Auto-generated mirror. Canonical sources: skills/canopy-runtime/references/runtime-claude.md and skills/canopy-runtime/references/runtime-copilot.md. Edits go to the canonical files; this page is regenerated by scripts/sync-runtime-docs.py.

Runtime Specs

Canopy uses an interpreter pattern: the same SKILL.md runs on Claude Code and GitHub Copilot — only the runtime-spec file loaded at execution time differs. The two specs below are loaded by canopy-runtime based on the active host.

Claude Code Runtime

Defines how Canopy skill constructs execute on the Claude Code platform.

<skills-root> resolves to whichever recognized skills directory canopy-runtime was installed into — .agents/skills/, .claude/skills/, or .github/skills/ (see SKILL.md → Skills root resolution). Claude Code natively reads from .claude/skills/; cross-agent installs at .agents/skills/ are exposed to Claude when the workspace mirrors them or the install script writes to .claude/skills/ directly.

Base paths

• Skills: <skills-root>/<name>/SKILL.md
• Primitive slices: <skills-root>/canopy-runtime/references/ops.md (index) → <skills-root>/canopy-runtime/references/ops/<slice>.md

Subagent dispatch

Native subagent invocation is supported via per-op markers (preferred) and via the legacy ## Agent section (soft-compat).

Marker-based dispatch (preferred)

When a tree node is **OP_NAME** << input >> output (bold around the op name), launch a subagent with the body of OP_NAME as its task:

• Resolve OP_NAME via the standard op lookup chain (skill-local <skill>/references/ops.md or <skill>/references/ops/<name>.md, falling back to <skill>/ops.md for legacy skills → consumer-defined cross-skill ops if any → canopy-runtime’s primitive slices).
• Verify the resolved op definition carries the marker > **Subagent.** Output contract: <schema-path> as the first content under its heading; if missing, halt with a contract-mismatch diagnostic.
• Launch the subagent via the Task tool, passing the bound << values as the subagent’s inputs; the subagent body uses only those inputs + static skill assets.
• Bind the subagent’s schema-shaped result to the call-site >> name; surface in the parent context.

Multiple bold-marked op calls inside a single PARALLEL block emit multiple Task tool calls in one assistant message — heterogeneous parallel fan-out. Sibling failures don’t abort each other (Promise.allSettled).

See ops/subagent.md for the full marker contract.

Soft-compat: ## Agent + EXPLORE

Existing skills with ## Agent declaring **explore** + EXPLORE >> context as the first tree node keep working — runtime treats this as a single-element marked op named EXPLORE:

• Launch an Explore subagent with the task described in the ## Agent body.
• Do NOT inline-read files yourself — the subagent handles all file access.
• Output contract is assets/schemas/explore-schema.json (or legacy schemas/explore-schema.json).
• First tree node must be EXPLORE >> context.

If the ## Agent body uses shape (C) — **explore** — execute NAMED_OP — resolve NAMED_OP via the standard op lookup chain and inject the op body as the subagent’s task.

See ops/explore.md for the soft-compat shapes.

Parallel subagent invocation

When a tree node says “spawn N subagents in parallel,” fan out by emitting N Task tool calls in a single assistant message. The harness runs them concurrently; each subagent has its own context window.

• Bind by name — assign each result to the >> name the prose specifies; outputs return in completion order.
• One assistant turn, N Task calls — preferred over N serial messages: keeps the prompt-cache prefix stable and avoids (N − 1) × R inter-turn reasoning overhead.
• Promise.allSettled semantics — a single failure does not abort siblings; surface all outcomes and let downstream nodes branch via IF.
• Heterogeneous fan-out only — different tasks, independent inputs. Data-parallel iteration over a list is not yet specified.
• PARALLEL block — when a PARALLEL node is the current tree position, emit one Task call per child in this assistant turn. Deterministic fan-out shape — no prose detection needed. Each child’s >> becomes its binding handle.
• Marker-based subagent calls inside PARALLEL — when children are **OP_NAME** << ... >> ... (bold around op name), each Task call uses the resolved op’s body as the subagent task and binds the schema-shaped result. See ## Subagent dispatch above for the contract.

Invocation

• Wrapper skill: /canopy <request> — invokes <skills-root>/canopy/SKILL.md
• Direct skill: Follow <skills-root>/canopy/SKILL.md and <request> — bypasses the wrapper
• Other skills: /skill-name — resolved from <skills-root>/<name>/SKILL.md

Op lookup

1. <skill>/references/ops.md or <skill>/references/ops/<name>.md — skill-local. Backward-compatible fallback: <skill>/ops.md at root.
2. Consumer-defined cross-skill ops (optional; consumers may package these as their own skill).
3. canopy-runtime’s primitive slices — <skills-root>/canopy-runtime/references/ops.md indexes the per-feature slice files.

GitHub Copilot Runtime

Defines how Canopy skill constructs execute on the GitHub Copilot platform.

<skills-root> resolves to whichever recognized skills directory canopy-runtime was installed into — .agents/skills/, .claude/skills/, or .github/skills/ (see SKILL.md → Skills root resolution). Copilot reads skills from .github/skills/ natively; gh skill install on recent gh CLI versions defaults to .agents/skills/ for cross-agent compatibility.

Base paths

• Skills: <skills-root>/<name>/SKILL.md
• Primitive slices: <skills-root>/canopy-runtime/references/ops.md (index) → <skills-root>/canopy-runtime/references/ops/<slice>.md

Dispatch path resolution

All subagent dispatch — whether from a marked op call, a PARALLEL block child, or a soft-compat ## Agent section — uses the same four-path chain (first match wins):

1. /fleet orchestration — when fleet mode or autopilot is active, the orchestrator handles fan-out; each op or child runs in its own context window.
2. @CUSTOM-AGENT-NAME reference — invoke a pre-defined custom agent inline by name; the result populates the call-site >> binding.
3. Native Copilot subagent dispatch — Copilot’s agent engine natively supports spawning subagents in parallel via its built-in subagent capability (e.g. runSubagent). This is the default path when fleet and named custom agents are not configured. Dispatch all subagents simultaneously; each runs in its own context window.

4. Sequential inline fallback — only when native dispatch is explicitly unavailable (e.g. tool disabled by the user). Before falling back, ask the user: “Parallel subagent dispatch is unavailable. These steps are intended to run in parallel — proceeding sequentially will be slower. Continue sequentially?

   Yes

   No”. Halt if the user says No. If Yes, evaluate each op body sequentially inline. Correctness is preserved; parallelism is lost.

Failure semantics: paths 1–3 follow Promise.allSettled — a single child failure does not abort siblings. Path 4 (sequential) short-circuits on first failure.

Subagent dispatch

Native subagent invocation is supported via per-op markers (preferred) and via the legacy ## Agent section (soft-compat).

Marker-based dispatch (preferred)

When a tree node is **OP_NAME** << input >> output (bold around the op name), dispatch the resolved op’s body as a subagent using the Dispatch path resolution chain above.

Resolution: standard op lookup chain (skill-local <skill>/references/ops.md or <skill>/references/ops/<name>.md, falling back to <skill>/ops.md for legacy skills → consumer-defined cross-skill ops if any → canopy-runtime’s primitive slices). The resolved op definition must carry the marker > **Subagent.** Output contract: <schema-path> as the first content under its heading; if missing, halt with a contract-mismatch diagnostic.

See ops/subagent.md for the full marker contract.

Soft-compat: ## Agent + EXPLORE

Existing skills with ## Agent declaring **explore** + EXPLORE >> context as the first tree node keep working — runtime treats this as a single-element marked op named EXPLORE. Apply the Dispatch path resolution chain above. In all cases the result populates context shaped to assets/schemas/explore-schema.json (or legacy schemas/explore-schema.json).

Output contract is identical across all paths.

If the ## Agent body uses shape (C) — **explore** — execute NAMED_OP — resolve NAMED_OP via the standard op lookup chain and dispatch the resolved op body via the selected path.

See ops/explore.md for the soft-compat shapes.

Parallel subagent invocation

When a tree node says “spawn N subagents in parallel,” apply the Dispatch path resolution chain above to all children simultaneously.

• Bind by name — assign each result to the >> name the prose specifies.
• Heterogeneous fan-out only — data-parallel iteration over a list is not yet specified.
• PARALLEL block — when a PARALLEL node is the current tree position, dispatch each child simultaneously via the resolved path. Each child’s >> becomes its binding handle.
• Marker-based subagent calls inside PARALLEL — when children are **OP_NAME** << ... >> ... (bold around op name), each child dispatches the resolved op’s body as a subagent and binds the schema-shaped result. See ## Subagent dispatch above for the contract.

Invocation

• Wrapper skill: /canopy <request> — invokes <skills-root>/canopy/SKILL.md
• Direct skill: Follow <skills-root>/canopy/SKILL.md and <request> — bypasses the wrapper
• Other skills: /skill-name — resolved from <skills-root>/<name>/SKILL.md

Op lookup

1. <skill>/references/ops.md or <skill>/references/ops/<name>.md — skill-local. Backward-compatible fallback: <skill>/ops.md at root.
2. Consumer-defined cross-skill ops (optional; consumers may package these as their own skill).
3. canopy-runtime’s primitive slices — <skills-root>/canopy-runtime/references/ops.md indexes the per-feature slice files.