Fable-5 for Codex

Before I merge, show what the coding agent actually finished.

Fable-5 is a completion-verification plugin for Codex from Agoragentic. Start with $fable-verify: map acceptance criteria to exact source, observed tests and remaining unknowns. Quick and deep modes produce a readable report and versioned JSON artifact. The six existing audit, review, fact-check, understanding, design and sweep workflows remain available.

Working example from this source checkout (Node 18+, Git):

node examples/completion-verification/demo.mjs

It runs a real local assertion in a disposable synthetic repository and correctly leaves deployment unverified. It does not install a plugin or prove effectiveness.

Each workflow is designed to record:

• scope and authority;
• evidence policy;
• verification performed;
• explicit unknowns;
• whether real subagents were used;
• a truthful final Workflow Trace.

To use the new skill from this tested source checkout (back up an existing copy first; see the upgrade guide):

node bin/install.mjs --no-codex-add
codex plugin add fable5-codex@personal

Then start a new Codex thread:

Use $fable-verify in quick mode.
Task: implement the acceptance criteria in issue #123.
Target: this branch. What is verified, incomplete or unverified before merge?
Return requirement-to-evidence JSON, a readable report and a compact handoff.

Install · See the workflows · Read sample outputs · Inspect benchmark evidence

Release status: v0.4.0-alpha.2 remains the latest public tagged release. Main was inspected at 0.4.0-alpha.3; this completion-verification upgrade is unreleased source 0.5.0-alpha.1. The tagged release does not contain the new primary skill. Test/install this checkout to use it; source, installed, activated, released and published are separate states.

Completion verification

Each criterion records its originating requirement, explicit/assumed status, source changes, exact head and working-tree/index snapshot, evidence method, observed command/result/context, conclusion, open work and smallest next action. PR head, base and merge-test commits stay distinct. Green CI never substitutes for requirement coverage. Skipped/pending/inaccessible checks are not passes; mocks do not establish live execution or deployment.

Saved reports are untrusted records. Structure validation and hashes prove no semantic correctness; rendering a saved artifact leads with unverified until observations and judgments are freshly checked. Verification never silently repairs, merges or deploys. Handoffs require revalidation and grant no authority.

See artifact/API details, model support and upgrade/rollback and privacy.

What Fable-5 changes

A normal coding-agent prompt can produce a plausible review without making it obvious:

• which files were actually inspected;
• whether runtime behavior was verified;
• which assumptions remained unresolved;
• whether parallel agents really ran;
• who owned the final write, commit, or deployment decision.

Fable-5 adds a repeatable evidence and authority contract around that work.

user goal
→ ECF-style run contract
→ one or more explicit review lenses
→ source/runtime evidence collection
→ main-agent verification
→ ranked findings or implementation
→ Workflow Trace + unknowns

Fable-5 provides workflow discipline. It does not make every finding correct, replace local verification, or turn model output into an audit opinion or security certification.

Install

Public tagged release

npx github:rhein1/fable5-codex#v0.4.0-alpha.2

On Windows, activate the copied personal plugin after the installer completes:

codex plugin add fable5-codex@personal

The installer prints the exact activation command because it deliberately does not launch the Codex shim through a Windows command shell.

For a project-local marketplace:

npx github:rhein1/fable5-codex#v0.4.0-alpha.2 --project

Local checkout

git clone https://github.com/rhein1/fable5-codex.git
cd fable5-codex
codex plugin marketplace add .
codex plugin add fable5-codex@fable5-local

Restart Codex if required, start a new thread, and call a skill directly.

npm status

The package metadata and installer are publish-ready, but the public install path remains GitHub until npm publication occurs. After publication, the intended shorter path is:

npx fable5-codex

Do not represent that npm command as currently published unless the registry confirms it.

Primary verification and six existing workflows

Skill	Use it for
$fable-verify	Requirement-mapped completion verification, quick/deep evidence, stale-state checks and handoff
$fable-audit	Ranked bugs, risks, integration gaps, and docs-vs-reality findings
$fable-deep-review	PR or branch review with verification passes and actionable findings
$fable-fact-check	Claim-by-claim verification against repository or runtime evidence
$fable-understand	Source-grounded explanation of how a system actually works
$fable-design-options	Alternative designs with tradeoffs, migration notes, and decision criteria
$fable-sweep	Repository-wide discovery, implementation, and verification workflow

Fastest audit prompt

Use $fable-audit.
Scope: this repository.
Focus: correctness, security, privacy, data, operations, tests, and docs-vs-reality.
Rank findings by severity.
Verify candidates locally before reporting them.
Include explicit unknowns and a Workflow Trace.

Fastest understanding prompt

Use $fable-understand.
Scope: this repository.
Question: What are the main execution paths, trust boundaries, and unresolved unknowns?
Cite exact files and include an UNKNOWNS section.

Configurable Codex model profiles

Balanced and deep profiles prefer gpt-6.1-sol with medium/max reasoning; economical work prefers gpt-6-luna. Preferences live in one versioned config. Wrappers discover the selected client's catalog, validate model-specific efforts and explicitly report supported fallbacks. An older npm shim and a desktop native executable may expose different catalogs. Catalog presence does not prove account access or that a model ran. Interactive sessions retain their selected parent; explicit overrides win.

Use the ready-to-copy template:

plugins/fable5-codex/config/model-profiles.json

Or run the PowerShell wrapper:

.\plugins\fable5-codex\scripts\fable5-codex.ps1 -Mode verify -Profile balanced -Scope . -DryRun

Override coordinator or worker settings when a task needs a different cost/capability balance. The model/runtime configuration does not weaken Fable-5's evidence, authority, and truthful-reporting requirements.

See current model/client support. Codex Ultra orchestration is not a portable API reasoning effort. Historical benchmark models/results below remain unchanged.

Subagents without fake parallelism

Fable-5 requests real Codex subagents for useful large or high-risk work when supported and not opted out. Quick verification avoids unnecessary delegation. Workers use the resolved supported profile and reports retain requested/actual settings and fallbacks.

Large/high-risk includes work such as:

• repository-wide audits or migrations;
• exhaustive review or broad sweeps;
• money, billing, wallets, or settlement;
• authentication, authorization, privacy, or secrets;
• data migrations or idempotency;
• public APIs or serialized contracts;
• deployment or production operations.

When no subagent tool is available, the workflow runs as single-agent multi-lens and says so. It must not imply independent parallel review without real subagent IDs.

Example:

Use $fable-audit with real Codex subagents and an ECF run contract.
I explicitly authorize parallel read-only subagents for this run.
Scope: src/billing.
Spawn four lenses:
1. correctness and integration
2. security, privacy, and authorization
3. data, migrations, and idempotency
4. operations, tests, and docs-vs-reality
The main agent must verify candidates locally before final findings.
Do not claim multi-agent mode unless real subagent IDs exist.
Include the ECF contract, unknowns, and Workflow Trace.

ECF run contracts

Fable-5 uses a public ECF-style run contract as the governance layer for each workflow. The contract records intent and evidence rules; the Codex runtime remains responsible for providing actual tools and subagent execution.

• Run-contract reference
• Starter contract JSON
• PR review contract
• Ledger schema

Authority split:

subagents
→ research, map, plan, draft, find, or verify inside assigned read-only lenses

main agent
→ verify findings and own final output

owner / authorized host
→ writes, commits, pushes, comments, deploys, publishes,
  changes credentials, or performs money/wallet actions

The public plugin contains Micro ECF-style contracts and reporting rules. It does not include private Full ECF internals.

Demo

The animation illustrates the packaged workflow. The repository files, examples, tests, and benchmark records remain the source of truth.

Examples gallery

• Audit sample: payment risk
• Fact-check sample: status claims
• Understand sample: boot flow
• Deep-review sample: pull request

Samples demonstrate format and evidence discipline. They are not proof that every future run will find the same issues or reach the same conclusion.

Benchmark snapshot

The latest published measured repository run is 20260713T234332Z. It used gpt-5.6-sol, matched ultra reasoning effort, a 600-second limit per trial, and three intentionally small fixtures.

The historical run reported:

• workflow-format/lexical rubric composite: 81.7 → 100.0;
• expected-concept recall: 93.3 → 100.0;
• evidence markers: 78.3 → 100.0;
• explicit unknowns: 0.0 → 100.0;
• average wall time: 144.5s → 344.0s (2.38×).

All six final trials completed.

Required qualifications

• The fixtures were intentionally small.
• The measurements evaluate workflow-format and lexical rubric signals, not broad model quality.
• Subagents were disabled to isolate workflow discipline.
• The run predates alpha.3 environment-isolation hardening.
• Its baseline ignored user config while the plugin arm used the active installed-plugin environment, so it is not clean plugin-only causal attribution.
• One plugin trial encountered provider capacity and was retried with the same model, effort, fixture, and mode.
• A complete isolated alpha.3 run is required before replacing the qualified charts.

See the benchmark methodology and raw outputs.

Repository layout

.agents/plugins/marketplace.json       repo-local plugin catalog
assets/brand/                          product and social images
assets/benchmarks/                     qualified benchmark charts
benchmarks/                            harness, raw outputs, and methodology
bin/install.mjs                        GitHub/npx installer
docs/                                  install, method, architecture, and schemas
evals/                                 audit, fact-check, and sweep fixtures
examples/                              prompts and expected reports
examples/gallery/                      polished sample outputs
plugins/fable5-codex/                  installable Codex plugin
plugins/fable5-codex/references/       ECF/run-contract guidance
plugins/fable5-codex/templates/        config and contract templates
scripts/                               validation, charts, demo, and sync tools
test/                                  package and behavior tests

Validation

npm test
npm run validate
npm run validate:artifact
npm run pack:dry-run

For a local installed-plugin smoke test:

Use $fable-fact-check.
Document: README.md.
Check every installed, supported, validated, and works claim against files on disk.
Include exact citations and explicit unknowns.

Treat plugins/fable5-codex/ in this repository as canonical. Use the provided sync script for a personal copy:

.\scripts\sync-personal-plugin.ps1

Product boundary

Fable-5 is:

• an installable Codex plugin;
• evidence-first workflow guidance and validation;
• explicit authority and verification discipline;
• compatible with single-agent and real-subagent runs;
• public Micro ECF-style run contracts;
• reproducible examples, evals, and benchmark artifacts.

Fable-5 is not:

• an independent security audit;
• a guarantee that all bugs will be found;
• a hosted agent runtime;
• wallet or settlement infrastructure;
• private Full ECF;
• automatic authority to write, commit, push, deploy, publish, or spend.

Where this fits

Fable-5 keeps its own product identity while belonging to the Agoragentic trust stack.

Fable-5
→ evidence-first engineering workflows inside Codex

Micro ECF / ECF Core
→ persistent local context, policy, provenance, and source routing

Harness Core
→ tool/action policy, approvals, evidence, and local receipts

Triptych OS
→ governed hosted-agent runtime

Marketplace / Interchange
→ agent work, payments, and cross-market reconciliation

• Agoragentic Integrations
• Micro ECF
• ECF Core
• Harness Core
• Triptych OS
• Marketplace
• Interchange

Use the canonical ecosystem profile for current portfolio metadata. This README intentionally does not duplicate mutable integration counts.

Brand assets

• assets/brand/fable5-hero.png — original README hero
• assets/brand/fable5-social.png — repository and link-card preview
• assets/brand/fable5-sol-ultra.png — Sol Ultra profile
• assets/brand/fable5-mark.png — compact plugin mark
• plugins/fable5-codex/assets/fable5-demo.gif — short install/run/trace demo
• assets/benchmarks/fable5-benchmark-*-qualified.png — visibly qualified historical charts

Repository owners still need to upload the chosen social card in GitHub settings; committing an image does not change the repository social-preview setting.

Security and license

See SECURITY.md for vulnerability reporting.

MIT. See LICENSE.