list of issue since 2.8.0

[Mitsu13Ion]

Checklist

• I searched existing issues and this hasn't been reported

Area

Fullstack

Operating System

Linux

Version

2.8.0

What happened?

• Planning can fail with EXDEV: cross-device link not permitted when implementation_plan.json is normalized inside a worktree,
  because the normalizer writes to /tmp and then tries to rename() across filesystems.
• Desktop agents can fail with Cannot find module .../auto-claude-mcp-server.js because core agent configs still try to launch a
  non-existent external auto-claude MCP server instead of using the built-in Auto-Claude tools.
• Valid MCP servers can be reported as missing because project-level MCP settings, custom servers, and MCP environment variables
  are not fully propagated from AgentManager to the worker-side MCP registry.
• Tasks with an existing but invalid or empty implementation_plan.json do not reliably trigger replanning and instead fail during
  execution with plan validation errors.
• Planning failures can be emitted as CODING_FAILED instead of PLANNING_FAILED, which leaves the task state machine in the wrong
  state and can cause the task to fall back to backlog incorrectly.
• QA retry loops can run far beyond the documented limit: prompts describe a 5-iteration cap, but the orchestration layer allows
  many more iterations, creating a high risk of excessive token consumption.
• QA and build sessions use overly large maxSteps budgets, so non-converging agent runs can burn a very large token budget before
  they stop.
• Hitting max_steps or context_window is treated as successful completion in multiple layers of the pipeline, allowing downstream
  stages to continue after the agent has effectively failed to finish the job.
• Build orchestration can continue to QA or fix-up stages even when some subtasks are stuck, as long as at least one subtask
  completed, which amplifies wasted runs and token usage.
• Fallback process-exit handling can promote ambiguous agent exits to success-like states, masking real failures and making token
  burn incidents harder to diagnose.
• Rapid project switching can cause stale async store writes across desktop stores, resulting in cross-project UI contamination
  and inconsistent state.
• reconcileLinkedFeatures can persist roadmap data into the wrong project if the active project changes while async task loading
  is in flight, leading to roadmap corruption.
• Review rejection cleanup can delete unrelated user files because it relies on a blind git clean style cleanup instead of
  deleting only files introduced by the merge attempt.
• Worktree-specific HOME rewriting can overwrite valid user environment configuration, breaking proxy settings, package manager
  configuration, or enterprise shell setup.
• Agent startup can hit avoidable 401 authentication failures because OAuth validity is not refreshed before spawning the agent
  process.
• Profiles with pre-existing auth terminal credentials can remain stuck in onboarding because onboarding completion logic assumes
  a prior login flag.
• Desktop production builds can fail at runtime because @libsql native binaries are not bundled correctly.

Steps to reproduce

Just use the beta

Expected behavior

Working ?

Logs / Screenshots

[ThalesMMS]

Same issue here.