SDLC Agent-Skill Wiki

Pattern: Fresh-context subagents

5 min read

Route heavy research and execution work through subagents that each start with a clean context window (GSD targets ~200k tokens per executor), while the main session stays lean. This directly counters context rot — quality degradation as a single agent’s context window fills. Each subagent gets only “a precisely scoped prompt and the subset of artifacts it needs.”

GSD:

Addy Osmani — Agent Skills:

BMAD:

  • bmad-dev-story — implements one story per fresh context, relying on the story file to carry everything it needs.
  • bmad-create-story — makes the fresh dev context viable by front-loading all context into the story.
  • bmad-code-review — reviews “in fresh context, ideally different LLM” than the one that wrote the code.

Compound Engineering:

gstack:

  • gstack-spec — --execute spawns a fresh-worktree agent; the browser sidebar agent runs in an isolated session.

Superpowers (a foundational principle — “they should never inherit your session’s context or history — you construct exactly what they need”):

  • sp-subagent-driven-development — a fresh implementer subagent per task; a fresh reviewer per review; briefs and reports move as files, never pasted into the controller’s context.
  • sp-dispatching-parallel-agents — one fresh-context subagent per independent problem domain, dispatched concurrently.
  • sp-requesting-code-review — the reviewer subagent gets precisely-crafted context, never the controller’s session history.

pstack — with an inversion worth recording: a generic fresh subagent is treated as a failure mode:

  • pstack-poteto-agent — the framework’s answer. A delegate spawned mid-playbook starts without the mode skill loaded and therefore drops the principles, the reply rules, and the autonomy policy, so pstack ships a subagent type whose sole instruction is to read pstack-poteto-mode in full first. “Substituting generalPurpose skips that read and drifts.” The exemption is deliberate: routed workflow skills set their own subagent types to get model diversity, and must not be overridden.
  • pstack-how · pstack-swarm · pstack-interrogate · pstack-reflect — readonly or cloud subagents that keep bulk reading out of the parent thread.
  • pstack-principle-guard-the-context-window — the principle behind the routing.

Tessl:

  • tessl-launch — runs an installed skill as an isolated local or cloud invocation; a --launchable skill can declare typed input and output JSON Schemas and return a structured result, which makes the capability a function call rather than a prompt.

builder-skills:

Matt Pocock — Skills for Real Engineers:

  • mp-implement-spec — exploration, implementer, and merger subagents, communicating through sparse context pointers (v1.3.0).
  • mp-grilling — environment facts dispatched to background sub-agents so research never blocks a question round (v1.2.0).
  • mp-wayfinder — a research subagent per research ticket, burned down in parallel while charting continues (v1.2.0).

Enabled by (infrastructure)

Where the harness offers fresh context as a primitive a skill may or may not use, an execution-layer runtime can make it non-optional:

  • bernstein (platform) — the review gate hard-asserts that the reviewer runs in a new session id and raises FreshContextViolation if the implementer’s transcript is threaded in, eliminating “the silent ‘we reused the session by accident’ failure mode.” The reason given is drift: “long-running implementer contexts accumulate drift that makes the implementer a poor reviewer of its own diff.” Context isolation enforced by the substrate, not requested by a prompt.

Provided by (harness)

The harness layer provides this pattern as a primitive — the sub-agent mechanism lives in the agent program itself, below any framework skill:

  • claude-code — the Agent / Task tool spawns a sub-agent with its own fresh context window and a scoped toolset, returning only a final result to the caller; agent types are defined in .claude/agents/*.md, and several launched in one turn run concurrently. This is the harness affordance every fresh-context skill above is standing on.
  • opencode — built-in build / plan / explore / scout agents plus background subagents, each with its own context and scoped tools.
  • factory-droid — Custom Droids (Markdown+YAML in .factory/droids/), each with a fresh context window and its own session, model, tool policy, and autonomy level; delegated via the Task tool with run_in_background for concurrency.
  • Not pi — pi’s minimal core deliberately omits sub-agents (extension-only); a framework that needs them is not portable to pi without an extension. The clearest case of a pattern that is a harness primitive in some harnesses and absent from another.

See Also

Graph View

Backlinks