Prompt Relay Stewardship Audit

What changed

Codex reviewed the repo's current prompt system, recorded DEC-0020 to assign prompt-relay stewardship to Codex, updated AGENTS.md to reflect that default ownership, and corrected stale develop references in prompt patterns so new handoffs match the repo's development branch model.

Why this changed

The prompt relay is no longer just a handoff folder. It is now the durable task ledger, external memory layer, and emerging PM substrate for the agent system. That makes drift in its docs, patterns, and validation assumptions more serious than ordinary docs rot.

The human asked Codex to get acquainted with the changed system, analyze it, provide feedback, and take ownership. This audit is the durable record of that handoff.

Assessment

Strong moves

The system has the right primitive. Prompts are a better durable unit than Plan mode, IDE todos, or chat memory for multi-agent work.
The schema expansion points in the right direction. work_type, verdict fields, model attribution, and cost/turn fields can support real operational learning.
The rejection and deferral lifecycle closes an important hole. Work can now be returned or intentionally held instead of being forced into false completion.
Telemetry and the internal dashboard give the PM layer somewhere concrete to land instead of keeping it purely conceptual.

Current risks

Validation debt is real. make prompt-validate TARGET=all reported 59 issues across 127 prompt files on 2026-04-03. The system cannot be treated as fully normalized while that remains true.
PM metadata adoption is still sparse. python3 scripts/agent-telemetry.py prompts showed only 3 prompts with work_type, 1 with routing/output verdicts, 1 with model, and 0 with estimated_cost_usd or execution_turns.
Documentation lag exists in load-bearing places. Before this audit, prompt patterns still told agents to pull develop, and AGENTS.md summaries lagged the live lifecycle states and CLI surface.
Dashboard confidence can outrun data quality. The ops center is valuable, but it should not imply cleaner process data than the underlying prompt corpus actually provides.

Follow-up

Normalize or explicitly grandfather legacy prompt records.
Decide which v1.1 metadata fields are mandatory, optional, or best-effort.
Keep prompt patterns tightly aligned with real branch/worktree practice.
Treat prompt-quality audits as routine maintenance, not as a one-time pass.