Workspace Setup

Purpose

Produce a safe, deterministic plan for the local profile .jm-adk.local.json. Default output is a dry-run preview; the skill validates runtime preferences, command policy, privacy boundaries, write safety, and evidence, and writes the file ONLY when an explicit apply mode is present and overwrite policy passes. [DOC]

Evidence tags in outputs use the Alfa core set (kit-facing audience): [CODE] [CONFIG] [DOC] [INFERENCE] [ASSUMPTION]. One spelling per document; never mix the Jarvis {...} family here. See references/verification-tags.md. [CONFIG]

When To Use

• User has approved creating local Alfa profile configuration. [DOC]
• First-use onboarding has collected goal, runtime, autonomy, command policy, privacy, and output preferences. [DOC]
• A developer wants reusable local defaults without committing personal state. [DOC]

When Not To Use (anti-scope)

• User has NOT approved writing local configuration — produce nothing past a preview offer. [DOC]
• The request includes secrets/credentials — reject input, do not plan. [DOC]
• A one-shot answer that needs no persistent local profile state. [DOC]
• Auditing/cleaning existing workspace session folders or task bridges — that is workspace-governance, a different skill. Do not touch it. [INFERENCE]

Inputs

• Primary goal with Alfa.
• Project/product type and known stack.
• Preferred runtime (e.g., Codex). [ASSUMPTION]
• Autonomy level.
• Commands allowed / prohibited / escalation-required.
• Privacy constraints.
• Workspace area and output format.

Missing a required input is {VACIO_CRITICO}-equivalent: stop and ask rather than auto-filling a security-relevant default (command policy, privacy). Cosmetic defaults (output format) may be auto-filled and tagged [ASSUMPTION]. [INFERENCE]

Outputs

• Dry-run preview of .jm-adk.local.json unless mode=apply is explicit. [DOC]
• A validated setup plan matching the contract.
• The local profile file ONLY when --apply is explicit AND overwrite policy passes (no existing file, or --force). [DOC]
• Setup summary: evidence, validation result, residual risks. [DOC]

Deterministic Contract

Validate every plan against assets/workspace-setup-plan-contract.json with scripts/validate_workspace_setup_plan.py. (Contract, validator, and fixtures are planned deliverables of this skill, not yet present in-repo — treat their paths as the build target.) [ASSUMPTION]

A valid plan MUST include:

• target_file exactly .jm-adk.local.json. [DOC]
• mode ∈ {dry-run, apply}, with dry-run as the default. [DOC]
• Runtime preferences: goal, runtime, autonomy, workspace area, output format. [DOC]
• Command policy: allowed, prohibited, and escalation-required classes. [DOC]
• Privacy policy: local-only storage, no secret storage, redaction categories, and a completed secret scan. [DOC]
• Write policy: explicit apply required, .gitignore coverage, --force for overwrite. [DOC]
• Evidence entries using approved tags. [CONFIG]
• Validation checks for: assets, deterministic scripts, quality criteria, runtime preferences, command policy, privacy policy, write policy, evidence. [DOC]

Workflow

1. Discover — read existing local profile state WITHOUT printing sensitive content (report presence/shape, not values). [DOC]
2. Analyze — run the secret scan on all inputs; detect overwrite risk (does .jm-adk.local.json already exist?). [DOC]
3. Plan — build the deterministic plan: target file, runtime preferences, command policy, privacy policy, write policy, evidence, validation checks. [DOC]
4. Validate — run the offline validator; a failing check blocks any apply. [DOC]
5. Execute — dry-run by default; write only with explicit --apply and a passing overwrite guard. [DOC]

Secret Detection (analyze gate)

Reject and do not store inputs matching credential shapes, including: API keys / tokens (sk-, ghp_, xox, long high-entropy strings), passwords, private keys (BEGIN ... PRIVATE KEY), bearer/Authorization headers, raw email addresses, and connection strings with embedded credentials. [INFERENCE] On a hit: reject the input, name the category (not the value), and continue with a redaction placeholder. Never echo the secret back. [DOC]

Safety Limits

• Never commit .jm-adk.local.json (require .gitignore coverage). [DOC]
• Never overwrite an existing profile without --force. [DOC]
• Never store secrets; use placeholders and policy text. [DOC]
• Never require network, wall-clock time, randomness, or live credentials for validation — validation MUST be reproducible offline. [DOC]
• Never widen command permissions silently; widening is escalation-required. [DOC]

Self-Correction Triggers

Stop and re-plan if any holds: planning to write while mode=dry-run; overwrite intended without --force; a secret reached the plan body; validation checks are incomplete or the validator was skipped; command policy widened without explicit escalation; profile would be writable into a git-tracked path. [INFERENCE]

Anti-Patterns

• Writing on the first pass "to save a round-trip" instead of previewing. [INFERENCE]
• Auto-filling command/privacy policy from assumptions instead of asking. [INFERENCE]
• Printing existing profile contents during Discover. [DOC]
• Tagging planned-but-absent assets as [CODE] instead of [ASSUMPTION]. [INFERENCE]
• Mixing Jarvis {...} and Alfa [...] tag families in one output. [CONFIG]

Decisions and Trade-offs

• Dry-run default trades one extra confirmation step for a hard guarantee that no profile is written without explicit intent. The friction is the feature. [INFERENCE]
• Offline-only validation trades richer live checks (e.g., verifying a runtime exists) for determinism and reproducibility across machines and CI. [INFERENCE]
• Stop-on-missing for security inputs trades convenience for the guarantee that no permissive default silently widens the trust boundary. [INFERENCE]

Acceptance Criteria (validation gate)

• Dry-run is the default; no write occurred without explicit --apply. [DOC]
• --apply produces a profile that passes the contract validator. [DOC]
• Existing profile preserved unless --force was passed. [DOC]
• Every secret-like input was rejected and never stored. [DOC]
• Offline validator accepts valid fixtures and rejects unsafe fixtures. [DOC]
• All evidence tags come from a single family, consistent spelling. [CONFIG]
• No network / clock / randomness was needed to validate. [DOC]

Edge Cases

• Existing profile + apply, no --force → block; return preview + the exact --force command. [DOC]
• Apply requested but secret detected → reject before planning; secret never enters the plan. [DOC]
• Network-required validation requested → refuse; substitute the offline validator. [DOC]
• Workspace-governance phrasing ("audit stale session folders") → do not activate; defer to that skill. [INFERENCE]
• One-shot answer requested → do not activate; answer without persisting. [DOC]

Fallback

If a safe write is not possible, return the JSON preview plus the exact command for a later authorized apply. Never partial-write. [DOC]

Scripts

# Validate a plan fixture offline (build target; see Deterministic Contract).
python3 skills/workspace-setup/scripts/validate_workspace_setup_plan.py \
  skills/workspace-setup/scripts/fixtures/valid-dry-run-profile.json
bash skills/workspace-setup/scripts/check.sh

Examples

• Dry-run — Codex runtime, evidence tags required, local-only state: emits a preview of .jm-adk.local.json, validator passes, nothing written. [DOC]
• Apply (safe) — approved profile, no existing file (or --force): writes the validated profile, then prints the setup summary. [DOC]
• Apply (blocked) — existing profile, no --force: returns preview + the --force command; no write. [DOC]
• Rejected — input contains an API token: secret scan rejects it, plans with a redaction placeholder, never stores the token. [DOC]