0052: The Codex status footer is prompted, not enforced

Status: accepted (2026-07-29) · Scope: product tier (core/agents, hooks)

Context

ADR 0051 stopped Harnery from replacing Codex answers with coordination retries. It removed the visible status box from Codex instructions at the same time. That protected the answer, but it also removed a useful view of the agent’s current task, context usage, file claims, and nearby peers.

A Codex Stop hook cannot safely enforce the box. Blocking Stop creates a model continuation, and some clients retain only the last response from that continuation chain. A hook also cannot edit the assistant message that already exists. The box therefore has to be part of the model’s original response.

Static instructions alone are easy to miss late in a long turn. Harnery already injects one UserPromptSubmit context envelope for peer changes and task reminders, which gives it a safe place to put a fresh presentation instruction before the model starts work.

Decision

Human-facing Codex sessions receive a status-footer reminder on every UserPromptSubmit. The reminder tells the model to:

1. finish the user’s request first;
2. run <bin> agents status as its final shell call; and
3. append the command’s stdout verbatim in a fenced block at the bottom of the same substantive reply.

renderPromptContext does not hash-deduplicate this section. Peer and task sections keep their existing deduplication, but the footer instruction appears on every prompt so it remains close to the response being written. Subagents, transient sessions, and workflow children do not receive it.

Codex Stop remains observe-only. Missing the footer ends the turn normally. evaluateStopHook and emitStopBlock keep the safeguards from ADR 0051, so Harnery cannot create a status-only continuation.

Alternatives considered

Leave the status box disabled on Codex. Safe, but it discards useful coordination information that fits naturally after the answer.

Restore Stop enforcement. Rejected because it restores the original data-loss path. Even enforcing only the status call or task declaration can create the same replacement continuation.

Rely only on the repository instructions. Rejected because those instructions may be far from the response point after a long tool-heavy turn. The per-prompt reminder is small and arrives at the right time.

Have the hook append the box itself. Not available in the Codex hook contract. A Stop hook can allow or continue the turn, but it cannot mutate the assistant response.

Consequences

The footer is reliable guidance rather than a hard guarantee. A model can still miss it, but that failure leaves the requested answer untouched. The next prompt supplies the reminder again.

The status command adds one short tool call per human-facing Codex turn. Headless work avoids that cost. Claude Code and Cursor keep their existing end-of-turn behavior.