compound

name	compound
description	Use at the end of a full cycle (after `/ship` or `/debug`) to harvest the run into durable signal — a solution doc in `docs/solutions/<category>/` plus curated memory files. Dispatches `docs-writer` then `/memory-curate` for families near the soft limit. Reads/writes all agent memory at the skill layer via MCP tools before/after subagent dispatch.
team_pattern	solo

compound

You are the durability pipeline. You take the transient outputs of a cycle (code, reviews, tests, scattered memory appends) and turn them into two persistent assets: a solution doc future orchestrators will route against, and consolidated family memories that fit within their soft limits.

Inputs you will be given

• User prompt (verbatim) under ## Original prompt in the brief file.
• Input artifact path — the driving artifact for the cycle being compounded. Usually a docs/ship/<slug>-ship.md or docs/solutions/bug-fixes/<slug>.md. May point at a plan doc if compounding directly after /plan+/work.

Memory protocol (skill layer)

This skill follows the canonical memory protocol in skills/_shared/memory-protocol.md. See that file for the read-before / persist-after contract, the _shared write serialization rule, and the deprecated ## Memory findings legacy path.

Memory deltas for this skill

• Pre-dispatch reads at the skill layer (for the docs-writer dispatch): _shared, docs-writer, plus every family memory touched during the cycle. Typically planner, developer, reviewer, tester, debugger, researcher; also design, framework, engineering, marketing whenever those families were dispatched or cross-read during the cycle. All are passed as cross-read context for docs-writer.
• Stage 2 delegates to /memory-curate, which handles its own memory read/write cycle. Do not duplicate curation policy here.
• Subagents now call memory_findings_submit directly per _shared/findings-schema.md. The legacy ## Memory findings YAML block is DEPRECATED but still parsed by the substrate in v0.4.

Stages

Stage 1: Categorize and author solution (docs-writer)

Determine the solution category from references/categories.md based on the driving artifact:

• /ship cycles with new capability → features
• /ship cycles with structural change only → refactors
• /debug cycles → bug-fixes (usually already written by /debug; docs-writer enriches)
• Other triggers map via the descriptions in references/categories.md

1. Read memory: call mcp__agent-substrate__memory_read_shared(), mcp__agent-substrate__memory_read(agent_name="docs-writer"), and read all family memories touched during the cycle via mcp__agent-substrate__memory_read(agent_name="<family>") for each — including any of design, framework, engineering, marketing that were dispatched.
2. Dispatch docs-writer with the driving artifact, cycle sub-artifacts, references/solution-schema.md, and all memory content under ## Memory context. Produces docs/solutions/<category>/<YYYY-MM-DD>-<slug>.md with sections ## Problem + ## Root cause (bug-fixes) or ## Motivation (all other categories), ## Solution, ## Related patterns, ## Applies to.
3. Parse the ## Memory findings YAML block from the docs-writer's response. For each finding, call mcp__agent-substrate__memory_append(agent_name="docs-writer", section=finding.section, item=finding.item). Handle warning / needs_curation responses.

Stage 2: Curate memories

Enumerate every family whose memory was appended during the cycle (read memory_findings from each sub-skill's summary). For each family, dispatch /memory-curate if its memory file is near or over the soft limit. The /memory-curate skill handles its own memory read/write cycle (reading the current YAML, passing it to the curator subagent, then writing the curated result back) — you do not override or duplicate curation policy here.

Stage 3: Consolidated summary

Produce a summary artifact listing: the solution doc path, the list of curated family memory files (with before/after char counts where available), and any families skipped because their memory was well under the soft limit. Confirm docs-writer's memory findings were persisted in stage 1.

Write back

Canonical artifact path: docs/solutions/<category>/<YYYY-MM-DD>-<slug>.md (primary) plus docs/compound/<YYYY-MM-DD>-<slug>-compound.md (summary).

artifact_path: docs/solutions/<category>/<YYYY-MM-DD>-<slug>.md
status: complete          # complete | blocked | needs_human
memory_findings: [docs-writer]
memory_curated: [planner, developer, reviewer, tester]   # actual families vary by cycle
next_skill_hint: null

Invariants (never violate)

• This skill must persist every subagent's memory findings via memory_append before returning (docs-writer findings in stage 1; curation handled by /memory-curate in stage 2). If a subagent returns no ## Memory findings section, log a warning — the agent may need its prompt updated.
• Solution category must come from references/categories.md — never invent a new category inline. If none fits, return status: needs_human.
• Solution doc must include every section in references/solution-schema.md.
• Never overwrite an existing solution doc — if the slug collides, suffix with -v2.
• Never skip curation on a family whose memory file is over the soft limit. Dispatch the curator even if the cycle was small.

References

• references/categories.md — canonical solution categories.
• references/solution-schema.md — required sections for a solution doc.