Imported from mthds-ai/mthds-plugins (mthds/skills/mthds-recursive/SKILL.md). Install upstream with npx skills add mthds-ai/mthds-plugins --skill mthds-recursive. Copyright stays with the author.

Build a MTHDS bundle top-down (recursive refinement)

Build a .mthds method top-down by stepwise refinement. Capture the whole job as a single pipe signature (the client contract — inputs, output, semantics). Then refine one signature at a time, one level down, into operators and controllers, leaving the not-yet-built parts as further signatures. Validate leniently (--allow-signatures) after each layer; the library is always valid and always resumable. Stop when no signatures remain (strict validation passes → runnable), or stop early and keep the leniently-valid scaffold.

Scope (what this skill can emit)

• Bundle headers: domain, description, main_pipe, system_prompt.
• Concepts: simple, refining, structured (field types text, integer, boolean, number, date, concept, list).
• Pipes: PipeSignature (contract-only header), PipeLLM, PipeCompose, PipeSequence, PipeBatch, PipeParallel, PipeCondition, PipeExtract, PipeSearch, PipeImgGen, PipeFunc.

Outside this skill's scope: dict field types, PipeStructure, inline templating_style blocks, and other advanced features. When the user asks for those, write the closest in-scope equivalent and call out the deviation; do not silently emit unsupported constructs.

How it works — read this first

• The artifact is a library, not one file. A directory of same-domain .mthds files, loaded together with -L. The runtime merges them into one domain; pipes and concepts reference each other across files by bare code.
• Construction is additive. A PipeSignature is a forward declaration ("header"); the concrete pipe is its "definition". Each refinement adds a new <code>.mthds file; no existing file is ever rewritten — no merge step. At merge, a concrete satisfies the signature of the same code (the definition wins), so the same code legitimately appears as a header in one file and a concrete in another.
• Invariant: every unbuilt pipe is a reachable signature, so the assembled library passes lenient validation (--allow-signatures) at every step.
• Operation (one refinement step): take one unimplemented signature, add its definition file one level down — an operator (done), or a controller that wires sub-pipes, forward-declares each not-yet-built sub-pipe as a new header, and owns any intermediate concepts it introduces — then re-validate.
• The backlog is the bundle's own todo list. It is exactly the ## Pending signatures list that validate reports (declared signatures with no concrete definition yet). Drain it round by round until empty → strict validation passes → runnable.

See recursive-cheat-sheet.md for the PipeSignature syntax, the signature_for hint, the operator-vs-controller decision, and all pipe-type field rules — it is the syntax source of truth.

Step 0 — Environment Check (mandatory, do this FIRST)

Run this command to check toolchain status:

# Wrapped in `bash -c` so the bash array syntax below works even when the
# session shell is zsh.
bash -c '
# Pick the cached env-check from the plugin version with the highest semver.
# Matches both `mthds` (prod) and `mthds-dev` (dev) plugin caches. The padded
# segment trick keeps lex order = semver order so 0.10 does not sort below 0.9.
_best_f=""; _best_k=""
for f in "$HOME/.claude/plugins/cache/"*/mthds*/*/bin/mthds-env-check; do
  [ -x "$f" ] || continue
  _v="${f%/bin/*}"; _v="${_v##*/}"
  _k=""; IFS=. read -ra _parts <<<"${_v%%[-+]*}"
  for _p in "${_parts[@]}"; do _p=${_p%%[!0-9]*}; _k="${_k}$(printf %06d "${_p:-0}")"; done
  [[ "$_k" > "$_best_k" ]] && { _best_f="$f"; _best_k="$_k"; }
done
[ -n "$_best_f" ] && exec "$_best_f" "0.22.1"
echo "MTHDS_ENV_CHECK_MISSING"
'

Interpret the output:

• MTHDS_AGENT_MISSING → STOP. Do not proceed. Tell the user:

The mthds-agent CLI is required but not installed. Install it with:

npm install -g mthds

Then re-run this skill.

• MTHDS_AGENT_VERSION_UNKNOWN → STOP. The installed mthds-agent returned an unparseable version. Tell the user:

Could not parse the output of mthds-agent --version. Your installation may be corrupt. Reinstall with:

npm install -g mthds@latest

Then re-run this skill.

• MTHDS_AGENT_OUTDATED <installed> <required> → The installed mthds-agent is too old for this plugin. Do not hard-stop. Instead, tell the user their mthds-agent (v<installed>) is older than the required v<required>, then follow the upgrade flow to offer upgrading mthds-agent via npm install -g mthds@latest. After the upgrade flow completes (whether the user upgraded or declined), proceed to Step 1. The upgrade flow's "Not now" and "Never ask" options let users continue with current versions.

• MTHDS_UPDATE_CHECK_FAILED ... → WARN. The update check command failed. Show the error output to the user. Suggest checking network connectivity and mthds-agent installation. Proceed to Step 1 with current versions.

• UPGRADE_AVAILABLE ... → Read upgrade flow and follow the upgrade prompts before continuing to Step 1.

• JUST_UPGRADED ... → Announce what was upgraded to the user, then continue to Step 1.

• UP_TO_DATE ... → Proceed to Step 1. The line is a terse list of verified installed versions (e.g. UP_TO_DATE mthds-agent=0.10.0 plxt=0.4.0 plugin=0.12.0); if you mention the env-check in your preamble acknowledgement, relay the agent and plugin versions you saw. Two "explicit-quiet" variants share the same prefix and are also clean — proceed to Step 1 without warning, and do not relay the quiet state unless the user is troubleshooting:

  • UP_TO_DATE update-check=disabled — the user has turned update-check off via config.
  • UP_TO_DATE update-check=snoozed — the user has an active snooze on the current version key; an upgrade would otherwise be available, but they explicitly asked for quiet.

• No output → WARN. The env-check produced no output at all, which usually means mthds-agent itself is broken or the wrapper script bailed before printing. Tell the user the environment check could not be confirmed, then proceed cautiously to Step 1.

• MTHDS_ENV_CHECK_MISSING → WARN. The env-check script was not found at either expected path. Tell the user the environment check could not run, but proceed to Step 1.

• Any other output → WARN. The preamble produced unexpected output. Show it to the user verbatim. Proceed to Step 1 cautiously.

Do not write .mthds files until the environment check passes. The CLI is required for validation and formatting — without it the output will be broken and the PostToolUse hook will fail.

No backend setup needed: building and validating never run the method, so no inference backends or API keys are required. Backend configuration is only needed for live execution — use /mthds-runner-setup when ready.

Step 1 — Capture the whole job as one signature (Layer 0)

Read recursive-cheat-sheet.md before writing.

Determine the three things that are the requirement:

• Input concept(s) — what the client provides.
• Output concept — what the client gets back.
• Description — the semantics, in prose precise enough to implement against.

Specify the boundary concepts fully now. The top input and output concepts are the client-facing data contract — give them their structure at Layer 0. Intermediate concepts come later, each introduced by the controller that owns it.

Write the root file mthds-wip/<bundle_dir>/bundle.mthds:

1. mkdir -p mthds-wip/<bundle_dir>/.

2. Write bundle.mthds with the Write tool. It carries the domain header, description, main_pipe, optional system_prompt, the fully-specified boundary concepts, and the top pipe as a single PipeSignature whose code is the main_pipe:

   domain      = "<snake_case_domain>"
   description = "<what the job means>"
   main_pipe   = "<top_pipe_code>"
   
   # boundary concepts — specified fully (structure if structured)
   # ...
   
   [pipe.<top_pipe_code>]
   description   = "<precise semantics of the whole job>"
   inputs        = { <name> = "<InputConcept>" }
   output        = "<OutputConcept>"
   signature_for = "<intended impl type, e.g. PipeSequence>"

3. The root file is written once and persists — refinement never touches it again. The concrete main pipe is added later in its own <top_pipe_code>.mthds file (like any other definition; main_pipe is a reference, not the implementation).

Validate leniently (the one-signature library passes):

mthds-agent validate bundle mthds-wip/<bundle_dir>/bundle.mthds -L mthds-wip/<bundle_dir>/ --allow-signatures --graph

Announce the captured contract in one line (inputs → output, one-sentence semantics) before recursing, so the user can interject — non-blocking. This is the only step with optional interactivity (see Autonomy below).

Step 2 — Refine layer by layer (auto)

Drain the signature backlog breadth-first, serially (one signature at a time — no parallel workers in this version). Repeat this loop until the backlog is empty:

1. Read the backlog. Validate leniently; validate states runnability in plain English and lists what remains:

   mthds-agent validate bundle mthds-wip/<bundle_dir>/bundle.mthds -L mthds-wip/<bundle_dir>/ --allow-signatures --graph

   • Still pending → the markdown shows a ## Pending signatures (N) heading, a ⚠️ This method is NOT yet runnable … line, and one bullet per library-wide pipe still declared as a signature. That bullet list is the backlog.
   • Done → the markdown shows ✅ All pipes are concretely implemented … this method is runnable. and no ## Pending signatures section. The backlog is empty → go to Step 3.

   (No --format json — the agent reads this markdown directly; the verdict line and the backlog are both plain text. Errors on failure stay markdown too.)

2. Expand each pending signature, one at a time (see "To expand one signature" below). Each expansion adds exactly one new <code>.mthds file and never edits an existing one. --graph re-renders dry_run.html each layer so the structure is visible as it grows.

3. Re-validate, recompute pending_signatures, repeat. Expanding a controller forward-declares its children, which become newly pending; expanding an operator removes one. Draining the whole current pending set each round walks the tree breadth-first by construction.

To expand one signature

Given one pending signature S (its frozen contract — inputs, output, description, signature_for):

1. Decide operator or controller. The signature_for hint usually pre-answers it. Heuristic when absent or wrong: a single cognitive/IO step → operator; multiple steps, iteration, branching, or parallelism → controller.

2. Add <code>.mthds to the bundle dir, where <code> is S's bare pipe code — strip the domain prefix. pending_signatures lists each signature namespaced as domain.code, but the file and the [pipe.<code>] it defines use the bare code so they satisfy the forward-declared header (a namespaced [pipe.domain.code] would mint a different pipe and leave the signature pending — the recursion would never converge). Every non-root file carries only domain = "<same_domain>" for membership — omit description, system_prompt, main_pipe (those live in the root).

   • Operator (leaf): write the concrete operator (PipeLLM, PipeExtract, PipeSearch, PipeImgGen, PipeCompose, PipeFunc) and fill its type-specific fields. No new signatures — this branch is done.
   • Controller (composite): write the controller (PipeSequence, PipeBatch, PipeParallel, PipeCondition), wire its sub-pipes, forward-declare each not-yet-trivial sub-pipe as a new PipeSignature header (contract + signature_for) in this same file, and declare any intermediate concepts the wiring needs. The new headers join the backlog.

3. Declare inputs and output explicitly, repeating S's contract. Pipes never infer inputs from prompt sigils, so a definition that omits them mismatches its header. Spelling is free — the contract reconciles by concept identity (bare↔qualified, native equivalents, multiplicity structural), so output = "Brief" and output = "thisdomain.Brief" match.

4. Prevent concept collisions (do this before introducing a new intermediate concept). Check the concept code is not already declared anywhere in the assembled library or pending set; if it is, derive a unique code by namespacing from the parent pipe code. Each concept is declared once, in the file of the pipe that introduces it. (A duplicate is the loud backstop: validate fails with Concept '<code>' is declared in two different bundle files … rename one of the concepts.)

5. Fix each concept's shape when you introduce it — you cannot add fields later. A concept is declared once, in its introducing file; the additive model never lets a later file add structure (a second declaration is the same hard "declared in two different bundle files" error). So decide the shape from the consumers this controller wires: if any downstream consumer field-reads the concept ($x.field, or a construct from = "x.field"), declare it structured now — a field-read on a simple concept fails even under lenient validation. If the concept is only ever consumed whole (@x in a prompt, or mapped wholesale), keep it simple (often refines a native). When a sibling field-reads a concept, hoist that concept up to the common parent controller that wires both its producer and that consumer, and structure it there.

If validation fails after an expansion

The failure is, by construction, in the file you just added — bounding the fix.

• Contract mismatch (the definition's inputs/output diverges from its header): conform the definition to the frozen header. Never edit the header — the parent depends on that contract. (Changing a contract is a propagating change that means revising the parent too; if the header itself is genuinely wrong, stop and flag it rather than silently editing it.)
• Other semantic errors: map the error to the relevant cheat-sheet section, fix the added file with the Edit tool, re-validate. See Error Handling.

Step 3 — Finalize (strict validation)

Once validate reports the ✅ … this method is runnable verdict (no signatures remain), validate strictly (drop --allow-signatures):

mthds-agent validate bundle mthds-wip/<bundle_dir>/bundle.mthds -L mthds-wip/<bundle_dir>/ --graph

Strict validation rejects any reachable signature, so passing it is the gate that says runnable (it reprints the same ✅ verdict). Fix any remaining whole-bundle semantic errors and re-run until it passes.

Step 4 — Deliver

Once strict validation passes:

1. Input schema — Run mthds-agent inputs bundle mthds-wip/<bundle_dir>/bundle.mthds -L mthds-wip/<bundle_dir>/ and show the user the input JSON schema so they can see what the method expects. Do NOT save it to inputs.json — input preparation is handled exclusively by /mthds-inputs.

2. Flowchart — Mention that dry_run.html was generated next to the bundle (refreshed at every layer).

3. Next steps — Suggest:

   Test with mock inference (no real inputs needed):

   mthds-agent run bundle mthds-wip/<bundle_dir>/ --dry-run --mock-inputs
   

   Prepare real inputs with /mthds-inputs, then run:

   mthds-agent run bundle mthds-wip/<bundle_dir>/
   

Early-stop variant. If the user stops before the backlog is empty, deliver the leniently-valid scaffold instead: confirm it passes lenient validation, list the unimplemented signatures (validate's current ## Pending signatures list), and explain that resuming means expanding them — no external state is needed, the bundle is its own todo list. A validated design skeleton is a legitimate deliverable.

NEVER write inputs.json manually. If the user provides files, paths, or wants to run with real data, invoke /mthds-inputs — it handles path resolution (paths must be relative to inputs.json, not CWD), placeholder formatting, and file copying.

Invariants & rules (keep these true at every step)

• Contract stability. A definition preserves its header's inputs/output contract, matched by concept identity (not byte-string). Choose a signature's internals freely; never change its surface without revising the parent too.
• main_pipe is the anchor. The top signature's code and its inputs/output are frozen after Layer 0. Its body arrives as a separate definition file; its identity and surface never change.
• Additive writes. One concrete definition per file; headers persist after they're satisfied. Never overwrite an existing file — always add a new <code>.mthds.
• A concept's shape is fixed at introduction. Boundary concepts structured at Layer 0; an intermediate concept's shape is decided in its introducing file — structured if any consumer field-reads it, simple if consumed whole — and cannot be changed in a later file.
• Backlog = {signatures} − {concretes}. Recomputed each layer from validate's ## Pending signatures list — never hand-tracked, never reconstructed as a depth tree.

Autonomy

This skill is automatic by default.

• Requirements (Step 1) — infer the contract and proceed; always announce it (non-blocking). Engage in discussion only when the request is genuinely ambiguous about inputs/output/semantics, or when the user signals they want to discuss.
• Recursion (Step 2+) — always auto. No per-layer approval prompts. The leniently-valid checkpoints and dry_run.html are the review surface; the user can interrupt at any time.
• Pause and ask only if validation fails twice on the same construct, or a header itself appears wrong (a propagating contract change).

Reference

• Recursive Cheat Sheet — read before writing. The MTHDS code subset this skill writes, including the PipeSignature header and lenient-vs-strict validation.
• Native Content Types — attributes of native concepts (Image.url, Page.text_and_images, ...) for $var.field references and construct from paths.
• Error Handling — read when validate returns errors to determine recovery.
• MTHDS Agent Guide — full CLI command syntax, including --allow-signatures and reading pending_signatures.