v0.23.0 — harness eval + concrete /gs L3 worktree isolation

Measure the loop, isolate the loop.

Added

• 8sync harness eval — loop quality probe. Runs a fixed task-suite through omp (omp -p --no-session --auto-approve) and scores each task with a deterministic verify.sh the agent can't game. Fixtures: fix-failing-test, add-fn-with-test, locate-symbol. JSON scorecard + --baseline reference in the gitignored .cache/8sync/eval/; later runs print the pass delta. Non-deterministic quality signal, not a CI gate. Verified 3/3.

Changed

• /gs L3 worktree isolation is now concrete: git worktree add .gs/wt/<slug> -b gs/<slug> → implement+verify+commit there → git worktree remove; never edit main's tree.

Phase 3b (gstack omp host) deferred — additive with a working bundled fallback; the host lives in the deinitialized gstack submodule, not su-code.

Install: curl -fsSL https://raw.githubusercontent.com/8-Sync-Dev/su-code/main/install.sh | sh

v0.22.0 — doc-hygiene audit + AI-engine health + loop fixes

Make the engineering-loop harness verifiably correct.

Added

• 8sync harness audit — code-backed doc-hygiene: stale repo-relative paths in docs, oversized docs (>400 lines / >120-line force-load block), 30-day churn hotspots. Report-only. 8sync doctor shows a one-line summary; /gs doc-hygiene runs it.
• 8sync doctor AI-engine health check — codegraph + codebase-memory-mcp + headroom must be installed AND registered in omp (enforces STEP 0 token discipline).

Fixed

• codegraph STEP 0 verbs corrected to query/callers/callees/impact (were search/deps/defs, nonexistent in 0.9.6).
• force-load now dedups skills by frontmatter name (the stale karpathy dir double-listed karpathy-guidelines).
• impeccable .agents/skills → agents/skills (28 refs) so its setup scripts run under 8sync's layout.

Changed

• managed .gitignore ignores .gs/.

Install: curl -fsSL https://raw.githubusercontent.com/8-Sync-Dev/su-code/main/install.sh | sh

v0.21.0 — right-size /gs (fix post-/gs quality regression)

Eval + deep-research showed the post-/gs quality drop was process over-engineering, not tokens (harness bench: ~8.5k upfront, 79% saved, KV-cache stable). /gs was a 93-line command forcing a team + full Closeout on every task, with auto "never asking".

Changed — /gs redesigned to right-size effort

• Right-size first — 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).
• 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 — 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 — detect stale paths / junk / superseded docs → fix or delete (no addition without the matching deletion); keep docs lean.
• Codebase-history — git log/blame + DECISIONS + cbm before load-bearing edits.
• Leaner command — 93 → 56 lines (lower constraint density → better instruction-following).

Research brief: outputs/gs-eval-improve-research-brief.md (+ provenance). Redeploy with 8sync harness. Upgrade: curl -fsSL https://raw.githubusercontent.com/8-Sync-Dev/su-code/main/install.sh | sh

v0.20.1 — /gs truly unattended + QA/Closeout + reference submodules

Fixed

• /gs auto now truly runs unattended. Autonomy contract: in auto the team NEVER asks — it resolves unknowns by research (codegraph/cbm → memory/PLAYBOOKS → skills → web_search/autoresearch/deep-research), picks the boring/reversible option, logs it under ## Assumptions in STATE, and proceeds. A "blocker" is now only a missing credential / external approval / destructive-irreversible action. (omp's default tools.approvalMode: yolo already auto-approves tools; a slash command cannot bypass the approval gate.)
• /gs argument hint when you type /gs (front-loaded description + argument-hint frontmatter).
• QA + test are first-class — per-slice verify-gate runs tests + a QA pass (no skipping/weakening), plus a mandatory Closeout (full test suite + end-to-end QA + independent re-review vs Definition-of-Done + handoff summary) before the loop reports done.

Added

• reference/gstack + reference/gsd-pi git submodules for studying the patterns behind /gs. Pointers committed; working trees deinitialized by default so they never bloat the index (codegraph honors no exclude). Study on demand: git submodule update --init reference/<name>.

Bare 8sync harness = full auto-setup (MCP + skills + /gs + memory + inject + index) in one command.

v0.20.0 — /gs autonomous engineering-team loop

Added

• /gs — one command, a whole engineering team that runs until done. /gs <goal> plans + runs · bare /gs resumes · /gs auto runs unattended · /gs status|next|stop. Drives the loop-engineering v2 (A–E) machinery 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 + guardrailed. codegraph + codebase-memory-mcp + headroom mandatory; verify-gate before every commit; worktree isolation + no push/PR at L3; hard-stop via /gs stop. Modeled on gsd-pi /gsd auto.
• Team-shared. 8sync harness/init/up deploy it to ~/.omp/agent/commands/gs.md (global) and <repo>/.omp/commands/gs.md (committed → whole team gets /gs). Run 24/7 with 8sync harness up --timer 30m. New on-demand gs skill documents the protocol.

bench on this repo: A1 stable-prefix PASS, upfront ~7.3k tok, A2 progressive-disclosure saved 81%.

v0.19.0 — loop-engineering v2 + harness bench

Loop-engineering v2 (Phases A–E) grounded in 2026 context-engineering research, plus a deterministic benchmark verb.

Added

• 8sync harness bench — deterministic loop-engineering scorecard (no model calls): upfront context budget (force-load prefix + CORE skill bodies + memory spine) vs deferred (SPECIALIST + on-demand), A2 progressive-disclosure saving, A1 KV-cache stable-prefix gate. This repo: upfront ~7.1k tok, A2 saved ~32.3k tok (81%), A1 PASS.
• agents/PLAYBOOKS.md — Voyager-style procedural memory (validated runbooks indexed by When:).

Changed (loop engineering v2)

• A token & stable-prefix: CORE/SPECIALIST progressive disclosure, mandatory headroom_compress (>~50 lines), KV-cache-stable KNOWLEDGE breadcrumb (no volatile epoch).
• B live memory: structured agents/STATE.md live-plan recitation + compaction + budget-awareness.
• C maker/checker: task implementer ↔ independent verifier (verify-gate before commit) + Reflexion failure: capture.
• D procedural memory + tiering (KNOWLEDGE / PLAYBOOKS / DECISIONS).
• E phased autonomy L1→L3 + guardrails (no auto push/PR at L3); timer-loop per-tick job.

Plan + provenance: outputs/harness-loop-engineering-v2-plan.md.

v0.18.1

Fixed

• 8sync harness init now pulls registered manifest skills — init runs 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.

v0.18.0 — Headroom context-compression wired as an omp MCP

8sync harness now auto-sets-up Headroom (context compression) as a third token-efficiency engine alongside codegraph + codebase-memory-mcp.

Added

• Headroom MCP — 8sync harness/init auto-installs headroom-ai[mcp] (uv → pipx → pip) and registers headroom mcp serve (stdio) in ~/.omp/agent/mcp.json. 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, skipped

• PixelRAG (screenshot-RAG) — overlaps 8sync shot / omp read / browser.
• LocateAnything3D — 3D-vision model, out of scope for a coding harness.

Restart omp (or /mcp reload) to load the new server. Upgrade: curl -fsSL https://raw.githubusercontent.com/8-Sync-Dev/su-code/main/install.sh | sh

v0.17.1 — fix: skills propagate to other machines (committed agents/skills.toml)

Fix: 8sync harness now reproduces your skills on a second machine.

Root cause

The skill registry that records what to pull (~/.config/8sync/skills.toml) is machine-local and never travels. On a fresh machine, 8sync harness re-pulled nothing beyond the 15 binary-embedded skills + 2 hardcoded external packs — every skill add-ed source (e.g. git collections like feynman) was invisible.

Fix

harness / skill update now write a committed project manifest agents/skills.toml (union of the machine-local registry) and read it back on any machine. A fresh git clone + 8sync harness re-pulls the exact same skills. agents/skills.toml is a file, so it travels even when the agents/skills/ directory is gitignored.

To propagate skills to another machine

1. On machine A: 8sync harness (writes agents/skills.toml), then commit + push it.
2. On machine B: git clone → 8sync up (get ≥ v0.17.1) → 8sync harness → skills re-pulled from the manifest.

Upgrade: curl -fsSL https://raw.githubusercontent.com/8-Sync-Dev/su-code/main/install.sh | sh

v0.17.0 — codebase-memory-mcp engine + code-intelligence STEP 0 + loop engineering

8sync harness now auto-sets-up a second code-intelligence engine alongside codegraph, and force-injects both as the mandatory STEP 0 for token-efficient code work.

Highlights

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

Upgrade: curl -fsSL https://raw.githubusercontent.com/8-Sync-Dev/su-code/main/install.sh | sh