github nyldn/claude-octopus v9.3.0
v9.3.0 — Agent Execution Improvements

latest releases: v9.18.0, v9.17.1, v9.17.0...
16 days ago

What's New

Three invisible-to-users improvements that make individual agents smarter:

Search Spiral Guard

Research agents get a prompt-level instruction preventing them from looping through searches without synthesizing. Stops token waste on search-heavy tasks.

Per-Role Token Budget Proportions

enforce_context_budget() now scales by role — implementers/researchers get 60%, planners/reviewers 40%, verifiers/synthesizers 25%. Prevents one chatty agent from starving others in multi-agent workflows.

Heuristic Learning

Records file co-occurrence patterns from successful agent runs. Before future runs, injects hints like "When modifying auth.ts, successful runs usually first read types.ts." Builds knowledge over time with zero user involvement. Kill switch: OCTOPUS_HEURISTIC_LEARNING=off.

Also Included (merged before release)

  • fix: prepend subagent preamble to Codex exec dispatches (#178) — closes #176
  • fix: resolve GEMINI_API_KEY in non-interactive shells (#179)
  • deps: bump vendors/ui-ux-pro-max-skill (#180)

Full Changelog: v9.2.2...v9.3.0

Don't miss a new claude-octopus release

NewReleases is sending notifications on new releases.