@agimon-ai/po-model-guidance

Per-model system prompt guidance for Pi agents, layered across global and repository scope.

Packages

Package details

extension

Install @agimon-ai/po-model-guidance from npm and Pi will load the resources declared by the package manifest.

$ pi install npm:@agimon-ai/po-model-guidance
Package
@agimon-ai/po-model-guidance
Version
0.0.0-alpha.183
Published
Nov 22, 2026
Downloads
31.3K/mo · 5,095/wk
Author
agiflow-ai
License
BUSL-1.1
Types
extension
Size
61 KB
Dependencies
3 dependencies · 1 peer

Security note

Pi packages can execute code and influence agent behavior. Review the source before installing third-party packages.

README

@agimon-ai/po-model-guidance

Per-model system-prompt guidance, selected through Po's optional llm layer. Guidance is keyed by the exact model ID reported by the host, so a new model's adjustments do not change older models' prompts.

Enable

Add @agimon-ai/po-model-guidance to a layer selected by your preset in .po/presets.yaml, then run po sync. Listing it in default.packages enables it for every preset. Preserve the other package entries.

The built-in preset currently supplies guidance for gpt-6-astra to continue agreed plans and scoped tasks without repeated confirmation of minor decisions. It does not remove security or approval requirements.

Configure

Custom guidance lives in ~/.po/model-guidance.yaml and <repository>/.po/model-guidance.yaml:

modelGuidance:
  claude-opus-5: |
    Keep the task focused. Never go off rails.
  gpt-6-astra: |
    Prefer fewer, larger tool batches.

Model IDs match exactly. Use the ID shown by /model; a newly versioned ID needs its own entry.

The built-in preset applies first, then the home file, then the repository file. Each later entry replaces the earlier text for that model only. Models without an entry receive no guidance. A malformed or unreadable file warns on stderr and falls back to the built-in preset rather than ending the session.

The native before_agent_start hook reads the files for each turn and appends the selected guidance to the system prompt. An edit therefore applies on the next turn without a restart. This package publishes a server entry, not a standalone Pi extension.

Development

pnpm exec nx run-many -t lint typecheck build test -p @agimon-ai/po-model-guidance

The hook lives under src/extensions; matching and merging live in src/services/modelGuidance, and layered file reads in src/services/modelGuidanceStore.