Install dev-skills

Install from curvedinf/dev-skills, preserving existing instructions and customizations.

Required opening: ask, then wait

After reading this skill, your first response to an unscoped installation request must ask:

Where would you like Dev Skills installed?

1. Project-wide — in the target repository, shared with collaborators.
2. Agent-wide — in your agent’s personal skill library, available across projects.

Reply 1 or 2. This installer requires your choice before proceeding.

Ask through an interactive question tool or a normal reply. Stop here until the user answers. Sending a question is not receiving a choice: do not inspect destinations, clone/download the collection, copy files, or modify instructions while waiting. A tool returning no answer, a preselected option, or elapsed time leaves the choice pending.

Do not infer scope from the working directory, an open repository, the agent host, an existing installation, or the phrase “install dev-skills.” Skip the question only if the user has explicitly selected project-wide, agent-wide, 1, or 2 for this installation; acknowledge that selection before proceeding. If the answer is ambiguous, clarify it. Record the selected scope in the installation record.

After the scope is selected

1. Inspect the destination. For option 1, read applicable project instructions, skills, and Git state; resolve the project root and existing layout, defaulting to agents/skills/<passive|activated|triggered>/<name>/. For option 2, resolve the selected host’s configured user-wide skill directory, discovery format, and existing conventions skills. Do not mistake the current project’s agents/ for a global directory; resolve host ambiguity before writing.
2. Acquire and reconcile. Fetch the requested revision, otherwise upstream main, and record the resolved commit. Read source AGENTS.md; copy complete skill folders, adapters, supporting resources, and license attribution. Exclude upstream .git, scratch history, and temporary files. Preserve category layout where supported; when flattening, map destinations and repair inter-skill links. Merge updates against the prior source revision and local customizations; inspect unknown same-name collisions before replacement. Keep changes recoverable and reinstallations idempotent.
3. Integrate guidance for the chosen scope. Apply the destination rules below. Merge by topic, preserve explicit target policy, deduplicate equivalent guidance, and report unresolved conflicts. Adapt repository-specific assumptions and repair imported relative links, using revision-pinned source permalinks for uncopied material. Keep activated workflows deliberate and link passive conventions within their scopes. Wire triggered security-hygiene-check before every project commit, including amendments and merges: inspect the final candidate and apply the consuming project’s natural-language scratch disclosure policy and wait for explicit user disposition of uncovered findings before committing; never import upstream disclosure permissions, independently of self-improvement opt-in. Wire triggered maintain-scratch into the first-project-context event: once per fresh session check its seven-day entry in shared agents/scratch/timers.json, compact if due, and stamp only completion. Preserve all existing timer entries; initialize missing entries from declared defaults with null completion. Migrate a legacy maintenance.json into the maintain-scratch entry without losing its interval or timestamp; remove the legacy file only after verifying migration, and reconcile conflicts if both exist. Never import upstream completion history. Agent-wide installation creates timers only in consuming projects. Always load triggered refine-skill as a metaskill and wire its check to completed skill use, coalescing related uses before the final response; preserve existing enablement until the final choice, never inherit upstream’s opt-in. For a project-wide installation, also insert the harness opt-in section declared in context-paging into the project AGENTS.md and execute it during installation: skip when its cp_subscribed.py gate exits 0 or agents/tmp/context-paging.json records "harness_install": "declined"; otherwise ask the user, and on yes code and install the harness plugin per that skill this session, subscribing last. Agent-wide installations neither insert the section nor subscribe; each project’s own opt-in governs. Wire triggered context-compaction-procedure into the same fresh-session startup that loads maintain-scratch: project AGENTS.md directs reading it once per session before substantive work, where it overrides harness compaction instructions when context-paging is enabled.
4. Verify and record. Check inventory, frontmatter, invocation settings, resources, links, and request-term routing. Verify host discovery where available; distinguish installation from proven loading and report reload requirements. Record source URL/commit, destination mapping, adaptations, and conflicts in project agents/scratch/dev-skills-install.md for option 1, or a maintenance reference beside the conventions skill for option 2. Installing Git guidance does not itself authorize committing or publishing the installation; follow existing authorization.
5. Ask the final setup question. After installation and verification, ask: “Would you like skill self-improvement enabled? It refines skills from corrections and reusable lessons after skill use.” Name the scope and exact AGENTS.md to update; explain this installer requires the final opt-in choice. For option 1 use the project file; for option 2 use a host-supported agent-wide AGENTS.md, or ask which project file should hold the choice if no global file is loaded. Set Self-improvement: enabled. for yes or Self-improvement: disabled. for no. Retain the installed refine-skill link and skill-completion check, reconcile duplicate declarations, and verify the saved choice. Project declarations override agent-wide defaults. No answer preserves the prior setting and leaves the choice pending; an explicit choice already supplied satisfies the question.

Destination rules

• Project-wide: create or merge the project’s AGENTS.md with request-term routing, scoped passive defaults, authoring conventions, scratch policy, and reference libraries. Preserve project commands and constraints. Explicit skill links work without assuming native host discovery. Initialize or reuse Git-versioned agents/scratch/ and its index; add /agents/tmp/ and /agents/context-paging/ to the root .gitignore without altering unrelated rules.
• Agent-wide: merge source AGENTS guidance into existing agent-level conventions/routing skills, or create passive dev-skills-conventions when no owner exists. Link procedures rather than copying them. Configure supported discovery/startup references; if automatic loading is unavailable, provide explicit activation instructions. Keep project-specific precedence and unrelated policy intact. Project scratch remains Git-versioned agents/scratch/, with ignored agents/tmp/, never global session history. Project opt-in alone does not authorize global skill edits.