Prompt Iteration

The running session's prompt cannot change: personas are loaded when api-rs starts and pinned when a session is created, and compose-system-prompt builds AGENTS.md once at sandbox startup. To test an edit, recompose the prompt into a scratch directory and run a fresh, one-shot harness process there.

The child is a local process, not a durable Centaur session. It reaches the model through this sandbox's normal proxy path. Nothing is deployed.

Workflow

1. Edit the prompt in a writable checkout (git-branch <org/repo> <slug>, then edit under ~/branches/<org>/<repo>):

   • persona: tools/personas/<id>/PROMPT.md
   • base or overlay: services/sandbox/SYSTEM_PROMPT.md

2. Compose the prompt. By default this uses the session's current base, overlays, and persona:

   S=.agents/skills/prompt-iteration/scripts/compose_preview.sh
   # Persona edit:
   $S --persona ~/branches/<org>/<repo>/tools/personas/<id>/PROMPT.md /tmp/prompt-preview
   # Overlay-repo edit (replaces ~/github/<org>/<repo> as an overlay):
   $S --repo ~/branches/<org>/<repo> /tmp/prompt-preview
   # Base prompt edit (Centaur's own services/sandbox/SYSTEM_PROMPT.md):
   $S --base ~/branches/<org>/<repo>/services/sandbox/SYSTEM_PROMPT.md /tmp/prompt-preview
   

   Flags combine. Use --no-persona to compose without a persona. Read /tmp/prompt-preview/AGENTS.md to confirm the sections and their order before you run anything. The preview directory also links this session's installed skills, so the child sees the same skills.

3. Run a fresh child on the composed prompt, from the preview directory so it reads that AGENTS.md. Use the harness the persona targets. If the persona pins a model in [tool.centaur.models], pass that model.

   cd /tmp/prompt-preview
   claude -p --append-system-prompt-file AGENTS.md "<probe message>"
   codex exec --skip-git-repo-check --ephemeral "<probe message>"
   

   These flags match how harness-server passes the prompt to each harness. Add --model <model> to either command if you need a specific model.

4. Compare runs. Send the same probe messages before and after the edit (or with --no-persona), and judge the replies on the behavior the edit is meant to change. Repeat steps 1–3 until it's right. Then commit the edit and open a PR as usual.

Limits

The child gets the same composed prompt, skills, harness config, tools, and proxy credentials as this session. It differs in these ways:

• It runs on the harness config's default model and effort. Real turns use the model api-rs picks: the persona's pin, the channel default, or the user's flags. Pass --model to match.
• It gets one bare message: no thread history, and none of the context the chat ingress adds (sender, channel, attachments). If the prompt reacts to that context, write it into the probe message.
• It bypasses harness-server, so Nanocodex subagent tools and the other harnesses (Pi, Amp, Hermes) are not covered.
• It does not show persona selection, persona model pinning, or Slack rendering. Those require a deploy, and persona registry edits also need an api-rs restart.
• Keep probe messages free of secrets. The child runs with this sandbox's credentials, so give it the same tasks you would run here.