FableCodex
Adds evidence-backed goals, review findings, and completion gates to Codex workflows where missed steps are costly.
Per-dimension scores and reasoning
The evidence limits CI to contents: read, identifies .codex-fable5/ as local state, marks the provider bridge optional, and describes the update command's mutation scope. The security policy prohibits committing credentials and exfiltrating local or connector data. Upstream projects and exact commits are attributed thoroughly. Deductions apply because no enforced confirmation gate for high-impact operations is shown, sensitive-data protection is mainly policy rather than code-level control, GitHub Actions use mutable major-version tags, and update fetches and switches code without a complete rollback facility. Publisher identity is unknown, but no unrelated risk was inferred from that fact.
The README, CI, and tests are internally consistent about commands, Python versions, the pinned upstream commit, and coverage validation. Tests also cover PATH wrappers, version reporting, dirty-checkout refusal, and stable-tag selection. Deductions apply because no execution results are supplied, handling or offline alternatives for unavailable upstream downloads, the Codex plugin system, and the optional gateway are limited, and the provided tests establish clear errors only for some paths rather than consistent actionable diagnostics across every command.
The documentation distinguishes suitable scenarios such as multi-step implementation, debugging, migrations, and security changes from short answers, tiny edits, and brainstorming. It supplies precise prompts for strict, analysis-only, constrained implementation, and debugging modes, while clearly stating that it cannot alter model weights, context length, or hidden safety systems. Environment-fit points were deducted because support is demonstrated mainly for Codex, shell, Git, and Python 3.11–3.13, without evidence for Windows, non-Git workspaces, or other hosts.
The material has strong information architecture covering quick start, commands, installation, local state, coverage, provider routing, tests, provenance, and licensing. Plugin, skill, wrapper, and state-directory names are stable; examples are extensive and limitations are explicit. Deductions apply because there is no dedicated FAQ or systematic troubleshooting section; the README says AGPL-3.0-or-later while the supplied LICENSE is the AGPLv3 text and the materials do not show an explicit or-later grant or NOTICE contents; an update/version path exists but no changelog is provided; and the security contact route depends on optional GitHub private reporting, with no clearly identified maintainer or response timeline.
The goal ledger, findings gate, evidence checkpoints, verification commands, and residual-risk reporting create directly usable outputs. The documentation also says when ledger overhead is not worthwhile, giving a strong cost-benefit boundary. Marginal-value points were deducted because improvement over an ordinary Codex workflow is supported by design and examples rather than comparative evidence or adoption outcomes.
The README identifies upstream repositories and exact commits; CI fetches the same pinned source and passes it to coverage validation; tests cross-check the README pin, CI pin, fetched path, and validator argument. The documentation explicitly separates workflow adaptation and coverage accounting from claims of model capability or runtime parity. The static evidence fully supports traceability, corroboration, and fact/inference separation, so no points were deducted on these criteria, although this does not imply execution in this review.
- This assessment uses only the supplied static files; tests, installation, update behavior, goal/findings flows, and the provider bridge were not executed.
- Before installation, inspect the update implementation and complete dependency surface, and pin GitHub Actions to immutable commits rather than only @v6 major tags.
- Do not place secrets in project files, ledger entries, finding evidence, or prompts. Separately verify credential storage, log redaction, and network destinations for the optional provider bridge.
- The update command changes the checkout/plugin package. Keep a recoverable commit or tag and confirm a clean worktree first because no complete automated rollback mechanism is documented.
- Confirm consistency among the AGPL-3.0-or-later statement, LICENSE, and NOTICE, particularly before network deployment or distribution of modified versions.
What does this agent do, and when should you use it?
FableCodex is a Codex plugin that improves process discipline without replacing or modifying the underlying model. Its repository supplies the `@codex-fable5` skill, the `codex-fable5` command-line helper, and local ledgers for goals, review findings, and event history. During a task, the workflow calls for inspecting the workspace and available evidence before acting, then verifying results through tests, linting, type checks, screenshots, command output, source inspection, or connector readback. Goals and findings are stored in `.codex-fable5/goals.json` and `.codex-fable5/findings.json`, while `ledger.jsonl` keeps an append-only event history. It is aimed at implementations, debugging, migrations, releases, CI work, and reviews where an omitted step matters, but it does not clone Fable 5 or alter model weights, context length, training, or hidden safety systems.
When a user invokes @codex-fable5, the skill classifies the task and inspects the workspace, files, tools, or cited sources. For longer work, codex-fable5 goals create builds a multi-step ledger, goals next advances it, and goals checkpoint closes a goal with evidence plus optional verification commands and results. Review issues are recorded with findings add, prioritized with findings next, and closed with findings resolve only after resolution and verification evidence are supplied; findings gate fails while any finding remains open or blocked. The status command summarizes progress, version reports the installed version, paths, and Git checkout state, and update moves the checkout or package to a stable tag or an explicitly requested main. The included fable_coverage.py script can also account for headings in a local CLAUDE-FABLE-5.md as implemented, adapted, unsupported, or not applicable.
- A Codex user implementing a multi-file feature or refactor who wants explicit stages and evidence checkpoints.
- A developer debugging a failure with no obvious root cause who wants to reproduce it, retain competing hypotheses, and gather disconfirming evidence before fixing it.
- A maintainer handling CI failures, migrations, or releases who needs tests and recorded verification before declaring completion.
- A team conducting security-sensitive or high-risk reviews that wants unresolved accepted findings to block closeout.
- A user with Claude/Fable-style prompts who wants those operating habits expressed as Codex-native guidance and commands.
- A team with authorized Anthropic access that wants to explore the optional routing path through an OpenAI-compatible gateway such as LiteLLM.
What are this agent's strengths and limitations?
- Goal checkpoints carry evidence, and final checkpoints can preserve the verification command and observed result.
- The separate findings ledger and
findings gatekeep accepted review issues from disappearing during closeout. - Local JSON state plus an append-only
ledger.jsonlhistory makes workflow progress inspectable. - Prompt-level controls support strict execution, analysis-only review, lightweight review, and implementation with explicit limits.
- The project clearly distinguishes procedural discipline from model capability and does not claim to unlock or reproduce Fable 5.
- The core delivery is tied to the Codex plugin ecosystem rather than being a general-purpose agent runtime for ChatGPT, Claude Code, or other platforms.
- Goal and findings ledgers add process overhead; the documentation advises skipping them for short answers, tiny edits, and brainstorming without verification.
- Task state is written into the repository under
.codex-fable5/, so teams must decide whether to ignore or intentionally preserve those files. - The optional Anthropic routing path supplies no model access; adopters must provide authorized credentials and a compatible gateway such as LiteLLM.
- A 100% source-heading coverage result means every heading received a disposition, not that the workflow matches Fable 5 model behavior or hidden runtime capabilities.
How do you install or deploy this agent?
Install the stable release with:
codex plugin marketplace add baskduf/FableCodex --ref v0.5.1
codex plugin add codex-fable5@fablecodexRestart Codex afterward. For the development version, use --ref main in the marketplace command. For local development, run codex plugin marketplace add ~/Desktop/FableCodex followed by the same codex plugin add command. No model credentials are documented for the base plugin; the optional provider bridge separately requires valid Anthropic access and an OpenAI-compatible gateway such as LiteLLM.
How do you use this agent?
A first Codex invocation can be:
@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.To use the helper directly, run:
export PATH="$PWD/plugins/codex-fable5/bin:$PATH"Then check it with codex-fable5 status. A starter ledger can be created with:
codex-fable5 goals create --brief "Migration" --goal "inspect::Find current behavior and tests" --goal "change::Implement the migration" --goal "verify::Run tests and inspect output"Run codex-fable5 goals next, supply evidence through goals checkpoint, and execute codex-fable5 findings gate before final completion. The documented test command is python3 -m unittest discover -s tests -v.
How does this agent compare with similar options?
Compared with an ordinary Codex session that does not load the skill, FableCodex adds goal ledgers, evidence checkpoints, review-finding gates, and structured closeout expectations. It adapts Claude/Fable-inspired working habits into Codex-native procedures, but it is neither a replacement for nor a clone of the Fable 5 model.
FAQ
Does this provide access to the Fable 5 model?
Are Anthropic credentials required for the base plugin?
What happens when review findings remain unresolved?
codex-fable5 findings gate fails while findings are open or blocked, and blocking findings also prevent completion of the final goal.Where does it store task information?
.codex-fable5/ directory. The documentation does not describe an external upload mechanism.