Field notes from the frontier of AI.

Bug report: {{bug}} Relevant code: ``` {{code}} ``` 1. Restate the bug as expected vs. actual behaviour. 2. Write a failing test that reproduces it (same framework as the codebase). 3. Explain the root cause in at most 3 sentences, citing line numbers. 4. Provide the smallest fix as a unified diff. 5. List any other places with the same pattern that may share the bug. Do not refactor unrelated code.

Files: {{file_tree}} Key files content: {{snippets}} Write an onboarding guide for a new engineer: - What the system does, in 3 sentences. - Request lifecycle: trace one typical request through the files, in order. - The 5 most important modules and what each owns. - Where to make common changes (add an endpoint, add a field, add a job). - Gotchas: surprising behaviour, implicit conventions, known debt. Reference file paths for every claim. Mark guesses as (unverified).

Review this code for security issues only: ``` {{code}} ``` Check for: injection (SQL, shell, template), missing authorisation checks, unsafe deserialisation, secrets in code, SSRF, path traversal, weak crypto, missing input validation, sensitive data in logs. For each finding return: severity (critical/high/medium/low), line(s), exploit scenario in one sentence, fix as a minimal code change. If nothing is found in a category, do not mention it. End with: "Not reviewed:" listing what you could not assess from this snippet.

Document: {{document}} Audience: {{audience}} with 2 minutes to read. Write: 1. Headline: the single most important conclusion, max 15 words. 2. Three bullets: what happened, why it matters, what is decided or needed. 3. The ask: one sentence with a deadline, or "No action needed." 4. Risks: up to 2, each with likelihood and owner. Use numbers from the document; never introduce new ones. Max 150 words total.

You are reviewing a plan before it is executed. Plan: {{plan}} 1. List every assumption the plan depends on (explicit or implicit). Number them. 2. For each assumption, rate: likelihood it is false (low/med/high) and impact if false (low/med/high). 3. For every high-impact assumption, propose the cheapest test that would confirm or refute it within one week. 4. Rewrite the plan so the riskiest untested assumption is checked first. Output sections: ASSUMPTIONS · RISK TABLE · TESTS · REVISED PLAN. No preamble.