FableCodex Evidence-Gated Workflow Skill
Instills Fable-style discipline in Codex: inspect first, ledger goals, capture evidence, close review findings, and verify before claiming done.
Skill is mostly local workflow docs and stdlib scripts with strong least-privilege discipline: no hidden credentials assumed, user confirmation required for Anthropic keys, local state scoped to .codex-fable5/, and SECURITY.md explicitly bans exfiltration. Deducted because scripts were not line-verified statically, provider bridge touches credential configuration, and rollback/isolation rely on documented conventions rather than enforced mechanisms.
Repository includes a unittest suite and CI (py_compile, shell syntax checks, pinned-source validation, wrapper smoke tests); key paths are auditable by design. Deducted because this is a static read with nothing executed, and failure-feedback quality on abnormal input can only be inferred; capped at the static maximum of 10.
Triggers (fablize, VFF, Fable-style workflow, prompt conversion) are explicit in the description and README clearly states fit/non-fit scenarios. Deducted because the core workflow assumes the Codex platform and some overseas services (Anthropic/LiteLLM gateway, raw.githubusercontent.com source fetch), creating mainland-China reachability risk not mitigated in the skill itself.
Layered docs (SKILL.md → references → multilingual README), provenance.md with upstream commit pins and licensing notes, version v0.5.1, AGPL-3.0, SECURITY.md and release guidance present. Deducted for unverified publisher identity, maintenance responsibility resting on a personal repo, and no visible changelog detail.
Core value (goal/findings ledgers, verification gates, 2-pass review) is concretely designed with directly usable commands and complete README examples. Deducted because static review cannot confirm actual output quality or marginal value over manual alternatives; kept within the static cap of 7.
Commit-pinned sources, an independently runnable fable_coverage.py validator over the coverage matrix, and CI that validates the matrix against the pinned upstream source give an unusually strong evidence chain. Deducted because all evidence is static files with no independent third-party execution reproduction; capped at 5.
- Publisher is an unverified personal repository; long-term maintenance continuity is uncertain.
- Provider-bridge functionality requires valid Anthropic access and an OpenAI-compatible gateway; verify credentials and compliance before configuring.
- CI fetches the upstream source from raw.githubusercontent.com, which may be unreachable from mainland-China networks.
- This assessment is a static source review only; no scripts or tests were executed.
- The skill emulates workflow discipline, not model capability; do not expect actual Fable 5 performance.
What does this skill do, and when should you use it?
FableCodex is a Codex plugin that turns Fable-inspired operating habits into an executable Codex workflow: inspect the workspace before acting, run long work against a goal ledger with evidence checkpoints, record review findings behind a mandatory findings gate, and verify with real evidence before declaring completion. It is explicitly procedural — it changes workflow discipline, not model weights, context length, training, or hidden safety systems, and does not replicate actual Fable 5 capability. It ships local stdlib-only Python scripts and a CLI, with state stored under .codex-fable5/. It fits work where the cost of a missed step exceeds the cost of a little process.
Reads the skill instructions into the Codex session: inspects workspace and files with rg; builds a multi-story goal ledger via scripts/codex_goals.py (.codex-fable5/goals.) where completion requires evidence; records review findings via scripts/codex_findings.py (findings.) with a gate that fails while findings are open; exposes codex-fable5 status/goals/findings commands; verifies source-heading coverage with fable_coverage.py; and generates a LiteLLM config via make_litellm_config.py (user supplies their own Anthropic key).
- A developer running a multi-step migration or refactor who wants evidence at each step before proceeding
- An engineer debugging an unclear root cause, wanting reproduce-first, multi-hypothesis, disconfirming-evidence workflow
- A team doing security-sensitive or pre-release review where unresolved findings should block final completion
- Anyone converting Claude/Fable-style system prompts into Codex-native AGENTS.md guidance
- An operator with authorized Anthropic access wiring Codex model routing through an OpenAI-compatible gateway
- Long-horizon tasks like CI failures or release work that benefit from a verification gate
What are this skill's strengths and limitations?
- Scripts depend only on the Python standard library — no extra install burden
- Goals and findings both require evidence, with a hard gate before final completion
- Honest scoping: workflow emulation only, never claims Fable capability or Claude identity
- Coverage tool audits that every source heading got a Codex-native decision
- Multilingual README (including Simplified/Traditional Chinese) and a CI test suite
- Ledger overhead outweighs the work for small tasks, short answers, or tiny single-file edits
- Tightly bound to Codex plugin mechanics and the @codex-fable5 trigger; other platforms need adaptation
- Improves discipline, not raw model capability — expectations must be managed
- Optional model routing requires the user's own Anthropic key and a LiteLLM gateway; no model access is provided
- Coverage verification against CLAUDE-FABLE-5.md only works if the user has a local copy of that source
How do you install this skill?
Stable: codex plugin marketplace add baskduf/FableCodex --ref v0.5.1, then codex plugin add codex-fable5@fablecodex, then restart Codex. For the development version use --ref main; for local development point the marketplace at a local path (codex plugin marketplace add ~/Desktop/FableCodex). The repo does not document installation on platforms other than Codex.
How do you use this skill?
Invoke explicitly in the prompt, e.g.: "@codex-fable5 Use this skill to implement the change. Create a goal ledger if the work has multiple steps. Track findings before final completion. Run the project tests before saying it is done." Also supports strict mode (ledger plus findings gate must pass), analysis-only mode (no file edits), or scoped mode (no commit/push/delete). For long work use codex-fable5 goals create/checkpoint with evidence; for review use findings add/resolve/gate. Skip it for short answers and tiny edits.