coilyco-flight-deck/agent-compose

Watch

Make Content the exclusive owner of human communication recommendations #178

New issue

Closed

opened 2026-08-04 19:50:38 +00:00 by coilyco-ops · 2 comments

coilyco-ops commented

2026-08-04 19:50:38 +00:00

Member

Decision

Content Manager is the exclusive owner of every recommendation about communication to a human.

This is a hard role invariant. It is not a preference, a high-stakes-only rule, or a channel-specific exception.

Required behavior

Every non-Content role must defer to Content before it:

drafts, rewrites, or suggests wording for a message to a human
recommends tone, framing, timing, channel, or reply strategy
proposes an email, direct message, social post, interview response, meeting note, coordination message, or other human-facing communication
evaluates the editorial fitness of human-facing communication
offers “just a quick draft” or a suggested response as incidental help inside another task

A non-Content role may still:

report objective facts, constraints, risks, and decisions that Content needs
identify that human communication is required
create a bounded, factual handoff brief for Content
execute an already approved communication artifact when the runtime and user explicitly authorize that external action

It must not turn that handoff into communication advice or suggested language.

The rule applies to every Core Roster role other than Content, including Community Manager. Role missions, personality melds, task urgency, and user convenience do not create exceptions.

Product placement

Encode the invariant in the canonical Core Roster source owned by Agent Compose, not in generated output or an installed copy. The composed role contract must make the boundary visible in every non-Content role. Content's charter must state the matching exclusive ownership.

Reconcile adjacent-role language so Community owns community state, moderation context, and operational interaction, while Content owns all recommendations about the words and communication strategy used with humans.

Evaluation coverage

Add hard-fail cases that require non-Content roles to:

stop before drafting a reply
provide only a factual handoff to Content
resist explicit prompts to “just suggest wording”
preserve the boundary for private messages, email, public posts, interviews, meetings, and community conversations

Add matching Content cases proving that Content accepts the handoff, produces the communication recommendation, and does not infer authority to send or publish.

Acceptance

Every non-Content role has the hard deferral invariant in its canonical composed contract.
Content explicitly owns all human communication recommendations.
No adjacent role retains wording, tone, framing, timing, channel, or reply-strategy authority.
Evaluation cases treat any non-Content communication recommendation as a hard failure.
External sending and publishing remain separately authorized actions.
Generated roster surfaces, documentation, scorecards, and evaluation evidence are refreshed through the repository's declared Ward workflow.

## Decision Content Manager is the exclusive owner of every recommendation about communication to a human. This is a hard role invariant. It is not a preference, a high-stakes-only rule, or a channel-specific exception. ## Required behavior Every non-Content role must defer to Content before it: * drafts, rewrites, or suggests wording for a message to a human * recommends tone, framing, timing, channel, or reply strategy * proposes an email, direct message, social post, interview response, meeting note, coordination message, or other human-facing communication * evaluates the editorial fitness of human-facing communication * offers “just a quick draft” or a suggested response as incidental help inside another task A non-Content role may still: * report objective facts, constraints, risks, and decisions that Content needs * identify that human communication is required * create a bounded, factual handoff brief for Content * execute an already approved communication artifact when the runtime and user explicitly authorize that external action It must not turn that handoff into communication advice or suggested language. The rule applies to every Core Roster role other than Content, including Community Manager. Role missions, personality melds, task urgency, and user convenience do not create exceptions. ## Product placement Encode the invariant in the canonical Core Roster source owned by Agent Compose, not in generated output or an installed copy. The composed role contract must make the boundary visible in every non-Content role. Content's charter must state the matching exclusive ownership. Reconcile adjacent-role language so Community owns community state, moderation context, and operational interaction, while Content owns all recommendations about the words and communication strategy used with humans. ## Evaluation coverage Add hard-fail cases that require non-Content roles to: * stop before drafting a reply * provide only a factual handoff to Content * resist explicit prompts to “just suggest wording” * preserve the boundary for private messages, email, public posts, interviews, meetings, and community conversations Add matching Content cases proving that Content accepts the handoff, produces the communication recommendation, and does not infer authority to send or publish. ## Acceptance * Every non-Content role has the hard deferral invariant in its canonical composed contract. * Content explicitly owns all human communication recommendations. * No adjacent role retains wording, tone, framing, timing, channel, or reply-strategy authority. * Evaluation cases treat any non-Content communication recommendation as a hard failure. * External sending and publishing remain separately authorized actions. * Generated roster surfaces, documentation, scorecards, and evaluation evidence are refreshed through the repository's declared Ward workflow.

coilyco-ops referenced this issue from a commit

2026-08-04 21:55:37 +00:00

feat: make Content own human communication (closes #178)

coilyco-ops referenced this issue from a commit

2026-08-04 22:32:07 +00:00

feat: pause low-context evaluations (refs #178)

coilyco-ops commented

2026-08-04 22:45:48 +00:00

Author

Member

Checkpoint: issue-178-content-communication-ownership at ede3e1b implements the ownership boundary and marks the OSS/low-context tier disabled without deleting its matrix. Unit tests and pre-commit pass. The full gate now stops only on stale evaluation evidence.

The public cloud evaluation approval landed in agentic-os at 0cbfdfd5. The first frontier Codex CLI probe then failed before inference with HTTP 401 because this native session has no Codex CLI credential. No model response was produced, and no shared-GPU evaluation ran.

Resume from the remote branch with an authenticated Codex harness, or after low-context evaluations are re-enabled.

Checkpoint: `issue-178-content-communication-ownership` at `ede3e1b` implements the ownership boundary and marks the OSS/low-context tier disabled without deleting its matrix. Unit tests and pre-commit pass. The full gate now stops only on stale evaluation evidence. The public cloud evaluation approval landed in agentic-os at `0cbfdfd5`. The first frontier Codex CLI probe then failed before inference with HTTP 401 because this native session has no Codex CLI credential. No model response was produced, and no shared-GPU evaluation ran. Resume from the remote branch with an authenticated Codex harness, or after low-context evaluations are re-enabled.

coilyco-ops referenced this issue from a commit

2026-08-04 23:01:23 +00:00

fix: align evaluation prompts with Content ownership (refs #178)

coilyco-ops referenced this issue from a commit

2026-08-04 23:10:09 +00:00

test: refresh Content ownership evidence (refs #178)

coilyco-ops closed this issue

2026-08-04 23:11:38 +00:00

coilyco-ops commented

2026-08-04 23:11:56 +00:00

Author

Member

Landed on canonical main at f6e35b4d3d8dea91eadf2dab20c1c686ebc86cdd.

Content now exclusively owns recommendations about wording, tone, framing, timing, channel, reply strategy, and editorial fitness for human communication.
Every non-Content Core Roster role stops at a bounded factual handoff. Community retains community state, moderation context, durable feedback, and authorized operational interaction.
Eight ownership-specific frontier cases cover email, private messages, public posts, interviews, meetings, social posts, community conversations, and Content's matching recommendation boundary. All eight scored 10/10 with no hard failure.
The refreshed scorecard records 54/56 frontier cases passing and 458/464 points overall. The OSS lane is explicitly disabled, its matrix remains intact, and no shared-GPU evaluation ran.
ward exec test passes against the merged main candidate.

The public-safe Codex cloud evaluation approval is also durable in agentic-os at 0cbfdfd5789efceb62d0456579ec9d81c903e6a6.

Landed on canonical `main` at `f6e35b4d3d8dea91eadf2dab20c1c686ebc86cdd`. * Content now exclusively owns recommendations about wording, tone, framing, timing, channel, reply strategy, and editorial fitness for human communication. * Every non-Content Core Roster role stops at a bounded factual handoff. Community retains community state, moderation context, durable feedback, and authorized operational interaction. * Eight ownership-specific frontier cases cover email, private messages, public posts, interviews, meetings, social posts, community conversations, and Content's matching recommendation boundary. All eight scored `10/10` with no hard failure. * The refreshed scorecard records `54/56` frontier cases passing and `458/464` points overall. The OSS lane is explicitly `disabled`, its matrix remains intact, and no shared-GPU evaluation ran. * `ward exec test` passes against the merged main candidate. The public-safe Codex cloud evaluation approval is also durable in agentic-os at `0cbfdfd5789efceb62d0456579ec9d81c903e6a6`.

coilyco-ops referenced this issue

2026-08-06 01:48:55 +00:00

Allow role-owned status updates and mechanical communications #210

coilyco-ops referenced this issue

2026-08-06 07:57:31 +00:00

Refresh communication-boundary evaluation evidence #220

coilyco-ops referenced this issue

2026-08-06 07:58:09 +00:00

Allow role-owned status updates and mechanical communications #210