Skip to content

armgit / Ix

Public Imported from GitHub
0

Understand any codebase instantly. System intelligence for codebases, built for humans and AI.

3.3 KB
Raw Blame

Installing the skill into a harness

Where scripts/install-skill.sh puts skills/ix, how it decides a harness is present, and the two environment seams its probe battery reads.

This lived in skills/ix/SKILL.md until it was moved here. It is contributor documentation: an agent that loads the skill to answer a question about a codebase pays for every byte of it and can act on none of it.

The skill installer (scripts/install-skill.sh) reads an explicit, small harness table from ix-cli/scripts/skill-harnesses.mjs — claude, agents, codex and cursor only. Each entry's skills directory is verified against what that harness actually reads (Cursor uses ~/.cursor/skills-cursor, not ~/.cursor/skills); gemini, opencode, openclaw and vscode have no skills convention, so they are deliberately not install targets. Adding a harness is a one-line edit in that table, after checking where the harness really looks. ix mcp install uses the separate MCP host registry (ix-cli/src/mcp/hosts.ts) — a different table answering a different question.

When TOOLSCAN_PATH is set, both surfaces consult its discovery output so a harness CLI installed outside PATH can still be found — and ix mcp install executes the absolute path toolscan reports, so the off-PATH harness is inspected and registered through the binary toolscan found. Toolscan is optional and additive: if it is unset or fails, the embedded PATH and config-directory probes remain authoritative. TOOLSCAN_PATH is opt-in: neither surface ever looks toolscan up on PATH (the CLI executes whatever TOOLSCAN_PATH names, so only set it to a binary you trust).

The probe battery has two environment seams:

Seam Used by Default
TOOLSCAN_PATH skill-harnesses.mjs --probe, ix mcp install Unset — no toolscan, embedded probes decide
HARNESS_HOME skill-harnesses.mjs --probe os.homedir(); install-skill.sh uses the shell's $HOME

For a hermetic probe with no real user configuration:

HARNESS_HOME="$(mktemp -d)" \
  node ix-cli/scripts/skill-harnesses.mjs --probe

To drive the same probe with a local or npm-installed toolscan bundle:

TOOLSCAN_PATH=/path/to/toolscan/dist/toolscan.mjs \
  HARNESS_HOME="$(mktemp -d)" \
  node ix-cli/scripts/skill-harnesses.mjs --probe

--probe emits id|label|bin|config-dir|skill-dir|present|detectedVia, where present is 1 or 0 and detectedVia names the probe that decided (toolscan | path | config-dir | none). scripts/install-skill.sh --dry-run --json emits the same per-harness view as machine-readable JSON: one object per host with action (would-install | would-refuse | installed | refused | skip), dest, and detectedVia — so CI can assert how a harness was found, not just that it was. A would-refuse in the preview exits 1, matching the real run's conflict exit, so scripts can rely on the two agreeing.

The install report's --format json|llm output carries detectedVia per host (toolscan | path | config-dir | none), so a consumer can see which probe decided presence — a host toolscan found reads toolscan even when its CLI also happens to be on PATH, and a host found by the PATH probe reads path rather than being misattributed to its config directory.