Changelog

Release history. Full details on GitHub Releases.

v0.7.11

  • Custom model ids: route any new model from Settings the day it ships — no app update required
  • Per-channel model params: sampling temperature and max output tokens are now user-facing dials
  • Disabled tools are no longer advertised to the model — smaller prompts, no wasted blocked calls
  • Security: dependency audit clean — transitive qs DoS advisory resolved via scoped override
  • Docs and audit-board consolidation: living surfaces carry only live work

v0.7.10

  • Windows builds return: Setup + Portable installers ship alongside Linux for the first time since v0.7.3
  • Tool naming standardized to MCP resource_verb convention across the catalog
  • Evolution-ready configuration: 17 frozen context constants become operator-tunable settings
  • Search relevance: recency decay + usage boost in knowledge ranking
  • Analytics: recent-window error tracking with consecutive-OK streaks, percentile latencies
  • Documentation drift audit: generated single-source product facts, zero hand-synced counts

v0.7.9

  • Agent context-budget self-awareness: the agent sees its own token occupancy and trims proactively
  • Project knowledge sub-folders for organizing large projects
  • Routine outputs simplified: each routine maintains one living knowledge entry, updated in place
  • Provider errors labeled by actual provider instead of a hardcoded name
  • Log retention: daily re-prune across both log stores

v0.7.8

  • edit_file corruption fix: dollar-prefixed replacement patterns no longer splice file content
  • Typed error hierarchy harmonized across ProviderError, StorageError, TranscribeError
  • Scheduler exponential backoff with jitter + per-routine tokenBudget cap
  • MCP surface expansion: memory_list.project, session_get.checkpoint, routine tokenBudget/dry_run
  • Mobile UI: three-tier viewport (portrait/landscape/desktop) + expandable composer

v0.7.6

  • Cloudflare Account ID injection via Connections — no more wrangler hangs
  • New MCP tools: session_delete, project knowledge management
  • Release pipeline consistency: Windows download links fixed, doc sync
  • Source map archiving for production debugging
  • Infra panel and credential status improvements

v0.7.5

  • FTS drift-detection at boot — automatic index repair
  • Identity file mtime revalidation — no stale identity cache
  • Observability parity: emitOp on chat-turn errors

v0.7.4

  • Memory archive with search — archived entries stay searchable
  • memory_unarchive tool for restoring archived entries
  • Auto-consolidation with fuzzy dedup in proposals
  • Zombie mastery-subfolder migration cleanup at boot
  • Stale Atlas-era name sweep across configs and tests

v0.7.3

  • 30 scheduled routines with compounding knowledge
  • Full routine portfolio: engineering audits, growth research, competitive intel
  • Claude Haiku 4.5 model for all routines — reliable, cost-effective
  • Improved error handling and structured JSONL logging
  • GEO foundation: robots.txt, JSON-LD schema, llms.txt

v0.7.2

  • Multi-inbox email management with domain admin
  • Agent self-switch between inboxes
  • AgentMail integration improvements

v0.7.1

  • Data root migration (~/.atlas → ~/.onestone)
  • Zero-tolerance quality gate (eslint . --max-warnings 0)
  • Memory overhaul: FTS search, unified ranking
  • 17-area architecture audit — all PASS

v0.7.0

  • Clean-break rename: Atlas → OneStone AI
  • New identity, domain (onestoneai.com), and branding
  • All identifiers, infra, and data records migrated

v0.6.36

  • AgentMail domain + inbox admin tools
  • Custom domain email (send/receive from your domain)
  • Thread-aware email reading and replying

Older releases available on GitHub.