list of issue since 2.8.0 #1953
Description
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 ?