AGENTS.md declares no agent-to-agent channel exists, and one is live in the Claude Code harness #1324

New issue

Open

opened 2026-08-27 09:50:52 +00:00 by coilyco-ops · 0 comments

coilyco-ops commented

2026-08-27 09:50:52 +00:00

Owner

The defect

AGENTS.md:214 states, as operative doctrine composed into every
session on every public and work host:

That covers a human recipient. There is no autonomous agent-to-agent
command channel: the o2r relay was archived in the June 2026 surface
reduction (agentic-os-kai#677, revival tracked at ward#104). Delivery is
human-mediated - route the request through the human, who relays it
upstream, with no command crossing an agent boundary on its own.

The Claude Code harness now ships SendMessage and ListAgents as first-class
tools. ListAgents enumerates in-process subagents, teammates, other local
Claude sessions on the same machine, cloud sessions, and Remote Control sessions
on other machines. SendMessage addresses any of them by name and delivers
without a human in the path.

I confirmed this from inside a session on this host by loading the tool schema,
not by inference. The sentence is false as written for this harness.

Why it matters rather than being a stale footnote

This is not a dead cross-reference in a doc nobody opens. It is in the composed
base, so it is in front of every agent on every session, and it tells them a
capability they are holding does not exist. An agent that reads it will route
work through Kai that it could have routed itself, which is exactly the
human-in-the-loop tax the rule was written to justify when the capability really
was absent.

Second defect, in the same sentence

ward#104 does not track an o2r revival. It is
Add ward ops passthrough surface (blocker for coily removal),
opened 2026-06-17 and closed 2026-07-01 under the coily-to-ward consolidation
milestone. Read literally, the citation resolves to an unrelated closed issue.

I am reporting this as observed rather than diagnosed. It may be a miscite, it
may point at a different repository's #104, or the revival tracker may have
moved. Whoever fixes the sentence should establish which before rewriting the
reference, rather than deleting it and losing the thread.

What the fix is not

Do not simply invert the sentence. Three things stay true and the rewrite has to
carry them, or it trades a false rule for a reckless one.

Cross-session messaging carries no authority. The SendMessage contract
says so itself: permission boundaries are per-session, and asking a peer to
perform an action that was denied or blocked in your own session is
cross-session permission laundering. That prohibition maps directly onto the
estate's boundary model, where a seat that defers a boundary does not acquire
it by asking a seat that holds it.
It is one harness. SendMessage is Claude Code. The roster runs seats on
codex, openhands, goose, holmesgpt, plandex, hermes, anythingllm, mixpost,
penpot, and discord. A rule written as though every agent can be messaged is
wrong everywhere except one harness.
Command delivery to a human is untouched. The five-tier rule above it
about committing reusable scripts and handing back paths is about a human
recipient and stays exactly as it is.

Acceptance criteria

AGENTS.md:214 describes the actual state: a harness-scoped agent-to-agent
channel exists in Claude Code, carries no authority, and does not exist on
the other harnesses the roster runs.
The permission-laundering prohibition is stated, not implied.
The ward#104 reference either resolves to the right issue or is removed
with the o2r archival still traceable through agentic-os-kai#677.
The kai-command-handover skill is reconciled with the rewritten section,
or an issue is filed against whichever repository owns it.
dead-cross-links and the catalog suite pass.

Ownership

Filed from the Portfolio Director seat. AGENTS.md is doctrine rather than
running software, so this is not a boundary-build-foundational-software
handoff. Whoever picks it up needs the answer to the ward#104 question before
the rewrite lands.

housecast#1 - unrelated, filed the same session.
The companion issue in this repository adding a skill for the harness
orchestration surface this rule is currently wrong about.

## The defect [`AGENTS.md:214`](AGENTS.md) states, as operative doctrine composed into every session on every public and work host: > That covers a **human** recipient. There is **no autonomous agent-to-agent > command channel**: the o2r relay was archived in the June 2026 surface > reduction (`agentic-os-kai#677`, revival tracked at `ward#104`). Delivery is > **human-mediated** - route the request through the human, who relays it > upstream, with no command crossing an agent boundary on its own. The Claude Code harness now ships `SendMessage` and `ListAgents` as first-class tools. `ListAgents` enumerates in-process subagents, teammates, other local Claude sessions on the same machine, cloud sessions, and Remote Control sessions on other machines. `SendMessage` addresses any of them by name and delivers without a human in the path. I confirmed this from inside a session on this host by loading the tool schema, not by inference. The sentence is false as written for this harness. ## Why it matters rather than being a stale footnote This is not a dead cross-reference in a doc nobody opens. It is in the composed base, so it is in front of every agent on every session, and it tells them a capability they are holding does not exist. An agent that reads it will route work through Kai that it could have routed itself, which is exactly the human-in-the-loop tax the rule was written to justify when the capability really was absent. ## Second defect, in the same sentence `ward#104` does not track an o2r revival. It is [Add ward ops passthrough surface (blocker for coily removal)](https://forgejo.coilysiren.me/coilyco-flight-deck/ward/issues/104), opened 2026-06-17 and closed 2026-07-01 under the coily-to-ward consolidation milestone. Read literally, the citation resolves to an unrelated closed issue. I am reporting this as observed rather than diagnosed. It may be a miscite, it may point at a different repository's #104, or the revival tracker may have moved. Whoever fixes the sentence should establish which before rewriting the reference, rather than deleting it and losing the thread. ## What the fix is not Do not simply invert the sentence. Three things stay true and the rewrite has to carry them, or it trades a false rule for a reckless one. * **Cross-session messaging carries no authority.** The `SendMessage` contract says so itself: permission boundaries are per-session, and asking a peer to perform an action that was denied or blocked in your own session is cross-session permission laundering. That prohibition maps directly onto the estate's boundary model, where a seat that defers a boundary does not acquire it by asking a seat that holds it. * **It is one harness.** `SendMessage` is Claude Code. The roster runs seats on codex, openhands, goose, holmesgpt, plandex, hermes, anythingllm, mixpost, penpot, and discord. A rule written as though every agent can be messaged is wrong everywhere except one harness. * **Command delivery to a human is untouched.** The five-tier rule above it about committing reusable scripts and handing back paths is about a human recipient and stays exactly as it is. ## Acceptance criteria - [ ] `AGENTS.md:214` describes the actual state: a harness-scoped agent-to-agent channel exists in Claude Code, carries no authority, and does not exist on the other harnesses the roster runs. - [ ] The permission-laundering prohibition is stated, not implied. - [ ] The `ward#104` reference either resolves to the right issue or is removed with the o2r archival still traceable through `agentic-os-kai#677`. - [ ] The `kai-command-handover` skill is reconciled with the rewritten section, or an issue is filed against whichever repository owns it. - [ ] `dead-cross-links` and the catalog suite pass. ## Ownership Filed from the Portfolio Director seat. `AGENTS.md` is doctrine rather than running software, so this is not a `boundary-build-foundational-software` handoff. Whoever picks it up needs the answer to the `ward#104` question before the rewrite lands. ## Related * [housecast#1](https://forgejo.coilysiren.me/coilyco-flight-deck/housecast/issues/1) - unrelated, filed the same session. * The companion issue in this repository adding a skill for the harness orchestration surface this rule is currently wrong about.

coilyco-ops referenced this issue

2026-08-27 10:03:53 +00:00

feat(skills): add tooling-agent-orchestration carrying Ward and o2r patterns #1326

coilyco-ops referenced this issue

2026-08-27 10:04:59 +00:00

Add tooling-harness-orchestration: the catalog has no skill for the harness-native orchestration surface #1325