Agent Native

Launch and observe caller-selected agent sessions as explicit roles, and return runtime facts per packet. Execution does not validate output: no adapter issues a binding verdict or turns factory completion into delivery or validation proof. Orchestrate decides what to dispatch and when; this skill launches and observes.

Before launch, and while running

• Validate the whole batch first. Every packet needs its packet identity, caller intent, role, selected executor, workspace, starting commit, authorized source scope, evidence destination, all transitive effects, the actions it must not take (such as push, merge, history rewrite, tracker or external writes) and a canonical workspace-relative write scope, generated companions and regeneration outputs included, in separate isolation. If any packet is invalid, launch none: starting the valid ones and fixing the rest later is a partial launch.
• Compare canonical scopes. Resolve symlinks, normalize paths and compare case-insensitively so an alias cannot hide a collision. A lexical check alone cannot prove symlink or runtime isolation. Disjoint files are not independence: workers, readers included, contend on a resource one of them writes, installs into or holds exclusively, such as a bound port, a database it writes, a package install it mutates or a locked browser profile; give each its own instance or serialize only those. A shared model-provider limit is backpressure, launched in waves, not serialized (shared runtime).
• Delivery is not engagement. A delivered prompt or an acknowledgement proves transport only; reading it as a working worker is prompt-send optimism. Prove engagement from observable state: session output, tool activity, changed files.
• Capture before restart. Before a nudge, replacement or restart, capture the worker's observable state. A restart destroys the evidence of why it stalled, and rescue is usually cheaper than rerun. Silence or impatience alone does not justify a restart; any replacement stays within authority and remaining bounds.
• Dispatch once. Keep each packet's identity with its result. An executor error is a result, not a retry trigger; repair and follow-up are the caller's authority. A launch call the runtime refuses, so that its worker never starts, is not a launch: correct a malformed call once, and treat a concurrency or capacity limit as backpressure, waiting natively for a slot and launching the validated batch in waves under it; what each freed slot runs first stays the caller's decision (Orchestrate). Running a packet inline or elsewhere is the caller's disclosed choice (refused launches).
• Account for every input. When a packet carries many items (files, records, issues), give each a stable ID before launch and join the returned results back to that list. An item with no result is missing, not done; an output with no matching input is unexplained. Report both by ID. Name each return field and its allowed values in the packet when results will be counted, sorted, deduplicated or compared; parsed prose is where counts drift.
• Prove integration before removal. Before a caller-authorized removal of a finished worker's workspace, branch or stash, compare its content with the integration target as it stands now, by patch or file diff, not by merged status or by matching the commit that was merged: squash and rebase hide ancestry. Preserve uncommitted, unique or unintegrated content in a recoverable form, such as an archive ref or bundle, first. Report what was kept, what was removed and the evidence for each.

Roles

• Orchestrator: passes focused intent, scope and evidence references, names the integration/final-review owner, and reports runtime facts. Retrieve extra history only for a consequential uncertainty; a fresh context is not necessarily small.
• Implementer: may modify only its packet's declared subject.
• Validator: receives exact candidate content in a fresh, read-only context and supplies judgment to Validate, which alone writes verdict.v2.

Reader and Writer are bounded cheap delegations, not roles with authority. A Reader returns line-referenced bullets over files the caller never loads, in slices of at most 350 lines. A Writer lands one patterned file from a spec plus a required reference file and returns a receipt the caller never reads back. Both are caller-selected per call and yield runtime facts only; a receipt is not validation. For Codex, use the source-owned bulk-reader and code-writer native roles; their model and effort are pinned in agents/bulk-reader.toml and agents/code-writer.toml, and each runs in the parent session's sandbox, which a role file cannot override. Context-budget delegation covers installation, native invocation, opt-in refusal hooks and the limits of role instructions.

Launch and observe

1. Start only a packet that passed the batch check above, whose field list is the one packet contract. Record the dispatch association in the caller's native work comments or metadata, or in existing runtime facts, before execution can fail. A requested ID is not an observed ID; worker identity stays unknown until the runtime reports it.
2. At startup, before substantive work, capture the observed runtime, session and context identity and return it to that same record. Before its first edit, the worker confirms its HEAD is that starting commit; a mismatch is a launch failure, never repaired by reset. Report launch failures, recording gaps and unknowns; never fill them in. An authentication-shaped error from a sandboxed launch does not prove a logged-out account (launch and isolation). Session associations owns parent and resume links, multi-work spans and frozen source bounds.
3. Keep concurrent writers disjoint and isolated, using the selected runtime's own isolation where it has one and never nesting a worktree inside the session's own (workspace isolation). Runtime coordination is not a claim, lease, queue or completion state in AgentOps.
4. Observe through native waits or status notifications; observation costs time and context. Unchanged state is no reason for another analysis, review or retrospective, while a known blocking failure deserves action even as other jobs run. Stop at terminal status or the end of the caller's window.
5. A headless worker cannot ask: on a choice its packet leaves open it takes no irreversible or out-of-scope step on a guess and returns the choice under gaps (open choices). For new authorized work after a worker completes, including an answer that settled intent supplies, use the runtime's documented follow-up or resume operation that starts a turn. A message operation only queues text for a running worker; a queued repair request is not a resumed attempt.
6. Record provider state, transcript references, artifacts and terminal status. Provider retries, reconnects, idle states and failures stay runtime facts, never Plan, Candidate or verdict state. A result whose observed tool activity shows an excluded action is out of authority, not a candidate, whatever its checks show.

Adapters and judgment

NTM, native processes, Agent Mail, Gas City and the one-shot headless runners (codex-exec, claude-exec, agy-native) are replaceable adapters. Use one only when the caller selected that execution shape; a single local agent pays no factory coordination cost.

Judgment defaults to a fresh context in the author's model family; cross-model legs are explicit caller selections. Model dispatch owns each leg's authorization, isolation, bounds and cleanup, and Agent Mail is never the judgment path. Role requests declare authority, but only native runtime/OS filesystem and egress controls enforce it. Native judgment receipts check caller-required model diversity; missing native identity never satisfies a leg. For exact native source spans, follow bounded raw source reads.

Per-packet result

packet_id:        the caller's packet id
executor:         selected runtime; requested model; observed model if reported
observed_session: session/context id the runtime reported, else unknown
engagement:       observable evidence the worker started, else unobserved
last_activity:    time and kind of the last observed tool call, output or file change, else unobserved
terminal_status:  exit or terminal state | running | launch failed | not launched
result:           candidate <ref> | evidence <ref> | out of authority <action> | error <verbatim>
gaps:             recording failures and unknowns