inizio/groundwork

Enforce the completion gate for the orchestrator, producing APPROVE/GAPS/CORRECTION/STOP/REPLAN verdicts before any task is declared done (orchestrator-only; subagents do not invoke this).

Analyze codebase architecture in isolated subagents and produce a prioritised HTML report of shallow modules, coupling smells, and testability gaps — findings never enter main context.

Audit and rewrite a groundwork skill against eight named failure modes. Load before editing any SKILL.md under skills/groundwork/.

Resume an in-flight motive from the motive spine — charter, compiled journal, and run ledger — without relying on the transcript.

Debug a reported failure using the six-phase protocol: loop → reproduce → hypothesise → instrument → fix → post-mortem.

Apply senior-engineer judgment — choose toolchain-enforced structure and test the product from the user's perspective against real hosted dependencies.

"Run and author claude plugin eval suites for this plugin. Triggers on: running a plugin eval, writing eval cases, a case scoring unexpectedly low, recording a baseline score."

Start here to plan a feature — runs the interview primitive then dispatches to the planner.

Route to the right groundwork skill — on-ramps, the feature flow, and context-window boundaries.

Remove AI-generated slop and hygiene debt — deslop (default), deps, lint-debt, docs-staleness, learnings-staleness. Triggers on: deslop, anti-slop, ai slop, cleanup, tidy, housekeep.

Decompose a motive into vertical slices, fan out parallel agent waves, and gate completion with behavioral evidence. Use after a plan is approved or after quick-interview for a small change.

Capture intent relentlessly, one question at a time; produce a motive charter when the frontier is empty.

Author and track a named, persistent work thread with an Objective, open-items register, Decision Log, and Baselines.

Record a PAUSE event into the motive journal — pointer, summary, and ordered next actions — so /continue reconstructs state without the transcript.

Audit plan coverage — map charter ACs to ledger slices and flag gaps before fan-out.

Build throwaway prototypes to answer a design question, then delete or absorb them. Two branches: LOGIC (interactive TUI) or UI (variant switcher). Never promote to production.

Verify a check can fail before reporting its result as evidence.

Run the interview primitive for an ambiguous small change — tight question limit, then briefs general-purpose.

Capture requirements or clarify scope — runs the interview primitive and produces a charter only, no implementation dispatch.

Capture durable lessons from session mistakes and improvements into the Learnings KB.

Manage requirement specifications under doc/specs/ — create, verify, query, and build concept nodes and traceability data. Triggers on: spec out, write requirements, add requirements, spec upkeep, spec is stale, build spec, spec build.

Execute maximum parallel fan-out mode — slice, write ledger, dispatch all independent slices simultaneously. Triggers on "ultrawork", "ulw", "max fan-out", "go parallel".

Bootstrap the groundwork workflow suite for Codex. Load once at session start and again only after context compaction.

Decompose a task into conflict-free parallel slices, one file owned by exactly one slice per wave. Writes a run ledger when the host supports one.