Skip to content

v0.2.0 — harness registry, dry-run, JSON status, doctor probes, adopt mode

Latest

Choose a tag to compare

@jroell jroell released this 10 Jul 17:51
· 1 commit to main since this release

Highlights

Harness registry. Every harness is one declarative record (detect dir, instruction path + strategy, skills dir + strategy, CLI, note). All commands are registry-driven; adding a harness is one line. Warp is now honestly modeled as manual — the fabric no longer creates ~/.warp/AGENTS.md, since Warp's global Rules are cloud-managed.

fabric sync --dry-run. Prints every would-link/would-backup action and writes nothing. Use it before any bulk change.

fabric status --json. Machine-readable state (hub, skills, per-harness detection/strategy/wiring: wired / conflict / missing / manual / not-detected) for scripts, CI, and backup jobs.

fabric doctor. Effective-loading probes beyond symlink topology: resolves each instruction path against canonical content, passes genuine @-import overlays, warns on byte-identical copies, fails stale/divergent copies, checks broken skill links and CLI presence. Manual harnesses are reported as unverifiable — never falsely passed.

install.sh --adopt [PATH]. Adopts an existing hand-rolled hub in place (defaults to ~/.shared-agent-memory): validates layout, preserves your AGENTS.md/skills, adds the CLI + starter skill, wires all harnesses to your hub, and bakes the hub path into the installed CLI so FABRIC_HOME never needs exporting. The competing-hub guard now points here.

Also

  • Hermes support with documented edge cases (memory-repair de-symlinks MEMORY.md; curator archiving) — validated against a live installation
  • Competing-hub installer guard (v0.1.x) hardened with tests
  • README: worked examples captured from live E2E runs against the published repo

Testing

61 sandboxed assertions (throwaway $HOME; real home never touched) + macOS CI, plus end-to-end verification of the published curl-pipe install, adopt flow, dry-run no-op, JSON schema, and doctor stale-copy/repair cycle.

Platform: macOS only (Apple Silicon, macOS 13+).