[kyle944]

Problem

Since #28410, every Claude process T3 Code starts is labeled CLAUDE_CODE_ENTRYPOINT=t3-code. On a Console organization that holds only the included API credits for subscribers, Anthropic answers that label with Credit balance is too low, so an API-key Claude instance fails every turn while the balance is available. T3 Code runs through the Agent SDK, which those credits cover.

Fixes #29393

Change

A Claude process that authenticates with an API key now leaves the entrypoint unset, so the Agent SDK applies its own sdk-ts, and text generation through claude -p gets sdk-cli.

The label is chosen when each process starts, from the settings that process will load:

• resolveClaudeSessionEnvironment runs for turns in ClaudeAdapterV2 and for commands in ClaudeTextGeneration, using their working directory. Before this, the label was chosen once, when the adapter was created.
• Those two paths start from mergeClaudeBaseEnvironment, which holds only a label the instance configured, so removing a label from settings takes effect on the next session.
• readClaudeSessionSettingsEnv reads settings in Claude's order: user, project, local, --settings, then managed. Files are read each time, so a new session sees edits.
• --project-config-root moves project and local settings to that directory, as in Claude Code.
• Settings left out by --setting-sources do not change the label. --settings and managed settings always apply. Root local settings apply only when the current user owns the root, .git and .claude, as in the Agent SDK. Tool search shares that lookup.
• Managed settings now match the Agent SDK: managed-settings.json, then sorted managed-settings.d/*.json, skipping hidden files. On Windows the directory is C:\Program Files\ClaudeCode, not PROGRAMDATA. This Windows path also applies to the existing skill settings reader.
• An empty ANTHROPIC_API_KEY in a higher layer counts as no key, because Claude's settings env beats the process env.
• A wrongly typed value in a settings file no longer hides the other values in that file.
• On Windows, all inherited names are upper-cased in the base environment, as instance variables already are, because a copy of process.env loses the case-insensitive lookup. This also lets T3 Code find CLAUDE_CONFIG_DIR however Windows spells it. Code in this repository that reads PATH accepts the upper-case name.
• This reader now parses strict JSON, because Claude Code skips a settings file with comments or trailing commas. It also serves ANTHROPIC_BASE_URL, which follows the same rule.
• The local settings lookup moves from ClaudeAdapterV2 to ClaudeSkills so both paths share it, with no change in behavior.

Nothing changes for subscription logins, which keep t3-code. A configured label from the instance or from settings still wins, and an inherited host value is still dropped, as #28410 intended.

The cost is that API-key requests no longer carry the t3-code label.

Scope and approval

There is no maintainer approval to link, and #29393 has not been triaged yet. This began under the small bug fix exception as one condition in one function. It grew when CodeRabbit showed that a key in project or local settings, or in --settings, would still fail, so it may now need a maintainer's view of scope. All of it serves the one problem in #29393. It keeps the t3-code default for subscription logins, which #28410 and #25597 were about.

This does not cover keys from apiKeyHelper or a Console /login, because they are not in the environment or the settings env. They keep today's behavior. Capability probes keep the label chosen at startup, because their never-yielding prompt sends no model request, according to the comment on probeClaudeCapabilities.

Verification

The problem, on one organization holding only these credits, with Claude Code 2.1.316 and desktop 0.1.7-nightly.20261102.4080:

Path	Entrypoint	Result
T3 Code, API-key instance	t3-code (default), Haiku 5.5 and Opus 5.5	Credit balance is too low
T3 Code, same instance and key	sdk-ts set in the instance environment, Haiku 5.5 and Opus 5.5	Worked, and the credit balance went down
claude -p with an API key	t3-code	Credit balance is too low
claude -p with an API key	sdk-cli, or unset	Worked

The Agent SDK 0.3.276 source sets sdk-ts only when the variable is unset, so the T3 Code row with sdk-ts is the value this change produces for turns.

An empty key in settings beats the process env, checked with Claude Code 2.1.316: with an invalid ANTHROPIC_API_KEY in the environment, claude -p returned 401 Invalid API key, and with --settings '{"env":{"ANTHROPIC_API_KEY":""}}' added, it succeeded on the subscription login.

Claude Code 2.1.316 also skips a non-strict settings file: with an invalid key in project settings, valid JSON returned 401 Invalid API key, while the same file with a line comment or a trailing comma succeeded on the subscription login.

Claude Code 2.1.316 also ignores project settings under --setting-sources user: an invalid key in the project file returned 401 Invalid API key with --setting-sources project, and authenticated normally with --setting-sources user.

The fix, from a checkout on macOS, with a failing test committed before each fix:

Test commit	Fails with	Fix commit
5d3e4f0	expected { …(3) } to deeply equal { ANTHROPIC_API_KEY: 'key', KEEP: '1' }	9613eff
fbe6dda	expected undefined to be 't3-code'	8a89037
069304e	Adapter: expected 't3-code' to equal undefined. Text generation: CLAUDE_CODE_ENTRYPOINT was unset.	cddc8c7
66118e3	mergeClaudeBaseEnvironment missing (TypeError)	bb5068f
6934847	expected undefined to be 't3-code'	307bcb5
d8d4f9a	expected { Anthropic_Api_Key: 'key', …(1) } to deeply equal { ANTHROPIC_API_KEY: 'key' }	339b177
eda0e6c	expected { Claude_Config_Dir: 'config', …(2) } to deeply equal { ANTHROPIC_API_KEY: 'key', …(2) }	f69ac89
a1dedaa	expected 't3-code' to be undefined	5d02ac7
ed320ef	expected undefined to be 't3-code' and expected 't3-code' to be undefined	2e5ae7b
e330887	expected undefined to equal 't3-code'	562c2c7

On 562c2c7, with main synced by the merge in 9d37d87:

• vp run check:changed passes: format, lint, knip, graphql, typecheck, and the 5 test files beside the change, 346 tests, which include ClaudeAdapterV2.test.ts.
• vp test run apps/server/src/provider/ProviderRegistry.test.ts passes, 50 of 50, because the same reader serves ANTHROPIC_BASE_URL.
• Those results used a resolved TMPDIR. With the default macOS TMPDIR, ClaudeSkills.test.ts > merges plugin enablement and selects installs in native registry order fails. It also fails on unmodified main (ac8dce0), so it is not related to this change.

I could not check:

• This build running end to end. I tested the resulting label through the released desktop app, not a dev build of this branch.
• A key in project or local settings against the real API. Those cases are covered by unit tests only.
  The Agent SDK types say that omitting settingSources loads all sources, and the adapter omits it, so a session does load those files.
• Real managed settings files, or MDM, registry and server-managed sources. The tests use a fake filesystem for the fixed managed paths.
• Skill settings still read managed-settings.json without drop-ins, as before this PR.
• A real Windows machine. The Windows name matching is tested with a simulated platform.
• The ClaudeDriver line that passes the base environment to text generation. The tests cover the functions on each side of it, but not that call.
• An organization that also has purchased credits, or one on a different plan.
• How Anthropic classifies entrypoints. I only observed that the result follows the label.

visible: yes

Model: Claude Opus 5.5 · Harness: Claude Code in T3 Code

🤖 Generated with Claude Code

[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	⚠️ Failed 2026-11-02T15:29:29.021729Z We couldn’t complete this review. Comment @codex review to try again.	9613eff	PR opened
🔒 Security Review	⚠️ Failed 2026-11-02T15:29:34.905450Z We couldn’t complete this review. Comment @codex security review to try again.	9613eff	PR opened

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.

[chatgpt-codex-connector]

You have reached your Codex usage limits for code reviews. You can see your limits in the Codex usage dashboard.

[chatgpt-codex-connector]

The account paying for this security review has reached its Codex usage limits. The payer can check the Codex usage dashboard. For personal accounts, using credits requires enabling “Use credits for security reviews” in Code review settings. If you do not manage the paying account, contact this repository's admins.

[macroscopeapp]

Approvability

Verdict: Needs Human Approval

Macroscope's review found this PR requires human approval — API-key Claude instances now default to the Agent SDK entrypoint, enabling subscriber API-credit usage while retaining explicit overrides. Human review is required to confirm this change to default credit-consumption behavior.

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

[coderabbitai]

Note

Reviews paused

It looks like this branch is under active development. To avoid overwhelming you with review comments due to an influx of new commits, CodeRabbit has automatically paused this review. You can configure this behavior by changing the reviews.auto_review.auto_pause_after_reviewed_commits setting.

• Last reviewed commit: 339b177
• Latest commit, not reviewed: 562c2c7

Use the following commands to manage reviews:

• @coderabbitai resume to resume automatic reviews.
• @coderabbitai review to trigger a single review.

Use the checkboxes below for quick actions:

• Resume reviews
• Trigger review

Walkthrough

Claude environment resolution now reads API keys and entrypoint values from settings and launch arguments. Provider, adapter, and text-generation paths use resolved environments. Tests cover entrypoint selection and environment-source precedence.

Changes: Claude session environment

Layer / File(s)	Summary
Settings loading and entrypoint selection apps/server/src/provider/Drivers/ClaudeSkills.ts, apps/server/src/provider/Drivers/ClaudeHome.ts, apps/server/src/provider/Drivers/ClaudeHome.test.ts	Settings loading accepts supported string values, including ANTHROPIC_API_KEY, from user, project, local, worktree, launch-argument, and managed sources. Environment merging selects the entrypoint based on settings, the current environment, and API-key presence. Tests cover source precedence and entrypoint behavior.
Base environment construction apps/server/src/provider/Drivers/ClaudeDriver.ts, apps/server/src/orchestration-v2/Adapters/ClaudeAdapterV2.ts	Provider and adapter setup uses a base environment that removes inherited entrypoint and Claude host-session variables before merging instance settings.
Session environment integration apps/server/src/orchestration-v2/Adapters/ClaudeAdapterV2.ts, apps/server/src/orchestration-v2/Adapters/ClaudeAdapterV2.test.ts, apps/server/src/textGeneration/ClaudeTextGeneration.ts, apps/server/src/textGeneration/ClaudeTextGeneration.test.ts, apps/server/src/textGeneration/testing/ClaudeTextGeneration.fixture.mjs	The adapter and text-generation paths use resolved session environments when launching Claude. The adapter uses the shared local-settings-directory resolver for tool-search settings. Tests and the fixture check expected entrypoint values.

Estimated code review effort: 3 (Moderate) | ~20 minutes

Change: Bug fix · Severity of issue fixed: Medium

Sequence Diagram(s)

sequenceDiagram
  participant ClaudeAdapterV2
  participant resolveClaudeSessionEnvironment
  participant readClaudeSessionSettingsEnv
  participant ClaudeProcess
  ClaudeAdapterV2->>resolveClaudeSessionEnvironment: Claude settings, environment, cwd, and launch arguments
  resolveClaudeSessionEnvironment->>readClaudeSessionSettingsEnv: load session settings
  readClaudeSessionSettingsEnv-->>resolveClaudeSessionEnvironment: merged settings environment
  resolveClaudeSessionEnvironment-->>ClaudeAdapterV2: resolved process environment
  ClaudeAdapterV2->>ClaudeProcess: launch with resolved environment

Loading

Suggested reviewers: t3dotgg, maria-rcks

Merge Risk: Low · up to 339b1

On Windows, a host that spells the Claude config directory variable in mixed case can still hit the low-credit error for API-key instances. This is a narrow edge case, and adding the variable to the canonicalized set fixes it. The rest of the change is low risk.

Pre-merge checks: 4

Check name	Status	Explanation
Title check		The title clearly identifies the main change: enabling API-key Claude instances to use subscriber API credits. It is concise, specific, and follows the repository's conventional commit format.
Linked Issues check		Issue #29393 requires API-key Claude instances to run with included subscriber API credits. mergeClaudeBaseEnvironment removes inherited CLAUDE_CODE_ENTRYPOINT, and `resolveClaudeSessionEnvironmen…
Out of Scope Changes check		The changes remain within issue #29393. Settings-file resolution extends the same API-key decision to user, project, local, launch, and managed settings. Per-session resolution is required because set…
Description check		The description follows the template's Problem, Change, Scope and approval, and Verification sections, and it is detailed. It states the bug, links #29393, and explains the behavior change for API-key…

Full details: Linked Issues check

Explanation

Issue #29393 requires API-key Claude instances to run with included subscriber API credits. mergeClaudeBaseEnvironment removes inherited CLAUDE_CODE_ENTRYPOINT, and resolveClaudeSessionEnvironment leaves it unset when the effective API key comes from the instance or Claude settings. ClaudeAdapterV2 and ClaudeTextGeneration use this per-session environment. Subscription logins retain t3-code. Tests cover API-key sources, settings precedence, Windows name matching, adapter turns, and text generation. These changes implement the issue requirement and add automated coverage.

Full details: Out of Scope Changes check

Explanation

The changes remain within issue #29393. Settings-file resolution extends the same API-key decision to user, project, local, launch, and managed settings. Per-session resolution is required because settings depend on the working directory and launch arguments. Shared local-settings resolution supports both affected execution paths. Strict parsing, precedence handling, Windows case matching, base-environment cleanup, and tests support the entrypoint fix. No unrelated change is demonstrated.

Full details: Description check

Explanation

The description follows the template's Problem, Change, Scope and approval, and Verification sections, and it is detailed. It states the bug, links #29393, and explains the behavior change for API-key and subscription logins. It also gives a thorough verification record with test commits, command results, and a clear list of what was not checked. The Scope and approval section does not link a maintainer approval. The author says the issue is untriaged and that the change grew past the small-fix exception, so the scope question is openly raised. The template requires either an approval link or an explicit justification for the small-fix exception. The author gives a partial justification, but the scope is larger than a one-line fix, so this section is incomplete. The template also asks for the model and harness, and the description includes them. Overall the description is mostly complete, with one required section only partly met.

Finishing Touches

Generate unit tests (beta)

• Create a new PR

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

Tip

🎯 Customize your risk assessments

Guide how CodeRabbit assesses merge risk and security architecture risk. Highlight sensitive changes and explain your security boundaries using Merge Risk Instructions and Security Risk Instructions in your review settings. Security architecture risk assessments require the Advanced plan or above.

✅ Automatic approvals, with your rules

Let CodeRabbit approve pull requests once required threads are resolved, pre-merge checks pass, and your risk limits are met. Enable Automatic approvals in review settings, with overrides for teams or file paths.

Recent review info

Run configuration

• Source mode: Auto
• Repository: pingdotgg/t3code
• Branch: main
• File: .coderabbit.config.ts
• Commit: ac8dce0
• Layers: Repository TS config
• Review profile: CHILL
• Plan: Advanced
• Run ID: c886c5a5-1c2f-45be-a6e3-b4c6a8fa61c1

Commits

Reviewing files that changed from the base of the PR and between 307bcb5 and 339b177.

Files selected for processing: 3

• apps/server/src/provider/Drivers/ClaudeHome.test.ts
• apps/server/src/provider/Drivers/ClaudeHome.ts
• apps/server/src/textGeneration/ClaudeTextGeneration.test.ts

Included review availability: This review used your included allowance. Your plan provides up to 10 included reviews per hour; 5 remain after this review.

[coderabbitai]

• Fix comments: Fix CodeRabbit comments on this PR

[coderabbitai]

• Fix comments: Fix CodeRabbit comments on this PR

[coderabbitai]

• Fix comments: Fix CodeRabbit comments on this PR

[coderabbitai]

• Fix comments: Fix CodeRabbit comments on this PR

[maria-rcks]

Note

Written by gpt-6.1-sol on behalf of Maria

Verified in a real dev environment

verified the launch behavior as-is at e9f63e5134086ad35ef703d2bf54e4b267f60dfd on a linux boat vm. no code changes or pushes were needed.

used separate main and pr checkouts, installed dependencies with vp i, and ran one background dev server at a time with isolated state. drove the actual web client using the repo's playwright and /usr/bin/google-chrome. created a project pointing at a real git repo through the app. all provider turns below used the real local proxy, with no mocks or replay provider.

• on github/main (3368dd70f2), an api-key claude haiku 5.5 turn running printenv CLAUDE_CODE_ENTRYPOINT returned t3-code.
• the same prompt on the pr returned sdk-ts. inspecting only that variable in the real claude child's process environment confirmed both values.
• an explicit instance override of t3-code remained effective with the api key.
• codex gpt-6.1-sol created real .claude/settings.local.json through the app. after removing the instance api key, a fresh claude session still completed with sdk-ts, without restarting the server.
• creating a new worktree through the app generated branch t3/check-claude-entrypoint through claude text generation. its cli child launched with the entrypoint unset. the real claude turn in that worktree also returned sdk-ts.
• with the api key restored on the instance, claude title generation succeeded. codex gpt-6.1-sol completed web turns on both main and the pr, returning the requested CODEX_MAIN_OK and CODEX_PR_OK responses.
• 395 tests passed through blacksmith in the five focused files: ClaudeHome, ClaudeSkills, ClaudeTextGeneration, ProviderRegistry, and ClaudeAdapterV2. ci reports 29 successful contexts, none pending or failing.

before, main with the api-key instance:

after, the same prompt on the pr:

recorded follow-up turn through the real claude provider:

short runtime transcript

the uploaded screenshots were inspected, and the uploaded 10.8-second video played to completion and supported seeking in t3's browser.

Not verified

actual subscriber-credit spending and the original Credit balance is too low response remain unverified. the proxy can run real claude turns but does not expose an anthropic console organization holding only subscriber api credits. the before/after proves the process-label change, not anthropic's billing decision.

native subscription login, windows environment-name handling, macos/electron shell behavior, native ios/android clients, managed settings, and direct remote/relay/tunnel connections were not exercised. haiku 4.5's short model alias was rejected by the proxy; the successful claude app checks used supported haiku 5.5.