Skip to content

v2.71.0 — Opus 4.8 agent-layer tuning

Choose a tag to compare

View all tags
@avelikiy avelikiy released this 13 Jun 10:12
v2.71.0
This tag was signed with the committer’s verified signature.
avelikiy Alex
SSH Key Fingerprint: nD9xQzAb/3YFTUEHwDpi26wMHeq89yGJsV5XW/FadYQ
Verified
Learn about vigilant mode.
5149b2e
This commit was signed with the committer’s verified signature.
avelikiy Alex
SSH Key Fingerprint: nD9xQzAb/3YFTUEHwDpi26wMHeq89yGJsV5XW/FadYQ
Verified
Learn about vigilant mode.

Tuned the agent layer for Claude Opus 4.8 (and the Fable/4.x family)

Four behavioral-delta fixes (plan: docs/plans/PLAN-opus48-tuning.md):

  • Review recall — /review separates finding (coverage: surface every issue + confidence, no self-filter) from filter (skeptical triage). 4.8 under-reports low-severity/uncertain findings when told to 'report blockers'; this restores recall without more false-positive gate blocks.
  • Effort = xhigh — board agent runner sets CLAUDE_EFFORT=xhigh (was stripped → under-thinking) + CLAUDE_CODE_MAX_OUTPUT_TOKENS=64000, both overridable. senior-dev/architect → XHIGH.
  • Subagent fan-out steering — 4.8 under-spawns; architect + senior-dev get explicit fan-out guidance.
  • Frontend-aesthetics guard — steer off 4.8's editorial cream/serif default (wrong for regulated dashboards); architect sets UI direction in ARCH + <frontend_aesthetics> guard; /review flags the default.
Assets 2
Loading