xxiaoxiong/dsh-ci

dsh-ci - Continuous Integration (CI) plugin for DeepSeek Harness (DSH)

dsh-ci is a generic, provider-neutral CI/CD plugin for DeepSeek Harness (DSH). It enables agents to inspect CI runs, extract bounded failure evidence, locate source errors, and safely rerun or cancel pipelines.

The current version supports GitHub Actions as the primary CI provider. Its architecture is explicitly designed to be provider-neutral, allowing future implementations for GitLab CI, Jenkins, CircleCI, etc. without modifying the generic CI tool layer.

Architecture

DSH Agent
   ↓
Generic CI Tools (ci_status, ci_failure, etc.)
   ↓
ctx.ci (CIService Context interface)
   ↓
GitHubActionsProvider (implements CIService)
   ↓
GitHub Actions REST API

Key Features

• Generic CI Tools: A stable set of 9 tools exposed to the agent loop.
• Ecosystem-neutral Service Layer: Abstract CIService and AbstractCIProvider core contracts.
• Automatic Workspace Detection: Parses local git repository (origins, remotes, branches, HEAD) to automatically map owner and repository without user input.
• Failure evidence extraction: Bounded logs (default 8KB), error-focused extraction (only lines around errors), and de-duplication.
• Failure Parser Registry: Specialized regex parser patterns to map compile errors, test suite failures, docker fails, and generic exit codes to exact source files, lines, and functions. Supports:
  • TypeScript compilation errors (tsc)
  • Jest / Vitest test runner failures
  • pytest (Python) failures
  • Maven / Gradle (Java) build failures
  • Go test failures
  • Cargo (Rust) compile errors
  • ESLint errors
  • Docker build errors
  • npm install/dependency errors
  • Generic command exit codes
• Secret Redaction: Automatic scrubbing of sensitive tokens, private keys, passwords, and AWS keys from all log output returned to the LLM.
• Write Approval Gates: Write operations (ci_rerun, ci_cancel) are approval-gated and disabled by default (writeEnabled: false). Cancel operation always requires approval.

Installation

Add this plugin directly via DeepSeek Harness CLI:

dsh plugin add dsh-ci

Configuration

Add the following block to your settings.json or settings.local.json configuration file:

{
  "plugins": {
    "dsh-ci": {
      "writeEnabled": false,
      "maxLogSize": 8192,
      "maxEvidenceItems": 50,
      "apiBase": "https://api.github.com"
    }
  }
}

Credentials

The plugin integrates with the standard DSH credential manager. It will automatically check for credentials in the following order:

1. GH_TOKEN environment variable
2. GITHUB_TOKEN environment variable
3. Local gh auth credentials (via GitHub CLI helper)

Tool Reference

The plugin registers 9 provider-neutral tools inside DSH:

ci_status

Gets the overall CI status of the repository, current branch, or commit.

• Parameters: repository?, branch?, commit?

ci_runs

Lists recent workflow runs in the repository.

• Parameters: repository?, branch?, status?, limit?

ci_run

Retrieves a detailed description of a single run.

• Parameters: runId (required), repository?

ci_jobs

Lists all jobs and their execution steps inside a workflow run.

• Parameters: runId (required), repository?

ci_failure

Looks up a failed run, identifies the failing jobs/steps, parses their logs, and compiles a structured failure evidence block containing exact file paths, line numbers, test names, and a bounded log excerpt.

• Parameters: runId (required), jobId?, repository?, maxExcerpts?

ci_log

Fetches log excerpts for a given job or step. Excerpts are bounded in size and redact all secrets automatically.

• Parameters: jobId (required), stepId?, repository?, maxBytes?

ci_artifacts

Lists metadata about build artifacts.

• Parameters: runId (required), repository?

ci_rerun

Triggers a rerun of a workflow run. Requires write authorization.

• Parameters: runId (required), dryRun?, repository?

ci_cancel

Cancels an in-progress workflow run. Always requires approval.

• Parameters: runId (required), repository?

Security Model

1. Untrusted Input: All log outputs are treated as untrusted data. We filter and redact strings matching high-risk signatures (e.g., token prefixes, private keys) to prevent prompt injection and credential leaking.
2. Context Preservation: Full logs are never returned to the model. We slice logs to focus on errors, avoiding context window explosion.
3. Approval Flow: Triggering runs or cancels will prompt the user with DSH Approval seam for explicit consent before issuing the REST requests.

License

MIT License.