Fable Experience Layer

    1

    Gives your coding agent a memory: past lesson cards and public agent traces before a task, a new card after it.

    Free

    1 useSecurity scanned

    Works with the AI Agents you already use

    Claude CodeClaude CodeCursorCursorCodex CLICodex CLIMuseMuseDotsDots+21 more

    Fable Experience Layer

    Example session with this skill installed

    /fable fix the flaky login test in e2e/login.spec.ts. It passes locally and fails about one run in five on CI.

    • Read your context and instructions
    • Compiled the fable experience layer

    Summary

    Routed the task, found a matching lesson card in the local corpus, applied it, and recorded the result as a new card.

    Step 1: route

    fable_route returned these engines to use, in order: corpus_search, external_retrieve, laya_boost (skipped on Linux), qa_architect, self_learning_record. It skipped headroom_compress because no full transcripts will be pulled.

    Step 2: search the local corpus

    fable_search with "flaky login test":

    [100% match] Fix flaky login test (e2e/login.spec.ts)
      context:   web app, Playwright, GitHub Actions
      outcome:   20/20 green after replacing the fixed sleep with a wait for networkidle
      steps:     1) reproduce 10x locally with --repeat-each → 2) replace page.waitForTimeout(2000) with page.waitForLoadState('networkidle') → 3) rerun 20x locally and once on CI
      learnings: wait for networkidle before asserting the dashboard • never use fixed sleeps in Playwright e2e • the CI runner is slower than local, so time-based waits fail there first
      gotchas:   the failure only showed on CI
    

    Step 3: apply

    Replaced page.waitForTimeout(2000) with page.waitForLoadState('networkidle') before the dashboard assertion in e2e/login.spec.ts. Ran the spec 20 times locally with --repeat-each 20: 20 of 20 passed. Pushed and the CI run passed.

    Step 4: record the lesson card

    node scripts/record.js --task "Fix flaky login test (e2e/login.spec.ts)" --outcome "20/20 green after replacing the fixed sleep with a wait for networkidle" --learnings "...":

    
    # Fable local corpus — 1 card(s)
    
    - [2026-10-20T15:11:09.222Z] Fix flaky login test (e2e/login.spec.ts) → 20/20 green after replacing the fixed sleep with a wait for networkidle
      learnings: wait for networkidle before asserting the dashboard • never use fixed sleeps in Playwright e2e • the CI runner is slower than local, so time-based waits fail there first
      gotchas: the failure only showed on CI
    

    Next steps

    • Run grep -rn waitForTimeout e2e/ and treat each hit the same way.
    • Next time a test flakes, fable_search finds this card first, before any external retrieval.

    Connects securely to your tools. The creator never sees your data.

    About this skill

    Fable is an open source experience layer for coding agents, with a stdio MCP server. It gives an agent two sources of past work and a loop that adds to them.

    What it does

    • Keeps a local corpus of lesson cards at ~/.fable/corpus.jsonl. A card holds the task, the outcome, key steps, gotchas and learnings. The agent writes the card at the end of a task; Fable stores, dedupes and ranks it. No model is called for this.
    • Retrieves agent session traces from five public Hugging Face datasets through the public datasets-server API. No account or token is needed.
    • Routes a task: fable_route returns which engines to use and which to skip, with a reason each.
    • Ships qaforge test tooling in the same repo: a risk score, change impact on the git diff, failure triage, a test healing proposal, and a release gate verdict.

    What you get

    The SKILL.md tells the agent when to use Fable, how to install it, how to register the MCP server, and the five step loop: route, search your own corpus, retrieve external sessions, do the work, record a card. Two reference files list every MCP tool with its input fields and the lesson card format with the exact commands.

    Requirements

    Node.js 20 or newer, Python 3.10 or newer, git, and the repository cloned locally. The core scripts run with Node and Python alone. npm install is needed for the qaforge tools. The MCP server is scripts/mcp_server.py and uses only the Python standard library.

    Network and data

    The only network access is read only requests to the Hugging Face datasets-server from fable_retrieve and fable_boost. The corpus stays on your machine as plain JSONL. Results from external datasets come from other projects, so the skill tells the agent to check each hint against the code in front of it.

    Source: https://github.com/romangalaxys10-spec/fable

    Laya triage runs on macOS with Apple Silicon only; other platforms use lexical ranking. The qa_* tools need npm install in the repo. Traces from the public datasets come from other projects and need checking against your code. Keep secrets out of lesson cards; the corpus is plain JSONL. HF_TOKEN is optional and only raises the rate limit.

    How to use

    1. 1

      Connect your agent to the Agensi MCP

      Copy the prompt and paste it into Claude, ChatGPT, Cursor or any agent that supports MCP. It sets up the connection for you. Manual setup

    2. 2

      Sign in to Agensi

      Your agent shows a sign-in link. Sign in and click Allow.

    3. 3

      Ask for what you need

      Your agent loads this skill and uses it on your task. It always asks before any payment.

    Reviews

    No reviews yet

    Be one of the first to try it. Every listed skill goes through the trust checks below.

    Security scanned

    Checked by our 8-point scan before listing

    1 use

    Times used to date

    Free forever

    No account required to browse

    Trust & safety

    Security scan

    Security scan passed 27 days ago

    • Free with an Agensi account

    Requires

    Node.js 20+
    Python 3.10+
    Git
    fable repository
    Python
    macOS
    tsx
    Node.js

    Permissions

    Terminal / Shell
    Read Files
    Network Access
    Write Files

    Terminal: the MCP server runs the repo's own Node and Python scripts. Read files: the qa_* tools read the current repository and git diff. Write files: record.js appends lesson cards to ~/.fable/corpus.jsonl. Network: read only requests to datasets-server.huggingface.co for fable_retrieve and fable_boost. No API keys are needed; HF_TOKEN is optional.

    Listed27 days ago

    What's inside

    Frequently Asked Questions

    At the end of a task the agent writes a lesson card (task, outcome, steps, gotchas, learnings) to a local JSONL corpus. Before the next task, fable_search ranks those cards against the new task so the agent reads its own past lessons first.

    Any MCP host that can run a stdio server: Claude Code, Claude Desktop, Cursor, Windsurf, Codex CLI and others. The server is a single Python file in the repo, scripts/mcp_server.py.

    Yes. fable_search reads your local corpus at ~/.fable/corpus.jsonl. fable_retrieve reads five public Hugging Face datasets through the public datasets-server API. The local corpus is searched first.

    A SKILL.md that tells the agent when to use Fable, how to install the repo, how to register the MCP server, and the route, search, retrieve, record loop, plus two reference files: every MCP tool with its inputs, and the lesson card format. The code itself lives in the public GitHub repository, which you clone.

    Lesson cards are plain JSONL on your machine. Nothing is uploaded. The only network calls are read only requests to the Hugging Face datasets-server. Do not put secrets into lesson cards.

    No. No model is called by Fable itself and no account is needed. HF_TOKEN is optional and only raises the Hugging Face rate limit.

    New in LLM Ops