Rhylthyme CLI Runner

Command-line interface for validating and running Rhylthyme real-time program schedules, and for turning a plain-language request into one through the hosted Rhylthyme MCP server.

pip install rhylthyme
rhylthyme login
rhylthyme generate "roast chicken, potatoes and green beans for 6" -e kitchen --by 19:00 --with "one oven"

A timeline in five commands: a birthday party

Parents collect at 4:00. Working backwards from that: when does the cake go in the oven, when do the balloons go up, and do the pizzas and the cake ever want the oven at the same time?

pip install rhylthyme

# 1. a program: the party, the kitchen and the setup crew, 13 steps
curl -sO https://raw.githubusercontent.com/rhylthyme/rhylthyme-cli-runner/main/examples/recipe/birthday-party.json

# 2. is it well formed?
rhylthyme validate birthday-party.json --strict

# 3. any clashes, and what time does everything start if we finish at 4?
rhylthyme analyze birthday-party.json --finish-at 16:00

**Makespan:** 3h 10m across 3 tracks, 13 steps.
**Critical path:** Guests arrive → Party games → Pizza → Candles and cake → Piñata, until it breaks → Party bags and goodbyes
**Binding constraints:** step dependencies only — no resource or in-flight limit gates the critical path.
**Peak concurrency:** 2 steps at 10:00.
**Resource conflicts:** none.

**Tracks that finish early:** Kitchen (1h 10m early), Setup crew (50m early)

Start Sat 12:50, finish Sat 16:00 (local time)
  Sat 12:50  Bake the cake
  Sat 13:00  Balloons and banner
  Sat 13:25  Cake cools
  Sat 13:30  Hang the piñata
  Sat 13:50  Guests arrive
  Sat 14:05  Frost and decorate
  Sat 14:20  Party games
  Sat 14:35  Pizzas in the oven
  Sat 14:50  Fill party bags
  Sat 14:50  Pizza
  Sat 15:15  Candles and cake
  Sat 15:30  Piñata, until it breaks
  Sat 15:45  Party bags and goodbyes

# 4. a live timeline for the fridge door (or a phone): timers, cues, and you end the piñata when it breaks
rhylthyme publish birthday-party.json --open --image party-preview.png

# 5. a picture for the group chat, on the wall clock (needs Node.js)
rhylthyme render birthday-party.json -o birthday-party.png \
  --style web --palette vivid --label-overflow outside --start-at "$(date +%F)T12:50:00"

The pizzas go in 15 minutes after the games start, so they come out as the games end; the party bags get filled while everyone is eating. Try --finish-at 17:30, or turn the cake into a two-hour fruitcake ("seconds": "2h" on bake) and analyze reports the clash: oven needs 2 but max is 1 from 1:45:00 to 2:00:00 (bake, heat-pizza). rhylthyme run birthday-party.json --time-scale 60 rehearses the whole afternoon in the terminal at a minute per second.

This repository and rhylthyme-mcp

Two repositories do related things and are easy to confuse. This one is the command a person types. rhylthyme-mcp is the server an AI assistant talks to.

	rhylthyme-mcp	rhylthyme-cli-runner
What it is	The MCP server: the tools an AI assistant calls	A command-line program: the rhylthyme command
Who uses it	Claude, ChatGPT, Cursor or any MCP client, on a person's behalf	A person at a terminal, a script, or CI
Where it runs	Hosted at mcp.rhylthyme.com; nothing to install	On your machine: pip install rhylthyme
Language	JavaScript (Node 20+)	Python 3.12+
Input	A program the assistant builds in conversation	A program file on disk (JSON or YAML)
Validate a program	validate_program	rhylthyme validate (works offline)
Import a recipe or protocol	import_from_source	rhylthyme import
Timing, conflicts, deadlines	analyze_schedule	rhylthyme analyze (asks the server)
Publish a live timeline	visualize_schedule	rhylthyme publish (asks the server)
Run a schedule with timers	no: it hands back a link to the web timeline	rhylthyme run, an interactive terminal runner
Recorded runs, calibration	reads runs saved to an account	records runs locally; rhylthyme runs, rhylthyme calibrate
Catalog search, account library	yes	no
Also in the repository	the rhylthyme-mcp PyPI package (a stdio bridge to the hosted server), the Claude plugin marketplace	the Claude skill's source, the prompt-evaluation harness and its results, rhylthyme mcp-test

How they fit together: this tool is one of the MCP server's clients. rhylthyme analyze, publish, generate and mcp-test are MCP calls to mcp.rhylthyme.com; rhylthyme validate, run, runs and calibrate never touch the network. Each has its own validator for the same program schema (Python here, JavaScript there), so a program is checked again when it is published.

Use this repository if you have a program file and a terminal, want timers in the terminal, or keep run records. Use rhylthyme-mcp to connect Claude, ChatGPT or another assistant, or to read or self-host the server.

Two names to keep apart: rhylthyme-cli-runner on PyPI is this tool (command rhylthyme); rhylthyme-mcp on PyPI is the server's stdio bridge (command rhylthyme-mcp). The program format itself is defined in rhylthyme-spec, with examples in rhylthyme-examples.

Installation

From Source

# Clone the repository
git clone https://github.com/rhylthyme/rhylthyme-cli-runner.git
cd rhylthyme-cli-runner

# Install the package
pip install -e .

# Or install with development dependencies
pip install -e ".[dev]"

From PyPI

pip install rhylthyme            # this package, the importers and the renderer
pip install rhylthyme-cli-runner   # this package alone

Getting the Examples

The example programs referenced in this documentation are maintained in a separate repository. To use them:

Option 1: Clone the Examples Repository Alongside

# Navigate to the parent directory of rhylthyme-cli-runner
cd ..

# Clone the examples repository
git clone https://github.com/rhylthyme/rhylthyme-examples.git

# Run examples using relative paths
cd rhylthyme-cli-runner
rhylthyme run ../rhylthyme-examples/programs/breakfast_schedule.json

Option 2: Clone Examples as a Subdirectory

# From within the rhylthyme-cli-runner directory
git clone https://github.com/rhylthyme/rhylthyme-examples.git examples

# Run examples directly
rhylthyme run examples/programs/breakfast_schedule.json

Option 3: Download Individual Examples

# Download a specific example file
curl -O https://raw.githubusercontent.com/rhylthyme/rhylthyme-examples/main/programs/breakfast_schedule.json

# Run the downloaded file
rhylthyme run breakfast_schedule.json

Quick Start

The Rhylthyme CLI provides commands for working with real-time program schedules defined using the Rhylthyme JSON or YAML schema.

Generate a Program from Plain Language

generate sends your request to the hosted MCP server at mcp.rhylthyme.com, which builds a validated multi-track program and publishes it as a live timeline. Sign in once first; the session is kept in ~/.config/rhylthyme/credentials.json and renews itself.

rhylthyme login     # opens rhylthyme.com in your browser

# Prints the Gantt chart, the itinerary and the live-timeline URL
rhylthyme generate "roast chicken, potatoes and green beans for 6" \
    -e kitchen --by 19:00 --with "one oven, four burners, one cook"

# From a file, save the program, then run it here in the terminal
rhylthyme generate -e lab -f western_blot.txt -o blot.json --run

# From stdin, print only the URL
pbpaste | rhylthyme generate -e events --by "doors at 18:30" -q

The server runs several model turns per request (20–60 s, capped per day). Headless machines can set RHYLTHYME_TOKEN to an access token from https://www.rhylthyme.com/mcp/auth instead of running login.

Validate a Program

Validate a program file against the schema to ensure it's properly formatted:

# Validate a JSON or YAML program file
rhylthyme validate my_program.json

# Validate with verbose output
rhylthyme validate my_program.json --verbose

# Validate with JSON output for CI/scripting
rhylthyme validate my_program.json --json

# Validate in strict mode (requires all tasks in resourceConstraints)
rhylthyme validate my_program.json --strict

Run a Program

Run a program with an interactive terminal UI for monitoring and controlling execution:

# Run a program file
rhylthyme run examples/programs/breakfast_schedule.json

# Run with automatic start (no manual trigger needed)
rhylthyme run examples/programs/breakfast_schedule.json --auto-start

# Run with a different time scale (2x faster)
rhylthyme run examples/programs/breakfast_schedule.json --time-scale 2.0

# Run with a specific environment
rhylthyme run examples/programs/breakfast_schedule.json --environment kitchen

# Run without validation (if you're sure the file is valid)
rhylthyme run examples/programs/breakfast_schedule.json --no-validate

Analyze and Publish a Program

Neither needs an account; both are computed by the hosted MCP server.

# Total length, critical path, resource conflicts
rhylthyme analyze examples/programs/breakfast_schedule.json

# When does each step start if breakfast is at 8:30?
rhylthyme analyze examples/programs/breakfast_schedule.json --finish-at 8:30am

# A live, shareable timeline with timers; prints the URL
rhylthyme publish examples/programs/breakfast_schedule.json

validate checks structure only. Two steps that want the only oven at the same time pass validation; analyze reports them, and --strict makes that a non-zero exit.

Optimize a Program

plan is an older stagger heuristic that reads the pre-0.2 program format; on programs written with stepId and task (all current examples) it writes the program back unchanged. Use analyze to find contention.

# Optimize a program and save to new file
rhylthyme plan examples/programs/breakfast_schedule.json optimized_breakfast.json

# Optimize with verbose output
rhylthyme plan examples/programs/breakfast_schedule.json optimized_breakfast.json --verbose

Work with Environments

List and validate environment catalogs:

# List all available environments
rhylthyme environments

# List environments in JSON format
rhylthyme environments --format json

# Validate all environment files
rhylthyme validate-environments

# Show information about a specific environment type
rhylthyme environment-info kitchen

Claude Skill

skills/rhylthyme is an Agent Skill that teaches Claude (Claude Code, the Claude apps, the Agent SDK) to turn a protocol, recipe or run sheet into a validated program with this CLI, check it for conflicts, and hand back a live timeline. It follows the layout of K-Dense scientific skills: a SKILL.md plus references/.

The easiest install is the Claude Code plugin, which also connects the hosted MCP server:

/plugin marketplace add rhylthyme/rhylthyme-mcp
/plugin install rhylthyme@rhylthyme

Or copy the skill by hand:

# Claude Code, for one user
mkdir -p ~/.claude/skills
cp -r skills/rhylthyme ~/.claude/skills/

# or for one project
mkdir -p .claude/skills && cp -r skills/rhylthyme .claude/skills/

Then ask, for example, "time a Western blot so imaging is at 4 pm" or "two PCR protocols, one thermocycler: when do I start each?". tests/test_skill.py validates every program in the skill and checks that every command it names exists.

Program File Examples

Simple Breakfast Schedule

{
  "programId": "breakfast-schedule",
  "name": "Breakfast Schedule",
  "description": "Coordinated breakfast preparation",
  "environmentType": "kitchen",
  "startTrigger": {
    "type": "manual"
  },
  "tracks": [
    {
      "trackId": "eggs",
      "name": "Scrambled Eggs",
      "steps": [
        {
          "stepId": "crack-eggs",
          "name": "Crack and Whisk Eggs",
          "startTrigger": {
            "type": "programStart"
          },
          "duration": {
            "type": "fixed",
            "seconds": 60
          }
        }
      ]
    }
  ],
  "resourceConstraints": [
    {
      "task": "stove-burner",
      "maxConcurrent": 2,
      "description": "Maximum stove burners"
    }
  ],
  "version": "1.0.0"
}

Interactive UI Controls

When running a program, the interactive UI provides these controls:

• Space: Start/stop the program
• Enter: Trigger manual steps
• Arrow keys: Navigate between steps
• q: Quit the program
• r: Refresh the display
• s: Sort by different criteria

Command Reference

rhylthyme generate [REQUEST...]

Turns a natural-language request into a program and a live timeline using the hosted MCP server's import_text and visualize_schedule tools. Requires rhylthyme login (or RHYLTHYME_TOKEN).

Options:

• -e, --env: generic (default), kitchen, lab, events or gym
• --by TEXT: when everything must be finished, e.g. 19:00
• --with TEXT: equipment and people limits, e.g. one oven, two cooks
• -f, --file PATH: read the request or source text from a file (- for stdin)
• -o, --output PATH: save the program JSON
• --run: run the program in the terminal UI afterwards
• --open: open the live timeline in a browser
• --no-publish: build the program only
• --json: print {url, program, ...} as JSON
• -q, --quiet: print only the URL

Environment: RHYLTHYME_TOKEN (access token, overrides the stored session), RHYLTHYME_MCP_URL (default https://mcp.rhylthyme.com/mcp), RHYLTHYME_SITE_URL (default https://www.rhylthyme.com).

rhylthyme mcp-test

Smoke-tests a Rhylthyme MCP server and exits 1 if any check fails. By default it runs the read-only checks against all five hosted endpoints (/mcp, /kitchen/mcp, /lab/mcp, /events/mcp, /gym/mcp):

Check	What it proves
initialize	server name, protocol version, tools/resources/prompts capabilities, instructions
tools	core tools and the endpoint's one-shot tools are listed, each with a description and schema
validate-good / validate-bad	a valid program passes (including a type: "compound" trigger); a dangling step reference is rejected with a fix hint
analyze	makespan, critical path and wall-clock itinerary for a known program
resources / prompts	schema and guides are readable, a bundled example validates, plan_schedule substitutes its arguments
json-accept	clients that do not accept SSE (*/*, application/json) get JSON, not 406
bad-requests	unknown method gives -32601, unknown tool gives an error, never a 5xx
login-gate	account tools refuse without a token and point at login
catalog	public search returns entries that load with a URL (empty catalog = warning)
publish (--publish)	visualize_schedule returns a URL on the right site; the page and PNG load
generate (--generate)	the model-backed import_text returns a program that validates (needs login)

rhylthyme mcp-test                         # everything read-only
rhylthyme mcp-test -e lab --publish        # one endpoint, plus a real share
rhylthyme mcp-test -k catalog -k tools     # only some checks
rhylthyme mcp-test --url http://localhost:3000/mcp -e generic
rhylthyme mcp-test --json --strict         # cron / CI: warnings fail too

Requests carry mcp-test in the User-Agent so the hosted server logs the errors these checks provoke without alerting anyone. The same suite is importable (rhylthyme_cli_runner.remote.checks.run_suite), and RHYLTHYME_MCP_LIVE=1 pytest tests/test_mcp_checks.py runs it live.

rhylthyme login / logout / whoami

login opens the rhylthyme.com sign-in page and receives the session on a one-shot listener bound to 127.0.0.1; --token TOKEN stores a pasted access token instead and --no-browser only prints the URL. whoami shows the stored account and expiry; logout deletes the stored session.

rhylthyme validate

Validates program files against the schema.

Options:

• --schema PATH: Path to schema file (default: built-in schema)
• --verbose, -v: Show detailed validation information

rhylthyme run

Runs programs with interactive terminal UI.

Options:

• --schema PATH: Path to schema file (default: built-in schema)
• -e, --environment TEXT: Environment ID to use
• --time-scale FLOAT: Time scale factor (default: 1.0)
• --validate / --no-validate: Validate before running (default: True)
• --auto-start: Automatically start without manual trigger

rhylthyme analyze

Total length, critical path, what gates each link of it, resource conflicts and tracks that finish early. No sign-in; nothing is published.

Options:

• --finish-at TEXT: when everything must be finished (19:00, 7:30pm or ISO 8601); prints local clock times per step
• --start-at TEXT: when the program starts; ignored with --finish-at
• --strict: exit non-zero when there are resource conflicts
• --json: the full analysis

rhylthyme publish

Publishes a program (a file, an http(s) URL, or - for standard input) as a live timeline and prints its URL. No sign-in. A published timeline is reachable by anyone who has the link.

Options:

• -e, --env: generic, kitchen, lab, events, gym (default: from environmentType)
• -q, --quiet: print only the URL
• --json: url, shareId, imageUrl, makespanSeconds, warnings
• --image PATH: also save a PNG of the timeline
• --open: open it in a browser

rhylthyme import

Imports a recipe, protocol or slide deck as a program, validates it, and writes <programId>.json. Needs rhylthyme-importers (included in pip install rhylthyme).

rhylthyme import https://www.seriouseats.com/the-best-chili-recipe   # any of ~580 recipe sites
rhylthyme import 52772 -i themealdb --publish                        # a TheMealDB id, straight to a live timeline
rhylthyme import https://www.protocols.io/view/western-blot-...      # needs PROTOCOLS_IO_TOKEN
rhylthyme import https://raw.githubusercontent.com/Opentrons/Protocols/develop/protocols/007992/rna_isolation.ot2.apiv2.py   # an Opentrons protocol
rhylthyme import "Neapolitan Pizza.cook" -i cooklang
rhylthyme importers                                                  # what is installed
rhylthyme search "pad thai" -i spoonacular                           # then import a hit's URL

Options: -i IMPORTER (default: chosen from the URL), -o PATH, --stdout, --publish (and --open), --no-validate. - reads pasted source from stdin (with -i).

rhylthyme render

Draws a program as an SVG, PNG or PDF figure through the rhylthyme-timeline package (included in pip install rhylthyme; needs Node.js 18+). Every option of rhylthyme-render is accepted: rhylthyme render --help.

rhylthyme render dinner.json -o dinner.png --style web --palette vivid --color-by task
rhylthyme render blot.json -o figure.pdf --style publication --legend right

rhylthyme plan

An older stagger heuristic. It reads the pre-0.2 program format and leaves programs written with stepId and task unchanged; use analyze to find contention and edit the triggers.

Options:

• --verbose, -v: Show detailed planning information

rhylthyme environments

Lists available environment catalogs.

Options:

• --format, -f: Output format (table, json, yaml)

rhylthyme validate-environments

Validates environment catalog files.

Options:

• --environments-dir PATH: Directory containing environment files
• --verbose, -v: Show detailed validation information

rhylthyme environment-info

Shows information about a specific environment type.

rhylthyme eval-prompts

Scores agent-authored programs against the expert gold set. See "Evaluating prompts" below for the full flag list.

Evaluating prompts

eval-prompts measures how well a prompt gets a model to author a Rhylthyme program. It compares a predicted program with the expert gold program of the same slug in rhylthyme-examples/gold/ and reports per-component precision/recall: steps, durations, resources, actors, relationships (triggers), track Rand index, an end-to-end pass and the unsupported-step rate.

Two modes:

# Offline: score programs already on disk (no model calls, no cost).
rhylthyme eval-prompts --score-only \
  --gold ../rhylthyme-examples/gold --predicted /tmp/predicted

# Live: render a prompt pattern per gold source, call a model, extract
# the program JSON, validate it, score it.
rhylthyme eval-prompts \
  --gold ../rhylthyme-examples/gold \
  --model claude-haiku-4-5-20251001 --patterns baseline,four-turn

Patterns (--list-patterns):

pattern	what it sends
baseline	The single plan_schedule message the MCP server sent before Phase 3, frozen in eval/patterns/baseline.py so the committed numbers stay reproducible.
four-turn	The four messages plan_schedule sends today, byte-copied from rhylthyme-server/mcp-api/prompts.js: T1 read-back → T2 model check → T3 extraction (steps + source spans, no relationships) → T4 relationships (tracks, triggers, the refined dependency question, then the program).

Both patterns end in the same validate-and-repair loop and get the same system prompt, so the only variable between them is the shape of the user turns. four-turn also records the paper's 2/1/0 score for T1 and T2 in results.json under extras (t1_score, t2_score): 2 when the reply carries the expected JSON shape first time, 1 when a single retry turn recovers it, 0 when it never does — plus unsupported_step_ids, the steps the model itself marked inferred in T3.

Flags

Flag	Meaning
--gold PATH	Gold set directory (required).
--score-only	Offline mode; needs --predicted.
--predicted PATH	Directory of predicted programs (<slug>.json or <slug>/program.json).
--model ID	Model id for a live run (required unless --score-only). Pin it; do not let it drift.
--patterns NAMES	Comma-separated prompt patterns (default baseline). --list-patterns prints the registry.
--only SLUGS	Comma-separated gold slugs to run; applied before --limit.
--limit N	Run only the first N gold programs (slug order).
--concurrency N	Programs run in parallel (default 1).
--from-cache	Never call the model: replay cached responses, and fail loudly on a missing key.
--cache-dir PATH	Where raw responses are cached (default <out>/cache).
--max-fix-iterations N	How many times Python-validator findings are sent back as the next user turn (default 2; 0 disables the fix loop).
--max-tokens N	max_tokens per model call (default 16000).
--write-baseline PATH	Also write {model, pattern, date, git_note, results} there.
--out PATH	Where results.json, results.md, responses/ and predicted/ are written (default ./eval-results).
--format table|json	What to print to stdout.
--skip-js	Skip the JavaScript validator even when node is available.
--threshold FLOAT	Minimum source-span overlap for two steps to match (default 0.5).
--verbose, -v	Per-program breakdown under the table.

Sub-command eval-prompts compare gates a run against the committed numbers instead of printing them:

Flag	Meaning
--baseline FILE	The stored reference. Use eval/four-turn.json — that is what ships.
--results FILE	results.json from the run under test.
--rel-f1-drop N	Largest tolerated fall in relationships F1, in points (default 5).
--e2e-drop N	Largest tolerated fall in the end-to-end pass rate, in points (default 0: any drop fails).
--allow-missing	Do not fail when the results file is missing a program the reference has.
--format table|json	Diff table, or the same comparison as JSON.

Cost. A live run spends real money: one call per turn per gold program, plus one per retry and one per fix iteration. The published baseline run cost $1.10 for 24 programs and 49 calls on claude-haiku-4-5-20251001 (358k input / 148k output tokens); the four-turn run over the same 24 cost $2.20 for 111 calls (969k / 246k) — four turns instead of one, each carrying the whole conversation so far. The run prints its own token totals and an estimated cost from the price table in src/rhylthyme_cli_runner/eval/llm.py (prices as of 2026-06-24); an unpriced model reports unknown instead of a number. --limit/--only keep a trial run cheap.

Cache. Every call is cached under --cache-dir (default <out>/cache) keyed by sha256 of the model, the pattern name, the system prompt and the full message list, so each turn of a multi-turn pattern is cached separately and re-scoring never re-spends. --from-cache replays and refuses to call the API; a missing key is an error naming the slug and the key. The committed cache for both published runs is eval/cache/ (1.44 MB, 160 files). Only the completion is replayed, so the committed files keep the completion in full and record the prompt side as digests (system_sha256, messages_sha256) plus a role/length shape — the cache is in the repository, so it is kept small on purpose.

Reading the table. One row per program, then a mean row:

program                               steps P/R/F1    dur acc  res P/R    actors  rel P/R/F1      tracks RI  e2e   unsupp
kitchen-weeknight-stir-fry            1.00/0.87/0.93  1.00     0.00/0.00  1.00    0.62/0.47/0.53  0.97       0.00  0.19

• steps — matched by source-span overlap ≥ --threshold, else by normalised name. Precision near 1.00 with low recall means the model wrote real steps but far fewer of them than the expert did (it lumps).
• dur acc — of matched steps, how many got the duration kind and value right.
• res P/R — resourceConstraints[].task names.
• actors — program actors count (1 or 0).
• rel P/R/F1 — triggers: same owning step, same anchor step, same type, offset within 10 %. This is the number the four-turn work in Phase 3 has to move; the paper this work follows found relationships are the weak component, and the baseline agrees.
• tracks RI — Rand index of the step-to-track partition.
• e2e — passes both validators, makespan within 10 % of gold, critical chain shares ≥ 50 % of gold's steps.
• unsupp — predicted steps with no gold match and no locatable span, as a share of predicted steps. Not counted as errors (PRD open question 4).

A program whose reply carried no parseable program, or one that still fails the validator after the fix loop, is scored zero on every component and listed under the table, so a failed run lowers the mean instead of shrinking the sample.

eval/baseline.json and eval/four-turn.json are the committed numbers, in the same shape: the model pin, the pattern, the date, a git note for the commit they were produced from, and the full per-program results. eval/four-turn.json is the one CI gates on — it is the pattern plan_schedule ships; eval/baseline.json is kept as the historical single-message number so this table can show what the four turns bought.

Published numbers

All 24 gold programs, --max-fix-iterations 2, --concurrency 3, both patterns on the same model on the same day.

baseline pattern, claude-haiku-4-5-20251001, 2026-09-14 (49 calls, $1.10):

domain	n	steps P/R/F1	dur acc	res P/R	actors	rel P/R/F1	tracks RI	e2e	unsupp
kitchen	10	0.90/0.43/0.56	0.88	0.50/0.48	0.90	0.38/0.20/0.25	0.78	0.00	0.47
laboratory	6	0.83/0.24/0.37	0.81	0.67/0.64	1.00	0.36/0.10/0.15	0.81	0.00	0.76
event	4	1.00/0.66/0.78	0.97	0.47/0.39	0.25	0.47/0.32/0.37	0.90	0.25	0.37
fitness	4	0.75/0.11/0.17	0.75	0.69/0.69	0.75	0.42/0.05/0.08	0.75	0.25	0.89
overall	24	0.88/0.37/0.48	0.85	0.57/0.54	0.79	0.39/0.17/0.22	0.80	0.08	0.60

four-turn pattern, same model, same 24 programs, 2026-09-14 (111 calls, $2.20; T1 and T2 scored 2 on every program, no retries):

domain	n	steps P/R/F1	dur acc	res P/R	actors	rel P/R/F1	tracks RI	e2e	unsupp
kitchen	10	0.94/0.98/0.96	0.94	0.58/0.59	1.00	0.68/0.67/0.67	0.94	0.60	0.01
laboratory	6	0.84/0.99/0.91	0.85	0.57/0.56	0.83	0.59/0.61/0.59	0.90	0.67	0.01
event	4	0.93/0.99/0.95	0.96	0.39/0.39	0.50	0.65/0.68/0.66	0.91	0.50	0.01
fitness	4	0.93/0.76/0.83	0.94	0.70/0.66	0.75	0.66/0.48/0.55	0.93	0.75	0.00
overall	24	0.91/0.95/0.92	0.92	0.56/0.56	0.83	0.65/0.62/0.63	0.93	0.62	0.01

Baseline vs four-turn, means over all 24 programs:

metric	baseline	four-turn	change
steps precision	0.88	0.91	+0.04
steps recall	0.37	0.95	+0.58
steps F1	0.48	0.92	+0.44
durations accuracy	0.85	0.92	+0.06
resources P/R	0.57/0.54	0.56/0.56	−0.00/+0.02
actors accuracy	0.79	0.83	+0.04
relationships F1	0.22	0.63	+0.41
tracks Rand index	0.80	0.93	+0.12
end-to-end pass	0.08	0.62	+0.54
unsupported-step rate	0.60	0.01	−0.59
calls / cost	49 / $1.10	111 / $2.20	×2.3 / ×2.0

The hypothesis holds on both numbers the PRD names, and by more on 24 programs than it did on the first 6: relationship F1 nearly triples (0.22 → 0.63) and the end-to-end pass rate goes from 2/24 to 15/24. The mechanism is extraction, not inference. The single-message prompt lumps a source into just over a third of the steps the expert wrote (recall 0.37) and fills the gaps with work the text does not support (unsupported rate 0.60), so there is little left for a trigger to point at — and its one end-to-end failure mode, a program that never passed the validator at all, is the run's only missing prediction (kitchen-two-burner-three-course-dinner, scored zero). Asking for the steps and their source spans before any relationship exists raises recall to 0.95 and all but eliminates unsupported steps (0.01); tracks and triggers then follow (Rand index 0.80 → 0.93). The effect is largest where the source is longest and most repetitive: baseline recall is 0.11 on the four fitness programs (interval sets it collapses into one step) and 0.24 in the lab, and four-turn lifts both above 0.75. Resource constraints are the one component the four turns do not move — naming what a step occupies is a vocabulary problem, not a structure problem — and actors stay a coin-flip on the event and fitness sources, where the expert counts people the text never numbers.

Regression policy / re-baselining. CI re-runs the four-turn pattern on the pinned model and fails on a relationships-F1 drop of more than 5 points or any end-to-end pass-rate drop against eval/four-turn.json:

rhylthyme eval-prompts compare \
  --baseline eval/four-turn.json --results eval-results/results.json \
  --rel-f1-drop 5 --e2e-drop 0

It prints a metric-by-metric diff, names the programs that moved, warns (rather than fails) when the two files disagree about the model, and exits 1 on a regression. Growing the gold set is not a regression; a program the results file is missing is, because a program that was not run cannot be said not to have regressed (--allow-missing waives it). Never widen a threshold to land a merge. Re-baseline quarterly to rotate the model pin, and always in a single commit that updates the pin, both JSON files, the cache and this table — see development/contributing.md for the procedure.

Reproduce both patterns, all 24 programs, without spending anything:

rhylthyme eval-prompts --gold ../rhylthyme-examples/gold \
  --model claude-haiku-4-5-20251001 --patterns baseline,four-turn \
  --concurrency 3 --max-fix-iterations 2 \
  --from-cache --cache-dir eval/cache --out /tmp/eval-replay

Prompt under test. eval/patterns/four_turn.py holds byte-identical Python copies of the four templates in rhylthyme-server/mcp-api/prompts.js, so the prompt the harness measures is the prompt the MCP server ships. A parity test extracts the JS exports with node and fails on any drift; check it by hand with

python -m rhylthyme_cli_runner.eval.patterns.four_turn

eval/patterns/baseline.py is frozen at the single-message prompt as of Phase 2 and is deliberately not in step with today's index.js; that is what keeps eval/baseline.json reproducible from the cache. eval/patterns/baseline.md is its rendered copy (server instructions + rhylthyme://guide/authoring + the plan_schedule message), regenerated with python -m rhylthyme_cli_runner.eval.patterns.baseline; a test fails if it goes stale.

Tests. The harness is unit-tested against a fake client; nothing in make test calls an API. The one real-API test carries the llm marker and is skipped unless RHYLTHYME_EVAL_LIVE=1:

make test-unit                 # excludes llm
RHYLTHYME_EVAL_LIVE=1 pytest -m llm      # one real call, opt in

Other models and a spending cap

--model also accepts models served through an OpenAI-compatible API. The provider is picked from the model id and the key is read from the environment:

Model id	Provider	Key
claude-*	Anthropic	ANTHROPIC_API_KEY
deepseek-* (e.g. deepseek-flash, deepseek-v4-pro)	api.deepseek.com	DEEPSEEK_API_KEY
qwen*	Alibaba DashScope (international)	DASHSCOPE_API_KEY
kimi-*, moonshot-*	Moonshot	MOONSHOT_API_KEY
glm-*	Z.ai	ZAI_API_KEY
anything with a slash, e.g. deepseek/deepseek-flash	OpenRouter	OPENROUTER_API_KEY
any id, with RHYLTHYME_EVAL_BASE_URL set	that endpoint (vLLM, Ollama, ...)	RHYLTHYME_EVAL_API_KEY, optional for localhost

Costs come from the price table in eval/llm.py (DeepSeek at its peak rates, so estimates are upper bounds). For a model the table does not know, set RHYLTHYME_EVAL_PRICE_IN and RHYLTHYME_EVAL_PRICE_OUT in USD per million tokens.

eval/run_model_comparison.py runs a model under a hard budget. It runs one gold program at a time, both prompts per program, records each program's cost in eval/models/spend-ledger.json, and stops before the next program could cross --cap. The ledger is cumulative across models and invocations, and the script refuses a model with no known price (it would be recorded as $0 and the cap would never trip).

One OpenRouter key reaches OpenAI, Qwen and Meta models (ids with a vendor prefix). Run the cheapest first and raise the cumulative cap as you go, so no single model can spend the whole allowance. Reasoning models need a higher per-call output ceiling than the default 16,000 tokens:

export OPENROUTER_API_KEY=...
L=eval/models/spend-ledger-openrouter.json
python eval/run_model_comparison.py --ledger $L --model meta-llama/llama-4-maverick --cap 0.80 --first-guess 0.05
python eval/run_model_comparison.py --ledger $L --model qwen/qwen3.8-flash --cap 2.00 --first-guess 0.08 --max-tokens 64000
python eval/run_model_comparison.py --ledger $L --model openai/gpt-5.6-luna --cap 4.50 --first-guess 0.15 --max-tokens 64000
python eval/compare_models.py

export DEEPSEEK_API_KEY=...
python eval/run_model_comparison.py --model deepseek-flash --cap 6.70 --dry-run
python eval/run_model_comparison.py --model deepseek-flash --cap 6.70
python eval/compare_models.py        # like-for-like table across every model run so far

Development

1. Clone the repository
2. Install development dependencies: pip install -e ".[dev]"
3. Run tests: pytest

License

Apache License 2.0