RuFlo Files

Plain-language summary: https://gist.github.com/ruvnet/496fe146754796522f133b31d526bc02

How to get it: the feature lives in the ruflo-mods 0.15.0 Claude Code plugin, delivered through the ruflo plugin marketplace (update it with /plugin). The npm CLI tarball bundles only ruflo-console and ruflo-metaharness, so npx ruflo@3.71.0 alone does not add it. This corrects a line in [#4652]'s description.

Minor release: @claude-flow/cli, claude-flow, ruflo 3.71.0 — automatic context delegation through ruflo-mods 0.15.0 (ADR-522). Off by default.

Added

• contextDelegation mod option (off | observe | enforce). ruflo-mods automates ADR-521:
• What qualifies: a main-loop, whole-file Read of a 14–120 KB file, for a prompt that routes to research or review and asks for no change.
• What happens: it is answered by a no-tools haiku completion on the session's own credentials, so no API key is needed. Every quoted line is checked against the file, and the model sees only the cited passages (the file's own lines), with a header saying how to read the rest.
• Why edits still work: the real Read always runs first. Any failure shows the whole file.
• observe changes nothing and records what would have been delegated in .claude-flow/metrics/context-delegation.json.
• Live A/B result (real binary, haiku main loop, 6 real repo files of 17–83 KB): 6/6 correct with delegation off and with it on, 6/6 reads delegated, total cost (main loop plus worker) −22.5%, main-loop input tokens −21.5%, wall time +6.5%. These are samples, not a statistical result; raw rows are in benchmarks/context-delegation/results-mod-live.json.

Security

• Mod contract: the "no model calls" check allows exactly one $.model.complete, in the opt-in module, and fails if that module is not gated on the option.
• No forged quotes reach the model: a quote whose words differ from the file never reaches it, because what is shown is always the file's own lines.

Verified after merge

• Live run against ruflo-mods from main, with the real claude binary and enforce, on agent-tools.ts (48,368 bytes):
• the read was delegated, and the model saw 1,360 bytes (−97.2%);
• the worker took 3.3 s and used 16,741 input + 328 output tokens;
• the answer was correct.
• Fresh npm i ruflo@3.71.0: a single agentdb (3.3.0), and npm audit 0 critical.
• Dist-tags: latest is 3.71.0 for all three packages; alpha/v3alpha are unchanged at 3.66.2.