Prompt-only

Prompt-only learns one instruction file. The adapter uses the prompt-only@1.0 template: a single ADAPTER.md given to the acting agent as context, with no tools. After every step a fresh session of the built-in Codex Evolver reads the step’s learning material (each task’s input, the agent’s output, the judge’s score and feedback, and the trajectory) and rewrites ADAPTER.md. Start here for an instruction-only adapter whose changes are easy to inspect.

Setup

The learner is a Codex agent learner, so it needs the Codex CLI on PATH, Linux with bubblewrap for its sandbox, and a learner model on a backend that serves the Responses API.

template: prompt-only@1.0
evolver:
  name: codex
  resume: false                  # a fresh session every step
  params:
    model: fireworks/accounts/fireworks/models/glm-5p3-flash   # the learner: <backend>/<model id>
    args: ["-c", 'model_reasoning_effort="high"']
    home: []                     # no login files: the provider key is the backend's variable
    tools: [check_spec]          # the specification check only
    hide: []                     # the framework's examples, tests, docs and .git
sandbox:
  enabled: true
  network: restricted
trainer:
  output_dir: ~/runs/my-task     # outside the checkout: see `hide` below

This fragment belongs in a complete run config. The acting model is configured by your agent, separately from the learner.

How an update works

Each step is one Codex session. It reads the task description and the adapter’s EVOLVER.md, studies the step’s results, edits ADAPTER.md, calls check_spec, and writes a report whose first line says what it changed. The run records each session as one call with its token totals; to count the session’s individual model requests, put a metering relay in front of the learner, as the examples’ launch.sh does (hermit/runtime/wire_meter.py). The Trainer then applies its specification checks. There is no validation-based improvement gate in this method: a structurally valid update can reduce task performance. The final adapter is the current one.

Parameters and defaults

params:
  backend: codex
  prompt: START.md
  resume_prompt: RESUME.md
  args: []
  model: null                # the agent's model; null: its default
  price: null                # USD per million tokens {input, cached_input, output}
  home: [auth.json, .credentials.json]   # files of the agent's home copied into the run's
  tools: null                # the only Trainer tools the agent may use, e.g. [check_spec]; null: all
  hide: null                 # paths hidden from the agent besides examples/tests/docs/.git; null: nothing hidden

model binds the learner to a backend, which supplies its endpoint, key variable and price. args passes Codex settings, such as the reasoning effort or the model’s context window (-c model_context_window=...). tools: [check_spec] leaves the session only the specification check. hide lists paths the session must not read; any value, even [], also hides the framework checkout’s examples/, tests/, docs/ and .git, and with them a run written there, so keep trainer.output_dir outside the checkout.

examples/clbench_poker/prompt_only/, examples/clbench_database/prompt_only/ and examples/servelearnbench/prompt_only/ run this method as their streams’ baseline; their launch.sh sets the learner’s model and the run’s paths. Use a small smoke run before a long one, and evaluate with the same held-out protocol.