v0.3.1: hard-wired edge cases

Three fix families, all mirrored in node + python:

• Corrupt or misshapen .harness/config.json now fails gently everywhere (doctor, skill add/update/remove, optimize, instinct toggles), no tracebacks.
• Staged-file secret scans report true line numbers (hunk-header aware), tolerate non-UTF-8 staged files, and the auditor skips its own deploy copies so fresh installs do not self-flag on their first git add -A.
• One run at a time in the GUI run endpoint; hook toggles record disabled hooks in the manifest so doctor --fix no longer re-enables them; --harness/--port flag values fail loudly; memory prune and optimize are separate commands; fixes for animated install deadlocks, escape-key handling, and frontmatter ending at the closing fence.
• Dev dependency refresh: npm audit clean (vite 8.3.5, vitest 5.0.5).

Full sweep green: tsc, eslint, 30 vitest, 24 pytest, audit clean, ci + security + CodeQL workflows.

v0.3.0: web GUI, menu CLI, 13 features

Local web GUI (Overview, Agents, real-terminal Commands with menu+shell PTY modes, Memory, Skills, Project, Activity, Settings with 19 themes, About, Update modal). Menu-only interactive CLI. 13 new subcommand features. haveBin PATH scan, update-info caching. Install hints moved to theelephantcoder.com/agent-harness. M6 local-model evidence in docs/benchmarks.md. Full sweep green: tsc, eslint, 27 vitest, 20 pytest, 3h soak zero failures.

v0.2.1

agent-harness v0.2.1

The interactive prompt grows up: status snapshot, per-project shell history, tab completion down to installed skill and hook names, two-digit menu entry, !! repeat, clear, and a one-line project context on entry. Everything mirrored in both the node and python CLIs.

• harness status: init state, MEMORY size, findings, last-bench age, optimizations, skills/hooks/adapters
• Shell history in .harness/history (initialized projects only, never fakes init state)
• Tab completion for subcommands, flags, skill names, hook paths, toggle names
• Python CLI accepts harness version and -v like the node CLI

Install: npm (tarball), brew tap, apt repo, pip. See README Quick start.

v0.2.0

agent-harness v0.2.0

The CLI does real work now: init installs per-harness native files, doctor verifies with exit codes, bench measures with baselines, optimize prunes memory to budget, upgrade self-updates with a progress bar.

• Per-agent installs verified against vendor docs (Cursor .mdc, Kiro steering, Cline rules, Codex .agents/skills, Aider conf)
• Benchmark evidence: 130+ runs across kimi-k3, nemotron-3-super, muse-glimmer-30b, failures published alongside passes
• Community benchmark tab with a submission flow
• Secret scanning with file:line reports, honest pip limits, slim init payload

Install: npm (tarball), brew tap, apt repo, pip. See README Quick start.

v0.1.2

Linux support and apt install. Compiled JS bin, portable bump-brew.sh, build-deb.sh, deb.yml workflow, docs/apt.md. Install via apt repo, npm, pip, brew or source.

v0.1.1

Fix brew install on Node 26 by shipping compiled JS (bin now points at dist/harness/cli.js). Also includes Kiro CLI and Kiro support.

v0.1.0

Initial release. Skills, instincts, memory, security, research-first for Claude Code, Codex, Opencode, Cursor, Kiro CLI and Kiro. Install via npm, pip or brew.