Claude Auto-Dev

Autonomous development workflow for Claude Code. Say what you want to build — Claude handles the rest.

Distributed as a plugin marketplace. Claude Code installs, updates, and removes
it; there is no install script and nothing is copied into ~/.claude.

Install

In Claude Code — the desktop app, the CLI, or an IDE session:

/plugin marketplace add djnsty23/claude-auto-dev@stable

/plugin install autodev-core@autodev

That is the whole install. Then say brainstorm.

@stable pins the install to the last release. Without it the marketplace
follows main, which carries unreleased commits under the last version
number, so two installs of "the same version" can run different code.
Already installed without it? In ~/.claude/settings.json, add
"ref": "stable" to the source of extraKnownMarketplaces.autodev, then
run the marketplace add line above. Your plugins stay installed. Running
the add line alone is refused while settings declare a different source.

Two optional add-ons:

/plugin install autodev-memory@autodev

/plugin install autodev-stack@autodev

If the install summary says Run /reload-plugins to activate., run that.

Upgrading from 7.x? The old installer copied files into ~/.claude and
appended a function to your shell profile. Read MIGRATION.md
before installing — it clears those out.

What's in each plugin

autodev-core stands alone. The other two are additive and can be removed
without touching it.

Commands

Plugin skills are namespaced, so /autodev-core:audit always works even if you
have another audit skill installed. See docs/commands.md
for the full list.

Start with /autodev-init. It measures how your codebase is actually
written — component style, data fetching, where auth is enforced, tokens vs raw
colors — and writes .claude/project-rules.md. review, audit, and
standards all defer to that file, so the tool enforces your conventions
rather than the defaults this plugin happens to ship. Every rule it writes cites
a count from your code; anything genuinely split is recorded as undecided and
never flagged.

These build on Claude Code, they do not replace it. review, security, and
fix each run the matching built-in command first (/code-review,
/security-review, /debug) and then add only what is specific to this
project — its design tokens, its RLS rules, its definition of "verified".

Quick fixes — skip the ceremony. For small tasks, just describe what you
want. No auto, no sprints, no prd.json:

fix the button overflow on mobile
add loading state to the dashboard

Workflow

brainstorm → scans codebase + live site, proposes improvements
auto       → implements all pending stories + visual verification
ship       → review + security + deploy + post-deploy scan
iterate    → brainstorm → fix → re-scan loop until clean

Framework Radar. framework radar collects official changes across Claude
Code, Codex, Gemini CLI, coding agents, agent SDKs, orchestration frameworks,
MCP, evaluation harnesses and adjacent automation tooling. It discovers relevant
captioned videos, keeps raw transcripts outside the repository, and corroborates
each lead against primary sources and the current code. Every selected hypothesis
is then executed as an A/B/simpler experiment in an isolated worktree. A winning
variant may become a reviewable PR, but scheduled runs never merge, deploy, tag or
release. Reports and raw measurements live under .claude/reports/. The
deterministic collector is also available as npm run radar in this repo.

Visual verification. Claude opens your app in the built-in Browser pane,
reads the page, checks the console, and screenshots desktop and mobile after each
UI change. There is no skill to invoke — the browser tools are used directly.

The separate browser skill and the agent-browser CLI steps were dropped in
8.79/8.80. The binary itself is unrelated and may still be installed for other
tools, which is why agent-browser-cleanup.js still runs: it clears zombie
Chromium processes and a stolen Win+Shift+S hotkey on Windows.

Phone screenshots and other out-of-band files. Save anything into
~/Library/Mobile Documents/com~apple~CloudDocs/claude-inbox — iCloud, so an iOS
Shortcut can drop a screenshot there in one tap — and the next prompt announces
it with filename, path, and arrival age. Measured at ~30ms per prompt and zero
context when the inbox is empty, which is almost every turn; the cost is flat
whether one file is waiting or twenty-five, because the hook stats the directory
and never opens a file. Each arrival is announced exactly once. Claude reads the
image only when the arrival time makes it plausibly relevant — auto-injecting
every screenshot would cost roughly a thousand tokens each.

Set AUTODEV_INBOX to use a different folder, AUTODEV_INBOX_DISABLED=1 to turn
it off, and /inbox to list what is waiting.

Image auto-scan. Attach a screenshot to any turn and Claude surfaces every
distinct issue it sees, not only the one you asked about. Add [focus] in your
message to opt out.

Verifying the framework itself

This repo ships checks and is therefore held to its own standard: coverage measures
execution, mutation measures verification. A function can be entered on every run
while nothing asserts anything about it.

npm test                  # every suite, then validate. The gate.
npm run check:hooks       # wired hooks no suite drives — a hard gate in validate
npm run check:functions   # functions never entered (~20s)
npm run check:vacuity <subject.js> <suite.js>   # code no assertion depends on
npm run check:suites      # suites that cannot fail
npm run check:superseded  # guidance a later decision has overtaken
npm run check:versions    # the six files that must agree on a version
npm run check:runtime     # asserts the version EXECUTING is the one you edited
npm run check:agent-cost  # what a subagent really costs, from real transcripts
npm run check:agent-budget --lenses 4 --verify adversarial   # how many agents to spawn now
npm run actions:cost      # CI spend, from GitHub's own usage CSV

Four of those answer different questions and none substitutes for another: scripts
nobody runs, hooks no suite drives, functions never entered, and code no assertion
depends on.

check:vacuity rewrites its subject with mutants. It refuses a dirty subject, and
validate fails while a *.vacuity-backup exists. If you kill a run, pkill -9 then
pgrep to confirm — a survivor rewrites the file underneath you.

Running a fleet

When several sessions run at once, they cannot see each other. These read the
same transcripts the app writes and answer the questions that causes.

They live in plugins/autodev-core/scripts/. Two design rules they all follow,
learned by getting them wrong first:

Print the population, not a verdict. "212 of 212 files read, 9 names, clean"
can be judged. "clean" cannot be told apart from a check that ran on nothing.

Distinguish "checked, none found" from "could not check". Silence must never
read as clean. Each script reports a missing dependency or an unreadable input in
place rather than returning an empty result.

fleet-overlap.js scores three signals separately rather than blending them,
because they mean different things: two sessions on one branch will collide, two
in one repo might, and two sharing a word in their titles probably will not.

watch-panels.js takes --self <sessionId> or AUTODEV_SELF_SESSION so it does
not report your own questions back to you, and persists its dedup state to disk so
restarting it does not re-raise what you already answered.

Updates

/plugin marketplace update autodev

/plugin update autodev-core

There is no update-dev command any more — Claude Code owns the update.

Settings

Plugins cannot change your permissions or your model, by design. If you want the
permission set this workflow assumes, merge
docs/recommended-settings.json into your own
settings — see docs/settings.md, which also explains which
rules from the old 7.x template were removed and why.

Files

Per project (created as needed):

prd.json                    # Tasks and sprint history
.claude/archives/           # Archived prd snapshots
.claude/reports/            # Scan and audit reports
.claude/screenshots/        # Visual verification output

Tasks use passes: null (pending), true (done), or "deferred".

Global: ~/.claude/auto-dev-memory.db — the autodev-memory SQLite store.
It deliberately lives outside the plugin so uninstalling does not delete your
project memory.

Uninstall

/plugin uninstall autodev-core@autodev

Repeat for any add-ons. Claude Code removes exactly what it installed. To drop
the marketplace too:

/plugin marketplace remove autodev

Your prd.json files, memory database, and settings are untouched.

Repository layout

.claude-plugin/marketplace.json   # The catalog
plugins/autodev-core/             # Plugin: skills, agents, hooks, templates
plugins/autodev-memory/           # Plugin: skills, hooks, runtime scripts
plugins/autodev-stack/            # Plugin: skills
docs/                             # Docs, settings + CLAUDE.md templates
tooling/                          # Repo tooling — validate, tests, bump. Never shipped.
VERSION                           # Single source of truth; tooling/bump.js propagates it

Contributions: see CONTRIBUTING.md. Run npm test before
opening a PR — it runs every suite plus tooling/validate.js.

Troubleshooting

Skills not showing up. Run /plugin and confirm autodev-core is listed and
enabled. If the install said so, run /reload-plugins. Hook changes need a new
session.

Hook errors. Hooks need Node 18+ (node -v). They fail quietly by design;
node tooling/validate.js checks that every hook in every hooks.json points at
a file that exists.

Memory commands do nothing. autodev-memory needs node:sqlite, which is
built in on Node 22+. On older Node the memory skills no-op rather than error.

Still have update-dev in your shell. That is from the 7.x installer.
MIGRATION.md removes it.

License

MIT