[only21mil]

Problem

The claude CLI declares --add-dir <directories...>. The flag can repeat, and its space form takes several directories: the CLI's own example is claude --add-dir ../apps ../lib. If you put that in Settings → Providers → Claude → Launch arguments, T3 keeps only one directory.

The cause is parseCliArgs, which stores one value per flag. A repeated --add-dir keeps only its last value, and the extra words of the space form become positionals that nothing reads. Two things break:

• The Claude process gets only one of the launch-arg directories. The others are outside its permission scope, so Claude asks before touching files there.
• Skill discovery scans only that one directory. Skills under the other directories' .claude/skills are missing from the $ and / menus.

The SDK's extraArgs is a Record<string, string | null>, so it can't carry a second --add-dir value anyway.

Change

• claudeLaunchArgDirectories (in ClaudeSkills.ts) reads every --add-dir value the way the CLI does: repeated flags, the = form, and each word after the flag until the next option or --. It uses the existing tokenizeCliArgs, so quoting works as before.
• discoverClaudeSkills scans each of those directories, with the same relative-path rule as before.
• makeClaudeQueryOptions takes add-dir out of extraArgs, the same way it already handles append-system-prompt. It adds the directories to the SDK's typed additionalDirectories after the cwd and the attachments directory. The SDK sends each entry as its own --add-dir argument, and relative paths still resolve against the session cwd.

A bare --add-dir with no directory after it now adds nothing.

Size: +39/−16 production lines in two files. Tests: two new launch-arg rows in the existing discovery test and one new makeClaudeQueryOptions test.

Scope and approval

This is a focused bug fix for an existing setting. No issue covers the multi-directory case. #20156 reported that skills from --add-dir were missing from the menus, and #20167 fixed that for one directory. This PR extends the same path to every directory the CLI accepts. Two open PRs touch nearby lines. #21336 makes parseCliArgs stop at --, and the new helper also stops there. #22864 handles --settings, and its diff still applies cleanly on this branch.

Verification

New tests against main's source, then on the branch:

I ran the real app on Linux, using production builds of main and this branch (vp build for web, vp pack for server). Each run had fresh state and its own HOME. Claude was the real claude CLI binary, version 2.1.306. ANTHROPIC_BASE_URL pointed it at a local Messages API stub, with a placeholder key, so no account was used. Both builds used the same launch arguments, and each of the two directories held one skill:

Typing $ in a Claude thread:

Next I sent a message. While the turn ran I read the --add-dir arguments of the running Claude process, and I also read the environment block Claude sent to the stub. On main, ../shared-b is missing from both. On the branch, both directories reach Claude.

Other checks on the branch:

• Full ClaudeSkills.test.ts and ClaudeAdapterV2.test.ts: 228 passed. The existing --add-dir missing --add-dir "<dir>" row still passes.
• Typecheck, lint and format: pass. The one lint warning, an unused layer in ClaudeAdapterV2.ts, is already on main.

# in apps/server
pnpm exec vp test run src/provider/Drivers/ClaudeSkills.test.ts src/orchestration-v2/Adapters/ClaudeAdapterV2.test.ts
pnpm run typecheck
# in the repo root
FILES="apps/server/src/provider/Drivers/ClaudeSkills.ts apps/server/src/provider/Drivers/ClaudeSkills.test.ts apps/server/src/orchestration-v2/Adapters/ClaudeAdapterV2.ts apps/server/src/orchestration-v2/Adapters/ClaudeAdapterV2.test.ts"
pnpm exec vp lint $FILES
pnpm exec vp fmt --check $FILES

Not checked:

• macOS and Windows. The change has no platform-specific code, but I only ran Linux.
• A tool call that reads a file in the second directory. The stub only answers with text, so the evidence is the arguments and the environment block Claude reported, not a permission prompt.
• --allowedTools and --disallowedTools lose extra words through the same parser. That is a separate change.

Prior PRs

#20167 by @ScottN-PV made skill discovery read --add-dir from launch args, for issue #20156 by @Young-caveman. This PR keeps that discovery path and reads every directory instead of one.

Production fix, tests, PR description and images written with Claude Opus 5.5 in T3 Code Nightly.

🤖 Generated with Claude Opus 5.5 in T3 Code Nightly

[macroscopeapp]

Approvability

Verdict: Approved at f32a0e9

Macroscope's review found this PR approvable — This is a focused fix that preserves the default Claude path while correctly forwarding every explicitly configured --add-dir directory and discovering skills from those locations. The production change is localized and covered by targeted tests, with no schema, deployment, or static-analysis configuration impact.

^{You can add or adjust custom eligibility rules. Learn more.}

[coderabbitai]

No actionable comments were generated in the recent review.

Recent review info

Run configuration

• Source mode: Auto
• Repository: pingdotgg/t3code
• Branch: main
• File: .coderabbit.config.ts
• Commit: 62245d8
• Layers: Repository TS config
• Review profile: CHILL
• Plan: Advanced
• Run ID: 9bd6a73a-1a15-4888-8828-dfa934ad66c2

Commits

Reviewing files that changed from the base of the PR and between 441c08c and 79feda1.

Files selected for processing: 3

• apps/server/src/orchestration-v2/Adapters/ClaudeAdapterV2.test.ts
• apps/server/src/orchestration-v2/Adapters/ClaudeAdapterV2.ts
• apps/server/src/provider/Drivers/ClaudeSkills.ts

Included review availability: This review used your included allowance. Your plan provides up to 10 included reviews per hour; 0 remain after this review.

Walkthrough

Claude launch-argument handling now collects values from repeated and multi-value --add-dir flags. Skill discovery uses eligible directories. Claude query options include parsed directories in additionalDirectories instead of extraArgs.

Changes: Claude add-dir handling

Layer / File(s)	Summary
Collect directories for discovery apps/server/src/provider/Drivers/ClaudeSkills.ts, apps/server/src/provider/Drivers/ClaudeSkills.test.ts	claudeLaunchArgDirectories collects values from repeated and multi-value --add-dir arguments. Skill discovery adds eligible values as roots. Tests cover arguments adjacent to missing directory paths.
Pass directories to Claude query options apps/server/src/orchestration-v2/Adapters/ClaudeAdapterV2.ts, apps/server/src/orchestration-v2/Adapters/ClaudeAdapterV2.test.ts	Claude query options remove --add-dir from extraArgs and include parsed directory values in additionalDirectories. The test checks the workspace and attachments directories alongside the supplied values.

Priority: Normal

Estimated code review effort: 2 (Simple) | ~10 minutes

Change: Bug fix

Suggested reviewers: maria-rcks, juliusmarminge

Merge Risk: Minimal · up to 79fed

The reviewed directory handling has no identified merge-blocking issue; merge after normal checks.

Pre-merge checks: 4

Check name	Status	Explanation
Description check		The description includes all required sections: Problem, Change, Scope and approval, and Verification. It explains the bug, implementation, rationale for scope, test results, manual verification, and …
Title check		The title clearly and concisely describes the primary fix: forwarding every Claude --add-dir launch directory.
Linked Issues check		Check skipped because no linked issues were found for this pull request.
Out of Scope Changes check		Check skipped because no linked issues were found for this pull request.

Full details: Description check

Explanation

The description includes all required sections: Problem, Change, Scope and approval, and Verification. It explains the bug, implementation, rationale for scope, test results, manual verification, and known limitations.

Finishing Touches

🧪 Generate unit tests (beta)

• Create a new PR

• Autofix · Keep fixing CodeRabbit findings and required CI, and resolving merge conflicts

_{Comment @coderabbitai help to get the list of available commands.}

[maria-rcks]

Note

Written by gpt-6.1-sol on behalf of Maria

@only21mil, i updated your writable head and kept your commits. your shared directory reader and typed sdk forwarding fix both launch permissions and skill discovery. i merged current main, resolved the discovery conflict, and kept main's shorter comments. the final diff is four files, +58/-15; no unrelated behavior was added.

verification at 79feda1e39851644a88a17a5614ecffd36f37c08:

• both regressions fail with current main's production source, then pass with this fix.
• blacksmith: all 256 tests in ClaudeSkills.test.ts and ClaudeAdapterV2.test.ts pass. vp run check:changed passes format, lint, knip, graphql, affected typechecks, and tests.
• real web client and claude 2.1.310: identical launch args with variadic alpha/beta and repeated gamma show only gamma on main, and all three on the fix. verified the real process argv and directories in the messages api request. selecting alpha inserts the correct skill mention.
• all 28 ci checks are green. coderabbit reviewed this head with no actionable findings. two independent reviews approve it with no blocking findings.

unverified: paid anthropic responses and a tool permission prompt, macos/windows, and native desktop/mobile clients. runtime verification used the real claude cli against a local messages api stub.

known low limit from review: a first space-form directory starting with - is skipped; use --add-dir=-scratch for that unusual directory name. other variadic flags remain outside this fix.

before, current main:

after, same directories and settings:

[maria-rcks]

Note

Written by gpt-6.1-sol on behalf of Maria

Verified in a real dev environment

verified as-is at 79feda1e39851644a88a17a5614ecffd36f37c08 on this disposable linux boat vm. no code changes or pushes.

• reproduced on github/main at 9c6429d74b in /home/user/main, then ran the exact pr head in /home/user/t3code. installed both with vp i; ran one background dev server at a time.
• drove the actual web app with repository playwright and /usr/bin/google-chrome. created a local git project through the app. a real codex gpt-6.1-sol turn created its documentation and sibling skill directories. reused those exact files on the pr.
• set identical claude launch arguments in settings: --add-dir ../shared-alpha ../shared-beta --verbose --add-dir="../shared gamma". main showed gamma and the local skill, omitting alpha and beta. the pr showed all four project skills in both $ and / menus. selecting alpha inserted the correct skill mention.
• ran the real claude 2.1.312 cli with claude-haiku-5-5 through the vm's authenticated local proxy. no mocks or messages-api stub. on the pr, selected alpha and requested reads of all three external notes files plus local notes. supervised mode (--permission-mode default) completed all four Read calls without an approval prompt and returned ALPHA-23281, BETA-23281, and GAMMA-23281.
• neighboring behavior: the cwd-local skill remained discoverable, the quoted = directory worked, a resumed claude turn completed, and a separate real codex gpt-6.1-sol turn on the pr read README.md and local notes successfully.
• blacksmith: vp test run apps/server/src/provider/Drivers/ClaudeSkills.test.ts apps/server/src/orchestration-v2/Adapters/ClaudeAdapterV2.test.ts passed all 256 tests. existing ci reports 28 successful, zero failing or pending.

short runtime transcript, captured from /proc/<claude-pid>/cmdline while each real turn ran:

main, pid 27429, --add-dir values:
  /home/user/main/.t3/projects/add-dir-verification
  /home/user/main/.t3/userdata/attachments
  ../shared gamma

pr, pid 32954, --permission-mode default, --add-dir values:
  /home/user/main/.t3/projects/add-dir-verification
  /home/user/.t3/dev/attachments
  ../shared-alpha
  ../shared-beta
  ../shared gamma

pr web activity:
  Read /home/user/main/.t3/projects/shared-alpha/notes.txt
  Read /home/user/main/.t3/projects/shared-beta/notes.txt
  Read /home/user/main/.t3/projects/shared gamma/notes.txt
  Read add-dir-verification/notes.txt
  completed; all three external markers returned

before, main with the same skill files and launch arguments:

after, the exact pr head:

inspected the captures and verified the uploaded recording plays and seeks in chrome.

Not verified

native electron, ios/android, macos/windows, remote relay/tunnel connections, and edits outside the workspace. the provider verification used real responses through the vm proxy rather than direct anthropic subscription login. the directory permission check covered reads in supervised mode; main's baseline turn used full access, so it proves the missing arguments and skills rather than reproducing a read permission prompt.

[maria-rcks]

Note

Written by gpt-6.1-sol on behalf of Maria

Verified in a real dev environment

verified as-is at 79feda1e39851644a88a17a5614ecffd36f37c08. no code changes or pushes.

• installed github/main (eb751fa0730d6a51d56547cb7fcaf4822a7e0e3b) in a separate worktree and the exact pr head with vp i. ran one background dev server at a time with isolated state.
• drove the real web client using repository playwright and /usr/bin/google-chrome. created a project for a real git repository through the app. a real codex gpt-6.1-sol turn created the documentation and sibling skill directories. both revisions used those same files.
• entered identical claude launch arguments through settings: --add-dir ../shared-alpha ../shared-beta --verbose --add-dir="../shared gamma".
• main omitted alpha and beta from the skill menu. a real supervised claude turn then stopped both reads with "Path is outside allowed working directories". its process argv included only gamma.
• the pr showed alpha, beta, gamma, and the cwd-local skill in both $ and / menus. selected alpha from the menu and ran a real claude-haiku-5-5 turn using claude cli 2.1.312 through the vm’s authenticated local proxy. all four Read calls completed in supervised mode without an approval prompt and returned the expected markers. no mocks or api stub.
• nearby behavior: the quoted = directory and cwd-local skill still worked; a resumed claude turn completed; a separate real codex gpt-6.1-sol turn on the pr read README.md and notes.txt and reported LOCAL-23281 without edits.
• blacksmith: both existing ClaudeSkills.test.ts and ClaudeAdapterV2.test.ts passed, 256 tests total. current ci reports 28 successful, zero failing or pending.

short runtime transcript, read from the real process argv during each turn:

main supervised, pid 162685:
  --permission-mode default
  --add-dir /home/user/pr23281-workspace
  --add-dir /home/user/main/.t3/userdata/attachments
  --add-dir ../shared gamma
  alpha and beta: Path is outside allowed working directories

pr supervised, pid 168100:
  --permission-mode default
  --add-dir /home/user/pr23281-workspace
  --add-dir /home/user/t3code/.t3/userdata/attachments
  --add-dir ../shared-alpha
  --add-dir ../shared-beta
  --add-dir ../shared gamma
  completed four Read calls, no approval prompt:
  ALPHA-23281 BETA-23281 GAMMA-23281 LOCAL-23281

main, same files and launch arguments:

the exact pr head:

inspected the captures and verified uploaded video playback and seeking in t3’s browser player.

Not verified

native electron, ios/android, macos/windows, remote relay/tunnel connections, and writes outside the workspace. provider responses used the vm’s real authenticated local proxy rather than direct anthropic subscription login. attachment directory forwarding was observed in argv; attachment reads were not exercised.