Skip to main content
← Back to list
01Issue
FeatureTriagedSwamp CLI
Assigneesstack72

#299 Provide a CLI-shape primer for AI agents to reduce rediscovery overhead

Opened by bixu · 5/8/2026

Problem

AI coding agents working with swamp (Claude Code, Cursor, etc.) spend significant time on each session rediscovering the CLI shape — running exploratory --help calls, guessing flag names, re-learning subcommand hierarchy, and figuring out how subsystems (models, workflows, extensions, vaults, data, reports) relate to each other.

swamp help already returns machine-readable JSON, but:

  • Agents don't always know to consult it.
  • The full schema is large and doesn't give a conceptual map.
  • It tells you what exists, not how the pieces fit together or which command to reach for first.

This results in wasted context window, slower task completion, and inconsistent CLI usage across sessions.

Proposed solution

Ship a CLI-shape primer designed for agent consumption. Options (not mutually exclusive):

  1. Bundled meta-skill — a swamp-cli (or swamp-overview) Anthropic-style skill that gives a high-level conceptual map of the CLI surface and points at the existing per-subsystem skills (swamp-model, swamp-workflow, etc.) with a decision tree: "if you want X, reach for skill Y."
  2. swamp help --agent / --primer mode — returns a curated, opinionated tour (mental model + the 10–20 commands an agent will actually use) rather than the full schema.
  3. Versioned AGENTS.md — a markdown file shipped alongside the binary at a discoverable path (e.g. swamp help --agents-md or $(swamp --prefix)/share/AGENTS.md) that harnesses can ingest.
  4. Auto-loaded project hint — when .swamp.yaml is detected in a repo, surface a one-shot agent primer (e.g. printed by swamp doctor or a new swamp agent init).

Alternatives

  • Status quo: each agent harness / project maintains its own swamp prompt (current approach — this repo's CLAUDE.md duplicates a lot of conceptual framing that swamp itself could provide canonically).
  • Better discoverability of swamp help JSON during onboarding output.
  • Document in swamp.club docs only (doesn't help agents that don't browse).

Why this matters

The swamp-* skill set already exists and is excellent for deep per-subsystem work. What's missing is the shallow, top-level map that orients an agent on its first command of the session — the equivalent of a one-page README for the CLI's mental model.

02Bog Flow
OPENTRIAGEDIN PROGRESSSHIPPED+ 1 MOREASSIGNED+ 5 MOREREVIEW

Triaged

5/8/2026, 9:13:08 PM

Click a lifecycle step above to view its details.

03Sludge Pulse
stack72 assigned stack725/8/2026, 9:11:10 PM

Sign in to post a ripple.