stemmata 0.0.3

Hierarchical prompt resolver CLI

stemmata

Prompt reuse across repositories is a mess. You copy a YAML prompt into a new project, tweak it, and within a week the original and the copy have diverged. Multiply that by a dozen services and you're maintaining the same boilerplate in twenty places.

stemmata fixes this with hierarchical composition: prompts declare ancestors (by relative path or by registry coordinate), and the CLI resolves the full inheritance chain into a single, deterministic YAML document. Ancestor prompts are distributed as npm packages through any private registry you already run.

Table of Contents

• Features
• Installation
• Quick Start
• CLI Reference
• Prompt Format
• Merge Semantics
• Exit Codes
• Configuration
• Testing

Features

• Hierarchical composition: prompts declare ancestors as paths or (package, version, prompt) coordinates; the full transitive closure is resolved eagerly via breadth-first search.
• Deterministic merging: nearest-wins for scalars and lists, deep-merge for maps, with breadth-first search distance plus reference occurring-ordering (for breaking ties) so the output is reproducible.
• Placeholder interpolation: ${path} references resolve against the merged namespace, with structural, textual, and list-splat forms.
• Abstract placeholders: a mid-graph prompt may declare required "holes" via ${abstract:<dotted-path>}, and any descendant is free to fill them — like the template-method pattern in OOP. resolve hard-fails on unfilled holes (exit 16); publish, describe, tree, and validate report them as contract information and keep working.
• Markdown resource embedding: packages may ship Markdown payloads alongside prompts and splice them in as opaque text via ${resource:...}, resolved eagerly on the same cache and registry rails as prompts.
• npm registry transport: speaks the standard npm REST API; credentials read from ~/.npmrc.

Installation

pip install stemmata

Requires Python 3.12+ (for tarfile.data_filter). Third-party dependencies: PyYAML and jsonschema.

Quick Start

# You have a local prompt that inherits from a base — resolve it:
stemmata resolve ./prompts/onboarding.yaml

# Or resolve a prompt published to your registry by coordinate:
stemmata resolve '@acme/prompts-core@1.2.3#onboarding'

# Describe every prompt in a published package (or a single one):
stemmata describe '@acme/prompts-core@1.2.3'
stemmata describe '@acme/prompts-core@1.2.3#onboarding'

# Print the ancestor DAG as an ASCII tree:
stemmata tree ./prompts/onboarding.yaml
stemmata tree '@acme/prompts-core@1.2.3#onboarding'

# Need machine-readable output for a script or pipeline:
stemmata --output json resolve ./prompts/onboarding.yaml

# Validate a prompt (or an entire directory) against its $schema:
stemmata validate ./prompts/onboarding.yaml
stemmata validate ./prompts/

# Wipe the local cache (by default stored under ~/.cache/stemmata):
stemmata cache clear

# Scaffold a package.json for an existing directory of prompts and resources:
stemmata init ./my-package

# Install a local package into the cache so it can be resolved offline by coordinate:
stemmata install ./my-package

CLI Reference

stemmata [GLOBAL FLAGS] <subcommand> [ARGS]

Global flags

Flag	Default	Description
--output {yaml,json,text}	yaml (text for tree)	Output format.
--verbose	off	Timestamped diagnostics on stderr.
--offline	off	Forbid network access; exit 22 if a fetch would be needed.
--refresh	off	Re-fetch artifacts even if cached.
--cache-dir <path>	~/.cache/stemmata	Override the cache root. Honours $PROMPT_CLI_CACHE_DIR when the flag is absent.
--npmrc <path>	~/.npmrc	Override the npmrc file used for registry routing and credentials.
--version	—	Print version and exit.

resolve <target>

Resolves a single prompt. Target is either a local path (./prompts/onboarding.yaml) or a registry coordinate (@<scope>/<name>@<version>#<prompt-id>).

Resource limits: --max-prompts (default 1000), --max-depth (default 50), --max-download-size (default 64 MiB per package), --max-total-size (default 512 MiB per invocation), --http-timeout (default 30s), --timeout (default 5m).

--set <dotted-path>=<yaml-value> overrides the merged value at that path. Repeatable; last-wins on duplicate paths. Values are parsed as YAML, so --set port=5432 is an int, --set enabled=true is a bool, --set tags=[a,b,c] is a list, and --set body= is null. Overrides merge at BFS distance -1 — nearer than the root prompt — so they beat every ancestor and can satisfy ${abstract:…} markers. They show up in the JSON envelope's ancestors[] as {canonical_id: "<overrides>", distance: -1}. --set is only accepted by resolve.

On success, stdout carries the resolved YAML (or a JSON envelope with {root, content, ancestors[]}). On failure, stdout carries a JSON error envelope regardless of --output, and stderr gets a one-line human-readable summary.

publish [path]

Builds and uploads the package at path (default .) to the registry routed by ~/.npmrc. Before any bytes leave the machine, every prompt listed in package.json is checked for: (1) ancestor cycles, (2) intra-document type conflicts, (3) placeholder resolvability against the fully resolved namespace, (4) dependencies consistency with the cross-package references found in the prompts (including those inside ${resource:...} placeholders), (5) manifest closure under relative-path references — every local ancestors entry must resolve to a path that is itself declared in prompts, since only manifest-listed files are bundled, (6) $schema validation against the prompt's content contract, and (7) resource-graph integrity — every ${resource:...} occupies an allowed position, every local resource reference resolves to an entry in the resources array, and the Markdown-embedding graph contains no cycles. All errors discovered in the pass are aggregated into a single envelope; the headline exit code is the most severe one (cycle > schema > reference > merge > placeholder).

Abstract placeholders (${abstract:<dotted-path>}) are not treated as failures by publish: a library package whose prompts contain unfilled abstracts is the whole point. For every prompt that still has unfilled abstracts, publish logs a warning: line to stderr listing the holes, records them under abstracts in the success payload, and defers $schema content validation for that prompt (per the per-prompt all-or-nothing rule). Real placeholder / reference / merge errors found alongside abstracts still fail publish as before.

Flags: --dry-run (build the tarball but skip upload), --tarball <path> (write the built tarball to path). The tarball is deterministic: identical inputs produce byte-identical output.

$schema enforcement is always on. An unfetchable $schema URI (missing local file, offline with no cache, or network failure) aborts with exit code 10.

validate <target>

Validates prompt files against their $schema. Target is a file path or a directory (recursively discovers .yaml, .yml, .json files). For YAML prompts with ancestors, the full resolve → merge → interpolate pipeline runs before validation so inherited and interpolated values participate.

Multi-document YAML files (separated by ---) are supported — each sub-document is validated independently against its own $schema. Files without $schema are silently skipped.

All violations are collected and reported together. Error payloads include the natural source line number of the offending value.

Flags: the same resource-limit flags as resolve. An unfetchable $schema URI (missing file, offline with no cache, or network failure) aborts with exit code 10.

$schema enforcement supports file://, http://, and https:// URIs, as well as bare relative paths (resolved against the validated file's directory).

describe <target>

Resolves every prompt in a published package, or a single prompt inside it. Target is either @<scope>/<name>@<version> (describe the whole package) or @<scope>/<name>@<version>#<prompt-id> (describe one prompt). Each prompt is resolved with ancestors merged and placeholders interpolated, using the same pipeline as resolve.

Default YAML output emits one document per prompt, separated by --- start markers. Each sub-document is prefixed with a # <canonical-id> comment (e.g. # @acme/prompts-core@1.2.3#onboarding) so the reader can tell which prompt is which. --output json returns an array of {root, content, ancestors[]} entries in manifest declaration order (length 1 when targeting a single prompt). Package artifacts are fetched through the usual cache (~/.cache/stemmata by default), so repeated invocations reuse downloaded tarballs; --offline and --refresh behave as with resolve.

Resource-limit flags match resolve.

tree <target>

Prints the ancestor DAG rooted at <target>, which takes the same two forms as resolve (a local YAML/JSON path or a @<scope>/<name>@<version>#<prompt-id> coordinate). The resolver runs the same eager pipeline as resolve, so cycles, missing ancestors, and version conflicts surface with the usual exit codes; --offline / --refresh and the resource-limit flags all apply.

Default --output text produces an ASCII tree (|-- / `-- connectors). Markdown resources reached via ${resource:...} are rendered inline under the prompt (or resource) that references them and are prefixed with resource: to disambiguate them from prompt coordinates. Diamond inheritance — across both ancestor and resource edges — is rendered once in full and subsequent visits are marked (seen) so the output stays finite:

root.yaml
|-- a.yaml
|   `-- x.yaml
|-- b.yaml
|   `-- x.yaml  (seen)
`-- resource:@acme/prompts-core@1.2.3#playbook
    `-- resource:@acme/prompts-core@1.2.3#safety

--output yaml / --output json emit a {root, nodes[], edges[]} envelope instead, with each node carrying its canonical id, source file, BFS distance from the root, and kind (prompt or resource). Edges carry kind (ancestor or resource).

init [path]

Scaffolds (or updates) a package.json at path (default .). Scans ./prompts recursively for .yaml, .yml, and .json files and ./resources recursively for .md files, deriving each entry's id from the basename and setting contentType from the extension. Entries are sorted alphabetically by path and rendered one-per-line.

install [path]

Installs the package at path (default .) into the local cache.

cache clear

Evicts every cached entry.

Prompt Format

A prompt is a structured mapping (YAML or JSON) with reserved envelope keys plus arbitrary content:

ancestors:
  - ../base.yaml                         # relative path (within package)
  - package: "@acme/common"              # cross-package coordinate
    version: "1.0.4"
    prompt: "defaults"
$schema: "https://schemas.example/foo.v1.json"   # optional, enforced at publish time if present

database:
  host: "db.internal"
  port: 5432
body: |
  Region is ${vars.region}; DB is ${database.host}:${database.port}.

ancestors and $schema are stripped from the namespace; every other key is addressable via dotted path.

Package manifest (package.json)

{
  "name": "@acme/prompts-core",
  "version": "1.2.3",
  "license": "UNLICENSED",
  "dependencies": { "@acme/common": "1.0.4" },
  "prompts": [
    { "id": "base",       "path": "prompts/base.yaml",             "contentType": "yaml" },
    { "id": "onboarding", "path": "prompts/extra/onboarding.yaml", "contentType": "yaml" }
  ],
  "resources": [
    { "id": "overview", "path": "resources/overview.md", "contentType": "markdown" }
  ]
}

name must be @<scope>/<n>. version is strict SemVer, no ranges. prompts is non-empty; id defaults to basename without extension and must match [a-z0-9][a-z0-9_-]*. resources is optional; ids, paths, and case-folded path uniqueness are shared across the union of prompts and resources.

Markdown resources

Prompts may embed opaque Markdown payloads via ${resource:<POSIX-relative-path>} (same-package) or ${resource:@<scope>/<name>@<version>#<id>} (coordinate). The reference must stand alone — either as the sole content of a line inside a block scalar or a Markdown file, or as the entire trimmed text of a flow-style YAML scalar or JSON string. Resource payloads are substituted verbatim after ancestor-namespace interpolation; they do not contribute keys to the merged namespace and any ${...} sequences inside them are left literal.

Abstract placeholders

An author can mark a dotted-path as a required "hole" that any descendant must fill before the graph becomes resolvable:

# @acme/prompts-core#persona (a reusable mid-graph prompt)
persona:
  tone: "${abstract:persona.tone}"         # required: any descendant must set persona.tone
  greeting: "Hi — my tone is ${abstract:persona.tone}."

A descendant fills the hole by providing a concrete value at the same dotted path:

# a concrete child
ancestors:
  - package: "@acme/prompts-core"
    version: "1.0.0"
    prompt: "persona"

persona:
  tone: "friendly"

Semantics:

• Syntax. ${abstract:<dotted-path>} — the prefix mirrors ${resource:…} so dispatch is unambiguous and the form is JSON-safe.
• Usable positions. An abstract marker stands in for a string value. It may appear as the sole content of any scalar (flow or block) or positionally inside a larger string scalar ("prefix ${abstract:x} suffix"). Map/list-shaped abstracts are not supported in v1.
• A hole is unfilled iff, after BFS merge, the nearest value at the referenced dotted path is (a) absent, (b) explicit null (null shadowing does not satisfy an abstract), or (c) itself another ${abstract:…} marker (an abstract does not satisfy another abstract). The per-case reason is reported in the error envelope as not_provided, null_shadow, or abstract_inherited.
• Per-prompt all-or-nothing validation. For any given prompt, if its merged namespace has zero unfilled abstracts, placeholder interpolation and $schema content-contract validation run as normal. If one or more abstracts remain unfilled, both checks are deferred for that prompt — nothing about that prompt's content contract can be enforced until the contract is fulfilled.

Subcommand behaviour:

Command	Unfilled abstracts present
resolve	Hard-fails with exit 16. The resolved artefact is not produced while any hole remains.
validate	Does not fail. Structural checks and cycle detection still run. Abstracts are reported under abstracts in the success payload.
publish	Does not fail. A warning: line is logged to stderr listing the unfilled abstracts, and each one is recorded under abstracts in the success payload. Schema validation is deferred for any prompt that still has holes.
describe	Always works. Emits two labelled buckets per prompt: abstracts.declared (markers introduced by this prompt) and abstracts.inherited (declared in an ancestor and still unfilled here). When abstracts remain, content is the merged (pre-interpolation) namespace so the reader can see where the holes sit.
tree	Always works. Each prompt node is annotated with [abstracts: a, b, c] listing what markers it introduces; the JSON/YAML envelope adds abstracts to each node.

Merge Semantics

Reachable prompts are layered by breadth-first search distance from the root (distance 0 = root, wins everything). Ties at the same distance break by enqueue order.

Maps are deep-merged, with the nearer value winning at each leaf:

# ancestor (distance 1)           # root (distance 0)
database:                         database:
  host: "base.internal"             host: "override.internal"
  port: 5432                        ssl: true

Resolved: database.host = "override.internal" (nearer wins), database.port = 5432 (survives from ancestor), database.ssl = true (only root provides it).

Lists replace wholesale — no element-level merge. null at a nearer layer shadows the entire subtree beneath it.

For the full interpolation reference (structural vs. textual placeholders, list splat, non-splat ${=...} form, escaping, version conflict resolution), see docs/interpolation.md.

Exit Codes

Code	Meaning
0	Success
1	Generic / unexpected failure
2	Usage error
10	Schema validation error
11	Unknown ancestor, prompt, or resource reference
12	Cycle detected (ancestor or resource graph)
14	Unresolvable placeholder
15	Merge / interpolation type mismatch
16	Abstract placeholder unfilled (from resolve)
20	Network / registry error
21	Cache error
22	Offline-mode violation

On failure, stdout always carries a JSON error envelope with {status, exit_code, command, error: {code, category, message, ...}} regardless of --output. Stderr gets a single-line human summary.

Configuration

Registry routing and credentials come from ~/.npmrc for both fetch and publish.

Testing

PYTHONPATH=src python -m pytest tests/ -q