Fakeble is the best local approximation we can build today of a Fable/Fusion-style agent panel using tools people can actually run:
- Pi as the host shell
- Claude Code as one guest
- Codex as one guest
- Salon for the multi-agent panel
- tmux for panes and process control
- subscription login flows instead of API keys
It is not real Fable. It is Fakeble: good enough to run an independent Claude plus Codex review loop from the terminal when the official product or shared API surface is not available.
For a new machine:
git clone https://github.com/Runpoint-Partners/fakeble.git
cd fakeble
./bin/fakeble install --yes
fakeble doctorThen launch a panel from the repo you want inspected:
fakeble .Inside the host pane, use:
/fakeble Compare the architecture and implementation risks of this change. Claude should focus on ambiguity and failure modes. Codex should focus on repo structure, tests, and edge cases. Cross-review before synthesis.
Point an assistant at this repo and say:
Read AGENTS.md or CLAUDE.md, run the Fakeble doctor, install what is missing, stop at any browser/passkey/MFA login boundary, then teach me how to use fakeble.
The assistant should:
- Run
fakeble doctor. - If setup is incomplete, run
fakeble install --yes. - Stop and ask the human to finish Claude/Codex subscription login if required.
- Run
fakeble doctoragain. - Start with
fakeble .in the target repo. - If it starts a panel, keep monitoring it until final synthesis or a real human-owned blocker.
The useful pattern is a small panel:
- one host that coordinates the review
- one Claude guest that is good at ambiguity, assumptions, architecture, and risk
- one Codex guest that is good at implementation reality, repo mechanics, tests, and edge cases
- an explicit cross-review step before synthesis
The fragile part is that the current stack is glued together through CLIs and tmux. Fakeble packages the glue, the patches, and the operating instructions so the setup is repeatable.
Fakeble reduces the stuck-pane problem, but it is still terminal automation. If an assistant launches a Fakeble panel for a human, the assistant should supervise it:
- capture the host pane every 30-60 seconds while work is active
- confirm each guest receives the prompt and actually answers
- nudge a stuck Codex pane with
Enterif it shows[Pasted Content ...] - ask the host to continue, cross-review, or synthesize when all guests have answered
- report only when the panel reaches synthesis, final review, or a real auth/trust/MFA blocker
That supervision loop is the best current substitute for first-party Fable reliability.
Fakeble applies a small Salon patch that:
- submits Codex prompts with
Enter, which avoids the stuck[Pasted Content ...]prompt problem - retries tmux host launch when tmux has a transient respawn failure
- makes Salon use the real
pion your PATH instead of an older bundled copy - allows
SALON_CODEX_MODELandSALON_CODEX_EFFORToverrides - chains Codex notify hooks instead of discarding a user's existing notify command
The fakeble wrapper also unsets ANTHROPIC_API_KEY and OPENAI_API_KEY before launching Salon so subscription-backed Claude/Codex logins win over API-key behavior.
Run health checks:
fakeble doctorInstall or repair local setup:
fakeble install --yesLaunch in the current repo:
fakeble .Launch against another folder:
fakeble /path/to/projectUse cheaper/lighter settings:
SALON_HOST_MODEL=openai-codex/gpt-5.4-mini \
SALON_GUEST_EFFORT=low \
SALON_CODEX_EFFORT=low \
fakeble .- This depends on local CLI auth. Fakeble cannot complete browser login, passkeys, or MFA for you.
- Panels use more quota than a single-agent turn.
- tmux/TUI automation is inherently more fragile than a first-party API.
- Letting multiple agents edit the same dirty worktree is risky. Use review-only prompts, branches, or worktrees for anything serious.