v1.101.0

What's New in v1.101.0

Added

• Hook-Intent Abstraction Layer — Defines little-loops hooks in terms of host-agnostic intents (PreCompact, SessionStart) backed by LLHookEvent/LLHookResult dataclasses. Python handlers in little_loops.hooks.* replace per-host shell scripts; adapters under hooks/adapters/<host>/ translate each host's native event into the wire format and back. (FEAT-1116)
• PreCompact and SessionStart Python Handlers — Ported precompact-state.sh and the session-start injector to pure-Python core handlers, with a Claude Code bash shim adapter and an OpenCode TypeScript adapter (hooks/adapters/opencode/index.ts). (FEAT-1449, FEAT-1450, FEAT-1455)
• LLHookIntentExtension Protocol — Third-party packages can contribute hook intent handlers via a provided_hook_intents() method; detected via hasattr() in wire_extensions() and merged into the global hook-intent registry. (FEAT-1452)
• OpenCode Adapter for Hook Intents — OpenCode TypeScript plugin calls python -m little_loops.hooks <intent> via Bun.spawn, setting LL_HOOK_HOST=opencode so dispatched handlers can identify their caller. (FEAT-1451)
• "Write a Hook" Authoring Guide — New docs/claude-code/write-a-hook.md covers the intent model, handler signature, LLHookIntentExtension registration, adapter flow, and pure-function + subprocess testing patterns. (FEAT-1458)
• type: learning FSM State — Learning states prove external-API/SDK assumptions against the learning-tests registry before advancing. Iterates targets in order; missing or stale records trigger /ll:explore-api; refuted records or exhausted retries route to on_blocked. Exempt from hard_max enforcement. (FEAT-1283)
• /ll:verify-issue-loop Skill — Generates a ready-to-run FSM verification loop YAML from a single issue's acceptance criteria. Each criterion becomes a dedicated verify state that fails fast if unmet. (FEAT-1310, FEAT-1446)
• /ll:explore-api Skill — Explores an external API or SDK target and writes a LearnTestRecord to the learning-tests registry (.ll/learning-tests/<slug>.md). Used by type: learning FSM states. (FEAT-1287)
• ll-learning-tests CLI — Query and manage learning-test registry records: check, list, mark-stale. Entry point at scripts/little_loops/cli/learning_tests.py. (FEAT-1286)
• Per-Edge Cycle Detection — max_edge_revisits (default 100) terminates a loop immediately with terminated_by="cycle_detected" when any state→state edge fires more than the limit, preventing tight two-state oscillations from draining the entire max_iterations budget. Edge counts survive --resume. (dd26036)
• Goals Discovery Fallback for scan-product — When .ll/ll-goals.md is absent, scan-product synthesizes a temporary goals context from existing docs (README, roadmaps, vision files) instead of hard-stopping. (ENH-1442)
• Hook-Intent Reference Documentation — Updated docs/reference/API.md, EVENT-SCHEMA.md, CONFIGURATION.md, and ARCHITECTURE.md with full field tables and dispatch contract for hook intents. (FEAT-1453, FEAT-1459)

Changed

• Progressive Throttling for FSM Loop Tool Calls — FSM loops now apply exponential back-off when tool-call rates exceed configurable thresholds, preventing runaway tool use in tight loops. (ENH-1115)
• autodev Routes Dead-End Decisions Before Size-Review — When a decide state fails, autodev now routes through triage before size-review, preventing incorrect decomposition of well-scoped issues whose bottleneck is an unresolved decision. (ENH-1415)
• New-Skill Classification Policy — Added a skill classification decision tree to CONTRIBUTING.md and ll-generate-skill-descriptions CLI for managing the skill listing budget. (ENH-1395, ENH-1396)
• Product Analyzer Improvements — Fixed output schema inconsistencies, removed double-deduplication between product-analyzer and scan-product, and wired product setup into /ll:init. (ENH-1401, ENH-1402, ENH-1403)

---

Full Changelog: v1.100.0...v1.101.0