Papersmith AI in Pi — commands, skills and agents Diagram

How Pi sees it

.pi/APPEND_SYSTEM.md is the system prompt: Pi appends it to the project one
/<skill> = .pi/prompts/<skill>.md: it loads the SKILL.md and passes $ARGUMENTS
The 29 .pi/agents/ are projected from .claude/agents/; Pi core has no subagents, so they need pi-subagents
Each agent declares tools (read, find, grep, write, edit, bash, mcp) and its stretch
refuse-offpath-push.js intercepts bash tool_call; it is a warning that fails open, not a gate

Harness commands

npm run setup:harnesses — links .pi/skills -> ../skills
python scripts/sync-repo-harness.py — regenerates .pi/prompts, .pi/agents, .pi/extensions (--check detects drift)
smith init · ui · status · ingest · deliberate · implement · run · remote · novelty · audit
npm run test:all — the quality gate

Per-skill CLIs

paper-ingestion: python scripts/setup_env.py install
deliberation: node skills/<skill>/cli.mjs --serve
the proposal graph: merge_proposal_sky.py → check_atlas.py → render_atlas.py
implementation: implementation_cli.py materialize · verify · fidelity · audit · step
kaggle-accounts: accounts_cli.py list · discover · validate · remove · materialize
remote-execution: remote_cli.py submit · status · poll · fetch · readiness …
paper-writing: paper_cli.py status · plan · write · verify … (31 verbs)
journal-finder: journal_cli.py search · sjr-import · status
venue adaptation: paper_cli.py venue select · requirements · acquire · diff · plan · approve · build
datasets: datasets_cli.py synth · search · license · qc · prepare · publish
figure-review: review_cli.py probe · raster · skill-audit: audit_cli.py

Flow rules

The order is optional: each skill refuses with a code when its input is missing
Starting with /plausibility, before ingestion, is the recommendation
The proposal is built in order: the central problem and its pillars, then the hypotheses, then the mathematics; each step stays locked until the one before it exists (Mapa Conceptual tab, one Generar y deliberar button per step)
paper-reviewer classifies the NeurIPS contribution type of the proposal (model_spec.json) and of each stage's findings (discussion.json)
Phase 5 runs protocol → toy set → replication → proof of concept → its verification (G-EXP-POC) → timing → optional search → full experiments → review → rehearsal → launch
The full experiments stay locked until the operator verifies the proof of concept (smith novelty approve --gate poc, or the Novedad tab card); the writing waits for G-NOVELTY
/datasets gives a seeded toy set first (dataset-scout: sourced → licensed; dataset-curator: toy → checked → prepared → published); publishing is private and only on the operator's consent
The proof of concept reads the toy set; the full experiments read the one official subject-wise preparation, mounted by remote notebooks
paper-replicator (implementation fidelity) searches for published code first: adopt it and audit its correspondence to the paper, else reimplement from the ingested paper; its baselines feed the experiments
experiments-reviewer reviews the instrumented code before any rehearsal or launch
No walk can launch a campaign: the operator authorizes it
contract-auditor, redactor and style-sampler have no code path: the orchestrator launches them and write judges
/figure-review and /skill-audit only report; they never repair
Once the paper compiles, /journal-finder ranks journals by quartile, access, APC and an estimated review time; nothing is submitted, and without an SJR file the quartile is a labelled OpenAlex proxy
The operator picks a journal in the PDF Preview tab; venue-adapter reads its guidelines and every requirement carries a quote and a url
The operator approves the adaptation plan by its sha256; text changes go only through write, and venue build never writes paper/main.tex
smith ui: each pipeline phase card sends its /<skill> with the rules as a guide, or opens the panel that owns the work