meridian-model-call

meridian-model-call — call a model fast from our hooks/gates/judges (via Meridian, not claude -p)

Why this exists (read before writing any model-call code)

Anything in this repo that needs a model DECISION inside a hook/gate/judge/engine script (the FR-49b stop-gate judge, FR-8 semantic drift, anchor-fix fallback, future LLM gates) MUST call Meridian — the local subscription proxy — directly over HTTP. Do NOT shell out to claude -p. Measured on this machine (.dev-pomogator/.tmp/latency-probe):

Transport	Latency / call	Why
claude -p --model haiku	~13 000 ms	cold-starts the whole Claude Code CLI (MCP servers, hooks, plugin tree, CLAUDE.md) EVERY call
claude -p --bare (via proxy)	~5 000 ms	--bare skips MCP/hooks/plugin, but the CLI wrapper still costs ~5s
Meridian /v1/messages, thinking ON	~7 000 ms	the model reasons before the verdict
Meridian /v1/messages, thinking OFF	~2 400 ms	direct HTTP, no CLI, no reasoning — use this

claude -p for a binary judgment is the slow/dumb wheel. ~5× slower for the SAME model. Meridian is already the project's blessed path (see proxy-up, use-claude-subscription).

The recipe (fast model call, no API key)

curl -s -m 20 http://127.0.0.1:3456/v1/messages \
  -H "content-type: application/json" \
  -H "anthropic-version: 2023-06-01" \
  -H "x-api-key: sk-dummy" \           # proxy injects the subscription auth; any value works
  -d '{
    "model": "claude-haiku-4-5-20251001",
    "max_tokens": 64,
    "thinking": {"type": "disabled"},
    "messages": [{"role": "user", "content": "<your prompt — ask for ONE JSON line>"}]
  }'

• Response = standard Anthropic Messages shape: content[0].text holds the reply.
• thinking: {type: disabled} is the speed knob — leaving it on adds ~3-5s of reasoning you don't need for a classifier/judge.
• In Node, prefer fetch(...) (Node ≥18) over spawning curl.
• Model: claude-haiku-4-5-20251001 for cheap/fast judging; bump to Sonnet only when the judgment is genuinely hard.

FAIL-OPEN — never hard-depend on Meridian

A plugin-distributed hook runs on machines where Meridian may be down or not installed. The caller MUST degrade, never block on the proxy:

1. Probe first: curl -s -m 3 http://127.0.0.1:3456/health → expect {"status":...,"auth":...}.
2. If unreachable / non-200 → skip the model call and fall back (e.g. the stop-gate keeps its instant fact-gate + regex; FR-8 marks SEMANTIC_SKIPPED). Do not error, do not hang.
3. To bring it up: invoke Skill("proxy-up") (health → start → reauth decision tree).

When to use which skill

Need	Skill
Make a model call from OUR hook/judge/engine	this (the recipe above)
Proxy down / 503 / start-restart-reauth Meridian	proxy-up
Wire an EXTERNAL project's .env to the subscription	use-claude-subscription
Install/autostart Meridian for plugin users	infra ships in the plugin at <plugin-root>/tools/claude-subscription-proxy/; a SessionStart hook (ensure-up.cjs) auto-starts it (fail-open, MERIDIAN_AUTOSTART=false to opt out); container reuses the host Claude login (no separate claude login); proxy-up for manual ops

Anti-patterns

• ❌ spawnSync('claude', ['-p', ...]) for a judgment — ~13s cold-start; use Meridian (~2.4s).
• ❌ Leaving thinking on for a binary classifier/judge — wastes ~3-5s.
• ❌ Hard-failing when the proxy is down — always fail-open to the dep-free path.
• ❌ Putting an ANTHROPIC_API_KEY in code/hooks — Meridian uses the subscription; no key.
• ❌ Burning subscription budget on health checks — /health is local, doesn't hit upstream.

History

Created 2026-06-14 after the FR-49 stop-gate-judge prototype measured claude -p at ~13s (unusable per-stop) vs Meridian-direct-thinking-off at ~2.4s. The owner: «встрой Meridian в плагин… и скил сделай как и нахуя юзать, чтоб не придумывать велосипеды медленные». This skill is the canonical answer so no future agent reinvents the claude -p slow path.