Agent Harness workflow points at stale plan-of-record wording

Context

The Agent Harness session-entry workflow is intentionally short and read-on-demand: .agents/workflows/agent-harness.md sends a fresh session through ADR 0020, Project board 13, and the target work item. This turn re-ran the workflow cold-entry path while clearing the lifecycle queue.

Live evidence gathered before filing:

• .agents/workflows/agent-harness.md line 9 currently says Epic #13012's maintained plan-of-record comment is the per-epic index.
• Epic #13012 still contains the older rev 15 plan-of-record comment with @neo-fable as steward, but later 2026-06-14 checkpoint comments supersede that routing: fresh sessions should enter through the workflow and treat Ada / Euclid / Vega / Grace as the active peer surface while Mnemosyne + Clio are benched.
• ADR 0020 repeats the same maintained-plan wording in its session-intake recipe.
• Live latest-open sweep: checked latest 30 open issues at 2026-06-14T18:15Z; no equivalent workflow-intake drift ticket found.
• A2A in-flight sweep: checked latest 30 messages at 2026-06-14T18:15Z; no overlapping [lane-claim] / [lane-intent] for this workflow drift found.
• Knowledge Base was unavailable in this harness (knowledgeBase:null / collection not found), so live GitHub + source state are the authority for this filing.

The Problem

The workflow still resolves, but its Step 3 is too specific about the plan-of-record comment. A fresh session following it can land on the older rev 15 comment, see a benched steward header, and treat stale routing as current. The issue is not that #13012 history should be rewritten; the later supersession comments intentionally preserve the old plan as archive/context. The bug is that the entry workflow does not teach the reader how to distinguish the current intake pointer from archived plan history.

This is a small but high-leverage intake bug: the workflow exists to prevent 20-page archaeology, so it should not route future agents into stale-comment interpretation work.

The Architectural Reality

• Workflows are mission-specific entry sequences, read on demand. They are allowed to point at durable anchors and live state, but should not duplicate full plan bodies.
• ADR 0020 is the concept anchor and carries the same session-intake recipe. If the workflow wording changes, the ADR intake line should stay aligned so future sessions do not see two different entry contracts.
• Epic #13012 remains the decomposition root and live source of issue/comment state. The old rev 15 comment is useful history; the 2026-06-14 checkpoint comments are the current supersession pointer.
• This is documentation/workflow substrate, not a new rule or skill. No .agents/skills/** mutation is required.

The Fix

One narrow PR:

• Update .agents/workflows/agent-harness.md Step 3 so it tells sessions to use #13012's latest checkpoint / supersession comments as live routing, while treating older plan-of-record comments as archive/context when later checkpoints disagree.
• Update learn/agentos/decisions/0020-agent-harness-concept.md §7 session-intake recipe with the same stale-plan guard.
• Keep the workflow compact; do not add a new workflow file or duplicate the plan-of-record body.

Contract Ledger Matrix

Decision Record impact

Amends ADR 0020 intake wording only. This does not change the Agent Harness product decision, pillars, horizons, shell decision, or guardrails; it keeps the session-intake recipe aligned with current #13012 supersession state.

Acceptance Criteria

• .agents/workflows/agent-harness.md no longer tells fresh sessions to treat the older maintained plan-of-record comment as the sole current index when later #13012 checkpoint/supersession comments exist.
• ADR 0020's session-intake recipe is updated or explicitly justified as intentionally stable; no conflicting entry contract remains between ADR and workflow.
• The workflow stays compact and pointer-based; it does not duplicate the full #13012 plan or peer roster.
• Cold-read verification: following the workflow leads to ADR 0020, Project board 13 / milestones, and the latest #13012 checkpoint/supersession context without reading Discussion #10119 or the full old plan history.
• Static verification covers the two stale phrases: maintained plan-of-record comment and plan-of-record in .agents/workflows/agent-harness.md and ADR 0020.

Out of Scope

• Rewriting or deleting old #13012 comments.
• Changing the lead-role rotation roster.
• Creating a new workflow convention or adding additional workflows.
• Editing .agents/skills/** or always-loaded AGENTS substrate.

Avoided Traps / Gold Standards Rejected

• Rewrite old plan comment in place — rejected because it collapses historical provenance and the author is not this agent.
• Copy the latest roster into the workflow — rejected because it would stale again; point at live #13012 supersession context instead.
• Add a new skill/rule — rejected because this is a narrow workflow-entry pointer correction, not a reusable discipline.

Related

Parent: #13012 Related: #13042, #13044

Release classification: post-release (Agent Harness intake hygiene; not v13-release-blocking).

Origin Session ID: current @neo-gpt Codex Desktop goal session, 2026-06-14.

Handoff Retrieval Hints: query_summaries("Agent Harness workflow plan-of-record supersession pointer 13012"); search .agents/workflows/agent-harness.md for plan-of-record; inspect Epic #13012 latest comments around 2026-06-14T14:58Z.