feat: prove Skill utility with paired GLM-5.2 evaluation

.env.example

@@ -13,21 +13,32 @@
 # -----------------------------------------------------------------------------
 # LLM Configuration
 # -----------------------------------------------------------------------------
-# API key for the LLM provider. Used by all glm-4 agents and as the fallback
-# for CoderAgent. Treat as a top-level secret — it is injected into the
-# credential gateway, never passed directly to agent prompts.
+# API key for the OpenAI-compatible LLM provider used by every agent.
+# Treat it as a top-level secret: inject it into the credential gateway and
+# never pass it directly to agent prompts.
 LLM_API_KEY=
 
-# Base URL of the OpenAI-compatible endpoint serving GLM-4 / Codex.
+# Base URL of the Z.AI OpenAI-compatible endpoint serving GLM-5.2.
 # Examples:
-#   - ZhipuAI:        https://open.bigmodel.cn/api/paas/v4
+#   - Z.AI general:   https://api.z.ai/api/paas/v4/
 #   - Local vLLM:     http://127.0.0.1:8000/v1
-#   - Azure OpenAI:   https://<resource>.openai.azure.com
-LLM_BASE_URL=https://open.bigmodel.cn/api/paas/v4
+LLM_BASE_URL=https://api.z.ai/api/paas/v4/
 
 # Default model name. Individual agents override this in config/agents.yaml
-# (e.g. CoderAgent uses "codex" with a "glm-4" fallback).
-LLM_MODEL=glm-4
+# Keep this identical to the model registered in AgentTeams.
+LLM_MODEL=glm-5.2
+
+# Embeddings are a separately metered Z.AI capability. By default they reuse
+# the LLM key and endpoint; set these when using a distinct provider/account.
+# `embedding-3` must be enabled in the selected account's resource package.
+EMBEDDING_API_KEY=
+EMBEDDING_BASE_URL=https://api.z.ai/api/paas/v4/
+EMBEDDING_MODEL=embedding-3
+
+# openai-compatible for semantic provider embeddings; local-hash is a
+# credential-free deterministic degraded mode for offline/test deployments.
+EMBEDDING_PROVIDER=openai-compatible
+LOCAL_EMBEDDING_DIMENSION=384
 
 # -----------------------------------------------------------------------------
 # GitHub Configuration
@@ -57,6 +68,10 @@ CHROMADB_PATH=.devflow/chromadb
 # remote collector (e.g. Tempo, Jaeger, or the OTel Collector) in prod.
 OTEL_EXPORTER_OTLP_ENDPOINT=http://127.0.0.1:4317
 
+# Loopback Prometheus exposition started with the CI/CD MCP service.
+DEVFLOW_METRICS_HOST=127.0.0.1
+DEVFLOW_METRICS_PORT=9090
+
 # Deployment environment tag attached to every trace/metric as a resource
 # attribute. Useful for filtering prod vs staging vs dev.
 # Default: development
@@ -87,6 +102,37 @@ MCP_CICD_PORT=8766
 # Default: github_actions
 CICD_PROVIDER=github_actions
 
+# Repository and server-owned argv used by the isolated CI/CD MCP server.
+# Agents never receive permission to submit an arbitrary command.
+DEVFLOW_REPOSITORY_ROOT=
+CICD_TEST_COMMAND_JSON=["python","-m","pytest","-q"]
+CICD_TEST_TIMEOUT_SECONDS=600
+
+# Optional server-owned rollback provider. The command receives environment and
+# target only through DEVFLOW_DEPLOYMENT_ENVIRONMENT/DEVFLOW_TARGET_RELEASE.
+CICD_ROLLBACK_COMMAND_JSON=
+CICD_ROLLBACK_TIMEOUT_SECONDS=120
+CICD_RELEASE_STATE_PATH=.devflow/releases.json
+
+# Optional built-in atomic symlink rollback provider for Linux deployments:
+# CICD_ROLLBACK_COMMAND_JSON=["python3","scripts/rollback_release.py"]
+CICD_RELEASES_ROOT=
+CICD_RELEASE_LINK_TEMPLATE=
+HEALTH_CHECK_URL=
+
+# At least 32 random bytes used only by the trusted approval service and MCP
+# policy layer to authenticate high-risk human approvals. Never expose it to an
+# Agent, model prompt, or client-side environment.
+DEVFLOW_APPROVAL_HMAC_KEY=
+
+# Separate key used by the trusted runtime and internal MCP server to prevent
+# forged Agent/Skill call context. Use at least 32 random bytes.
+DEVFLOW_MCP_CONTEXT_HMAC_KEY=
+
+# Signing key for short-lived credential capability handles. Provider secrets
+# are resolved only inside trusted adapters and never returned to an Agent.
+DEVFLOW_CREDENTIAL_BROKER_HMAC_KEY=
+
 # -----------------------------------------------------------------------------
 # Credential Gateway (Security)
 # -----------------------------------------------------------------------------

.gitattributes

@@ -0,0 +1,2 @@
+* text=auto eol=lf
+*.ps1 text eol=crlf

.github/workflows/ci.yml

@@ -23,21 +23,25 @@ jobs:
           python-version: ${{ matrix.python-version }}
           cache: pip
       - name: Install
-        run: python -m pip install -e ".[dev]"
+        run: python -m pip install -e ".[dev,mcp]"
       - name: Validate configuration
         run: python -m devflow.cli validate
       - name: Lint
         run: ruff check .
       - name: Type check
         run: mypy src scripts tests
+      - name: Build AgentTeams package
+        run: python scripts/build_agentteams_package.py
       - name: Test
-        run: python -m pytest -q
+        run: python -m pytest --cov=devflow --cov-report=term --cov-fail-under=80 -q
       - name: Evaluate Skills
         run: python scripts/evaluate_skills.py
-      - name: Build AgentTeams package
-        run: python scripts/build_agentteams_package.py
+      - name: Validate behavioral evaluation suite
+        run: python scripts/run_behavior_evals.py --validate-only
       - uses: actions/upload-artifact@v4
         with:
-          name: devflow-worker-python-${{ matrix.python-version }}
-          path: dist/devflow-worker.zip
+          name: devflow-role-packages-v1.3.0-python-${{ matrix.python-version }}
+          path: |
+            dist/devflow-*-v1.3.0.zip
+            dist/devflow-*-v1.3.0.zip.sha256
           if-no-files-found: error

.gitignore

@@ -1,4 +1,6 @@
 .env
+.env.*
+!.env.example
 .venv/
 .devflow/
 .pytest_cache/
@@ -11,5 +13,12 @@ build/
 dist/
 htmlcov/
 .coverage
+coverage.json
 logs/
-
+tmp/
+.tmp/
+outputs/*.pptx.inspect.ndjson
+outputs/*/
+outputs/*.zip
+outputs/*.zip.sha256
+agentteams/systemd/

CHANGELOG.md

@@ -2,6 +2,45 @@
 
 All notable changes follow semantic versioning.
 
+## 1.3.0 - 2026-07-28
+
+- Added a typed Tester-to-TeamLeader-to-Coder repair route with sanitized,
+  digest-bound failure evidence, one route claim per execution attempt, and a
+  three-call issue-global Coder model budget shared by validation and test
+  retries.
+- Hardened Coder input/output boundaries with shared credential redaction,
+  Locator-derived file scope, patch consistency checks, syntax validation, and
+  whole-candidate secret scanning.
+- Upgraded `patch-generator` to 2.3.0 and `test-runner` to 3.2.0 with strict
+  runtime-aligned contracts, PatchCandidate 1.2 evidence scopes and model-call
+  ordinals, standalone
+  source-bound validators, and cross-Skill mediation checks.
+- Added a canonical local task router, Leader-owned dispatch claims, deep-copy
+  event isolation, bounded test-result ingress, and generic named-credential
+  redaction.
+- Bound T4/T5 resume approval to a deployment domain, policy key, project
+  incarnation, action, task, risk, target, and one exact request digest; added
+  legacy-policy and cross-deployment replay rejection tests, race-safe key
+  creation cleanup, and stale approval-ledger lock recovery.
+- Prepared deterministic 1.3.0 release-candidate role packages and updated the
+  package server, controller cache policy, source attestation, CI artifacts,
+  and deployment documentation. Final commit/tag binding and server deployment
+  evidence remain separate release gates.
+
+## 1.2.0 - 2026-07-27
+
+- Published six deterministic, role-scoped AgentTeams Worker packages and
+  added a seventh `github-evidence` Skill with an envelope-bound local
+  authorizer.
+- Added guarded TeamHarness role enforcement, idempotent task transitions,
+  digest-bound T4/T5 approval, and scope-bound GitHub capability issuance.
+- Added a pinned Higress read-only MCP route, Broker receipts, NetworkPolicy,
+  Redis-backed MCP sessions, and live positive/negative boundary evidence.
+- Added package, runtime, MinIO, Matrix compatibility, and native OpenClaw
+  boundary reconcilers with fail-closed checks.
+- Added the final preliminary-round PPT/PDF, a fixed sample request and
+  machine-checkable expected output, and the complete Apache-2.0 license.
+
 ## 1.0.0 - 2026-07-24
 
 - Added a six-agent AgentTeams software-delivery design.
@@ -11,4 +50,3 @@ All notable changes follow semantic versioning.
   and terminal experience distillation.
 - Added a credential-free deterministic demonstration and AgentTeams worker
   package builder.
-