sc-ax

# sc-ax — Agent Experience ## Language Keep durable instructions in English. Reply in the user's language unless requested otherwise. ## AX definition AX asks: how easily can an agent understand the project, find the correct action, call it with minimal ambiguity, know what changed, recover from failure, resume work, and produce evidence without wasting context? Optimize for: `discoverability → navigability → operability → observability → recoverability → resumability → verifiability`. ## Project orientation An agent should be able to answer quickly: - what this project is, - canonical source vs generated output, - safe commands and verification commands, - important constraints/exclusions, - current state and recent relevant changes, - where domain-specific knowledge lives, - what actions require confirmation or credentials. Avoid forcing an agent to crawl the whole repository for facts that can live in a compact canonical index/knowledge file. ## Tool/action design Agent-facing tools/functions should: 1. Have one clear purpose and bounded side effects. 2. Use stable, explicit names and typed/structured inputs. 3. Prefer identifiers returned by previous calls over ambiguous re-search. 4. Return structured state plus concise human-readable impact. 5. Distinguish `not found`, `permission denied`, `invalid input`, `conflict`, `transient failure`, and `partial success`. 6. Be idempotent where repeated calls are plausible. 7. Refuse destructive ambiguity rather than guessin...