Version 1.0.93 du 16/09/2026- Memory system now understands life domains — research, intent, and temporal context boost recall so stored facts surface when they are actually relevant.
- Input history recall lets you cycle through previous messages instantly, and zero-yield memory recovery prevents lost facts when extraction returns nothing new.
- Automatic token refresh keeps authenticated sessions alive — access tokens renew in the background before expiry so work is never interrupted mid-task.
- Theme system relabelled with real-life concept names and richer multi-accent palettes — dark chrome now bakes accent richness in at generate time for a more polished look.
- Panel chrome unified across the app — theme-accent tint now covers every chrome role, split sashes are flattened, and window widths snap cleanly to fitted columns on resize.
- Trace view polished — bottom status lines wrap into multiple rows, dead space removed, and agent trace layout enforces consistent spacing between content and scrollbar.
- Chat fallback wording is honest and chat-scoped — empty answers and clarification delivery now show clear status messages instead of generic placeholders.
- Reviewer retry is budget-aware — when the turn budget cannot fit a quality review, the engine skips it instead of forcing a late retry that cannot finish.
- Retry-waste trace counters make slow turns debuggable — wasted iterations are visible in the trace so you can see where the engine gave up on a dead path.
- Web research is more reliable — layered deadlines and chain observability prevent silent wedges, and the learned-patterns store survives busy files without data loss.
- Prompt-prefix caching wired for fan-out and retry rounds — repeated LLM calls reuse cached prefixes so retries are faster and cheaper.
- Citation and evidence handling hardened — truncated RAG evidence no longer triggers false-FAIL verdicts, and citation references use a single source of truth for consistency.
- Fixed web_search silent wedge — layered deadlines and chain observability prevent the search lane from hanging silently when a provider is slow or rate-limited.
- Fixed learned-patterns store corruption on busy files — the quarantine mechanism now survives WinError 32 sharing violations without losing learned routing data.
- Fixed OAuth secret leakage in logs — MCP config log dumps now redact sensitive paths and credentials before anything reaches disk.
- Fixed web-research loop on blocked pages — pages stuck in redirect loops or bot protection are remembered and skipped on retry.
- Fixed zero-gain dynamic-chain iterations — the planner no longer counts the same iteration twice, preventing phantom progress in multi-step research.
- Fixed clarification-turn delivery — when the LLM returns an empty response during a clarification question, the pending question is delivered instead of ending the turn.
- Fixed reviewer false-FAIL on truncated evidence — the quality reviewer now handles partial RAG evidence gracefully instead of rejecting valid answers.
- Fixed malformed pseudo-XML stripping — synthesis retries tolerate strip-to-empty edge cases without corrupting the prompt envelope.
- Fixed planner argument carriage — planned tool arguments now reach the executed call instead of being dropped on the native tool path.
- Fixed memory extraction collisions — BUG_601 SSOT pass verifies enablement reader, floor aliases, ineligible kinds, and memory-db-path as a single source of truth.
- Fixed model-fallback routing — the memory readback judge is now correctly mapped to the quick role, and orphan task keys produce a warning instead of failing silently.
- Extensive audit documentation added — session closeout notes cover the Einhell under-extraction arc, BUG_436 open-task audit, web-research reliability sweeps, theming overhaul, and prompt-cost caching arc.
- Zero-tolerance baseline completed — all tool execution sites converted to strict fail-open behavior with audit documentation and changelog coverage.
- Regression coverage expanded — SSOT behaviour tests, pattern-audit gates, web-research reliability pins, and memory extraction quality tests guard against future regressions.
(0 votes)
