v0.4.3

One fix to the premortem skill, measured before and after, and two new columns in the harness that made the problem visible.

os-what-could-go-wrong hands the fresh agent's report over whole. Step 3 used to say "pass the report through as it came". Measured on 2026-09-13 across three briefs and three runs each, Sonnet 5 shrank the fresh agent's 18-25k-character report to 0.8-8.5k characters in its final message, Haiku 4.5 22k to 4k, and Opus 5 passed it through. A first rewrite that only said "the final message is the report" changed nothing on Sonnet; one run pointed the reader to "the agent's report above". The model assumes the user can see a tool result. Step 3 now says that the user never sees what the agent returned, that the final message is that report copied whole, first line to last, with one line before it naming the depth and whether a fresh agent ran, and anything else after it, never instead of it; hard rule 8 repeats the point. Re-measured on 2026-09-14: Opus 5 copies 91-96% of the agent's lines (43-97% before), dispatched a fresh agent in 8 of 9 runs (5 of 9 before) and named the planted flaw in all 6 runs that carry it (5 of 6 before). Sonnet 5 now keeps the headings and every risk card but rewrites their bodies shorter: 11-35% of the lines copied (2-10% before), and its own fresh agent found the flaw in 1 run of 6. Haiku 4.5 was not re-measured. SKILL.md stays at 150 lines.

The harness now sees the two things a well-shaped report hides. A "Fresh agent" column: how many runs handed the brief to a fresh agent, an Agent call that was not denied. A "Report copied" column: the share of that agent's lines that reach the final message unchanged. The restraint check follows the skill's own Quick look, which allows three to five risk cards: it fails above five or on a heavy verdict, and the line names the rule that fired; the 2026-09-13 Opus result that failed at 4.0 cards was the harness being stricter than the skill. evals/results.md gains the premortem section. Tests: evals/test.sh 53 checks, hooks/test.sh 53 checks, one of them pinning the step 3 sentences and the 150-line ceiling. Closes #14: levels 1-3 are measured; level 4, a backtest against decisions with known outcomes, is not built.

A known environment effect, not a skill defect. On the measuring machine Opus 5 sometimes declines to dispatch, saying the session's instructions forbid the Agent tool unless asked for: 4 of 9 runs on 2026-09-13, 1 of 9 on 2026-09-14. The skill then says so in its first line and writes the report itself. The likeliest source is a global instruction against creating subagents ahead of need; that is a candidate, not a proven cause.

Update: git pull inside your clone, then claude plugin update open-steps@open-steps.