mirror of
https://github.com/ruvnet/ruflo.git
synced 2026-11-02 18:17:51 +09:00
rUv
5f3dacc54b
* feat(mods): automatic context delegation (ruflo-mods 0.15.0, ADR-522) + release 3.71.0 New contextDelegation option (off | observe | enforce, default off). A main-loop whole-file Read of a 14-120 KB file for an analysis prompt is answered by a no-tools haiku $.model.complete on the session's own credentials. Every quote is checked against the file and the model sees only the cited passages (the file's own lines) plus how to read the rest. The real Read always runs first, so a later Edit works (spike: answering the Read without next(e) breaks Edit); any failure shows the whole file. Live A/B (real binary, haiku main, 6 repo files 17-83 KB): 6/6 correct both ways, 6/6 delegated, total cost -22.5% incl. worker, main input -21.5%, wall +6.5%. First run delegated 0/6: haiku drops indentation in quotes; the verifier now trims per-line whitespace and relocates up to 3 lines, showing the file's own lines. Tests: 8 plugin tests, 3 real-binary e2e (extract + Edit, forged quote, off), CLI parity test vs DEFAULT_CONTEXT_LIMITS and core verdicts. Smoke step 4 now allows exactly one gated $.model.complete. Co-Authored-By: RuFlo <ruv@ruv.net> Claude-Session: https://claude.ai/code/session_01MfPHVHXWyvDvDtYMTsA6MP * ci: keep the claude-code test-kit plugin test out of the vitest ratchet Co-Authored-By: RuFlo <ruv@ruv.net> Claude-Session: https://claude.ai/code/session_01MfPHVHXWyvDvDtYMTsA6MP