v0.1.7

• Add opt-in readOnlyCommands shell patterns and readOnlySkills skill IDs for explore, reviewer, and scout. Other commands and skills stay denied, including the fusion skill.
• Include configured shell patterns in lane prompts and lead guidance; defaults keep the existing prompts and no-shell guidance.

v0.1.6

• Raise the sidekick default from 40 to 80 steps. Over 30 days of recorded sessions, 62 of about 350 sidekick reports (18%) stopped at the 40-step limit. Finishing the work took a median of 23 more steps (75th percentile 34) after the lead resumed it, and each resume cost an extra lead turn.

v0.1.5

• The fusion skill now says how to redirect a background lane mid-run: call subagent again with its session ID. OpenCode steers the message into the running child, and the parent still gets a single report (checked live on OpenCode 2.1.4).
• Type-check and test against @opencode/plugin 2.1.4. The minimum supported OpenCode version is still 2.0.26.

v0.1.4

• Add the lead guidance to session generation requests (session.generate) too. They replay the agent's request, so without the same system section they missed the session's prompt cache after the tool definitions. In a live test, a cache-reusing compaction read 4,806 of 22,874 input tokens from cache before this change and 22,104 of 23,596 after.

v0.1.3

• Raise the reviewer default from 30 to 40 steps, the same as sidekick. In 63 recorded reviews, 7 hit the old 20-step limit, and the resumed ones needed about 26, 29, and 32 steps in total.

v0.1.2

• Raise the reviewer default from 20 to 30 steps; reviews of large diffs were running out of steps before the write-up.
• explore, reviewer, and scout no longer stop when repository instructions require a tool they don't have, such as a shell-only search CLI; they use their own tools instead.
• Releases are published to npm from version tags with trusted publishing and provenance.