Adds opt-in `ccsx auth import-default <name>` to migrate the existing
~/.codex/auth.json into a new profile, plus the cross-system integration
tests and user-facing documentation.
- import-default-command (C3 torn-write protection):
- readFileSync + JSON.parse with 3x retry / 100ms backoff to survive
Codex's truncate-then-write auth.json refresh race
- decode-id-token sanity-check on JWT shape (catches mid-write JWT
corruption that JSON.parse alone wouldn't notice)
- pgrep -f codex best-effort detection; warns + refuses without
--force-while-running flag if a live codex process is found
- rejects cliproxy-format auth files ({type: "codex", ...} wrapper)
with a clear "use ccs cliproxy ..." pointer
- atomic write to <dest>.tmp.<pid>.<rand> + rename
- --with-history defaults to false per D8 (auth-only is the safer
default; opt in for bulkier data)
- --force backs up existing auth.json to .bak-<ts> before overwrite
- non-destructive — never modifies ~/.codex/; legacy mode keeps
working without ever running this command
- integration tests:
- two-terminal-isolation: two profiles with separate CODEX_HOMEs
write to their own auth.json/history.jsonl with no crosstalk
- ccsxp-independence: codex-auth profile set; ccsxp still uses its
own CCSXP_CODEX_HOME / ~/.codex pool (H5 stderr notice present)
- legacy-fallback: no profiles registered → codex-runtime-router
leaves CODEX_HOME unset → codex falls back to ~/.codex
- import-default.integration: real fs copy + decode + register
- docs/codex-auth.md: user guide covering quick start, two-terminal
example, migration, dashboard, and caveats (cmd.exe, Windows
symlinks, ccsx vs ccsxp distinction)
155 codex-auth-scope tests green (45 Phase 1 + 57 Phase 2 + 19 Phase 3
+ 15 Phase 4 + 19 Phase 5). Full suite 3051/3082 — the 1 failure is a
pre-existing test-pollution issue between ccsxp-runtime.test.ts and
codex-runtime-integration.test.ts that exists on dev today; the test
passes in isolation.
Adds a read-only dashboard surface for the codex-auth profile registry.
Power users can see which profile is active and its decoded email/plan
without leaving the browser; mutations stay CLI-only.
- codex-auth-dashboard-service: builds the response shape (active +
default + profiles[]), reads each profile's auth.json, decodes id_token
via Phase 1 decoder (nested URI claims per C1), 5s in-memory single-key
cache plus exported invalidateCodexAuthProfilesCache() hook for
in-process Phase 2 callers (D7); strict field whitelist — id_token /
access_token / refresh_token NEVER appear in response body or logs
- GET /api/codex/profiles registered INSIDE the
requireLocalAccessWhenAuthDisabled middleware (H6) — emails are PII
and must not leak when dashboard is exposed remotely; integration
test asserts 403 from non-localhost origin
- NEW Auth Profiles tab (D5) in ui/src/pages/codex.tsx — distinct from
the existing codex-profiles-card (which edits config.toml [profiles],
a different concept); active profile + email + plan tier highlighted,
table of all profiles below, disabled Switch/Remove buttons redirect
to terminal commands
- accountId returned by API for power users (curl) but hidden from
the default UI (D6)
- 11 service unit tests + 4 endpoint integration tests, all green
Upgrades the previously-stub ccsx binary entry (src/bin/codex-runtime.ts)
into an argv router: `ccsx auth <cmd>` dispatches to the Phase 2 router;
any other argv resolves the active codex-auth profile and spawns codex
with CODEX_HOME pointed at the profile dir.
- resolve-active-profile: sync, hot-path-safe (<5ms), reads YAML
registry via Phase 1 helpers; precedence is CODEX_HOME (explicit)
> CCS_CODEX_PROFILE (env) > registry default > null (legacy
~/.codex fallback); fails open on any error (silent for missing
registry, stderr warn for corrupt/missing-profile)
- codex-runtime-router: extracted main() for testability; entry
script is a thin 3-line wrapper; returns -1 sentinel for the
CCS branch so the spawn lifecycle isn't terminated
- ccsxp-runtime: H5 defensive stderr notice when CCS_CODEX_PROFILE
is set, surfacing the boundary between codex-auth (native codex)
and ccsxp (cliproxy pool) without changing functional behavior;
CLIProxyAPI does not read CODEX_HOME so no pool contamination
possible
- 14 unit tests (8 resolver + 6 router); ccsxp regression suite
(5 tests) untouched and still green
Implements the user-facing surface for ccsx auth profile management.
After `ccsx auth create work` (auto-spawns codex login with CODEX_HOME
pinned per D11), users can `eval "$(ccsx auth use work)"` in any shell
to scope all subsequent codex invocations to that profile — letting
two terminals run two different Codex accounts concurrently.
- codex-auth-router: dispatches argv to subcommand handlers
- create: idempotent, --force re-links config.toml preserving auth.json
(D9), then auto-spawns codex login with CODEX_HOME pinned (D11);
filesystem ops happen before registry write to avoid registry orphans
on EACCES/ENOSPC
- login: standalone re-auth for an existing profile
- switch: persistent default in YAML registry
- use: STDOUT-DISCIPLINED — emits only shell-evalable exports;
bash/zsh/fish/PowerShell/cmd syntaxes via shell-detect; sets
CCS_NO_PRE_DISPATCH=1 at module load to suppress recovery/migration
banners that would otherwise contaminate eval (C2)
- show: list (active(missing) row at top per D14) + detail views
- remove: default-profile guard, active-shell warn, --yes / --force
- ASCII-only output, NO_COLOR honored, all errors to stderr via
exitWithError
- pre-dispatch.ts: early-return when CCS_NO_PRE_DISPATCH=1, placed
before autoMigrate which is itself a stdout writer
57 unit tests, all green. Help text cross-references the ccsxp/ccsx
distinction since the binaries differ by one character (H5).
Adds the storage substrate for ccsx auth profile isolation.
Each Codex profile gets its own CODEX_HOME dir under
~/.ccs/codex-instances/<name>/ with isolated auth.json and
history.jsonl; config.toml is shared via symlink to ~/.codex/config.toml
so two terminals can run two Codex accounts simultaneously without
duplicating user config.
- CodexProfileRegistry (YAML, atomic write tmp.<pid>.<rand> + rename,
orphan cleanup, full CRUD + default pointer)
- decode-id-token: pure base64 JWT decoder for OpenAI id_token,
reads nested https://api.openai.com/auth claims (chatgpt_plan_type,
chatgpt_account_id) and dual-path email
- ensureSharedConfigSymlink: self-healing, idempotent, overwrites
stale entries with stderr warning
- 45 unit tests, all green
Foundation only — no CLI, no runtime injection, no dashboard.
Subsequent commits wire those in.
* fix(ci): breaking-change-guard handles \${VAR:-default} compose image syntax (REV12)
The sed 's/:.*//' pattern truncated at the first colon in the
\${CCS_IMAGE:-ghcr.io/...} expression, yielding "\${CCS_IMAGE" as the
image name instead of the actual registry path. Any change to the default
image namespace was therefore undetectable.
Introduce extract_image_name() that first strips the \${VAR:-default}
wrapper with a sed -E expression, then strips only the trailing :tag
suffix using a pattern that preserves internal colons (e.g. registry:5000/
owner/repo). Applied to both OLD_RAW and NEW_RAW extraction paths.
* fix(ci): breaking-change-guard runs on ubuntu-latest to cover forked PRs (REV13)
The trusted-author gate (COLLABORATOR|MEMBER|OWNER) caused forked-PR
contributors to bypass the breaking-change check entirely. A forked
contributor could rename services.ccs or change the image namespace
without a feat!/fix! marker and the guard would never run.
This workflow is a documented exception to the self-hosted-first policy:
it performs ONLY pure YAML diff parsing (git show / awk / sed). No build,
install, or arbitrary PR-branch scripts are executed. The checkout uses
persist-credentials: false. There is no untrusted code execution, so
ubuntu-latest is safe and necessary for universal fork coverage.
Update self-hosted-runner-policy.test.ts to:
- Introduce GITHUB_HOSTED_RUNNER_EXCEPTIONS registry with required
justification comments for each entry
- Skip exception workflows in the "keeps active workflows on local runners"
and "gates pull-request workflows" assertions
- Add a new "documented exceptions use github-hosted runners" test that
verifies each exception entry actually uses a GitHub-hosted runner
(prevents stale entries accumulating without cleanup)
* fix(test): tighten compose-parity image-name match to exact expected values (REV7)
Replace loose substring grep (grep -q "ccs") with exact equality checks
against declared EXPECTED_CANONICAL_IMAGE and EXPECTED_INTEGRATED_IMAGE
constants. The integrated compose builds locally as ccs-cliproxy:latest
(not ghcr.io/kaitranntt/ccs), so both expected names are explicitly
documented at the top of the assertion block.
Fixes: a drift in integrated compose to a wrong owner/registry could
slip through the old "grep -q ccs" check; now any name other than the
declared constant is a hard failure.
* feat(ci): breaking-change-guard catches services.ccs rename — public DNS contract (REV8)
Docker's service-name DNS uses the compose service KEY as the hostname.
Sibling containers on ccs-net reach CCS via http://ccs:8317; renaming
services.ccs: to anything else silently breaks that contract even when
image name, network name, and container_name are unchanged.
Add check 4 to the guard:
- Extract top-level service keys from both base and HEAD versions of
docker/compose.yaml using awk (no external YAML parser required).
- Fail if the "ccs" key is absent from HEAD.
- Fail if the sorted set of service keys differs from base.
Wraps inside the existing git cat-file guard so new-file PRs skip it.
* fix(docker): parameterize image with CCS_IMAGE env so smoke-test exercises new digest
docker/compose.yaml hardcoded image: ghcr.io/kaitranntt/ccs:latest, causing
network-contract.sh to always use the old published image regardless of what
IMAGE_OVERRIDE was passed. Switching to \${CCS_IMAGE:-ghcr.io/kaitranntt/ccs:latest}
lets CI pass CCS_IMAGE=<digest> so the smoke-test actually exercises the
just-built image. Default end-user behaviour is unchanged.
Resolves REV4 (reviewer loop 2, PR #1261).
* fix(test): image-size.sh fails loudly on manifest inspection failure
The --platform branch previously exited 0 when imagetools inspect returned
empty or zero bytes, silently passing the budget check. A broken image
could reach production undetected if buildx had any inspection issue.
Changed to exit 1 with a clear diagnostic message explaining possible causes
(old buildx, manifest format mismatch, image not yet pushed). No --allow-inspect-failure
escape hatch is provided; CI must fix the root cause.
Resolves REV5 (reviewer loop 2, PR #1261).
* test(docker): cover --platform branch in image-size-logic tests
Prior tests only exercised the local docker image inspect path. The
--platform branch (imagetools inspect) had zero coverage, meaning the
REV5 silent-pass regression would not have been caught by CI.
Added 5 mock-based test cases for the --platform branch:
- pass when platform-scoped compressed size < budget
- fail when platform-scoped compressed size > budget
- fail (exit 1) when imagetools inspect errors (REV5 guard)
- fail (exit 1) when reported size is "0" (REV5 guard)
- fail (exit 1) when size output is empty (REV5 guard)
Mocks follow the existing pattern (override docker in PATH with a temp
wrapper), extended to handle buildx imagetools inspect subcommand.
Resolves REV6 (reviewer loop 2, PR #1261).
* fix(test): compose-parity image_name() handles \${VAR:-default} syntax
After parameterizing compose.yaml's image field with \${CCS_IMAGE:-...},
the image_name() sed pipeline cut at the first colon in the shell variable
syntax (:-) instead of the tag separator, returning '\${CCS_IMAGE' and
failing the kaitranntt/ccs grep.
Added a sed step to unwrap \${VAR:-default} by extracting just the default
value before stripping the tag. Existing plain image: name:tag references
are unaffected.
* fix(ci): breaking-change-guard skips missing base files (CI-1)
Guard each git-show call with git cat-file -e existence check before
attempting to read docker/compose.yaml from the base branch. When the
file doesn't exist on the base (new file in this PR), the script would
crash with "fatal: path exists on disk but not in origin/dev". New
files can't cause a contract regression, so we exit early with
breaking=0.
* style: prettier reformat unrelated drift (CI-2)
Two files had minor formatting drift from earlier umbrella PRs:
- src/cliproxy/quota/quota-manager.ts
- src/management/checks/image-analysis-check.ts
No logic changes — formatter-only pass to unblock CI format:check.
* fix(test): update trusted-author gate count to 4 in ci-workflow test (CI-3)
PR #1260 added a compose-parity job to ci.yml. That job runs on
self-hosted runners using PR-provided checkout, so it legitimately
requires the trusted-author guard (same as validate, build, test).
The test expected 3 occurrences; the correct count is now 4:
validate (matrix), build, test, compose-parity.
* fix(ci): smoke-test passes compose path + image ref to network-contract (REV-1)
network-contract.sh signature is: <compose-file> [image-ref]
The smoke-test job was calling it as:
bash tests/docker/network-contract.sh "${{ steps.image.outputs.ref }}"
which placed the image ref in the compose-file position ($1), causing
the script to try docker compose -f <image-ref> which fails.
Corrected to:
bash tests/docker/network-contract.sh docker/compose.yaml "${{ steps.image.outputs.ref }}"
Also removes publish-dashboard from smoke-test.needs (REV-2): when
publish-dashboard is SKIPPED on prerelease events, GitHub Actions
propagates the skip to downstream jobs, so smoke-test and
promote-mutable-tags were silently skipped on every rc.N publish.
smoke-test only verifies the integrated image; it has no dependency
on the legacy dashboard image job.
* docs(docker): annotate /root/.ccs path in compose volume (REV-3 clarification)
The reviewer raised a concern that the compose volume mounts /root/.ccs
but the entrypoint might default to /home/node/.ccs. This is a false
positive: the integrated image uses entrypoint-integrated.sh (not
entrypoint.sh), which runs under supervisord with user=root and
explicitly mkdir -p /root/.ccs. HOME is /root inside the container.
The volume mount at /root/.ccs is correct.
Added an inline comment documenting the reasoning so future reviewers
do not confuse entrypoint.sh (legacy dashboard image) with
entrypoint-integrated.sh (integrated image).
* fix(ci): gate docs-parity pull_request job to trusted authors
docs-parity.yml runs on a self-hosted runner and checks out PR code.
The self-hosted-runner-policy test requires any such workflow to include
the trusted-author guard. The workflow was missing the guard, causing
bun test:fast to fail with 1 failure.
Allow push events (no author check needed — push is to own branch)
and trusted-contributor PRs only.
* docs: hoist Docker zero-install quickstart above npm install path
- Create docs/quickstart-snippet.md as canonical source for the
two-command flow (curl + docker compose up -d), wrapped in
<!-- quickstart-snippet-start/end --> markers
- Hoist the snippet into README.md immediately below the deprecation
banner, above all other install paths
- Rename old npm-only "## Quick Start" to "## Install on Host (npm)"
and move it below the Docker quickstart
* docs(docker): restructure README with zero-install first and migration section
- Reorder top-level sections: zero-install (canonical snippet with
markers), choosing an image, power-user ccs docker, prebuilt image,
connect your app to CLIProxy, migration, env vars, troubleshooting
- Add deprecation banner at the top pointing at the migration section
- Add ## Migration from ccs-dashboard:latest section with step-by-step
instructions covering compose down, data preservation, named volume
vs bind-mount path, and compose up with the new image
- Keep P1's Choosing an image table and P5's Connect Your App to
CLIProxy section intact, just repositioned
* test(docs): parity check for quickstart snippet across README files
Assert README.md and docker/README.md both contain the canonical
quickstart block verbatim, anchored by marker comments. Exits non-zero
and prints a diff on any drift.
* ci(docs): wire quickstart-parity test on push and PR
Runs tests/docs/quickstart-parity.sh on self-hosted runner whenever
docs/quickstart-snippet.md, README.md, docker/README.md, or the
test/workflow files themselves change. Fails fast on snippet drift.
Add "Connect your app to CLIProxy" section to docker/README.md with:
- Public contract table (network=ccs-net, service DNS=ccs, ports 8317/3000)
- Pattern A: same compose file with external network reference
- Pattern B: docker run --network ccs-net
- Troubleshooting subsection (DNS, missing network, conflict, Podman, MTU)
Add one-line link in README.md pointing to the new section.
Add CHANGELOG entry under Unreleased noting the contract as SemVer-major stable.
Add CONTRIBUTING.md note that changing services.ccs or networks.ccs-net requires major bump.
* feat(docker): parameterize Dockerfile.integrated with ARG FLAVOR=minimal|full
Add FLAVOR build arg that gates the AI CLI install layer (claude-code,
gemini-cli, grok-cli, opencode) so one Dockerfile produces both the
minimal (< 350 MB) and full (< 600 MB) integrated images.
Use BuildKit cache mount for /root/.npm to speed up repeated builds.
Part of #1251 (P1 — publish integrated images).
* feat(docker): emit startup deprecation warning in legacy ccs-dashboard entrypoint
Prepend a [WARN] line to stderr on every container start so operators
running ghcr.io/kaitranntt/ccs-dashboard:latest are notified to migrate
to ghcr.io/kaitranntt/ccs:latest. Sunset window: 2 releases. See #1251.
* test(docker): add image-size.sh budget assertion + unit test suite
image-size.sh: asserts docker image inspect .Size against a byte budget.
Usage: image-size.sh <image:tag> <max-bytes>
Budgets: minimal=350 MB (367001600), full=600 MB (629145600)
image-size-logic.test.sh: 6 mock-docker unit tests covering pass/fail
boundaries, bad arg count, and non-integer input. All 6 pass locally.
Part of #1251 (P1).
* ci(docker): extend docker-release.yml to publish ccs:latest and ccs:full
- Add publish-integrated job with matrix flavor: [minimal, full]
- minimal -> ghcr.io/kaitranntt/ccs:<ver> + :latest (when promoted)
- full -> ghcr.io/kaitranntt/ccs:full-<ver> + :full (when promoted)
- Multi-arch: linux/amd64 + linux/arm64 via buildx
- Scoped GHA cache per flavor to avoid cross-contamination
- Add promote_to_latest workflow_dispatch input (default false)
- rc.1 soak: first publish only pushes immutable version tag
- Mutable :latest/:full promoted only on release event OR explicit opt-in
- Add smoke-test job (post-publish) for each flavor
- Pulls the just-published version tag
- Asserts image size via tests/docker/image-size.sh
- Boots container, waits for healthcheck, probes :3000 and :8317
- Keep publish-dashboard job unchanged (legacy 2-release sunset)
- Updated labels to note deprecation status
All jobs run on self-hosted cliproxy runners. Part of #1251 (P1).
* docs(docker): document new image tags, add ccs-dashboard deprecation notices
docker/README.md:
- Add "Choosing an image" table (ccs:latest / ccs:full / ccs-dashboard deprecated)
- Update Quick Start section to use ccs:latest as primary example
- Add legacy image note with sunset timeline
CHANGELOG.md:
- Add Unreleased > ### Deprecated entry for ccs-dashboard:latest
pointing to migration path and #1251
README.md:
- Add one-line deprecation banner near top routing users to ccs:latest
Part of #1251 (P1).
Squash merge PR #1240.\n\nValidated locally from the PR head because trusted-author GitHub CI was skipped for this fork PR:\n- bun run validate\n- bun run build:all\n- bun run test:all\n- CCS_E2E_SKIP_BUILD=1 bun run test:e2e
Closes#1218.
When the user invokes `ccs <profile> agents` (and other Claude subcommands
like `mcp`, `doctor`, `plugin`, ...), CCS was unconditionally injecting
session-only Claude flags (`--append-system-prompt`, `--disallowedTools`,
`--settings`, official-channels plugin specs) and forwarding the
`DISABLE_TELEMETRY=1` env var from profile settings. Each of those
either errored out the subcommand (`error: unknown option
'--append-system-prompt'`) or silently flipped Claude into
non-interactive list mode, so the new `claude agents` agent view never
opened under CCS.
Add a small `claude-subcommand-detector` that recognizes the documented
Claude subcommand set after skipping known value-taking flags, then have
the three steering-prompt injectors (websearch, image-analysis, browser),
the cliproxy and settings launchers, and the official-channels plan
short-circuit when a subcommand invocation is detected. Also strip
`DISABLE_TELEMETRY` from the spawned env only for subcommand
invocations — upstream Claude Code uses that var as a kill switch for
the subcommand TUIs, and the user's telemetry preference still applies
to every normal interactive session.
Verified end-to-end against `ccs glm agents` and `ccs ck agents`: the
agent view opens correctly, no `unknown option` error, and all 1938
existing tests pass.
Keep active CCS workflows on self-hosted runners, gate self-hosted PR execution to trusted authors, and scope privileged release credentials to the exact git operations that need them.
Two pre-existing test bugs surfaced under CCS CI on the new claw self-
hosted runners but stayed hidden locally:
1. tests/unit/commands/cliproxy-routing-subcommand.test.ts mocks the
routing-subcommand module with only 3 exports, but
src/commands/cliproxy/index.ts statically imports 6. Bun resolves
the static import graph against the mock and reports
SyntaxError: Export named 'handleRoutingAffinitySet' not found
for every test in the file. Add the missing 3 exports
(handleRoutingAffinityStatus, handleRoutingAffinityHelp,
handleRoutingAffinitySet) to the mock and document why the mock
must mirror every named export of the target module.
2. tests/unit/hooks/websearch-transformer.test.ts builds the
"disallowed" trace path from process.cwd() and from os.homedir().
Both fall under the os.tmpdir() safe-prefix in two real
environments: CI runners with cwd == /tmp/runner/work/... and Bun
test isolation that re-roots HOME under tmpdir. The hook treats
them as safe, writes the trace, and the assertion that the file
does NOT exist fails. Anchor disallowedTracePath under /etc/...
instead so it cannot satisfy the tmpdir, /var/log, or
<CCS_HOME>/.ccs/logs prefixes in any host environment.
Both fixes are independent of the OAuth callback traceability change
that this branch otherwise carries, but ship together so the PR
clears CI on the new runner stack.
This reverts the runs-on change for ai-review.yml. All CI/CD must
stay on home infra (claw, docker LXC, etc); GitHub-hosted runners
are forbidden. The qodo-ai/pr-agent docker action mount issue will
be addressed by either configuring the self-hosted runner image to
use path-identical bind mounts so nested-docker volume sources
resolve correctly, or by switching to native CLI invocation
(pip install pr-agent) instead of the docker action.
Tracked separately; not blocking PR #1209.
The qodo-ai/pr-agent Docker action requires the GitHub event payload
at /github/workflow/event.json inside the action container. On
self-hosted runners that themselves run inside a Docker container
(e.g. myoung34/github-runner), the host docker daemon resolves the
volume mount /tmp/runner/work/_temp/_github_workflow:/github/workflow
against the host filesystem, where the runner's /tmp path does not
exist. The action container starts with an empty /github/workflow
mount and fails with FileNotFoundError on event.json.
AI review jobs do not need self-hosted runner access (no bun cache,
no internal infra). Switch both dispatch-review and pr-agent jobs
to ubuntu-latest so the volume mount resolves on the same host where
the action expects it.
* fix(dispatcher): preserve anthropic auth env for settings profiles on non-proxy path
API profiles whose `ANTHROPIC_BASE_URL` is classified as `'anthropic'`
(anthropic.com, paths containing `/anthropic`, ollama.com) skip the
local OpenAI-compat proxy. The non-proxy launch path stripped
`ANTHROPIC_BASE_URL` / `ANTHROPIC_AUTH_TOKEN` / `ANTHROPIC_API_KEY`
from the subprocess env without re-injecting them, so Claude Code
launched with no routing/auth in `process.env` and failed with
`Not logged in - Please run /login`. The `--settings` env block does
not satisfy Claude Code's auth check.
Pre-existing for `anthropic.com` and `/anthropic` profiles. Newly
broken in v7.77.0 for Ollama Cloud profiles - PR #1175 reclassified
`ollama.com` from `generic-chat-completion-api` to `anthropic`,
moving it from the proxy path onto this broken non-proxy path.
Fix by extending `stripAnthropicRoutingEnv` with an optional
`preserveFrom` parameter. Routing keys present in `preserveFrom`
survive the strip (with values from `preserveFrom`). Settings-type
profiles pass their own `settings.env` as the preserve source so
routing they explicitly supplied is kept while routing leaked from
the parent shell or `global.env` is dropped.
Wired into both call sites:
- `headless-executor.ts` (the `-p` headless executor)
- `settings-flow.ts` (the interactive flow, which then calls
`execClaude` - whose own `stripAnthropicRoutingEnv` pass on the
merged env now also takes `envVars` as the preserve source so the
caller-supplied routing is not stripped a second time before spawn)
Native Anthropic / Bedrock / Vertex profiles are unaffected (they
don't put routing keys in `settings.env`). The OpenAI-compat proxy
path is unaffected because `buildOpenAICompatProxyEnv` overrides
`BASE_URL` / `AUTH_TOKEN` with localhost values and explicitly deletes
`API_KEY` after this strip.
Replaces the Apr 21 defensive double-strip test (which was the
mechanism causing this regression on the interactive path) with a
test asserting the new contract: caller-supplied routing in `envVars`
survives, parent-process routing is still stripped.
* test: harden anthropic settings env preservation coverage
* fix: preserve explicit blank anthropic routing env
---------
Co-authored-by: Tam Nhu Tran <kaitran.ntt@gmail.com>
Tam Nhu Tran