init: rewrite discovery prompt + add LLM-driven metadata enrichment

[jakejimenez]

Summary

The init flow today uses a single LLM call to suggest more command paths to probe, then fills description / system_prompt / safety / synonyms from hardcoded heuristics — every tool ends up with description: <tool> CLI, the same generic system prompt, and empty safety lists. The model has the help text in hand and can do better.

This PR implements both halves of the approved plan: an improved discovery prompt (§A) and a new metadata-enrichment call (§B).

§A — Discovery prompt rewrite

buildInitInferencePrompt and initInferenceSystemPrompt:

• XML-tagged input (<tool>, <root_help>, <verified>) per Anthropic's prompting guidance
• Long help text placed at the top of the user prompt (and trimmed to 2KB to fit Apple Intelligence's 4096-token ceiling)
• One concrete few-shot example showing the expected output shape
• Positive instructions ("Look for: …") instead of "Never invent…"

Output format unchanged — still newline-separated paths that DiscoverFromSeeds verifies.

§B — New metadata enrichment call

After the verified command tree is finalised, a second LLM call (enrichInitMetadata) asks for description, tool-specific system_prompt, safety patterns, and synonyms in a single YAML block wrapped in <yaml>...</yaml> sentinel tags.

The prompt includes:

• Two few-shot examples — docker (command_tree) and curl (flag_driven) — so the model generalises to either shape
• A leading <evidence> block where the model quotes help-text lines that justify each destructive pattern (grounds the safety list in reality; the parser ignores it)
• Mode-aware rules so synonym values map to subcommand paths or capability names depending on the tool's shape

parseEnrichmentYAML extracts the block via regex, validates with yaml.Unmarshal, and rejects placeholder leakage like <your-tool> in the description.

writeScaffold and writeFlagDrivenScaffold accept an optional *EnrichmentMetadata. When present, LLM-produced description/system_prompt replace the hardcoded defaults; LLM safety patterns populate the previously-empty arrays; LLM synonyms union with the heuristic ones via mergeSynonyms.

Apple Intelligence handling

nlci-apple's Swift bridge constrains output to a Generable CommandResult struct (apple/Sources/NLCIApple/App.swift:6-13), which cannot return free-form YAML. runMetadataEnrichment detects this via Router.ActiveName() == "apple" and skips §B silently with a clear stderr message:

Metadata enrichment skipped: Apple backend constrains output to single commands.
Use --backend ollama|llamacpp|lmstudio for richer metadata.

§A still runs on Apple — it produces newline-separated paths which fit within the bridge's command-shaped output.

Test plan

• go test ./... passes
• go vet ./... clean
• parseEnrichmentYAML covered: happy path, missing tag, empty block, junk-before-tag, placeholder-leak rejection
• mergeSynonyms covered: union semantics, nil-LLM passthrough
• nlci init docker on Apple: skips §B with the documented message; falls back to hardcoded
• Reviewer with Ollama: verify nlci --backend ollama init docker produces a real description and populated safety list
• Reviewer: verify nlci --backend ollama init curl produces a flag-driven-aware system_prompt

Files changed

• cmd/nlci/main.go — §A prompt rewrite, §B implementation, parser, merge helper, integration
• cmd/nlci/main_test.go — parser + merge unit tests; signature update for the existing writeFlagDrivenScaffold test

🤖 Generated with Claude Code