Kōan v0.8.0

Added

• Chat stays responsive during missions — your chat messages now get a priority lane, so the bot answers you even while it is busy running a mission (#1084).
• Split /rebase — a bare /rebase <pr> now only rebases the PR onto its base branch. Use /rebase --fix (implied when you add a focus area or severity after the URL) to also apply review feedback (#2439).
• More CLI providers — added the Grok Build and Haze providers, so you can run missions on additional backends. /models slots can now be annotated with a per-role CLI, and MCP tools can be enabled per execution role (#2422).
• Plain-English GitHub mentions — mentioning the bot in natural language now routes to the right skill instead of requiring an exact command (#2402).
• Authoritative mission store — missions now live in a SQLite store (missions.md becomes a read-only export). New /list <state> shows done/failed history, and make missions is a break-glass CLI when the bridge is down.
• Live mission timeline — the dashboard /progress view now shows a structured, live timeline; you can edit a project's GitHub URL and auto-merge settings from the dashboard and via PATCH /v1/projects/<name>.
• GitHub "Running" indicator — a label and commit status now mark a PR while a mission is working on it (#2365).
• /claudemd <project> keeps a managed learnings block in CLAUDE.md; added the /rereview alias and a configurable label to pause automatic PR reviews.
• Review upgrades — severity-graded verdict alerts, stale-HEAD warnings on posted reviews, and per-mission token usage/cost reported over the REST API.
• .koan/ steering tree — koan-only KOAN.md and per-skill instructions that apply to the autonomous agent without affecting interactive sessions (#2293).
• Cloud deploy — hydrate instance/ from a repo on cold boot, with an opt-in periodic pull.
• Non-blocking CI checks — CI-fix missions no longer stall the queue behind other work.
• Prettier make logs output (set raw=1 to bypass), and per-mission-type reasoning-effort configuration.

Changed

• Large review diffs are no longer silently truncated — the old 32k-character cap is now a configurable budget, and a review that couldn't cover the whole diff says so with a partial-coverage warning (#2292).
• The bot stays up when the CLI binary is missing at startup, entering a degraded no-mission mode instead of failing outright.
• Unified release pipeline — releases are now cut from the incubating branch through the release workflow, and CHANGES.md is the published changelog.

Fixed

• No more duplicate command output when Telegram re-delivers an update, and no more double-resume when pausing/resuming in quick succession (#2437).
• Missions no longer wrongly resurrect after removal, and a mission is re-queued when verification fails (#2210).
• Grok: credit/spending-limit exhaustion is now correctly treated as quota, false burn-rate alerts and pauses are gone, and headless tool calls are approved so /implement can commit.
• The Telegram bridge no longer grows its memory footprint over long uptime (#2360), and long /report messages now deliver reliably.
• Slack: the bot no longer replies in threads addressed to someone else.
• Security: closed a shell-injection path in CI dispatch-tag validation.