fable-sweep git:20261108.8f3234fA

Immutable. This exact content never changes and is served at /api/v1/blob/fe609add95be1b79.

---
name: fable-sweep
description: Implement renames, migrations, generated-surface sync, cross-cutting cleanup and connected-workflow repairs. Select automatically for complex Agoragentic repairs, including repeated integration failures, multi-process launchers, generated/hash-bound packages and timed reviewer or installation handoffs, even when the user does not name Fable. Diagnose and qualify locally before an attended or paid attempt. Routine isolated edits do not need the full workflow; review-only requests use fable-deep-review or fable-verify.
---

# Fable Sweep

Perform broad changes with discovery first and scoped edits second. Do not mutate unrelated worktree changes.

For ECF-style governed runs, use `../../references/ecf-run-contract.md`. For large sweeps, use real Codex subagents when the runtime exposes a subagent tool and the user has not opted out; treat repo-wide renames, migrations, consistency passes, generated-surface syncs, docs/tests syncs, and "find every place" tasks as large by default. Otherwise run `single-agent multi-lens` and say why no subagents were used when workflow trace is requested.

Use the centralized capability-aware profiles in `../../references/model-support.md` and `../../config/model-profiles.json`. Keep model and reasoning selections configurable, preserve explicit overrides, and record the requested and actual coordinator model and effort and requested and actual worker model plus any fallback. Request parallel delegation only for useful disjoint lenses; report `single-agent multi-lens` when no workers ran. The coordinator owns synthesis and final verification. Never claim a model, effort or subagent ran without runtime evidence.

## Workflow

1. Restate the target change and authority boundaries.
2. Declare the ECF run mode when the user asks for ECF, subagents, or a receipt. Use `multi-agent` only when real subagents will be spawned.
3. Read repo instructions and identify generated, public, config, docs, tests, and serialized surfaces that may need updates.
4. Build a complete hit list using fast search first:
   - `rg` for identifiers, strings, docs, config, tests, migrations, snapshots, and generated files
   - language-aware search or AST tools when available and useful
   - import/caller lookup for behavior changes
5. Classify hits:
   - must edit
   - must preserve for compatibility
   - generated or external
   - uncertain
6. Make scoped edits only after the discovery map is clear. Preserve existing style and local helper APIs.
7. Run targeted validation. Include tests, lint, type checks, snapshot regeneration, or schema validation when relevant.
8. Report changed files, preserved compatibility points, validation, and remaining uncertain areas.

## Evidence Safety

Never print raw secrets, tokens, private keys, wallet keys, credential files, or `.env` values. Redact secret-like values and cite only the file/path/key name needed to explain the issue.

## Safety

Stop and ask before irreversible changes, external publishing, deploys, credential mutation, wallet spend, trust-state mutation, or destructive cleanup. For normal code edits, proceed when the user asked for implementation.

For write-enabled parallel sweeps, assign each worker a disjoint write scope and tell workers they are not alone in the codebase. The main agent owns final integration, conflict resolution, validation, staging, commits, pushes, and external side effects.

When requested, include a compact `Workflow Trace` with mode, ECF contract status, lenses covered, spawned agents or no-subagent reason, verification method, and coverage gaps.

## Automatic complex-repair routing

When an authorized repair concerns Agoragentic (including Agent Marketplace,
Triptych OS / Agent OS, Interchange and Fable) and has repeated integration
failures, spans process/reviewer boundaries, changes generated/hash-bound packages,
or couples live/attended actions to timed handoffs, select this workflow without
requiring the user to name Fable or repeat the procedure. Also use it for equivalent
connected-workflow repairs in other projects when this skill is selected. Read
[Repair and qualify a connected workflow](../../references/complex-workflow-qualification.md).
Measure the selected route, repair its producers, and qualify the assembled
package locally before an attended or paid attempt. Apply only relevant stages and
reuse unchanged evidence. A routine typo or isolated low-risk edit does not trigger
the full procedure. Review-only requests stay with `fable-deep-review` or
`fable-verify`; they do not become repair authorization.

Automatic selection means choosing guidance within the current task, not starting
background work or adding authority. Honor current instructions, paused goals and
the designated repair owner. Do not resume Interchange or compete with another
operator merely because these instructions match; do not solicit MFA or restart
approval while it is paused. Complete local preparation
before asking for MFA or proposing another paid run.

## Optional development playbooks

When the user requests an indexed delivery workflow, scaffold completion, or a
named development recipe, consult `../../prompts/development-index.md` and
`../../references/development-playbooks.md`. Use
`../../scripts/development-prompts.mjs` to search or show one recipe when Node is
available; otherwise read the matching entry in the bundled catalog. Do not load
all recipe bodies by default. Route each stage to the existing skill named in its
entry and retain the normal discovery, edit, and verification workflow above.

A playbook is guidance, not an execution engine or an authorization grant. Resolve
actual refs and edit scope, preserve current ECF/model/worker limits, stop dependent
stages when prerequisites fail, and end with the completion-proof reconciliation.
For read-only requests, return a plan rather than executing scoped-edit stages.