Implementation Plan

Implementation Plan

Orchestrate a fresh Slice-first planned-feature set under the selected artifact root: SPEC.md, lightweight tasks.json, package Markdown, and declared report_path result files. “Fresh” describes the artifact set, not whether the target system is new. Source inspection and helpers use the separate code root.

This skill is the implementation-plan orchestrator, not the planner worker. Preserve approved requirements, roots, decisions, and workflow state here; delegate artifact writing with references/planner-agent-contract.md. Mode is initial by default. Use implementation-continuation only when implement supplies approved requirements, roots/ref/slug, current artifacts, the Execution Contract, a stage-sourced plan-defect packet, and accepted empirical reports or explicit none.

Always

• Plan from approved requirements, safe Conceptualize/diagnosis material, verified repository/official evidence, and, when empirical evidence applies, accepted bounded empirical-spike reports only. Preserve planned-hotfix context without inventing an artifact field or feature ref.
• Prefer static/official evidence. Inventory a bounded set of distinct material empirical questions tied to the current approved decisions; routine work, cost alone, and statically resolved questions do not trigger a spike.
• Preserve planning context while evidence runs. Give each falsifiable question one stable logical-question ID. Attempt 1 is one fresh empirical-spike invocation; follow-ups are fresh invocations with the same question ID, incremented attempt IDs 2–3, and a named corrected packet or changed method/signal. Unchanged attempts are forbidden; three total attempts or continually emerging/unbounded questions are non-convergence. Independent questions may run in parallel; only accepted evidence may create the next sequential question.
• A spike is evidence-only and cannot write plan artifacts, choose workflow, or invoke planning. Because this is already the implementation-plan invocation, resume this orchestrator; never recursively invoke implementation-plan.
• Accept resolved-static, supported, or rejected only after validating identity, provenance, method, authority, bounds, limitations, and cleanup. Correct blocked/inconclusive only through an authorized changed attempt; unresolved initial mode stops and continuation returns protected/out-of-contract gaps to implement. Evidence never authorizes behavior, scope, architecture, deferral, or risk acceptance.
• Delegate all planned-feature artifact writing to a fresh planner. A planner that finds unresolved material empirical behavior must return BLOCKED: empirical_evidence_needed; only this orchestrator may resolve it.
• In initial mode, retain every existing planning/overwrite/user-decision gate. In continuation, repair supplied plan defects autonomously while semantics/scope/visible behavior/risk/manual exceptions stay fixed. Any new continuation package supplies BASE_KIND, exact BASE_REF, focused-review-bound REVIEWED_BASE_SHA, and prerequisite ref/SHAs: approved original base when independent, or exact feature/integration SHA containing prerequisite ancestors when dependent. Reports may be none; return only semantic/risk/ manual/protected expansion to implement, never prompt here.
• Slices are product/design authority only, never workflow/tool/review authority. Index-only planning is allowed when no Slice is independently useful; otherwise the planner inventories and reads every safe Slice in full.
• Registry is bookkeeping; package Markdown owns assignment, and report_path names the independent result. Boundaries keep material requirements observable from files alone.
• Define done objectively: executable feature ## Acceptance in SPEC.md and per-package ## Acceptance Checklist, with only explicit human-approved manual (approved) exceptions. Surface missing runnable commands before authorization rather than writing unverifiable acceptance.
• The planner applies the complete shared Module/Interface/Seam model and all smell heuristics while shaping packages, but persists only material requirement/risk-traced implications in existing scope, boundaries, risks, dependencies, and verification fields; it creates no standalone quality result artifact or per-smell rows.
• Carry explicit artifact root/ref, code root, and resolved feature/artifact slug through packets, validation, and summaries. A Conceptualize slug remains the default absent approved full rename/migration metadata.
• Resolve Semgrep before planner dispatch. Use supplied state without reopening opt-in; for direct invocation, resolve preferences. Disabled requires no setup/scan/internet. Load policy only at its action point, disclose clone or fast-forward pull side effects, and keep artifact authoring scan-free.
• Validate returned artifacts before success.

Do

1. Load ../../references/artifact-store.md. Resolve mode, artifact root/ref, code root, slug, and source. For initial, select direct requirements/evidence or one Conceptualize workspace and ask once if ambiguous. For implementation-continuation, validate the caller binding and use only its requirements, current artifacts, Execution Contract, stage/defect provenance, and report set or none; return conflicts without prompting. Create a missing direct-planning sidecar through worktree; publish only when separately authorized.
2. Check unsafe paths, unresolved decisions, overwrite, empirical uncertainty, and risk acceptance before delegation. In continuation mode, treat same-root routine artifact overwrite/repair as covered and fail closed on semantic or authority expansion. Resolve testing authority only when material execution feasibility requires it: accepted/current workflow, routine-safe bounded-local fallback, or exact task-local authorization. If insufficient, initial mode may invoke testing; continuation returns the gap to implement for contract-owned repair or an existing protected stop. For nontrivial/risky plans, apply references/design-preflight.md; reuse equivalent current adversarial analysis only for identical approved scope/evidence with complete requirements and overengineering coverage. Resolve every COVERAGE_GAPS, MUST_DECIDE, and BLOCKERS item before artifact writing.
3. Inventory bounded material empirical questions unresolved after repository/official evidence. Start each logical-question ledger at attempt 1 and invoke empirical-spike once with its decision, outcomes, constraints, safe paths, authority, and report contract. Parallelize independent questions; sequence only evidence-created questions. Validate every report. A corrected packet or changed method/signal may invoke attempts 2–3 under the stable question ID; never retry unchanged or exceed three total. Initial mode retains semantic gates; continuation applies same-requirement evidence autonomously and returns only Stop-if expansion to implement.
4. Resolve the planner packet's Semgrep state. Supplied state is authoritative and required in continuation mode. Otherwise, only in initial mode, load ../../references/model-preferences.md and conditional ../../references/semgrep.md, present opt-in/setup, and name clone/pull effects. Continue disabled if declined; never run Semgrep scans while authoring.
5. Dispatch a fresh planner with mode, roots/ref, slug/migration metadata, delivery context, approved source, overwrite/continuation authority, stage/defect plus reports or none, original base, current integration state, and BASE_KIND, exact BASE_REF, candidate REVIEWED_BASE_SHA, and prerequisite ref/SHAs for each continuation-created package. Include testing provenance when triggered and Conceptualize paths only when applicable. Pass labeled paths for planner contract, artifact store, Slice authority/projection, preflight evidence, SPEC template, clean-code, work-package, canonical artifact model, artifact authoring, validation, tool usage, and optional Semgrep. Include resolved Semgrep state, the accepted empirical reports (or none), stop conditions, and output.
6. If the planner returns BLOCKED: empirical_evidence_needed, verify no artifacts were written and reconcile its one question with the ledger. A distinct bounded question starts attempt 1; an existing question advances only through step 3's materially changed attempts. Redispatch a fresh planner with the original packet and reports. Stop on an unchanged/over-cap question, unbounded emergence, bundled questions, or malformed status.
7. On a normal return, re-open SPEC.md, tasks.json, and every package Markdown from the artifact root. From the code root run python3 "${SUPER_DEVELOPER_PLUGIN_ROOT}/assets/sliceproof.py" validate-plan \ --artifact-root <artifact-root> --code-root <code-root> .tasks/<feature>/tasks.json. Route semantic repair through a fresh planner packet; never patch planned-feature artifacts inline.
8. Report roots/ref, delivery context, feature paths, packages/dependencies, closure-complexity and sequencing rationale, empirical question/report-set status and provenance or static-resolution note, execution-feasibility/testing authority, Slice inventory, approved deferrals, assumptions, validation, and next gate.

Load if needed

• Conceptualize input applies → references/conceptualize-inputs.md and ../../references/conceptualize-slice-authority.md
• Nontrivial/risky design challenge → references/design-preflight.md
• Package shaping → ../../references/work-packages.md
• Drafting SPEC.md → references/spec-template.md
• Drafting registry/packages/result declarations → references/artifact-authoring.md and ../../references/slice-first-artifacts.md
• Before artifact writes/completion claims → references/validation-checklist.md
• Helper syntax or command safety is unclear → ../../references/tool-usage.md
• Semgrep preference/cache/policy/evidence applies → ../../references/semgrep.md

Stop if

• Any slug, root/ref, artifact/source path, or supporting-contract path is missing, unsafe, or contradictory.
• A Conceptualize slug changes without approved migration metadata; in initial mode, existing plan state would be overwritten.
• Initial mode needs a requirement/Slice/deferral/boundary/risk decision; continuation mode would change approved semantics/scope/visible behavior/manual exceptions, accept risk, or exceed the Execution Contract.
• Slices exist but the delegated planner cannot complete the full safe inventory.
• Required empirical evidence is blocked/inconclusive/malformed or lacks provenance, method, authority, or cleanup.
• A logical empirical question exhausts attempt 3 without an accepted result, repeats unchanged, or distinct blockers continually emerge or cannot be bounded.
• sliceproof.py validate-plan fails and cannot be repaired within scope.
• Semgrep would require unapproved network/setup, unavailable safe cache, hidden cloud behavior, or mandatory scans.

Output

Return mode, artifact root/ref, code root, feature/artifact paths, packages/dependencies and closure rationale, Acceptance checks with flagged manual exceptions, empirical report-set status/provenance, execution-feasibility/testing authority, Slice inventory, deferrals, assumptions, validation, and next step. Initial mode retains review-plan after confirmation; continuation returns repair status to implement for focused review-plan without fresh plan or implementation approval.