Summary

Archon's Gitea and GitLab webhook adapters are implemented, but all default command files (.md) and workflow files (.yaml) hardcode gh CLI calls. Running workflows against a Gitea or GitLab repository fails because gh pr create, gh issue view, etc. only work with GitHub.

Proposed Solution

Three layers of change:

1. Forge Detection (@archon/git)

• New detectForge(repoPath) utility that parses git remote get-url origin to determine forge type (github | gitea | gitlab | unknown)
• Matches against GITEA_URL and GITLAB_URL env vars for self-hosted instances
• Returns API base URL (e.g., https://git.example.com/api/v1)

2. Variable Injection + Env Passthrough

• New workflow variables: $FORGE_TYPE, $FORGE_API_BASE, $FORGE_CLI
• Detected at workflow start, substituted in all prompts and scripts
• Added to SUBPROCESS_ENV_ALLOWLIST so AI subprocesses can read them
• GITLAB_TOKEN and GITLAB_URL added to the allowlist alongside existing GITEA_TOKEN/GITEA_URL

3. forge-cli.ts Helper Script

• TypeScript script (runs via bun, zero additional dependencies)
• Unified interface: bun forge-cli.ts pr create, bun forge-cli.ts issue view, etc.
• Routes to gh CLI for GitHub, native fetch() API calls for Gitea/GitLab
• Bash nodes in workflows call it deterministically (no AI reasoning needed)
• For AI command prompts, a compatibility preamble is auto-injected for non-GitHub forges

Files Changed

• New: packages/git/src/forge.ts (detection), .archon/scripts/forge-cli.ts (CLI wrapper)
• Modified: executor.ts, executor-shared.ts, dag-executor.ts (variable injection + plumbing)
• Modified: env-allowlist.ts (GITLAB_TOKEN, GITLAB_URL, FORGE_TYPE, FORGE_API_BASE)
• Modified: 3 workflow YAMLs (bash nodes: gh → bun "$FORGE_CLI")
• Modified: 3 command MDs (explicit forge notes)

Testing

• Unit tests for forge detection (11 tests, all pass)
• Full bun run validate passes (type-check + lint + format + all tests)
• End-to-end tested against production Gitea instance (git.axiodev.com):
  • ✅ repo info — returns nameWithOwner, defaultBranch
  • ✅ issue create — created test issue Model stucked at response stream text #1
  • ✅ issue comment — posted automated comment
  • ✅ issue view — fetched issue details
  • ✅ pr create — created PR updated code to use locally hosted llama LLM, nomic-embed-text model. #2 from test branch
  • ✅ pr comment — posted PR comment
  • ✅ pr diff — retrieved unified diff
  • ✅ pr list — listed open PRs
  • ✅ All test artifacts cleaned up (issue closed, PR closed, branch deleted)

Backwards Compatibility

• GitHub repos: zero behavior change (forge detection defaults to github)
• Auto-injected preamble only activates for non-GitHub forges
• All existing tests pass without modification (except index offsets for new params)