fable-dispatch

Fable dispatch

How every fable skill reaches Fable 5.1 on the current harness. The calling skill owns the prompt, the subagent type, and what happens with the result. This skill owns the path to Fable 5.1. A downgrade to another model is the ladder's last step, and it is always reported.

1. Detect the harness

Run one Bash check before dispatching:

[ -n "$CLAUDECODE" ] && echo claude-code || echo other

Claude Code sets CLAUDECODE=1. Every other result means "not Claude Code". Do not probe for Cursor or Codex variables (their names are unverified), and never infer the harness from an Agent-call error: a harness that maps fable to another model raises no error. That silent substitution is the defect this ladder closes.

2. The fallback ladder

1. Claude Code ($CLAUDECODE set): the Agent tool with model: fable, as the calling skill's dispatch step describes.
2. Other harness, and command -v claude succeeds: the CLI shim (section 3). The work still runs on Fable 5.1 at the intended effort.
3. Neither path, or a shim call that failed after the section 7 retry: the most capable model available, with the downgrade reported. In order: (a) the harness's own subagent facility with its most capable model, keeping the calling skill's isolation pattern; (b) a completed shim result that another model served, adopted per section 5; (c) no subagent facility at all: do the work inline in the main context and make no writes during it, so the calling skill's read-only promise holds. The report names both what failed and which model ran instead.

3. The CLI shim

Write the prompt to a file, then feed that file to stdin:

claude -p --model fable --effort <tier> \
  --output-format json --permission-mode plan \
  --allowedTools <read-only command list> \
  < <prompt-file>

• Never interpolate the prompt into the command line. A prompt containing ", `, or $ breaks the quoting or executes (the class tests/prompt-shell-safety.test.js guards). Stdin from a file passes the prompt as data.
• --permission-mode plan keeps the subagent read-only. --dangerously-skip-permissions must never appear in the shim.
• --allowedTools names every read-only command the calling skill's procedure runs. Example: fable-validate runs gh issue view --comments and gh pr list, so its list is "Bash(gh issue view *)" "Bash(gh pr list *)". Add git log/git diff forms when the procedure reads history, and -R owner/repo forms when it reads another repo. An empty list is valid for a subagent that only reads files. Plan mode auto-approves built-in reads. A headless run has no user to answer any other prompt, so a missing entry lands in .permission_denials and the result is incomplete. The list carries read-only commands only: never a bare Bash, never an edit, comment, or push command. Plan mode does not block an allow-listed Bash write (verified live), so the read-only list is the shim's only guard against writes.
• --effort carries the tier directly. Fable defaults to high; pass xhigh only when the user asked for it or stamped it.
• Timeout. A Fable 5.1 run at high takes minutes and exceeds a host's default Bash timeout, which kills the call mid-run. Set the host's maximum timeout on the call, or start the CLI in the background with output redirected to a file and poll for exit.