`terminal_synth / endless_terminals`¶

Endless Terminals generates terminal tasks from a category, complexity and scenario. Its native category list covers files, text, databases, configuration, software tools and other terminal workflows.

Pipeline, step by step¶

flowchart TD
  S["Seeded sampler JSON"] --> A["Uniform category + complexity + scenario draws"]
  A --> P1["P1 · Description and private truth"]
  P1 --> P2["P2 · Initial-state tests"]
  P2 --> P3["P3 · Final-state tests"]
  P3 --> P4["P4 · Environment fixtures and reference"]
  P4 --> Q["Q1 · Optional draft consistency review"]
  Q --> I["Initial tests on unsolved environment"]
  Q -->|"Blocking issue"| F
  I --> R["Final tests: fresh nop + oracle"]
  I -->|"Fixture failure"| F["Bounded builder repair"]
  R -->|"Wrong reward or error"| F
  F --> P4
  R -->|"0 / 1"| H["Export sampled terminal task"]

P1, P2, … identify actual model calls. Unlabelled stages are code or remote execution.

Sample without a repository. Category, complexity and scenario are independent uniform draws from retained native lists. The seed contains the request constructed from those choices.

Generate the task and tests. The description/truth call is followed by separate initial-state and completion-test calls. The final-test author sees the initial tests.

Construct a runnable environment. The fourth call writes starting text fixtures and a reference solution. The same remote preflight and bounded builder-repair machinery as TMax executes the result.

Every prompt and its data¶

With review_drafts: true, the shared runner adds Q1 consistency review after P4 and before the initial-state check. Each repaired draft gets a new review. The default is false; existing run configurations record whether it ran.

Four calls on the first successful attempt; the environment builder is the stage repeated during repair.

Call	System prompt composition	User / input material	Output	Retry or branch
P1 · Template	template_prompt.md + shared adaptation	Sampled category, complexity, scenario and constructed request.	TaskTemplate: description, truth	One call per design.
P2 · Initial tests	initial_prompt.md + shared adaptation	Description and truth.	TestProgram: code	Five to ten top-level pytest tests.
P3 · Final tests	final_prompt.md + shared adaptation	Description, truth and initial_tests.	TestProgram: code	One separate call.
P4 · Build / repair	environment_prompt.md + templates.builder_prompt additions + common materialization	TemplateDesign and actual failures.	TerminalDraft	Initial call plus max_repairs.

Read the complete endless_terminals prompt reference for every retained template, appended instruction, substitution, example and output schema. The shared prompt guide explains how to inspect the fully resolved request from a real run.

Follow one task¶

Illustration: the SQLite category and a backup scenario become a database export task with supplied records and a precise archive format. The initial checks assert that inputs exist; final checks inspect the requested archive.

What repeats, what is checked¶

This shares execution machinery with TMax, but not TMax’s taxonomy or weighted language sampling. No training, adaptive sampler or blind-agent acceptance step is part of this generation recipe.

An exported bundle is a generation result. Independent leakage review, shortcut probes and blind solver traces belong to the later quality campaign.

Implementation map¶

Run and supported profile¶

Run repo2rlenv generate --config examples/owned-endless-terminals.yaml. The input is a sampler JSON file:

{"count":40,"seed":24,"categories":["text processing and manipulation","backup and archiving","SQLite database operations via CLI"]}

Omit categories to sample the full native list. All three axes use uniform sampling, preserving the original method. TMax's separate recipe uses its own domain/skill taxonomy and language weights. They share the metered authoring and remote execution stages.

The first profile uses text fixtures in an offline CPU Docker container. The solver runs as user; dependencies are installed at build time. Initial tests execute on a fresh environment before the final baseline/reference pair. The builder can repair inconsistent fixtures or a failing reference within max_repairs. Every attempted input and execution outcome remains in the run.

The measured generation sample completed 100 tasks, now published as HuggingEnvs/repo2rlenv-endless-terminals. All 80 newly generated tasks have hash-matched baseline-0/reference-1 controls; independent semantic quality evaluation and blind rollouts remain separate. Upstream's sampled solutions and training run are outside this generation milestone. Options and cloud setup follow the shared interface; the configuration records the actual author model rather than claiming the original Qwen settings.

Credit: Endless Terminals, Apache-2.0, commit 99f4c74b75faacf21e53d3dc01df170902e924cb. See RFC 0020 and packaged recipes/endless_terminals/provenance.md for the source map and adaptations.

terminal_synth / endless_terminals¶