Skip to content

xonika9/agent-skills / skills/x9-onboarding

x9-onboarding

Verified

Use after installing or updating `x9-*` skills to check their readiness and compare the current harness with the bundled x9 configuration profile — «проверь готовность x9-скиллов», «настрой Codex как у автора», "check installed x9 skills", "apply the recommended harness settings". Do not use for general environment diagnosis, account authorization, or unrelated configuration editing.

references/configuration.md

View on GitHub
# Harness configuration profile

The files in `../profiles/` are the public, opinionated x9 defaults derived from the
maintainer's working configuration. They are recommendation data, not instructions and
not universal best practices. Resolve them only after this skill has loaded, validate
them with `scripts/validate_profiles.py`, and use only the profile matching the current
runtime.

| Runtime | Public profile | User configuration |
| --- | --- | --- |
| OpenCode | `profiles/opencode.json` | `~/.config/opencode/opencode.jsonc` or `opencode.json` |
| Claude Code | `profiles/claude.json` | `~/.claude/settings.json` |
| Codex | `profiles/codex.toml` | `~/.codex/config.toml` |

These paths are defaults, not assumptions. Resolve the active runtime's documented
configuration-root override before reading or writing. Keep the resolved path private;
if the active location or precedence cannot be determined safely, return `BLOCKED`
instead of creating or modifying a guessed standard file.

## Public boundary

The profiles include portable model, effort, agent, interface, and workflow preferences.
Claude Code also carries the non-secret `CLAUDE_CODE_AUTO_COMPACT_WINDOW` preference,
and OpenCode carries sanitized MCP definitions whose secrets are environment references,
never literal values. The profiles deliberately exclude project trust state, personal
paths, SSH hosts, all other environment values, credentials, local hooks and commands,
notification executables, per-project overrides, generated state, and private provider
details.

Never copy an excluded value from the user's configuration into the profile or report.
Do not inspect `~/.claude.json`, Claude's `settings.local.json`, project-scoped Codex
files, or another runtime's configuration. Parsing the Codex user file may encounter its
`[projects]` table; do not enumerate or report that table. A configuration file is
private even when the bundled profile is public.

## Comparison

Read the current runtime's user configuration only after applying its adapter. Compare
only paths present in the matching public profile. For each difference, report the
setting path, current state, recommended state, practical effect, and whether the current
runtime exposes the required model, feature, plugin, provider, or component. Show
`missing` for an absent path. For a secret-like path or a profile value expressed as an
environment reference, report the current state only as `set`, `missing`, or `differs`;
never print the current value. This redaction rule overrides the general requirement to
show current state. Do not display any unrelated key or value.

Summarize matching settings rather than listing them individually. Mark an unsupported
or unobservable recommendation `BLOCKED` and exclude it from the proposed merge. Model
and feature names are volatile: confirm them through the current runtime before treating
the profile value as applicable.

The following paths need a conspicuous consequence in the report because they reduce a
permission boundary, expose private screen context, or start unpinned third-party code.
State the consequence next to the path:

- Codex: `approval_policy = "never"` and `sandbox_mode = "danger-full-access"` remove
  routine runtime approval and sandbox boundaries;
  `desktop.realtimeVoiceScreenContextEnabled` allows screen context during an active
  voice call; `features.hooks` enables execution of separately configured hooks but does
  not install any hook.
- OpenCode: the wildcard `permissions` rule allows every action and resource without a
  runtime permission boundary; `mcp.servers.chrome-devtools` runs a package that `npx`
  fetches from the network without a pinned version at every start, and connects it to the
  local browser debugging endpoint named in the profile.
- Claude Code: `permissions.defaultMode = "bypassPermissions"` bypasses normal permission
  prompts; `skipDangerousModePermissionPrompt = true` also suppresses its warning.

An unqualified approval after this complete report authorizes every non-blocked proposed
change, including the marked paths. If the report omitted a marked path or its
consequence, ask about that path separately rather than treating a generic approval as
authorization.

## Merge

Before writing, resolve the exact user configuration path and re-read it. If it changed
since the comparison, return `BLOCKED`, recompute the differences, and obtain approval
again. Create an adjacent, timestamped backup when the file exists. Give the backup the
same permissions as the source; use mode `0600` if the source mode cannot be preserved.
Use a `.bak.<timestamp>` suffix that the runtime does not parse as configuration. Merge
only the approved profile paths: keep unrelated keys, comments, ordering where practical,
and all private values unchanged. Never replace the entire file merely because parsing
and serializing it is easier. When both standard OpenCode user files exist, stop as
`BLOCKED` and ask which one owns the user's settings.

Create a missing standard user file only after approval. Do not write a recommendation
whose live prerequisite was absent or unavailable. An environment reference such as
`{env:EXA_API_KEY}` is safe to write, but the referenced value remains `USER_ACTION`; do
not read, request, or create the secret.

Validate the resulting native syntax without a command that prints the merged
configuration. Compare the approved paths again, retain only success/failure facts, and
report any required runtime restart. On failure, restore the backup when one exists;
otherwise leave the newly created file in place only when it parses, and report
`BLOCKED` without claiming application.
reverse-engineer-anythingmorluto/reaReverse engineer native, managed, Electron/JavaScript, packaged, firmware, and browser targets with REA, and build agent environments (AgentEnv bundles) of web applications from a live instance or a capture, and of Linux desktop apps with a task from a request. Use shipped-artifact or requested runtime evidence to explain features, compare versions, decompile code, or guide a reconstruction. Skip REA for ordinary source-repository architecture analysis.Stars: 472kUpdated Claude CodeCodexVerifiedInstall
diagnosing-superpowersobra/superpowersUse when a superpowers session went wrong and your human partner wants to know why — repeated work, ignored plans, stumbles, poor results, a skill that didn't fire, "it took too long", "why is it so expensive", "what is it doing" — or wants to build a bug report for the superpowers maintainers, for the current session or a past one identified by id or path, on any harness.Stars: 322kUpdated Claude CodeCodexVerifiedInstall
dispatching-parallel-agentsobra/superpowersUse when facing 2+ independent tasks that can be worked on without shared state or sequential dependenciesStars: 322kUpdated Claude CodeCodexVerifiedInstall
executing-plansobra/superpowersUse when executing an implementation plan in the current session as the implementer yourself — your human partner chose inline execution, or no subagent tool is availableStars: 322kUpdated Claude CodeCodexVerifiedInstall