/suede-graph-flo-xr · Suede Thought Graph shipping search
Competing plans. One selected build.
Suede Graph Flo XR generates competing implementation plans, scores and prunes them, adversarially refutes and improves the survivors, aggregates compatible lanes, then builds, reviews, and gates only the selected plan. Light, standard, and deep runs are capped at 55, 110, and 200 total agent calls. The bundled runner requires Claude Code on macOS, registered Suede Graph Flo XR agent profiles, and sandbox-exec. It never deploys.
"ship this change across the whole surface"->/suede-graph-flo-xrAsk in plain words. The router reads the request and loads this lane; MCP agents find it with list_suede_skills.
For this manual copy, pass an empty agentNamespace string in the workflow args. The caller must pass agentNamespace: "suede-skills" for the full plugin or agentNamespace: "suede-agent-workflows" for the focused plugin. Missing and unknown values fail before the first agent call.
Create one exact origin/main worktree through a clamped Scout setup command, then research with separated local-reader and public-web-reader profiles.
Generate competing implementation plans, then Score and KeepBestN to prune the beam deterministically.
Refute surviving plans adversarially and Improve the ones whose objections are repairable.
Aggregate compatible lanes without merging conflicting file ownership, then Select one winner.
Allow only the selected plan to mutate; rejected, pruned, and unselected thoughts remain evidence.
Gate the integrated winner on exact allowlisted local checks for web, native iOS, and Android projects, then write an evidence handoff without deploying.
Keep every applied blocker repair on hold until an independent follow-up verifies the original failure scenario.
When to use it
Any multi-file change where competing implementation plans should be compared before the worktree changes.
Work in a repo where several sessions run at once and a lane collision would otherwise clobber a sibling branch.
Changes where you want the findings refuted before you read them, not after you ship them.
Choose the budget before launch: light permits 55, standard 110, and deep 200 total agent calls.
Not for: high-volume, well-specified work that splits into independent worker-sized tasks (run those as a separate private worker-fleet pass).
Not for: findings-only review with no code change (use suede-code-review).
Not for: wiring CI and branch protection (use suede-ci-gate).
Runtime and evidence boundary
Patch authors have no mutation tools. Appliers and verifiers have Bash plus structured output only, and their Bash forms are clamped. The requested Scout command probes macOS sandbox-exec before fetch or worktree creation, parses NUL-delimited Git status and exact lsof -Fn CWD fields, fails closed on safety-manifest overflow, and creates a private per-run temp root. Gate changes into the attested worktree and runs allowlisted validation under sandbox-exec with network denied. Reads are limited to runtime and system roots, the worktree, the Git common directory derived inside the exact Gate clamp, and the private temp root; the model-reported common directory cannot widen sandbox permissions. Writes are limited to generated artifacts, validated nested module build roots, and that temp root. The source digest includes reported file contents, including untracked additions, and is attested before and after Gate. An applied blocker repair remains pending semantic verification and forces a hold until its original failure scenario is checked independently.
The distinction matters: Claude Workflow does not return a trusted required-tool-call receipt. A clamp limits Bash when invoked, while the structured result remains a model attestation. The metadata fields named authority, allowedFiles, and allowedCommands are audit labels, not host filesystem permissions. Do not describe the result as host-certified.
Best prompt
Use $suede-graph-flo-xr on /absolute/path/to/my-app to add per-account rate limiting to every API route. Scope: preserve the existing auth contract and unrelated WIP. Budget: standard. Search competing plans and mutate only the selected winner. Read production if needed, but do not deploy.
Use the full or focused Claude plugin on macOS to run the hardened workflow. The clone installer also copies its six fixed agent profiles. Codex installs the orchestration contract, not this Claude Workflow runtime.
Skill-only and generic skills-CLI installs do not register the six Claude agent profiles. Use the Claude plugin or clone installer for the bundled runner. Every path is on the install page, and every folder is readable first on GitHub.
Proof is part of the release.
First install takes two commands. After the marketplace is added, install is one. Every skill is plain Markdown.