- README: front door states the full system (core + experimental shadow-only
inspection model + distributed fleet learning); new "Fleet learning" usage
section; learn/ + fleet_export/merge + governor in Architecture; a proper
"Documentation" map; license corrected MIT -> GPLv3 (matches LICENSE).
- AGENTS.md: crisp YAMS work-loop (retrieve/inspect/persist/report), replaced
the stale ~64-line dated "Current Project Reality" with an accurate concise
state + pointers (PLAN/DEVELOPER/git/YAMS), trimmed stale work-order/milestone,
added "Keep This File Lean".
- CLAUDE.md: work-loop + evidence rule up top.
- gitignore the local validation/attack-range harness (deploy/validation/{docker,vm}).
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
2.4 KiB
2.4 KiB
CLAUDE.md
Read AGENTS.md first: it holds this repository's rules, current reality, guardrails and the per-change check table. This file only points at the shared workflow.
Work loop: retrieve → inspect → act → verify → persist → report (YAMS-first; see AGENTS.md "Operating Model"). Validate material claims — code behavior, paths, command syntax, versions, counts, completion status — against current source or tool output before asserting them; a successful edit is not a working feature and a passing test covers only what it exercises.
- Memory and retrieval: YAMS-first loop in docs/prompts/PROMPT-eng-codex.md (claim → checkpoints → handoff; plans, decisions and evidence live in YAMS, hydrated by hash before use).
- Audits: docs/prompts/PROMPT-strawberry-audit.md:
brick (structure), fold (
scripts/fold_changed.sh, bounded symbolic SAST), pipe (enum classstate flow). - Before calling a change done: the "Per-change checks" table in AGENTS.md
(meson tests incl. source invariants and fault injection, clang-tidy on
changed files,
make analyze(clang static analyzer, strict) on changed TUs,make fuzz, fold on changed TUs, live Docker tests,make perf-kpi). Index the change in YAMS at each checkpoint, before and after. - Performance & C++: heavy or hot-path changes carry a constrained-lane KPI
(
make perf-kpi,scripts/perf_kpi_report.py --compare, 3 repeats) and Tracy zones (STRAWBERRY_ZONE,-Denable_profiling=true) recorded inPerfStatsand surfaced in statustimings. Code reading is not evidence. - CLI changes: follow docs/cli-ux.md (yams-style
noun verbtable insrc/cli_commands.cpp,--jsoncontract, exit codes, terminal-safe output, "adding a command" checklist). - Never leave a
catchempty: handle it, or record it withSTRAWBERRY_NOTE_SWALLOWED()(include/strawberry/swallowed.hpp). Enforced by source invariant R7 and clang-tidybugprone-empty-catch(.clang-tidy). - This machine: builds and tools run niced and capped (
nice -n 19,-j2, Docker--cpus=2); an uncapped fold build has frozen it before. - Plans in flight: YAMS
task=strawberry-data-engine(staged data engine + online learning),task=strawberry-settings-ux,task=strawberry-encoder.