Internal error: directory mismatch for ".../claude-code-action/<ref>/tsconfig.json", fd 4 — fails on every run

[linxdigitalstudio-png]

Symptom

Every triggered run logs an "Internal error" before any tool dispatch and returns exit 1:

Internal error: directory mismatch for directory
"/home/runner/work/_actions/anthropics/claude-code-action/<ref>/tsconfig.json", fd 4.
You don't need to do anything, but this indicates a bug.

The <ref> portion of the path matches whichever ref the workflow pinned (v1, a SHA, etc.). Switching the pin does NOT fix it — same error message, only the path changes.

Reproduction in our repo

linxdigitalstudio-png/linx-edge-functions — workflow .github/workflows/claude.yml, action invocation:

- name: Run Claude Code
  uses: anthropics/claude-code-action@<ref>
  with:
    anthropic_api_key: ${{ secrets.ANTHROPIC_API_KEY }}

Tested refs (all fail identically):

• @v1 (floating tag) — 7+ failures on 2026-04-26
• @567fe954a4527e81f132d87d1bdbcc94f7737434 (SHA pin) — 2 failures on 2026-04-26 (issue Add OpenTelemetry environment variable support #48 smoke test)

Notable: one earlier run on the same repo succeeded

Run id 24969076246 at 2026-04-26T22:50:53Z succeeded against @v1. Adjacent runs (24969081289, 24969086042) were skipped. The next several runs against the same @v1 ref started failing. So the regression appears to have landed within a narrow window during 2026-04-26 ~22:50 UTC.

Affected runs (chronological)

Run ID	Ref	Result	Trigger
24969076242	@v1	failure (directory mismatch)	issue #41
24969076246	@v1	success	(sibling concurrent)
24969081289	@v1	skipped	(sibling concurrent)
24969086042	@v1	skipped	(sibling concurrent)
24969382421	@v1	failure	re-poke #39
24969383100	@v1	failure	re-poke #40
24969383700	@v1	failure	re-poke #41
24969687285	@v1	failure	issue #45
24970527005	@567fe95	failure	issue #48 (after pin attempt)

Workaround in place

We disabled the workflow trigger entirely (commented out the on: block, kept workflow_dispatch for manual runs). All AI tasks now route to a local cursor-agent dispatch script which is reliable.

What we'd like

1. Fix or rollback in the action.
2. Confirmation of whether the bug is environment-specific (composite-action depth, runner image, repo size, etc.) or universal — so we know if a workaround like adjusting actions/checkout fetch-depth would help.
3. A pinnable known-good SHA we can use until the fix lands.

Repo permissions

Happy to grant you read-only access to linxdigitalstudio-png/linx-edge-functions if needed for diagnosis.

[obroshnij]

also experiencing this issue

[vmohir]

I had the same issue and it was fixed by removing display_report: true input from github workflow

[obroshnij]

I had the same issue and it was fixed by removing display_report: true input from github workflow

that did not help in my case as we're not even using that config. The config I'm having is rather minimal

name: Claude Code

on:
  workflow_dispatch:
  workflow_call:
    inputs:
      model:
        description: Claude model to use
        type: string
        default: claude-sonnet-4-6
      claude_args:
        description: Extra CLI args passed to claude-code-action
        type: string
        required: false
        default: ""
    secrets:
      anthropic_api_key:
        required: true

concurrency:
  group: claude-${{ github.event.issue.number || github.event.pull_request.number || github.ref }}
  cancel-in-progress: true

jobs:
  claude:
    if: |
      (github.event_name == 'issue_comment' && contains(github.event.comment.body, '@claude')) ||
      (github.event_name == 'pull_request_review_comment' && contains(github.event.comment.body, '@claude')) ||
      (github.event_name == 'pull_request_review' && contains(github.event.review.body, '@claude')) ||
      (github.event_name == 'issues' && (contains(github.event.issue.body, '@claude') || contains(github.event.issue.title, '@claude'))) ||
      github.event_name == 'workflow_dispatch'
    runs-on: ubuntu-latest
    timeout-minutes: 20
    permissions:
      contents: write
      pull-requests: write
      id-token: write
      issues: write
      actions: read # Required for Claude to read CI results on PRs
    steps:
      - name: Checkout repository
        uses: actions/checkout@v6
        with:
          fetch-depth: 1

      - name: Run Claude Code
        id: claude
        uses: anthropics/claude-code-action@v1
        with:
          anthropic_api_key: ${{ secrets.anthropic_api_key }}

          claude_args: |
            --model ${{ inputs.model || 'claude-sonnet-4-6' }}
            ${{ inputs.claude_args }}

[minghsuy]

Independent confirmation of #1266, fresh as of 2026-05-08.

Hitting Internal error: directory mismatch for directory ".../claude-code-action/<ref>/tsconfig.json", fd 4 consistently on claude-code-action@v1 invoked from pull_request events. The action exits 1 before producing any review.

Possibly useful diagnostic vector: on the same repos, the same action+ref+runtime invoked from issue_comment events (the @claude mention handler pattern from the action's docs) is reliably succeeding at the same time. Same Bun runtime, same actions/checkout@v4 fetch-depth: 1, same OAuth token. Only the trigger event differs. Whatever's specific to the pull_request invocation shape that's tickling the directory check, the issue_comment invocation shape isn't.

Workaround in place matches what's reported above: when the auto-review path fails silently, manually commenting @claude review reliably produces a review via the mention path.

[digeshnanda]

+1 — same bug, self-hosted Ubuntu runner.

Adds another data point for the impact assessment:

Environment:

• Self-hosted runner on Ubuntu 7.0.0-14-generic (X64), runner v2.327.1
• Bun 1.3.13 (installed via official https://bun.sh/install script)
• Action ref pinned to v1
• runs-on: self-hosted (not GitHub-hosted)
• Workflow files: canonical claude.yml + claude-code-review.yml from /install-github-app

Symptom:

"is_error": false                         ← LLM call succeeds
Internal error: directory mismatch for directory
"/home/github-runner/actions-runner/_work/_actions/anthropics/claude-code-action/v1/tsconfig.json", fd 4.
You don't need to do anything, but this indicates a bug.

Net behavior:

• Workflow run is marked success (exit 0)
• But no review comment is posted to the PR
• Reproducible across multiple PRs; not transient
• Re-running the workflow on the same PR yields identical result
• LLM API call to Anthropic clearly succeeded (is_error: false + "model": "claude-sonnet-4-6" in logs)
• The crash happens at the post-LLM "post review back to GitHub" step

Things tried that don't help:

• Switching runs-on from ubuntu-latest → self-hosted
• Reinstalling bun + unzip + zstd on runner
• Restarting runner systemd service
• Running on a freshly-bootstrapped runner

Workaround: None known. Bot is effectively half-functional — @claude mention path completes but auto-review on PR open silently no-ops at the comment-post step.

Happy to provide additional logs / runner config / workflow YAML if useful for diagnosis.