[renmengye]

Release 0.6.0.

• __version__, CITATION.cff (version and date) and the pinned-version test move to 0.6.0.
• The [Unreleased] entries move under ## [0.6.0] - 2026-10-26, with an ### Upgrading summary at the top, in the same layout as 0.5.0.

Since 0.6.0rc1: OUTERLOOP_AUTHOR_EFFORT sets the author's effort level on every backend (#598). The label-triggered advisory review now covers fork PRs (#599, CI configuration only). Everything else is as in #597.

State that outlives a release (RELEASING step 1; each PR carries its compatibility statement, fixture and Upgrading: line):

• trajectory file (new): older kernels ignore it; old runs read as "no trajectory".
• session events gain role: older events count as author sessions.
• inbox: the new sibling: notes use the existing note kind.
• ledger: pending records are unchanged; the reader now ignores unknown fields.
• review envelopes: head_sha added; old envelopes still post.
• OUTERLOOP_AUTHOR_EFFORT: set a Claude author's effort level #598 adds no state: the effort setting is read when each session starts and is not stored.

The legacy fixtures run in the gate on this head.

Canary refusals (RELEASING step 3): the 0.6.0rc1 fleet test ran three author runs (Claude on claude-opus-5-5, quickstart-trial #25). outerloop trajectory reports 0 refusals in each (0 of 1 decisions per run). All three ended as negative results: the issue's idea is already on main, and after three claims the issue went to a human, as designed. 0.5.0 runs have no trajectory to compare with. The state and outbox of its last six fleet runs record no refusal either. So refusals did not rise.

Rollback: safe. The only operator step is in the Upgrading section: an author switched to a Claude subscription token goes back to an API key.

Gate: ruff check, ruff format --check, mypy, pytest (3098 passed, 7 skipped).

Prepared by Claude.

[github-actions]

Round 1 — reviewed head b39866d7 — reviewer summarizer:hermes/gpt-5.6-terra over coverage+credentials+deployment+general+lifecycle+prose.

terra
Advisory findings from outerloop — the code owner decides. Reply to disagree; the outerloop:no-review label opts this PR out.

Verdict: nothing blocking — 1 advisory note.

1 finding attached to the lines below.

One advisory prose suggestion in the release notes. Rejected findings: none.