SDLC Agent-Skill Wiki

sd plan submit

3 min read

sd plan submit <seed-id> --plan <file> — the framework’s centre of gravity. It validates a filled plan against its template, spawns one child seed per step, wires the dependency edges between them, writes the plan row to .seeds/plans.jsonl with status approved, and blocks the parent seed on its children. Flags: --overwrite to replace an existing non-draft plan, --name for a human-readable label, --record-decision to write the chosen approach back to mulch, --domain to pin domain inference.

✓ plan pl-a1b2 created (status: approved)
✓ spawned 4 child seeds: seeds-aa01, seeds-aa02, seeds-aa03, seeds-aa04
✓ seeds-9c4d now blocked by 4 children

The gate

Validation is an AJV schema generated from the template config, not a prompt: required sections present and non-empty, min_length on text, min on lists and steps, every steps[].blocks index inside 1..steps.length with self-references rejected, each step declaring either a title (fresh spawn) or an existing_seed (adoption), object-spec fields matching their declared kind, and the template name resolving. Failure is fail-fast and pre-write — both JSONL files are untouched — and stderr carries a patchable partial-state diff:

{ "errors": [ { "path": "sections.steps", "code": "min", "fix": "add at least 1 more step entry" } ],
  "current": { "…plan-as-submitted…" } }

so the agent patches and resubmits without re-running seeds-plan-prompt. That is pattern-plan-verification-loop with the checker replaced by a program: the cluster’s other five gates (gsd-plan-checker, speckit-analyze, bmad-check-implementation-readiness, ce-doc-review, gstack-plan-eng-review) are all models asked to judge a plan and can be argued with; this one exits non-zero (pattern-deterministic-gates). The trade is equally sharp — it checks shape, never quality. Nothing here notices that an approach is wrong, only that the approach field is under fifty characters.

The spawn

Each steps[] entry becomes a child seed carrying the step’s title, type, priority, and normalized labels; a step’s 1-based blocks: [j] means this step blocks step j, translated on submit into real ids on both endpoints’ blocks / blockedBy. The template’s own instruction for the section — “Decompose into ordered, independent implementation steps” — is the pattern-vertical-slice edge. A step may instead declare plan_template: to spawn a child requiring its own sub-plan (requires_plan: true, hidden from seeds-ready until that plan is approved), or existing_seed: to adopt an already-open seed at that index rather than duplicating it.

--overwrite re-plans in place: the row is replaced atomically, revision increments, and new steps are diffed against current children id-first (existing_seed), then by title, then spawned fresh. Children whose step is gone are reported as obsolete with a suggested close command but are never auto-closed — “the LLM has the context to decide whether the existing work is still useful.”

Cross-framework equivalents

The decomposition cluster spans six frameworks: gsd-plan-phase, speckit-tasks, addy-planning, sp-writing-plans, mp-to-tickets, and this. Five of them write markdown a later step re-reads; seeds writes a validated record and spawns the work items itself in the same call, so there is no gap between “the plan says step 3 depends on step 1” and the queue enforcing it.

See Also

Graph View

Backlinks