Paseobility

Fifteen reusable skills for projects, browsers, artifacts and computers connected through Paseo.

English · 한국어

A first resident Paseo bot manager is available under tools/paseo-bot: project registry across connected machines, durable task intake, worktree dispatch and independent review. See Paseo bot v1.

At a glance

Current package: v3.0.1. Existing skills are now named ps-<name> instead of paseo-<name>, with ps added as the catalog and explicit ps-plan/ps-write harnesses; release history explains how to move an existing installation. The plugin supports Paseo 0.12.x–0.15.x and the checked 0.16.0-beta.1 prerelease, fixing the rejection of 0.14+ in v2.12.11. The 0.15.2 check records the reproduction, compiler/SDK matrix, live RPC and adapter results, and remaining UI limits. /ps-multi was exercised with Paseo 0.11.2 on the listed Mac/Windows combinations. The existing package compatibility baseline is 0.10.2; this is not a claim that all fifteen skills were verified on 0.11.2. The 0.12.0 report is a static source/CLI check. The 0.13.0 report adds live runs of the plugin install and the cleanup/multi adapters against throwaway 0.12.0 and 0.13.0 daemons, not a package-wide runtime certification. Multi validation · Package baseline · 0.12.0 static check · 0.13.0 check.

Upgrading from v2.x — every installation needs this once. The update button in a v2 plugin cannot install v3.0.0: it stops with "Invalid skill name" and leaves the old skills in place. From a checkout of this repository:
./scripts/paseobility-init.sh --migrate-skills --with-claude --no-context   # macOS, Linux
.\scripts\paseobility-install.ps1 -MigrateSkills -WithClaude                # Windows
Old paseo-<name> folders move to the backup folder the installer prints. Then replace the plugin with paseo plugin remove paseobility followed by paseo plugin add git:wilgon456/Paseobility:plugin, start a new agent session, and rename any skill you mention in your own prompts or settings (/paseo-god is now /ps-god).

One-line summary

Paseobility is an agent-installable Paseo slash-skill pack: you hand this GitHub repo URL to Codex, Claude, or a Paseo agent and have it install.

Once installed, you can pull frequently used Paseo patterns out like slash commands: browser use (web browser manipulation), multi-agent orchestration, agent tournaments, session briefs, project bootstrap, pre-install repo security checks, archiving inactive subagents, and cross-device artifact sharing.

With base Paseo alone you can combine built-in tools to do similar work. But letting the agent re-decide that combination every time is slow and produces inconsistent results, so the pack bundles frequently used patterns for immediate reuse.

This repo is a skill package for reproducibly composing Paseo's built-in tools. It ships a macOS/Linux bootstrap helper and a Windows PowerShell install helper.

/ps <task> reads the local skill catalog across agent homes, chooses a fitting specialist and reads its original SKILL.md. /ps list only shows the list. Requires Node 20+; it creates no cache or usage log and copies no skills.

Quick start

Install one skill first. This example needs Git and Bash on macOS/Linux, or Git and PowerShell on Windows; it does not install the optional Cua Driver.

git clone https://github.com/wilgon456/Paseobility.git
cd Paseobility
./scripts/paseobility-init.sh --skill ps-research --no-context
test -f ~/.agents/skills/ps-research/SKILL.md && echo 'Skill file installed'

git clone https://github.com/wilgon456/Paseobility.git
cd Paseobility
.\scripts\paseobility-install.ps1 -Skill ps-research
Test-Path (Join-Path $env:USERPROFILE '.agents\skills\ps-research\SKILL.md')

Open a new agent session or reload integrations, then try /ps-research Explain how this project works using its source code. File presence confirms the copy, not that the current session loaded it. See installation for the full pack, optional Claude support and driver requirements, and verification for tested platforms.

`/ps-god` — explicit task mode (v0)

Define observable completion criteria, use only the needed debugging/review/spike procedure, and retain a separate TODO, evidence and resume point for each run. Learning is proposed for review; native memory and skills are not automatically rewritten. The optional Node helper creates/resumes records, not autonomous jobs.

./scripts/paseobility-init.sh --skill ps-god --no-context

.\scripts\paseobility-install.ps1 -Skill ps-god

/ps-god Fix this bug, verify the original trigger, and save a resume point if blocked.

See the God guide for scope, local records and validation limits. Existing version and platform reports retain their original scope.

Release history contains the dated changes and their original validation limits.

How to use

The primary user of this project is not someone installing from a shell directly, but someone who gives an AI agent the GitHub URL and delegates installation and verification.

To install only ps-share, send the following sentence verbatim to Codex or Claude.

https://github.com/wilgon456/Paseobility

Read the instructions in this repository, install only ps-share on this computer, and verify it.
Back up any existing install and do not restart the Paseo daemon.
When it is done, tell me how to start a new session/reload and how to make the first share request.

https://github.com/wilgon456/Paseobility

Read this repo and install into my local Paseo skills directory following AGENTS.md.
First test the install with a temporary --target-home/-TargetHome path, and install into the real skills directory once it passes.
After installation, also check that /ps-project is recognized in a new Paseo session.

The agent reads AGENTS.md in Codex/Paseo and CLAUDE.md in Claude Code, then follows the OS-specific paths and install procedure. This Paseobility repository is the install source. Actual shared artifacts live in the private paseo_share repository of the authenticated account; installing never creates a repository by itself.

When you explicitly request these tasks, you can use the following capabilities:

Open, read, click, type into, and verify web pages with screenshots.
Split work across multiple agents and synthesize the results.
Compare answers from several models such as GPT/Claude/DeepSeek/Grok and pick a winner or a merged plan.
Before installing a GitHub URL or local repo, read-only inspection of spyware/supply-chain risk signals.
Archive the inactive agents you name or that carry test-only markers, preserve ordinary idle sessions, and archive workspaces after approval.
Publish work artifacts to a private Git share and open or fetch them from another computer or mobile.
On request, summarize repo context, commands, instructions, and risks on one page.
When initial setup is requested, gather README, docs, Claude/Codex/Cursor-style instructions to build working context.
Design cross-provider collaboration such as implementing with Codex and reviewing with Claude.
Put clear guardrails on infinite loops, dangerous submissions, and account changes.

Included skills

Skill	Role	Strong at requests like
`/ps`	On-demand local skill catalog	`/ps <task>` chooses one fitting specialist; `/ps list` shows skills across agent homes. Reads originals without copying or installing
`/ps-plan`	Explicit planning harness	One Markdown plan: graded current state, options, testable criteria, independently verifiable phases and risks
`/ps-write`	Explicit writing harness	Reader-facing documents with source tracing and cold-reader checks. Korean drafts are revised using a pinned `humanize-korean` rulebook.
`/ps-toss`	Apps in Toss mini-app workflow	Planning, scaffolding, browser debugging, phone testing and pre-launch review; authenticated console tools required for uploads
`/ps-god`	Explicit task execution and resumption	Bounded work, acceptance evidence, debugging/review/spike and learning proposals
`/ps-multi`	Operate selected Paseo computers	Named hosts or `all` (every registered host, in parallel), remote status/agents, inline or file Bash/PowerShell commands (`exec --command`, `--wait` up to 300 s), verified updates, `restart --force` only when you asked for an immediate restart, merging `import-saved`
`/ps-design`	Web and app frontend workflow	Project design rules, existing components and frontend review
`/ps-browser`	Browser workflow	Default to one Ego Lite TaskSpace; Chrome requires an explicit user choice. Preserve user tabs and verify completion
`/ps-cua`	Native desktop GUI driving	App window manipulation/snapshot/verification with the trycua Cua Driver, whenever a task needs a native app GUI
`/ps-orchestration`	Multi-agent coordination and comparison	Explicitly requested coordination or tournament, mode-specific guidance
`/ps-project`	Project summary and setup	Requested brief (writes or refreshes stale `.paseobility` context) or environment/context setup
`/ps-share`	Computer/mobile artifact sharing	Auto-prepare private `paseo_share`, any file type up to 50 MiB (several per publish), clickable preview/download links, share-ID-based verification and fetch
`/ps-research`	Research with primary sources	Code at a pinned commit, history, official versioned docs; each claim graded observed/documented/inferred/unknown
`/ps-spyware-check`	Pre-install security/spyware static check	Checking GitHub URLs, local repos, install scripts, secrets, exfiltration, supply-chain risk
`/ps-agent-cleanup`	Inactive agent/workspace cleanup	Preview by default, preserve ordinary idle, archive named or test-marked candidates directly, protect active agents, archive workspaces only with `--yes`

Install

The recommended install method is to give an AI agent this repository URL and let it install.

https://github.com/wilgon456/Paseobility
Read this repo and install into my local Paseo skills directory.
Copy skills/* for my Windows/Mac environment and tell me how to reload afterward.

The agent checks the Mac/Windows skills paths based on AGENTS.md in Codex/Paseo and CLAUDE.md in Claude Code. A ps-share request installs only that skill instead of the full pack.

You can also install manually.

git clone https://github.com/wilgon456/Paseobility.git
cd Paseobility

# Install the Paseo / Codex skills only
./scripts/paseobility-init.sh --no-context

# Update just one skill
./scripts/paseobility-init.sh --skill ps-agent-cleanup --no-context
./scripts/paseobility-init.sh --skill ps-spyware-check --no-context
./scripts/paseobility-init.sh --skill ps-share --no-context

# Also install for Claude Code
./scripts/paseobility-init.sh --with-claude --no-context

# Bootstrap a specific project right away
./scripts/paseobility-init.sh --root /path/to/your/project

On Windows PowerShell:

git clone https://github.com/wilgon456/Paseobility.git
cd Paseobility

# Install the Paseo / Codex skills
.\scripts\paseobility-install.ps1

# Update just one skill
.\scripts\paseobility-install.ps1 -Skill ps-agent-cleanup
.\scripts\paseobility-install.ps1 -Skill ps-spyware-check
.\scripts\paseobility-install.ps1 -Skill ps-share

# Test-install into a temporary home first
.\scripts\paseobility-install.ps1 -TargetHome $tmp.FullName

# Also install for Claude Code
.\scripts\paseobility-install.ps1 -WithClaude

Cua Driver runtime

Selecting ps-cua, or installing the full package, also ensures the trycua Cua Driver runtime. Other skill selections have no driver side effects. The runtime step is idempotent: an existing working driver is reused and never auto-upgraded.

# ps-cua selected -> driver auto-ensured
./scripts/paseobility-init.sh --skill ps-cua --no-context

# Skills only, no runtime (docs-only / offline)
./scripts/paseobility-init.sh --skill ps-cua --skip-cua-driver --no-context

# Custom target homes skip the real-host runtime by default and log why
./scripts/paseobility-init.sh --target-home "$tmp" --no-context

# Explicitly allow real host runtime install with a custom target-home
./scripts/paseobility-init.sh --skill ps-cua --target-home "$tmp" \
  --allow-host-runtime --no-context

.\scripts\paseobility-install.ps1 -Skill ps-cua
.\scripts\paseobility-install.ps1 -Skill ps-cua -SkipCuaDriver
.\scripts\paseobility-install.ps1 -TargetHome $tmp.FullName
.\scripts\paseobility-install.ps1 -Skill ps-cua -TargetHome $tmp.FullName -AllowHostRuntime

The driver helper can also be run directly:

./scripts/paseobility-cua-driver.sh --check       # presence only (exit 3 when missing)
./scripts/paseobility-cua-driver.sh --dry-run     # print the plan, execute nothing
./scripts/paseobility-cua-driver.sh               # ensure the pinned driver

If the runtime step fails, the installer exits non-zero and reports that skills were copied but the runtime setup failed (installation may be incomplete); re-run or pass --skip-cua-driver / -SkipCuaDriver to install skills only.

Doctor and tests

The doctor is read-only, bounded, and cross-platform (Python 3 stdlib). Text mode is for a human; --json is the shareable artifact carrying only allow-listed fields (statuses, counts, validated skill names, sanitized semantic versions) — never raw command output. Exit code 0 means "a report was generated", not "everything is ready". --source-root overrides the skill pack; --target-home points at an isolated install root.

python3 scripts/paseobility-doctor.py --root /path/to/project
python3 scripts/paseobility-doctor.py --root /path/to/project --json
# Explicit-only MCP protocol/tools discovery (needs a running daemon; not a GUI test):
python3 scripts/paseobility-doctor.py --root /path/to/project --check-mcp
# Isolated install check against a temporary home:
python3 scripts/paseobility-doctor.py --target-home /tmp/isolated-home --json
# Thin wrapper with the same passthrough:
./scripts/paseobility-doctor.sh --root /path/to/project

.\scripts\paseobility-doctor.ps1 -Root .
# Same flags as the Python helper (-SourceRoot/-TargetHome/-CuaBin/-CuaBinDir):
.\scripts\paseobility-doctor.ps1 -SourceRoot . -TargetHome $tmp.FullName -Json

Offline test suites:

# macOS/Linux
python3 -m unittest discover -s tests -p "test_*.py" -v

# Windows (no network, injected fixture installer)
pwsh -NoProfile -File tests/test_cua_driver_runtime.ps1

The installer backs up an existing same-name skill before overwriting.

OS	Backup path
macOS/Linux	`~/.agents/skills-backups/Paseobility-<version>-<timestamp>/`
Windows	`%USERPROFILE%\.agents\skills-backups\Paseobility-<version>-<timestamp>\`

If you must force-replace without a backup, macOS/Linux uses --no-backup and Windows uses -NoBackup.

After installation, start a new agent in the Paseo app, or reload integrations/skills in Settings.

If you prefer a manual copy (documents only — it cannot auto-install the Cua Driver runtime):

mkdir -p ~/.agents/skills
cp -R skills/* ~/.agents/skills/

Windows manual install:

New-Item -ItemType Directory -Force "$env:USERPROFILE\.agents\skills" | Out-Null
Copy-Item -Recurse -Force ".\skills\*" "$env:USERPROFILE\.agents\skills\"

Requirements

Paseo desktop app
Paseo Settings -> Agents -> Enable Paseo tools enabled
One or more AI providers
- e.g. Codex, Claude Code, etc.
/ps-share needs Node.js and Git. Default GitHub auto-onboarding needs an authenticated GitHub CLI (gh); Forgejo and custom remotes connect with an explicit repository URL. GitHub Actions is not used.
/ps-cua needs the trycua Cua Driver. The installer auto-ensures it for ps-cua or the full package; macOS Accessibility and Screen Recording grants stay manual.
Orchestration prefers Paseo's list_profiles. The legacy ~/.paseo/orchestration-preferences.json is read only as optional extra guidance.
Install target skills path
- macOS/Linux: ~/.agents/skills
- Windows: %USERPROFILE%\.agents\skills
The bootstrap context script targets macOS/Linux shell environments
- On Windows, scripts/paseobility-install.ps1 only assists with skill installation.

Verification status

The CLI/MCP compatibility record for the existing six skills is the v2.7.0 compatibility report, which does not cover ps-cua. The six-skill record must not be read as covering all seven skills. The current package compatibility baseline is Paseo 0.10.2; its re-verification record is docs/compatibility-0.10.2.md.

ps-cua is preview / limited validation — not broad stable Mac and Windows support. Current dated status (2026-09-22):

Environment	Native GUI (Cua)	Browser result and limits
Apple Silicon macOS 26.6.2 · Cua 0.28.2 · direct local	Calculator input/read-back + PNG pass; `verify_state` unknown	viewport pass; built-in `fullPage` fails
Intel Mac mini 16 GiB · macOS 15.8 · Paseo 0.8.0 · Cua 0.28.2	user-reported: native GUI + PNG pass; `verify_state` unknown	browser input/viewport pass; built-in `fullPage` fails; same-tab scroll+stitch whole-PNG passed for two DPR-1 fixtures only (not a general helper)
Intel MacBook 8 GiB · macOS 15.7.9 · Paseo 0.7.2 · Cua 0.17.0 (reused)	user-reported prior E2E: native GUI/PNG pass; `verify_state` unknown	browser input/viewport pass; built-in `fullPage` fails; installed 0.9.0-beta.2 left unchanged and not revalidated
MacBook follow-up · isolated upstream PR `#3197` build (reported 0.3.1)	—	user-reported: real MCP `browser_screenshot(fullPage:true)` passed 2573px/3511px at DPR 2 twice each, repeat SHA identical
Windows 11 25H2 · Paseo 0.9.0-beta.2 · Cua 0.28.2	user-reported: MCP 57 tools + Calculator input/AX/PNG/`verify_state` pass	viewport/full-page PNG failed (`screenshot_no_frame` then 15s timeout, no image); DOM-fixture input/click/scroll pass; localhost HTTP failed; physical browser host unknown

Notes:

Only the Apple Silicon row was directly verified in this review; the Intel Mac and Windows rows are user-reported and were not independently reproduced. The MacBook follow-up used the isolated upstream PR build without additional product edits and is not a fix released or installed by Paseobility — the 0.9 backport was aborted on conflicts, and a 0.3.1 downgrade is not recommended.
Native passes cover narrow tested flows, not blanket platform support. The 8 GiB sample is not a RAM minimum or guarantee, and memory was not measured at peak.
Windows automatic runtime installation had a code-level conflict: the shared non-junction .local/bin was pre-created before the pinned upstream installer, which refuses an existing non-junction directory. v2.8.1 installs into a private staging bin and publishes only the executable set, and tests/test_cua_driver_runtime.ps1 covers the conflict, rollback, and shared-bin preservation offline. A real-host fresh Windows install was not reproduced for this release, so the earlier manual hardlink recovery remains the last real-host evidence. Two explicit-only skills remain unconfirmed in the Windows catalog.
Test counts are as reported and were not re-run for this README update: 41 automated tests on the Intel Mac mini (cleanup 11 + share 15 + scanner 13 + shell 2); Windows 63 passed / 2 failed / 1 skipped, unchanged.
The PowerShell wrappers were source-reviewed only on macOS (pwsh unavailable); the supplied Windows report confirms skill migration and manual runtime recovery. Isolated-path install/migration and helper regression tests were confirmed.
The v2.7 six-skill compatibility doc does not prove all seven skills at runtime. This documentation update does not modify the Paseo app or capture engine. Do not read binary presence as permission readiness. Other macOS versions and Linux remain unverified.

Current release testing (v2.8.1) — separate from the device rows above

The rows above are older device evidence. The v2.8.1 release ran its offline suites on macOS and Windows CI; CI run 35708474356 (head 9932ea7590ef137552d8b2b4e065e7fbb1792254) is SUCCESS for both the macOS and Windows jobs. Exact scope:

macOS: 49 Python tests discovered, 49 pass, 0 skip (Cua Driver runtime 18, doctor 21, E2E assets 3, skill migration 7), plus Node cleanup 11, Node share 15, Python scanner 13, and shell 2 — 90 automated tests total.
Windows shared Python suite: 49 discovered, of which 14 pass and the other 35 are POSIX-only skips. The Windows job also runs Node cleanup 11, Node share 15, and Python scanner 13, all passing.
Native PowerShell fixture suite: runs under both pwsh and Windows PowerShell 5.1, each reporting 109 assertions, 0 fail, 0 skip. This counts assertions, not 109 test cases, and it is a fixture suite — not a real downloaded-Cua GUI test.
Still unverified: a fresh native Windows Cua install from the official download path was not reproduced. The fixture suite proves the staged install / shared-bin preservation / rollback / skill-migration contract, not a live download.
./scripts/paseobility-doctor.sh --root . was run in text and --json modes against this checkout; the JSON is deterministic, carries only allow-listed fields, and reports the live host as Cua 0.28.2 (permissions granted, daemon running) and Paseo 0.9.0-beta.2 reachable.
With an explicit --check-mcp, the live Cua 0.28.2 MCP probe was observed negotiating the protocol and listing 56 tools on one connection. The probe spawns its own stdio server child, and the 56 tools come from that parent-run probe. This is a protocol/tools check only, never a GUI result, and it does not prove a Paseo agent client's MCP config or tool discovery.
GUI end-to-end is blocked, not passed. A manual attempt to drive the e2e/ browser fixture on this host failed: browser_new_tab http://127.0.0.1:<port> timed out waiting for tab registration, no connected tab appeared, and the fixture server was terminated with no listener left on the port. The current browser host could not register the tab, so no GUI pass is claimed. The historical device rows above are unchanged.
The updated ps-cua skill was installed locally for this release into both the Codex/Paseo and Claude skill directories, with backups; all seven source trees match, the other twelve installed trees were untouched, and the Cua runtime stayed at 0.28.2.

Rollback

These changes are scripts, tests, docs, and E2E assets; no Cua or Paseo binary is modified. To restore the previous helper behavior, the v2.8.0 tree is referenced but is not a guaranteed tag; the safest option is to check out the previous main commit 970ea8cab7c3fef757d2bd13f58c8389758aad1d in a separate clone, preserving your existing work. Skill rollback uses the installer backups.
There is no automatic runtime rollback. Paseobility never auto-upgrades the installed Cua Driver, so a driver that already worked is left as it was.
Diagnostics need Python 3, but installation never does: both paseobility-doctor.sh and paseobility-doctor.ps1 require Python 3 and exit non-zero with a clear message when it is absent, and the installer treats that as "diagnostics unavailable" and still completes. Without Python 3, skip diagnostics.

Details, including the optional Playwright fallback and per-host probes, are in the Cua platform validation status and the compatibility report. The entries below are prior-version records and are not evidence for the current version.

Past verification record — separate from this consolidation

Environment	Status	What was checked
Paseo 0.6.1 on Windows	Tested locally	CLI/daemon 0.6.1 agreement, agent/workspace/provider JSON, worktree/schedule/heartbeat/archive CLI schema, cross-check against current MCP/profile/browser sources, PowerShell temporary install
Apple Silicon macOS	Tested	Paseo CLI 0.2.5 detected `Darwin/arm64`, temporary HOME install, real `~/.agents/skills` install, context generation, package script detection, `/paseo-session-brief` recognized by a new Paseo agent
Intel macOS	Tested	Paseo CLI 0.2.5 detected `Darwin/x86_64`, temporary HOME install, real `~/.agents/skills` install, context generation, package script detection, `/paseo-session-brief` recognized by a new Paseo agent
Windows	Tested	Windows 11 x64, Windows PowerShell 5.1, Paseo CLI 0.2.5: PowerShell installer, `-TargetHome` temporary install, real `%USERPROFILE%\.agents\skills` install, `/paseo-session-brief` recognized by a new Paseo agent. Native bash context script not verified
`/ps-spyware-check` on Apple Silicon macOS	Tested	On Paseo 0.3.0: temp HOME install, helper script run, fixture risk-pattern detection, recognition by a new Paseo agent, scanner dry-run. Required report I/O fail-closed and optional scanner failure continuation regression tests passed
`/ps-spyware-check` on Intel macOS	Tested	On Darwin x86_64 / Paseo 0.3.0: install, helper script run, fixture risk-pattern detection, recognition by a new Paseo agent, Gitleaks secret scan no finding
`/ps-spyware-check` on Windows	Tested	On Windows 11 x64 / PowerShell 5.1 / Paseo 0.3.0: native PowerShell static-search workflow regex compile, fixture risk-pattern detection, recognition by a new Paseo agent. Bash helper not verified on native Windows
`/ps-agent-cleanup` helper	Tested locally	Node regression tests covering bare dry-run, ordinary idle preservation, disposable marker/explicit ID/user-pattern restriction, running/initializing protection, Paseo 0.6 archive acknowledgement and post-archive list verification, workspace approval gate, no dangerous command execution
`/ps-agent-cleanup` on Windows	Previous install verified	On Windows 10.0.26200 x64 / Node v24.14.0 / Paseo 0.3.0: `%USERPROFILE%\.agents\skills` install and dry-run JSON. v2.5.2 policy regression verified by cross-platform Node unit tests
`/ps-share` on Apple Silicon macOS	Tested	Official skill validator, Node security/onboarding/help and `--help` regression tests, private repo create/reuse and public/irrelevant repo rejection, real private GitHub publish/read/auto-fetch/original SHA-256 comparison, isolated Codex/Claude install
`/ps-share` on Windows	Tested	Windows private checkout, PowerShell install, private GitHub connection, real TXT publish and remote file read, mobile GitHub preview

Intel Mac testing observed no install failure, agent recognition failure, or Intel-specific error. Windows testing passed the PowerShell installer, -TargetHome temporary home install, and real skill recognition. Temporary home installs isolate on -TargetHome or $env:USERPROFILE instead of $HOME.

Quick usage examples

Sharing artifacts with another computer and mobile

On the first share request, GitHub connectivity is checked and the authenticated account's private paseo_share repository is prepared automatically.

/ps-share Share this report so I can view it on my phone.

If GitHub authentication is missing, it first guides you to gh auth login --hostname github.com. To use Forgejo or a separate repository, use setup <repo-url> explicitly.

Then publish or fetch in natural language.

/ps-share Share this report so I can view it on my phone.
/ps-share Show me the file I just shared from another computer.
/ps-share Fetch the latest shared file into the current project and summarize it.

The publish result includes a share ID and clickable preview/download links. Private repository links require GitHub/Forgejo login in the mobile browser.

Driving a web UI directly

/ps-browser
Open the login page at https://example.com, inspect the form structure,
then verify with a screenshot whether a test account logs in.

A possible workflow:

Open a new tab -> page snapshot -> find input fields -> enter values
-> click the button -> wait -> verify again with snapshot/screenshot

Commanding several agents

/ps-orchestration
Split this feature into implementation, testing, and review, run them in parallel,
and have a review agent decide pass/fail at the end.

A possible workflow:

Decompose task -> create worker agents -> run in parallel
-> collect results -> review gate -> pass/block decision

Comparing answers from several models

/ps-orchestration
On this README direction, have GPT argue for it, OpenCode DeepSeek against it,
and Grok compare the two and produce a final judgment.

A possible workflow:

Define participant roles -> run agents in parallel
-> collect results -> create a judge agent
-> produce winner / merged plan / risks

Making a session brief

/ps-project
Assume I'm seeing this project for the first time and produce a brief I can work from.

Output scope:

Project -> Current State -> Instructions -> Commands
-> Project Map -> Risks -> Suggested First Moves

Setting up new project context

/ps-project
Assume I'm seeing this repo for the first time; read docs, CLAUDE.md, AGENTS.md,
and package scripts to build working context.

You can also prepare the same thing with shell scripts first.

./scripts/paseobility-doctor.sh
./scripts/paseobility-context.sh

Generated files:

.paseobility/
├── context.md        # summary of README/docs/instruction files
├── commands.md       # candidate install/dev/build/test commands
├── project-map.md    # map of key files/directories
└── bootstrap-log.md  # OS, arch, Paseo state, warnings

Pre-install spyware check

/ps-spyware-check
Is it safe to install https://github.com/owner/repo?
Focus on install scripts, secret access, remote code execution, and data exfiltration risk.

Core principles:

It reads the repo read-only and does not execute its code.
On macOS/Linux it uses the bundled helper, if present, to make a static report after a temp clone.
On Windows it can make a native static report with the bundled PowerShell helper.
If gitleaks, trufflehog, semgrep, osv-scanner, trivy, shellcheck, or yara are installed it uses them; otherwise it falls back to rg-based heuristics.
If a scanner is missing and the extra coverage is worth it, the agent may install it with the bundled installer and says what it installed (--dry-run shows the plan).
Results are classified by severity as High, Medium, and Info, and self-references inside scanner docs/regex are marked Info.
The helper report includes a Finding Summary that first shows High / Medium / Info counts and a verdict hint.
Results are summarized with a Low / Medium / High / Critical verdict plus file/line evidence.

Cleaning up test agents

/ps-agent-cleanup
First show a dry-run of inactive agents with disposable/test markers and test workspace candidates.
Preserve ordinary idle sessions and do not archive agents outside the IDs or patterns I specify.
Do not touch running agents, and only show workspace candidates before approval.

Core principles:

The default run is a dry-run and does not archive ordinary idle agents based on status alone.
--auto is allowed only for inactive agents with an explicit ID, a user-specified pattern, or a clear disposable/test/validation marker.
It never deletes, only archives, and always protects active agents.
Agents you name by ID or pattern are archived directly (--archive, no --yes); workspace archive needs an explicit ID and --yes.
It confirms the target ID and status in the archive response and re-queries that the agent disappeared from the active list.

`/ps-share`

Publishes files of any type up to 50 MiB as immutable artifacts to a dedicated private Git repository. Without GitHub Actions or a background daemon, each computer only runs local git fetch, rebase, commit, and push.

onboard: check GitHub authentication, then create <login>/paseo_share private repo or safely reuse it
setup: configure the per-machine name and the same remote repository
publish: one or more files in one commit; return a share ID, preview URL, and download URL for each
list / latest: view another computer's artifacts after remote sync
fetch: verify and copy a share ID or latest to the current computer; an identical existing file is reused, a different one is kept and the copy is written as name (2).ext (--force overwrites only when you asked to replace it)

Artifacts are stored under artifacts/<machine>/<year>/<month>/<artifact-id>/. Any file type up to 50 MiB is allowed (zip a folder first); secret-looking file names (.env, *.pem, *.key, SSH keys, credentials, secrets) and text content with private-key or access-token signatures are refused. Fetching checks the metadata schema, path containment, symlinks, file size, and SHA-256, skips malformed entries with a warning, and safely handles Windows CRLF and macOS/Linux LF conversion. If a push fails (offline, auth, hook), the commit stays in the managed clone and the next list, latest, fetch or publish pushes it automatically; do not publish the same file again just to retry.

The supported model is one user trusting their own private repository across their own devices. Multi-tenant exchange where untrusted multiple users share one repository is not supported. Sensitive information inside binaries is not inspected, and Git deletion still leaves history, so sensitive files must not be published.

Auto-onboarding aborts without attempting to change visibility, rename, delete, or overwrite when a same-name repository is public or is not a Paseo Share structure.

`/ps-orchestration` comparison mode

This skill gives the same problem independently to several agents, then a separate judge compares them and picks the final answer.

Mode	Purpose
Debate	One model argues for, another against, and a judge synthesizes
Competing Plans	Compare several design/modification plans
Competing Implementations	Compare implementation candidates in isolated workspaces

Key rules:

First read the notes from list_profiles and materialize the profile into create_agent arguments. If no profile fits, confirm with list_providers, inspect_provider, and, if needed, list_models.
Analysis-only work may use the same workspace.
A tournament that modifies files uses a separate workspace per participant.
The judge preferably uses a different provider from the participants.
Results are summarized as winner, runner-up, best merged plan, and risks.

`/ps-project`

Selects project summary and environment setup from a single entry point.

Summary mode: summary of the requested repo overview, commands, and handoff; writes or refreshes .paseobility/ context when it is missing or stale.
Setup mode: performs only the requested initial setup, environment change, or context generation.
Reuses information already confirmed and does not re-investigate the same documents in full.
Queries version/daemon/workspace only when the Paseo runtime is relevant. If a --cwd lookup fails due to a path alias, it uses the existing workspace ID.
A summary installs no dependencies and starts no servers. .paseobility/ is local state; if the repo does not ignore it, it is added to .git/info/exclude, not .gitignore.

The procedure for the mode you need is in the skill and its linked references.

`/ps-spyware-check`

Before installing a GitHub URL or local repo, statically inspects for spyware, malicious install scripts, secret theft, remote code execution, and supply-chain risk signals.

What it checks:

preinstall, install, postinstall, prepare in package.json
remote download/execute, hidden processes, and persistence in shell/PowerShell scripts
access to .ssh, .aws, .npmrc, browser profiles, keychain, and API tokens
obfuscation/execution patterns such as eval, Function, base64, EncodedCommand, child_process
GitHub Actions risks: pull_request_target, unpinned actions, secret exposure
optional scanner results: gitleaks, trufflehog, semgrep, osv-scanner, trivy, shellcheck, PSScriptAnalyzer, yara
the helper report severity classes: High, Medium, Info
the helper report Finding Summary count and verdict hint
self-reference patterns inside scanner docs/scripts

Key rules:

It does not execute the code of the repo under inspection.
It does not run dependency install/build/test.
It does not print sensitive values; it redacts around file/line/key names.
Even a clean result is phrased as "no high-risk signals in the static scan", not "absolutely safe".
Self-references caught in scanner documentation or regex descriptions are not hidden but classified as Info.

Third-party scanner note:

/ps-spyware-check calls external open-source scanner CLIs installed locally.
Paseobility does not bundle or redistribute those scanners' binaries or rule sets.

Bundled helpers:

# Cross-platform JSON security receipt
python skills/ps-spyware-check/scripts/spyware-check.py https://github.com/owner/repo --json

# macOS/Linux
skills/ps-spyware-check/scripts/spyware-check.sh --target https://github.com/owner/repo
skills/ps-spyware-check/scripts/install-scanners.sh --dry-run

# Windows PowerShell
.\skills\ps-spyware-check\scripts\spyware-check.ps1 -Target https://github.com/owner/repo
.\skills\ps-spyware-check\scripts\install-scanners.ps1 -DryRun

`/ps-agent-cleanup`

Safely cleans up Paseo's inactive agents and test workspaces.

What it checks:

a complete agent and workspace snapshot read through the installed Paseo client (including background workspaces when the daemon supports them)
the workspace list from paseo workspace ls --json
inactive agents limited by an explicit ID, a user-specified --pattern, or a disposable/test/validation marker
ordinary idle agents to preserve and active agents that are protected
explicitly specified agent/workspace IDs

Key rules:

The default run is a dry-run and does not archive ordinary idle agents based on status alone.
--auto archives only inactive agents with an explicit ID, a user-specified pattern, or a clear disposable/test/validation marker.
Agents in running, initializing, working, active, starting, queued, pending, busy, executing, or in-progress states are not archived.
It does not use delete/stop/kill/restart or history/timeline/resume verification.
After archiving it verifies the Paseo CLI JSON acknowledgement and removal from the active list.
Agent archive (a soft delete) needs only --archive; workspace archive needs an explicit ID and --archive --yes.

Bundled helper:

node skills/ps-agent-cleanup/scripts/agent-cleanup.js
node skills/ps-agent-cleanup/scripts/agent-cleanup.js --dry-run --pattern 'cleanup-validation|fixture'
node skills/ps-agent-cleanup/scripts/agent-cleanup.js --agent <agent-id> --archive
node skills/ps-agent-cleanup/scripts/agent-cleanup.js --auto --agent <agent-id>
node skills/ps-agent-cleanup/scripts/agent-cleanup.js --auto --pattern 'cleanup-validation|fixture'
node skills/ps-agent-cleanup/scripts/agent-cleanup.js --workspace <workspace-id> --archive --yes

`/ps-browser`

Default: Ego Lite → one TaskSpace → reuse its p1 → finish. Read ego-browser skill for the installed API contract; see the Ego reference for multi-round reuse, completion receipts, user control, and shell portability. Do not run Chrome's helper against Ego. Failed tasks retain their space for recovery rather than opening a replacement space or silently switching browsers.

Missing Ego, a CLI/PATH gap or a connection failure does not select Chrome. Check the target computer's CLI and app paths using the Ego reference, then report any setup gap. Updated skills require a new agent session or integration reload on each affected computer; restarting Paseo is unnecessary.

For explicitly selected Chrome: existing Chrome window → one task tab → reuse → close only that tab. The macOS scripts/chrome-tabs.py adapter targets an explicit existing window, records tab ownership and Chrome incarnation, and recovers a lost creation reply without making another tab/window. Cleanup is verified by re-listing.

python3 ~/.agents/skills/ps-browser/scripts/chrome-tabs.py list
python3 ~/.agents/skills/ps-browser/scripts/chrome-tabs.py open --task '<unique-task-id>' --window '<existing-window-id>' --url 'https://example.com/'
# Bind the returned tabId through the installed Chrome integration for page work.
python3 ~/.agents/skills/ps-browser/scripts/chrome-tabs.py close --task '<unique-task-id>'

User tabs, login state and other agents' work are preserved. An unhandled provider/process crash requires journal recovery on resumption; no watchdog is installed. Existing-window adapters outside macOS are not yet verified. A connection failure never silently launches another browser.

CloakBrowser, Paseo workspace browser and Playwright CLI are requested alternatives for isolation/headless work, or required by a specific workflow such as ps-design verification. Installation starts no browser or daemon. See the skill, Chrome lifecycle reference, CloakBrowser recipes and Playwright CLI recipes.

`/ps-cua`

Drives a native desktop GUI through the trycua Cua Driver (cua-driver CLI or MCP) when the user asks to operate a native app or a task can only be done through a desktop app's GUI; a CLI, API or file edit that reaches the same result is preferred. Preview / limited validation — see the Cua platform validation status. Current dependency pins and host checks: 2026-11-06.

Preconditions: cua-driver --version. If missing, it uses a non-GUI route when one exists, otherwise gives the install command and runs it if you ask. macOS Accessibility and Screen Recording grants stay with the human.
Read the installed contract first: cua-driver list-tools, cua-driver describe <tool>.
cua-driver mcp-config --client opencode only prints a config snippet; it does not register a server.
Core loop: find/launch window -> snapshot -> act (several actions may chain from one snapshot while the UI is unchanged) -> verify with a fresh read.
Background delivery is the default; if an action does not land it switches to delivery_mode: "foreground" and says so.
Acts freely inside the task; asks first before spending money, irreversible deletion, account-security changes, or sending/publishing something you did not ask for.

`/ps-orchestration`

A pattern collection where one agent creates several agents, splits roles, and synthesizes results.

Pattern	Purpose
Fan-out	Give independent work to several agents at once
Task DAG	Step 1 -> Step 2 -> Step 3 sequential execution
Hybrid DAG	Mixed flow such as parallel work then synthesis then review
Decision Gate	A review agent decides pass/block before the next step
Coordinator Loop	Periodically check long work with a heartbeat
Blocking Ask/Reply	Proceed after receiving another agent's answer
Escalation	Clearly report permissions, ambiguity, and hard failures to the user

Recommended composition:

Codex       -> implementation / refactoring / test writing
Claude      -> review / UX copy / risk check
Coordinator -> decomposition / progress management / final synthesis

Safety rules:

If several workers may modify the same file, create a separate workspace.
Read the notes from list_profiles first, then materialize the chosen profile into provider/settings. Do not reuse stale provider strings verbatim.
Set maxRuns or expiresIn on heartbeats and schedules.
On failure, retry with a changed approach or reassign to another profile/provider; stop after a repeated same-cause failure and report it.
Approve worker permission requests within the task. Ask first before pushing, merging to shared branches, deploying, publishing, spending money, or irreversible deletion.
When done, archive the agents and workspaces the workflow created unless they hold unintegrated work or you want them kept.

Optional extra guidance

In Paseo 0.6, the agent profiles configured in the app are the basis for provider/model/mode/thinking/feature selection. The legacy file below can be used to pass extra user guidance, not as a provider source.

~/.paseo/orchestration-preferences.json example:

{
  "preferences": [
    "Write task instructions as a self-contained briefing.",
    "Prefer a different provider for review agents than the implementing agent.",
    "Do not silently bypass hard failures; report them to the user."
  ]
}

Repository layout

skills/
├── ps/                     # SKILL.md + read-only local catalog/tests
├── ps-agent-cleanup/       # SKILL.md + CLI helper/tests
├── ps-browser/            # SKILL.md + references (cloakbrowser, paseo, playwright, screenshots)
├── ps-cua/                # SKILL.md + implicit-invocation policy
│   └── references/           # setup.md, workflow.md, recovery.md
├── ps-design/             # SKILL.md + references, templates, scripts/design-check.mjs, vendored guides
├── ps-god/                # SKILL.md + explicit-only policy, procedures, run helper/tests
├── ps-multi/              # SKILL.md + host adapter/tests + references
├── ps-orchestration/      # SKILL.md + explicit-only policy
│   └── references/           # coordination.md, tournament.md
├── ps-plan/               # SKILL.md + explicit-only policy
├── ps-project/            # SKILL.md
│   └── references/           # brief.md, setup.md
├── ps-research/           # SKILL.md + primary-source workflow
├── ps-share/              # SKILL.md + CLI helper/tests
├── ps-spyware-check/      # SKILL.md + scanners/tests
├── ps-toss/               # SKILL.md + pinned upstream Harness
└── ps-write/              # SKILL.md + explicit-only policy, Korean style reference and vendored rules
scripts/                     # installers, doctor (.sh/.ps1/.py), context, cua-driver helpers
tests/                       # test_doctor.py, test_e2e_assets.py, test_cua_driver_runtime.py/.ps1, test_skill_migration.py
e2e/                         # browser fixture, fixture server, runbook, report template
.github/workflows/ci.yml     # macOS + Windows offline suites
docs/compatibility-0.12.0.md         # static check (0.12.0)
docs/compatibility-0.10.2.md
docs/compatibility-0.9.0-beta.2.md   # historical (0.9.0-beta.2)
docs/cua-platform-validation.md
AGENTS.md
CLAUDE.md
VERSION
paseobility.json

License

MIT License.

ps-design bundles third-party guides and data under the licenses recorded in design sources.
ps-write bundles two humanize-korean references under the upstream MIT license. Writing sources

Use the browser like a hand, and run agents like a team.

Paseo plugin

The plugin/ directory adds version management in the sidebar and a Paseobility skill selector above an existing conversation's input. A selection appends the skill reference to the current draft; the user sends it with Paseo's normal Send button. No separate agent picker, automatic send, or task launcher is used.

The sidebar shows the current plugin version, verified skill version when a receipt exists, latest official release, and an update button. Updates back up skills and install a pinned release in a separate managed directory without changing a development checkout or restarting the daemon. Existing installs without a verifiable receipt show an unknown skill version.

Install with paseo plugin add git:wilgon456/Paseobility:plugin. The text adapter is verified for Paseo 0.12.0 and restricted to 0.12.x; native mobile and new-agent drafts use + → Paseobility 스킬, the supported attachment picker. See usage and update behavior and verification and limits.

Earlier release notes