon to score transcripts and record failures outside the model.

Install with Rust and Cargo:

cargo install --git https://github.com/b2bvic/observer-daemon.git --locked

Sample output from `observer-daemon --config spec.toml --validate "The file is ready."` when no rules fail:

Class: generic
Score: 100/100
No violations.

Your rules determine the score. Configure `spec.toml` before you run the command. Transcript scoring alone does not block an outbound action.

What it does

• Watches transcript directories (watch_paths in spec.toml) with debounce, hot-reloading its spec on file change or SIGHUP.
• Classifies each prompt into a content class before scoring: legal, career_application, content_copy, technical_build, conversational, or generic fallback. Multi-signal prompts resolve by fixed priority. A career-strategy briefing that mentions a compliance role is career material, and the test suite pins that exact case as a named regression.
• Scores the response with the rubric for its class: a stack of deterministic validators (sycophancy, filler, crutch words, punctuation habits, lexical density, rhythm, model tells, policy-pack literals) each deducting from 100. Below passing_score, the response fails.
• Records every validation to a JSONL ledger, and every correction to a corrections ledger. Corrections are class-scoped: a rule promoted from legal failures does not silently constrain conversational writing. Records from before class scoping remain valid as generic.

Why a daemon and not a prompt

You run the same configured writing checks outside each model's prompt. The checks record rule matches and deductions. A model upgrade does not change those rules. You must review rules and examples when your requirements change.

Worked example

Feed it a transcript directory and a