Pipelex Organize

Organize a designed MTHDS bundle

The signature-driven mode of /pipelex-design is deliberately additive: every refinement adds a new <code>.mthds file, and satisfied PipeSignature headers linger in the files that declared them. A converged library can therefore be correct but shaped by construction history — one file per pipe, contracts duplicated as stale headers, no relationship between file layout and how the method reads. Direct designs are normally coherent already and skip this skill; they may still use it later on explicit request or when their layout genuinely needs regrouping.

This skill regroups that library into a layout organized for comprehension — the way source code is spread across files in any language: related pipes gathered into one file per coherent unit, so a reader (human or agent) can discover the method progressively — open the root to see the contract and the top-level flow, then descend into just the module they care about. The layout scales with the method: a simple method fits entirely in main.mthds; a complex method with many pipes gets one file per subtree or functional area. Consolidation into fewer files is the usual effect, but the goal is grouping, not a single file.

It is a content-preserving transformation — the method's semantics never change, and the validation verdict before and after must be identical. That equivalence is proven with the mthds_validate tool, never assumed.

Requirements — the Pipelex MCP tool

Equivalence is checked through the mthds_validate tool, served by the plugin's pipelex MCP server. It is required — this skill never reorganizes without proving the verdict is preserved.

• If the tool is absent from this session (the MCP server isn't connected), STOP and tell the user in one line: "The Pipelex MCP server isn't connected — the plugin manifest spawns the local workshop (npx -y @pipelex/mcp@latest), so its absence usually means node/npx is unavailable or the spawn failed. Check the plugin's MCP connection (/mcp)." Do not touch the bundle files without validation available.
• If a call returns status: "error" with an error of class config (missing or rejected PIPELEX_API_KEY, unreachable API), STOP the same way and surface the error's hint verbatim. Never reorganize unvalidated.
• The server authenticates to the validation API with PIPELEX_API_KEY from the session environment — the same variable the plugin's validation hook documents.

Formatting is automatic. Every write of a .mthds file triggers the plugin's validation hook: it lints, rewrites the file in canonical formatting, and blocks on syntax errors. Don't hand-format, and re-read a file before editing it again after the hook reformatted it.

What this skill may and may not change

May change (layout only):

• Which file a concept or pipe declaration lives in, and the file names of non-root files.
• The order of declarations.
• Dropping a PipeSignature header whose code has a concrete definition anywhere in the library (the definition supersedes it — the header is construction scaffolding).
• Absorbing per-file domain = "..." membership lines into whichever files survive.

Must NOT change (semantics):

• Pipe codes, concept codes, the domain, main_pipe, description, system_prompt.
• Any pipe's inputs/output contract, type, prompts, steps, branches, outcomes, or any other body field.
• Any concept's description, refines, or structure fields.
• Unsatisfied signatures. A PipeSignature with no concrete definition (an early-stopped scaffold's backlog) is kept — deduplicated to one header per code, never dropped. The bundle stays its own todo list.

If reorganizing seems to require a semantic edit (a rename, a contract fix, a missing declaration), STOP and report it — that is /pipelex-design territory, not organization.

The target layout — group for comprehension

Think of it exactly like organizing source code across files. The unit of grouping is a coherent piece of the method: a controller together with the sub-pipes only it wires (its private subtree), or a functional area several small pipes serve.

• Root main.mthds is the entry point and table of contents. It carries the bundle header (domain, description, main_pipe, system_prompt), the boundary concepts (the client-facing contract), and the main pipe. Reading it alone tells you what the method does, what goes in and out, and what the top-level steps are. The root file is always named main.mthds unless the user says otherwise — a root inherited under another name (e.g. a legacy bundle.mthds) is renamed as part of organizing.
• One file per module. Each top-level subtree (or clearly-named functional area) gets its own <snake_case_name>.mthds file — named after the controller that heads it or the area it implements — containing that subtree's pipes in flow order (controller first, then the pipes it wires) plus the intermediate concepts introduced for that subtree. Each non-root file starts with domain = "<same_domain>" only.
• Shared declarations go up. A pipe or concept used by several modules lives in the root file (or, if the shared surface is large, a dedicated shared.mthds), never duplicated.
• Scale the file count to the method, in both directions. A simple method — a handful of pipes — belongs entirely in main.mthds; don't scatter it. A complex method with tens of pipes needs several module files; don't cram it into one. And never keep a file per pipe: that is the construction sprawl this skill exists to clean up. Each file should read like a chapter, not a line.
• Progressive discovery is the test. An agent that needs to understand or modify one step should be able to read the root plus one module file and have everything relevant — nothing important hidden in an unrelated file, no file that can't be understood without opening all the others.

Within every file: satisfied signature headers are dropped; a still-pending signature (scaffold case) sits in the module of the controller that wires it.

Procedure

Step 1 — Baseline verdict

1. Gather all .mthds files in the bundle directory (e.g. pipelex-wip/<bundle_dir>/).
2. Call mthds_validate with files for every file. Prefer the path form {path: <absolute path to the file>} — it keeps the real path as provenance in diagnostics and spares copying whole bundles into the request; the workshop resolves a path against its own working directory, so pass an absolute one. Inline {content: <file content>, uri: <path relative to the bundle dir>} is the fallback, and the only form the hosted console accepts.
3. Record the baseline: is_valid, is_runnable, and the exact pending_signatures set.

Branch on the structured verdict:

• is_valid: true → proceed (runnable or scaffold — both are organizable).
• is_valid: false → STOP. Do not reorganize a broken library — report the verdict and point to /pipelex-design to fix it first. Organization must start from, and preserve, a passing verdict.
• status: "error" → class input_domain: fix the call; class config: stop per the rule above; class runtime: report and retry once before stopping.

Step 2 — Plan the layout, compose the files

1. Map the structure: from the main pipe down, identify the subtrees and which pipes/concepts belong to each; identify shared declarations.
2. Decide the file set using the target-layout rules above — possibly just main.mthds, possibly root + several module files. State the plan in one line (e.g. main.mthds + extract.mthds + analyze.mthds + report.mthds).
3. Compose every file in memory (do not write yet): copy each declaration verbatim into its assigned file, one declaration per concept, one entry per pipe, ordered for top-down reading within each file.

Step 3 — Prove equivalence before touching disk

Call mthds_validate with the composed candidate set (all planned files, nothing else). The verdict must match the baseline exactly: is_valid: true, the same is_runnable, and an identical pending_signatures set.

• Match → proceed to Step 4.
• Mismatch or failure → the composition dropped or duplicated something; fix the composition (never the semantics) and re-validate. If it still fails after two fix attempts, STOP, leave the original layout untouched, and report the discrepancy.

Step 4 — Swap the layout

Only after the candidate verdict matches:

1. Write every file of the new layout (Write tool — the hook lints and reformats each in place; a new file may legitimately reuse an old file's name, e.g. main.mthds).
2. Delete every .mthds file that is not part of the new layout. Delete only .mthds files; leave inputs.json, input files, and anything else in the directory alone.
3. Confirm on disk: re-gather the directory's .mthds files and validate once more — this catches anything the formatting hook changed.
4. If that confirmation fails, restore the original layout — never leave the directory unconfirmed. On status: "error" (no verdict — e.g. the MCP dropped mid-swap) or a verdict that does not match the baseline, roll back: you still hold every original file's content from Step 1 — rewrite the original files, delete the new-layout files that were not in the original set, and report the failure with the layout left as it was. The swap ends either proven equivalent or fully rolled back.

Step 5 — Report

One short summary: the layout (which files, what each contains, one line per file), the preserved verdict (runnable, or valid scaffold with its pending list). No approval prompts — by the time you report, the bundle is organized and proven equivalent.

Autonomy

This skill is fully automatic once invoked — no per-step approval, including the layout decision (announce it, don't ask). /pipelex-design auto-invokes it after converged signature-driven construction or re-entry when the layout still reflects construction history; an already coherent direct result does not invoke it solely for process compliance. It can also be invoked explicitly on any designed bundle directory. The only stops are the ones above: missing/misconfigured MCP tool, a failing baseline verdict, or a candidate that cannot be proven equivalent. If the user has expressed a layout preference (single file, specific grouping), honor it — the equivalence proof works the same for any layout.