mthds-run（运行方式）

Run MTHDS methods

Execute MTHDS method bundles and interpret their JSON output.

Process

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

Run mthds-agent --version. The minimum required version is 0.1.2 (declared in this skill's front matter as min_mthds_version).

• If the command is not found: 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.

• If the version is below 0.1.2: STOP. Do not proceed. Tell the user:

This skill requires mthds-agent version 0.1.2 or higher (found *X.Y.Z*). Upgrade with: `` npm install -g mthds@latest `` Then re-run this skill.

• If the version is 0.1.2 or higher: proceed to the next step.

Do not write .mthds files manually, do not scan for existing methods, do not do any other work. The CLI is required for validation, formatting, and execution — without it the output will be broken.

Step 0.5 — Pipelex Runtime Check (mandatory)

Running methods requires the Pipelex runtime to be installed and configured.

First, check if pipelex-agent is available:

pipelex-agent --version

• If the command is not found: STOP. Tell the user:

The Pipelex runtime is not installed. Install it with: `` curl -fsSL https://pipelex.com/install.sh | sh ``

Then, if pipelex-agent is installed, check configuration health:

mthds-agent pipelex doctor

• If the doctor reports config issues (missing backends, missing API keys) AND the user is requesting a live run (not --dry-run): STOP. Tell the user:

Pipelex needs to be configured with inference backends before running methods. Use /pipelex-setup for guided configuration.

• If the user is requesting a dry run (--dry-run): config issues are OK — dry runs work without backend configuration. Proceed.
• If healthy: proceed to Step 1.

Step 1: Identify the Target

Directory mode (recommended): Pass the pipeline directory as target. The CLI auto-detects bundle.mthds, inputs.json, and sets -L automatically — no need to specify them explicitly. This also avoids namespace collisions with other bundles.

Step 2: Prepare Inputs and Check Readiness

Fast path — inputs just prepared

If inputs were already prepared during this conversation — via /mthds-inputs (user-data, synthetic, or mixed strategy), or by manually assembling inputs.json with real values earlier in this session — skip the schema fetch and readiness check. The inputs are ready. Proceed directly to Step 3 with a normal run.

This applies when you just wrote or saw inputs.json being written with real content values. It does NOT apply after /mthds-build (which saves a placeholder template) or after /mthds-inputs with the template strategy.

Full check — cold start

If /mthds-run is invoked without prior input preparation in this session, perform the full readiness check:

Get the input schema for the target:

mthds-agent pipelex inputs bundle bundle.mthds

Output:

{
  "success": true,
  "pipe_code": "process_document",
  "inputs": {
    "document": {
      "concept": "native.Document",
      "content": {"url": "url_value"}
    },
    "context": {
      "concept": "native.Text",
      "content": {"text": "text_value"}
    }
  }
}

Fill in the content fields with actual values. For complex inputs, use the /mthds-inputs skill.

Input Readiness Check

Before running, assess whether inputs are ready. This prevents runtime failures from placeholder values.

No inputs required: If mthds-agent pipelex inputs bundle <file>.mthds returns an empty inputs object ({}), inputs are ready — skip to Step 3.

Inputs required: If inputs exist, check inputs.json for readiness:

1. Does inputs.json exist in the bundle directory?
2. If it exists, scan all content values for placeholder signals:

- Template defaults: "url_value", "text_value", "number_value", "integer_value", "boolean_value", or any value matching the pattern *_value - Angle-bracket placeholders: values containing <...> (e.g. <path-to-cv.pdf>, <your-text-here>) - Non-existent file paths: url fields pointing to local files that don't exist on disk

Readiness result:

• Ready: inputs.json exists AND all content values are real (no placeholders, referenced files exist) → proceed to Step 3 with normal run
• Not ready: inputs.json is missing, OR contains any placeholder values → proceed to Step 3 with dry-run fallback

Step 3: Choose Run Mode

If inputs are not ready

Default to --dry-run --mock-inputs and inform the user:

"The inputs for this pipeline contain placeholder values (not real data). I'll do a dry run with mock inputs to validate the pipeline structure."

After the dry run, offer the user these options:

• Prepare real inputs — use /mthds-inputs to fill in actual values, then re-run
• Provide files — if the pipeline expects file inputs (documents, images), ask the user to supply file paths
• Keep dry run — accept the dry-run result as-is

Run modes reference

Graph by default: Execution graphs (live_run.html / dry_run.html) are now generated automatically. Use --no-graph to disable.

Inline JSON for Inputs

The --inputs flag accepts both file paths and inline JSON. The CLI auto-detects: if the value starts with {, it is parsed as JSON directly. This is the fastest path — no file creation needed for simple inputs.

# Inline JSON
mthds-agent pipelex run bundle <bundle-dir>/ --inputs '{"theme": {"concept": "native.Text", "content": {"text": "nature"}}}'

# File path (auto-detected in directory mode)
mthds-agent pipelex run bundle <bundle-dir>/

Step 4: Present Results

After a successful run, always show the actual output to the user — never just summarize what fields exist.

Output format modes

The CLI has two output modes:

• Compact (default): stdout is the concept's structured JSON directly — no envelope, no success wrapper. This is the primary output of the method's main concept. Parse the JSON directly for field access.
• With memory (--with-memory): stdout has main_stuff (with json, markdown, html renderings) + working_memory (all named stuffs and aliases). Use this when piping output to another method.

The output_file and graph_files are written to disk as side effects (paths appear in logs/stderr), not in compact stdout.

4a. Determine what to show

In compact mode (default), the output is the concept JSON directly. Show the fields to the user:

{
  "clauses": [...],
  "overall_risk": "high"
}

In --with-memory mode, the output structure depends on the pipe architecture:

if main_stuff is non-empty (not {} or null):
    → main_stuff is the primary output (single unified result)
else:
    → working_memory.root holds the primary output (multiple named results)

4b. Show the output content

In compact mode: show the JSON fields directly. For structured concepts, format for readability.

In --with-memory mode when main_stuff is present (most pipe types):

• Show main_stuff.markdown directly — this is the human-readable rendering. Display it as-is so the user sees the full output.
• For structured concepts with fields, also show main_stuff.json formatted for readability.

In --with-memory mode when main_stuff is empty (PipeParallel without combined_output):

• Iterate working_memory.root and present each named result.
• For each entry, show the content field with its key as a label.
• Example: "french_translation: Bonjour le monde" / "spanish_translation: Hola mundo"

For dry runs: Show the same output but clearly label it as mock/simulated data.

4c. Output file

• The CLI automatically saves the full JSON output next to the bundle (live_run.json or dry_run.json).
• The output file path appears in runtime logs (stderr), not in compact stdout.

4d. Present graph files

• Graph visualizations are generated by default (live_run.html / dry_run.html). Use --no-graph to disable.
• The graph file path appears in runtime logs (stderr), not in compact stdout.

4e. Mention intermediate results

• If the pipeline has multiple steps, briefly note key intermediate values from working_memory (e.g., "The match analysis intermediate step scored 82/100").
• Offer: "I can show the full working memory if you want to inspect any intermediate step."

4f. Suggest next steps

• Re-run with different inputs
• Adjust prompts or pipe configurations if output quality needs improvement

Step 5: Handle Errors

When encountering runtime errors, re-run with --log-level debug for additional context:

mthds-agent --log-level debug pipelex run bundle <bundle-dir>/ --inputs data.json

For all error types and recovery strategies, see Error Handling Reference.

Execution Graphs

Execution graph visualizations are generated by default alongside the run output. Use --no-graph to disable.

mthds-agent pipelex run bundle <bundle-dir>/

Graph files (live_run.html / dry_run.html) are written to disk next to the bundle. Their paths appear in runtime logs on stderr, not in compact stdout. When using --with-memory, graph_files is included in the returned JSON envelope.

Piping Methods

The run command accepts piped JSON on stdin when --inputs is not provided. This enables chaining methods:

mthds-agent pipelex run method extract-terms --inputs data.json --with-memory \
  | mthds-agent pipelex run method assess-risk --with-memory \
  | mthds-agent pipelex run method generate-report

When methods are installed as CLI shims, the same chain is:

extract-terms --inputs data.json --with-memory \
  | assess-risk --with-memory \
  | generate-report

• Use --with-memory on intermediate steps to pass the full working memory envelope.
• The final step omits --with-memory to produce compact output.
• --inputs always overrides stdin when both are present.
• Upstream stuff names are matched against downstream input names. Method authors should name their outputs to match the downstream's expected input names.

Reference

• Error Handling — read when CLI returns an error to determine recovery
• MTHDS Agent Guide — read for CLI command syntax or output format details
• MTHDS Language Reference — read for.mthds syntax documentation
• Native Content Types — read when interpreting pipeline outputs or preparing input JSON, to understand the attributes of each content type