plat 263

← Pulse platform index

PLAT-263 — Retire the redundant per-step learning lock

Coordination	Value
Assigned agent	Codex
Ticket state	implemented
Last synchronized	2026-08-31

• Priority: product simplification, severity medium.
• Related: PLAT-055, PLAT-059, PLAT-258), PLAT-260), PLAT-265.

Problem

The platform represented “this step may read shared learnings but must not write them” twice: learnings_access="read" and lock_learnings=true layered over read-write access. The overlap produced a second reason field, a lock audit/version-upgrade turn, Builder and HTTP mutation paths, runtime bootstrap exceptions, deletion-time unlocking, and a separate Pulse lifecycle proposal. Those paths could disagree even though the desired state was already expressible directly.

Implemented contract

• learnings_access is the only learning permission:
  • read consumes shared guidance without a contribution turn;
  • read-write also contributes and requires a concrete learning_objective;
  • none neither reads nor writes.
• lock_code remains independent and continues to protect proven scripted main.py artifacts.
• Pulse and plan-drift review judge whether read-write access is still useful; they do not create or manage a second learning lock.

Migration

AgentConfigs retains the two old JSON names only as migration inputs. On step-config read/repair:

• legacy lock_learnings=true becomes learnings_access="read";
• an explicit learnings_access="none" remains none (migration never widens access);
• legacy false locks and all legacy reasons are removed;
• normal persistence writes no retired keys.

The HTTP config endpoints reject new attempts to set either retired field and point callers to learnings_access="read".

Removed surfaces

• runtime lock checks, bootstrap exception, metadata projection, and background completion hints;
• Builder tool schema/handler/status prompts and the frontend lock toggle;
• delete-learning “unlock” requests;
• the obsolete 1.0.22 learning-lock audit upgrade turn;
• current guidance that recommended lock/unlock workflows.

Verification

• Migration tests cover true → read, explicit none preservation, false-lock removal, and clean serialized JSON.
• Runtime tests prove read-only access suppresses message-sequence learning turns.
• Tool-schema and HTTP tests ensure retired fields cannot be newly authored.
• Focused Go tests, the full Go build, frontend typecheck/production build, and repository-search checks passed before handoff.