linear-code-agent 0.5

The coder-agent is working. It clones repos, it edits code, it opens PRs, it marks the issue done, most of the time. Most of the time is where this project starts.

linear-code-agent 0.5 is the sharpening pass. Not a rewrite, not a rebrand — just the set of upgrades that make the agent less likely to stumble on the cases that have already tripped it up:

• Playwright MCP sidecar so the agent can drive a browser, verify UI, take screenshots, run ad-hoc E2E flows — not just run the tests that already exist.
• Serena MCP to replace the Grep + Read + guess pattern with AST-aware symbol navigation: find_symbol, find_referencing_symbols, rename_symbol. Fewer wasted iterations on refactors.
• Sequential Thinking MCP for structured reasoning on hard tasks — a tool the model can use to think in explicit steps before touching files. Complements the TodoWrite planning pass.
• Bugfixes from real-world runs: ghost sessions that receive a session_created but never emit; the resolveRepo ambiguity when an issue has multiple GitHub attachments; the spurious “Authentication required” error that marks successful PR creations as failures.

The scope is deliberately a point-release, not a 1.0. Ship the fixes, learn from the next batch of failures, then plan the bigger things.