claude-code-action

This usually means the Claude process failed during startup, not that the code review completed and failed. Inspect claude-execution-output.json and read the result field; in the reported case it contained Credit balance is too low. Fix it by checking the Anthropic API account's billing/credit balance, adding credits, and re-running the workflow. A failed startup typically shows is_error: true with no review aggregation or findings, such as spawned: 0. This is distinct from a completed review that reports findings.

The mcp__github_file_ops__delete_files tool schema accepts paths, but the tag-mode prompt example incorrectly uses files. This causes Zod validation to reject the first deletion attempt when use_commit_signing: true is enabled.

Fix the prompt example in src/create-prompt/index.ts (around line 827):

- mcp__github_file_ops__delete_files: {"files": ["path/to/old.js"], "message": "chore: remove deprecated file"}
+ mcp__github_file_ops__delete_files: {"paths": ["path/to/old.js"], "message": "chore: remove deprecated file"}

Then rebuild/release the action or use a release containing that fix. In an existing run, retrying the tool call with paths works around the failed call:

{"paths": ["path/to/old.js"], "message": "chore: remove deprecated file"}

Note that commit_files still uses files; only delete_files uses paths. A regression test should import the MCP tool schemas and verify that parameter names in prompt examples match the corresponding schemas.

For pull_request events, the action auto-detects tag mode, which only merges user-provided MCP tools starting with mcp__github_ into the default allowlist. Custom rules such as Bash(npx markdownlint:*) or Bash(gh pr:*) are parsed but filtered out.

To use custom Bash tools, explicitly set mode: agent in the workflow:

- uses: anthropics/claude-code-action@v1
  with:
    mode: agent
    claude_code_oauth_token: ${{ secrets.CLAUDE_CODE_OAUTH_TOKEN }}
    prompt: |
      Run `npx markdownlint *.md` to check markdown files.
    claude_args: |
      --allowedTools "Bash(npx markdownlint:*),Read,Glob"

Use the camel-case flag --allowedTools; --allowed-tools may not be recognized. This behavior was observed in claude-code-action v1.0.22.

URL-encode each branch-name path segment before inserting it into a GitHub tree or compare URL, while preserving / as the path separator:

const encodedBranch = branchName
  .split("/")
  .map(encodeURIComponent)
  .join("/");

For example, claude/fix#123 becomes claude/fix%23123, so the generated link is:

https://github.com/owner/repo/tree/claude/fix%23123

Append the [1m] context selector to the model IDs in your workflow configuration:

env:
  ANTHROPIC_MODEL: claude-opus-5[1m]
  ANTHROPIC_SMALL_FAST_MODEL: claude-sonnet-5[1m]

This is required when an ANTHROPIC_BASE_URL gateway prevents Claude Code from automatically detecting native 1M context support. The workaround was verified on action v1.0.185 with Claude Code 2.1.222: claude-opus-5[1m] resolves modelUsage.contextWindow to 1000000, while claude-opus-5 resolves to 200000 and causes premature auto-compaction.

To verify the resolved limit, enable show_full_output: true and inspect the modelUsage object in the job log:

"contextWindow": 1000000

Note: maxOutputTokens may still report 64000 even with the [1m] suffix; that value is resolved separately from the context-window selector.