CHANGELOG.md

Changelog

Mọi thay đổi đáng kể của 8sync ghi vào đây. Format theo Keep a Changelog, versioning theo SemVer. 8sync rule: mỗi PR cập nhật mục Unreleased.

[Unreleased]

[0.47.0] — 2026-07-06

Added — cross-platform builds: macOS + Windows (was Linux/Arch-only)

• New crate::platform module — the OS seam. platform::os() (compile-time constant per target), os_name(), require_linux() (clean no-op guard for Linux-only verbs), pkg_manager() (pacman ⁄ brew ⁄ winget), install_core_pkg() (per-manager package-id map), and a cross-platform periodic timer — install_timer/remove_timer backed by systemd user timer (Linux) ⁄ launchd LaunchAgent (macOS) ⁄ Scheduled Task (Windows). Uses only cross-platform std/crate APIs (no std::os::unix) so one body compiles on every target.
• Portable prebuilts: dropped target-cpu=native from .cargo/config.toml (it baked the build host's ISA into the binary → SIGILL on older CPUs of the same arch). Release binaries now run on any CPU of the target arch.
• harness up --timer + clean --timer now route through platform::* — identical systemd behavior on Linux (incl. the 0.46.2 cgroup memory bounds), launchd StartInterval on macOS, schtasks /SC MINUTE on Windows.
• setup Stage A is cross-platform: gh installs via the native package manager (github-cli on pacman, gh on brew/winget); paru/AUR is skipped off-Linux; Arch-only Stage B profiles are skipped with a clear note on macOS/Windows. omp/codegraph keep their curl installers (POSIX shells).
• sec, bt, clean guard with require_linux — a clean "Linux-only" message + no-op on macOS/Windows instead of shelling out to absent tools.
• CI (.github/workflows/release.yml): on a v* tag, a matrix builds real binaries on native runners — musl-static Linux x86_64/aarch64 (aarch64 via cross), macOS x86_64 (macos-13) + arm64 (macos-14), Windows x86_64 (MSVC) — and publishes them to a GitHub Release. Assets keep the 8sync-<tag>-<os>-<arch> scheme (--locked, contents: write).
• Installers: install.sh extended to the full os×arch matrix (linux x86_64/aarch64, darwin x86_64/arm64) keeping the source-build fallback; new install.ps1 (Windows) — irm …/install.ps1 | iex, -Uninstall, $env:SUSYNC_VERSION pin, User-PATH wiring.
• Verified: Linux release build clean (0 warnings, 6.16 MB); Linux runtime smoke (sec/clean/help) unaffected. mac/Windows compilation + binaries are produced and verified by the CI native runners (a Linux host cannot build MSVC/Apple-SDK targets or the C deps rusqlite/zstd-sys without each platform's toolchain).

[0.46.2] — 2026-07-06

Fixed — harness up --timer OOM-killed the machine

• The generated 8sync-harness-up.service had no cgroup resource limits, so the per-tick codegraph index (peak RSS ~5.3 GB on a large repo, e.g. zus) tripped the kernel OOM killer — reported as Result: oom-kill, Mem peak: 5.3G every 10 min — thrashing swap and taking down other apps.
• The service unit is now bounded to its own cgroup + de-prioritized: MemoryHigh=2G (reclaim throttle), MemoryMax=4G (hard ceiling — kills only this unit, never the machine), MemorySwapMax=512M, OOMPolicy=stop, Nice=15 / CPUWeight=10 / IOWeight=10, TimeoutStartSec=900.
• Verified: codegraph now held at ~2 GB by reclaim pressure instead of ballooning to 5.3 GB. Re-run 8sync harness up --timer <dur> in a project to regenerate the bounded unit (already-installed unbounded units are overwritten).

[0.46.1] — 2026-07-06

Fixed — sweep now redeploys the project-level /auto command

• 8sync harness global --sweep migrated a project's memory folder (agents/ → su-code/) but left the project's .omp/commands/auto.md (and 8sync-engine.ts) untouched — so /auto in a swept repo kept reading agents/STATE.md from a stale copy deployed by an older binary (project commands take precedence over the global one in omp).
• stamp_project (the per-repo sweep layer) now calls deploy::ensure_engine, refreshing both the /auto command and the engine extension in every swept project. Byte-identical writes stay quiet. Verified: all projects under ~/Projects now have su-code/-only /auto commands (0 stale).

[0.46.0] — 2026-07-06

Changed — the agent-memory folder is now su-code/ (was agents/), a distinctive project marker

• BREAKING (auto-migrated): every project's agent-memory dir is renamed agents/ → su-code/ so a su-code-managed repo is unambiguously identifiable (the old agents//AGENTS.md markers are generic — every repo has them). The AGENTS.md anchor file stays (open-standard entry point); only the folder moves, and its links are rewritten to point at su-code/.
• Detection now keys on su-code/. is_omp_project (sweep) + the harness project-root detection recognise su-code/ (with agents/ kept only as a legacy migration trigger). 8sync harness global --sweep [DIR] (the sweep command — not harness all up) migrates every legacy project it finds.
• Auto-migration (memory::migrate_legacy_layout, runs in here/init/up/bare-harness/sweep): renames agents/ → su-code/ and rewrites agents/ → su-code/ references in the anchor + live memory markdown. Guarded: only fires on a real 8sync memory dir (identified by STATE.md/KNOWLEDGE.md/PROJECT.md/PLAYBOOKS.md/skills.toml/skills/), so a source package literally named agents/ is never touched. .agents/ and subagents/ are protected from the text rewrite. Idempotent.
• All 8sync-authored code, assets, skills, docs, and the recall hook (with an agents/ fallback for un-migrated repos) updated to su-code/. Historical CHANGELOG entries left as-written (they document the agents/ era). Verified E2E: legacy project migrated (folder + refs), guard skipped a non-memory agents/ source dir, this repo dogfood-migrated, build clean.

[0.45.0] — 2026-07-06

Added — MCP server.json standard conformance (marketplace install) + the spec as a machine-wide default

• The marketplace MCP install now conforms to the official MCP registry server.json spec (schema 2025-12-11, modelcontextprotocol/registry). official_install (crates/cli/src/verbs/harness/marketplace.rs) rewritten to honor:
  • registryType → runtime: npm→npx -y · pypi→uvx · oci→docker run -i --rm (+ -e NAME env forwarding) · nuget→dnx (was: everything defaulted to npx).
  • version pinning — identifier@version (or identifier:version for docker images).
  • runtimeArguments + packageArguments rendered into the command line (named/positional Argument shapes).
  • transport.type — a package/remote with streamable-http/sse becomes a remote ({type,url,headers}), not stdio.
  • BUGFIX: environmentVariables → a {NAME: value} MAP, not the array of descriptors it used to write (which produced an unusable mcp.json env for any server needing secrets). Required-but-empty vars are surfaced in the install note.
• /api/mcp/add (McpAddBody, web.rs) + the dashboard install flow (api.ts, App.tsx) thread env/headers maps end-to-end (were dropped before). Verified live against the registry: docker (apithreshold → docker run … -e … img:0.1.0), pypi (armor-mcp → uvx armor-mcp@0.6.1 + env:{ARMOR_API_KEY:""}), npm — all via the UI, 0 console errors.
• The standard is now a machine-local default that forces AI adherence. A distilled spec reference ships in the binary (assets/specs/mcp-server.md) and 8sync harness global/init/up deploy it to ~/.omp/specs/mcp-server.md (ensure_mcp_spec). A short rule in the always-on APPEND_SYSTEM.md points every omp session at it: when writing/editing mcp.json, follow the on-disk standard — env/headers are maps, runtime from registryType, pin version — never invent fields.

Added — /auto engine: gitleaks gate before every autonomous commit

• engine_advance {commit:true} (assets/extensions/8sync-engine.ts) now runs a secret gate (gitleaks protect --staged, matching the 8sync pre-commit hook) before committing — a finding aborts the commit and unstages, so an unattended /auto run can't leak a secret when the pre-commit hook is absent. No-op when gitleaks isn't installed (best-effort, no regression).
• /auto reviewed + independently functional-tested (Bun harness): all 6 engine_* tools register; verify-gate FAIL→WARN(2×)→BLOCK(3× doom-loop); engine_advance refuses an unverified task; pass→advance→done; trivial no-verify advance; commit path.

Added — plan: Agent Terminal App (deep-researched, build later)

• agents/plans/agent-terminal-app.md — Tauri v2 + xterm.js/WebGL + portable-pty + Zellij-backend (resurrection sau reboot) + omp sidecar. Positioning: "cmux for every OS" (cmux: 22.3k stars/4 tháng, macOS-only → khe cross-platform đang mở). Gồm stack đã chốt, 4 phase MVP, trend mechanics, rủi ro, metrics, next actions + sources.

Fixed — omp startup error providers: must be an object (was null)

• ~/.omp/agent/models.yml was left with a bare providers: key (YAML null) when the local-model registry became empty (e.g. after add-local-model rm <last>), making omp print a schema error + disable custom providers on every start. insert_block now finalizes the file: no real children ⇒ providers: {} (valid empty object); a later add reopens {} and inserts under it. Both branches proven live via the real binary and A/B'd against omp (omp models list): bare ⇒ error, {} ⇒ clean.

Added — harness help: LOCAL GGUF MODEL real-flow example block

• Copy-paste flow: add from .gguf path / HF repo id / URL → list → use once via 8sync ai --model local/<name> → set as default/code model → rm. Points at the TSV registry + sentinel-managed provider block.

[0.44.0] — 2026-07-05

Added — loop-engineering stop signals in the 8sync engine (doom-loop guard + real gate)

• Per the loop-engineering literature (Avi Chawla, "Prompt, Context, Harness & Loop Engineering", Jul 2026): an agent's own "done" is not a stop signal. Two code-enforced fixes in assets/extensions/8sync-engine.ts:
• engine_advance now actually enforces the gate — it REFUSES a task that has verify commands but no passing engine_verify run (new verified flag per task; previously the description claimed "code-enforced" but advance set done unconditionally). Tasks with zero verify commands keep the documented trivial-advance path.
• No-progress detector (doom-loop guard) — engine_verify fingerprints each failure output (FNV-1a); 2 consecutive identical failures WARN ("change the approach"), 3 BLOCK the task early even below maxRetries (a byte-identical failure means the retry did nothing but burn tokens). New per-task failStreak/lastFailureHash state; old state.json files load via zod defaults (backward compatible).
• /auto command updated: different-fix-per-retry rule, advance-refusal note, and unattended runs now require a hard token ceiling (omp budget +Nk!) as the third stop signal (turn/token cap). Verified end-to-end in Bun: refuse-unverified, warn-at-2, block-at-3 (retries 3/10), pass→advance, trivial-advance, old-state load.

Added — 8sync harness global: omp rules machine-wide, one key (Anthropic token-optimized)

• 8sync harness global — applies the omp rule layer MACHINE-WIDE so every project that runs omp gets it without a per-project run: ~/.omp/skills + 00-force-load.md, ~/.omp/agent/APPEND_SYSTEM.md (appended to EVERY omp system prompt), MCP servers (codebase-memory · headroom · serena · zai-vision), recall hook, capabilities snapshot, workflow extension + engine. CWD-independent — never touches the current project.
• Anthropic token-optimizer defaults: compaction.thresholdPercent = 50 written only when unset (never overrides the user), headroom compression for >50-line outputs, and byte-stable APPEND_SYSTEM.md deploys (identical ⇒ skip) so the system prefix stays hot for Anthropic prompt caching. New compaction::ensure_threshold_default helper.
• --sweep [DIR] (default ~/Projects) — stamps the per-project layer into every omp project under DIR (a git repo with agents/ or AGENTS.md/CLAUDE.md — repos not using omp are skipped + reported): mirror skills (additive), inject force-load into AGENTS.md/CLAUDE.md, seed agents/ memory, install the gitleaks hook. Skips node_modules/target/hidden dirs, depth ≤ 4, found repos are not descended into. --pull re-pulls registered skills first.
• Dedup: bare 8sync harness's global block now calls the shared global::global_pass() (crates/cli/src/verbs/harness/global.rs) — one source of truth for the machine-wide layer.
• Overwrite policy made explicit (default = NEVER overwrite, only add what's missing): documented in 8sync harness help (new OVERWRITE POLICY section) + AGENTS.md §8 as a repo-wide invariant. Audited: agents/*.md seed-if-missing, CHANGELOG created once, skills mirror additive (--force only), AGENTS.md sentinel-block only, gitleaks hook only-if-absent, config key-detect. Proven live: hand edits to a mirrored SKILL.md + STATE.md survive a sweep re-run byte-for-byte.

[0.43.0] — 2026-07-05

Added — codegraph canvas capture (?shot=1) + automatic locate for non-vision models

• /codegraph?shot=1 on the dashboard renders ONLY the React-Flow package call graph, full-viewport (no sidebar/cards) — made for 8sync shot 'http://127.0.0.1:8731/codegraph?shot=1': one big, legible graph image (~2k vision tokens) instead of a full-page capture. Everything else on that page stays text via /api/codegraph/overview|search|trace — image for the layout, API text for the details.
• Auto-rule baked into the always-read layer (APPEND_SYSTEM.md + codegraph/image-routing/ locate-anything skills): a non-vision model (GLM-5.2) that needs positions/layout/distribution from an image uses 8sync locate (LocateAnything-3B, on-device ggml, CPU or CUDA) automatically — zai-vision answers what it says, locate answers where it is.

Changed — README + GitHub Pages are English-first (100%)

• README.md fully rewritten in English with the dashboard demo: hero (State page) + Bench / Codegraph / Models / Marketplace screenshots (docs/assets/dashboard-*.png, freshly captured, leak-checked). docs/index.html (Pages) translated to English with a Dashboard gallery.

Changed — bench now DRIVES optimization (breakdown + spine advisory, CLI/API/web)

• BenchMetrics (/api/bench) exposes the upfront breakdown — core_tok, spine_tok, naive_tok — plus spine_advice: a warning set when the memory spine (agents/*.md) eats more than half the upfront budget (the single lever bench exposes; prefix + CORE are fixed by design, the spine grows unbounded between consolidations).
• CLI 8sync harness bench prints the same advisory as a ! line with the concrete fix (trim agents/STATE.md / let 8sync harness auto-archive KNOWLEDGE >200 lines).
• Dashboard Bench page rebuilt: auto-loads on mount (was an empty state until a manual "Run bench" click — the compute is 40 ms and deterministic), upfront breakdown meters (prefix / CORE / spine share of upfront), advisory callout, naive-baseline row. New .meter-val-wide for long meter values. Browser-verified: 0 console errors.
• Warning sweep: removed dead assets::web_asset_iter, LocalModel → pub(crate) (private-interface warning), unused ctx in api_state. cargo build is warning-free.

Added — 8sync harness add-local-model: local GGUF models for omp (Rust runtime)

• New subcommand 8sync harness add-local-model <path> [name] loads a GGUF model through mistral.rs (pure-Rust, memory-safe inference — no C++ llama.cpp) and registers the served OpenAI endpoint as an omp provider local/<name>, so omp routes to on-device models exactly like a cloud one (8sync ai --model local/<name>).
• <path> auto-classifies: an existing *.gguf FILE, a HuggingFace repo id (org/repo), or a *.gguf URL (downloaded to ~/.cache/8sync/models/). GGUF-only this version (validated by magic bytes); other formats auto-detect later. GGUF chosen for speed.
• Runner is auto-installed via the official mistral.rs install.sh (prebuilt per-GPU CUDA or CPU binary — no Rust/CUDA toolkit needed, just the NVIDIA driver for GPU).
• Each model runs as a systemd user service 8sync-llm-<name>.service. For a local .gguf the served command is mistralrs serve --host 127.0.0.1 --no-ui --format gguf -m <dir> -f <file> (mistral.rs --model-id is a directory, not a file); an HF repo goes straight to -m. Port auto-allocated from 8770. A TSV registry (~/.config/8sync/local-models.tsv) is the source of truth; the omp provider lives in a managed sentinel block inside ~/.omp/agent/models.yml that gateway apply preserves across re-deploys. list / rm <name> manage the set.
• The provider sends model id: default upstream (the only alias mistral.rs serves besides the model-dir path); the clean local/<name> handle lives in the name: field (omp fuzzy-matches both), so --model local/<name> still selects it.
• doctor reports the registered local-model count; ~/.omp/capabilities.md lists them so the agent knows they exist.
• E2E-validated on this machine: mistral.rs 0.8.23 (auto-selected the prebuilt cuda131-sm120 RTX-5080 binary) serves a 135M GGUF; add-local-model → systemd unit → /v1/chat/completions returns real text; rm tears down unit + block + registry clean.

Added — 8sync locate: visual grounding (NVIDIA LocateAnything-3B)

• New verb 8sync locate <image> "<prompt>" returns labeled bounding boxes + click-center coordinates from an image — GUI element grounding (click points), OCR/text localization, open-set detection. Grounding, not captioning: complements zai-vision (describe) and the browser tool (act). Pipeline: 8sync shot → locate → click. --annotated out.png, --mode hybrid|slow|fast.
• Runs NVIDIA LocateAnything-3B through mudler/locate-anything.cpp (MIT C++/ggml port, prebuilt GGUFs, no Python). --setup clones + cmake-builds the CLI (CUDA if the toolkit is present, else CPU) and downloads the q8_0 GGUF (~6.3 GB) to ~/.cache/8sync/locate-anything/. Model license: NVIDIA research / non-commercial.
• New always-on specialist skill locate-anything (inject.rs::always_on_rank, rank 8 after image-routing): force-loaded into the AGENTS.md block every session and re-deployed/refreshed on every 8sync harness run (bundled). Body read on trigger (token-lean); an APPEND_SYSTEM pointer surfaces it when exact coordinates are needed.

[0.42.0] — 2026-07-04

Added — modality routing: read STRUCTURE as an image, PRECISE things as text

• 8sync shot is now real (was a no-op stub) — renders any URL / local HTML to PNG via system or omp's bundled Chromium (~/.omp/puppeteer/chrome/…), prints a vision-token estimate (ceil(w/28)*ceil(h/28), 28×28 patch — no stale cap). Also fixes the image-routing skill which already pointed at it.
• New Modality routing directive forced across the harness: APPEND_SYSTEM.md (always-on), the image-routing skill (rewritten with honest economics), and the capabilities.md snapshot, enforced per-turn by --advisor. Rule: vision models render a codegraph / diagram / dashboard / big PDF to ONE image (modality-fit); code / exact config / line-numbered data stay TEXT (cheaper AND lossless).
• Grounded, not hyped: the 10×/90% token cut (DeepSeek-OCR, arXiv 2510.18234) needs a DEDICATED optical encoder — NOT a screenshot to Opus/GLM. Claude bills images per 28×28 patch (pay-per-pixel on Opus 4.7+). Measured on this repo: STATE.md as image = 0.87× (LOSES vs text); the 12k-edge codegraph as image ≈ 25× (structure win). The gate captures exactly that. OCR-Memory pattern (arXiv 2604.26622) documented: image to LOCATE, exact text to READ.
• Dashboard deep-link — web/src/App.tsx now reads ?page=<id> (or /<id>) for the initial page (pageFromUrl() + history.replaceState on nav). Nav was in-memory only, so 8sync shot .../codegraph used to render State; now 8sync shot http://127.0.0.1:8731/?page=codegraph captures the real graph. build.rs already rebuilds the Vite bundle on web/src change, so a plain cargo build re-embeds.

[0.41.0] — 2026-07-03

Added — dashboard Marketplace: discover + install skills & MCP servers

• New Marketplace nav page (Discover group) in 8sync harness web: browse, search, sort (Top by stars/uses · New by recency), and one-click install skills and MCP servers from public registries into the current project.
• MCP sources (4): the official registry (registry.modelcontextprotocol.io, REST API), Smithery (registry.smithery.ai), Glama (glama.ai JSON API), and mcp.so scraped with the pure-Rust scraper crate (HTML DOM via a[href^="/server/"], fetched through curl — no reqwest, Rust-first). 135+ merged/deduped entries; install writes a real ~/.omp/agent/mcp.json stdio (npx/uvx) or remote (http/sse) entry.
• Skills source: GitHub repo search ranked by stars; install shells the existing collection-aware 8sync skill add <url>.
• Catalog cached under .cache/8sync/marketplace/*.json (1h TTL — the MCP registry maintainers ask aggregators to poll infrequently + persist). (crates/cli/src/verbs/harness/marketplace.rs, web.rs, web/src/*)

Added — import buttons across the dashboard (were plumbing-only)

• Skills page: skillAdd/skillUpdate were wired in the API client but had no UI — added an Import toolbar (github URL · gh:owner/repo · path:/abs/dir folder · builtin:name) + Update all.
• MCP page: Install-from-link (npx -y pkg, uvx pkg, or an https remote URL → merged into mcp.json) + a per-server Remove.
• Rules page: Import from a folder or GitHub repo (.md/.mdc, recursively; prefers a rules//.cursor/rules/… subdir), shallow-cloned to a RAII temp dir. Complements the existing inline text-add.
• New routes: /api/marketplace, /api/mcp/{add,remove}, /api/rules/import.
• New dep: scraper 0.20 (pure-Rust html5ever + CSS selectors) for the mcp.so aggregator — HTTP still shells out to curl.

[0.40.0] — 2026-07-03

Changed — advisor default-ON (per-turn rule/tool-use reviewer)

• omp's --advisor (passive per-turn reviewer that checks each turn against the always-on rules — code-intel first, correct MCP tool names, open SKILL.md — and injects corrective notes) is now passed by default by 8sync ai and the 8sync . / resume session. Closes the last anti-forget gap: layers 1-4 (APPEND_SYSTEM rules, recall-hook live context, Mnemopi memory, capabilities catalog) declare + remind but nothing checked whether the last turn actually obeyed — advisor is that live reviewer.
• Token-optimal gating: skipped for trivial-class prompts. Opt out per run with 8sync ai --no-advisor, or globally via advisor = false in ~/.config/8sync/models.toml. New advisor key in ModelConfig (default true).
• Docs: models.toml, APPEND_SYSTEM.md, and 8sync ai --help document the toggle + tradeoff.

Fixed — 8sync doctor self-heals stale profile state

• profile::mark_applied() was append-only — a profile deleted from the repo (e.g. caelestia.toml, removed in e761c31) stayed in ~/.config/8sync/profile.toml's applied list forever, and doctor printed it back as a false positive. New profile::prune_stale() diffs applied against load_all(), drops entries that no longer resolve, rewrites state only if changed. Wired into doctor (warns once, then clean). Verified against a real stale state.

Docs — README + GitHub Pages refreshed to current surface

• README + docs/index.html: TL;DR now leads with the one-liner install → 8sync harness → 8sync harness web; new Dashboard section (with screenshots) documents the CRUD control surface (models/skills/memory/rules/engines/Codegraph); full harness subcommand table (web/gateway/bench/audit/eval/toolstats); added the machine verbs (bt/clean/theme/bg) that were missing. Fixed stale numbers (binary ≈ 5.0 MB, 35 bundled skill). Landing page gains a Dashboard nav link + feature card + two live screenshots (docs/assets/).

[0.39.0] — 2026-07-02

Added — dashboard Codegraph page: visualize the codebase-memory-mcp knowledge graph

• The web dashboard (8sync harness web) had zero visibility into the codegraph/ codebase-memory-mcp engines it lists on the Engines page — search_graph/ trace_path/get_architecture were agent-only. New Codegraph nav item (Runtime group) renders the real graph: package call graph (elk auto-layout, box size ≈ node count, edges = call counts between packages), Leiden cluster cards (de-facto modules, cohesion %, top symbols — the actual architectural seams, not just folders), a BM25 symbol search, and a caller/callee trace subgraph for the selected symbol or a fan-in hotspot.
• Backend: 3 new routes (/api/codegraph/{overview,search,trace}) shell out to codebase-memory-mcp cli <tool> <json> (same binary+slug harness up already indexes against) — no MCP client embedded, stdout-only JSON parsing verified log-noise-free. Honest 404 ("not indexed yet — run harness up") when the project has no graph. (crates/cli/src/verbs/harness/web.rs, web/src/App.tsx, web/src/api.ts)

Fixed — dashboard UI/UX audit (browser-verified against a real project)

• Engines page: the codebase-memory-mcp 0.8.1 tile title rendered one character per line — overflow-wrap: anywhere collapsed the flex item's intrinsic width to zero next to a wide version tag. Fixed .tile-head to give the title flex: 1 1 auto; min-width: 0 and wrap on word boundaries instead of mid-word.
• Version tags were inconsistent/redundant across tools (on codebase-memory-mcp 0.8.1, on headroom, version 0.27.0 — duplicating the already-visible tool name). api_engines now extracts just the semver token.
• Skills page (67 skills, no way to find one): added a filter input + tier dropdown (all/always/on-demand/off) with a live "N of M" count.

Added — 8sync harness gateway — deploy/verify the omp model-gateway

• New subaction: 8sync harness gateway [apply|key <KEY>|verify|status] — deploys ~/.omp/agent/models.yml from a bundled template so the 9router gateway config (provider URL, models, API key, thinking.mode = anthropic-budget-effort) is reproducible by one command instead of hand-editing.
• apply is idempotent (backs up a differing file to models.yml.bak, preserves the existing key on refresh; key from $NINE_ROUTER_KEY or gateway key <KEY>).
• verify pings cc/claude-sonnet-5 through the gateway — the exact path that 400'd before the thinking fix; HTTP 200 = healthy. status masks the key + flags a missing fix.
• Fixes recurring 400 thinking.enabled.budget_tokens: Field required on claude-sonnet-5: omp's default thinking:{type:adaptive} is rejected by the gateway; the bundled template forces enabled + budget_tokens. (crates/cli/src/verbs/harness/gateway.rs, assets/configs/omp/gateway-models.yml)

Added — 18 feynman research skills ported to omp-native (were unusable stubs)

• Audited agents/skills.toml's 20 companion-inc/feynman-sourced skills (submodule-inspected at reference/feynman, then removed). Found 12 were 12-line stubs pointing at feynman's OWN slash-commands (/deepresearch, /lit, /recipe, /audit, /draft, /review, /compare, /watch, /replicate, /jobs, /log, /autoresearch) — those commands only exist in feynman's own pi-coding-agent runtime (prompts/*.md + extensions/research-tools.ts), NOT in omp. Deployed as-is they were completely inert. 2 more (session-search, preview) had the same problem behind a documented fallback. Ported all 14 (deep-research, literature-review, autoresearch, ml-training-recipe, paper-code-audit, paper-writing, research-review, source-comparison, watch, replication, jobs, session-log, session-search, preview) into self-contained assets/skills/<name>/SKILL.md using omp's real tools (task in place of feynman's subagent, web_search/read in place of fetch_content, ask in place of ask_user_question, job/retain where feynman had no equivalent). Also re-bundled 4 genuinely-portable CLI skills (eli5, docker, modal-compute, runpod-compute — only cosmetic "Feynman" naming, no runtime dependency) as builtin: too. alpha-research is kept pointed at the real companion-inc/feynman source since it's a legitimate CLI wrapper (feynman alpha ..., needs @companion-ai/feynman installed via the existing ensure_feynman_cli()). Dropped contributing (feynman-repo-only, no value for su-code users).
• Bug fixed in the process: agents/skills.toml had [peer-review] pointing at feynman, but feynman renamed that skill upstream to research-review — the entry never resolved to anything on disk. Fixed to [research-review].
• Bug fixed in update_skills (crates/cli/src/verbs/skill/update.rs): registering ANY single skill from a git collection repo (e.g. just alpha-research from feynman's 20-skill repo) silently reinstalled EVERY sub-skill in that repo on every 8sync harness/skill update run, regardless of registry membership — contributing kept reappearing after being deliberately dropped from the manifest, because the git-source loop treated filter.is_none() as "install everything found". Fixed: a sub-skill is only (re)installed when the URL/repo was explicitly targeted, the skill name was explicitly filtered, or (bulk run) it already has its own registry key — a collection repo no longer silently grows the registry.
• 18 new on-demand skills registered in assets/skills/00-force-load.md's lookup table (55 on-demand total, up from 37).

Fixed — kitty terminal zoom (ctrl+shift+minus) silently stolen by vsplit binding

• 8sync setup --profile terminal mapped ctrl+shift+minus to launch --location=vsplit for the gsd-style 3-pane layout. That's kitty's DEFAULT font-zoom-out binding (change_font_size all -2.0) — user maps override defaults, so zoom-out silently stopped working with no error. Moved vsplit to ctrl+shift+backslash (unclaimed by any kitty default); ctrl+shift+minus/+equal/+backspace now behave stock. Re-ran 8sync setup --profile terminal to regenerate the live ~/.config/kitty/8sync.conf on this machine — user must reload/reopen kitty (font-zoom maps apply live via kitty @ load-config if remote control is on, no window-recreate needed unlike the earlier decoration fix). (crates/cli/src/verbs/setup.rs:665-668)

Added — ~/.omp/capabilities.md now embeds EXACT MCP/builtin/memory tool catalogs

• Previously the snapshot only said "4 server(s) registered" — no tool names. Agents had to guess, which is exactly how the earlier "codegraph verb" hallucination bug happened (see KNOWLEDGE.md). Now 8sync harness writes the FULL exact tool catalog for every registered MCP server: codebase-memory-mcp (14), headroom (3), serena (23), zai-vision (8) — plus omp's own built-in tools (parsed live from omp --help's "Available Tools" block) and the Mnemopi memory tools (recall/reflect/retain/ memory_edit, listed only when the backend is ON). (crates/cli/src/verbs/skill/deploy.rs::known_mcp_tool_catalog + ensure_omp_capabilities_snapshot rewrite.)
• APPEND_SYSTEM.md RULE #0 now names the 4 connected servers explicitly and points at ~/.omp/capabilities.md as the ground truth for exact names/params — "never guess/invent an MCP tool name". The 8sync-recall.ts hook (injected every before_agent_start + compaction) carries the same pointer so it survives past 50% context.

Fixed — kitty lost its title bar/min-max-close/resize border on KDE (stacking WM)

• 8sync setup --profile terminal unconditionally wrote hide_window_decorations yes into ~/.config/kitty/8sync.conf. That's correct on a tiling Wayland compositor (Hyprland/HyDE — the project's primary target) which draws no chrome and expects clients to hide their own, but on a stacking desktop (KDE/kwin, GNOME/mutter, …) the compositor ALSO does not add server-side decorations for an undecorated client — the window ends up with no title bar, no traffic-light buttons, and no drag-to-resize border at all.
• New env_detect::is_tiling_wm() checks is_hyde() first, then XDG_CURRENT_DESKTOP/DESKTOP_SESSION against known tiling WMs (hyprland/sway/river/wayfire/qtile/i3/bspwm/awesome). render_kitty_conf now only emits hide_window_decorations yes when is_tiling_wm() is true; stacking desktops (verified live on KDE/Plasma/ kwin/Wayland) keep normal window chrome. (crates/cli/src/verbs/setup.rs, crates/cli/src/env_detect.rs)
• Re-running 8sync setup --profile terminal (idempotent) regenerates ~/.config/kitty/8sync.conf with the fix; requires closing/reopening the kitty window (decorations are negotiated at window-creation time, not live-reloadable).

Added — Z.AI vision MCP (zai-vision) bridges GLM-5.2's text-only gap + dedicated skill

• GLM-5.2 (omp's default model) is text-only. 8sync harness now auto-installs @z_ai/mcp-server (npm, via bun add -g) and registers it as the zai-vision omp MCP server, exposing 8 GLM-5V tools (ui_to_artifact, extract_text_from_screenshot, diagnose_error_screenshot, understand_technical_diagram, analyze_data_visualization, ui_diff_check, analyze_image, analyze_video). Auth reuses the SAME Z.AI key already configured for zai/glm-5.2 (pulled via omp token zai, no separate signup). (crates/cli/src/verbs/skill/deploy.rs::ensure_zai_vision_mcp + resolve_zai_api_key; wired into harness auto/harness init; reported by doctor.)
• register_omp_mcp now supports per-server env (only emitted when non-empty, so existing env-less entries stay self-heal-stable).
• Verified end-to-end (not illustrative): real browser screenshots run through the actual zai-mcp-server stdio process via JSON-RPC tools/call. Found and fixed a real gap — 8sync harness now defaults Z_AI_VISION_MODEL=glm-4.6v-flash, the ONLY vision model that works on a stock Z.AI key with no vision resource package (paid models 400 with 1113 insufficient balance; verified against Z.AI's live pricing table).
• New skill zai-vision (assets/skills/zai-vision/SKILL.md, auto-deployed by install_bundled_global) documents the full combination matrix: browser screenshots, 8sync shot/pdf-img/diff-img, codegraph/cbm diagrams, serena, headroom compression, inspect_image fallback, retain/recall, and advisor — plus the real verified tool-call output and a troubleshooting table for Z.AI error codes (1113/1211/1301/1305).
• ~/.omp/capabilities.md — new live snapshot of omp's surface (advisor, thinking, inspect_image, adaptive model roles, retain/recall, registered MCP count, skill count), refreshed every 8sync harness run, surfaced by doctor (ensure_omp_capabilities_snapshot).
• APPEND_SYSTEM.md and image-routing SKILL now point to zai-vision as the mandatory bridge step after routing to "image".

[0.36.0] — 2026-06-30

Added — 8sync bg search: find wallpapers online (no API key) + pick with live preview

• New 8sync bg search <query> sub-action. Searches wallhaven.cc via its public API (no API key needed, SFW, ≥1920×1080) — wallpaper-focused (incl. anime/dark), a good fit for the project's aesthetic without imposing an Unsplash/Pexels registration on the user.
• Interactive (kitty): stages thumbnails, then opens fzf with a live kitten icat preview pane showing each candidate + its wallhaven source link. Enter downloads the full-res image, adds it to the collection, and sets it live; Esc cancels. Only the full image you pick is fetched.
• Non-interactive: prints the result list (id + resolution + source link) for scripting/agents.
• RAII temp cleanup; reuses the existing add+set path. No new Rust deps. (crates/cli/src/verbs/bg.rs)

[0.35.0] — 2026-06-30

Added — 8sync bg: manage the kitty wallpaper at runtime (live swap + inline preview)

• New 8sync bg verb: show | get | set [file] | list | add <url|file>. Brings back the wallpaper control that was removed in the slim-down — now without HyDE overlap (kitty's in-terminal background_image ≠ HyDE's desktop wallpaper).
• Inline preview: bg show renders the current wallpaper in the terminal via kitten icat (kitty graphics protocol — same mechanism omp uses); bg list/bg set (no arg) open an interactive fzf picker with a live kitten icat preview pane → scroll, see each image, Enter to set.
• Live swap: bg set <file> rewrites the background_image line in 8sync.conf + SIGUSR1-reloads kitty (instant, no restart). The choice is recorded in ~/.config/8sync/wallpaper and 8sync setup honors it (re-setup no longer resets your wallpaper). Collection lives in ~/.config/8sync/wallpapers/ (bg add <url> populates it).
• Zero new Rust deps (shell-outs to kitten/fzf/curl) — binary stays lean. (crates/cli/src/verbs/bg.rs)

[0.34.0] — 2026-06-30

Added — 8sync theme: switch kitty palettes live (readable on any wallpaper)

• New 8sync theme verb: list | set <name> | show [name]. Six curated dark palettes (tokyo-night default · catppuccin-mocha · gruvbox-dark · nord · rose-pine · dracula), each a pure color fragment tuned for wallpaper-overlay readability (foreground + bright-black verified at WCAG-AA contrast ≥ 4.5:1 against the theme bg). Switching writes ~/.config/kitty/8sync-theme.conf and SIGUSR1-reloads kitty — instant, no restart, no remote-control socket. hydectl theme still owns Hyprland/UI; this owns kitty (distinct surfaces). (crates/cli/src/verbs/theme.rs)

Fixed — kitty config: readable text + restored allow_remote_control + structure/palette split

• Readability root-cause: deployed 8sync.conf had background_tint 0.55 (image 45% visible → bright wallpaper washed out the foreground). Raised to 0.86 (image subtle, text crisp).
• The glass structure (background_opacity/blur/font/splits/tabs) is now separated from the palette (8sync-theme.conf, swappable); render_kitty_conf no longer emits colors inline.
• Restored allow_remote_control yes in the managed config — it had been dropped in the slim-down, breaking kitty @ live control. (crates/cli/src/verbs/setup.rs)
• 8sync setup --profile terminal now deploys both files (structure + active palette); the active theme is recorded in ~/.config/8sync/kitty-theme and survives re-runs.

[0.33.0] — 2026-06-29

Added — dashboard surfaces the live /auto engine run (real, not demo)

• New /api/engine reads the real gsd-pi state machine the engine drives at <root>/.cache/8sync/engine/state.json; the Engines page renders a live board — goal · progress bar · slice/task tree with ✓/▸/○/✗ status + retries · current task (4 s refresh, read-only mirror of the terminal board). Closes the gap where the dashboard showed the workflow editor + engine binaries but never the actual /auto run. {active:false} when none. Browser-verified, 0 console errors. (crates/cli/src/verbs/harness/web.rs, web/src/{api.ts,App.tsx})

Added — AFFiNE in the alexdev profile

• affine-bin — official prebuilt of the open-source Community Edition (AGPL/custom: free, self-hostable, no cloud lock-in). The from-source affine AUR pkg fails upstream (electron-packager zip step), so the prebuilt is used. (assets/profiles/alexdev.toml)

Changed — always-on directives also prime recall/retain + browser

• APPEND_SYSTEM.md (every system prompt, never compacted) + the recall hook now explicitly prime recall/reflect before · retain durable facts after (Mnemopi) and browser to verify any web/UI change for real — on top of RULE #0 (code-intel MCPs) + skill ref-paths. Stays terse by design (the system prompt isn't headroom-compressed; headroom is for tool OUTPUTS). (assets/configs/omp/APPEND_SYSTEM.md, assets/hooks/8sync-recall.ts)
• kitty tab bar moved to the bottom (tab_bar_edge bottom) — easier tab switching. (setup.rs renderer)

Fixed — 8sync harness up now redeploys the recall hook

• harness up refreshed APPEND_SYSTEM/engine/workflow but not the recall hook (only init/bare-harness did), so hook changes never reached existing machines via up. Now it does. (crates/cli/src/verbs/harness/up.rs)

[0.32.1] — 2026-06-29

Fixed — 8sync harness auto-installs the token-optimization MCPs (no startup error)

• headroom (and serena) were registered in ~/.omp/agent/mcp.json even when their binary wasn't installed — so omp failed at startup with Executable not found in $PATH: "headroom". Now 8sync harness bootstraps uv (Astral, user-level curl install — no sudo), installs headroom-ai[mcp] through it, and only registers an MCP whose executable actually exists — a still-missing tool has its stale entry purged so omp never errors at startup. uv also ships the uvx serena needs, so both engines come up from one 8sync harness, no manual steps. (crates/cli/src/verbs/skill/deploy.rs)

[0.32.0] — 2026-06-29

Performance — binary back under control (offsets bundled rusqlite)

• Enabled rust-embed's compression feature (transparent include-flate decompress on .data — both the assets/ skills tree and the embedded web/dist FE shrink) and set the release profile to opt-level = "z". Roughly halves the binary, offsetting the bundled rusqlite (toolstats) + impeccable + the Vite FE. (crates/cli/Cargo.toml, Cargo.toml)

Fixed — wallpaper self-heal (no more kitty "render to RGB: EOF")

• setup::deploy_wallpaper trusted exists(), so a transient/blocked download left a 0-byte wallpaper.png kitty can't render (blank background) — and the early exists() return meant it never re-tried. Now validates the file (size>0 + PNG/JPEG/WEBP magic via is_valid_image), adds a Mozilla/5.0 UA + --retry 2, and purges a corrupt file so a re-run re-downloads. (crates/cli/src/verbs/setup.rs)

[0.31.1] — 2026-06-29

Changed — toolstats now reports the actionable ratio

• The headline is now optimizer vs raw-search of code-lookup calls only (optimizer = codegraph / cbm / serena; raw-search = grep / search / find / glob) — instead of "% of all calls", which was misleading (most calls are edit / bash / read-before-edit, not lookups). read is shown separately (often legitimate, not shamed) and headroom is labelled background/auto-compress (not an agent-called tool). Measured: su-code optimizer 34% of lookups, agentic-cloudgo 25% — vs the old "2% of all calls" framing. The DB is rebuilt from current sessions each run (re-categorizes).

[0.31.0] — 2026-06-29

Added — 8sync harness toolstats (SQLite tool-call tracker)

• New verb that tracks how the agent actually uses tools, parsed from omp's own session JSONL, into a per-project SQLite DB (.cache/8sync/toolstats.db, gitignored). Reports the optimizer (codegraph / codebase-memory-mcp / serena / headroom) vs fallback (grep / read / search / find / glob) call ratio + per-tool failures, so you can see whether the STEP-0 token-optimization stack is being used and catch failing calls (e.g. a dead MCP server).
• Idempotent (keyed on session+seq → re-run only adds new calls); inspectable with any SQLite tool.
• Motivation: across this machine's 68 omp sessions / 28k calls, the optimizer stack was 1.1% of calls (serena/headroom 0) vs 35% raw fallback — toolstats makes that visible per project.

[0.30.0] — 2026-06-29

Changed — default 8sync setup is AI-core only

• Stage A now installs only the AI coding harness: omp, codegraph, MCP servers + skills, github-cli, paru, PATH bootstrap, configs. The terminal/editor polish (kitty glass theme + helix
  • JetBrains Nerd font + wallpaper) is no longer installed by default — a fresh 8sync setup is pure AI now.
• New opt-in terminal stack: 8sync setup --profile terminal (also offered in the y/N menu and applied by --full). docker moved out of the terminal stack — it lives in dev-stack.
• doctor reports the terminal stack (kitty/helix/docker) as advisory/opt-in — no longer warns when it's absent.
• Personal/hardware profiles (vietnamese/unikey, warp, hardware-*, displaylink, …) stay opt-in as before. Nothing personal is installed unless you pick it.

[0.29.3] — 2026-06-29

Fixed — serena MCP "Transport closed"

• serena's executable was renamed. The registered command uvx … serena-mcp-server no longer exists (serena now ships serena with a start-mcp-server subcommand), so the MCP process exited instantly → omp reported serena: Transport closed. Now registers uvx … serena start-mcp-server --context claude-code (ide-assistant was also deprecated). Verified it launches (22 tools exposed, no error).
• MCP registration now self-heals. register_omp_mcp previously skipped any server already in mcp.json, so a stale entry never got corrected. It now updates in place when the command/args changed, and 8sync harness up also refreshes MCP servers (was init/bare-harness only) — so 8sync harness up fixes the stale serena entry on existing machines.

[0.29.2] — 2026-06-29

Fixed — Context page is now correct for ALL models (not just GLM)

• Per-model context window. /api/context hardcoded a 1,000,000-token window, so models with a smaller real window (e.g. claude-haiku-4-5 200k, glm-4.x 131–205k) showed an artificially low % and looked like they never hit the compaction threshold — while 1M models (glm-5.2, claude-opus) looked fine. Now the window is parsed per active model from omp models (cached via LazyLock), so the %, threshold marker, and "will compact" are accurate for every model. Falls back to an explicit assumed estimate only when the model isn't in omp's catalog.
• Honest compaction copy. omp's threshold compaction is turn-triggered (fires after a completed turn / safe mid-turn point once usage exceeds thresholdPercent of the real window) — not a hard cap, so a paused/ended session legitimately sits above the line until resumed. The page now says "compacts on next turn", flags idle/ended sessions (stale), surfaces the explanation, and only shows the "assumed window" badge when the window is truly unknown.
• build.rs shipped stale FE. It rebuilt the Vite bundle only when web/dist was missing and watched only web/dist, so edits to web/src were silently embedded stale. Now it rebuilds when any FE source is newer than dist and emits rerun-if-changed for web/src + configs.

[0.29.1] — 2026-06-29

Fixed — dashboard project switcher

• Switching projects now actually switches the data. activate only wrote an advisory web-session.json; every handler still read detect_current_project_root() (the launch cwd), so pages never changed. Now apply_active_project chdir's into the activated project (at startup
  • on activate) so all cwd-based handlers (State/Context/Skills/Memory/Rules/Submodules/Workflow) resolve to it. Verified in-browser: switch → State path + content + trigger label all update.
• /api/projects cleanup — dedup by resolved path; drop junk slugs (no session file / non-dir); widened the green-dot "active" window to 2h + added a current flag for the project being viewed (a project open but idle >30 min now shows correctly).

[0.29.0] — 2026-06-29

Added — 8sync harness web dashboard: full redesign + Models/Projects

• Models page (/api/models get+post) — view/edit the adaptive model routing live: [roles] (default/plan/smol/slow) + [tasks] (plan/review/debug/code/trivial), inline selects write ~/.config/8sync/models.toml immediately. Surfaces the routing philosophy: thinking → Opus (plan/review/debug/vision), mechanical → GLM (code/edit/default/trivial).
• Project switcher (/api/projects) — sidebar-top dropdown lists every omp project with a green (active) / gray (off) status dot; activate + refresh without cd.
• Workflow templates (/api/workflows/templates) — 3 starter graphs (research→plan→build, review, qa) loadable in the editor.
• Markdown rendering — new XSS-safe renderer (web/src/markdown.tsx); State/Memory/Context now render headings, lists, GFM checkboxes, code, emphasis (was raw text).

Fixed

• serena engine showed "off" wrongly — detection now checks mcpServers.serena in ~/.omp/agent/mcp.json + uvx/uv on PATH (serena is uvx-launched, no PATH binary), not which serena. Reports {present,registered,runner}.
• Context window honesty — /api/context now exposes assumed:true, windowTok, thresholdPct, willCompact; the FE labels the 1M window as an estimate (no false precision).
• Workflow canvas — react-flow viewport fixed (was a tiny broken box) to a usable 560px panel with fit/zoom.

Changed

• Dashboard FE redesigned to a product-register design system (impeccable): solid surfaces, violet brand preserved, legible chips, grouped nav, dark + light. 14 pages, zero console errors.

[0.28.0] — 2026-06-29

Changed — ONE command: /auto (retired /gs)

• Unified the autonomous entry to a single /auto — removed /gs (command + skill + ensure_gs_command + all wiring + help/force-load refs). /auto (8sync-engine) is the only automation path; deploy::cleanup_legacy_gs removes the retired /gs from machines that had it.
• /auto upgraded to gsd-pi-grade (grounded in reference/gsd-pi): research INTEGRATED into planning (codegraph/cbm/serena scout + feynman/deep-research), fresh scoped context per task, mechanical verify-gate, hard Closeout (full suite + QA/UAT in a browser + independent re-review vs DoD + doc-hygiene), and a context-budget/handoff rule.
• Verify UI for real: web → browser at the dev URL; Tauri v2 / WRY-WebKit desktop → run with its web-inspector/remote-debug port + point the same browser tool at it.
• harness up now deploys the full harness (APPEND_SYSTEM + /auto engine + workflow), matching bare 8sync harness.

Added

• 8sync harness model — view/edit ~/.config/8sync/models.toml (single model-routing source): bare shows roles+tasks; 8sync harness model <key> <value> sets one (e.g. review opus). omp resolves names fuzzily + falls back to an authenticated model.

[0.27.0] — 2026-06-29

Added — adaptive model routing

• Per-prompt model selection (no more single fixed model). assets/configs/models.toml (deployed → ~/.config/8sync/models.toml) maps [roles] default/plan/smol/slow + [tasks] plan/review/debug/code/trivial → models (defaults: codex main · glm plan · opus review/debug · haiku smol). New crate::models classifies the prompt heuristically and passes omp --model/--plan/--smol/--slow (omp resolves fuzzy: "glm","codex","opus"). Wired into 8sync ai (+--model override) and 8sync . (resume flags). omp owns the catalog — 8sync only steers.

Added — gsd-pi-style automation engine (on omp core, no patch)

• 8sync-engine omp extension (~/.omp/agent/extensions/ + project) — durable slice/task state machine (.cache/8sync/engine/state.json) + model-callable tools engine_plan/status/ next/verify/advance/worktree. Code-enforced verify-with-retry gate (counts attempts, BLOCKs at maxRetries — the agent can't skip it) and git worktree open/squash-merge/remove.
• /auto command orchestrates the engine to run a goal to DONE (right-sized, token-lean). Closes the gsd-pi gaps (verify/worktree as CODE, not prose). /gs stays a lighter skill.

Added — context engineering (always-read + serena + tunable compaction)

• APPEND_SYSTEM.md deployed to ~/.omp/agent/ — RULE #0 (code-intel before grep/CRUD) + always-on skill manifest (name·purpose·ref-path) injected into EVERY system prompt (never compacts away) → fixes "skills/rules defined but ignored past 50%". Recall hook rewritten to the LIVE half (skill index + STATE Current/Next).
• serena MCP registered (ensure_serena_mcp, via uvx) — symbol-level code intel, prioritized over native search/file-CRUD. Surfaced on the dashboard Engines page + force-load RULE #0.
• 8sync harness compaction [pct] — view/set compaction.thresholdPercent (auto-clean at 50%).

Added — terminal: kitty glass + helix + docker (Stage A defaults)

• 8sync setup now installs kitty + helix + docker + docker-compose + JetBrains Nerd font and deploys a glass kitty theme (~/.config/kitty/8sync.conf, included from kitty.conf — no clobber): transparency + blur + wallpaper + 3-pane split keymaps + violet accent. Wallpaper deployed to ~/.config/8sync/wallpaper.png from assets/wallpapers/default.png (bundled) or [ui].wallpaper_url. Transparent helix config (base16_transparent) deployed if absent. 8sync doctor checks hx/kitty/docker.

Changed / Fixed

• Web dashboard redesigned to a dark glassmorphism / Hyprland aesthetic (translucent blurred panels, layered gradient, icon sidebar, refined type scale, light-mode + a11y fallbacks). 13 pages, react-flow workflow editor intact. Browser-verified: all pages render, zero console errors. (web/src/{styles.css,App.tsx,icons.tsx})
• build.rs now builds the FE with bun → pnpm → npm (first available); on no toolchain it embeds a styled, instructive fallback page instead of a bare one-line stub.
• Helix command fixed to hx — dropped the dead "helix" fallback (Arch ships /usr/bin/hx, no helix binary); note/find now share one pick_editor() ($VISUAL→$EDITOR→hx→vi).

[0.26.0] — 2026-06-27

Added (dashboard FE enhancement)

• Context tracker page — live omp session token usage for the current project (reads the session JSONL's contextSnapshot.promptTokens, auto-refresh 4s). Gauge + 50% threshold marker + compaction-observed badge (detects the token drop = empirical proof auto-compact fired). /api/context. Verified real: tracks this very session 440k→447k live; detected last compact at 575k.
• MCP servers page — visualize ~/.omp/agent/mcp.json (name/command/args/present). /api/mcp.
• Rules CRUD page — list/add/delete omp rule files (.omp/rules/* project + ~/.omp/agent/rules/* global), add from pasted content (link/file/folder source). /api/rules (+add/delete).
• Dashboard now 12 pages (State · Context · Skills · Memory · Engines · Bench · Readiness · Workspaces · Team · Submodules · MCP · Rules). Anti-slop per impeccable (no gradient text / glassmorphism / over-round; verb+object buttons). Browser-tested (Chromium): all pages render real data, Context live-tracking + Rules add-end-to-end verified.

[0.25.0] — 2026-06-27

Added (Phase A — anti-forget)

• Anti-forget: compaction@50% + idle + recall hook. 8sync harness giờ ensure ~/.omp/agent/config.yml có compaction.thresholdPercent: 50 + idleEnabled: true (snapcompact vẫn là default), và deploy ~/.omp/hooks/pre/8sync-recall.ts — hook inject lean ref bundle (skill index + live STATE) tại mỗi before_agent_start + vào mọi compaction summary → agent giữ index skills/rules/workflow qua 50% context & sau compact. 8sync doctor báo "anti-forget ON/OFF". Key-based config detection (robust khi omp rewrite/strip comments config.yml — bỏ sentinel strategy). Verified: omp 16.2.1 load OK.

Added (Phase B — 8sync harness web)

• 8sync harness web — dashboard Vite+React (embedded qua rust-embed) do axum serve tại http://127.0.0.1:8731 (--port, --no-open). API: /api/state · /api/skills (list + toggle tier qua agents/skills.toml) · /api/memory/:file (get/set, allowlist) · /api/engines (codegraph/cbm/headroom/serena + mnemopi) · /api/bench · /api/eval. Refactor B5: tách bench_metrics()/eval_project_data() (home: &Path) cho cả CLI lẫn web reuse. Build.rs tự build FE qua pnpm khi thiếu + stub fallback. Deps: axum 0.7 + tokio + tower-http (override có chủ đích rule "tránh tokio" trong AGENTS.md §8, gated harness web). Verified real: 6 endpoint trả data sống (eval 96% 28/29, bench A1 PASS).

Added (Phase C — full manage)

• Workspace + team + submodule + skill install qua dashboard: /api/workspaces (list omp profiles + project + activate ghi web-session.json), /api/team (subagent roster 8 loại + readiness reuse eval_project_data), /api/submodules (parse .gitmodules + add/pull/remove qua git shell-out), /api/skills/add|update (self-shell-out 8sync skill). FE: 3 page mới (Workspaces, Team, Submodules) + nav. Verified real: workspaces/team/submodules trả data, skill add validate spec.

[0.24.0] — 2026-06-25

Added

• 8sync harness eval --project — agent-team readiness scorecard (% per vai). Deterministic + offline: chấm capability coverage trên repo hiện tại theo dev · qa/testing · research · ba/po · fe · be · docs · memory/learn · token-opt (engine on PATH + skill present + memory spine + stack signals). Honest READINESS (team được trang bị gì Ở ĐÂY), KHÔNG phải output-quality (đó là harness eval loop probe model+network). Run thật: su-code 89%, 8syncdev-pro-v2 79%.
• token-bench skill (bundled) — chứng minh token-saving của code-intel trên repo thật. scripts/token_bench.py (uv/PEP723, stdlib-only): mỗi symbol thật so codegraph-query+slice vs grep+read-whole-file, có def-kind correctness gate. Đo trên codebase lớn thật: 8syncdev-pro-v2 −96.6%, gsd-pi −78.6% (range 54–98%; symbol dùng rộng / file lớn → 95–98%), correctness gsd-pi 10/10. Cần ANSI-strip (codegraph tô màu cả khi pipe). Bundled qua deploy.rs (16 skills).
• 6 reference submodule (inspect/track upstream; deinit, content gitignored): gstack · gsd-pi · agent-reach · addyosmani/agent-skills · DietrichGebert/ponytail · DeusData/codebase-memory-mcp.
• outputs/agent-team-workflow-automation-plan.md — operating plan để vận hành su-code như một agent team: map sprint 23-specialist của gstack + loop slice/auto/worktree của gsd-pi lên /gs + skills + subagents, kèm UI/UX Design Lane riêng (impeccable + Clouds F + Lighthouse 4-tiêu chí quality gate).
• 8sync help dẫn đầu bằng AI TEAM (harness + /gs). Cheatsheet (8sync / 8sync help / 8sync flow) trước đây mở đầu bằng install + vibe loop, không hề nhắc 8sync harness (all-in-one) lẫn /gs (team lead) — giờ là section ĐẦU TIÊN. Fix dòng stale: 8sync skill sync (đổi thành skill update; regen là 8sync harness) và 8sync up ("binary + omp" → chỉ 8sync; omp qua omp update).
• /gs <goal> scope handshake (chỉ assisted). Goal medium+/mơ hồ không dive thẳng: GS ground (codegraph/cbm) rồi đề xuất 2–4 phương án cụ thể (scope · team size + roles/skills · effort · tradeoff, rút từ bench senior: impeccable+Lighthouse / senior-frontend / code-review-and-quality / senior-security / performance-optimization) kèm recommended default + 2–4 câu hỏi sắc qua AskUserQuestion — một vòng rồi chạy. auto vẫn unattended (no questions); trivial/small bỏ qua handshake. (assets/commands/gs.md §1b.)
• 8sync harness eval báo % (eval.rs:114) — score: N/M passed (X%). 3/3 = 100%.
• outputs/omp-customization-memory-platform-research.md — research grounded từ omp docs: cơ chế nhớ THẬT = Mnemopi memory + cbm + spine, dùng model API (không local — máy yếu vẫn chạy), thay cho ngộ nhận GGUF/fine-tune (không khả thi); custom command/workflow trên ĐÚNG base omp (.omp/commands native, update không conflict); submodule auto-pull là ngộ nhận (skill đã auto-latest qua manifest+harness up --pull); agent-reach = capability layer (đọc internet), thêm làm skill.
• Mnemopi memory wired vào 8sync harness (deploy.rs::ensure_mnemopi_memory) — harness/init bật memory.backend: mnemopi (+ scoping: per-project-tagged, llmMode: smol API, noEmbeddings: true FTS, polyphonicRecall) trong ~/.omp/agent/config.yml (idempotent sentinel-block, KHÔNG clobber memory: của user). 0 local model → máy yếu chạy. 8sync doctor báo memory ON/OFF (doctor.rs). Verified: omp 16.1.20 load config OK, doctor "mnemopi memory ON". Tradeoff: recall inject token/phiên (user đã chốt bật).
• 5 reference repo = git submodule (reference/, content gitignored, deinit mặc định): gstack · gsd-pi · agent-reach · addyosmani/agent-skills · DietrichGebert/ponytail. Đăng ký để inspect/track upstream (git submodule update --init --remote reference/<name> để pull-latest đọc khi cần). Submodule = nguồn-tra-cứu; deploy auto-latest cho skill LIVE vẫn qua manifest + harness up --pull.

Changed

• Declutter skill registry — bỏ pack research companion-inc/feynman (20 skill on-demand). Manifest (agents/skills.toml committed + ~/.config/8sync/skills.toml machine-local) đăng ký 20 skill research/ML/academia (paper-writing, ml-training-recipe, literature-review, runpod/modal-compute, peer-review, jobs, eli5, …) — sai domain cho một coding harness + là prefix noise inject vào AGENTS.md mỗi phiên. Cắt cả 20 (collection re-pull là all-or-nothing theo URL — update.rs:49, giữ 1 cái là re-clone cả pack). Kết quả: on-demand 55 → 35, force-load prefix ~1998 → ~1717 tok, deferred −5k tok (8sync harness bench), A1 stable-prefix PASS. Giữ nguyên addyosmani coding-eng + design payload (impeccable/taste/assp) + bundled always-on.

[0.23.0] — 2026-06-24

Added

• 8sync harness eval — loop quality probe. Runs a fixed task-suite through omp non-interactively (omp -p --no-session --auto-approve) and scores each task with a deterministic verify.sh (the verifier OWNS the assertion, so the agent can't game the check). Three bundled fixtures: fix-failing-test (correct a wrong impl until cargo test is green), add-fn-with-test (implement slugify; the verifier appends the assertions), locate-symbol (answer path:line for a symbol). Writes a JSON scorecard + a --baseline reference into the gitignored .cache/8sync/eval/; later runs print the pass-count delta vs baseline. Model + network, non-deterministic — a periodic quality SIGNAL, not a CI gate. Verified end-to-end: 3/3 on this machine.

Changed

• /gs L3 worktree isolation is now concrete. The guardrail named "git-worktree isolation" with no mechanism; it now prescribes the exact flow — git worktree add .gs/wt/<slug> -b gs/<slug>, implement
  • verify + commit on that branch inside it, then git worktree remove (merge/PR only if asked); never edit main's working tree directly. (.gs/ is gitignored, v0.22.0.)

[0.22.0] — 2026-06-24

Added

• 8sync harness audit — code-backed doc-hygiene (was prompt-only advice with zero code behind it). Scans committed docs (AGENTS.md/CLAUDE.md/README/CHANGELOG + agents/*.md) for stale path references (repo-relative paths in docs that no longer exist), oversized docs (>400 lines / >120-line force-load block), and 30-day churn hotspots (history-awareness — docs near churned code are likeliest stale). Report-only: never auto-deletes (heuristic; illustrative paths flagged "review before editing"). Skips absolute / ~-rooted / URL paths so the harness's own machine-generated refs don't false-positive. 8sync doctor surfaces a one-line summary; /gs + the gs skill doc-hygiene step now run the audit instead of eyeballing.
• 8sync doctor AI-engine health check — verifies the token-optimization stack is installed AND registered with omp ("luôn xài"): codegraph (local index) · codebase-memory-mcp (semantic graph) · headroom (output compression). A missing or unregistered engine silently defeats STEP 0 token discipline, so doctor now flags it with the one-command fix (8sync harness).

Fixed

• codegraph STEP 0 verbs were wrong in the force-load prefix, the subfolder-index block, and the KNOWLEDGE breadcrumb: they taught codegraph search/deps/defs, none of which exist. Corrected to the real CLI surface codegraph query/callers/callees/impact (verified against codegraph 0.9.6) so the agent's first explore call doesn't error out.
• Duplicate always-on skill in the force-load list. A stale/external karpathy dir alongside the canonical karpathy-guidelines (identical frontmatter name) made the skill appear twice — once in CORE, once in on-demand. build_force_load now dedups by frontmatter name, keeping the higher-ranked dir, so each logical skill is listed exactly once. Future-proof against any dir/name collision.
• impeccable setup scripts couldn't run under 8sync's layout. The bundled design skill referenced .agents/skills/impeccable/scripts/*.mjs (leading dot) but 8sync mirrors skills to agents/skills/ (no dot). Fixed 28 references across SKILL.md + 4 reference docs → agents/skills/.

Changed

• Managed .gitignore block now ignores .gs/ (per-run worktree + /gs stop marker — machine-local).

[0.21.0] — 2026-06-24

Changed

• /gs redesigned to right-size effort (fixes the post-/gs quality regression). Eval + deep-research (outputs/gs-eval-improve-research-brief.md) found the drop was process over-engineering, not tokens (harness bench: ~8.5k upfront, 79% saved, KV-cache stable): the 93-line command forced a team + full Closeout on every task and auto "never asked".
  • Right-size first — classify trivial/small → solo (no team, no Closeout) · medium → solo + one verifier · large → full loop + roles + Closeout. A team is the exception you justify (Cognition/Anthropic: single-agent default; multi-agent only when it clears the bar).
  • Solo-by-default delegation — subagents only for parallel-independent / context-isolation / specialization; scoped objective + summary return (never free-form, never inline transcript).
  • Autonomy confidence-gated — strong auto, but a high-stakes hard-to-undo low-confidence call is now a blocker (Anthropic 2026: "agents learning when to ask"); prefer reversible, never compound.
  • Doc-hygiene step — detect stale paths / junk / superseded docs → fix or delete (no addition without the matching deletion); keep docs lean. Stale docs poison agent context.
  • Codebase-history — git log/blame + DECISIONS + cbm detect_changes before load-bearing edits.
  • Leaner command — assets/commands/gs.md 93 → 56 lines (lower constraint density → better instruction-following); full protocol stays in the gs skill (progressive disclosure).

[0.20.1] — 2026-06-23

Fixed

• /gs auto actually runs unattended now. Added an Autonomy contract to the /gs command + gs skill: in auto/L3 the agent NEVER calls ask or stops on ambiguity — it resolves unknowns by research (codegraph/cbm → agents/*/PLAYBOOKS → skills → web_search/autoresearch/deep-research), picks the boring/reversible option, logs it under a new ## Assumptions section in agents/STATE.md, and proceeds. "Blocker" is tightened to ONLY missing credential / external approval / destructive- irreversible action; design choices, naming and scope are no longer stops. Note: a slash command cannot bypass omp's approval gate — keep tools.approvalMode: yolo (default) for true unattended runs.
• /gs argument hint. Added argument-hint frontmatter and front-loaded the description with [auto | <goal> | status | next | stop] so the autocomplete dropdown shows the modes when you type /gs (omp renders per-argument hints only for built-ins; the description is what surfaces for file-based commands).
• QA + test are now first-class gates in /gs. Per-slice verify-gate explicitly runs tests + a QA pass and forbids skipping/weakening tests; added a mandatory Closeout step — full test suite + end-to-end QA + independent re-review against the Definition-of-Done + a handoff summary — that must pass before the loop reports "done". Never hands back unverified work.

Added

• Reference submodules reference/gstack + reference/gsd-pi (git submodules, MIT) for studying the engineering-team + autonomous-loop patterns that informed /gs. Pointers are committed (reproducible) but the working trees are deinitialized by default so they never bloat the codegraph/cbm index (codegraph honors no exclude/ignore — populating them ballooned the index to ~3k files / 110 MB). Study on demand: git submodule update --init reference/<name>; re-shrink with git submodule deinit -f reference/<name>. reference/ is also gitignored as a cbm-index guard.

[0.20.0] — 2026-06-23

Added

• /gs — one-command autonomous engineering-team loop (omp slash command). /gs <goal> plans + runs, bare /gs resumes, /gs auto runs unattended (L3), /gs status|next|stop. Drives the loop off agents/STATE.md: plan → delegate to specialist roles (task subagents / gstack roles if installed) → verify-gate → commit → record (KNOWLEDGE/PLAYBOOKS) → advance until Definition-of-Done or a blocker. Token-lean (codegraph + codebase-memory-mcp + headroom mandatory) and guardrailed (verify-gate before commit, worktree isolation + no push/PR at L3, hard-stop via /gs stop). Modeled on gsd-pi /gsd auto.
• Deploy + team-sharing. 8sync harness/init/up write it to ~/.omp/agent/commands/gs.md (global) and <repo>/.omp/commands/gs.md (committed → whole team gets /gs). New on-demand gs skill documents the protocol; 8sync harness up --timer runs it 24/7.

[0.19.0] — 2026-06-23

Changed

• Loop engineering v2 — Phase A (token & stable-prefix discipline).
  • Force-load block (inject.rs) + master 00-force-load.md split always-on skills into CORE (codegraph · karpathy · ponytail · 8sync-cli — đọc body upfront) và SPECIALIST (assp · impeccable · taste · image-routing — biết khả năng, đọc body khi task khớp / progressive disclosure). Thu nhỏ tập đọc-ngay; impeccable vẫn bắt buộc ngay khi có việc UI/design.
  • headroom_compress nâng từ khuyến nghị → bắt buộc cho output > ~50 dòng (STEP 0 + invariants).
  • KNOWLEDGE breadcrumb (memory.rs) bỏ timestamp epoch: volatile → byte-stable giữa các lần harness (thân thiện KV-cache, hết git churn). now_stamp() vẫn dùng cho tên file archive.
  • Plan + provenance: outputs/harness-loop-engineering-v2-plan.md.
• Loop engineering v2 — Phase B (live memory & recitation).
  • agents/STATE.md seeded as a structured live plan (Goal · DoD · Checklist · Current · Next · Open-questions · Handoff) — recitation anchor (Manus todo.md pattern): read at session start, rewritten at each phase boundary to keep the plan in recent context.
  • Loop section (00-force-load.md) + generated block (inject.rs) gain recitation, compaction (near-limit → structured handoff to STATE + lessons to KNOWLEDGE → reinit, with headroom_compress as summarizer), and budget-awareness rules.
  • harness bench now counts the memory spine in the upfront budget (more honest accounting).
• Loop engineering v2 — Phase C (maker/checker + Reflexion).
  • Loop section + generated block: task implementer ↔ independent verifier (build/test in its own context, verify-gate before commit), explicit objective/boundaries/output per subagent, share-full-trace for dependent work, parallel only when subtasks are independent.
  • Reflexion failure-capture: a failed verify writes a failure: entry to KNOWLEDGE (symptom
    • cause + fix); recent failures are read at session start to avoid repeating them.
• Loop engineering v2 — Phase D (procedural memory / playbooks).
  • agents/PLAYBOOKS.md seeded (Voyager-style skill library): validated multi-step procedures distilled into reusable runbooks indexed by a When: line — retrieved + adapted, not re-derived.
  • Memory tiering: KNOWLEDGE = verbal lessons · PLAYBOOKS = verified procedures · DECISIONS = ADR. harness bench now counts PLAYBOOKS in the spine (6 files).
• Loop engineering v2 — Phase E (phased autonomy + guardrails).
  • L1 report · L2 assisted · L3 unattended defined, with guardrails (verify-gate before commit, gitleaks, commit scoped to agents/+docs, no auto push/PR at L3). harness up --timer per-tick job documented (read STATE → Next → verify → update spine → optional commit).

Added

• 8sync harness bench — deterministic loop-engineering benchmark (no model calls): upfront context budget (force-load prefix + CORE skill bodies) vs deferred (SPECIALIST + on-demand), the A2 progressive-disclosure saving, and an A1 KV-cache stable-prefix gate. Refactors a shared inject::build_force_load() (single source of truth for inject + bench). Baseline on this repo: upfront ~5.5k tok vs naive ~37.9k tok → 85% upfront cut; A1 PASS.

[0.18.1] — 2026-06-23

Fixed

• 8sync harness init now pulls registered manifest skills — init calls skill update against agents/skills.toml (git collections like feynman: deep-research, autoresearch, …) before mirroring, making it a true superset of bare 8sync harness. Previously init only deployed the bundled skills + 2 hardcoded external packs (ponytail, addyosmani), so manifest-only skills never reached agents/skills/ via init — only bare 8sync harness / up --pull did.

[0.18.0] — 2026-06-21

Added

• Headroom context-compression wired as an omp MCP — 8sync harness/init auto-installs headroom-ai[mcp] (uv → pipx → pip fallback) and registers it in ~/.omp/agent/mcp.json (headroom mcp serve, stdio). Tools headroom_compress / headroom_retrieve / headroom_stats compress long tool outputs / logs / diffs 60–95% before they reach the model. Force-injected into STEP 0 + 00-force-load.md. Researched alongside PixelRAG + LocateAnything3D — skipped: PixelRAG (screenshot-RAG) overlaps 8sync shot/read/browser; LocateAnything3D is a 3D-vision model (out of scope for a coding harness).

[0.17.1] — 2026-06-21

Fixed

• Skills now propagate to other machines. 8sync harness / skill update write a committed project manifest agents/skills.toml (mirroring the machine-local registry) and read it back on any machine — so a fresh clone re-pulls the exact same skills. Previously only the machine-local ~/.config/8sync/skills.toml recorded skill add-ed sources, so custom skills (e.g. git collections like feynman) never reached a second machine via harness — only the 15 binary-embedded skills + 2 hardcoded external packs did. (agents/skills.toml is a file, so it travels even when the agents/skills/ directory is gitignored.)

[0.17.0] — 2026-06-21

Added

• codebase-memory-mcp = first-class code-intelligence engine — 8sync harness/init auto-installs the binary (upstream installer, binary-only), sets auto_index true, and registers it as an omp MCP server in ~/.omp/agent/mcp.json (idempotent, preserves other servers). harness/up index the repo. Mirrors ensure_codegraph — zero manual MCP config.
• Code intelligence FIRST (STEP 0) — the injected force-load block + 00-force-load.md mandate codegraph + codebase-memory-mcp BEFORE grep/read for all code exploration (~99% token saving); raw Read only for read-before-edit.
• Loop-engineering principles (Addy Osmani / Boris Cherny) in 00-force-load.md: STATE/KNOWLEDGE spine, maker/checker via task sub-agents, verify-gate, phased L1→L3 autonomy via harness up --timer.

[0.16.0] — 2026-06-21

Added

• 8sync harness (bare) = ONE command — idempotent driver that makes a project agent-ready in a single pass: deploy/update skills + mirror (additive) + inject force-load + seed memory & gitleaks hook + consolidate learnings + re-index codegraph. harness init = explicit full bootstrap (progress UI); harness up = light refresh; harness up --timer 30m = background loop.
• Additive skill mirror + --force — harness/harness init never clobber an already-vendored (possibly edited) agents/skills/<name>; only missing skills are written. harness init --force re-mirrors everything. harness up now also seeds the gitleaks pre-commit hook.
• 8sync skill update [name] — re-pull registered skills from their recorded source in skills.toml (git URL / builtin: / path:). Git sources are deduped per URL (a collection repo is cloned once, all sub-skills reinstalled); best-effort per source (offline / missing git warns + skips, exit 0). name updates just one.
• 8sync harness up --pull — refresh AND re-pull every registered skill before re-injecting. Default up stays network-free + fast (timer/loop unaffected).
• 8sync harness up --commit — close the self-learning loop: stage + git commit ONLY the refreshed agent memory (agents/, AGENTS.md, CLAUDE.md, CHANGELOG.md, .gitignore; never your code) so learnings persist to git in the same pass. No-op when nothing changed (no empty-commit spam on --timer); default off.
• 8sync harness help — one-screen cheatsheet: commands, skill tiers, the commit-vs-ignore file taxonomy, and the new-machine runbook.
• Portability: harness init/up seed a managed .gitignore block (between # >>> 8sync (managed) >>> sentinels) — ignore derived (.codegraph/, .cache/8sync/)
  • secrets (.env, .env.*, keep !.env.example), keep agent memory + agents/skills/ committed. 8sync doctor now errors if any durable agents/*.md / AGENTS.md / CHANGELOG.md is gitignored (learnings wouldn't survive a move to a new machine).
• agents/KNOWLEDGE.md seeded with an append-only ## Learnings zone below the managed breadcrumb block (overwritten each harness up) so learnings persist.

Hardened (research-driven — see outputs/harness-selfimprove-research-brief.md)

• Lean force-load context — the injected on-demand skill list is now names+path only (one line each); full descriptions live in each SKILL.md (progressive disclosure). 8sync doctor warns if the AGENTS.md force-load block exceeds 120 lines. Why: Gloaguen et al. arXiv 2602.11988 (138 repos) — bloated/duplicative context files cut agent success and add >20% inference cost.
• Skill version pinning (lockfile) — 8sync skill add <url>@<ref> pins a git commit/tag/branch; the resolved SHA is recorded as rev in skills.toml and skill update checks out exactly that rev (reproducible). Unpinned entries track latest. Why: mirrors Claude Code plugin marketplace (SHA pin = reproducible).
• Secret-scanned auto-commit — harness up --commit runs gitleaks protect --staged (if installed) and ABORTS on detection; harness init installs a gitleaks pre-commit hook (non-destructive); 8sync doctor reports gitleaks. Why: GitGuardian 2026 — AI-assisted commits leak secrets ~2× baseline.
• Bounded memory (anti context-rot) — harness up consolidates the ## Learnings zone past ~200 lines, archiving older entries to agents/archive/ with a pointer. Why: 4-lever consolidation; "remember everything → remember nothing".
• Verifier-gated learnings — seeded KNOWLEDGE.md instructs prefixing entries validated: (test/build confirmed) vs hypothesis:. Why: Reflexion verifiability constraint — no reliable improvement beyond what's objectively verified.

[0.15.1] — 2026-06-17

Added

• impeccable house design references (assets/skills/impeccable/references/house/): bundled frontend-agent-workflow.md (senior coding-agent workflow) + clouds-f.md (senior front-end orchestration) + clouds-f-rules/*.mdc (design-redesign / responsive / performance / fix / refactor / security keyword routers). impeccable's SKILL.md auto-references them.

Changed

• Emphasised impeccable as THE house design system across the force-load flow (AGENTS.md / CLAUDE.md block, 00-force-load.md, sub-folder index, KNOWLEDGE breadcrumb): mandatory for any UI / design / redesign / audit, read with references/house/*.

[0.15.0] — 2026-06-16

Added

• 8sync harness verb — one command to stand up the full agent harness.
  • harness init: deploy mọi bundled skill + codegraph binary + external skill packs (best-effort clone), mirror vào agents/skills/, codegraph init, seed agents/* memory + CHANGELOG.md, inject force-load vào AGENTS.md/CLAUDE.md
    • một index gọn vào mọi sub-folder code (progressive disclosure). Có progress UI [i/N] + thời gian.
  • harness up: refresh theo state hiện tại (re-inject + refresh agents/KNOWLEDGE.md breadcrumb + codegraph index). --loop <dur> chạy foreground; --timer <dur>|off cài/gỡ systemd user timer (đúng cách cho chạy nền, mirror 8sync clean --timer).
• 6 bundled skill mới: ponytail (always-on, lazy-senior YAGNI), code-review-and-quality, senior-security, senior-frontend, full-flow, encore-deploy (on-demand). Trước đó (0.14.x → nội bộ) đã thêm assp-skill, impeccable, taste-skill. Tổng 15 bundled.
• Always-on order (đọc top-down, ưu tiên): codegraph → karpathy → ponytail → assp → impeccable → taste → 8sync-cli → image-routing. Inject block dạy rõ cách tận dụng từng skill.
• Tech-gated skills: encore-deploy chỉ hiện trong force-load block khi project dùng Encore (encore.app / encore.dev).
• Opt-in skills: social-growth (chiến dịch social/branding/lead-gen cho FB/YouTube/TikTok, page setup, insight, monthly plan + target) — KHÔNG auto-bật; bật bằng 8sync skill add builtin:social-growth.
• 8sync skill add collection-aware: clone repo rồi cài mọi skills/<name>/SKILL.md (vd addyosmani/agent-skills 24 skill, ponytail full); builtin:<name> deploy bundled skill từ embedded assets.
• Sub-folder AGENTS.md index + agents/KNOWLEDGE.md breadcrumb + CHANGELOG.md seeding tự động, để agent không bỏ sót rule và tự học theo state dự án.

Changed

• 8sync skill sync → 8sync harness init (clean cutover, không giữ alias). skill sync in cảnh báo trỏ sang lệnh mới.
• crates/cli/src/verbs/skill.rs (~1340 dòng) tách thành module tree verbs/skill/ (mod · meta · discover · list · spec · add · gen · deploy · inject · index), mỗi file < 500 dòng. Harness logic ở verbs/harness/ (mod · init · up · memory · external).
• 8sync . giờ cũng inject sub-folder index (nearest-AGENTS.md wins).
• Binary size target: < 4 MB (binary ~3.8 MB stripped, gồm 15 bundled skill).

[0.14.2] — 2026-06-02

• fix(bt): Bluetooth vanishing after cold boot (USB autosuspend).

[0.14.1] — 2026-05-31

• clean is project-safe: never touches models / Playwright / download caches.

[0.14.0] — 2026-05-31

• 8sync clean: disk/RAM reclaim + CPU/GPU report + periodic timer.

[0.13.0] — 2026-05-31

• 8sync bt bluetooth verb; Caelestia desktop install removed.

[0.12.1] — 2026-05-30

• two-tier skill injection (always-on vs on-demand).