SMALL

AGENTS.md - Agent Execution Rules

This document is the canonical, authoritative AGENTS.md for SMALL.

Any agent operating in a SMALL-managed repository MUST follow these rules exactly. Failure to do so invalidates the run.

SMALL Protocol Agent Guide

This is the operating contract for any AI agent or human acting like one. Follow it exactly. If anything here conflicts with small --help, the CLI wins.

SMALL governs state, not execution. The CLI is the only valid way to mutate .small/.

Target profiles: SMALL v1.0.0 and v2.0.0

Scope: all repos that claim SMALL compliance

Zero-trust rules

  • Never manually edit .small/progress.small.yml
  • Never invent timestamps
  • Never mark tasks complete with plan toggles
  • Completion is only via small checkpoint
  • Never run small handoff unless small check --strict passes
  • Never assume chat memory; always read state from disk and CLI

V1 artifact roles

  • .small/intent.small.yml - purpose and scope
  • .small/constraints.small.yml - hard limits and policy
  • .small/plan.small.yml - planned tasks
  • .small/progress.small.yml - append-only audit trail
  • .small/handoff.small.yml - resume context
  • .small-runs/ - run snapshots
  • .small-archive/ - archived lineage

V2 authoritative state

V2 uses human policy plus immutable session, event, resolution, import, and receipt records under .small/. Do not create or edit those records manually. Use small policy, small session, small mode, small reconcile, and the v2-aware lifecycle commands. Local locks, active-session pointers, transaction journals, and indexes stay under ignored .small-cache/.

Agent-safe command set

Read-only

  • small version (also small --version or small -v)
  • small selftest
  • small status --json
  • small emit --check --workspace root
  • small doctor
  • small run list
  • small run diff
  • small policy show --json
  • small session list --json
  • small reconstruct --resume --session <id> --json

Plan and progress

  • small plan --add "Task title"
  • small progress add --task <task-id> --status in_progress --evidence "..." --notes "..."
  • small checkpoint --task <task-id> --status completed|blocked --evidence "..." --notes "..."

Execution capture

  • small apply --cmd "<command>" --task <task-id>

A successful apply records command outcome only. Accept the task separately with small checkpoint after reviewing the required evidence.

Gates and lifecycle

  • small check --strict
  • small handoff --summary "..."
  • small archive (milestones only)

Forbidden for agents

  • Any timestamp override flags (-at, -after)
  • Plan-only completion (small plan --done)
  • Manual edits to .small/*.yml (except intent or constraints when explicitly authorized)
  • Handoff before strict check passes

Session lifecycle

Always begin clean. Always end with a gated handoff.

Session start

Read artifacts on disk. Do not assume chat memory.

Then run:

bash 
small status --json
small check --strict

If strict fails, fix the workspace before any new work.

Ralph loop

Repeat until all tasks are completed or blocked.

Refresh state

bash 
small status --json
small check --strict

If strict fails, fix immediately.

Select task

Choose a pending task with no unmet dependencies.

If none exists, add one:

bash 
small plan --add "Clear task title"

Start task with evidence

bash 
small progress add \
  --task <task-id> \
  --status in_progress \
  --evidence "Starting: what will be done" \
  --notes "optional"

Execute bounded changes

For any mutating command:

bash 
small apply --cmd "<command>" --task <task-id>

Read-only exploration may run outside apply.

Any command that changes files, deps, builds, tests, or migrations must use apply.

Complete atomically

Completed:

bash 
small checkpoint \
  --task <task-id> \
  --status completed \
  --evidence "What changed and why" \
  --notes "optional"

Blocked:

bash 
small checkpoint \
  --task <task-id> \
  --status blocked \
  --evidence "Why blocked and what is needed" \
  --notes "optional"

Enforce gate

bash 
small check --strict

Do not continue with a failing workspace.

Session end

Always gate, then hand off:

bash 
small check --strict
small handoff --summary "Current state in one paragraph"

Milestones only

When a meaningful milestone is reached:

bash 
small check --strict
small archive

ReplayId threading

  • Treat the replayId from small status --json as the primary key
  • All progress and handoff context must match the active replayId
  • Strict check should fail if replayId is missing or mismatched

Evidence quality

Good evidence:

  • Precise
  • Names files or systems touched
  • Explains why the change was made

Bad evidence:

  • "Fixed it"
  • "Build works now"

Stop rules

Stop and block if:

  • Intent or constraints are missing or unclear
  • Secrets or external access are required
  • A human decision is needed
  • Strict check cannot pass

Blocked handoff example:

bash 
small checkpoint \
  --task <task-id> \
  --status blocked \
  --evidence "Need decision on X before proceeding"

small check --strict
small handoff --summary "Blocked on decision X"

Glossary

Agent

An LLM or automation acting through the SMALL CLI

Evidence

A factual description of what changed and where

Checkpoint

Atomic plan plus progress update with evidence

Strict check

The enforcement gate before any handoff

This contract exists to prevent drift. Follow it exactly.