Hey, nice addition — the architecture is clean and it doesn't touch any existing code. A few issues in scripts/retrieve_wiki_context.py that should be fixed before merge though:

1. Silent data loss — malformed JSON files are skipped with no warning

load_json() catches JSONDecodeError/OSError and returns None silently. If a campaign's concepts.json has a syntax error (realistic — these are written by LLM agents), that campaign vanishes from recommendations with zero indication. The user gets suggestions based on incomplete data without knowing.

Fix: print a warning to stderr when a file exists but can't be parsed.

2. Bare dict["key"] access will crash on missing fields

Entity/concept/parameter iteration uses entity["name"], e["principles"], e["definition"] without .get(). A single missing key in any campaign's data crashes the entire script with an unhelpful KeyError.

Meanwhile other parts of the same script already use safe access (d.get('id', '?'), f.get('title', '')), so it's inconsistent. Should use .get() with defaults everywhere.

3. JSONL metrics parsing has no per-line error handling

A truncated or corrupted line in llm_metrics.jsonl (common — campaigns get interrupted) crashes the whole script. Also accesses entry["model"], entry["num_turns"], entry["duration_ms"] without .get().

Fix: wrap json.loads(line) in try/except per line, and use .get() for field access.

4. Missing campaigns produce empty output with no warning

If --campaigns foo is passed and foo doesn't exist on disk, the script silently outputs an empty context block that looks valid. The LLM then generates recommendations grounded in zero data — exactly what this tool exists to prevent.

Fix: emit a stderr warning per skipped campaign, and if zero campaigns loaded, exit with an error.

Everything else looks good — the skill prompt is consistent with the wiki data model, paths are portable, and the design isn't overengineered. Just needs defensive error handling since the input data is LLM-generated and imperfect by nature.

All four issues addressed in 490c664:

1. Silent data loss — load_json() now prints to stderr when a file exists but can't be parsed
2. Bare dict access — all dict["key"] replaced with .get() defaults throughout retrieval and formatting
3. JSONL parsing — per-line try/except with stderr warning + .get() for field access
4. Missing campaigns — per-campaign stderr warning, hard exit if zero campaigns load successfully

Pattern: warn-and-degrade for partial failures (one bad line, one missing field), fail-hard for total failures (zero usable campaigns).