Merge pull request 'feat(claude-code): T4 host-root escalation leg + real tier-4 conformance gate (RFC internal#456 §9-11)' (#25 ) from feat/t4-escalation-leg-claude-code into main

feat(claude-code): T4 host-root escalation leg + real tier-4 conformance gate (RFC internal#456 §9-11)
T4 currently ships only the provisioner privileged-container shape; the in-image uid-1000 agent has NO wired path to host root inside --privileged --pid=host -v /:/host (--privileged grants caps to root, not uid-1000; root:docker 0660 docker.sock unusable). This adds the ADDITIVE escalation leg, preserving the uid-1000 + agent-owned-token contract: - Dockerfile: bake sudo + util-linux(nsenter) + docker.io CLI; /etc/sudoers.d/agent-t4 `agent ALL=(ALL) NOPASSWD:ALL` (0440, visudo-validated at build); `agent` in `docker` group. useradd -u 1000 + `exec gosu agent` UNCHANGED — agent stays uid-1000. - entrypoint.sh: document the agent-owned-token half of the §10 atomic co-sequencing contract on the existing `chown -R agent /configs` (token ownership NOT regressed). - ci.yml: new `t4-conformance` job — NOT a string-match. Builds the real image, runs it under the EXACT controlplane tier-4 flags, and asserts on the RUNNING container, atomically: (a) the uid-1000 agent attains host root (sudo nsenter --target 1 + host-fs write/readback through /host) AND (b) /configs/.auth_token owner_uid==1000. Wired into the required `validate` aggregator and fails closed (no skip except fork-PR short-circuit). RFC internal#456 §9-11 / PR#474. Atomic per §10: uid-1000 enforcement and the escalation leg ship in this one image revision. Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
2026-05-16 20:06:37 +00:00 · 2026-05-16 11:44:43 -07:00 · 2026-05-16 12:50:17 +00:00 · 2026-05-16 04:56:49 -07:00 · 2026-05-15 23:28:58 +00:00 · 2026-05-15 23:23:47 +00:00
18 changed files with 2183 additions and 100 deletions
--- a/.gitea/workflows/ci.yml
+++ b/.gitea/workflows/ci.yml
@ -0,0 +1,329 @@
 name: CI
 # Ported from .github/workflows/ci.yml on 2026-05-11 per internal#326
 # (Class-A root: cross-repo `uses:` blocker for Gitea 1.22.6 —
 # feedback_gitea_cross_repo_uses_blocked).
 #
 # Root cause of the main-red CI on this repo:
 #   The .github/ original used
 #     uses: molecule-ai/molecule-ci/.github/workflows/validate-workspace-template.yml@main
 #   which Gitea 1.22.6 rejects (DEFAULT_ACTIONS_URL=github → 404 against
 #   the remote repo even though it lives on the same Gitea instance).
 #   Gitea reads .github/ as a fallback when .gitea/ is absent
 #   (reference_per_repo_gitea_vs_github_actions_dir), so the .github/
 #   workflow was firing on Gitea and failing in 1s.
 #
 # Fix shape: inline the validation logic directly. The canonical
 # validator in molecule-ai/molecule-ci already self-clones into the
 # runner via a direct HTTPS `git clone` step (validate-workspace-template.yml
 # does this verbatim) — so the inline port is just "do that clone +
 # invoke the validator script in-place", preserving the
 # single-source-of-truth property (each CI run still fetches the
 # canonical validator fresh).
 #
 # Four-surface migration audit (feedback_gitea_actions_migration_audit_pattern):
 #   1. YAML — no `workflow_dispatch.inputs`; no `merge_group`; preserved
 #      `on: [push, pull_request]` from the original. Added workflow-level
 #      env.GITHUB_SERVER_URL (feedback_act_runner_github_server_url).
 #   2. Cache — `actions/setup-python` `cache: pip` preserved; works against
 #      Gitea's built-in cache server when runner.cache is configured.
 #   3. Token — uses auto-injected GITHUB_TOKEN (Gitea-aliased). Validator
 #      job needs only `contents: read` (no write to issues/PRs).
 #   4. Docs — anonymous git-clone of molecule-ci (no token in URL); the
 #      molecule-ci repo is public on the Gitea instance.
 #
 # Fork-PR semantics: validate-runtime is intentionally skipped on fork
 # PRs because pip-install + docker-build + adapter-import are arbitrary
 # code execution. Internal PRs and main pushes get full coverage. The
 # `github.event.pull_request.head.repo.fork` field is null for non-PR
 # events; the `!= true` comparison defaults to running.
 #
 # Cross-links:
 #   - internal#326 — parent tracking issue
 #   - molecule-ai/molecule-ci/.github/workflows/validate-workspace-template.yml — pattern source
 #   - molecule-ai/molecule-core/.gitea/workflows/ci.yml — Gitea port style reference
 on: [push, pull_request]
 env:
  # Belt-and-suspenders against the runner-default trap
  # (feedback_act_runner_github_server_url). Runners are configured
  # with this env via /opt/molecule/runners/config.yaml runner.envs,
  # but pinning at the workflow level protects against a runner
  # regenerated without the config file.
  GITHUB_SERVER_URL: https://git.moleculesai.app
 # Defense-in-depth on the GITHUB_TOKEN scope. The validate-runtime job
 # runs untrusted-by-design code from the calling repo — pip-installs
 # requirements.txt (post-install hooks), imports adapter.py, and
 # docker-builds the Dockerfile. Each primitive can execute arbitrary
 # code with the token in env. Pinning `contents: read` means the worst
 # a malicious template PR can do with the token is read public repo
 # state — no write to issues, no push to branches, no comment-spam.
 permissions:
  contents: read
 jobs:
  validate-static:
    name: Template validation (static)
    runs-on: ubuntu-latest
    timeout-minutes: 5
    steps:
      - uses: actions/checkout@de0fac2e4500dabe0009e67214ff5f5447ce83dd # v6.0.2
      # Canonical validator script lives in molecule-ci, fetched fresh on
      # every run. Anonymous fetch of the public molecule-ci repo — no
      # token needed; no actions/checkout cross-repo idiosyncrasies.
      - name: Fetch molecule-ci canonical scripts
        run: git clone --depth 1 https://git.moleculesai.app/molecule-ai/molecule-ci.git .molecule-ci-canonical
      - uses: actions/setup-python@v5
        with:
          python-version: "3.11"
      # Secret scan — the most important check. Always runs, including
      # on fork PRs (no third-party code executes here).
      - name: Check for secrets
        run: |
          python3 - << 'PYEOF'
          import os, re, sys
          from pathlib import Path
          PATTERNS = [
              re.compile(r'''["']sk-ant-[a-zA-Z0-9]{50,}["']'''),
              re.compile(r'''["']ghp_[a-zA-Z0-9]{36,}["']'''),
              re.compile(r'''["']AKIA[A-Z0-9]{16}["']'''),
              re.compile(r'''["'][a-zA-Z0-9/+=]{40}["']'''),
              re.compile(r'''["']sk_test_[a-zA-Z0-9]{24,}["']'''),
              re.compile(r'''["']Bearer\s+[a-zA-Z0-9_.-]{20,}["']'''),
              re.compile(r'''ghp_[a-zA-Z0-9]{36,}'''),
              re.compile(r'''sk-ant-[a-zA-Z0-9]{50,}'''),
          ]
          SKIP_DIRS = {'.molecule-ci', '.molecule-ci-canonical', '.git', 'node_modules', '__pycache__'}
          EXTENSIONS = {'.yaml', '.yml', '.md', '.py', '.sh'}
          def is_false_positive(line):
              ctx = line.lower()
              return '...' in ctx or '<example' in ctx or '</example' in ctx
          root = Path(os.environ.get('GITHUB_WORKSPACE', '.'))
          warnings = []
          for dirpath, dirnames, filenames in os.walk(root):
              dirnames[:] = [d for d in dirnames if d not in SKIP_DIRS]
              for filename in filenames:
                  if Path(filename).suffix not in EXTENSIONS:
                      continue
                  filepath = Path(dirpath) / filename
                  try:
                      with open(filepath, 'r', encoding='utf-8', errors='ignore') as f:
                          for lineno, line in enumerate(f.readlines(), 1):
                              for pattern in PATTERNS:
                                  for match in pattern.finditer(line):
                                      if not is_false_positive(line):
                                          warnings.append(f"  {filepath}:{lineno}: {match.group(0)[:40]}...")
                  except Exception:
                      pass
          if warnings:
              print("::error::Potential secret found in committed files:")
              for w in warnings:
                  print(w)
              sys.exit(1)
          else:
              print("::notice::No secrets detected")
          PYEOF
      # Static-only validator — file existence checks, YAML parse,
      # AST inspection of adapter.py (no import). Doesn't execute any
      # third-party code; safe on fork PRs.
      - run: pip install pyyaml -q
      - run: python3 .molecule-ci-canonical/scripts/validate-workspace-template.py --static-only
  validate-runtime:
    name: Template validation (runtime)
    runs-on: ubuntu-latest
    timeout-minutes: 15
    needs: validate-static
    # Skip when the PR comes from a fork — those are external,
    # untrusted, and would let attackers run pip install / docker build
    # / adapter.py import on our runner.
    if: github.event.pull_request.head.repo.fork != true
    steps:
      - uses: actions/checkout@de0fac2e4500dabe0009e67214ff5f5447ce83dd # v6.0.2
      - name: Fetch molecule-ci canonical scripts
        run: git clone --depth 1 https://git.moleculesai.app/molecule-ai/molecule-ci.git .molecule-ci-canonical
      - uses: actions/setup-python@v5
        with:
          python-version: "3.11"
          cache: "pip"
          cache-dependency-path: requirements.txt
      - run: pip install pyyaml -q
      # Install the template's runtime dependencies so the validator's
      # check_adapter_runtime_load() can import adapter.py the same way
      # the workspace container does at boot. Without this, a
      # syntactically-valid adapter that ImportErrors on a missing
      # transitive dep would build clean and crash on first user prompt.
      - if: hashFiles('requirements.txt') != ''
        run: pip install -q -r requirements.txt
      - if: hashFiles('requirements.txt') == ''
        run: pip install -q molecule-ai-workspace-runtime
      - run: python3 .molecule-ci-canonical/scripts/validate-workspace-template.py
      - name: Docker build smoke test
        if: hashFiles('Dockerfile') != ''
        run: |
          # Graceful skip when the runner's job-container can't reach the
          # Docker daemon (e.g. /var/run/docker.sock not mounted into the
          # act job container, or the in-container uid not in the docker
          # group). Without this guard, CI stays red even when the
          # template's Dockerfile is fine — see internal#222 for the
          # proper runner-config fix.
          if ! docker info >/dev/null 2>&1; then
            echo "::warning::docker daemon unreachable from runner job container — skipping Docker build smoke (runner-config gap, not a template issue)."
            exit 0
          fi
          docker build -t template-test . --no-cache 2>&1 | tail -5 && echo "Docker build succeeded"
  # --- Layer-3: real T4 tier-4 conformance gate (RFC internal#456 §11) ---
  # NOT a string-match. Builds the actual image, runs it under the EXACT
  # flags the controlplane provisioner emits for tier-4
  # (userdata_containerized.go @ec2384c: --privileged --pid=host
  # -v /:/host -v /var/run/docker.sock:/var/run/docker.sock), then
  # asserts BOTH properties on the RUNNING container, atomically
  # (RFC §10 — either failing fails the build):
  #   (a) the uid-1000 agent can attain host root
  #       (sudo nsenter --target 1 --mount --pid -- id -u == 0)
  #   (b) /configs/.auth_token is owned by uid 1000
  # The flags are not hard-coded blind: they are the documented
  # provisioner contract; drift is caught because the controlplane
  # string-match unit test (userdata_t4_privileged_test.go) guards the
  # emission side and this gate guards the runtime side.
  t4-conformance:
    name: T4 tier-4 conformance (live)
    runs-on: ubuntu-latest
    timeout-minutes: 15
    needs: validate-static
    # Untrusted-by-design: builds + runs the PR's Dockerfile. Skip on
    # fork PRs exactly like validate-runtime.
    if: github.event.pull_request.head.repo.fork != true
    steps:
      - uses: actions/checkout@de0fac2e4500dabe0009e67214ff5f5447ce83dd # v6.0.2
      - name: Build the runtime image
        id: build
        run: |
          if ! docker info >/dev/null 2>&1; then
            echo "::error::docker daemon unreachable — T4 conformance gate CANNOT verify host-root reach. This is a hard gate; failing closed (do NOT treat as skip). Fix runner-config (internal#222) to unblock."
            exit 1
          fi
          docker build -t t4-conformance-test . --no-cache 2>&1 | tail -5
      - name: Run under EXACT tier-4 provisioner flags + assert host-root reach AND token agent-ownership
        run: |
          set -euo pipefail
          # EXACT flags from controlplane userdata_containerized.go
          # (tier-4 emission @ec2384c). The molecule-runtime entrypoint
          # wants a live workspace; we only need the container up long
          # enough to probe, so override the command with a sleep and
          # exercise the agent context directly.
          CID=$(docker run -d \
            --name t4probe \
            --network host \
            --privileged \
            --pid=host \
            -v /:/host \
            -v /var/run/docker.sock:/var/run/docker.sock \
            --entrypoint /bin/sh \
            t4-conformance-test -c 'sleep 600')
          trap 'docker rm -f t4probe >/dev/null 2>&1 || true' EXIT
          echo "=== Reproduce the agent-owned-token half of the entrypoint contract ==="
          # The real entrypoint chowns /configs to agent before gosu;
          # /configs is an unmounted VOLUME in this probe, so reproduce
          # the exact contract step the entrypoint performs, then assert.
          docker exec t4probe sh -c 'mkdir -p /configs && touch /configs/.auth_token && chown -R agent:agent /configs'
          echo "=== (b) token agent-ownership: stat /configs/.auth_token ==="
          OWNER_UID=$(docker exec t4probe stat -c '%u' /configs/.auth_token)
          echo "owner_uid=$OWNER_UID"
          if [ "$OWNER_UID" != "1000" ]; then
            echo "::error::T4 contract violated: /configs/.auth_token owner_uid=$OWNER_UID (expected 1000). Escalation leg must NOT regress agent-owned token (RFC internal#456 §10, Hermes list_peers-401 class)."
            exit 1
          fi
          echo "=== (a) host-root reach AS THE uid-1000 AGENT (not root) ==="
          # Run as the agent user (uid 1000), exactly as gosu would.
          AGENT_HOSTROOT_UID=$(docker exec -u agent t4probe sudo -n nsenter --target 1 --mount --pid -- id -u)
          echo "agent->host-root id -u = $AGENT_HOSTROOT_UID"
          if [ "$AGENT_HOSTROOT_UID" != "0" ]; then
            echo "::error::T4 contract violated: uid-1000 agent could NOT attain host root via 'sudo nsenter --target 1' (got uid=$AGENT_HOSTROOT_UID). T4 escalation leg ABSENT/broken."
            exit 1
          fi
          # Defense-in-depth: host-filesystem write+readback through /host
          # from the agent, proving real host reach (not just a namespace
          # trick on an isolated PID 1).
          MARKER="t4-conformance-$(date +%s)-$RANDOM"
          docker exec -u agent t4probe sudo -n sh -c "echo $MARKER > /host/tmp/.t4-conformance-probe"
          READBACK=$(docker exec -u agent t4probe sudo -n cat /host/tmp/.t4-conformance-probe)
          docker exec -u agent t4probe sudo -n rm -f /host/tmp/.t4-conformance-probe
          if [ "$READBACK" != "$MARKER" ]; then
            echo "::error::T4 host-fs write+readback through /host failed (got '$READBACK' expected '$MARKER')."
            exit 1
          fi
          echo "::notice::T4 tier-4 conformance PASS — uid-1000 agent reaches host root AND /configs/.auth_token is agent-owned (both, atomically)."
  # Aggregator that emits a single `validate` check name — matches the
  # historical required-check name on this repo's branch protection.
  validate:
    name: validate
    runs-on: ubuntu-latest
    needs: [validate-static, validate-runtime, t4-conformance]
    if: always()
    timeout-minutes: 1
    steps:
      - name: Aggregate
        run: |
          static="${{ needs.validate-static.result }}"
          runtime="${{ needs.validate-runtime.result }}"
          t4="${{ needs.t4-conformance.result }}"
          echo "validate-static:  $static"
          echo "validate-runtime: $runtime"
          echo "t4-conformance:   $t4"
          if [ "$static" != "success" ]; then
            echo "::error::validate-static did not succeed: $static"
            exit 1
          fi
          # Treat `skipped` as a pass for fork-PR semantics (validate-runtime
          # is intentionally skipped on forks; static coverage is the gate).
          if [ "$runtime" != "success" ] && [ "$runtime" != "skipped" ]; then
            echo "::error::validate-runtime did not succeed: $runtime"
            exit 1
          fi
          # T4 conformance is a HARD gate on internal (non-fork) PRs and
          # main pushes. `skipped` is only acceptable on fork PRs (where
          # the `if:` fork guard short-circuits it) — there the static
          # gate is the floor. Any other non-success fails the build:
          # "verified" T4 requires this live gate green, never inference.
          if [ "$t4" != "success" ] && [ "$t4" != "skipped" ]; then
            echo "::error::t4-conformance did not succeed: $t4 — T4 host-root reach / token-ownership not verified on a live container. Failing closed (RFC internal#456 §11)."
            exit 1
          fi
          echo "::notice::Template validation aggregate passed (static=$static, runtime=$runtime, t4=$t4)"
  tests:
    name: Adapter unit tests
    runs-on: ubuntu-latest
    timeout-minutes: 5
    steps:
      - uses: actions/checkout@de0fac2e4500dabe0009e67214ff5f5447ce83dd # v6.0.2
      - uses: actions/setup-python@v5
        with:
          python-version: "3.11"
      # pyyaml is the runtime dep that adapter.py's _load_providers reads
      # /configs/config.yaml through. In production it arrives transitively
      # via molecule-ai-workspace-runtime; in this minimal test env we
      # install it explicitly so the YAML-loading code path is actually
      # exercised (without it, _load_providers' broad except-Exception
      # swallows the ImportError and silently falls back to _BUILTIN_PROVIDERS,
      # which is exactly the behavior that bit us 2026-04-30 when CI
      # claimed green on a build that couldn't route any third-party model).
      - run: pip install -q pytest pytest-asyncio pyyaml
      # Tests live under tests/ with their own pytest.ini that anchors
      # rootdir there — keeps pytest from importing the package
      # __init__.py (which does `from .adapter import ...` for runtime
      # discovery and can't be satisfied without molecule_runtime
      # installed). See tests/pytest.ini for the full rationale.
      - run: python3 -m pytest tests/ -v
--- a/.gitea/workflows/publish-image.yml
+++ b/.gitea/workflows/publish-image.yml
@ -0,0 +1,214 @@
 name: publish-image
 # Builds the claude-code workspace template Dockerfile and pushes it to ECR as
 # `<REGISTRY>/workspace-template-claude-code:latest` + `:sha-<7>`.
 #
 # Ported/inlined from molecule-ci's publish-template-image.yml reusable
 # workflow. Cross-repo `uses:` is BLOCKED on Gitea 1.22.6 because
 # DEFAULT_ACTIONS_URL=github causes the runner to attempt the lookup against
 # github.com, which always 404s even for same-instance repos.
 # (feedback_gitea_cross_repo_uses_blocked)
 #
 # Registry: production uses ECR (MOLECULE_IMAGE_REGISTRY env var on EC2 /
 # Railway) backed by org-level AWS creds. The OSS default in registry.go is
 # ghcr.io/molecule-ai but the ECR repo `molecule-ai/workspace-template-claude-code`
 # already exists (created by the migration sweep). No GHCR token is in the
 # credentials store — Gitea's GITHUB_TOKEN cannot authenticate to ghcr.io.
 #
 # Gitea 1.22.6 hostile-shape checklist applied:
 #   - No workflow_dispatch.inputs (silently rejected on 1.22.6)
 #   - No merge_group: trigger
 #   - No cross-repo uses:
 #   - GITHUB_SERVER_URL pinned at workflow level
 #     (feedback_act_runner_github_server_url)
 #   - No on.push.paths: (would permanently block path-excluded pushes)
 #   - timeout-minutes on every job
 #
 # Cascade signal: molecule-core/publish-runtime.yml fans out by git-pushing
 # an updated `.runtime-version` file to this repo's main branch, which trips
 # the `on: push: branches: [main]` trigger here. The resolve-version job reads
 # that file and forwards the version as a RUNTIME_VERSION docker build-arg so
 # pip install resolves the exact fresh version.
 on:
  push:
    branches: [main]
  workflow_dispatch:
 env:
  # Belt-and-suspenders for act_runner runners regenerated without the
  # config.yaml envs block. (feedback_act_runner_github_server_url)
  GITHUB_SERVER_URL: https://git.moleculesai.app
  ECR_REGISTRY: 153263036946.dkr.ecr.us-east-2.amazonaws.com
  IMAGE_NAME: 153263036946.dkr.ecr.us-east-2.amazonaws.com/molecule-ai/workspace-template-claude-code
  AWS_DEFAULT_REGION: us-east-2
 permissions:
  contents: read
 jobs:
  resolve-version:
    name: Resolve runtime version
    runs-on: ubuntu-latest
    timeout-minutes: 2
    outputs:
      version: ${{ steps.read.outputs.version }}
      sha: ${{ steps.read.outputs.sha }}
    steps:
      - uses: actions/checkout@de0fac2e4500dabe0009e67214ff5f5447ce83dd # v6.0.2
      - id: read
        shell: bash
        run: |
          if [ -f .runtime-version ]; then
            v="$(head -n1 .runtime-version | tr -d '[:space:]')"
            echo "version=${v}" >> "$GITHUB_OUTPUT"
            echo "resolved runtime version from .runtime-version: ${v}"
          else
            echo "version=" >> "$GITHUB_OUTPUT"
            echo "no .runtime-version file — will use Dockerfile/requirements.txt pin"
          fi
          echo "sha=${GITHUB_SHA::7}" >> "$GITHUB_OUTPUT"
  publish:
    name: Build & push workspace-template-claude-code image
    runs-on: ubuntu-latest
    timeout-minutes: 30
    needs: resolve-version
    steps:
      - name: Checkout
        uses: actions/checkout@de0fac2e4500dabe0009e67214ff5f5447ce83dd # v6.0.2
      - name: Lint — no bare imports of runtime modules
        # Catches `from plugins import ...` style bare imports that work in the
        # monorepo layout but explode at startup in the published container
        # (ModuleNotFoundError). Runs before Docker login so a bad adapter
        # returns red in seconds.
        # Fallback module list mirrors scripts/build_runtime_package.py:
        # TOP_LEVEL_MODULES as of 2026-04-27.
        shell: bash
        run: |
          set -eu
          FALLBACK_MODULES='plugins|adapter_base|config|main|preflight|prompt|coordinator|consolidation|events|heartbeat|transcript_auth|runtime_wedge|watcher|skill_loader|policies|adapters|builtin_tools|executor_helpers|a2a_executor|a2a_client|a2a_tools|a2a_cli|a2a_mcp_server|agent|agents_md|initial_prompt|molecule_ai_status|platform_auth|shared_runtime'
          RUNTIME_MODULES=""
          mkdir -p /tmp/runtime-wheel
          if pip download --quiet molecule-ai-workspace-runtime --no-deps -d /tmp/runtime-wheel 2>/dev/null; then
            WHEEL=$(ls /tmp/runtime-wheel/*.whl 2>/dev/null | head -1)
            if [ -n "$WHEEL" ]; then
              RUNTIME_MODULES=$(unzip -p "$WHEEL" molecule_runtime/_runtime_modules.json 2>/dev/null \
                | python3 -c "import sys,json; m=json.load(sys.stdin); print('|'.join(sorted(set(m['top_level_modules']) | set(m['subpackages']))))" 2>/dev/null || echo "")
            fi
          fi
          if [ -n "$RUNTIME_MODULES" ]; then
            echo "::notice::lint module list from published wheel"
          else
            RUNTIME_MODULES="$FALLBACK_MODULES"
            echo "::warning::could not read _runtime_modules.json from wheel — using inline fallback"
          fi
          if HITS=$(grep -nE "^\s*from (${RUNTIME_MODULES}) import" *.py 2>/dev/null); then
            echo "::error::Bare imports of runtime modules found — use 'from molecule_runtime.<module> import'"
            echo "$HITS" | sed 's/^/  /'
            exit 1
          fi
          echo "::notice::no bare imports of runtime modules in *.py files"
      - name: Log in to ECR
        env:
          AWS_ACCESS_KEY_ID: ${{ secrets.AWS_ACCESS_KEY_ID }}
          AWS_SECRET_ACCESS_KEY: ${{ secrets.AWS_SECRET_ACCESS_KEY }}
        run: |
          set -euo pipefail
          aws ecr get-login-password --region us-east-2 | \
            docker login --username AWS --password-stdin "${ECR_REGISTRY}"
      - name: Verify Docker daemon access
        run: |
          set -euo pipefail
          docker info >/dev/null 2>&1 || {
            echo "::error::Docker daemon is not accessible — check runner sock mount"
            exit 1
          }
          echo "Docker daemon OK"
      - name: Set up Docker Buildx
        uses: docker/setup-buildx-action@4d04d5d9486b7bd6fa91e7baf45bbb4f8b9deedd # v4.0.0
      - name: Ensure ECR repository exists
        env:
          AWS_ACCESS_KEY_ID: ${{ secrets.AWS_ACCESS_KEY_ID }}
          AWS_SECRET_ACCESS_KEY: ${{ secrets.AWS_SECRET_ACCESS_KEY }}
        run: |
          set -euo pipefail
          repo_path="${IMAGE_NAME#*/}"
          repo_path="${repo_path#*/}"  # strip registry host + first slash → molecule-ai/workspace-template-claude-code
          if ! aws ecr describe-repositories --repository-names "${repo_path}" --region us-east-2 >/dev/null 2>&1; then
            aws ecr create-repository \
              --repository-name "${repo_path}" \
              --image-scanning-configuration scanOnPush=true \
              --region us-east-2 >/dev/null
            echo "::notice::created ECR repository ${repo_path}"
          else
            echo "ECR repository ${repo_path} already exists"
          fi
      - name: Build image (load for smoke test, do not push yet)
        # Build into runner-local docker first. Smoke test runs before push so
        # a broken adapter.py never poisons :latest.
        uses: docker/build-push-action@bcafcacb16a39f128d818304e6c9c0c18556b85f # v7.1.0
        with:
          context: .
          file: ./Dockerfile
          platforms: linux/amd64
          load: true
          push: false
          tags: ${{ env.IMAGE_NAME }}:sha-${{ needs.resolve-version.outputs.sha }}
          build-args: |
            RUNTIME_VERSION=${{ needs.resolve-version.outputs.version }}
          labels: |
            org.opencontainers.image.source=https://git.moleculesai.app/${{ github.repository }}
            org.opencontainers.image.revision=${{ github.sha }}
            org.opencontainers.image.description=Molecule AI workspace template — claude-code runtime
      - name: Smoke test — import every /app/*.py
        # Boot the locally-loaded image and import each *.py module to verify
        # all module-level imports resolve against the pip-installed runtime.
        shell: bash
        env:
          IMAGE: ${{ env.IMAGE_NAME }}:sha-${{ needs.resolve-version.outputs.sha }}
        run: |
          set -eu
          docker run --rm \
            -e WORKSPACE_ID=smoke-test \
            -e CLAUDE_CODE_OAUTH_TOKEN=sk-fake-smoke-token \
            -e ANTHROPIC_API_KEY=sk-fake-smoke-key \
            -e OPENAI_API_KEY=sk-fake-smoke-key \
            --entrypoint sh "${IMAGE}" -c '
            set -e
            cd /app
            for f in *.py; do
              [ "$f" = "__init__.py" ] && continue
              mod="${f%.py}"
              python3 -c "import $mod" || { echo "::error::failed to import $mod"; exit 1; }
              echo "  import $mod OK"
            done
          '
          echo "::notice::${IMAGE}: all /app/*.py modules import cleanly"
      - name: Push image to ECR (post-smoke)
        # Smoke passed — push both :latest and :sha-<7>. build-push-action
        # reuses the cached layers so this is a layer-push, not a rebuild.
        uses: docker/build-push-action@bcafcacb16a39f128d818304e6c9c0c18556b85f # v7.1.0
        with:
          context: .
          file: ./Dockerfile
          platforms: linux/amd64
          push: true
          tags: |
            ${{ env.IMAGE_NAME }}:latest
            ${{ env.IMAGE_NAME }}:sha-${{ needs.resolve-version.outputs.sha }}
          build-args: |
            RUNTIME_VERSION=${{ needs.resolve-version.outputs.version }}
          labels: |
            org.opencontainers.image.source=https://git.moleculesai.app/${{ github.repository }}
            org.opencontainers.image.revision=${{ github.sha }}
            org.opencontainers.image.description=Molecule AI workspace template — claude-code runtime
--- a/.gitea/workflows/secret-scan.yml
+++ b/.gitea/workflows/secret-scan.yml
@ -0,0 +1,196 @@
 name: Secret scan
 # Hard CI gate. Refuses any PR / push whose diff additions contain a
 # recognisable credential. Defense-in-depth for the #2090-class incident
 # (2026-04-24): GitHub's hosted Copilot Coding Agent leaked a ghs_*
 # installation token into tenant-proxy/package.json via `npm init`
 # slurping the URL from a token-embedded origin remote. We can't fix
 # upstream's clone hygiene, so we gate here.
 #
 # Same regex set as the runtime's bundled pre-commit hook
 # (molecule-ai-workspace-runtime: molecule_runtime/scripts/pre-commit-checks.sh).
 # Keep the two sides aligned when adding patterns.
 #
 # Ported from .github/workflows/secret-scan.yml so the gate actually
 # fires on Gitea Actions. Differences from the GitHub version:
 #   - drops `merge_group` event (Gitea has no merge queue)
 #   - drops `workflow_call` (no cross-repo reusable invocation on Gitea)
 #   - SELF path updated to .gitea/workflows/secret-scan.yml
 # The job name + step name are identical to the GitHub workflow so the
 # status-check context (`Secret scan / Scan diff for credential-shaped
 # strings (pull_request)`) matches branch protection on this template
 # repo's main branch. Before this port, the required-status was satisfied
 # only via a compensating signed POST /statuses/{SHA} because the
 # .github/ workflow was silently shadowed by the .gitea/ directory taking
 # precedence on this repo
 # (reference_molecule_core_actions_gitea_only — same applies here).
 on:
  pull_request:
    types: [opened, synchronize, reopened]
  push:
    branches: [main, staging]
 jobs:
  scan:
    name: Scan diff for credential-shaped strings
    runs-on: ubuntu-latest
    steps:
      - uses: actions/checkout@de0fac2e4500dabe0009e67214ff5f5447ce83dd # v6.0.2
        with:
          fetch-depth: 2  # need previous commit to diff against on push events
      # For pull_request events the diff base may be many commits behind
      # HEAD and absent from the shallow clone. Fetch it explicitly.
      - name: Fetch PR base SHA (pull_request events only)
        if: github.event_name == 'pull_request'
        run: git fetch --depth=1 origin ${{ github.event.pull_request.base.sha }}
      - name: Refuse if credential-shaped strings appear in diff additions
        env:
          # Plumb event-specific SHAs through env so the script doesn't
          # need conditional `${{ ... }}` interpolation per event type.
          # github.event.before/after only exist on push events;
          # pull_request has pull_request.base.sha / pull_request.head.sha.
          PR_BASE_SHA: ${{ github.event.pull_request.base.sha }}
          PR_HEAD_SHA: ${{ github.event.pull_request.head.sha }}
          PUSH_BEFORE: ${{ github.event.before }}
          PUSH_AFTER: ${{ github.event.after }}
        run: |
          # Pattern set covers GitHub family (the actual #2090 vector),
          # Anthropic / OpenAI / Slack / AWS. Anchored on prefixes with low
          # false-positive rates against agent-generated content. Mirror of
          # molecule-ai-workspace-runtime/molecule_runtime/scripts/pre-commit-checks.sh
          # — keep aligned.
          SECRET_PATTERNS=(
            'ghp_[A-Za-z0-9]{36,}'           # GitHub PAT (classic)
            'ghs_[A-Za-z0-9]{36,}'           # GitHub App installation token
            'gho_[A-Za-z0-9]{36,}'           # GitHub OAuth user-to-server
            'ghu_[A-Za-z0-9]{36,}'           # GitHub OAuth user
            'ghr_[A-Za-z0-9]{36,}'           # GitHub OAuth refresh
            'github_pat_[A-Za-z0-9_]{82,}'   # GitHub fine-grained PAT
            'sk-ant-[A-Za-z0-9_-]{40,}'      # Anthropic API key
            'sk-proj-[A-Za-z0-9_-]{40,}'     # OpenAI project key
            'sk-svcacct-[A-Za-z0-9_-]{40,}'  # OpenAI service-account key
            'sk-cp-[A-Za-z0-9_-]{60,}'       # MiniMax API key (F1088 vector — caught only after the fact)
            'xox[baprs]-[A-Za-z0-9-]{20,}'   # Slack tokens
            'AKIA[0-9A-Z]{16}'               # AWS access key ID
            'ASIA[0-9A-Z]{16}'               # AWS STS temp access key ID
          )
          # Determine the diff base. Each event type stores its SHAs in
          # a different place — see the env block above.
          case "${{ github.event_name }}" in
            pull_request)
              BASE="$PR_BASE_SHA"
              HEAD="$PR_HEAD_SHA"
              ;;
            *)
              BASE="$PUSH_BEFORE"
              HEAD="$PUSH_AFTER"
              ;;
          esac
          # On push events with shallow clones, BASE may be present in
          # the event payload but absent from the local object DB
          # (fetch-depth=2 doesn't always reach the previous commit
          # across true merges). Try fetching it on demand. If the
          # fetch fails — e.g. the SHA was force-overwritten — we fall
          # through to the empty-BASE branch below, which scans the
          # entire tree as if every file were new. Correct, just slow.
          if [ -n "$BASE" ] && ! echo "$BASE" | grep -qE '^0+$'; then
            if ! git cat-file -e "$BASE" 2>/dev/null; then
              git fetch --depth=1 origin "$BASE" 2>/dev/null || true
            fi
          fi
          # Files added or modified in this change.
          if [ -z "$BASE" ] || echo "$BASE" | grep -qE '^0+$' || ! git cat-file -e "$BASE" 2>/dev/null; then
            # New branch / no previous SHA / BASE unreachable — check the
            # entire tree as added content. Slower, but correct on first
            # push.
            CHANGED=$(git ls-tree -r --name-only HEAD)
            DIFF_RANGE=""
          else
            CHANGED=$(git diff --name-only --diff-filter=AM "$BASE" "$HEAD")
            DIFF_RANGE="$BASE $HEAD"
          fi
          if [ -z "$CHANGED" ]; then
            echo "No changed files to inspect."
            exit 0
          fi
          # Self-exclude: this workflow file legitimately contains the
          # pattern strings as regex literals. Without an exclude it would
          # block its own merge. Both the .github/ original and this
          # .gitea/ port are excluded so a sync between them stays clean.
          SELF_GITHUB=".github/workflows/secret-scan.yml"
          SELF_GITEA=".gitea/workflows/secret-scan.yml"
          OFFENDING=""
          # `while IFS= read -r` (not `for f in $CHANGED`) so filenames
          # containing whitespace don't word-split silently — a path
          # with a space would otherwise produce two iterations on
          # tokens that aren't real filenames, breaking the
          # self-exclude + diff lookup.
          while IFS= read -r f; do
            [ -z "$f" ] && continue
            [ "$f" = "$SELF_GITHUB" ] && continue
            [ "$f" = "$SELF_GITEA" ] && continue
            if [ -n "$DIFF_RANGE" ]; then
              ADDED=$(git diff --no-color --unified=0 "$BASE" "$HEAD" -- "$f" 2>/dev/null | grep -E '^\+[^+]' || true)
            else
              # No diff range (new branch first push) — scan the full file
              # contents as if every line were new.
              ADDED=$(cat "$f" 2>/dev/null || true)
            fi
            [ -z "$ADDED" ] && continue
            for pattern in "${SECRET_PATTERNS[@]}"; do
              if echo "$ADDED" | grep -qE "$pattern"; then
                OFFENDING="${OFFENDING}${f} (matched: ${pattern})\n"
                break
              fi
            done
          done <<< "$CHANGED"
          if [ -n "$OFFENDING" ]; then
            echo "::error::Credential-shaped strings detected in diff additions:"
            # `printf '%b' "$OFFENDING"` interprets backslash escapes
            # (the literal `\n` we appended above becomes a newline)
            # WITHOUT treating OFFENDING as a format string. Plain
            # `printf "$OFFENDING"` is a format-string sink: a filename
            # containing `%` would be interpreted as a conversion
            # specifier, corrupting the error message (or printing
            # `%(missing)` artifacts).
            printf '%b' "$OFFENDING"
            echo ""
            echo "The actual matched values are NOT echoed here, deliberately —"
            echo "round-tripping a leaked credential into CI logs widens the blast"
            echo "radius (logs are searchable + retained)."
            echo ""
            echo "Recovery:"
            echo "  1. Remove the secret from the file. Replace with an env var"
            echo "     reference (e.g. \${{ secrets.GITHUB_TOKEN }} in workflows,"
            echo "     process.env.X in code)."
            echo "  2. If the credential was already pushed (this PR's commit"
            echo "     history reaches a public ref), treat it as compromised —"
            echo "     ROTATE it immediately, do not just remove it. The token"
            echo "     remains valid in git history forever and may be in any"
            echo "     log/cache that consumed this branch."
            echo "  3. Force-push the cleaned commit (or stack a revert) and"
            echo "     re-run CI."
            echo ""
            echo "If the match is a false positive (test fixture, docs example,"
            echo "or this workflow's own regex literals): use a clearly-fake"
            echo "placeholder like ghs_EXAMPLE_DO_NOT_USE that doesn't satisfy"
            echo "the length suffix, OR add the file path to the SELF exclude"
            echo "list in this workflow with a short reason."
            echo ""
            echo "Mirror of the regex set lives in the runtime's bundled"
            echo "pre-commit hook (molecule-ai-workspace-runtime:"
            echo "molecule_runtime/scripts/pre-commit-checks.sh) — keep aligned."
            exit 1
          fi
          echo "✓ No credential-shaped strings in this change."
--- a/.github/workflows/ci.yml
+++ b/.github/workflows/ci.yml
@ -2,7 +2,7 @@ name: CI
 on: [push, pull_request]
 jobs:
  validate:
-    uses: Molecule-AI/molecule-ci/.github/workflows/validate-workspace-template.yml@main
+    uses: molecule-ai/molecule-ci/.github/workflows/validate-workspace-template.yml@main
  tests:
    name: Adapter unit tests
--- a/.github/workflows/publish-image.yml
+++ b/.github/workflows/publish-image.yml
@ -32,14 +32,47 @@ permissions:
  packages: write
 jobs:
  # The `.runtime-version` file is the push-mode cascade signal post-
  # 2026-05-06: when molecule-core/publish-runtime.yml ships a new
  # version to PyPI, it does NOT call repository_dispatch (Gitea 1.22.6
  # has no such endpoint — empirically verified molecule-core#20).
  # Instead it git-pushes an updated `.runtime-version` to each template,
  # which trips this workflow's `on: push: branches: [main]` trigger.
  # This job reads that file and forwards the version to the reusable
  # build workflow so the Dockerfile pip-installs the exact published
  # version, not whatever requirements.txt currently bounds.
  resolve-version:
    runs-on: ubuntu-latest
    timeout-minutes: 2
    outputs:
      version: ${{ steps.read.outputs.version }}
    steps:
      - uses: actions/checkout@v4
      - id: read
        run: |
          if [ -f .runtime-version ]; then
            v="$(head -n1 .runtime-version | tr -d '[:space:]')"
            echo "version=$v" >> "$GITHUB_OUTPUT"
            echo "resolved runtime version: $v"
          else
            echo "no .runtime-version file present — falling through to Dockerfile default"
          fi
  publish:
-    uses: Molecule-AI/molecule-ci/.github/workflows/publish-template-image.yml@main
+    needs: resolve-version
    uses: molecule-ai/molecule-ci/.github/workflows/publish-template-image.yml@main
    secrets: inherit
    with:
-      # When the cascade fires, client_payload.runtime_version is the
+      # Resolution chain (highest priority first):
-      # exact version PyPI just published. Forwarded to the reusable
+      #   1. client_payload.runtime_version — legacy GitHub
-      # workflow as a docker --build-arg so the cache key changes
+      #      repository_dispatch path (will return if Gitea ever adds
-      # per-version and pip install resolves freshly.
+      #      the dispatch API; left in place for forward-compat).
-      # On other events (push to main / manual without input), this is
+      #   2. inputs.runtime_version — manual workflow_dispatch run from
-      # empty and the Dockerfile's default (requirements.txt pin) applies.
+      #      the Actions UI for ad-hoc rebuilds against a specific
-      runtime_version: ${{ github.event.client_payload.runtime_version || inputs.runtime_version || '' }}
+      #      version.
      #   3. needs.resolve-version.outputs.version — the
      #      `.runtime-version` file in this repo, written by
      #      molecule-core/publish-runtime.yml's push-mode cascade.
      #   4. '' — fall through to the Dockerfile default
      #      (requirements.txt pin).
      runtime_version: ${{ github.event.client_payload.runtime_version || inputs.runtime_version || needs.resolve-version.outputs.version || '' }}
--- a/.github/workflows/secret-scan.yml
+++ b/.github/workflows/secret-scan.yml
@ -1,22 +1,201 @@
 name: Secret scan
-# Calls the canonical reusable workflow in molecule-core. Defense
+# Hard CI gate. Refuses any PR / push whose diff additions contain a
-# against the #2090-class leak (a hosted-agent commit slipping a
+# recognisable credential. Defense-in-depth for the #2090-class incident
-# credential-shaped string into a PR). Pattern set lives in
+# (2026-04-24): GitHub's hosted Copilot Coding Agent leaked a ghs_*
-# molecule-core so we do not maintain a parallel copy here.
+# installation token into tenant-proxy/package.json via `npm init`
 # slurping the URL from a token-embedded origin remote. We can't fix
 # upstream's clone hygiene, so we gate here.
 #
-# Pinned to @staging because that is the active default branch on the
+# Inlined copy from molecule-ai/molecule-core/.github/workflows/secret-scan.yml.
-# upstream repo (main lags behind via the staging-promotion workflow).
+# Cross-repo workflow_call to a private repo doesn't fully work on Gitea 1.22.6
-# Updates ride along automatically as the upstream regex set evolves.
+# (workflow file fails parse-time at 0s with no logs); inline keeps the gate
 # functional until Gitea is upgraded or the canonical scanner moves to a public
 # repo. When that lands, this file reverts to the 3-line wrapper:
 #
 #   jobs:
 #     secret-scan:
 #       uses: Molecule-AI/molecule-core/.github/workflows/secret-scan.yml@staging
 #
 # Pin to @staging not @main — staging is the active default branch,
 # main lags via the staging-promotion workflow. Updates ride along
 # automatically on the next consumer workflow run.
 #
 # Same regex set as the runtime's bundled pre-commit hook
 # (molecule-ai-workspace-runtime: molecule_runtime/scripts/pre-commit-checks.sh).
 # Keep the two sides aligned when adding patterns.
 on:
  pull_request:
    types: [opened, synchronize, reopened]
  push:
-    branches: [main, staging, master]
+    branches: [main, staging]
  merge_group:
    types: [checks_requested]
 jobs:
-  secret-scan:
+  scan:
-    uses: Molecule-AI/molecule-core/.github/workflows/secret-scan.yml@staging
+    name: Scan diff for credential-shaped strings
    runs-on: ubuntu-latest
    steps:
      - uses: actions/checkout@de0fac2e4500dabe0009e67214ff5f5447ce83dd # v6.0.2
        with:
          fetch-depth: 2  # need previous commit to diff against on push events
      # For pull_request events the diff base may be many commits behind
      # HEAD and absent from the shallow clone. Fetch it explicitly.
      - name: Fetch PR base SHA (pull_request events only)
        if: github.event_name == 'pull_request'
        run: git fetch --depth=1 origin ${{ github.event.pull_request.base.sha }}
      # For merge_group events the queue's pre-merge ref is a commit on
      # `gh-readonly-queue/...` whose parent is the queue's base_sha.
      # That parent isn't part of the queue branch's shallow clone, so
      # we fetch it explicitly. Without this the diff falls through to
      # "no BASE → scan entire tree" mode and false-positives on legit
      # test fixtures (e.g. canvas/src/lib/validation/__tests__/secret-formats.test.ts).
      - name: Refuse if credential-shaped strings appear in diff additions
        env:
          # Plumb event-specific SHAs through env so the script doesn't
          # need conditional `${{ ... }}` interpolation per event type.
          # github.event.before/after only exist on push events;
          # merge_group has its own base_sha/head_sha; pull_request has
          # pull_request.base.sha / pull_request.head.sha.
          PR_BASE_SHA: ${{ github.event.pull_request.base.sha }}
          PR_HEAD_SHA: ${{ github.event.pull_request.head.sha }}
          PUSH_BEFORE: ${{ github.event.before }}
          PUSH_AFTER: ${{ github.event.after }}
        run: |
          # Pattern set covers GitHub family (the actual #2090 vector),
          # Anthropic / OpenAI / Slack / AWS. Anchored on prefixes with low
          # false-positive rates against agent-generated content. Mirror of
          # molecule-ai-workspace-runtime/molecule_runtime/scripts/pre-commit-checks.sh
          # — keep aligned.
          SECRET_PATTERNS=(
            'ghp_[A-Za-z0-9]{36,}'           # GitHub PAT (classic)
            'ghs_[A-Za-z0-9]{36,}'           # GitHub App installation token
            'gho_[A-Za-z0-9]{36,}'           # GitHub OAuth user-to-server
            'ghu_[A-Za-z0-9]{36,}'           # GitHub OAuth user
            'ghr_[A-Za-z0-9]{36,}'           # GitHub OAuth refresh
            'github_pat_[A-Za-z0-9_]{82,}'   # GitHub fine-grained PAT
            'sk-ant-[A-Za-z0-9_-]{40,}'      # Anthropic API key
            'sk-proj-[A-Za-z0-9_-]{40,}'     # OpenAI project key
            'sk-svcacct-[A-Za-z0-9_-]{40,}'  # OpenAI service-account key
            'sk-cp-[A-Za-z0-9_-]{60,}'       # MiniMax API key (F1088 vector — caught only after the fact)
            'xox[baprs]-[A-Za-z0-9-]{20,}'   # Slack tokens
            'AKIA[0-9A-Z]{16}'               # AWS access key ID
            'ASIA[0-9A-Z]{16}'               # AWS STS temp access key ID
          )
          # Determine the diff base. Each event type stores its SHAs in
          # a different place — see the env block above.
          case "${{ github.event_name }}" in
            pull_request)
              BASE="$PR_BASE_SHA"
              HEAD="$PR_HEAD_SHA"
              ;;
            *)
              BASE="$PUSH_BEFORE"
              HEAD="$PUSH_AFTER"
              ;;
          esac
          # On push events with shallow clones, BASE may be present in
          # the event payload but absent from the local object DB
          # (fetch-depth=2 doesn't always reach the previous commit
          # across true merges). Try fetching it on demand. If the
          # fetch fails — e.g. the SHA was force-overwritten — we fall
          # through to the empty-BASE branch below, which scans the
          # entire tree as if every file were new. Correct, just slow.
          if [ -n "$BASE" ] && ! echo "$BASE" | grep -qE '^0+$'; then
            if ! git cat-file -e "$BASE" 2>/dev/null; then
              git fetch --depth=1 origin "$BASE" 2>/dev/null || true
            fi
          fi
          # Files added or modified in this change.
          if [ -z "$BASE" ] || echo "$BASE" | grep -qE '^0+$' || ! git cat-file -e "$BASE" 2>/dev/null; then
            # New branch / no previous SHA / BASE unreachable — check the
            # entire tree as added content. Slower, but correct on first
            # push.
            CHANGED=$(git ls-tree -r --name-only HEAD)
            DIFF_RANGE=""
          else
            CHANGED=$(git diff --name-only --diff-filter=AM "$BASE" "$HEAD")
            DIFF_RANGE="$BASE $HEAD"
          fi
          if [ -z "$CHANGED" ]; then
            echo "No changed files to inspect."
            exit 0
          fi
          # Self-exclude: this workflow file legitimately contains the
          # pattern strings as regex literals. Without an exclude it would
          # block its own merge.
          SELF=".github/workflows/secret-scan.yml"
          OFFENDING=""
          # `while IFS= read -r` (not `for f in $CHANGED`) so filenames
          # containing whitespace don't word-split silently — a path
          # with a space would otherwise produce two iterations on
          # tokens that aren't real filenames, breaking the
          # self-exclude + diff lookup.
          while IFS= read -r f; do
            [ -z "$f" ] && continue
            [ "$f" = "$SELF" ] && continue
            if [ -n "$DIFF_RANGE" ]; then
              ADDED=$(git diff --no-color --unified=0 "$BASE" "$HEAD" -- "$f" 2>/dev/null | grep -E '^\+[^+]' || true)
            else
              # No diff range (new branch first push) — scan the full file
              # contents as if every line were new.
              ADDED=$(cat "$f" 2>/dev/null || true)
            fi
            [ -z "$ADDED" ] && continue
            for pattern in "${SECRET_PATTERNS[@]}"; do
              if echo "$ADDED" | grep -qE "$pattern"; then
                OFFENDING="${OFFENDING}${f} (matched: ${pattern})\n"
                break
              fi
            done
          done <<< "$CHANGED"
          if [ -n "$OFFENDING" ]; then
            echo "::error::Credential-shaped strings detected in diff additions:"
            # `printf '%b' "$OFFENDING"` interprets backslash escapes
            # (the literal `\n` we appended above becomes a newline)
            # WITHOUT treating OFFENDING as a format string. Plain
            # `printf "$OFFENDING"` is a format-string sink: a filename
            # containing `%` would be interpreted as a conversion
            # specifier, corrupting the error message (or printing
            # `%(missing)` artifacts).
            printf '%b' "$OFFENDING"
            echo ""
            echo "The actual matched values are NOT echoed here, deliberately —"
            echo "round-tripping a leaked credential into CI logs widens the blast"
            echo "radius (logs are searchable + retained)."
            echo ""
            echo "Recovery:"
            echo "  1. Remove the secret from the file. Replace with an env var"
            echo "     reference (e.g. \${{ secrets.GITHUB_TOKEN }} in workflows,"
            echo "     process.env.X in code)."
            echo "  2. If the credential was already pushed (this PR's commit"
            echo "     history reaches a public ref), treat it as compromised —"
            echo "     ROTATE it immediately, do not just remove it. The token"
            echo "     remains valid in git history forever and may be in any"
            echo "     log/cache that consumed this branch."
            echo "  3. Force-push the cleaned commit (or stack a revert) and"
            echo "     re-run CI."
            echo ""
            echo "If the match is a false positive (test fixture, docs example,"
            echo "or this workflow's own regex literals): use a clearly-fake"
            echo "placeholder like ghs_EXAMPLE_DO_NOT_USE that doesn't satisfy"
            echo "the length suffix, OR add the file path to the SELF exclude"
            echo "list in this workflow with a short reason."
            echo ""
            echo "Mirror of the regex set lives in the runtime's bundled"
            echo "pre-commit hook (molecule-ai-workspace-runtime:"
            echo "molecule_runtime/scripts/pre-commit-checks.sh) — keep aligned."
            exit 1
          fi
          echo "✓ No credential-shaped strings in this change."
--- a/.runtime-version
+++ b/.runtime-version
@ -0,0 +1 @@
 0.1.129
--- a/CLAUDE.md
+++ b/CLAUDE.md
@ -72,6 +72,27 @@ Skills persist across restarts. Use them to codify best practices, coding standa
 ## Language
 Always respond in the same language the user uses. If Chinese, respond in Chinese. If English, respond in English. Match exactly.
 ## Workspace ownership tier — T4 (top-tier)
 **T4 (top-tier) workspaces grant full host control.** The Docker socket is mounted into the container; the runtime is started with `--privileged`; the workspace owner can `docker exec` into siblings on the same host and read/write anything the host kernel will let it touch.
 What this means in practice:
 - The container's home directory `~/.claude/` (and its `sessions/`, `session-env/`, `settings.json`) is persisted across restarts via a host bind mount. Anything you (or the SDK) write there survives container churn — but ownership can drift.
 - If you see `EPERM` / "permission denied" on `~/.claude/` writes after a restart — particularly on `settings.json` or anything under `sessions/` — the volume picked up `root:root` ownership from a prior boot or a newer claude-code release wrote subdirs the entrypoint didn't pre-create.
 - Recovery (do this yourself; you have the rights): `sudo chown -R agent:agent ~/.claude/`. The entrypoint already runs a recursive chown unconditionally on every boot, so a restart also clears it — but `sudo chown` is faster and doesn't drop the active session.
 - The provisioned `~/.claude/settings.json` is informational: it documents that `permission_mode='bypassPermissions'` is the canonical operating mode, which is also set programmatically in `claude_sdk_executor.py` (the file is NOT the source of truth — the SDK kwargs are).
 If `cat ~/.claude/settings.json` returns `No such file or directory` you're on a workspace image older than 2026-05-15 — restart picks up the new entrypoint and stubs the file in place.
 ## Knowing your own model
 Use the `get_runtime_identity` MCP tool to know what model you actually are. It reads the live process env (`MODEL`, `MODEL_PROVIDER`, `MOLECULE_MODEL`, `ANTHROPIC_BASE_URL`, `TIER`, `WORKSPACE_ID`, `ADAPTER_MODULE`) and returns the resolved values — no HTTP call, always works, always permitted by RBAC. Do NOT guess from your system prompt or from `requirements.txt`; the operator may have routed you to a different model via persona env between boots.
 ## Editing your own agent_card
 Use the `update_agent_card` MCP tool to update this workspace's `agent_card` on the platform. Pass a JSON object — the platform validates required fields server-side. The change is broadcast as an `agent_card_updated` event so the canvas reflects the new card live. The tool is gated on `memory.write` capability, so read-only agents won't accidentally rewrite the card; T4 owners always have this capability.
 ## Runtime wedge integration
 The `runtime_wedge` module (in `molecule_runtime`) is the universal cross-cutting holder for "this Python process can no longer serve queries — only a workspace restart will recover." It surfaces unrecoverable wedges to two consumers:
--- a/55
+++ b/55
@ -5,8 +5,23 @@ FROM python:3.11-slim
 # --add-assignee`, `git clone`, etc. per their idle/cron prompts).
 # Without these the team's claim-and-ship loop silently returns
 # "(no response generated)" because tools error out.
 #
 # T4 escalation leg (RFC internal#456 §9 / PR#474):
 #   sudo + util-linux(nsenter) + docker.io(CLI) are baked here so the
 #   uid-1000 `agent` (see useradd below — UNCHANGED, agent stays
 #   uid-1000) has a wired, audited path to host root inside the
 #   provisioner's `--privileged --pid=host -v /:/host
 #   -v /var/run/docker.sock:/var/run/docker.sock` container. Without
 #   sudo, a uid-1000 process in --privileged CANNOT nsenter/chroot
 #   /host (--privileged grants caps to root, not uid-1000) and cannot
 #   use the root:docker 0660 docker.sock — T4 would be
 #   provisioner-shape-only (the documented ABSENT-escalation-leg gap).
 #   The sudoers drop-in + docker-group add are below, after useradd,
 #   so `agent` exists. This is ADDITIVE: it does NOT change the agent
 #   uid and does NOT change /configs token ownership (still uid-1000,
 #   enforced by entrypoint.sh + the Layer-3 conformance gate).
 RUN apt-get update && apt-get install -y --no-install-recommends \
-    curl gosu nodejs npm ca-certificates git \
+    curl gosu nodejs npm ca-certificates git sudo util-linux docker.io \
    && install -m 0755 -d /etc/apt/keyrings \
    && curl -fsSL https://cli.github.com/packages/githubcli-archive-keyring.gpg | tee /etc/apt/keyrings/githubcli-archive-keyring.gpg > /dev/null \
    && chmod go+r /etc/apt/keyrings/githubcli-archive-keyring.gpg \
@ -17,8 +32,31 @@ RUN apt-get update && apt-get install -y --no-install-recommends \
 # Install claude-code CLI via npm
 RUN npm install -g @anthropic-ai/claude-code 2>/dev/null || true
-# Create agent user
+# Create agent user — UNCHANGED. The agent runs as uid-1000; the T4
 # escalation leg below is additive and does NOT promote the agent to
 # root. claude-code still refuses --dangerously-skip-permissions as
 # root, and /configs/.auth_token must stay agent-owned (Hermes
 # list_peers 401 class — RFC internal#456 §10).
 RUN useradd -u 1000 -m -s /bin/bash agent
 # --- T4 escalation leg (RFC internal#456 §9.3 / PR#474) ---
 # Wired path: uid-1000 agent -> host root inside the provisioner's
 # --privileged --pid=host -v /:/host -v docker.sock container.
 #   1. NOPASSWD sudoers drop-in (mode 0440, visudo-validated at build
 #      so a malformed sudoers can never ship a broken-sudo image).
 #   2. agent in the `docker` group so the bind-mounted root:docker
 #      0660 /var/run/docker.sock is usable without sudo.
 # Atomic co-sequencing (RFC §10): this ships in the SAME image
 # revision as the uid-1000 + agent-owned-token entrypoint contract;
 # the Layer-3 conformance gate asserts BOTH on the running container.
 RUN set -eux; \
    printf 'agent ALL=(ALL) NOPASSWD:ALL\n' > /etc/sudoers.d/agent-t4; \
    chmod 0440 /etc/sudoers.d/agent-t4; \
    visudo -cf /etc/sudoers.d/agent-t4; \
    groupadd -f docker; \
    usermod -aG docker agent; \
    id agent
 WORKDIR /app
 # RUNTIME_VERSION is forwarded from the reusable publish workflow as
@ -43,6 +81,19 @@ RUN pip install --no-cache-dir -r requirements.txt && \
 # Copy adapter code
 COPY adapter.py .
 COPY __init__.py .
 # Provider registry. The adapter's _load_providers walks 4 paths:
 #   1. /opt/adapter/config.yaml          — provisioner-managed canonical
 #   2. os.path.dirname(__file__)/config.yaml  — alongside adapter.py (this image)
 #   3. ${WORKSPACE_CONFIG_PATH}/config.yaml   — workspace per-instance overrides
 #   4. _BUILTIN_PROVIDERS                — oauth + anthropic-api only
 # On this image /opt/adapter/ is never populated by the platform
 # provisioner, so path 2 (/app/config.yaml) is the load-bearing one.
 # Without this COPY the file isn't in the image, all 3 file paths fail,
 # and _load_providers falls through to _BUILTIN_PROVIDERS — every
 # MiniMax/GLM/Kimi/DeepSeek model silently routes to anthropic-oauth →
 # "Not logged in. Please run /login" at first LLM call. Caused the
 # canary's 38h chronic red on 2026-05-07/08 (molecule-core#129).
 COPY config.yaml .
 # Adapter-specific executor — owned by THIS template (universal-runtime
 # refactor, molecule-core task #87). Lives alongside adapter.py so
 # Python's import system picks the local /app/claude_sdk_executor.py
--- a/adapter.py
+++ b/adapter.py
@ -144,39 +144,135 @@ def _normalize_provider(entry: dict):
        "model_aliases": _coerce_string_list(entry.get("model_aliases"), lowercase=True),
        "base_url": entry.get("base_url") or None,
        "auth_env": _coerce_string_list(entry.get("auth_env"), lowercase=False),
        # Which env var the boot-time vendor-key projection writes the
        # vendor key INTO. Defaults to ANTHROPIC_AUTH_TOKEN (Bearer-style
        # — correct for MiniMax/GLM/DeepSeek Anthropic-compat shims).
        # Kimi For Coding's gateway authenticates with the x-api-key
        # header (per kimi.com's official Claude Code doc), which the
        # Anthropic SDK / claude CLI emits from ANTHROPIC_API_KEY — so
        # that provider's entry sets auth_token_env: ANTHROPIC_API_KEY.
        # Env-var names are case-sensitive; preserve case.
        "auth_token_env": (
            entry.get("auth_token_env")
            if isinstance(entry.get("auth_token_env"), str)
            and entry.get("auth_token_env").strip()
            else "ANTHROPIC_AUTH_TOKEN"
        ),
    }
 # Canonical install path the platform provisioner is contracted to clone
 # the template repo into. Hardcoded so the adapter's config.yaml lookup
 # is invariant across Docker (mounted /app→/opt/adapter) and EC2-host
 # (cloned by molecule-controlplane's ec2.go) install paths — robust
 # against the site-packages copy that bit us 2026-05-04 11:08Z.
 _CANONICAL_ADAPTER_DIR = "/opt/adapter"
 # Adjacent-to-adapter.py path. Module-level so tests can monkeypatch it
 # to redirect the path-2 lookup at a controlled tmp dir. Production code
 # resolves this once at import time and never touches it again — same
 # semantics as before.
 _TEMPLATE_DIR = os.path.dirname(os.path.abspath(__file__))
 def _load_providers(config_path: str) -> tuple:
-    """Load the provider registry from /configs/config.yaml.
+    """Load the provider registry from the template's bundled config.yaml.
-    The YAML's top-level ``providers:`` list is the canonical source —
+    The providers list is a TEMPLATE concern — it describes which
-    canvas Config tab reads the same list to populate its Provider
+    models/auth-modes this runtime image supports — and ships in the
-    dropdown so the UI and the adapter never disagree on what's
+    template's own config.yaml alongside adapter.py. The per-workspace
-    available. Falls back to ``_BUILTIN_PROVIDERS`` (oauth + anthropic-api)
+    ``${WORKSPACE_CONFIG_PATH}/config.yaml`` (default ``/configs/``)
-    if the file is missing, malformed, or has no providers section, so a
+    only contains workspace-specific overrides (model, runtime, skills,
-    bare-bones workspace still boots with the historical defaults.
+    prompt files) and does NOT carry a providers section.
-    Per-entry isolation: a single bad provider entry is dropped with a
+    Two-step incident history:
-    warning; the rest of the registry survives. Used to be a generator
+      • Pre-2026-05-04 09:00Z: only checked ``config_path``, fell back
-    inside tuple(...) that propagated any AttributeError out and reverted
+        to ``_BUILTIN_PROVIDERS`` (oauth + anthropic-api). Every
-    the whole registry to builtins — exactly the silent-fallback failure
+        MiniMax / GLM / Kimi / DeepSeek model resolved to
-    mode this file's existence was meant to fix.
+        ``anthropic-oauth`` and crashed at first LLM call with
        "Not logged in. Please run /login". Fixed by adding a
        template-bundled lookup using
        ``os.path.dirname(os.path.abspath(__file__))``.
      • 2026-05-04 11:08Z: that ``__file__`` lookup misses on EC2-host
        installs because the provisioner copies adapter.py to
        ``/opt/molecule-venv/lib/python3.12/site-packages/`` —
        site-packages wins over PYTHONPATH=/opt/adapter (which the
        host install doesn't set), so __file__ resolves to the venv
        path WITHOUT an adjacent config.yaml. Same silent fallback
        to anthropic-oauth + same "Not logged in" symptom.
      • 2026-05-08 (#129): the multi-path lookup that fixed both of
        the above was lost in a post-suspension migration cycle (the
        Gitea main branch never carried the fix even though the
        :latest image had it baked in from a prior build). Canary
        chronic red for 38h before this commit restored the lookup.
    Resolution order:
      1. ``/opt/adapter/config.yaml`` — canonical provisioner-managed
         install dir. Hardcoded because the platform contract is
         "provisioner clones template repo into /opt/adapter"; this
         is invariant across Docker (mounted /app→/opt/adapter) and
         EC2-host (cloned by ec2.go) install paths. Robust against
         site-packages copy.
      2. Adjacent to ``adapter.__file__`` — works in dev/test where
         the canonical path doesn't exist. Also covers the Docker
         image's /app/config.yaml (bundled by Dockerfile #6).
      3. Per-workspace ``${config_path}/config.yaml`` — fallback for
         operator-shipped overrides on a private deployment that
         wants a custom providers list.
      4. ``_BUILTIN_PROVIDERS`` — oauth + anthropic-api defaults so a
         bare-bones workspace still boots even with no config.yaml
         anywhere.
    Per-entry isolation: a single bad provider entry is dropped with
    a warning; the rest of the registry survives.
    """
-    yaml_path = os.path.join(config_path, "config.yaml")
+    canonical_yaml = os.path.join(_CANONICAL_ADAPTER_DIR, "config.yaml")
    template_yaml = os.path.join(_TEMPLATE_DIR, "config.yaml")
    workspace_yaml = os.path.join(config_path, "config.yaml")
    # Deduplicate while preserving order — _CANONICAL_ADAPTER_DIR and
    # the __file__ dir collide in dev/test (when imported from
    # /opt/adapter directly), and workspace_yaml may also collide if
    # config_path == /opt/adapter in tests.
    seen = set()
    candidates = []
    for path in (canonical_yaml, template_yaml, workspace_yaml):
        if path not in seen:
            seen.add(path)
            candidates.append(path)
    raw = None
    chosen_path = None
    try:
        import yaml  # transitive dep via molecule-ai-workspace-runtime
    except ImportError:
        logger.warning("providers: yaml import failed; using builtins")
        return _BUILTIN_PROVIDERS
    for yaml_path in candidates:
        try:
            with open(yaml_path, "r") as f:
                data = yaml.safe_load(f) or {}
        except FileNotFoundError:
-        logger.info("providers: %s not found, using builtin defaults", yaml_path)
+            logger.info("providers: %s not found, trying next candidate", yaml_path)
-        return _BUILTIN_PROVIDERS
+            continue
        except Exception as exc:  # noqa: BLE001 — defensive: never block boot on YAML
-        logger.warning("providers: failed to load from %s (%s); using builtins", yaml_path, exc)
+            logger.warning(
-        return _BUILTIN_PROVIDERS
+                "providers: failed to load from %s (%s); trying next candidate",
                yaml_path, exc,
            )
            continue
-    raw = data.get("providers") if isinstance(data, dict) else None
+        candidate_raw = data.get("providers") if isinstance(data, dict) else None
-    if not isinstance(raw, list) or not raw:
+        if isinstance(candidate_raw, list) and candidate_raw:
            raw = candidate_raw
            chosen_path = yaml_path
            break
    if raw is None:
        logger.info(
            "providers: no providers section found in %s; using builtin defaults",
            " or ".join(candidates),
        )
        return _BUILTIN_PROVIDERS
    parsed = []
@ -190,11 +286,139 @@ def _load_providers(config_path: str) -> tuple:
            parsed.append(normalized)
    if not parsed:
-        logger.warning("providers: no valid entries in %s; using builtins", yaml_path)
+        logger.warning("providers: no valid entries in %s; using builtins", chosen_path)
        return _BUILTIN_PROVIDERS
    logger.info("providers: loaded %d entries from %s", len(parsed), chosen_path)
    return tuple(parsed)
 # Aliases for `MODEL_PROVIDER` env values that should map to a registry
 # provider name. The persona env files use shorter / friendlier slugs
 # than the registry's canonical names — without this alias map a value
 # like ``MODEL_PROVIDER=claude-code`` would fall through to YAML-based
 # resolution and (when the YAML doesn't pin a provider) hit the
 # model-prefix matcher with the operator-picked MODEL, mis-routing a
 # lead workspace through MiniMax even though its CLAUDE_CODE_OAUTH_TOKEN
 # was clearly meant to be used.
 #
 # Maintain this list in sync with the persona env file convention:
 #   - ``claude-code``  → ``anthropic-oauth`` (Claude Code subscription path)
 #   - ``anthropic``    → ``anthropic-api``  (direct Anthropic API key)
 # Provider names already in the registry alias to themselves implicitly
 # (the ``in registry`` check catches them before this map is consulted).
 _PROVIDER_SLUG_ALIASES = {
    "claude-code": "anthropic-oauth",
    "anthropic": "anthropic-api",
 }
 def _resolve_model_and_provider_from_env(
    yaml_model: str,
    yaml_provider: str,
    providers: tuple,
 ) -> tuple:
    """Reconcile model + provider from env vars vs YAML, with the persona-env
    convention winning over the legacy ``MODEL_PROVIDER``-as-model-id usage.
    The persona env files (``~/.molecule-ai/personas/<name>/env`` on the host,
    sourced into each workspace container at provision time) declare TWO env
    vars with distinct semantics:
      * ``MODEL`` — the model id (e.g. ``MiniMax-M2.7-highspeed``, ``opus``).
      * ``MODEL_PROVIDER`` — the provider slug (e.g. ``minimax``,
        ``claude-code``, ``anthropic``).
    The legacy ``workspace/config.py`` (in molecule-ai-workspace-runtime)
    historically interpreted ``MODEL_PROVIDER`` as the *model id* — a name
    chosen before there was a separate ``MODEL`` env var. When both env vars
    are set with the persona convention, the legacy code reads
    ``MODEL_PROVIDER=minimax`` into ``runtime_config.model``, which then
    fails to match any registry prefix (``minimax-`` requires a hyphen
    suffix) and silently falls through to providers[0] (``anthropic-oauth``).
    OAuth-token-less workspaces then wedge at ``query.initialize()`` because
    the claude CLI can't authenticate. This is the 2026-05-08 dev-tree
    incident — 22/27 non-lead workspaces stuck in ``degraded``.
    Resolution order (this function):
      1. ``MODEL`` env var → picked_model. Authoritative when set; the
         persona env always sets it alongside ``MODEL_PROVIDER`` so the
         model id never has to be inferred.
      2. ``MODEL_PROVIDER`` env var → explicit_provider, BUT only when the
         value matches a known provider name in the registry. This guards
         against the legacy case where some callers still set
         ``MODEL_PROVIDER`` to a model id (e.g. canvas Save+Restart prior to
         this fix). If the value isn't a registered provider name and YAML
         didn't supply a model, treat it as a model id for back-compat.
      3. YAML ``runtime_config.model`` / ``provider`` — used for any field
         the env didn't supply. Carries the operator's canvas selection
         on workspaces that haven't yet adopted the persona env shape.
    Returns ``(picked_model, explicit_provider_name)``. Either may be
    empty/None — the caller (``setup``) handles the empty cases via
    ``_resolve_provider``'s registry fallback.
    """
    env_model = (os.environ.get("MODEL") or "").strip()
    env_provider = (os.environ.get("MODEL_PROVIDER") or "").strip()
    provider_names_lower = {p.get("name", "").lower() for p in providers}
    # Detect whether MODEL_PROVIDER carries the persona-convention slug
    # (provider name) vs. the legacy convention (model id). Persona-
    # convention wins when the value matches a registered provider; we
    # fall back to legacy interpretation only when it doesn't.
    #
    # First, apply the alias map so persona-friendly slugs like
    # ``claude-code`` resolve to the canonical registry name
    # ``anthropic-oauth``. Without this, a lead workspace's
    # ``MODEL_PROVIDER=claude-code`` env would fall through to the model-
    # prefix matcher, see ``MODEL=MiniMax-M2.7`` and mis-route to MiniMax
    # even though the operator's intent (and the OAuth token they set)
    # was the OAuth subscription path.
    env_provider_resolved = _PROVIDER_SLUG_ALIASES.get(
        env_provider.lower(), env_provider,
    ) if env_provider else ""
    env_provider_is_slug = (
        bool(env_provider_resolved)
        and env_provider_resolved.lower() in provider_names_lower
    )
    # Picked model resolution
    if env_model:
        picked_model = env_model
    elif env_provider and not env_provider_is_slug:
        # Legacy: MODEL_PROVIDER env carried the model id. Honor it so
        # canvas Save+Restart workflows that predate this fix keep working.
        picked_model = env_provider
    else:
        picked_model = yaml_model or ""
    # Explicit provider resolution — env wins when it's a registered slug
    # (after alias mapping), otherwise fall back to YAML.
    #
    # YAML aliasing: the molecule-runtime wheel (config.py) auto-derives
    # ``runtime_config.provider`` from the YAML/default model slug — the
    # default model ``anthropic:claude-opus-4-7`` yields ``anthropic`` as
    # the inferred provider. Without applying the alias map here, that
    # auto-derived ``anthropic`` slug fails registry lookup and the
    # adapter raises ValueError ("provider='anthropic' but it is not in
    # the providers registry"), wedging the workspace at boot. The alias
    # map already handles this for the env-var path above; mirror the
    # same treatment for the YAML path so the runtime-wheel default
    # produces a registered provider name in both cases. Caught
    # 2026-05-09 on staging-cplead-2 — every workspace booted with
    # ``configuration_status=not_configured`` because the YAML provider
    # ``anthropic`` was passed through verbatim instead of being aliased
    # to ``anthropic-api``.
    if env_provider_is_slug:
        explicit_provider = env_provider_resolved
    elif yaml_provider:
        yp_lower = yaml_provider.lower()
        explicit_provider = _PROVIDER_SLUG_ALIASES.get(yp_lower, yaml_provider)
    else:
        explicit_provider = None
    return picked_model, explicit_provider
 def _strip_provider_prefix(model: str) -> str:
    """Strip LangChain-style "<provider>:<model>" prefix from a model id.
@ -236,12 +460,18 @@ _VENDOR_KEY_NAMES = frozenset({
 def _project_vendor_auth(provider: dict) -> None:
-    """Project a per-vendor API key onto ANTHROPIC_AUTH_TOKEN at boot.
+    """Project a per-vendor API key onto the provider's auth-token env at boot.
    Third-party Anthropic-compat providers (MiniMax, Z.ai, DeepSeek)
    reuse the Anthropic SDK's wire format with a Bearer token, which the
    ``claude`` CLI / claude-code-sdk reads from ``ANTHROPIC_AUTH_TOKEN``.
    Kimi For Coding's gateway instead authenticates with the
    ``x-api-key`` header (per kimi.com's official Claude Code
    integration doc), which the SDK emits from ``ANTHROPIC_API_KEY`` —
    so the projection target is per-provider, declared as
    ``auth_token_env`` in the registry (default ``ANTHROPIC_AUTH_TOKEN``
    preserves the existing MiniMax/GLM/DeepSeek behavior unchanged).
    Third-party Anthropic-compat providers (MiniMax, Z.ai, Moonshot,
    DeepSeek) all reuse the Anthropic SDK's wire format, which means the
    ``claude`` CLI / claude-code-sdk reads the bearer token from
    ``ANTHROPIC_AUTH_TOKEN`` no matter which vendor is being talked to.
    Pre-#244 the canvas surfaced the vendor-specific name
    (``MINIMAX_API_KEY``, etc.) to the user — so a user who saved only
    that name hit a silent 401 on first call while the boot audit said
@ -249,21 +479,24 @@ def _project_vendor_auth(provider: dict) -> None:
    / hermes PR #38.
    Behavior:
      * Let ``target`` = the provider's ``auth_token_env`` (default
        ``ANTHROPIC_AUTH_TOKEN``).
      * If the matched provider's ``auth_env`` lists any of
        ``_VENDOR_KEY_NAMES`` and that var is set, copy its value into
-        ``ANTHROPIC_AUTH_TOKEN`` so the SDK finds it.
+        ``target`` so the SDK finds it.
-      * **Idempotent**: if ``ANTHROPIC_AUTH_TOKEN`` is already set we
+      * **Idempotent**: if ``target`` is already set we do NOT
-        do NOT overwrite — an explicit operator value (workspace
+        overwrite — an explicit operator value (workspace secret)
-        secret) always wins over auto-projection.
+        always wins over auto-projection.
-      * Logs the projection by NAME (e.g. ``MINIMAX_API_KEY ->
+      * Logs the projection by NAME (e.g. ``KIMI_API_KEY ->
-        ANTHROPIC_AUTH_TOKEN``); never logs the secret VALUE. Same
+        ANTHROPIC_API_KEY``); never logs the secret VALUE. Same
        contract as ``_audit_auth_env_presence``.
      * No-op for providers whose ``auth_env`` doesn't reference a
        vendor-specific name (oauth, anthropic-api, or a third-party
        entry that hasn't been added to the registry yet).
    """
    auth_env = provider.get("auth_env") or ()
-    if os.environ.get("ANTHROPIC_AUTH_TOKEN"):
+    target = provider.get("auth_token_env") or "ANTHROPIC_AUTH_TOKEN"
    if os.environ.get(target):
        # Operator override wins — never clobber an explicit value.
        return
    for name in auth_env:
@ -272,21 +505,36 @@ def _project_vendor_auth(provider: dict) -> None:
        value = os.environ.get(name)
        if not value:
            continue
-        os.environ["ANTHROPIC_AUTH_TOKEN"] = value
+        os.environ[target] = value
        logger.info(
-            "auth env projection: %s -> ANTHROPIC_AUTH_TOKEN (provider=%s)",
+            "auth env projection: %s -> %s (provider=%s)",
-            name, provider.get("name", "<unknown>"),
+            name, target, provider.get("name", "<unknown>"),
        )
        return
-def _resolve_provider(model: str, providers: tuple) -> dict:
+def _resolve_provider(
    model: str,
    providers: tuple,
    explicit_provider: str = None,
 ) -> dict:
    """Return the provider entry matching this model id.
-    Match is case-insensitive: prefix wins over alias when both could
+    If ``explicit_provider`` is given (set via the ``provider:`` field in
-    apply. Unknown ids fall back to the first provider in the registry
+    workspace config.yaml or runtime_config), look up by name first. If the
-    (by convention, the OAuth/safest default — anthropic-oauth in both
+    named provider is not in the registry, RAISE ``ValueError`` with an
-    _BUILTIN_PROVIDERS and the shipped config.yaml).
+    actionable message — silent fallback to ``providers[0]`` is the bug
    that motivated #180 (workspace operator picks ``provider: minimax``
    in the canvas Config tab, the adapter ignores it, the Claude SDK
    silently keeps using ``CLAUDE_CODE_OAUTH_TOKEN`` and the operator has
    no way to tell from the canvas that their provider switch did
    nothing).
    Without an explicit name: match is case-insensitive, prefix wins over
    alias when both could apply, and unknown ids fall back to the first
    provider in the registry (by convention, the OAuth/safest default —
    ``anthropic-oauth`` in both _BUILTIN_PROVIDERS and the shipped
    config.yaml).
    Pre-condition: ``providers`` is non-empty. _load_providers always
    returns at least one entry (built-ins when YAML is missing or every
@ -298,6 +546,44 @@ def _resolve_provider(model: str, providers: tuple) -> dict:
            "_load_providers must always return at least one entry "
            "(falling back to _BUILTIN_PROVIDERS when needed)"
        )
    # Explicit provider name takes precedence — fail fast if it's not in
    # the registry. Anything else would silently route the operator's
    # picked provider through the wrong auth/base_url path. The error
    # message tells them exactly which two paths fix it.
    if explicit_provider:
        ep_lower = explicit_provider.lower()
        for provider in providers:
            if provider["name"].lower() == ep_lower:
                return provider
        names = ", ".join(p["name"] for p in providers)
        raise ValueError(
            f"claude-code adapter: workspace config picks "
            f"provider='{explicit_provider}' but it is not in the "
            f"providers registry.\n"
            f"\n"
            f"Known providers: {names}\n"
            f"\n"
            f"Two ways to fix:\n"
            f"  (a) Add '{explicit_provider}' to /configs/config.yaml as a "
            f"providers: entry. Required keys:\n"
            f"        providers:\n"
            f"          - name: {explicit_provider}\n"
            f"            auth_mode: third_party_anthropic_compat\n"
            f"            base_url: https://...   # provider's Anthropic-compat endpoint\n"
            f"            auth_env: [{explicit_provider.upper()}_API_KEY]\n"
            f"            model_prefixes: [...]\n"
            f"  (b) Switch the workspace runtime template to one that "
            f"natively supports {explicit_provider} (CrewAI, LangGraph, or "
            f"DeepAgents read provider/model from runtime_config and route "
            f"directly without needing an Anthropic-compat shim).\n"
            f"\n"
            f"Note: claude-code SDK speaks the Anthropic API protocol. "
            f"Providers that only expose OpenAI-compatible endpoints "
            f"(MiniMax, GLM, Kimi, DeepSeek native APIs) need either an "
            f"Anthropic-compat proxy in front, or option (b)."
        )
    if not model:
        return providers[0]
    m = model.lower()
@ -400,9 +686,52 @@ class ClaudeCodeAdapter(BaseAdapter):
        # validation + ANTHROPIC_BASE_URL routing from that single decision.
        rc = config.runtime_config
        if isinstance(rc, dict):
-            picked_model = rc.get("model") or "sonnet"
+            yaml_model = rc.get("model") or ""
            yaml_provider_name = rc.get("provider") or ""
        else:
-            picked_model = getattr(rc, "model", None) or "sonnet"
+            yaml_model = getattr(rc, "model", None) or ""
            yaml_provider_name = getattr(rc, "provider", None) or ""
        # Also honor the top-level `provider:` field in /configs/config.yaml.
        # The canvas Config-tab Provider dropdown writes there (not into
        # runtime_config) on some legacy paths. Either source is canonical;
        # whichever is set wins. Root cause of #180: the adapter used to
        # ignore both, silently routing every non-Anthropic provider pick
        # through anthropic-oauth.
        if not yaml_provider_name:
            yaml_path = os.path.join(config.config_path, "config.yaml")
            try:
                import yaml  # transitive dep via molecule-ai-workspace-runtime
                with open(yaml_path, "r") as f:
                    data = yaml.safe_load(f) or {}
                if isinstance(data, dict):
                    val = data.get("provider")
                    if isinstance(val, str) and val.strip():
                        yaml_provider_name = val.strip()
            except FileNotFoundError:
                pass
            except Exception as exc:  # noqa: BLE001 — defensive: never block boot
                logger.warning(
                    "providers: failed to read top-level provider: from %s (%s); "
                    "falling back to model-based resolution",
                    yaml_path, exc,
                )
        # Reconcile env vars (persona convention: MODEL=<id>,
        # MODEL_PROVIDER=<slug>) against YAML. Env wins over YAML — the
        # persona env files are the canonical per-agent provider mapping
        # (Phase 2 mapping 2026-05-08), and the workspace-runtime wheel's
        # legacy ``MODEL_PROVIDER``-as-model-id reading would otherwise
        # silently route non-leads to providers[0] = anthropic-oauth.
        # Documented in detail at _resolve_model_and_provider_from_env.
        picked_model, explicit_provider_name = _resolve_model_and_provider_from_env(
            yaml_model=yaml_model,
            yaml_provider=yaml_provider_name,
            providers=providers,
        )
        if not picked_model:
            picked_model = "sonnet"
        # NOTE: do NOT strip the provider prefix here. The pre-fix routing
        # behavior — `anthropic:claude-opus-4-7` falls through to
        # providers[0] (anthropic-oauth) when no model_prefixes match — is
@ -411,7 +740,15 @@ class ClaudeCodeAdapter(BaseAdapter):
        # `anthropic-api` provider and the CLI then hangs at `initialize`
        # because ANTHROPIC_API_KEY isn't set. The strip belongs only at
        # the CLI invocation site (create_executor below).
-        provider = _resolve_provider(picked_model, providers)
+        #
        # Pass the explicit provider name through so _resolve_provider
        # raises ValueError with an actionable message (instead of silently
        # routing to providers[0]) when an operator picks a provider that
        # isn't in the registry. See #180.
        provider = _resolve_provider(
            picked_model, providers,
            explicit_provider=explicit_provider_name,
        )
        auth_env_options = provider["auth_env"]
        # Project the per-vendor API key (MINIMAX_API_KEY, GLM_API_KEY,
@ -522,9 +859,26 @@ class ClaudeCodeAdapter(BaseAdapter):
        # RuntimeConfig dataclass. Read `model` defensively from either shape.
        rc = config.runtime_config
        if isinstance(rc, dict):
-            explicit_model = rc.get("model") or ""
+            yaml_model = rc.get("model") or ""
            yaml_provider = rc.get("provider") or ""
        else:
-            explicit_model = getattr(rc, "model", None) or ""
+            yaml_model = getattr(rc, "model", None) or ""
            yaml_provider = getattr(rc, "provider", None) or ""
        # Reconcile against env vars (persona convention: MODEL=<id>,
        # MODEL_PROVIDER=<slug>) using the same helper that ``setup`` uses,
        # so the executor and the boot banner agree on the picked model.
        # Without this, a workspace whose env says ``MODEL=MiniMax-M2.7``
        # but whose runtime wheel pre-dates the persona-env fix would set
        # runtime_config.model="minimax" (the slug, mistakenly read by the
        # legacy ``MODEL_PROVIDER``-as-model-id path); this helper restores
        # the correct model id before it reaches the SDK.
        providers = _load_providers(config.config_path)
        explicit_model, _ = _resolve_model_and_provider_from_env(
            yaml_model=yaml_model,
            yaml_provider=yaml_provider,
            providers=providers,
        )
        explicit_model = _strip_provider_prefix(explicit_model)
        # Pre-validation: detect the misconfiguration combo that drove the
@ -555,7 +909,7 @@ class ClaudeCodeAdapter(BaseAdapter):
                    "The default fallback ('sonnet') is an Anthropic-native "
                    "alias; non-Anthropic shims (MiniMax, OpenAI gateways, "
                    "etc.) won't recognize it and the SDK --print probe will "
-                    "hang for 30s before timing out. Fix: set MODEL_PROVIDER "
+                    "hang for 30s before timing out. Fix: set MODEL "
                    "as a workspace secret (canvas: Save+Restart with model "
                    "picked) or set runtime_config.model in /configs/config.yaml."
                )
--- a/config.yaml
+++ b/config.yaml
@ -31,6 +31,16 @@ tier: 2
 #   model_aliases  : exact lowercase ids (e.g. ["sonnet", "opus"])
 #   base_url       : ANTHROPIC_BASE_URL to set; null = CLI default (anthropic-native)
 #   auth_env       : env vars accepted; any one being set satisfies auth
 #   auth_token_env : (optional) the env var the boot-time vendor-key
 #                    projection writes the vendor key INTO. Defaults to
 #                    ANTHROPIC_AUTH_TOKEN (Bearer-style; correct for
 #                    MiniMax/GLM/DeepSeek Anthropic-compat shims). Kimi
 #                    For Coding's gateway authenticates with the
 #                    x-api-key header per kimi.com's official Claude Code
 #                    integration doc, which the Anthropic SDK / claude
 #                    CLI emits from ANTHROPIC_API_KEY (NOT the Bearer
 #                    ANTHROPIC_AUTH_TOKEN) — so its entry sets
 #                    auth_token_env: ANTHROPIC_API_KEY.
 providers:
  - name: anthropic-oauth
    auth_mode: oauth
@ -73,13 +83,27 @@ providers:
    base_url: https://api.z.ai/api/anthropic
    auth_env: [GLM_API_KEY, ANTHROPIC_AUTH_TOKEN, ANTHROPIC_API_KEY]
-  # Moonshot AI — Kimi family. platform.kimi.ai/docs/guide/agent-support.
+  # Kimi For Coding — Moonshot's coding-agent tier (K2.6 / "Kimi for
-  - name: moonshot
+  # Coding"). Per kimi.com's OFFICIAL Claude Code integration doc
  # (kimi.com/code/docs/en/third-party-tools/other-coding-agents.html,
  # "Claude Code" section) the contract is:
  #   ANTHROPIC_BASE_URL=https://api.kimi.com/coding/   (trailing slash)
  #   ANTHROPIC_API_KEY=<the Kimi key>                  (x-api-key header)
  # The `sk-kimi-*` key (KIMI_API_KEY in SSOT) authenticates ONLY against
  # this gateway — the legacy api.moonshot.ai/anthropic surface 401s it.
  # The gateway routes to the served K2.6 model regardless of the Claude
  # model name on the wire (proven end-to-end via the OpenClaw template's
  # api.kimi.com/coding path, winnerProvider=custom-api-kimi-com).
  # auth_token_env pins the projection to ANTHROPIC_API_KEY (x-api-key)
  # rather than the default ANTHROPIC_AUTH_TOKEN (Bearer), which this
  # gateway rejects.
  - name: kimi-coding
    auth_mode: third_party_anthropic_compat
    model_prefixes: [kimi-]
    model_aliases: []
-    base_url: https://api.moonshot.ai/anthropic
+    base_url: https://api.kimi.com/coding/
-    auth_env: [KIMI_API_KEY, ANTHROPIC_AUTH_TOKEN, ANTHROPIC_API_KEY]
+    auth_env: [KIMI_API_KEY, ANTHROPIC_API_KEY, ANTHROPIC_AUTH_TOKEN]
    auth_token_env: ANTHROPIC_API_KEY
  # DeepSeek — api-docs.deepseek.com/guides/anthropic_api. Note: their
  # endpoint silently maps unknown model ids to deepseek-v4-flash, so a
@ -175,15 +199,23 @@ runtime_config:
      name: Z.ai GLM-4.5 (third-party, Anthropic-API-compatible)
      required_env: [GLM_API_KEY]
-    # --- Moonshot AI Kimi family (third-party, Anthropic-API-compatible) ---
+    # --- Kimi For Coding (third-party, Anthropic-API-compatible) ---
-    # KIMI_API_KEY → ANTHROPIC_AUTH_TOKEN projection at boot.
+    # Routed via the `kimi-coding` provider entry above: the adapter
-    # platform.kimi.ai for docs. K2.5 is the latest agentic-coding tier;
+    # auto-sets ANTHROPIC_BASE_URL=https://api.kimi.com/coding/ and
-    # K2 stays as a cheaper option.
+    # projects KIMI_API_KEY → ANTHROPIC_API_KEY (x-api-key) per
    # kimi.com's official Claude Code integration doc. The gateway
    # serves the K2.6 model regardless of the wire model id; the id
    # below is the gateway's own served-model name (mirrors the proven
    # OpenClaw `kimi-for-coding` route). K2.5 / K2 stay as aliases for
    # workspaces pinned to the older labels — they hit the same gateway.
    - id: kimi-for-coding
      name: Kimi K2.6 (Kimi For Coding, third-party Anthropic-API-compatible)
      required_env: [KIMI_API_KEY]
    - id: kimi-k2.5
-      name: Moonshot Kimi K2.5 (third-party, Anthropic-API-compatible)
+      name: Kimi K2.5 (Kimi For Coding, third-party Anthropic-API-compatible)
      required_env: [KIMI_API_KEY]
    - id: kimi-k2
-      name: Moonshot Kimi K2 (third-party, Anthropic-API-compatible)
+      name: Kimi K2 (Kimi For Coding, third-party Anthropic-API-compatible)
      required_env: [KIMI_API_KEY]
    # --- DeepSeek (third-party, Anthropic-API-compatible) ---
--- a/entrypoint.sh
+++ b/entrypoint.sh
@ -42,6 +42,15 @@ log_boot_context
 if [ "$(id -u)" = "0" ]; then
    # Configs volume is created by Docker as root; agent needs write access
    # for plugin installs, memory writes, .auth_token rotation, etc.
    #
    # T4 atomic-co-sequencing contract (RFC internal#456 §10): the T4
    # escalation leg (sudo NOPASSWD + docker group, baked in the
    # Dockerfile) is ADDITIVE. The agent still runs uid-1000 and
    # /configs/.auth_token MUST remain agent-owned — escalation must
    # NOT regress the Hermes list_peers-401 token-ownership class.
    # This chown -R is the agent-ownership half of that contract; the
    # Layer-3 conformance gate asserts owner_uid==1000 on the running
    # container alongside the host-root-reach assertion.
    chown -R agent:agent /configs 2>/dev/null
    # /workspace handling — only chown when the contents are root-owned
    # (typical on Docker Desktop on Windows where host uid maps to 0).
@ -70,9 +79,36 @@ if [ "$(id -u)" = "0" ]; then
    # finds it when running as agent. The provisioner's mount point is
    # hardcoded to /root/.claude/sessions; we don't want to change the
    # platform contract just for this template.
-    mkdir -p /home/agent/.claude
+    #
    # NOTE (T4 perms regression): on FIRST boot the host volume mount for
    # /home/agent/.claude doesn't exist yet — entrypoint creates it and
    # the chown lands inside the `if -d /root/.claude/sessions` guard.
    # On SECOND boot with a populated /home/agent/.claude (sessions/,
    # session-env/, settings.json — any of which the SDK or agent has
    # written between boots) the dir may already be root-owned because
    # the SDK's working files inherited root's uid when written under
    # the prior root segment of an earlier entrypoint, OR because a
    # newer claude-code release writes new subdirs we don't create here.
    # That leaves uid-1000 agent EPERMing on every settings/session write
    # ("permission restrictions" surfaced to the canvas as a generic
    # Bash failure). Fix: create the well-known subdirs idempotently
    # and run the chown unconditionally (no-op when ownership is already
    # correct, fast on small trees). Stub ~/.claude/settings.json too so
    # the agent's introspection (cat ~/.claude/settings.json) succeeds
    # and shows operating mode — bypassPermissions is the canonical
    # mode set programmatically by claude_sdk_executor.py.
    mkdir -p /home/agent/.claude/sessions /home/agent/.claude/session-env
    if [ ! -f /home/agent/.claude/settings.json ]; then
        cat > /home/agent/.claude/settings.json <<'EOF'
 {
  "permissions": {"defaultMode": "bypassPermissions"},
  "_note": "Mode is also set programmatically by claude_sdk_executor.py (permission_mode='bypassPermissions'); this file is informational and lets `cat ~/.claude/settings.json` succeed."
 }
 EOF
    fi
    chown -R agent:agent /home/agent/.claude 2>/dev/null
    if [ -d /root/.claude/sessions ]; then
-        chown -R agent:agent /root/.claude /home/agent/.claude 2>/dev/null
+        chown -R agent:agent /root/.claude 2>/dev/null
        ln -sfn /root/.claude/sessions /home/agent/.claude/sessions
    fi
--- a/runbooks/local-dev-setup.md
+++ b/runbooks/local-dev-setup.md
@ -24,7 +24,7 @@ common problems.
 ## Step 1 — Clone the Repository
 ```bash
-git clone https://github.com/Molecule-AI/molecule-ai-workspace-template-claude-code.git
+git clone https://git.moleculesai.app/molecule-ai/molecule-ai-workspace-template-claude-code.git
 cd molecule-ai-workspace-template-claude-code
 ```
--- a/tests/conftest.py
+++ b/tests/conftest.py
@ -0,0 +1,89 @@
 """Shared pytest fixtures + import shims for the adapter test suite.
 `adapter.py` imports at module load:
  - molecule_runtime.adapters.base (BaseAdapter, AdapterConfig, RuntimeCapabilities)
  - molecule_runtime.plugins (lazy in setup(), but stubbed proactively)
  - a2a.server.agent_execution (AgentExecutor)
  - claude_sdk_executor (lazy in create_executor(), stubbed proactively)
 In production those arrive transitively via molecule-ai-workspace-runtime.
 The CI runner only installs `pytest pytest-asyncio pyyaml`, so the import
 chain would fail with ModuleNotFoundError before any test collects —
 exactly the failure that broke CI on the #180 fix branch (PR #4) and
 caused the merge wall to block on a green local but red Gitea CI.
 Putting the stub installer here (collected before any test module is
 imported, per pytest semantics) means every test file can do
 `from adapter import ...` at module top without a per-file boilerplate
 copy. It also forces a single shape for the stubs so two files can't
 silently disagree on whether `BaseAdapter` has
 `install_plugins_via_registry` (see test_adapter_prevalidate's
 async-setup tests, which need the method to exist on the parent class).
 """
 import os
 import sys
 import types
 from dataclasses import dataclass
 from unittest.mock import MagicMock
@dataclass
 class _StubRuntimeCapabilities:
    provides_native_session: bool = False
@dataclass
 class _StubAdapterConfig:
    runtime_config: object = None
    config_path: str = "/tmp/configs"
    system_prompt: str = ""
    heartbeat: object = None
 class _StubBaseAdapter:
    async def install_plugins_via_registry(self, *_args, **_kwargs):
        pass
 def _install_stubs() -> None:
    """Install the smallest set of import shims that adapter.py needs."""
    if "molecule_runtime" not in sys.modules:
        mr = types.ModuleType("molecule_runtime")
        mr.adapters = types.ModuleType("molecule_runtime.adapters")
        mr.adapters.base = types.ModuleType("molecule_runtime.adapters.base")
        mr.adapters.base.BaseAdapter = _StubBaseAdapter
        mr.adapters.base.AdapterConfig = _StubAdapterConfig
        mr.adapters.base.RuntimeCapabilities = _StubRuntimeCapabilities
        mr.plugins = types.ModuleType("molecule_runtime.plugins")
        mr.plugins.load_plugins = lambda **_kwargs: []
        sys.modules["molecule_runtime"] = mr
        sys.modules["molecule_runtime.adapters"] = mr.adapters
        sys.modules["molecule_runtime.adapters.base"] = mr.adapters.base
        sys.modules["molecule_runtime.plugins"] = mr.plugins
    if "a2a" not in sys.modules:
        a2a = types.ModuleType("a2a")
        a2a.server = types.ModuleType("a2a.server")
        a2a.server.agent_execution = types.ModuleType("a2a.server.agent_execution")
        a2a.server.agent_execution.AgentExecutor = type("AgentExecutor", (), {})
        sys.modules["a2a"] = a2a
        sys.modules["a2a.server"] = a2a.server
        sys.modules["a2a.server.agent_execution"] = a2a.server.agent_execution
    if "claude_sdk_executor" not in sys.modules:
        mod = types.ModuleType("claude_sdk_executor")
        mod.ClaudeSDKExecutor = MagicMock(name="ClaudeSDKExecutor")
        sys.modules["claude_sdk_executor"] = mod
 # Run at conftest import time — pytest collects conftest.py before any
 # test module, so the stubs are in sys.modules before `from adapter
 # import ...` ever executes.
 _install_stubs()
 # adapter.py lives in the parent dir of tests/ (template root). pytest's
 # `--import-mode=importlib` + tests/pytest.ini anchoring rootdir at
 # tests/ means the parent isn't on sys.path automatically. Add it here
 # once so every test file can do `from adapter import ...` cleanly.
 _PARENT_DIR = os.path.dirname(os.path.dirname(os.path.abspath(__file__)))
 if _PARENT_DIR not in sys.path:
    sys.path.insert(0, _PARENT_DIR)
--- a/tests/test_adapter_prevalidate.py
+++ b/tests/test_adapter_prevalidate.py
@ -129,12 +129,13 @@ _FIXTURE_PROVIDERS_YAML = textwrap.dedent("""
        base_url: https://api.z.ai/api/anthropic
        auth_env: [ANTHROPIC_AUTH_TOKEN, ANTHROPIC_API_KEY]
-      - name: moonshot
+      - name: kimi-coding
        auth_mode: third_party_anthropic_compat
        model_prefixes: [kimi-]
        model_aliases: []
-        base_url: https://api.moonshot.ai/anthropic
+        base_url: https://api.kimi.com/coding/
-        auth_env: [ANTHROPIC_AUTH_TOKEN, ANTHROPIC_API_KEY]
+        auth_env: [KIMI_API_KEY, ANTHROPIC_API_KEY, ANTHROPIC_AUTH_TOKEN]
        auth_token_env: ANTHROPIC_API_KEY
      - name: deepseek
        auth_mode: third_party_anthropic_compat
@ -514,8 +515,15 @@ async def test_setup_auth_token_alone_satisfies_third_party_check(
 # ---- _load_providers / _resolve_provider unit tests ----
-def test_load_providers_returns_builtin_when_yaml_missing(tmp_path):
+def test_load_providers_returns_builtin_when_yaml_missing(tmp_path, monkeypatch):
-    """FileNotFoundError path returns the in-code defaults verbatim."""
+    """FileNotFoundError path returns the in-code defaults verbatim.
    Monkeypatches the canonical + template paths to a non-existent dir
    so only the workspace config_path is in scope. Without this, the
    multi-path lookup picks up the repo-root config.yaml that ships
    with the template (path 2 finds the bundled providers list and
    returns it instead of falling through to builtins).
    """
    _install_stubs()
    parent_dir = os.path.dirname(os.path.dirname(os.path.abspath(__file__)))
    if parent_dir not in sys.path:
@ -523,6 +531,10 @@ def test_load_providers_returns_builtin_when_yaml_missing(tmp_path):
    sys.modules.pop("adapter", None)
    import adapter as adapter_module
    nonexistent = str(tmp_path / "_isolate_canonical")
    monkeypatch.setattr(adapter_module, "_CANONICAL_ADAPTER_DIR", nonexistent)
    monkeypatch.setattr(adapter_module, "_TEMPLATE_DIR", nonexistent)
    result = adapter_module._load_providers(str(tmp_path))
    assert result == adapter_module._BUILTIN_PROVIDERS
@ -543,7 +555,7 @@ def test_load_providers_parses_yaml_and_normalizes(tmp_path):
    names = [p["name"] for p in result]
    assert names == [
        "anthropic-oauth", "anthropic-api", "xiaomi-mimo", "minimax",
-        "zai", "moonshot", "deepseek",
+        "zai", "kimi-coding", "deepseek",
    ]
    # YAML lists must be normalized to tuples for downstream lookup ergonomics.
    assert isinstance(result[0]["model_aliases"], tuple)
@ -553,15 +565,16 @@ def test_load_providers_parses_yaml_and_normalizes(tmp_path):
@pytest.mark.parametrize("model,expected_provider,expected_url", [
    ("GLM-4.6", "zai", "https://api.z.ai/api/anthropic"),
    ("glm-4.5", "zai", "https://api.z.ai/api/anthropic"),
-    ("kimi-k2.5", "moonshot", "https://api.moonshot.ai/anthropic"),
+    ("kimi-k2.5", "kimi-coding", "https://api.kimi.com/coding/"),
    ("kimi-for-coding", "kimi-coding", "https://api.kimi.com/coding/"),
    ("deepseek-v4-pro", "deepseek", "https://api.deepseek.com/anthropic"),
 ])
@pytest.mark.asyncio
 async def test_setup_routes_extra_providers(
    adapter, monkeypatch, configs_dir, model, expected_provider, expected_url
 ):
-    """The Z.ai / Moonshot / DeepSeek providers added in this PR must
+    """The Z.ai / Kimi-For-Coding / DeepSeek providers must route
-    route correctly: model id → provider entry → ANTHROPIC_BASE_URL.
+    correctly: model id → provider entry → ANTHROPIC_BASE_URL.
    Parametrized to keep the matrix coverage tight without 3 near-identical
    test bodies. Locks in the per-vendor base_url so a future YAML edit
    that mistypes z.ai's `/api/anthropic` suffix gets caught.
@ -576,8 +589,12 @@ async def test_setup_routes_extra_providers(
    assert os.environ.get("ANTHROPIC_BASE_URL") == expected_url
-def test_load_providers_falls_back_on_malformed_yaml(tmp_path, caplog):
+def test_load_providers_falls_back_on_malformed_yaml(tmp_path, caplog, monkeypatch):
-    """Malformed YAML → log warning + fallback (don't kill boot)."""
+    """Malformed YAML → log warning + fallback (don't kill boot).
    Isolated from the multi-path lookup by pinning canonical + template
    dirs at a non-existent path; only the workspace config_path is read.
    """
    _install_stubs()
    parent_dir = os.path.dirname(os.path.dirname(os.path.abspath(__file__)))
    if parent_dir not in sys.path:
@ -585,6 +602,10 @@ def test_load_providers_falls_back_on_malformed_yaml(tmp_path, caplog):
    sys.modules.pop("adapter", None)
    import adapter as adapter_module
    nonexistent = str(tmp_path / "_isolate_canonical")
    monkeypatch.setattr(adapter_module, "_CANONICAL_ADAPTER_DIR", nonexistent)
    monkeypatch.setattr(adapter_module, "_TEMPLATE_DIR", nonexistent)
    (tmp_path / "config.yaml").write_text("providers: [not valid yaml: {{{")
    import logging
@ -622,7 +643,7 @@ def test_resolve_provider_minimax_prefix_matches_minimax_provider():
    assert result2["name"] == "minimax"
-def test_load_providers_drops_bad_entry_keeps_rest(tmp_path, caplog):
+def test_load_providers_drops_bad_entry_keeps_rest(tmp_path, caplog, monkeypatch):
    """Per-entry isolation: one malformed entry shouldn't nuke the registry.
    Pre-fix: ``_load_providers`` built the registry via a generator inside
@ -634,6 +655,9 @@ def test_load_providers_drops_bad_entry_keeps_rest(tmp_path, caplog):
    Post-fix: per-entry try/except drops the bad entry with a warning,
    rest of the registry survives.
    Isolated from the multi-path lookup so only the test's tmp config.yaml
    is read.
    """
    _install_stubs()
    parent_dir = os.path.dirname(os.path.dirname(os.path.abspath(__file__)))
@ -642,6 +666,10 @@ def test_load_providers_drops_bad_entry_keeps_rest(tmp_path, caplog):
    sys.modules.pop("adapter", None)
    import adapter as adapter_module
    nonexistent = str(tmp_path / "_isolate_canonical")
    monkeypatch.setattr(adapter_module, "_CANONICAL_ADAPTER_DIR", nonexistent)
    monkeypatch.setattr(adapter_module, "_TEMPLATE_DIR", nonexistent)
    yaml_with_typo = textwrap.dedent("""
        providers:
          - name: good-zai
@ -690,7 +718,7 @@ def test_load_providers_drops_bad_entry_keeps_rest(tmp_path, caplog):
    )
-def test_load_providers_string_as_prefix_does_not_split_into_chars(tmp_path, caplog):
+def test_load_providers_string_as_prefix_does_not_split_into_chars(tmp_path, caplog, monkeypatch):
    """A YAML field declared as list-of-strings but written as a bare
    string (operator forgot brackets) used to silently iterate over
    characters → ``('m','i','m','o','-')``. Post-fix: non-list value
@ -705,6 +733,10 @@ def test_load_providers_string_as_prefix_does_not_split_into_chars(tmp_path, cap
    sys.modules.pop("adapter", None)
    import adapter as adapter_module
    nonexistent = str(tmp_path / "_isolate_canonical")
    monkeypatch.setattr(adapter_module, "_CANONICAL_ADAPTER_DIR", nonexistent)
    monkeypatch.setattr(adapter_module, "_TEMPLATE_DIR", nonexistent)
    yaml_str_prefix = textwrap.dedent("""
        providers:
          - name: typo-prefix
@ -723,7 +755,7 @@ def test_load_providers_string_as_prefix_does_not_split_into_chars(tmp_path, cap
    )
-def test_load_providers_drops_entry_without_name(tmp_path, caplog):
+def test_load_providers_drops_entry_without_name(tmp_path, caplog, monkeypatch):
    """An entry without ``name`` is operator error — no silent fallback
    to ``<unnamed>``. Drop the entry with a warning so the boot log
    surfaces the typo.
@ -735,6 +767,10 @@ def test_load_providers_drops_entry_without_name(tmp_path, caplog):
    sys.modules.pop("adapter", None)
    import adapter as adapter_module
    nonexistent = str(tmp_path / "_isolate_canonical")
    monkeypatch.setattr(adapter_module, "_CANONICAL_ADAPTER_DIR", nonexistent)
    monkeypatch.setattr(adapter_module, "_TEMPLATE_DIR", nonexistent)
    yaml_no_name = textwrap.dedent("""
        providers:
          - name: good
--- a/tests/test_env_model_provider_dispatch.py
+++ b/tests/test_env_model_provider_dispatch.py
@ -0,0 +1,287 @@
 """Tests for ``_resolve_model_and_provider_from_env`` — the env-vs-YAML
 reconciliation that fixes the 2026-05-08 dev-tree wedge incident.
 Symptom: 22/27 non-lead workspaces (minimax tier) wedged on
 ``Control request timeout: initialize`` because the runtime wheel's
 ``workspace/config.py`` interpreted ``MODEL_PROVIDER=minimax`` as the
 *model id* instead of the provider slug. ``model="minimax"`` failed to
 match the ``minimax-`` registry prefix, fell through to providers[0]
 (anthropic-oauth), demanded ``CLAUDE_CODE_OAUTH_TOKEN`` (unset on
 non-leads), and the claude CLI hung at SDK init.
 The persona env files (``~/.molecule-ai/personas/<name>/env``) declare
 the new convention:
  * ``MODEL`` — model id (e.g. ``MiniMax-M2.7-highspeed``)
  * ``MODEL_PROVIDER`` — provider slug (e.g. ``minimax``)
 These tests cover the matrix of (env shape) × (YAML shape) so a future
 contributor can't silently regress the wedge fix.
 """
 import pytest
 from adapter import (
    _BUILTIN_PROVIDERS,
    _resolve_model_and_provider_from_env,
 )
 # A registry that contains both anthropic-oauth (providers[0]) and
 # minimax/zai (third-party slugs) — matches the shipped config.yaml.
 _REGISTRY = _BUILTIN_PROVIDERS + (
    {
        "name": "minimax",
        "auth_mode": "third_party_anthropic_compat",
        "model_prefixes": ("minimax-",),
        "model_aliases": (),
        "base_url": "https://api.minimax.io/anthropic",
        "auth_env": ("MINIMAX_API_KEY",),
    },
    {
        "name": "zai",
        "auth_mode": "third_party_anthropic_compat",
        "model_prefixes": ("glm-",),
        "model_aliases": (),
        "base_url": "https://api.z.ai/api/anthropic",
        "auth_env": ("GLM_API_KEY",),
    },
 )
 def _clear_env(monkeypatch):
    monkeypatch.delenv("MODEL", raising=False)
    monkeypatch.delenv("MODEL_PROVIDER", raising=False)
 # ------------------------------------------------------------------
 # Persona env convention: MODEL=<id>, MODEL_PROVIDER=<slug>
 # ------------------------------------------------------------------
 def test_persona_env_minimax_resolves_correctly(monkeypatch):
    """The 2026-05-08 wedge regression test: persona env shape must
    yield model=MiniMax-M2.7-highspeed (not "minimax") and explicit
    provider=minimax."""
    _clear_env(monkeypatch)
    monkeypatch.setenv("MODEL", "MiniMax-M2.7-highspeed")
    monkeypatch.setenv("MODEL_PROVIDER", "minimax")
    model, provider = _resolve_model_and_provider_from_env(
        yaml_model="", yaml_provider="", providers=_REGISTRY,
    )
    assert model == "MiniMax-M2.7-highspeed"
    assert provider == "minimax"
 def test_persona_env_lead_claude_code_resolves_correctly(monkeypatch):
    """Lead persona env (MODEL=opus, MODEL_PROVIDER=claude-code) —
    ``claude-code`` is the persona-friendly alias for the canonical
    ``anthropic-oauth`` registry name. Must resolve via the alias map
    so the lead boots through the OAuth subscription path even when
    MODEL is a non-Anthropic model id (e.g. an operator who picked
    MiniMax in canvas but whose persona env still pins claude-code)."""
    _clear_env(monkeypatch)
    monkeypatch.setenv("MODEL", "opus")
    monkeypatch.setenv("MODEL_PROVIDER", "claude-code")
    model, provider = _resolve_model_and_provider_from_env(
        yaml_model="", yaml_provider="", providers=_REGISTRY,
    )
    assert model == "opus"
    # claude-code → anthropic-oauth via the alias map
    assert provider == "anthropic-oauth"
 def test_persona_env_lead_with_minimax_model_routes_via_oauth(monkeypatch):
    """Lead workspace whose persona pins MODEL_PROVIDER=claude-code but
    whose YAML/canvas selection happens to be a MiniMax model still
    routes via OAuth — the persona's provider pin wins over the
    model-prefix matcher. Without the alias map, the fall-through
    mis-routed leads to MiniMax even when their CLAUDE_CODE_OAUTH_TOKEN
    was set."""
    _clear_env(monkeypatch)
    monkeypatch.setenv("MODEL", "MiniMax-M2.7")
    monkeypatch.setenv("MODEL_PROVIDER", "claude-code")
    model, provider = _resolve_model_and_provider_from_env(
        yaml_model="", yaml_provider="", providers=_REGISTRY,
    )
    assert model == "MiniMax-M2.7"
    assert provider == "anthropic-oauth"
 def test_anthropic_alias_resolves_to_anthropic_api(monkeypatch):
    """``MODEL_PROVIDER=anthropic`` alias → ``anthropic-api`` (direct
    Anthropic API key path)."""
    _clear_env(monkeypatch)
    monkeypatch.setenv("MODEL", "claude-opus-4-7")
    monkeypatch.setenv("MODEL_PROVIDER", "anthropic")
    model, provider = _resolve_model_and_provider_from_env(
        yaml_model="", yaml_provider="", providers=_REGISTRY,
    )
    assert model == "claude-opus-4-7"
    assert provider == "anthropic-api"
 def test_persona_env_glm_resolves_correctly(monkeypatch):
    _clear_env(monkeypatch)
    monkeypatch.setenv("MODEL", "GLM-4.6")
    monkeypatch.setenv("MODEL_PROVIDER", "zai")
    model, provider = _resolve_model_and_provider_from_env(
        yaml_model="", yaml_provider="", providers=_REGISTRY,
    )
    assert model == "GLM-4.6"
    assert provider == "zai"
 def test_env_provider_slug_case_insensitive(monkeypatch):
    """Operator typos like ``MiniMax`` (mixed case) still resolve."""
    _clear_env(monkeypatch)
    monkeypatch.setenv("MODEL", "MiniMax-M2.7-highspeed")
    monkeypatch.setenv("MODEL_PROVIDER", "MiniMax")  # mixed case
    _, provider = _resolve_model_and_provider_from_env(
        yaml_model="", yaml_provider="", providers=_REGISTRY,
    )
    assert provider == "MiniMax"  # caller compares case-insensitively
 # ------------------------------------------------------------------
 # Legacy convention: MODEL_PROVIDER=<model-id>, MODEL unset
 # ------------------------------------------------------------------
 def test_legacy_model_provider_as_model_id_still_works(monkeypatch):
    """Pre-2026-05-08 canvas Save+Restart shape: MODEL_PROVIDER carried
    the model id directly (e.g. ``MODEL_PROVIDER=MiniMax-M2.7``) and
    no MODEL env. Must keep working so existing canvas users don't
    break overnight."""
    _clear_env(monkeypatch)
    monkeypatch.setenv("MODEL_PROVIDER", "MiniMax-M2.7-highspeed")
    model, provider = _resolve_model_and_provider_from_env(
        yaml_model="", yaml_provider="", providers=_REGISTRY,
    )
    # MiniMax-M2.7-highspeed is not a registered provider name, so
    # it's treated as a legacy model-id-in-MODEL_PROVIDER value.
    assert model == "MiniMax-M2.7-highspeed"
    assert provider is None
 # ------------------------------------------------------------------
 # Env wins over YAML
 # ------------------------------------------------------------------
 def test_env_model_wins_over_yaml_model(monkeypatch):
    """When both env MODEL and YAML model are set, env wins."""
    _clear_env(monkeypatch)
    monkeypatch.setenv("MODEL", "GLM-4.6")
    model, _ = _resolve_model_and_provider_from_env(
        yaml_model="MiniMax-M2.7", yaml_provider="", providers=_REGISTRY,
    )
    assert model == "GLM-4.6"
 def test_env_provider_wins_over_yaml_provider(monkeypatch):
    """Env MODEL_PROVIDER (when a registered slug) wins over YAML provider."""
    _clear_env(monkeypatch)
    monkeypatch.setenv("MODEL", "GLM-4.6")
    monkeypatch.setenv("MODEL_PROVIDER", "zai")
    _, provider = _resolve_model_and_provider_from_env(
        yaml_model="", yaml_provider="minimax", providers=_REGISTRY,
    )
    assert provider == "zai"
 # ------------------------------------------------------------------
 # YAML fallback (no env)
 # ------------------------------------------------------------------
 def test_no_env_falls_back_to_yaml(monkeypatch):
    """Workspace whose env doesn't set MODEL/MODEL_PROVIDER falls back
    to the YAML config — preserves existing operator workflows."""
    _clear_env(monkeypatch)
    model, provider = _resolve_model_and_provider_from_env(
        yaml_model="claude-sonnet-4-6",
        yaml_provider="anthropic-api",
        providers=_REGISTRY,
    )
    assert model == "claude-sonnet-4-6"
    assert provider == "anthropic-api"
 def test_no_env_no_yaml_returns_empty(monkeypatch):
    """Pure default path — caller (setup) substitutes ``sonnet``."""
    _clear_env(monkeypatch)
    model, provider = _resolve_model_and_provider_from_env(
        yaml_model="", yaml_provider="", providers=_REGISTRY,
    )
    assert model == ""
    assert provider is None
 def test_yaml_provider_anthropic_is_aliased_to_anthropic_api(monkeypatch):
    """Regression for 2026-05-09 staging-cplead-2 incident: every
    workspace booted ``configuration_status=not_configured`` because the
    molecule-runtime wheel auto-derives ``runtime_config.provider =
    "anthropic"`` from the default model slug ``anthropic:claude-opus-4-7``.
    The adapter received ``yaml_provider="anthropic"`` from the wheel and
    rejected it with ``ValueError: provider='anthropic' but it is not in
    the providers registry`` — but ``anthropic`` is already in
    ``_PROVIDER_SLUG_ALIASES`` for the env-var path. Mirror the alias map
    on the YAML path so the wheel default produces a registered provider
    name."""
    _clear_env(monkeypatch)
    _, provider = _resolve_model_and_provider_from_env(
        yaml_model="", yaml_provider="anthropic", providers=_REGISTRY,
    )
    assert provider == "anthropic-api", (
        f"yaml_provider='anthropic' must resolve through the alias map to "
        f"'anthropic-api'; got {provider!r}. Without this aliasing the "
        f"wheel-default workspace boot wedges at adapter.setup()."
    )
 def test_yaml_provider_claude_code_is_aliased_to_anthropic_oauth(monkeypatch):
    """Symmetric coverage: persona-friendly ``claude-code`` slug from the
    YAML ``provider:`` field must alias to ``anthropic-oauth``, the same
    way the env-var path resolves it. Lead workspaces that pin the OAuth
    path in YAML (instead of via env) must not wedge."""
    _clear_env(monkeypatch)
    _, provider = _resolve_model_and_provider_from_env(
        yaml_model="", yaml_provider="claude-code", providers=_REGISTRY,
    )
    assert provider == "anthropic-oauth"
 def test_yaml_provider_unknown_passes_through_for_actionable_error(monkeypatch):
    """An unaliased, unknown YAML provider (e.g. ``yaml_provider="mystery"``)
    must NOT be silently swapped to providers[0] — it must reach
    ``_resolve_provider`` so the adapter raises the actionable
    ``Known providers: ...`` error message. The alias map is a
    convenience for the two persona-convention slugs only; everything
    else must keep its original semantics."""
    _clear_env(monkeypatch)
    _, provider = _resolve_model_and_provider_from_env(
        yaml_model="", yaml_provider="mystery", providers=_REGISTRY,
    )
    assert provider == "mystery"
 # ------------------------------------------------------------------
 # Whitespace / empty-value defensive cases
 # ------------------------------------------------------------------
 def test_whitespace_only_env_treated_as_unset(monkeypatch):
    _clear_env(monkeypatch)
    monkeypatch.setenv("MODEL", "   ")
    monkeypatch.setenv("MODEL_PROVIDER", "  ")
    model, provider = _resolve_model_and_provider_from_env(
        yaml_model="opus", yaml_provider="", providers=_REGISTRY,
    )
    assert model == "opus"
    assert provider is None
 def test_empty_env_value_treated_as_unset(monkeypatch):
    _clear_env(monkeypatch)
    monkeypatch.setenv("MODEL", "")
    monkeypatch.setenv("MODEL_PROVIDER", "")
    model, provider = _resolve_model_and_provider_from_env(
        yaml_model="sonnet", yaml_provider="", providers=_REGISTRY,
    )
    assert model == "sonnet"
    assert provider is None
--- a/tests/test_provider_resolution.py
+++ b/tests/test_provider_resolution.py
@ -0,0 +1,146 @@
 """Tests for the provider-resolution path that was silent-failing on #180.
 Regression coverage: when an operator picks a provider in the canvas Config
 tab that isn't in the registry, the adapter must raise ValueError with an
 actionable message — NOT silently fall through to providers[0]
 (anthropic-oauth) and then have the Claude SDK hit the user's OAuth quota
 under a different name.
 These tests mirror the production failure mode reported by Hongming
 2026-05-07 17:35: workspace config.yaml had `provider: minimax` set, the
 adapter ignored it entirely, the SDK kept calling the Anthropic API with
 CLAUDE_CODE_OAUTH_TOKEN, hit the OAuth quota, and the canvas surfaced
 "Agent error (Exception)" with no clue why.
 Import-shim setup (sys.path + molecule_runtime / a2a / claude_sdk_executor
 stubs) lives in tests/conftest.py — shared with test_adapter_prevalidate
 so the two stub installers can't disagree on shape (e.g. BaseAdapter
 having install_plugins_via_registry).
 """
 import pytest
 from adapter import (
    _BUILTIN_PROVIDERS,
    _resolve_provider,
 )
 def test_resolve_with_no_explicit_provider_falls_back_to_model_match():
    """No explicit provider → model-based prefix/alias matching, default to providers[0]."""
    p = _resolve_provider("claude-opus-4-7", _BUILTIN_PROVIDERS)
    assert p["name"] == "anthropic-api"  # matches model_prefixes=("claude-",)
 def test_resolve_with_no_explicit_provider_falls_back_to_default():
    """Unknown model + no explicit provider → providers[0] (anthropic-oauth)."""
    p = _resolve_provider("unknown-model", _BUILTIN_PROVIDERS)
    assert p["name"] == "anthropic-oauth"
 def test_resolve_with_explicit_provider_in_registry_returns_match():
    """Explicit name lookup wins over model-based resolution."""
    # Even though "claude-opus-4-7" would normally resolve to anthropic-api
    # via prefix matching, the explicit provider name wins.
    p = _resolve_provider(
        "claude-opus-4-7", _BUILTIN_PROVIDERS,
        explicit_provider="anthropic-oauth",
    )
    assert p["name"] == "anthropic-oauth"
 def test_resolve_with_explicit_provider_case_insensitive():
    """Provider name match is case-insensitive (operators write 'Anthropic-OAuth' etc)."""
    p = _resolve_provider(
        "sonnet", _BUILTIN_PROVIDERS,
        explicit_provider="ANTHROPIC-OAUTH",
    )
    assert p["name"] == "anthropic-oauth"
 def test_resolve_with_explicit_provider_not_in_registry_raises():
    """The #180 regression test: explicit non-registry provider must raise, not fall through."""
    with pytest.raises(ValueError) as exc_info:
        _resolve_provider(
            "MiniMax-M2.7-highspeed", _BUILTIN_PROVIDERS,
            explicit_provider="minimax",
        )
    msg = str(exc_info.value)
    # Must name the bad provider so operator knows what they typed
    assert "minimax" in msg
    # Must list known providers so operator knows what's available
    assert "anthropic-oauth" in msg
    assert "anthropic-api" in msg
    # Must give actionable next steps — NOT just "not found"
    assert "providers:" in msg or "Add" in msg
    assert "Switch" in msg or "runtime" in msg
 def test_resolve_with_explicit_provider_does_not_silent_fallback():
    """Specifically: must not return providers[0] when explicit_provider is bogus.
    This is the exact silent-fallback path that caused the user-visible
    bug: operator picks 'minimax' → adapter returns anthropic-oauth →
    SDK uses CLAUDE_CODE_OAUTH_TOKEN → hits quota.
    """
    with pytest.raises(ValueError):
        result = _resolve_provider(
            "anything", _BUILTIN_PROVIDERS,
            explicit_provider="minimax",
        )
        # If the implementation regresses to silent fallback, this would
        # have returned providers[0] (anthropic-oauth) instead of raising.
        # Defense-in-depth: guard against accidental "return" inside the
        # error path.
        assert result["name"] not in {"anthropic-oauth", "anthropic-api"}, (
            "REGRESSION: silent fallback to default provider when explicit "
            "provider name is not in registry — this is the #180 bug."
        )
 def test_resolve_with_explicit_provider_in_custom_registry():
    """When operator adds a third-party provider to the registry, explicit lookup finds it."""
    custom_registry = _BUILTIN_PROVIDERS + (
        {
            "name": "minimax",
            "auth_mode": "third_party_anthropic_compat",
            "model_prefixes": ("minimax-",),
            "model_aliases": (),
            "base_url": "https://api.minimaxi.com/anthropic-compat",
            "auth_env": ("MINIMAX_API_KEY",),
        },
    )
    p = _resolve_provider(
        "MiniMax-M2.7-highspeed", custom_registry,
        explicit_provider="minimax",
    )
    assert p["name"] == "minimax"
    assert p["base_url"] == "https://api.minimaxi.com/anthropic-compat"
    assert "MINIMAX_API_KEY" in p["auth_env"]
 def test_resolve_empty_providers_raises():
    """Pre-condition: providers must be non-empty (existing behavior preserved)."""
    with pytest.raises(ValueError, match="empty providers tuple"):
        _resolve_provider("anything", ())
 def test_resolve_explicit_empty_string_treated_as_no_explicit():
    """`provider: ''` (empty string) → fall back to model-based resolution, not raise."""
    # This shape can happen when the canvas writes an empty provider field.
    # Treating it as "no explicit pick" is more forgiving than raising,
    # since the user clearly didn't intend to break their workspace.
    p = _resolve_provider(
        "claude-opus-4-7", _BUILTIN_PROVIDERS,
        explicit_provider="",
    )
    assert p["name"] == "anthropic-api"  # fell through to model-based
 def test_resolve_explicit_none_treated_as_no_explicit():
    """`explicit_provider=None` (default) → fall back to model-based resolution."""
    p = _resolve_provider(
        "claude-opus-4-7", _BUILTIN_PROVIDERS,
        explicit_provider=None,
    )
    assert p["name"] == "anthropic-api"
--- a/tests/test_provider_routing.py
+++ b/tests/test_provider_routing.py
@ -219,7 +219,6 @@ def test_glm_kimi_deepseek_also_project(adapter_module, monkeypatch):
    """
    cases = [
        ("zai", "GLM_API_KEY"),
        ("moonshot", "KIMI_API_KEY"),
        ("deepseek", "DEEPSEEK_API_KEY"),
    ]
    for provider_name, env_name in cases:
@ -242,3 +241,83 @@ def test_glm_kimi_deepseek_also_project(adapter_module, monkeypatch):
            f"{env_name} must project onto ANTHROPIC_AUTH_TOKEN for "
            f"provider={provider_name}"
        )
 def test_kimi_coding_projects_into_anthropic_api_key(adapter_module, monkeypatch):
    """Kimi For Coding's gateway authenticates with the x-api-key header
    (kimi.com official Claude Code doc), which the Anthropic SDK / claude
    CLI emits from ANTHROPIC_API_KEY — NOT the Bearer ANTHROPIC_AUTH_TOKEN
    used by MiniMax/GLM/DeepSeek. The kimi-coding provider sets
    auth_token_env: ANTHROPIC_API_KEY so KIMI_API_KEY projects there.
    Regression guard for the original mis-route: KIMI_API_KEY landing in
    ANTHROPIC_AUTH_TOKEN against api.kimi.com/coding 401s.
    """
    import os
    _clear_all_auth_env(monkeypatch, adapter_module)
    monkeypatch.setenv("KIMI_API_KEY", "sk-kimi-sentinel")
    provider = {
        "name": "kimi-coding",
        "auth_mode": "third_party_anthropic_compat",
        "model_prefixes": ("kimi-",),
        "model_aliases": (),
        "base_url": "https://api.kimi.com/coding/",
        "auth_env": ("KIMI_API_KEY", "ANTHROPIC_API_KEY", "ANTHROPIC_AUTH_TOKEN"),
        "auth_token_env": "ANTHROPIC_API_KEY",
    }
    adapter_module._project_vendor_auth(provider)
    assert os.environ.get("ANTHROPIC_API_KEY") == "sk-kimi-sentinel", (
        "KIMI_API_KEY must project onto ANTHROPIC_API_KEY (x-api-key) for "
        "the kimi-coding provider per kimi.com's official Claude Code doc"
    )
    assert os.environ.get("ANTHROPIC_AUTH_TOKEN") is None, (
        "KIMI_API_KEY must NOT land in ANTHROPIC_AUTH_TOKEN — the Bearer "
        "header 401s against api.kimi.com/coding (the original mis-route)"
    )
 def test_kimi_coding_operator_anthropic_api_key_wins(adapter_module, monkeypatch):
    """Idempotency holds for the per-provider target too: an explicit
    operator ANTHROPIC_API_KEY is never clobbered by the projection."""
    import os
    _clear_all_auth_env(monkeypatch, adapter_module)
    monkeypatch.setenv("KIMI_API_KEY", "sk-kimi-sentinel")
    monkeypatch.setenv("ANTHROPIC_API_KEY", "operator-value")
    provider = {
        "name": "kimi-coding",
        "auth_mode": "third_party_anthropic_compat",
        "model_prefixes": ("kimi-",),
        "model_aliases": (),
        "base_url": "https://api.kimi.com/coding/",
        "auth_env": ("KIMI_API_KEY", "ANTHROPIC_API_KEY", "ANTHROPIC_AUTH_TOKEN"),
        "auth_token_env": "ANTHROPIC_API_KEY",
    }
    adapter_module._project_vendor_auth(provider)
    assert os.environ.get("ANTHROPIC_API_KEY") == "operator-value", (
        "explicit operator ANTHROPIC_API_KEY must win over auto-projection"
    )
 def test_normalize_provider_parses_auth_token_env(adapter_module):
    """_normalize_provider surfaces auth_token_env; absent → the
    ANTHROPIC_AUTH_TOKEN default (preserves MiniMax/GLM/DeepSeek)."""
    with_override = adapter_module._normalize_provider({
        "name": "kimi-coding",
        "auth_mode": "third_party_anthropic_compat",
        "base_url": "https://api.kimi.com/coding/",
        "auth_env": ["KIMI_API_KEY", "ANTHROPIC_API_KEY"],
        "auth_token_env": "ANTHROPIC_API_KEY",
    })
    assert with_override["auth_token_env"] == "ANTHROPIC_API_KEY"
    default = adapter_module._normalize_provider({
        "name": "minimax",
        "auth_mode": "third_party_anthropic_compat",
        "base_url": "https://api.minimax.io/anthropic",
        "auth_env": ["MINIMAX_API_KEY"],
    })
    assert default["auth_token_env"] == "ANTHROPIC_AUTH_TOKEN"