fix(code-server,vscode-web): parse JSONC in extensions.json and .code-workspace

[rodmk]

Fixes #708. Supersedes #707.

Description

auto_install_extensions fails to parse .vscode/extensions.json and .code-workspace files that contain valid JSONC (block comments, line comments, trailing commas). VS Code officially supports these via its built-in JSONC parser with default options.

This PR replaces the original sed 's|//.*||g' | jq pipeline with a small, portable strip_jsonc_for_extensions helper (defined once per script) that strips JSONC features before handing the file to jq — no new dependencies, no GNU-only extensions.

Type of Change

• Bug fix

Module Information

Path: registry/coder/modules/code-server and registry/coder/modules/vscode-web
New version: code-server 1.5.2, vscode-web 1.6.2
Breaking change: No

Problem

The original sed 's|//.*||g' | jq pipeline only handled line comments; block comments and trailing commas broke jq. Naive fixes introduce two portability traps and one ordering bug:

1. sed -z (slurp the whole file) is a GNU extension — it does not exist on BSD sed (macOS) and would break there.
2. [^\n] inside a bracket expression is a GNU extension — BSD sed reads \n in a bracket as a literal backslash and n, so //[^\n]* silently corrupts any extension ID containing the letter n (e.g. dbaeumer.vscode-eslint).
3. Stripping line comments before block comments corrupts a // that lives inside a block comment (e.g. a URL in /* see https://open-vsx.org */), leaving an unterminated /* that makes jq fail — installing zero extensions with no error surfaced. A single-line/minified file was also dropped entirely by the :a;N;$!ba slurp idiom.

Fix

strip_jsonc_for_extensions runs three ordered sed passes, each scoped to what it needs and portable across GNU (Linux), BSD (macOS), and BusyBox (Alpine) sed:

strip_jsonc_for_extensions() {
  sed -E ':a
$!{
N
ba
}
s#/[*]([^*]|[*]+[^*/])*[*]+/##g' "$1" \
    | sed -E 's#^[[:space:]]*//.*##; s#([^:])//.*#\1#' \
    | sed -E ':a
$!{
N
ba
}
s/,[^]}"]*([]}])/\1/g'
}

1. Block comments first — slurps the whole file so /* ... */ can span lines. Running first means a URL inside a block comment is removed as a unit before the line-comment pass ever sees its //.
2. Line comments per line — // ... stops at end of line without the non-portable [^\n] class. A // preceded by : is preserved, so URLs in string values (e.g. "http.proxy": "https://…" in a .code-workspace) survive.
3. Trailing commas — slurps the whole file so a comma and its closing bracket may sit on different lines.

The :a;$!{N;ba} slurp idiom is single-line safe (it falls through on the last/only line), unlike :a;N;$!ba, which drops single-line input.

Also makes the code-server recommendations query null-safe to match vscode-web (.recommendations[] → (.recommendations // [])[]), so a valid extensions.json with no recommendations key no longer errors.

Known limitation

This is not a general-purpose JSONC parser. A non-URL // inside a string value (e.g. a regex setting "a//b") can still be mis-stripped — only a real string-aware parser could distinguish it, and a hand-rolled parser was rejected in #707 as over-engineered. Extension IDs are restricted to [a-z0-9\-.] by vsce's nameRegex, and URLs (the common //-in-string case) are handled by the :// guard, so the recommendations arrays this feature actually reads are safe.

Alternatives considered

Approach	Tradeoff
sed -z	Simpler slurp, but -z is a GNU extension — breaks on macOS BSD sed
Single sed -E with //[^\n]*	Works on GNU sed, but [^\n] in a bracket is not portable — BSD sed treats it as literal \ + n, silently corrupting IDs containing n
perl -0777 -pe	Clean one-liner, but no module in the repo uses perl and Apple may remove it from future macOS
Node.js state machine	Correct JSONC parsing, but rejected in #707 as over-engineered
Python3	Could parse JSONC natively, but not guaranteed on all workspaces and would need a command -v guard
Three-pass sed (this PR)	Two extra pipes, but zero new dependencies, fully portable across GNU/BSD/BusyBox sed, and fixes comment/URL ordering + single-line + null-safety

Testing & Validation

• terraform test: code-server 8/8, vscode-web 3/3
• code-server + vscode-web TypeScript suites pass in CI (incl. the container installs and runs code-server test)
• prettier (with prettier-plugin-sh) and shellcheck --severity=warning clean; terraform fmt clean
• templatefile() renders both run.sh files; rendered scripts pass bash -n
• Version bumped

The strip_jsonc_for_extensions helper was validated with a 25-case matrix (the original 19 JSONC cases plus new regression cases for URL-in-block-comment, single-line/minified input, IDs containing n, a line comment containing a quote before a trailing comma, and URL-valued .code-workspace settings). The matrix was run against the exact bytes committed to both run.sh files and passes 25/25 on GNU sed (Ubuntu), BSD sed (macOS), and BusyBox sed (Alpine).

Related Issues

• Fixes auto_install_extensions fails on JSONC comments and trailing commas #708
• Supersedes fix: properly parse JSONC in extensions.json and batch extension installs #707

🤖 Generated with Claude Code using Claude Opus 4.8

[Copilot]

Pull request overview

This PR improves auto-installation of VS Code/code-server extensions by attempting to preprocess JSONC workspace/extension files (comments + trailing commas) before feeding them to jq, and updates the module README version pins accordingly.

Changes:

• Add a strip_jsonc_for_extensions() helper in both vscode-web and code-server modules, then use it when extracting extension recommendations via jq.
• Make jq parsing tolerant of missing recommendations arrays by defaulting to [].
• Bump module versions referenced in READMEs and add a typos dictionary entry to avoid false positives.

Reviewed changes

Copilot reviewed 5 out of 5 changed files in this pull request and generated 4 comments.

Show a summary per file

File	Description
registry/coder/modules/vscode-web/run.sh	Adds JSONC preprocessing helper and uses it for workspace/extensions parsing before jq.
registry/coder/modules/vscode-web/README.md	Updates example module version pins to 1.5.1.
registry/coder/modules/code-server/run.sh	Adds JSONC preprocessing helper and uses it for .vscode/extensions.json parsing before jq.
registry/coder/modules/code-server/README.md	Updates example module version pins to 1.4.4.
.github/typos.toml	Adds ba to the allowed word list (sed label pattern).

[DevelopmentCats]

I agree with you that the Copilot Comments were wrong.

This LGTM though thanks for the fix and quick response!! 😸

[35C4n0r]

@rodmk can you resolve the conflicts.

[rodmk]

@35C4n0r Whoa, I didn't realize this hadn't been merged yet, it was approved months ago 😅 I'll resolve the conflicts promptly.

[rodmk]

Resolved the conflicts against the latest main. The only real overlap was in the code-server/vscode-web run.sh scripts: I merged the new WORKSPACE-file support together with this PR's JSONC parsing, so both commented extensions.json files and .code-workspace files work. Version bumps re-based onto the current releases — code-server → 1.5.1, vscode-web → 1.6.1.

Validation: fmt, README validation, golangci-lint, and the full code-server + vscode-web test suites all pass. (The one failing test is an unrelated nexus-repository timeout flake.)

🤖 Generated with Claude Code using Claude Opus 4.8

[35C4n0r]

Two P1 bugs found in the strip_jsonc_for_extensions shell function introduced by this PR. Both are verified with live tests. Each inline comment above includes a suggested fix or a plain diff.

Summary of required changes:

#	File	Issue	Severity
1	code-server/run.sh lines 119–128	Line-comment pass runs before block-comment slurp — corrupts URLs in comments and settings; single-line files pass through unprocessed	P1
2	vscode-web/run.sh lines 153–162	Same stage-ordering bug	P1
3	code-server/run.sh line 137	RECOMMENDATIONS_QUERY not null-safe (vscode-web was fixed, code-server was not)	P1

Generated by Coder Agents on behalf of @35C4n0r.

[rodmk]

Thanks for the careful review — all three findings were valid and are fixed in f29505a. I reproduced each one on BSD sed (macOS) first, which also turned up a portability issue in the suggested patch, so I took a slightly different route to fix the same bugs. Details below.

1 & 2 — line-comment pass running before the block-comment slurp (both files)

Confirmed, both failure modes reproduce:

• URL inside a block comment — /* see https://open-vsx.org */ → the first-stage s|//.*||g strips //open-vsx.org */, leaving an unterminated /* see https: that jq rejects. Zero extensions, no error surfaced.
• Single-line / minified file — :a;N;$!ba drops single-line input entirely (on BSD it produced empty output).

On the suggested fix: collapsing into one sed -E with s#//[^\n]*##g fixes the block-comment case, but [^\n] inside a bracket expression is a GNU extension. On BSD sed (macOS, which this module supports via platform = "darwin") \n in a bracket is read as a literal backslash + n, so //[^\n]* stops at the first n and silently corrupts IDs — e.g. dbaeumer.vscode-eslint came out as nting extension. Verified locally.

So I kept the fix portable with three ordered passes instead:

strip_jsonc_for_extensions() {
  sed -E ':a
$!{
N
ba
}
s#/[*]([^*]|[*]+[^*/])*[*]+/##g' "$1" \
    | sed -E 's#^[[:space:]]*//.*##; s#([^:])//.*#\1#' \
    | sed -E ':a
$!{
N
ba
}
s/,[^]}"]*([]}])/\1/g'
}

• Block comments first (whole-file slurp) so a // inside /* ... */ is removed as a unit before the line-comment pass sees it.
• Line comments per line — no [^\n]; a // preceded by : is preserved so URLs in string values survive.
• Trailing commas last (slurp).
• :a;$!{N;ba} is single-line safe (falls through on the last/only line), fixing the minified-file case.

3 — code-server null-safety

Fixed: RECOMMENDATIONS_QUERY is now (.recommendations // [])[] (and the workspace query (.extensions.recommendations // [])[]), matching vscode-web.

On failure mode 2 (URL in a .code-workspace setting value)

Agreed this is the fundamental limit you called out — a purely regex-based stripper can't reliably tell a // in a string from a comment. The ://-guard in the line-comment pass handles the realistic case (URL-valued settings such as "http.proxy": "https://proxy.corp:8080" now survive). A non-URL // inside a string value is still out of reach without a real parser, which #707 established as over-engineered; I documented this as a known limitation in the PR description.

Validation

The helper was run against a 25-case matrix (the original 19 JSONC cases + regressions for URL-in-block-comment, single-line input, IDs containing n, a quote inside a line comment before a trailing comma, and URL-valued .code-workspace settings), using the exact bytes committed to both run.sh files. 25/25 on GNU sed (Ubuntu), BSD sed (macOS), and BusyBox sed (Alpine). terraform test, the code-server/vscode-web TS suites, shellcheck, and prettier all pass. Versions bumped to code-server 1.5.2 / vscode-web 1.6.1.

🤖 Generated with Claude Code using Claude Opus 4.8

[35C4n0r]

@rodmk can you resolve the merge conflict.

[rodmk]

Yup sorry just noticed, doing that now.

[rodmk]

All set - no more conflicts for now.

[35C4n0r]

@rodmk can you add some tests in main.test.ts for both the modules to test the changes.

[rodmk]

Added tests in main.test.ts for both modules (b0fc629) that drive the rendered script through the auto_install_extensions path with a mock CLI recording --install-extension calls — covering JSONC comments (incl. a URL inside a block comment), trailing commas, the null-safe query, and a .code-workspace with a URL-valued setting. They fail on the pre-fix code.

One thing worth flagging: code-server reaches the install path cleanly via use_cached, but vscode-web early-exits on use_cached/offline, so its test stubs curl/tar to no-op the download and let the pre-placed mock binary survive. All green in CI.

🤖 Generated with Claude Code using Claude Opus 4.8