[matifali]

Description

Adds an additional_args input to agent-relay-claude-code for passing extra flags to claude self-hosted-runner. coder/registry#1199 makes the same change to agent-relay-cursor.

additional_args is a map of flag to value. Each key goes to the CLI exactly as written, so the admin decides between a long flag (--flag) and a short one (-f). The value comes next as a separate argument, and an empty value passes the flag alone. The flags go after the ones the module sets and before --exec-path.

module "claude_code_runner" {
  source   = "registry.coder.com/coder/agent-relay-claude-code/coder"
  version  = "0.3.0"
  agent_id = coder_agent.main.id
  # Passed as: --another-flag --some-flag value
  additional_args = {
    "--another-flag" = ""
    "--some-flag"    = "value"
  }
}

• Verbatim values: the map is base64-encoded into the start script, one argument per line. The script shell-quotes each argument with printf %q before writing it into the detached supervisor, so values reach the CLI unchanged: spaces, quotes, $VAR and backticks are not split or expanded.
• Validation: keys must look like --name or -n, and values can't contain newlines.
• Flags the module sets are rejected: --capacity, --base-dir, --exit-if-unused-min, --exec-path. Use the base_dir and exit_if_unused_min inputs instead. A comment in main.tf and in start.sh.tftpl says to keep this list in sync with the runner command.

The name additional_args is the most common one in the registry; code-server and ttyd already use it.

Type of Change

• New module
• New template
• Bug fix
• Feature/enhancement
• Documentation
• Other

Module Information

Path: registry/coder/modules/agent-relay-claude-code
New version: v0.3.0
Breaking change: [ ] Yes [x] No — the new input defaults to {}, which leaves the runner command unchanged.

Testing & Validation

• Tests pass: terraform test, 12 passed (new: empty default, rendering and base64 encoding, rejecting a malformed key, rejecting a flag the module sets). I added a container test in main.test.ts, but it could not run here because the runner has no Docker daemon.
• Code formatted (bun fmt)
• Changes tested locally: rendered the start script with terraform console and ran it against a stub claude. Its argv received --verbose, --name followed by two words $HOME \id` 'q' "d"unchanged, and-tfollowed byx y. The default {}` produces the same command as before.

Related Issues

None

🤖 Generated with Claude Code

https://claude.ai/code/session_01R2TMgCPW8SL8kUCB28L4dt

[github-actions]

Module Scorecard Check

coder/agent-relay-claude-code: 52 → 59

✅ Score improvement: 52 → 59 (+7).

Theme	Before	After
Presentation & Onboarding	6 / 17	12 / 17
Integration	0 / 25	0 / 25
Credential Hygiene	20 / 20	20 / 20
Restricted-Environment	12 / 20	12 / 20
Engineering Quality	10 / 10	10 / 10
Overall	52 / 100	59 / 100

Full scorecard for this PR

Presentation & Onboarding	Agent Integration	Credential Hygiene	Restricted-Environment Readiness	Engineering Quality	Overall
12 / 17	0 / 25	20 / 20	12 / 20	10 / 10	59 / 100

Drilldown

Presentation & Onboarding — 12 / 17

Criterion	Max	Score	Notes
Configuration-mode examples	12	12	README documents the two major modes: install_cli = true (default, downloads CLI) shown in the primary example, and install_cli = false (baked image) documented in the Requirements section. additional_args has a dedicated example block. cli_binary override is documented. All variables carry sensible defaults.
Visual preview	5	0	No image, GIF, or video embedded in the README. The frontmatter icon field points to a Claude SVG, which is an icon, not a visual preview of the module in action.

Agent Integration — 0 / 25

Criterion	Max	Score	Notes
AI governance	10	0	No mention of Coder AI Gateway or Agent Firewall anywhere in the README or source. The credential is a work-order JWT stamped by Agent Relay, but there is no documented integration with AI Gateway routing or Agent Firewall policy enforcement.
Dashboard entry point	5	0	No coder_app resource is declared and the README does not document a dashboard entry point. The module is a headless runner dispatched by Agent Relay.
Session continuity	5	0	No documentation of resuming an existing agent session across reconnects or relaunches. No mention of session-ID-based resume, boo, tmux, screen, or any persistent session manager. The agent_relay_session_id parameter is for Agent Relay deduplication, not user-facing session continuity.
Managed configuration	5	0	No documented support for managed MCP servers, settings files, policies, or workdir configuration. base_dir sets the checkout directory and additional_args passes CLI flags, but neither constitutes managed configuration in the rubric sense.

Credential Hygiene — 20 / 20

Criterion	Max	Score	Notes
Secrets marked sensitive	16	16	agent_relay_credential is styled with mask_input = true in its coder_parameter block. The README states "the credential is masked." No inline secrets or placeholder keys appear in any README example.
Non-hardcoded auth path	4	4	The README documents that Agent Relay stamps a single-use work-order JWT on each build; the user never pastes a raw key. The credential flows through coder_env → SELF_HOSTED_RUNNER_ENVIRONMENT_SECRET, entirely managed by the relay.

Restricted-Environment Readiness — 12 / 20

Criterion	Max	Score	Notes
Mirrorable artifact source	5	0	The install URL https://claude.ai/install.sh is hardcoded in install.sh.tftpl. No module input variable overrides this URL. cli_binary overrides the binary path, not the download source; install_cli is a boolean toggle. No variable can point the installer at an internal mirror.
Bring-your-own binary	10	10	install_cli (default true) can be set to false to skip the download entirely. The README explicitly documents: "Bake the CLI into the image and set this to false for faster workspaces." The install script checks command -v and short-circuits when the binary is already present.
Egress transparency	3	0	No dedicated README section enumerates external endpoints contacted at install or runtime. The install script contacts claude.ai / downloads.claude.ai and the runner contacts Anthropic's API, but these are only visible in source code, not in a documented network/air-gapped section.
Runs without sudo	2	2	No script invokes sudo. The install step writes to ~/.local/bin; the start step creates directories under $HOME. All operations work as an unprivileged user.

Engineering Quality — 10 / 10

Criterion	Max	Score	Notes
Input quality	6	6	All nine variables carry clear descriptions and sensible defaults. cli_binary has a regex validation restricting it to a command name/path. exit_if_unused_min validates whole-number ≥ 0. additional_args has two validations: flag-format regex and reserved-flag rejection.
Test coverage	4	4	main.tftest.hcl covers the parameter contract (names, ephemeral flags, styling/masking), runner wiring, install-CLI behavior, additional-args encoding, and path overrides. main.test.ts runs the actual install and start scripts in a container with stub binaries, exercising the full supervisor lifecycle (idle, working, serving, done, failed, orphaned), the permissions wrapper, IS_SANDBOX enforcement, and shell-injection safety of serving_log_pattern.

Overall — 59 / 100

Raw 54 / 92 → round(54 / 92 × 100) = 59

Tip

You can run this locally by telling your agent: "review this module against .github/scorecard/SCORECARD.md".

---

Scored against SCORECARD.md with solstice-1. Language-model scores are advisory.

[dluman]

Will enable semantic, reasonable approach to accessing additional environment settings to customize module behavior for org. workflows. ✅

Looks reasonable to me providing (map)string