SWEny, YAML workflows for AI agents I'm running in prod (triage, E2E)

Major Changes

• 2f1a424: Rename triage-specific issue tracking interfaces to generic equivalents:

  • TriageHistoryCapable → LabelHistoryCapable
  • TriageHistoryEntry → IssueHistoryEntry
  • listTriageHistory() → searchIssuesByLabel() (signature: (projectId, labelId, opts?: { days? }))
  • canListTriageHistory() → canSearchIssuesByLabel()
  • Remove FingerprintCapable interface and canSearchByFingerprint() (dead code — never called from engine)
  • Make Issue.branchName optional (Linear returns real names; Jira/GitHub Issues synthesize them)

• ebbb5a7: Remove wrong-pattern MCP adapter providers (breaking change).

  linearMCP, githubMCP, and slackMCP have been removed. These adapters called MCP servers from recipe steps — the wrong architectural layer. MCP servers are agent tools accessed during reasoning, not recipe-step backends.

  Migration: Configure these MCP servers via mcpServers in CodingAgentRunOptions (now supported in all three coding agents). The agent gets access to Linear, GitHub, and Slack MCP tools during its reasoning session with zero custom provider code.

  Also removed: slack-mcp notification provider option from CLI and GitHub Action (previously required SLACK_BOT_TOKEN, SLACK_TEAM_ID, SLACK_CHANNEL). Use the slack webhook notification provider or configure the Slack MCP server for the agent directly.

Minor Changes

• 9313ff9: Add mcpServers injection to coding agent providers.

  CodingAgentRunOptions now accepts mcpServers?: Record<string, MCPServerConfig>. When provided, all three coding agents (Claude Code, OpenAI Codex, Gemini) serialize the config to a temp JSON file and pass --mcp-config <path> to the agent CLI. The agent receives all configured MCP tools during its reasoning session. Temp file is cleaned up after the agent exits.

  MCPServerConfig extended to support both transports:

  • type: "stdio" — local pre-installed binary (command, args, env)
  • type: "http" — remote Streamable HTTP server (url, headers)

  Type defaults to "http" when url is set, "stdio" when command is set.

• 556a53d: Add browser-safe PROVIDER_CATALOG — a structured list of all available provider implementations with display names, categories, env var specs (key, description, required, secret, example), and import paths. Exported as PROVIDER_CATALOG, getProvidersForCategory(), and getProviderById() from the main package entry.

• f33c74d: Add GitProvider (local git operations) and RepoProvider (remote API operations) interfaces.
  SourceControlProvider is now a type alias for GitProvider & RepoProvider — fully backward compatible.
  Enables partial implementations for contexts without a local checkout (cloud workers, MCP servers).

• 1df08e0: Task 03/04/05: wire slackMCP, add file providers to Action, shared factories.

  providers (minor — new exports):

  • createObservabilityProvider(name, credentials, logger) — shared factory for all 8 observability providers
  • createCodingAgentProvider(name, logger, opts) — shared factory for all 3 coding agents

  cli (patch):

  • notification-provider: slack-mcp now supported via slackMCP()
  • CLI and Action provider switches for observability and coding agent replaced with shared factory calls

  action (patch):

  • issue-tracker-provider: file, source-control-provider: file, notification-provider: file now supported
  • slack-mcp notification provider added
  • New output-dir input (default .github/sweny-output) for file-based providers
  • slack-bot-token, slack-team-id, slack-channel inputs added to action.yml

Patch Changes

• 68780d5: Fix ERR_MODULE_NOT_FOUND crash on import when optional peer deps are not installed. S3 storage classes (S3SessionStore, S3MemoryStore, S3WorkspaceStore) and MCPClient now lazy-load their peer dependencies (@aws-sdk/client-s3, @aws-sdk/s3-request-presigner, @modelcontextprotocol/sdk) on first use rather than at module load time. Also fixes a retry bug in MCPClient where a failed connect() would leave a stale client reference that blocked subsequent reconnect attempts.

• 207a317: Add index signature to ToolResult to satisfy updated @anthropic-ai/claude-agent-sdk tool callback type constraints.

• 010b6d7: Code review fixes for MCP adapters and triage dedup step.

  providers:

  • MCPClient: clear connectPromise on error to prevent stuck reconnections; use local variable before callTool to guard against concurrent disconnect
  • linearMCP: throw on empty id/identifier in toIssue(); expose limit option in searchIssuesByLabel() (default 100)
  • githubMCP: validate repo config as owner/repo format; guard against PR number=0; validate targetRepo in dispatchWorkflow
  • slackMCP: clarify channel config must be a Slack channel ID (e.g. C123456), not a name
  • LabelHistoryCapable: add limit?: number to searchIssuesByLabel opts

  engine:

  • dedup-check step: rename outcome from "notify" → "duplicate" so routing key is distinct from target node name
  • triage definition: update on: { duplicate: "notify" } to match