plat 076

← Pulse platform issue index

PLAT-076 — learning and scripted metadata record claims instead of runtime facts

Coordination	Value
Assigned agent	Codex
Ticket state	implemented — runtime reverify pending
Last synchronized	2026-08-10

• Priority: P1 — incorrect provenance can drive locking, maturity and optimization decisions
• Owner: learning detection and scripted metadata persistence
• Findings: build-in-public 03f6ed727fb67c9a, 9d1c0fe414871e37, 10eb995c50fb52c0

Evidence and root causes

The full workflow-local records and retained artifacts reproduced four defects:

1. step-route-selector/script_metadata.json recorded four successful scripted_fast_path executions under successful_runs.agentic. The writer literally incremented the "agentic" key for every validated saved-script execution.
2. Seven of nine .learning_metadata.json files retained creation-time step_path values after plan reordering. Both step-route-selector and step-reddit-scan-draft claimed step-1. The writer only assigned identity when creating or recovering an unreadable file.
3. has_new_learning was inferred from phrases such as “files changed” in the agent's response. A KB/DB-only change could therefore be recorded as a learning change even when the managed learning tree was byte-identical.
4. script_version and relearn_count increased every time execution code was copied back, including when main.py was byte-identical. The copy routine already held both versions but incremented before comparing them.

Fix

• Normalize saved-script success provenance to the single "scripted" key. Existing mirrored legacy keys use total_runs - failed_runs, avoiding double counting.
• Refresh step_id and the current positional step_path on every learning metadata write.
• Derive has_new_learning from before/after hashes of the complete managed learning artifact tree for normal reflection, recovered reflection and message-sequence learning turns. Agent prose is no longer evidence of a write.
• Increment script version/relearn count only when the saved script artifact bytes actually differ, including added or removed helper files.

Regression tests use the captured build-in-public metadata shapes and cover legacy mirrored counters, stale identity, identical scripts and real artifact changes.

Verification

• Focused regression tests: pass.
• go build ./...: pass.
• Required broad suite: exactly 22 known baseline failures; no additional failures.
• Runtime reverify: pending server restart and a scripted plus learning-writing run.

Ready to close after runtime verification: 03f6ed727fb67c9a, 9d1c0fe414871e37, 10eb995c50fb52c0.

Social-media dd2a48047c4d7993 required no second fix: PLAT-061 already removed the dead-field shape, and the current v1.0.22 workflow no longer contains global_skill_objective. It is ready to close as covered by PLAT-061 after the normal post-restart check.