skillZs
★ LIVE SKILL TAGS ★
>>> LIVE SKILLS INDEX <<<
* OPEN SOURCE *
NO LOGIN, NO TRACKING
※ REAL INSTALL DATA ※
← back to all skills
marcioaltoe/skills51 installs

setup-context-driven

Adopt, update, or verify a repository's Context-Driven Baseline through the public Roundfix Baseline Command. Use when preparing a repository for the write-prd/write-tasks/implement pipeline, changing its Baseline Profile, preserving existing instructions, restoring its Repository Skill Set, or confirming that Baseline-owned content is current.

How do I install this agent skill?

npx skills add https://github.com/marcioaltoe/skills --skill setup-context-driven
view source ↗

Is this agent skill safe to install?

  • Gen Agent Trust Hubpass

    The skill facilitates repository configuration management using the 'roundfix' CLI. It follows security best practices by requiring human review of generated plans and cryptographic digests before applying changes.

  • Socketpass

    No alerts

  • Snykpass

    Risk: LOW · No issues

What does this agent skill do?

Setup Context-Driven

Use the public roundfix baseline command family. The Roundfix binary is the only runtime authority for Baseline Profiles, repository inspection, Decision Plans, Change Plans, confirmation, apply, recovery, and Baseline verification. This skill supplies recipes and interpretation only. It has no independent setup engine or behavioral fallback.

Read the durable operating contract before the first adoption or a recovery: docs/user-guide/context-driven-development.md#adopt-or-update-the-context-driven-baseline.

Root instruction file

AGENTS.md is the root instruction file; CLAUDE.md is always a relative symbolic link to AGENTS.md, with no repository option to keep a separate file. Adoption creates the link or replaces a regular file or a safe link elsewhere. A managed refresh creates an absent link and replaces a regular file or safe link elsewhere only when it holds no passage of its own. A replaced regular file gets the root backup CLAUDE.<digest>.md; a link's target stays untouched.

When passages remain, the plan leaves CLAUDE.md unchanged and refreshes the rest. Text prints this warning; JSON carries its code, path and message in warnings:

Warning: baseline.root.claude-link.withheld: CLAUDE.md: CLAUDE.md holds <n> passage[s] of its own (<path>:<start>-<end>[, ...]); move each into docs/agents/specific-repository.md, then run roundfix baseline update to replace it with a symbolic link to AGENTS.md

Use 1 passage or <n> passages. For a link elsewhere the message begins CLAUDE.md links to <target>, which holds. When AGENTS.md is not a regular file, the same code instead says:

AGENTS.md is not a regular file, so CLAUDE.md cannot link to it; make AGENTS.md the root instruction file, then run roundfix baseline update

An unlinked checkout holds exactly AGENTS.md as regular-file bytes while Git records mode 120000. The plan leaves it unchanged and warns:

Warning: baseline.root.claude-link.unlinked-checkout: CLAUDE.md: this checkout holds the symbolic link CLAUDE.md -> AGENTS.md as a regular file; run git config core.symlinks true, then check CLAUDE.md out again

Both warnings preserve the update's exit code. Move every passage into docs/agents/specific-repository.md before the next update converts the file. The command never moves a passage itself.

When a plan warns baseline.root.claude-link.withheld, report each passage to the person and move it into docs/agents/specific-repository.md only with the person's approval; the next update then replaces CLAUDE.md with the link.

Who starts this skill

An Agent may start this skill itself when its dispatch trigger matches; its metadata does not turn off model invocation. Starting it authorizes no write. Present the Plan and its Plan Digest first, and pass --confirm-plan <digest> or --yes only after the person approves that Plan Digest or the repository's own instructions already authorize it.

Boundaries

Baseline owns declared root blocks, setup-owned guides, first-adoption root-instruction backups, the Claude Instruction Link, and docs/agents/setup-context.json. Except for the approved conversion of CLAUDE.md described above, it preserves repository-authored bytes outside managed boundaries.

Baseline never:

  • installs application or Agent dependencies;
  • connects to a database or live infrastructure;
  • runs repository formatter, build, test, lint, migration, or Verification commands;
  • follows unsafe links or mutates nested instruction carriers;
  • infers repository policy from implementation evidence;
  • lets ACP proposals authorize mutation;
  • combines profiles or loads user-scoped or remote profiles.

Treat profile expectations, repository commands, and recommendations as different facts. A profile expectation is portable policy, not implementation proof. A repository command is executable only when the result says repositoryExecutable: true and binds a local declaration. A recommendation is a command for the maintainer to run; Baseline did not execute it.

Interpret composed guidance

Read the generated root Instruction Hierarchy from universal instructions through context and documentation, Spec workflow, enabled autonomous work, stack guidance, surface guidance, and optional knowledge sources. A narrower guide may add constraints for its concern but cannot weaken a universal Normative Clause or confirmed project decision.

Greenfield composition emits only guides and pointers selected by the Baseline Profile. It creates no generic repository guide and no residual carrier when there are no Repository-Specific Normative Rules.

When Greenfield retains managed source that needs classification, Greenfield refuses and names preservation.mode=preservation as the route to choose.

For update or Baseline Readoption, interpret the complete retention ledger before approving it. Every existing rule must retain its exact source bytes in one active semantic owner, a recognized typed repository document, or docs/agents/specific-repository.md, or have an individual reasoned rejection or non-governed classification. An empty residual removes the carrier and root pointer; non-empty accepted residuals retain only their exact rules.

Generated docs/agents/docs-layout.md owns the copyable ADR and Findings contracts. Only accepted is active for ADRs; legacy ADRs without lifecycle frontmatter remain active unless their body marks them inactive. Findings use pending, partial, deferred, and done, plus the terminal statuses deprecated, superseded, closed, and cancelled, keep the original observation, and append later evidence as dated addenda. Copy those generated templates; this skill does not render or replace them.

Refresh an adopted repository

When the repository has a compatible Setup Manifest, use the non-interactive managed-refresh path:

roundfix baseline update --repo . --format json

Without confirmation, a changed Plan is presented and repository bytes remain unchanged. For a fleet sweep, approve the Plan Digest computed in the same invocation:

roundfix baseline update --repo . --yes --format json

The command derives the current Baseline Profile and compatible decisions from the Setup Manifest. It asks no settled question and does not invoke semantic classification. It changes Baseline-managed regions, the Repository Skill Set and, when no passage would be lost, CLAUDE.md into the symbolic link. Every other byte outside managed boundaries remains byte-identical.

Use --confirm-plan <digest> instead of --yes for a previously reviewed Plan Digest. A decision absent from the manifest exits 3 without writing unless the maintainer explicitly chooses --adopt-suggested. Use --no-skills only to suppress the Repository Skill Set refresh, and --skills-source-dir <path> only for a declared offline Git checkout or bare object store.

Repositories select their fast local command through verification.incremental, separately from the complete verification.gate. A single-gate Setup Manifest exits 3, names the missing incremental decision, and writes nothing. After the repository declares the suggested make verify-incremental command locally, rerun update with --adopt-suggested; later updates use the recorded decision without that flag.

A repository-owned Baseline Profile that does not list a decision a template it renders requires is refused first: baseline update exits 3 and names the decision, the entry to add to the profile's decisions and a --decision flag, and baseline profile validate reports the same gap with exit 2. Add the entry to the profile, then rerun the update with --adopt-suggested or that flag.

Human adoption or profile change

At an interactive terminal, run:

roundfix baseline --repo . --format text

The interactive workflow handles first adoption, a Baseline Profile change, or a decision the Setup Manifest does not carry. It selects exactly one Baseline Profile and shows one consolidated Change Plan with file changes first. With a compatible Setup Manifest it skips settled decisions and semantic classification. Do not authorize mutation before the complete plan and Plan Digest are visible.

Decision and preservation prompts mark one visible default; Enter confirms it. A valid stored Setup Manifest value wins even when a changed Profile Digest requires adoption again. Otherwise the CLI uses its embedded catalog suggestions, including codex gpt-6.1-sol high for backend work and claude opus high for design work. Existing root instructions default to Preservation, an empty instruction inventory defaults to Greenfield, and a recoverable existing profile is preferred. Classification, Plan approval, and apply still require an explicit non-empty choice.

For first adoption, ask the maintainer to choose one instruction mode:

  • Greenfield backs up safe root carriers and imports none of their rules.
  • Preservation backs up safe root carriers and requires one reviewed disposition per Source Baseline Entry.

Nested instruction carriers remain unchanged and appear only as warnings. Unsafe root carriers block planning. If sealed ACP classification is unavailable or discarded, continue through the public command's structured manual classification review. Never invent Source Baseline identities, classifications, or destinations.

Repository-Specific Normative Rules have one canonical carrier: docs/agents/specific-repository.md. The enabling decision permits this destination; it does not create an empty scaffold. Baseline creates the file and its managed root pointer only for approved non-empty rules. It migrates either legacy docs/agents/repository.md or docs/agents/repository-rules.md byte-for-byte, removes the known empty legacy scaffold, and blocks divergent non-empty legacy carriers for manual reconciliation. Existing 0.0.1 Decision Documents that name either legacy destination are normalized to the canonical path.

For a managed refresh, use roundfix baseline update; a moved Profile Digest alone does not force a new interview when the profile and stored decisions still resolve. A profile change produces a new full interactive plan. If the maintainer rejects the plan, use the command's decision-area revision flow. Every accepted correction must produce a new Plan Digest and another complete review.

When roundfix baseline update or roundfix doctor names a Misplaced Repository Rule, report each passage to the person and move it into docs/agents/specific-repository.md only with the person's approval.

Profile alignment runs before instruction classification. For profile-specific blockers, the interactive command offers Change Baseline Profile, a reviewed repository-owned Profile adaptation, or Decline without writing. An adaptation lists every module and profile-specific capability removal, validates a repository-owned Profile ID, re-audits the in-memory draft, and proceeds only when alignment is ready. The Profile file is part of the final Change Plan and is not written before Plan Digest approval.

Universal required capabilities cannot be removed or waived. Run the exact restoration preview named by the result, review its Plan Digest, confirm the same current preview, then rerun Profile alignment:

roundfix baseline skills restore --repo . --profile <built-in-id> --skill <skill-name> --format json
roundfix baseline skills restore --repo . --profile <built-in-id> --skill <skill-name> --confirm-plan <digest> --format json

Project decisions and Spec constraints

The CLI owns project-decision collection and rendering. This skill does not collect, derive, validate, or render decisions; it explains the public result and sends every correction back through roundfix baseline.

verification.incremental names the selected fast local check that generated guidance publishes. It is required independently of verification.gate, which names the complete repository Verification.

Changing only the HTTP mode retains every exception and the source.

For the Standard TypeScript Monorepo Profile, UUID version 7 is a visible suggestion for identifier.strategy. The human must explicitly keep it or provide one non-empty repository-defined rule. When Better Auth remains selected, auth.provider proposes GET and POST under /api/auth/*, owner Better Auth, with the Session, OAuth redirect, callback, and related provider protocol reason. The human must keep or change that complete typed proposal. /api/auth is Better Auth's default base path, not a Baseline rule: a repository that mounts Better Auth elsewhere records the optional auth.base_path decision, and the provider and HTTP exceptions then cover that path followed by /*.

Automation supplies both structured values through the strict Decision Document passed to --decision-file. If either value is unresolved, planning exits 3, names every missing decision in roundfix/baseline-result/v1, emits no partial Plan, and writes nothing. Never treat a catalog suggestion, repository evidence, setup approval, or an empty answer as authorization.

Every new PRD and TechSpec must contain Project Constraints rows for:

  • Identifier strategy from docs/agents/domain.md;
  • Authentication and HTTP from docs/agents/backend.md;
  • active ADR obligations from docs/agents/spec-routing.md; and
  • tooling authority from docs/agents/agent-instructions.md.

Each row must say applicable or not applicable with a reason. Protected tooling work remains blocked until the Spec records express maintainer authorization and the exact bounded repository-relative files. Task assignment and generic implementation requests do not grant that authority. Keep completed and archived legacy Specs byte-identical.

Non-interactive planning for adoption or profile change

For a profile change, baseline plan --profile <id> with preservation.mode carries each valid Carried Decision and names all missing answers in one refusal. Apply names each Root Backup with a baseline.preservation.backup.written warning; reconcile names the Setup Snapshot revision.

The interactive root command refuses redirected or absent terminal input. Automation and Agents use baseline update for an adopted repository. For first adoption or a Profile change, use:

roundfix baseline plan --repo . --profile <profile-id> --decision-file <decision-file> --format json

Automation can pass the equivalent reviewed strict draft:

roundfix baseline plan --repo . --profile-file <draft.json> --decision-file <decision-file> --format json

--profile-file and --profile are mutually exclusive. The CLI validates the draft against the embedded catalog, resolves it in memory, and includes its canonical repository path and exact bytes in the portable Plan. Do not write, normalize, classify, or render the draft through this skill.

Scalar answers may use repeatable --decision <id=value> flags. Structured answers use repeatable strict Decision Documents with schema setup-context-driven/decisions/0.0.1.

Interpret the result by exit category:

  • 0: stdout is one complete roundfix/baseline-plan/v1 document.
  • 2: repair invalid input, schema, Git state, or an unsafe carrier.
  • 3: follow the roundfix/baseline-result/v1 nextAction; no partial plan exists.

For preservation, use only a current, strict Decision Document derived from the public human classification review. It must bind the current Source Baseline and every Source Baseline Entry exactly once. Do not reuse identities or digests after repository bytes change.

Review at least:

  • repository, catalog, and the one resolved profile;
  • normalized decisions;
  • fileChanges and the complete canonical managedEntries ledger;
  • retention, warnings, and the proposed setupManifest;
  • when the plan relocates history, historyMoves and the baseline.history.citation warnings;
  • every exact preimage and postimage;
  • planDigest.

The plan can contain repository policy and exact generated bytes. Store and transport it as a sensitive review artifact.

Apply the approved plan

Ask the maintainer to approve the exact current planDigest. Then run:

roundfix baseline apply --repo . --plan <plan-file> --confirm-plan <plan-digest> --format json

Apply strictly parses the supplied roundfix/baseline-plan/v1 document and applies only its postimages. It never recalculates or substitutes a plan.

On exit:

  • 0: apply and Baseline verification passed, or the exact postimages were already verified.
  • 1: inspect execution, verification, recovery, output, or rollback failure.
  • 2: correct invalid input, plan schema, or an unsafe repository.
  • 3: confirmation, preimage, catalog, profile, or Git lineage is no longer current. Generate and review a new plan.
  • 130: the operation was canceled.

A portable plan can apply in another clone only when clone-stable Git lineage, object format, catalog identity, profile, and all bounded preimages match.

Baseline Profiles

Inspect and validate profiles through the public CLI:

roundfix baseline profile show <profile-id> --format json
roundfix baseline profile validate <profile-id> --format text

Create one repository-owned profile from one built-in profile:

roundfix baseline profile init --id <repository-profile-id> --from <built-in-id>
roundfix baseline profile validate <repository-profile-id> --format text

Repository-owned profiles live only under .roundfix/baseline/profiles/<id>.json, compose only embedded catalog entry IDs, and must be versioned with the repository. They cannot combine profiles, load user-scoped catalogs, or declare executable or remote content.

Repository Skill Set restoration

Restore preserves the Claude Skill Link layout, uses create-link for a missing link, repairs links for installed skills, and never moves an installed skill.

Adoption reports missing or drifted external skills but never restores them as a side effect. Preview an explicitly requested restoration:

roundfix baseline skills restore --repo . --profile <built-in-id> --skill <skill-name> --format json

A non-empty preview exits 3 with a Plan Digest and writes nothing. After the maintainer approves that exact preview, rerun:

roundfix baseline skills restore --repo . --profile <built-in-id> --skill <skill-name> --confirm-plan <plan-digest> --format json

Use --source-dir <path> only for a declared offline Git checkout or bare object store containing the exact immutable source commit. Never substitute a generic skill refresh.

After restoration, generate the same Baseline Plan again. Restoration success proves selected Repository Skill Set bytes, not repository-authored policy.

Canonical asset synchronization

This is a maintainer-only operation and is never part of ordinary adoption:

roundfix baseline assets sync --source-dir <canonical-setups> --check --format json

Run without --check only when the maintainer explicitly authorizes refreshing the Go-owned canonical setup snapshots. The command validates source provenance and the generated catalog before changing only canonical setup assets.

Recovery

  • Missing decisions or manual classification: follow nextAction, complete current structured input, and rerun planning.
  • Confirmation mismatch: copy the full current Plan Digest from the supplied plan after review.
  • Stale plan: generate and approve a replacement; never edit or force the old plan.
  • Unsafe carrier: repair the exact reported path; never follow or replace it through an ad-hoc workaround.
  • Interrupted transaction: rerun the same public apply command so Roundfix can lock and recover its Git-private journal. Never delete recovery state.
  • Incomplete rollback: stop writes, preserve checkout and recovery state, inspect every reported path, restore trustworthy bytes, and generate a new plan.

Completion

Do not claim adoption or update complete from planning alone.

  1. Require apply exit 0 and inspect verifiedPostimages.
  2. Run each reported formatter or Verification recommendation outside Baseline.
  3. Repair any repository failure outside managed Baseline output.
  4. Generate a fresh plan with the same current Profile input and decisions.
  5. Require no file changes, or perform an exact idempotent reapply at exit 0.

Report remaining warnings, recommendations not run, the approved Plan Digest, and the final Baseline result state.

Add the canonical catalog link to the repository README so users can inspect current installs and available audits. The publishing guide covers the complete discovery path.

<a href="https://skillzs.dev/skills/marcioaltoe/skills/setup-context-driven">View setup-context-driven on skillZs</a>