v0.34.0

0.34.0 (2026-03-29)

Features

• add CGo llama.cpp backend and wire into EmbeddedProvider (ed71564)
• add DB pair extraction, fix data loading, prep for EXP-14 (2ff222f)
• add deploy pipeline, embedding fine-tune script, and pre-register EXP-9 (bf534bc)
• add dismiss_abstraction tool, recall filtering, IDs in output (5279493)
• add embedding extraction via mean-pooled hidden states (6da2d0d)
• add lightweight D3 shim replacing CDN dependency (Phase 4) (9491930)
• add llama.cpp submodule and fix GGUF export for Felix architecture (93f1766)
• add llama.cpp submodule, CGo bridge, and Qwen spoke training docs (6c9e1d2)
• add LoRA support, gradient checkpointing fix, and EXP runner scripts (9ba9738)
• add per-token logit validation for embedded LLM quality gating (96775a2)
• add Phase 3 data pipeline scripts and pre-register EXP-6/7/8 (640123b)
• add Phase 3 fine-tuning pipeline (9c7df1c)
• add Phase 3 fine-tuning pipeline (498dd28)
• add Q8_0 quantization support and prefer quantized models (b7a2488)
• add Qwen spoke adapter, re-tokenization pipeline, and pre-register EXP-11-14 (371464a)
• add ROCm HIP link flags and encoding comparison script (8ae61c0)
• add server-side episode_id filter to /memories endpoint (934b6a6)
• add standalone command center dashboard for GitHub tracking (e750f60)
• add thread view, restyle SDK/LLM/Tools headers (Phase 3/7) (0651281)
• add training script and evaluation hooks for Qwen spoke experiments (36d8bdb)
• agent identity system — sources become forum users (20ba192)
• breadcrumbs update on every view switch (afe0d62)
• clickable live feed posts — navigate to relevant views (eb47ac2)
• clickable memory/pattern/abstraction rows — expand to show detail (9b9e757)
• complete Epic #339 dashboard cleanup — digest threads, associations, D3 removal (0939956)
• complete EXP-9 mixed fine-tune, fix embedding script, add v2 model (5bb94dc)
• complete structural rewrite — all views use phpBB patterns (26cf76c)
• episode-aware @mentions — agents know which episode you're asking about (7984ba9)
• extract CSS sections and add forum component library (Phase 2) (746e636)
• extract CSS tokens, remove D3 CDN, archive mockups (Phase 1) (354570d)
• fix instrumented model label for embedded provider, add nomic BERT test (f5e0c24)
• forum categories — sub-forum index page with phpBB layout (c2b3197)
• forum communication layer — posts, threads, agent personality, @mentions, internalization (fe7445d)
• forum UX — @mention autocomplete, quote button, @tag names, blank reply fix (d1b662b)
• forum-style dashboard redesign (Epic #339) (66ae0cd)
• functional depth — clickable memory sections, project routing, episode project tagging (05303c7)
• improve recall quality for LLM agents, fix Windows self-update (7a8bfa6)
• improve recall quality for LLM agents, fix Windows self-update (aa07982)
• live activity feed — agents post to forum in real-time (fb96d78)
• MCP agent UX — dismiss_abstraction, recall filtering, hook fixes (d747591)
• nested forum navigation — index > group > sub-forum > thread > post (fb9ca12)
• Phase 3-4 autoresearch — fine-tuning pipeline, CGo backend, experiments (49ad590)
• populate welcome panel with stats + last visit tracking (ef88f51)
• project auto-detection, data-aware agents, agent-to-agent chat, thread subscriptions (1354424)
• Qwen 3.5 2B + Felix spoke training infrastructure (21facd3)
• remove Mind/Graph view entirely (Phase 3 partial) (94056ca)
• render quoted text as styled blockquote boxes (a6147a9)
• restyle SDK, LLM, Tools views to forum aesthetic (#349) (ec8ce1a)
• rewrite Explore and Recall renderers to forum style (Phase 3-5) (2c05b8d)
• rewrite Timeline renderer to forum rows (Phase 6) (fb4e78f)
• run spoke gate analysis (EXP-8) and fix synthesis data generation (c43587c)
• scaffold modular dashboard structure (Phase 0) (6145e9a)
• standalone command center dashboard (ea0536e)
• strip coaching prompts from training data, tune LR to 3e-4 (68c2725)
• structural rewrite — phpBB-inspired dl/dt/dd forum layout (cca709a)
• thread view shows Episoding Agent post when no encoded memories (49f6b14)
• transform nav to forum top bar + navbar + footer (Phase 2-3) (ef1c48a)
• unified forum index — merge welcome+live feed, add Memory System section (ad634b3)
• update llama.cpp submodule with Qwen 3.5 spoke support (75fa3f5)
• verify nomic-bert embedding GGUF works via llama.cpp backend (f0b71dd)

Bug Fixes

• add missing MCP tools to web UI agent allowed_tools list (ddb419f)
• add scaleSqrt to D3 shim, unblocking SDK/Tools pages (b717df9)
• align forum layout with phpBB prosilver column proportions (c74d768)
• allow text wrapping in forum rows, widen title column (eec5d63)
• backfill category_id for existing forum posts (9b2db92)
• correct hook matchers, add work-first rule, align platform docs (fd165ec)
• D3 shim — stack().value(), datum(), attr(fn) for chart rendering (6d53442)
• D3 shim area/line accept constants, not just functions (bca0321)
• D3 shim selectAll parent tracking + axis tickValues/tickFormat (14c2d0d)
• disable thinking for forum @mention replies — root cause of truncation (cbafcd8)
• disable thinking project-wide on thinking models (7539c13)
• encoding agent checks closed episodes for episode_id linkage (542d7bf)
• episode memory count discrepancy — show observations not encoded count (8c85970)
• episode-memory linkage race condition — backfill on close and startup (f27e791)
• expand-zone toggle — remove inline display:none that overrode CSS class (48f86bb)
• float clearing for dl/dt/dd columns + title overflow (abbb48e)
• forum index visual hierarchy — column headers, cleaner nesting (4a48d68)
• increase @mention response token limit from 200 to 512 (a37baf9)
• ingest salience floor, JSON recall associations, graph edge priority, concept vocabulary (088dca0)
• live feed onclick — fix nested quote escaping in HTML attributes (ca16c33)
• make version number a clickable link to GitHub release (3ac7011)
• map source event types to display types, cap salience at 100% (ad369fa)
• memory section clicks — expose handler on window object (7592dd3)
• memory section clicks — use addEventListener instead of inline onclick (b5bbc88)
• memory section onclick — use data attributes to avoid escaping issues (2778ac0)
• navVersion not showing — null reference on renamed healthDot (223a3bd)
• prevent crashes from embedding failures and concurrent llama.cpp access (0c7f69e)
• quote button, clickable @tags, dropdown styling (fa12f7b)
• remove crashy switchExploreTab, add null safety to loadThread (5f63467)
• remove JS text truncation — let CSS handle wrapping (205ec55)
• remove old .tl-card CSS that broke timeline layout (06d185c)
• replace old fblock classes with forabg, remove dead CSS, null safety (a202266)
• resolve duplicate declarations breaking ES module loading (1f88e1a)
• resolve merge conflicts with main (679b22d)
• resolve NaN loss, dtype deprecation, and VRAM issues in training (0bc050d)
• revert ingest salience floor exemption, keep heuristic boost (ad0c3ab)
• skip pre-migration backup when schema is current, add retention (bed03d4)
• skip pre-migration backup when schema is current, add retention (75a74c3)
• split ROCm link flags, add context window safety, and wire GBNF grammar (9294aa0)
• stress test issues — ingest salience, JSON associations, graph edges, concept vocabulary (5ef4bfc)
• switchView crash + LLM ERR from missing navTokenCount element (7d5cf08)
• thread view filters memories client-side by episode_id (8006864)
• thread view loads memories via episode_id filter (8fd5199)
• thread view unwraps nested episode response (85d8dac)
• timeline layout, thread CSS, tag filtering compatibility (73269bf)
• update EmbeddedProvider prompt format to match Felix-LM fine-tuning (be284b3)

Performance Improvements

• composite indexes + strip embeddings (2.5s → 9ms) (5eec78b), closes #352
• strip embeddings from /memories list response (2MB → 49KB) (b27fa9e)

What's Changed

• fix: stress test issues — ingest salience, JSON associations, graph edges, concept vocabulary by @CalebisGross in #340
• fix: skip pre-migration backup when schema is current, add retention by @CalebisGross in #358
• feat: Qwen 3.5 2B + Felix spoke training infrastructure by @CalebisGross in #359
• feat: Phase 3-4 autoresearch — fine-tuning pipeline, CGo backend, experiments by @CalebisGross in #360
• feat: standalone command center dashboard by @CalebisGross in #353
• feat: improve recall quality for LLM agents, fix Windows self-update by @CalebisGross in #361
• feat: forum-style dashboard redesign (Epic #339) by @CalebisGross in #362
• feat: MCP agent UX — dismiss_abstraction, recall filtering, hook fixes by @CalebisGross in #363
• chore(main): release 0.34.0 by @appsprout-release-please[bot] in #341

Full Changelog: v0.33.0...v0.34.0