[Feature]: /speckit-implement-waves => run each phase in a subagent to prevent context rot

[denizmatar]

Problem Statement

/speckit-implement executes the entire tasks.md inside a single agent conversation. Every phase, every task, sequentially, in one context window. For anything beyond a small feature this causes context rot. By the time the agent reaches Phase 5, the context window is already saturated with things that are irrelevant to the current phase.

The symptoms compound as the feature grows:

• Degraded quality late in the run
• No true parallelism: tasks.md already encodes which phases/user stories are independent (via "Dependencies & Execution Order" and [P] markers), but a single-context run can't exploit it; independent user stories are done one after another.
• All-or-nothing blast radius: one long conversation implementing 20+ tasks is hard to supervise, and a mistake in phase 2 silently pollutes the context for every later phase.

The task structure is already a dependency graph. implement just doesn't execute it like one.

Proposed Solution

Add a wave-based way to run implement. It could be a new command like /speckit-implement-waves, or an opt-in flag on the existing implement. Either way it acts as a thin orchestrator on top of the implement command we already have.

The idea:

Read tasks.md and the "Dependencies & Execution Order" section, and group the phases into waves. Phases that don't depend on each other and touch different files go in the same wave. Phases that depend on others go in a later wave.
For each phase, run a phase-scoped implement (limited to that phase's task IDs) in its own fresh agent context. Each phase loads only the design docs it needs, runs its own tests, keeps its own TDD ordering and [P] parallelism, and marks only its own tasks off.
Run wave by wave. Phases in the same wave go in parallel, then the orchestrator waits, reconciles tasks.md, and starts the next wave.
The orchestrator never writes code itself. Its context only holds the wave plan plus short per-phase reports, so it doesn't rot no matter how big the feature is.
Since each phase is just implement running in isolation, all the existing implement behavior stays intact (design-doc loading, checklist gate, hooks, marking tasks off). The only thing that changes is where the work happens: lots of small clean contexts instead of one giant one.

This stays integration-neutral. "One isolated context per phase" maps onto Claude Code sub-agents, Cursor's multitask sub-agents, or just sequential fresh sessions for integrations that don't have a sub-agent primitive.

Alternatives Considered

• Running speckit.implement once per phase by hand with a scope argument. This works, but it's fully manual. You have to track the dependency graph yourself, decide what's parallel, kick off each phase, and fix up tasks.md afterwards. The whole value here is that the orchestration is automatic.
• Just making implement commit more often. That lowers the risk of losing work, but it does nothing about context rot. It's still one saturated context.
• Splitting the feature into several smaller specify runs. This fights the methodology. The point of tasks.md is one coherent feature with one dependency graph.
• Leaning on bigger context windows or summarization. Larger windows and compaction push rot further out, but they don't fix it, and summarization tends to drop the exact decisions later phases rely on.

Component

Spec templates (BDD, Testing Strategy, etc.)

AI Agent (if applicable)

None

Use Cases

• Large multi-story features (3+ user stories, 20+ tasks), which are exactly the ones where the last phases get sloppy today.
• Genuinely independent user stories. For example US1 (backend) and US2 (a separate settings screen) both only need the Foundational phase, so they can run in the same wave instead of back to back.
• Backend/frontend split features where one foundational phase unblocks two separate tracks.
• Runs you actually want to supervise. Reading five short per-phase reports is much easier than scrolling through one 200-message transcript.
• MVP-scoped runs. Orchestrate only Setup, Foundational, and US1 (the MVP slice SpecKit already recommends), each in a clean context, then stop.

Acceptance Criteria

• Given a tasks.md, the command builds a correct wave plan from "Dependencies & Execution Order" (independent phases share a wave, dependent phases are sequenced).
• Each phase runs in its own isolated context. No phase's context contains another phase's transcript.
• Independent phases in the same wave run at the same time. Dependent phases never start before their prerequisites finish.
• Every in-scope task ends up marked off in tasks.md, with nothing lost or overwritten when phases run at the same time (the orchestrator reconciles after each wave).
• Per-phase implement behavior stays intact: TDD ordering, [P] handling, checklist gate, hooks.
• Phases that are already fully done get skipped, so re-runs are safe.
• An optional scope argument can limit the run to a phase range or a single user story (like "US1 only"). Default is all unfinished phases.
• If one phase reports a blocker mid-run, the later waves stop instead of quietly carrying on.
• Output is a short summary per phase (what changed, tests run, blockers) plus a final rollup.

Additional Context

The data this needs is already in every tasks.md: "Dependencies & Execution Order", "Parallel Opportunities", the [P] markers, and the [US#] labels. This feature isn't about generating anything new. It's about executing that graph instead of flattening it.

I've already built this as an integration-side skill: a small orchestrator that dispatches phase-scoped implement runs into background sub-agents, with a quick blocker check before it starts and a tasks.md reconcile after each wave. Happy to share it as a reference.

[denizmatar]

I've already built this as an integration-side skill for Cursor: a small orchestrator that dispatches phase-scoped implement runs into background sub-agents, with a quick blocker check before it starts and a tasks.md reconcile after each wave. Designed to work with Cursor's Multitask mode. Happy to share it as a reference.

SKILL.md

[mnriem]

FYI just to raise awareness I showed how this can be tackled for Copilot using a preset (see https://github.com/mnriem/spec-kit-presets/tree/main/copilot-sub-agents). Note that there is no predefined contract on how to do this for every agent in an agent agnostic way. So feel free to deliver a preset here too or an extension whichever you feel works the best. Then submit it for inclusion in our community catalog (note you will host the preset/extension yourself).