Skip to content
Pi · mlx-community/MiniMax-M2.7-8bitfeatureCLI & tooling2 files · +235 −79

Add CLI flags to ppi wrapper for pi-mono coding agent

Session from combust-labs_pi-mono-docker on Hugging Face by rgruchalski · Apache-2.0 · condensed by Coders Talk

35msession
1interventions
GOAL

Extend the ppi Docker wrapper script to support additional pi-mono CLI flags (prompt, provider, system-prompt, session control, thinking level, etc.) and document them in the README.

  1. Prompt · +0
    Summarize the project, learn pi-mono's CLI args from source, and plan adding a --prompt flag to ppi
  2. Agent did · +4m
    Created feature/add-prompt-flag branch and added --prompt flag parsing to ppi, verified via grep/read
  3. Intervention · +6m
    Corrected the model name used for co-author credit in commit messages, twice (first to MiniMax-M2.7, then fixing 2.6 to 2.7)
    Why: the agent kept getting its own model name wrong in the commit credit, had to fix it manually
  4. Agent did · +7m
    Merged feature/add-prompt-flag into main with --no-ff and co-author credit, verified clean working tree
  5. Agent did · +9m
    On new branch feature/add-ppi-flags, added --provider, --system-prompt, --append-system-prompt, --continue, --resume, --no-session, --verbose, --offline flags with conditional passing, syntax-checked each time
  6. Agent did · +24m
    Added --thinking flag with is_valid_thinking() validation function, warning and clearing invalid values
  7. Agent did · +27m
    Listed supported/unsupported pi flags, merged feature/add-ppi-flags into main, then updated README with tables of 13 supported and 22+ unsupported flags plus usage examples
  8. Outcome · +35m
    README documents 13 supported and 22+ unsupported ppi flags; main branch is clean, 9 commits ahead of origin/main
VERDICT

I'd start by asking for the full flag list and a plan for all of them up front instead of adding flags one at a time across many prompts and commits. The one correction that mattered was fixing the model name used in the co-authored-by credit, since the agent got its own name wrong twice; everything else it handled cleanly with syntax checks after each edit.

I tried this · 0

Discussion · 0

How Builds work