First tagged release. Since the 1.0 baseline:
Honey harness
- SubagentStart hook (
hooks/honey-subagent.js): dispatched subagents never inherited the session's Honey directive — full-fat output, multiplied per dispatch. Now injected automatically at spawn; review-shaped agents get a prose-only variant (compress the report, never the verdict). Measured: −32% worker output at 100% test pass. - Loop cost discipline: new §Loops in the core skill +
honey-loopsatellite (cache-aware pacing, event-driven over polling, no-change short-circuit, stop condition). Measured: 95–100% pacing vs 81% off, 0 dead-zone picks, −66..74% tokens/tick. honey-superpowerssatellite: Honey directives for Superpowers-style subagent workflows (the hook is the mechanical form on Claude Code).- Regression gate (
bench/src/gate.js): editing SKILL.md now means re-verifying the numbers — tests/judge/output-reduction thresholds, mock-vs-live guarded. - CI: unit tests, mirror drift check,
verify-tests, three benches in mock mode, gate smoke test on every push/PR. - Two new benches: loop-tick (
bench:loop) and subagent-dispatch (bench:dispatch), mock + live.
ESON
- Honey's agent-to-agent format is now ESON — Efficient Structured Object Notation, developed in its own repo: https://github.com/Green-PT/honey-eson (normative spec v1.1, JS + Python reference implementations, conformance vectors, canonical LLM primer, Honey Wire Profile, negotiation).
- Vendored codec synced: wire header
!eson/1,encode(value, {number: true})for the reserved checksum-verifiednrow field. - CLI:
esonalias (BigInt-safe decode — fixes numeric-string corruption, #20). - Re-validated post-migration: live relay 100% lossless recovery at −52% output (baseline 83%).
Benchmarks (Opus 4.8 + GPT-5.5, committed)
Honey remains the only variant with no test regressions across all tiers while cutting output on every one: code −49%, user-facing −6% (top quality), agent-to-agent −51% at 100% recovery.
🤖 Generated with Claude Code