rig must symlink installed skills into the harness discovery dir (~/.claude/skills), not just skills_target

[alex-mextner]

Bug: rig installs skills to skills_target (default ~/.agents/skills), but the Claude Code harness discovers Skill-tool skills from ~/.claude/skills/ (symlinks resolving to ~/.agents/skills). rig never creates those symlinks, so rig-installed universal skills (dead-code-investigation, shell-timeouts, promise-durable-action, naming, …) are NOT loaded/listed by the harness — only the ~11 that the tools' own install-skill happened to symlink.

Impact: any 'extract universal rules into agent-tools skills, slim AGENTS.md/CLAUDE.md' work is UNSAFE while this is broken — the rule appears 'self-advertised via a skill' but the skill is never loaded → the rule is silently lost.

Fix: rig apply (skills section) must idempotently symlink each enabled skill into the harness's discovery dir, keyed by harness kind (claude-code → ~/.claude/skills). Don't clobber real non-symlink dirs there. rig status reports link drift. (Manual fix applied on this machine: 39 skills symlinked, harness now lists them live. PR in flight: branch skill-harness-link.)

[alex-mextner]

Extend to ALL supported harnesses (tg#3747). The fix must not be claude-code-only — rig must provision agent-tools skills/instructions into EACH supported harness's discovery mechanism, which DIFFER:

• Claude Code (cc) — ~/.claude/skills/<skill> symlink dirs (Skill-tool) + settings.json.
• Codex — ~/.codex/AGENTS.md (instructions/blurbs).
• opencode (oc) — ~/.config/opencode/AGENTS.md.
• Gemini — ~/.gemini/GEMINI.md.
• commandcode (cmd) — CONFIRM: currently a model provider in our stack, not a CLI harness — does it have an agent CLI with its own instruction dir? If yes, add it.
• pi — CONFIRM full name + its instruction/skill dir (CTO wasn't sure of the name).

A per-harness target table (harness → discovery dir + format: skill-dir vs AGENTS.md vs GEMINI.md) keyed by harness kind. The supported-harness list must be documented in the README. (install.py already targets cc/codex/opencode/gemini/cursor for the tool blurbs; this generalizes it to the rule-skills + rig provisioning.)

[alex-mextner]

Harnesses identified (tg#3749): commandcode (cmd) has its own agent CLI — CONFIRMED, provision it. pi = earendil-works/pi (pi-mono / badlogic) — a minimal, extensible terminal coding-agent harness (Read/Write/Edit/Bash core, multi-provider, loads AGENTS.md). So pi provisions like codex/opencode (AGENTS.md-based), not like cc (skill dirs). Full supported-harness set for rig provisioning + the README list: cc, codex, oc, gemini, cmd, pi — cc uses ~/.claude/skills (dir-skills); the rest use an AGENTS.md / GEMINI.md instruction file in their global config dir.