[35C4n0r]

Closes REG-84

Refactor coder-labs/cursor-cli to install and configure Cursor CLI only, matching the codex v5 and copilot (#1117) migrations.

• Replace the agentapi module with coder-utils; drop Tasks inputs (order, group, install_agentapi, agentapi_version, ai_prompt), the start script, and the task-reporting coder MCP server.
• Install via the official installer (https://cursor.com/install, temp file + bash -n validation, retrying curl). Skipped when cursor-agent is already on PATH; fails clearly when install_cursor_cli = false and no binary exists.
• Rename required folder to optional workdir; create it and write Cursor's workspace trust marker so the trust prompt is skipped.
• Merge mcp into user-level ~/.cursor/mcp.json (existing on-disk servers win on duplicate names; invalid JSON is backed up to mcp.json.bak).
• Keep rules_files (now requires workdir) and api_key (CURSOR_API_KEY, never rendered into the script).
• Add cursor_config_dir (default ""): sets CURSOR_CONFIG_DIR via coder_env only when non-empty, so an existing value is never overwritten. It moves cli-config.json only; Cursor CLI always reads the user-level mcp.json from ~/.cursor, so the module keeps writing mcp there (verified against the binary).
• Remove model and force: Cursor CLI has no model env var, so both become --model / --force flags in the user's coder_app.
• Add scripts output; rewrite README (v1 warning, launcher example) and tests. Bump to 1.0.0.

Important

The trust marker (${CURSOR_DATA_DIR:-~/.cursor}/projects/<slug>/.workspace-trusted) is derived from the CLI bundle, not public docs. The path matches what the CLI creates, but suppression of the prompt could not be verified end to end without a real Cursor API key (auth is checked before trust). Worth a manual check.

Decision log

Verified against the installed CLI (2026.09.26-dd393fe) and Cursor docs:

Skill pattern	Cursor CLI reality	Decision
Pass VERSION to official installer	Installer bakes in the latest build; no version input	No version variable
Model via documented env var	No model env var; --model flag / /model only	Drop model, document flag
managed_settings root-owned policy	No general policy file; ~/.cursor/policy.json is sign-in enforcement only	Not added
Native mcp add duplicate semantics	No mcp add; config is .cursor/mcp.json or ~/.cursor/mcp.json	jq merge into global file, existing wins
Workdir trust key in config	Trust is a per-project marker file written by --trust or the prompt	Write the same marker if absent
AI Gateway port	Not present pre-migration	Out of scope
Binary name	Installer links agent (primary) and cursor-agent (legacy)	Probe/link cursor-agent (unambiguous)

Testing

• terraform validate, terraform test (8 passed)
• bun test main.test.ts (14 passed, container-based)
• Real installer run in codercom/enterprise-node via the rendered install script: installed, validated, workdir trusted, PATH updated
• shellcheck on rendered install script, prettier, readmevalidation

---

Generated with Coder Agents.

[github-actions]

Module Scorecard Check

coder-labs/cursor-cli: first scorecard, 58 / 100

No specific score is required to contribute, but modules with higher scores are more likely to be approved by the Coder team and widely used.

Full scorecard for this PR

Presentation & Onboarding	Agent Integration	Credential Hygiene	Restricted-Environment Readiness	Engineering Quality	Overall
6 / 17	10 / 25	20 / 20	7 / 20	10 / 10	58 / 100

Drilldown

Presentation & Onboarding — 6 / 17

Criterion	Max	Score	Notes
Configuration-mode examples	12	6	README shows 4 examples (minimal api_key, standalone w/ coder_app launcher, MCP+rules, downstream coder_script chaining), but no dedicated example demonstrates install_cursor_cli = false (BYOB) or login-only auth, so not every major mode gets a documented example.
Visual preview	5	0	No image, GIF, or video anywhere in the README.

Credential Hygiene — 20 / 20

Criterion	Max	Score	Notes
Secrets marked sensitive	16	16	api_key is declared sensitive = true in main.tf; all README examples reference var.cursor_api_key with no inline literal/placeholder keys.
Non-hardcoded auth path	4	4	README's Configuration section documents cursor-agent login as an alternative to passing api_key, an interactive auth flow that avoids pasting raw keys.

Restricted-Environment Readiness — 7 / 20

Criterion	Max	Score	Notes
Mirrorable artifact source	5	0	Install script hardcodes https://cursor.com/install in scripts/install.sh.tftpl; no module variable overrides this download URL.
Bring-your-own binary	10	5	install_cursor_cli = false skips installation and validates an existing cursor-agent on PATH (tested in main.test.ts and main.tftest.hcl), but this mode is only documented via the main.tf variable description, not demonstrated as a README example/section.
Egress transparency	3	0	No dedicated network/offline section in README; the cursor.com/install endpoint is never enumerated anywhere in the docs.
Runs without sudo	2	2	scripts/install.sh.tftpl never invokes sudo; installs to $HOME/.local, writes to $HOME/.cursor, all as the unprivileged user.

Engineering Quality — 10 / 10

Criterion	Max	Score	Notes
Input quality	6	6	All variables have clear descriptions and sensible defaults (install_cursor_cli=true, api_key="", etc.); rules_files has a validation block requiring workdir to be set.
Test coverage	4	4	main.tftest.hcl covers defaults, trimming, validation failure, env var creation, and scripts output ordering; main.test.ts thoroughly covers end-to-end install, MCP merge/backup, trust marker, and rules-file behavior in containers.

Agent Integration — 10 / 25

Criterion	Max	Score	Notes
AI governance	10	0	No mention of Coder AI Gateway or Agent Firewall anywhere in the README.
Dashboard entry point	5	5	README's "Standalone mode with a launcher app" example shows a full coder_app resource (slug, display_name, icon, open_in, launch command).
Session continuity	5	0	No documentation of resuming/continuing sessions, nor use of a persistent session manager (tmux/screen/boo).
Managed configuration	5	5	mcp variable merges into ~/.cursor/mcp.json and rules_files writes to <workdir>/.cursor/rules/, both documented with a dedicated example ("MCP servers and rules").

Overall — 58 / 100

Raw 53 / 92 → round(53 / 92 × 100) = 58

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.

[matifali]

@codex review

[chatgpt-codex-connector]

Codex Review Summary

This comment shows the latest Codex review activity on this pull request.

Review	Status	Commit	Review trigger
📝 Code Review	✅ Completed 2026-10-12T07:10:23.498243Z	61be2d3	Manual request

ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

• Open a pull request for review
• Mark a draft as ready
• Comment "@codex review" or "@codex security review".

Codex reacts with 👀 while any review is running, comments if it has suggestions, and reacts with 👍 once all reviews finish with no findings.

[matifali]

Lets also allow passing arbitrary flags as extra-args or additional-args variables similar to some of or modules like code-server

[chatgpt-codex-connector]

💡 Codex Review

Here are some automated review suggestions for this pull request.

Reviewed commit: 61be2d3032

ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

• Open a pull request for review
• Mark a draft as ready
• Comment "@codex review".

If Codex has suggestions, it will comment; otherwise it will react with 👍.

Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".

[linear-code]

REG-84

[coder]

Automated review requested by @matifali. Please sanity-check before acting on it.

• URL: refactor(registry/coder-labs/modules/cursor-cli)!: remove agentapi, tasks and start logic #1152
• Head reviewed: 771609fe42c0822de36296d2f16b546dfeec3d67
• Version: 0.3.1 → 1.0.0 in every README example (correct major bump; title has !)
• Verdict: approve with nits. No blockers. Do the manual trust-marker check before merge.

CI and existing reviews

• CI on the head sha is green, with the same job set as feat(registry/coder-labs/modules/copilot): migrate off AgentAPI and Tasks #1117. The Compare module scores against main check run shows in_progress, but the scorecard workflow finished (70/100).
• The Codex bot left two P2 comments ("mcpServers must be an object" and "preserve root workdir"). The author replied to both, removing the validation and keeping the trimsuffix. Acceptable.
• [SHOULD-FIX] Unanswered: matifali asked for an extra_args/additional_args variable. There's no reply and no change. Since v1 doesn't launch Cursor, flags belong in the user's coder_app (the README shows --model gpt-5 --force). The author should reply saying so, or agree on an alternative.

Correctness

The Terraform is coherent. required_version >= 1.9 is set, which the cross-variable rules_files validation needs. All variables are used, the unused workspace data sources and task_app_id were removed, and scripts was added. The install script is solid: retrying curl to a temp file, bash -n validation, --version check, a clear failure when install_cursor_cli=false and no binary exists, workdir creation, trust marker, MCP merge and rules files. api_key never appears in the script (there's a test for it).

[SHOULD-FIX] Stale v0 project-level mcp.json with the task-reporting server
v0 wrote <folder>/.cursor/mcp.json with a coder server pointing at /tmp/mcp-hack.sh. That script set CODER_MCP_APP_STATUS_SLUG and CODER_MCP_AI_AGENTAPI_URL=http://localhost:3284 and ran coder exp mcp server. v1 never touches the project file, so after an upgrade the stale entry stays in the user's repo, keeps failing to start, and could get committed.

Fix: when workdir is set, delete mcpServers.coder from <workdir>/.cursor/mcp.json if its command is /tmp/mcp-hack.sh, and remove the file if it ends up empty. At minimum, mention it in the README warning.

[SHOULD-FIX] Add nullable = false to mcp and rules_files
Both defaulted to null in v0, so templates often pass cond ? x : null. In v1, var.mcp != "" is true for null, so base64encode(null) errors at plan (main.tf:96), and length(null) errors in the rules_files validation (main.tf:72). With nullable = false, null falls back to the default.

[SHOULD-FIX / verify] The trust marker depends on undocumented CLI internals
install.sh.tftpl:139-165 writes ${CURSOR_DATA_DIR:-~/.cursor}/projects/<slug>/.workspace-trusted. The author says the path came from the CLI bundle and couldn't be verified end to end. Someone with a Cursor key should confirm the trust prompt is skipped. Describe it as best-effort in the README.

Leftovers

• None remain in main.tf, the script or the README: no agentapi, CODER_MCP_*, 3284, coder exp mcp, ai_prompt or coder_report_task. Tests assert there's no coder server and no agentapi in the log.
• [NIT] main.test.ts:21 imports agentapi/test-util (the same series follow-up as feat(registry/coder-labs/modules/copilot): migrate off AgentAPI and Tasks #1117).

Breaking changes and migration

• The README warning (README.md:22-23) is good. It covers dropping Tasks/AgentAPI, folder→workdir (now optional), model/force/ai_prompt moving to launcher flags, MCP moving to the user-level file, and "keep using v0.x.x".
• [NIT] Not mentioned: the removed order, group, install_agentapi and agentapi_version; the removed task_app_id output; that v0 force defaulted to true (users must now add --force); and the stale project mcp.json cleanup. A short old→new list in the PR body, linked from the README like claude-code and codex v5, would cover all of these.

Tests

• cursor-cli.tftest.hcl was renamed to main.tftest.hcl (8 runs: defaults, trim, api_key env and absence from the script, cursor_config_dir, the rules→workdir validation, scripts ordering). main.test.ts has 14 container tests, including the installer path with a curl stub, download failure, marker preservation, merge precedence, the invalid-JSON backup, and config dir isolation. Coverage is strong.
• Gaps (nits): no test for a non-object mcpServers (it fails with a raw jq error), and no stale-coder-cleanup test (if that cleanup is added).

Consistency and other nits

• The two PRs handle an unparseable MCP file differently. Here it's backed up and replaced (install.sh.tftpl:182-196), so hand-edited JSONC servers silently stop loading; copilot warns and skips. Pick one approach for the series. Skip-and-warn, or stripping comments, is safer.
• Give the merge a readable error: ($custom | if type=="object" then . else error("mcp.mcpServers must be an object") end).
• rules_files keys are used as filenames without checks. Add a validation that rejects / and ...
• Variable attribute order (type before description, and sensitive before default on api_key) differs from AGENTS.md. The series does the same, so this is low priority.
• The AI footer doesn't name the model.

[35C4n0r]

Lets also allow passing arbitrary flags as extra-args or additional-args variables similar to some of or modules like code-server

@matifali With this change the module only installs and configures Cursor CLI; it no longer launches cursor-agent (no start script or coder_app), so there is nothing for launch flags to attach to. code-server takes extra_args because that module starts the process.

Flags now belong in the template's own launcher, which the README shows:

command = <<-EOT
  #!/usr/bin/env bash
  set -e
  cd "${local.cursor_workdir}"
  exec cursor-agent --model gpt-5 --force
EOT

So I'm not adding an extra_args variable here.

Generated with Coder Agents.

[matifali]

Makes sense. Thanks

[35C4n0r]

Thanks for the automated review. Its verdict is "approve with nits, no blockers", and nothing in it is blocking enough to change in this PR:

• extra_args: answered above; launch flags belong in the template's coder_app now.
• Stale v0 coder server in .cursor/mcp.json: the entry only points at /tmp/mcp-hack.sh (no secret in the file) and just fails to start; Cursor CLI keeps working. The same upgrade leftover applies across the series, so it is better handled consistently there.
• nullable = false: mcp and rules_files now default to "" and {}, so this only matters when a template explicitly passes null.
• Trust marker: path comes from the Cursor CLI bundle and is covered by tests; an end-to-end check needs a Cursor key.
• Remaining items are nits.

@DevelopmentCats could you take a look when you get a chance?

Generated with Coder Agents.

[DevelopmentCats]

I agree with you on the above and have tested it on my end