/caller-options

Caller Options (COP) makes a skill's hidden choices visible. When a participating skill runs, COP detects whether the invocation has material optionality along two dimensions, evaluates the live options, and surfaces them to the caller — instead of the skill silently picking. The caller (a human, or another agent) decides.

• Venue — who executes: inline (the active agent), delegated (one primed sub-agent, while the active agent orchestrates — the "agentic counterpart"), or fan-out (N parallel sub-agents).
• Mode — which path of the skill, when the skill already has more than one.

COP is invoked two ways: automatically, when a skill carrying the COP opt-in block is triggered; or directly, as /caller-options <task or skill> for an ad-hoc routing decision.

The protocol — five steps

COP always runs in the active agent — the venue decision cannot itself be delegated.

1. Triage. Read the triggered skill's caller-options declaration and the task. If exactly one venue and one mode clearly dominate, skip straight to execute — do not prompt the caller. This anti-nag gate resolves most invocations.
2. Enumerate. List the viable mode × venue combinations. Drop delegated/fan-out if the skill is on the never-delegate list, is self-orchestrating (spawns its own sub-agents — wrapping it would nest to depth 2), or if a resolved-invocation marker is present (depth guard).
3. Assess. Evaluate the survivors for cost, context hygiene, and fit. Two or fewer options → compare inline. Three or more → invoke /insight-assess.
4. Surface. Human caller → AskUserQuestion, recommended option first. Agent caller → state the options and the pick visibly, honoring the skill's default-policy.
5. Execute. Run the chosen combination. For delegated/fan-out: build a primed brief (reusing the handoff brief schema), prepend the COP-RESOLVED marker block (literal format in OPT-IN-BLOCK.md) — it carries depth: 1, so the sub-agent runs its skill and any COP participant it triggers inline and never re-enters COP; spawn; orchestrate; inspect the diff/output before accepting.

Venue heuristics

Whether to delegate is decided by references/VENUE-HEURISTICS.md. In short — delegate when the skill does heavy throwaway reading, produces a bounded artifact, writes files, or the main context is large; keep inline when the skill needs the live conversation, is trivial, is interactive, or the caller needs to see the reasoning. A skill that spawns its own sub-agents is inline only — COP never wraps an orchestrator.

The opt-in block & frontmatter contract

A skill joins COP by adding the standard opt-in block — see references/OPT-IN-BLOCK.md for both forms (a @SEC/@STEP for protocol.yon, a prose block for SKILL.md) and the frontmatter-contract spec:

caller-options:
  venue: [inline, delegated]      # venues this skill supports
  modes: [quick, deep]            # named paths; omit if single-mode
  default-policy: ask             # ask | recommend | auto

The frontmatter is metadata — the harness does not execute it; it is read by audit tooling and by humans. Behavior lives in the opt-in block + protocol.yon. Per-skill frontmatter is canonical for that skill's dimensions; references/CANDIDATES.md is a derived index.

Depth & safety

• Depth-1 only. COP-spawned sub-agents must not delegate again. Every primed brief carries a COP-RESOLVED marker block with depth: 1; the receiving sub-agent runs its skill — and any COP participant it triggers — inline, which closes the depth-2 path. No harness signal detects sub-agent nesting, so the marker is the mechanism.
• No orchestrator nesting. Self-orchestrating skills are inline-forced.
• Inspect before accept. Delegated output is inspected (diff/changed paths), never accepted on narrative alone — reuses orchestrate-mode's discipline.

Platform mapping

COP runs on every runtime, but two steps need a runtime-specific form:

• Surface (step 4) — Claude Code: AskUserQuestion. Codex, or any runtime without a structured-question tool: present the options + recommendation in prose and wait for the caller's reply. The agent-caller path (visible pick) is already runtime-agnostic.
• Execute delegated/fan-out (step 5) — resolve the capability the current runtime actually provides. Use proven worktree isolation when available. When workers share a filesystem or working tree, read-only work may still fan out; code-changing work requires disjoint write scopes and explicit shared-state coordination, otherwise serialize it or run inline. Never infer isolation from a vendor or tool name.

Triage, Enumerate, Assess, and inline Execute are runtime-agnostic.

Standalone use

/caller-options <task> runs the protocol ad-hoc on a task or a named skill, with no opt-in block involved — useful for a one-off routing decision or to smoke-test the protocol.

Boundary

Not a session mode — for that use orchestrate-mode (dispatch-only) or multi-agent-mode (lead + helpers). COP routes a single skill invocation. COP uses insight-assess to evaluate 3+ options — it does not replace it.

When orchestrate-mode or multi-agent-mode is active, the mode owns delegation — COP drops the Venue dimension and surfaces only Mode; the orchestrator decides whether the skill runs in a worker. This avoids COP and an active orchestrator both trying to route the same invocation.

Structured execution spec: protocol.yon. Read it for the canonical rules and step sequence; this file is explanation. The two must stay in sync — if you edit one, update the other and refresh the @STAMP date.

Self-improvement. On completion, run the Self-Improvement Protocol (self-improve/SKILL.md): apply its signal taxonomy and threshold to this run; surface any surviving proposal for handler approval. Conservative — silent otherwise. Never auto-apply.