v1.113.0

What's Changed

Added

• proof-first-task loop — Opt-in wrapper that gates any implementation loop on an assumption-firewall check before code changes begin. (FEAT-1738)
• learning-tests-audit loop — FSM loop for stale record detection and triage reporting in the Learning-Test Registry. (FEAT-1739)
• Learning-tests opt-in feature flag — /ll:init and config-schema.json now wire learning-tests as an opt-in feature flag. (FEAT-1743)
• ll-loop monitor subcommand — Live state polling and log tail for running loops. (FEAT-1764)
• Parallel-safe autodev for disjoint issues — autodev now declares scope: ["${context.run_dir}"], enabling concurrent instances with different issue sets to refine in parallel. (FEAT-1789)
• ll-loop run --baseline blind A/B comparison — Runs paired harness/baseline arms in parallel, feeds both outputs into an anonymized LLM judge, and aggregates results into ab.json with pass-rate deltas and token/duration ratios. (FEAT-1790, FEAT-1822)
• assumption-firewall --assume flag — Records untestable claims directly via CLI without interactive prompts. (ENH-1740)
• history_reader module — Typed read-only query API for history.db. (ENH-1752)
• Multi-profile design-tokens system — Active profile selector with profile-scoped token sets. (ENH-1768)
• W3C DTCG $value token format — Design-tokens resolver now supports the W3C DTCG $value field alongside the legacy value field. (ENH-1769)
• ll-loop diagnose-evaluators subcommand — Detects non-discriminating evaluator states from run history by computing per-state Bernoulli variance; flags states below threshold with pattern-matched recommendations. (ENH-1792)

Fixed

• ll-auto CONTEXT_HANDOFF signal forwarding — ll-auto now correctly forwards CONTEXT_HANDOFF signals to the outer FSM loop. (BUG-1759)
• Background loop scope conflict silent failure — Scope lock conflicts in background loops now surface a clear error message instead of failing silently. (BUG-1771)
• ll-loop list --json missing description field — The description field is no longer omitted from JSON output. (BUG-1779)
• Nested loop name crashes in background runs — Nested loop names no longer cause background run crashes. (BUG-1788)
• audit-issue-conflicts scans terminal issues — The command no longer includes done/deferred issues alongside active ones. (BUG-1799)
• audit-issue-conflicts unstaged files — git add .issues/ no longer stages unrelated untracked files. (BUG-1800)
• hitl-md generate state missing on_error routing — Missing on_error in the generate state no longer causes fatal loop termination. (BUG-1803)

Changed

• general-task retry hardening — Applies existing max_retries field uniformly across general-task.yaml loop states. (ENH-1677)
• FSM scope context template variables — FSM scope field now supports ${context.*} template variables for per-file locking. (ENH-1787)
• --no-lock flag for ll-loop run — Bypasses scope lock conflict detection when explicitly requested. (ENH-1778)
• hitl-md prompt extracted to shared fragment — 16 KB generate prompt moved to a shared file fragment for reuse. (ENH-1804)
• LLM model shown in ll-loop run and ll-loop monitor headers — Active model name now visible in run headers. (ENH-1805)

Full Changelog: v1.112.0...v1.113.0