Follow-up to the earlier false-positive reduction work (PR #1). A run on 2026-08-21 staged a proposal reporting 89 unrouted correction signals, but roughly a third are false positives from three remaining classes:
-
System scaffolding as user messages — Codex rollout files include <app-context>, <multi_agent_mode>, and <skills_instructions> blocks emitted as user-role messages at session start; all were flagged as user_correction (e.g. rollout-2026-08-02T10-10-53-019fc2d0…jsonl lines 3–5, 1281).
-
Duplicated user events — every user message in the Codex rollout format appears twice (adjacent lines, identical text), so each real correction is double-counted (e.g. lines 693/694, 742/743 in the same file).
-
Non-corrective messages — plain acknowledgements ("ok", "done", "approved", "great"), pasted JSON payloads, and bare URLs are counted as corrections.
Suggested fixes: skip user messages that start with a known scaffolding tag or match the session-preamble position; dedupe consecutive identical user messages before scoring; require some corrective signal (negation, imperative revision verb, contrast marker) rather than counting every user message after a tool result.
🤖 Generated with Claude Code
Follow-up to the earlier false-positive reduction work (PR #1). A run on 2026-08-21 staged a proposal reporting 89 unrouted correction signals, but roughly a third are false positives from three remaining classes:
System scaffolding as user messages — Codex rollout files include
<app-context>,<multi_agent_mode>, and<skills_instructions>blocks emitted as user-role messages at session start; all were flagged asuser_correction(e.g.rollout-2026-08-02T10-10-53-019fc2d0…jsonllines 3–5, 1281).Duplicated user events — every user message in the Codex rollout format appears twice (adjacent lines, identical text), so each real correction is double-counted (e.g. lines 693/694, 742/743 in the same file).
Non-corrective messages — plain acknowledgements ("ok", "done", "approved", "great"), pasted JSON payloads, and bare URLs are counted as corrections.
Suggested fixes: skip user messages that start with a known scaffolding tag or match the session-preamble position; dedupe consecutive identical user messages before scoring; require some corrective signal (negation, imperative revision verb, contrast marker) rather than counting every user message after a tool result.
🤖 Generated with Claude Code