From f0d71d3498cd5d9326fad8ef8a23804604fd317f Mon Sep 17 00:00:00 2001 From: garrytan Date: Mon, 2 Nov 2026 06:25:21 +0000 Subject: [PATCH] changelog: GST-938 fragment wording Co-authored-by: capy-ai[bot] <230910855+capy-ai[bot]@users.noreply.github.com> --- changelog.d/gsta251-investigate-drop-steps.md | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/changelog.d/gsta251-investigate-drop-steps.md b/changelog.d/gsta251-investigate-drop-steps.md index 95d9297de..fd8393d91 100644 --- a/changelog.d/gsta251-investigate-drop-steps.md +++ b/changelog.d/gsta251-investigate-drop-steps.md @@ -1,2 +1,2 @@ -- **`/investigate` drops its five-phase step procedure** (GST-938). Harness Evals watched Opus 5.5 run `/investigate` on 46 bug-fix sessions: the procedure loaded every time and the model followed none of its steps (0 of 30 with Claude Code's real system prompt), while a gstack session cost $0.46 against $0.22 without gstack (HARN-14, gn-hard-follow-2026-11). The skill keeps what still does a job: the Iron Law, the arguments table, a short "How to work" with the three lessons that cost real time before (read call sites before calling two paths duplicates, refresh the git index before timing, test a shared premise before a third fix), the 3-strike and more-than-5-files asks, the Scope Lock with its freeze hooks, the recipes and `--backlog`, the prior-learnings search and the DEBUG REPORT. The generated Claude skill shrinks from 29.8 KB to 20.1 KB. Gone: the Phase 1-5 walkthrough, the pattern table, the web-research block and the hypothesis-keyed learnings re-pull. +- **`/investigate` drops its five-phase step procedure** (GST-938). Harness Evals watched Opus 5.5 run `/investigate` on 46 bug-fix sessions: the procedure loaded every time and the model followed none of its steps (0 of 30 with Claude Code's real system prompt), while a gstack session cost $0.46 against $0.22 without gstack (HARN-14, gn-hard-follow-2026-11). The skill keeps the parts that still do a job. Those are the Iron Law, the arguments table, a short "How to work" with the three lessons that cost real time before (read call sites before calling two paths duplicates, refresh the git index before timing, test a shared premise before a third fix), the 3-strike and more-than-5-files asks, the Scope Lock with its freeze hooks, the recipes and `--backlog`, the prior-learnings search and the DEBUG REPORT. The generated Claude skill shrinks from 29.8 KB to 20.1 KB. Gone: the Phase 1-5 walkthrough, the pattern table, the web-research block and the hypothesis-keyed learnings re-pull. - Behavior change: the model is no longer told to work through Phases 1-5 in order, and `/investigate` no longer carries the Aside web-research block. Hook-enforced versions of the two steps that could change results (reproduce first, verify after) are planned as an opt-in Claude Code candidate that ships only if a preregistered eval shows it helps.