`terminal_synth / tmax`¶

TMax composes terminal skills into new tasks and checks that the starting environment contains the fixtures described by the task.

Pipeline, step by step¶

flowchart TD
  S["Legacy taxonomy + seeded sampler"] --> A["Domain, 3–5 skills, complexity, scenario and language"]
  A --> P1["P1 · Task description and private truth"]
  P1 --> P2["P2 · Initial-state tests"]
  P2 --> P3["P3 · Final-state tests; sees initial tests"]
  P3 --> P4["P4 · Starting fixtures and reference solution"]
  P4 --> Q["Q1 · Optional draft consistency review"]
  Q --> I["Remote initial-state tests must pass"]
  Q -->|"Blocking issue"| F
  I --> R["Fresh final-state nop + oracle"]
  I -->|"Fixture failure"| F["Bounded builder feedback"]
  R -->|"Wrong reward or error"| F
  F --> P4
  R -->|"0 / 1"| H["Export task with taxonomy lineage"]

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

Sample requirements. The owned sampler retains legacy TMax domain/skill axes, weighted languages and optional real-software anchors. This profile uses the first three task-complexity categories.

Separate starting and solved states. P1 produces a public description and private truth. P2 checks what must exist before solving. P3 sees those tests and describes the completed state.

Build and test both states. The builder creates the starting fixtures and reference; initial-state execution happens before final-state baseline/oracle trials. The common builder can repair an inconsistent expectation using execution feedback.

Every prompt and its data¶

The four native authoring stages below can be followed by Q1 consistency review when review_drafts: true. Q1 runs on each complete draft before initial-state tests and returns blocking defects to P4 within its existing repair limit.

Four calls on the first successful attempt: template, initial tests, final tests, environment/reference builder.

Call	System prompt composition	User / input material	Output	Retry or branch
P1 · Template	template_prompt.md with domain_label and DOMAIN_MODULES substitutions; v2_block is empty	Sampled taxonomy requirements.	TaskTemplate: description, truth	Legacy corpus only.
P2 · Initial tests	initial_prompt.md + shared adaptation	Description, private truth; initial_tests is null.	TestProgram: code	Five to ten top-level pytest tests.
P3 · Final tests	final_prompt.md + shared adaptation	Description, truth and generated initial_tests.	TestProgram: code	Separate model call.
P4 · Build / repair	environment_prompt.md + templates.builder_prompt additions + common materialization	Complete TemplateDesign and failure feedback.	TerminalDraft	Default initial build plus two repairs.

Read the complete tmax 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: sampled time-series, encoding and aggregation skills become a sensor-data task. Initial tests check the supplied input fixture; final tests check the requested cleaned aggregate; the reference performs the transformation.

What repeats, what is checked¶

P1–P3 run once per sampled design. Failed initial fixtures or final baseline/reference checks return to P4. The first runtime supports text fixtures and an unprivileged offline solver. The reward is binary all-required-tests success; stored weights do not make the current grader fractional.

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-tmax.yaml. The seed file is a JSON sampler configuration, for example:

{"corpus_kind":"legacy","count":40,"seed":24,"domains":["data_processing","file_operations","data_querying"],"languages":["Python","Bash"]}

Omit domains or languages to retain the full corresponding taxonomy. The sampler preserves the upstream legacy axes and weighted language selection; restrictions select a conditional subset. Configuration and sampled axes are recorded with the generated artifacts. Options include target, max_candidates, max_repairs, seed, test_timeout_sec and max_tokens.

This first profile generates text fixtures in a single CPU Docker container. All dependency installation happens during image build. The solver runs as user, without runtime internet, and can edit /workspace and /home/user. Initial-state tests, final-state tests and the reference stay outside the learner image. The builder receives execution errors for bounded repairs.

The published dataset contains 55 tasks. The measured sample added 35, averaging $2.10 per new task including failed attempts and estimated compute. See economics for the sample denominator and unresolved charges.

It runs the native initial-state check and a fresh Harbor baseline/reference pair. Detailed reward-hack review, blind solver trials and independent acceptance are separate from generation. TMax v2 multimodal fixtures, metric verifiers and the upstream large sampled-solution stage are outside this first profile.

Use the shared worker, budget and progress interface for Modal or Daytona. Credit: TMax, Apache-2.0, commit 7387d2f9142397a458dc39f0827a2ab0b4c03cda. See RFC 0018 and packaged recipes/tmax/provenance.md for retained files and adaptations.

terminal_synth / tmax¶