get-fable

Two coding agents running the same model can behave in diametrically opposite ways. One inspects the repository before touching anything, writes a failing test before a fix, and re-diagnoses the root cause after three failed retries. The other patches production code on a hunch, treats a stale green test run as proof, and ends the session with a confident "Done." get-fable targets exactly that gap. It does not swap your model; it wraps the model in a disciplined working layer built from 25 interconnected specialist skills and a lifecycle router that decides which skill takes over based on intent, durable state, and priority signals such as repeated failure, security sensitivity, and stale verification.

Every skill carries its own operational playbook: explicit activation boundaries, situational classification for ambiguous cases, staged red/green execution gates, non-negotiable architectural invariants, failure taxonomies that change the next action on error, anti-pattern guards, structured state handoffs, and comprehensive behavioral evals. The catalog spans understanding the work (repository path tracing, primary-source research, dependency-aware planning), building the change (TDD boundary proof, semantic-level parallel delegation, scoped execution, complexity reduction), proving correctness (failure-mode matrices, diff-level review, trust-boundary security tracing, independent oracle comparison, agent behavior evaluation), keeping long sessions healthy (causal rebuild after repeated failure, the Spark micro-suggestion that is permitted to stay silent, durable memory, resumable handoff, autonomous scoped work, state-machine polling), working with the environment (runtime artifact spawning with readiness and cleanup phases, harness configuration with least privilege), and producing evidence people can use (honest data visualisation, source-grounded documents and diagrams). A dedicated skill-creator skill lets you extend the harness to the same V2 standard.

get-fable is portable across more than 30 coding agents, platforms, and IDEs, from Claude Code, Codex, Cursor, and Gemini CLI to community tools like Aider, Cline, and Roo Code. You keep the model and editor you already use and add the discipline layer on top. It is built for engineers and teams who want their existing agent to work more like a senior partner: inspect before edit, prove before patch, re-diagnose before retry, and verify before declaring done.