Compound Engineering Plugin boundary

Type: agentic-system-analyses/types/agentic-system-boundary.md

Boundary and evidence

This run examines the complete shipped plugin, not just its converter API or its learning-capture skill. Its intended use is to structure coding-agent work and carry durable project learning into later work. The target is in scope as an extension or tool mechanism: model-facing skills direct a host agent, and bundled programs support dispatch, state handling, validation and integration. It is not the enclosing agent runtime. The whole-system boundary includes the plugin's workflows, producers and consumers of retained knowledge, maintenance and admission paths, host loading, converter/install tooling, and repository-local improvement and evaluation plane.

The initial trace connects native entry points to their consumers. package.json declares Pi resources and the root OpenCode module; index.js exports .opencode/plugins/compound-engineering.js, which loads skill files and registers them and commands with the host. .pi/extensions/compound-engineering.ts supplies skill paths through resource discovery. .codex-plugin/plugin.json declares the same skill tree. The separate src/index.ts CLI routes install and conversion to src/commands/install.ts, src/commands/convert.ts and src/targets/index.ts; these are shipped distribution paths, not excluded applications (SRC-1 at those paths).

export default function compoundEngineeringPiExtension(pi: any) { pi.on("resources_discover", async () => ({ skillPaths: [skillsDir], })) } --- .pi/extensions/compound-engineering.ts

The knowledge loop is included on both sides of storage. Capture eligibility, writing and grounding are specified by skills/ce-compound/SKILL.md and its assembly and grounding references. skills/ce-compound-refresh/SKILL.md and references/per-action-flows.md maintain, replace, consolidate and delete retained learning. skills/ce-plan/references/research.md dispatches the local learnings researcher and reads its output during consolidation. That researcher searches the retained corpus and declared packs. skills/lfg/SKILL.md is a shipped caller that conditionally invokes capture after implementation and review; it cannot be excluded merely because it delegates the write (SRC-1 at these paths).

Outcome: one qualifying solved problem is written as a durable learning under <root>/solutions/, grounded against the current tree, discoverable by the next agent. --- skills/ce-compound/SKILL.md

Open $SCRATCH_DIR/learnings.md when that dispatch ran, and show the user any Skipped pack files line once. --- skills/ce-plan/references/research.md

The improvement plane includes the repository-local authoring skill, standing authoring instructions, tests, behavioral evaluation drivers and CI/release gates. These are distinguished from instructions installed into user projects. The target explicitly states this distinction (SRC-1, AGENTS.md):

AGENTS.md, CLAUDE.md (symlink to AGENTS.md), and GEMINI.md are authoring context for this source repository. Skills are installed into end-user environments, where they run against the user's local instruction files, not this repo's. Behavioral rules that must affect a skill at runtime belong in that skill's SKILL.md or files under its own references/ directory. --- AGENTS.md

Frozen revision and evidence limits

The sole frozen source is SRC-1 at 50532b00de523980010d97fe950965b35382a3b7, with applicability cutoff 2026-10-13 from the opening input. The acquired checkout's HEAD matches the pin and its status was clean. The acquired source object is retained unchanged in frontmatter. The cutoff is the run date, not a claim that upstream changes after this commit were reviewed.

The overall tier is code-grounded: executable resource loading, CLI dispatch, pack resolution and evaluation drivers were inspected alongside model-facing instructions. Skill prose supplies doctrine/design evidence; it does not itself demonstrate that an agent obeyed it. No target execution, installation, tests or model calls were performed. Supplied tests and evaluation drivers are inspectable implementation, not newly observed runs. The source register does not claim observed-run or causal-experiment evidence merely because the repository contains fixtures and authored reports.

External dependencies include enclosing coding-agent hosts and their skill loaders, model providers, native subagent facilities or companion packages, and the Node/npx/acpx and local agent-CLI transport used by cross-model routes. The initial cross-model trace includes skills/ce-work/references/cross-model-execution.md, skills/ce-work/scripts/cross-model-work.sh and skills/ce-work/scripts/peer-job-runner.py (SRC-1). Their host-facing protocol is included; implementations of external hosts, transport packages and services are not frozen here. This prevents conclusions about host scheduling internals, provider-side memory, actual permission enforcement across every host, or actual served-model identity. Third-party packs, live project repositories and real user sessions are not supplied captures: their absence prevents conclusions about installed-project learning quality, successful later reuse, or improvement attributable to the plugin. The included resolver and session-history tooling may still establish how those inputs are requested and consumed.

There is no identified source-access gap. Uninspected files at this commit remain available evidence, not missing evidence. Initial inspection establishes the boundary and principal shipped loop; it is not a claim that every skill or every platform adapter has already received a behavioral audit.

Top-level coverage

Every top-level tracked entry is classified below. All rows refer to SRC-1. “Included” denotes functional coverage by this run, not a claim of exhaustive initial reading. “Indexed” means the tracked paths were enumerated but their bodies were not yet inspected. Directory rows identify material paths rather than treating a directory label as evidence of loop coverage.

Source register