/mode — LLM Routing Switch

Manage delegation across a multi-LLM setup. The pattern: Claude orchestrates, judges, reviews, and surgically fixes; cheaper or specialized CLIs handle the heavy lifting.

Other skills in this collection read .claude/llm-mode.json to decide whether to delegate. When mode is single, every skill runs natively in Claude.

Route by argument

Parse the argument after /mode:

• No args → show status
• multi → switch to multi
• single → switch to single
• stats → show delegation log
• reset → reset to defaults

Show status (default)

Read .claude/llm-mode.json and display:

## LLM Mode: <MULTI / SINGLE>

### Delegation targets (in multi mode)
  READING (Antigravity CLI / agy):  [available / not found]
  WRITING (Codex CLI):              [available / not found]
  WEB SEARCH (Perplexity):          [key set / key missing]

### Delegation stats
  Total:     N successful delegations
  Failed:    N (X% failure rate)
  Last:      YYYY-MM-DD HH:MM via <skill>

Switch with: /mode multi  or  /mode single
Full log:    /mode stats

To populate the display:

1. Read .claude/llm-mode.json for mode and stats.
2. Check agy --version for Antigravity CLI. Fall back to gemini --version for enterprise plans still on the legacy binary.
3. Check codex --version for Codex CLI.
4. Check $PERPLEXITY_API_KEY (or your provider's env var) for web search.
5. Count entries in .claude/delegation.log for the totals.

If .claude/llm-mode.json doesn't exist, show Mode: SINGLE (default — no config file yet).

Switch to multi

1. Verify the Google delegate CLI is installed and authenticated. Prefer Antigravity CLI (agy); fall back to the legacy gemini binary for enterprise plans.

   # Antigravity CLI (consumer + new enterprise).
   # The `</dev/null` redirect is mandatory: `agy -p "..."` deadlocks waiting for
   # TTY input even when a prompt arg is supplied.
   agy --version 2>/dev/null \
     || gemini --version 2>/dev/null
   
   agy -p "Reply with exactly: OK" </dev/null 2>/dev/null \
     || gemini -p "Reply with exactly: OK" -o text 2>/dev/null

   If the auth check fails, abort with installation/auth instructions. For agy, the install command is curl -fsSL https://antigravity.google/cli/install.sh | bash.

2. Update .claude/llm-mode.json: set mode to "multi", update updated_at to the current ISO timestamp.

3. Append to .claude/delegation.log:

   YYYY-MM-DDTHH:MM:SSZ | mode | ok | 0s | Switched to multi-LLM mode

4. Check Codex CLI and Perplexity availability non-blockingly. Report each as "available" or "not found / key missing" in the confirmation.

Switch to single

1. Update .claude/llm-mode.json: set mode to "single", refresh updated_at.
2. Log: ... | mode | ok | 0s | Switched to single-LLM mode.
3. Confirm: "Switched to single-LLM mode. All skills using Claude natively."

Show delegation log

Read .claude/delegation.log and show the last 20 entries plus summary stats: total, failure rate, average duration on success, most-delegated skill.

Reset

1. Write a fresh .claude/llm-mode.json with defaults (single mode, zeroed stats).
2. Clear .claude/delegation.log (keep any header comments).
3. Confirm.

Delegation logging protocol

Other skills that delegate should follow this protocol:

After a successful delegation:

echo "$(date -u +%Y-%m-%dT%H:%M:%SZ) | <skill> | ok | <duration>s | <notes>" \
  >> .claude/delegation.log

Also increment stats.delegations_total and update stats.last_delegation in .claude/llm-mode.json.

On failure:

echo "$(date -u +%Y-%m-%dT%H:%M:%SZ) | <skill> | fail | <duration>s | <reason>" \
  >> .claude/delegation.log

Increment stats.delegations_failed.

Schema for .claude/llm-mode.json

{
  "mode": "single",
  "updated_at": "2026-06-04T00:00:00Z",
  "stats": {
    "delegations_total": 0,
    "delegations_failed": 0,
    "last_delegation": null
  }
}

Known gotchas

• Multi mode without authenticated CLIs silently falls back to Claude. That can mask a real problem — the user thinks they're saving tokens but every delegation is failing. The status display surfaces both auth state and recent failure rate for this reason.
• The log file grows unboundedly. Rotate or truncate it before it crosses ~10k lines or the stats view gets slow.
• Mode is per-project. Each project has its own .claude/llm-mode.json. Setting it in one project does not affect siblings.
• agy -p "<prompt>" requires </dev/null. Without the stdin redirect, Antigravity CLI waits on a TTY even when a prompt arg is given. The probe in "Switch to multi" uses the canonical agy -p "..." </dev/null form; mirror it in every other call site.

Anti-patterns

• Flipping to multi mode for a quick win, then forgetting to switch back when the delegated CLIs go stale. Add the status check to a /brief-style routine.
• Trusting delegations_total as a token-savings proxy. The skills that delegate know their own savings; the mode skill just tracks counts.

Validated patterns

• Logging both success and failure (rather than just success) keeps the failure-rate metric honest. A 30% failure rate is a signal that the auth or model config needs attention.

Self-improvement

This skill ships with a lightweight feedback loop. Adopt or ignore — the skill works without it.

Trigger a review when:

• The status display is wrong about CLI availability (the detection command needs updating).
• A new external CLI joins the routing pattern.
• learnings.md crosses ~100 bullets (consolidation time).
• The skill mis-triggers or fails to trigger.

Consolidation pass (5–10 min, weekly or threshold-driven):

• Each entry gets one fate: apply, capture, or dismiss.
• Apply = merge into Known gotchas / Anti-patterns / Validated patterns.
• Capture = leave in learnings.md for now.
• Dismiss = delete.
• Bump last-consolidated: in frontmatter.