Backends¶

The engine is backend-abstracted. Set backend in config or PROMPT_ENHANCER_BACKEND, or pass it per call.

Backend	How it works	Needs
`auto` (default)	API key → `api`; else a `claude` binary → `cli`; else `heuristic`	nothing
`cli`	`claude -p` with Haiku, prompt on stdin, `--output-format json`	Claude Code CLI
`api`	Anthropic SDK (`messages.create`), system-prompt caching	`prompt-preflight[api]` + key
`openai`	OpenAI-compatible chat completions	`openai` SDK + key
`ollama`	Local `ollama` chat endpoint over stdlib HTTP	a running Ollama
`heuristic`	Rule-based whitespace/casing/structure cleanup, no model	nothing

CLI backend¶

Uses your existing Claude Code login — no separate API key. The prompt is passed on stdin (never argv), so it can't leak into process listings and there's no shell involved. cli_bare (off by default) adds --bare to skip hook/skill/MCP/memory discovery for a faster start — but on some Claude Code versions --bare bypasses the interactive login (claude reports "Not logged in"), so it's opt-in; if an enabled --bare call fails, the engine automatically retries without it.

API backend¶

pip install "prompt-preflight[api]"
export ANTHROPIC_API_KEY=sk-...
enhance-cli --json "..."   # backend auto-selects api when the key is present

The large system prompt is sent with cache_control: ephemeral to cut cost and latency on repeated calls. Token usage and cost are surfaced in EnhanceResult, --json, and the proxy /stats.

Bedrock / Vertex¶

Set api_provider to bedrock or vertex (or PROMPT_ENHANCER_API_PROVIDER). These use their own cloud credential chains — no ANTHROPIC_API_KEY — and require the matching Anthropic SDK extra to be installed.

Offline: Ollama & heuristic¶

ollama gives fully local, zero-cost enhancement for confidential or air-gapped use. heuristic needs nothing at all and is the auto fallback when no model is reachable; it only tidies whitespace, casing, and terminal punctuation, so it can never distort meaning.

Plugin backends¶

Third parties can register a backend without editing the engine, via an entry point:

# in your package's pyproject.toml
[project.entry-points."prompt_preflight.backends"]
mybackend = "my_pkg.module:enhance_fn"

enhance_fn(raw_prompt, cfg, *, start) must return an EnhanceResult. Select it with backend = "mybackend". A crashing or misbehaving plugin fails open — it can never break enhancement.

Profiles¶

--profile concise|detailed|coding|research (or the profile config field) selects a tuned system-prompt variant. coding preserves code, identifiers, paths, and error text exactly; research frames the request as a precise analysis question; concise prefers the shortest faithful rewrite.