Skip to content

Agents Operating SkillMeat

An interactive walkthrough of how a coding agent uses SkillMeat, in lifecycle order.

Open the walkthrough →

Every command, every number and every screenshot in it came from a live enterprise instance running 0.82.0 — 828 artifacts across 47 registered projects. Where the product could not do what a step needed, the walkthrough says so instead of staging it.

What it covers

Stage What the agent does
01 Scaffold skillmeat init registers a project, scaffolds a harness profile (claude_code, codex, gemini, cursor), and --pull-pending pulls waiting deployments under conflict mode skip-if-different
02 Discover & deploy skillmeat discover takes intent in plain language — quick mode calls no model at all — then deploy --dry-run before the real deploy
03 Drift & versions sync-check detects divergence against the recorded content_hash; sync-pull --auto-link matches untracked files by hash; real version lineage with parent chaining
04 Propagate project reconcile's real pipeline: derive expected → diff on-disk → classify → resolve policy → act

It also documents the MCP contract, which is deliberately narrow: exactly two tools, discover_artifacts and install_artifact. Deploy, sync, reconcile and version operations are not exposed over MCP — an agent reaches those through the CLI.

Why it shows real drift

Stage 03 renders an actual diff — 84 additions and 6 deletions on dev-execution.content, with real @@ hunk headers side by side. Earlier builds of the enterprise edition could not display a diff for any artifact: the Sync Status view always reported no changes, even when a project copy had genuinely diverged. That is fixed, and stage 03 is a live capture of the working view.

It documents its own limits

The closing section carries five measured limits as first-class content rather than omissions — three open defects found while building the page, and two commands that simply do not exist:

  • There is no skillmeat sync-push. Collection→project is deploy; project→collection is sync-pull. The asymmetry is intentional.
  • There is no one-shot fleet fan-out. project reconcile is the real primitive and it is per-project; fanning a fix across many projects is orchestrated outside this CLI.
  • No fleet-wide deployment total is claimed, because only per-project counts were measured.

That section is the most useful part of the page for anyone deciding what to build on.