v0.8.0

[0.8.0] - 2026-07-14

Added

• Added signed-in WorkBuddy account models to /model, a3s model, and
  a3s code models. The integration reuses WorkBuddy's bundled CodeBuddy CLI
  and local account state without reading or copying private tokens, discovers
  the models enabled for the account, streams responses, and preserves native
  A3S tool execution.

Changed

• Consolidated Claude Code, Codex, and WorkBuddy integrations under the
  account_providers boundary. Claude and WorkBuddy share one cancellable
  account-CLI stream/tool bridge, while all three share account detection,
  client construction, model switching, persistence, and session restore.
• Unified TUI /compact with the direct, tool-free compactor used by Code Web.
  Repeated manual compaction now includes the previous summary without creating
  a temporary tool-capable agent session, while Core's rolling auto-compaction
  remains re-armed for long-running conversations.
• Rebuilt /goal as a durable Ultracode goal loop. Setting a goal now creates
  a complete .a3s/loops/goal-* Loop Engineering workspace, forces planning
  and goal tracking, runs separate maker/verifier guidance, and continues
  across normal ends and retryable errors until a matching GoalAchieved
  arrives. Esc and /goal clear invalidate delayed retries immediately, and
  normal Ultracode message-gated planning is restored when the goal closes.
• Aligned fenced Markdown code with Codex-style terminal highlighting: known
  languages retain distinct token colors, unknown languages stay plain, CRLF
  is normalized, and 512 KiB / 10,000-line guardrails avoid render stalls.
• Replaced width-unstable colorful emoji in the transcript, task queue,
  thinking indicator, and /ide tree with monochrome terminal-safe marks and
  consistent hair-space padding.
• Made the footer the single owner of live context usage; the composer status
  chip keeps effort and mode information without duplicating context fill in
  the input border.
• Enabled Core's model-aware rolling compaction for TUI and Code Web sessions.
  Each selected model supplies its actual context window, requests compact
  before overflow, and can compact repeatedly throughout a long-running task.
  A3S Code Core 5.2.4 budgets the retained suffix by estimated message tokens,
  bounds oversized summaries, and refuses replacements that would not reduce
  context. Core summaries are written back to each host's durable timeline so
  later turns continue from the latest generation instead of compacting it
  again.
• Moved model, effort, Ultracode, goal, auth, reload, fork, and clear session
  changes onto an async atomic replacement path. The UI no longer blocks the
  Tokio runtime, failed reconfiguration keeps the old session usable, and
  /goal can reliably enter forced-planning Ultracode before its first turn.
• Routed A3S Code 5.2.2 native structured output through TUI launch, configured
  model selection, effort rebuilds, and headless DeepResearch. Codex Responses
  and Responses Lite now force the schema function through tool_choice, while
  providers with verified JSON Schema support retain that path and unknown
  custom OpenAI-compatible endpoints keep the safe prompt fallback.
• Clarified that web-and-workspace scope describes available evidence tools,
  not mandatory tracks. The semantic planner now reserves workspace collection
  for queries that explicitly depend on a repository or local artifacts.
• Replaced the normal public-investigation direct_then_maker route with
  direct_then_review. Multi-query retrieval now flows into one structured
  synthesis-and-coverage review, removing a redundant slow model turn while
  preserving direct_then_maker replay compatibility for older journals.
• Expanded the LLM-selectable public evidence envelope to four searches and
  eight parallel fetches. Query-specific candidates are fetched before seed
  URLs, page excerpts are ranked against their owning evidence question, and a
  zero-result unconfigured search receives one bounded Brave fallback.
• Gave the independent checker a 180-second clock within the unchanged
  300-second workflow fuse and carry observed checker latency into subsequent
  scheduling decisions. Public-source gaps route to direct retrieval; makers
  remain reserved for evidence production or required local/non-web work.

Fixed

• Corrected the real-LLM compaction integration test to compare matched
  compressed and uncompressed histories. It now proves provider-reported
  prompt reduction on the compacted request and again after session restore,
  instead of comparing two already-compacted turns.
• Preserved traceable structured evidence when the independent DeepResearch
  checker times out. The workflow now completes with an explicit degraded
  verification state and publishes a provisional evidence-derived report;
  only runs without reportable evidence fall back to a Recovery artifact.
• Kept explicit checker URLs out of finding prose and bound each one to its
  matching source card, preventing Chinese terminal punctuation from becoming
  part of an auto-linked URL and avoiding unrelated citations on a finding.
• Made long report headings prefer a complete semantic clause, bounded the
  caveat section to eight reader-relevant items, and added a mobile horizontal
  scroll cue to wide evidence matrices.
• Prevented structured no-tool makers from being used for checker-requested
  evidence collection. Existing-evidence synthesis is now an explicit legacy
  optimization, while new gaps retain tool-capable collection semantics.
• Tightened checker output so findings state supported facts instead of merely
  announcing that sources or comparisons exist, and requested recommendations
  must give a conditional answer or remain an explicit evidence gap.