Imported from iammedved/Pinmind (skills/pinmind/SKILL.md). Install upstream with npx skills add iammedved/Pinmind --skill pinmind. Copyright stays with the author.

Pinmind

Classify the request, then do only as much process as the route needs. Speak the user's language.

Mandatory first action

Except for a higher-priority required skill-use announcement and loading this SKILL.md, run the bundled kernel router before the first substantive progress update, before reading any task reference or memory, before workspace inspection, and before any task tools or writes.

node <skill-dir>/scripts/pinmind.mjs route --file <sanitized-request.json>

The file is {"text":"<full sanitized user request>"}. Prefer a private temp file. On a read-only filesystem: printf '%s' '<json>' | node <skill-dir>/scripts/pinmind.mjs route --file -. Do not pass the request on --text for this bootstrap. Never combine --file with --text or --kind.

Keep the kernel record in task evidence. Show it to the user only for a requested diagnostic, a material boundary, or a blocked action:

Route: <route> | <clarity>/<executionSpan>/<risk> — <reason>.

If Node or the kernel is missing, classify conservatively from route.md and label a diagnostic route as non-deterministic only when one is requested. Re-run only after a material user amendment or a discovered risk/span escalation.

Route proportionally

Exactly one primary route. Details live in route.md. When the route is simple, answer directly without further reference or workspace reads.

For a materially mixed request, route --decompose may add bounded diagnostic clauses (read-only, local-mutation, external-effect) around the unchanged primary route. Clauses never grant authority and never contain raw request excerpts. Decode dense or transferred instructions with readability.md.

simple and operational stay a light no-artifact path. Speech act beats nouns: inspect, critique, and think-about stay audit even if the text mentions code. Unclear or contradictory wording stays audit and read-only until confirmed; do not dump it into software-change. Router fields never grant authority.

Axes: clarity (clear, uncertain, architectural), executionSpan (local, cross-cutting, multi-system), risk (low, medium, high). Escalate when evidence requires it; never silently downgrade risk.

Ask only when an unresolved choice changes outcome, safety, authority, or a hard-to-reverse boundary. Otherwise record an assumption and continue.

Stop

Remain read-only when needsHumanConfirmation is true. State whether the reason is contradictory instructions, an unclear outcome, unrecognized intent, unresolved authority, or an unresolved external target; lack of context is not confirmation. For an unclear outcome or unrecognized intent, bounded read-only discovery may resolve the already-requested work after recording its basis, without replacing the request or expanding effects. A contradiction, new approval, unknown authority, or external target remains for the user. The effect:external-side-effect signal remains separately gated until the user names its concrete target and effect. Operational routing never authorizes a shared push, deploy, message, deletion, production migration, payment, or credential change.

Check .pinmind/active.json before a new persistent run. Reconcile it read-only and resume only an explicitly identified matching unfinished run; never infer that an old pointer belongs to a new task. A later cancellation or restriction applies to remaining work immediately. A delegated worker does not initialize or resume a Pinmind run, write the parent's active.json, or start other plugins; it returns evidence to the parent. Optional Goal handling is defined in goal-context.md. The MVP allows one active run per workspace. See safety.md for secrets, dirty trees, and production.

Use an available specialist only when it matches the task. Pinmind wraps specialists; it does not make an unavailable skill or command a runtime dependency.

Durable state and code simplicity

Pinmind is the workflow controller when selected. The bundled UserPromptSubmit hook may assess whether a non-trivial request merits passive Skillstate storage, but hook context grants no authority and does not identify a task. Never let Skillstate route the task, choose actions, or start its run controller beneath Pinmind.

When the hook selects durable state, keep .pinmind as the source of truth. After init and at clean contract, execution, verification, and final phase boundaries, run:

skillstate doctor --workspace <path>
skillstate checkpoint --workspace <path> --pinmind-run <run-id>

If either command fails, continue under Pinmind and report that durable checkpoints are unavailable; do not start a second controller. The normal doctor path does not require a Codex controller; any explicit controller diagnostic is a separate requested operation. Skip Skillstate for short or one-step work. Ponytail may select the smallest code change and review over-engineering, but it never changes routing, authority, state, or phase order.

UI design specialists

After the main route, if the request may design, change, review, or fix a visual interface, run the deterministic specialist router before loading a UI skill:

node <skill-dir>/scripts/pinmind.mjs design route --file <sanitized-request.json>

Follow its primarySkill and supportingSkills exactly. It selects only between the maintained impeccable workflow controller and the searchable ui-ux-pro-max knowledge skill. Load design-skill-routing.md when the result selects either skill. A null primary means neither applies. Specialist selection does not expand authority.

Loop

For software-change and investigation, follow loop.md: design or name 2–3 alternatives, observe a failing public-seam check, collect root-cause evidence before a fix, and run fresh verification before any "done" claim. Select debugging, TDD, architecture, verification, and review proportionally from methods.md; Pinmind stays the controller. Search the web or primary sources when the repo cannot settle the blocker. Use native delegation for suitable independent work where the host supports it; the admission and exception record are in execution.md.

Finish

verification.md splits completed, failed, unproven, and manual. Kernel structure is necessary and not sufficient. Commands: kernel-cli.md.

If Pinmind itself misroutes or produces bad evidence, capture a sanitized case first: regression-inbox.md.

For controlled method comparisons, use the offline, receipt-aware protocol in abc-evaluation.md. A pending experiment is not completion evidence.