0.2.1b7

.github/workflows/pr.yml

@@ -69,7 +69,7 @@ jobs:
     strategy:
       fail-fast: false
       matrix:
-        python-version: ["3.10", "3.11", "3.12"]
+        python-version: ["3.10", "3.11", "3.12", "3.13"]
 
     steps:
       - uses: actions/checkout@v4

HISTORY.rst

@@ -3,6 +3,93 @@
 Release History
 ===============
 
+0.2.1b7
++++++++
+
+Build stage re-entry fix
+~~~~~~~~~~~~~~~~~~~~~~~~~~
+* **QA remediation failure now retries on re-entry** — fixed a bug where
+  ``mark_stage_generated()`` was called after each remediation attempt
+  inside ``_run_stage_qa()``, leaving the stage with status ``"generated"``
+  even when QA subsequently failed.  On re-entry, the stage was skipped
+  instead of retried.  Changed to ``mark_stage_validating()`` so failed
+  stages remain in the retry list.
+
+QA checklist hardening
+~~~~~~~~~~~~~~~~~~~~~~~~
+* **Aligned response_export_values directive** — QA checklist now requires
+  ``response_export_values = ["*"]`` on EVERY ``azapi_resource``, matching
+  the terraform agent's mandatory rule (was conditional on output usage).
+* **Added deploy.sh -state= flag check** — QA checklist now flags use of
+  ``terraform output -state=`` which was removed in Terraform 1.9.
+* **Added UUID hex validation** — QA checklist now checks that UUID values
+  in role assignment names contain only valid hex characters ``[0-9a-f]``.
+
+Full stage retry on QA exhaustion
+~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+* **Full stage retry when QA remediation fails** -- when QA remediation
+  exhausts all attempts for a stage, the build now retries the entire
+  stage from scratch (clean artifacts, regenerate, QA) instead of
+  stopping the build immediately.  Previous QA findings are injected
+  into the new generation prompt — framed as guidance rather than
+  file-specific instructions — so the model avoids the same classes
+  of mistakes on the fresh attempt.
+
+  In practice, the same generation prompt produces passing code ~90%
+  of the time.  The remaining ~10% failure rate is stochastic — not a
+  systematic prompt deficiency — meaning a fresh generation with
+  knowledge of what went wrong almost always succeeds.  Without this
+  retry, that 10% forces the user to manually re-run the entire build,
+  losing the progress of all previously generated stages.  The retry
+  doubles the token cost of one stage in the worst case, but saves
+  the full cost of restarting a 16-stage build from scratch.
+
+  Controlled by ``_MAX_FULL_STAGE_ATTEMPTS`` (default 2: 1 initial
+  + 1 fresh retry).
+
+Generation prompt improvements
+~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+* **Front-loaded remote state no-dead-code directive** — when upstream
+  stages exist, a ``CROSS-STAGE DEPENDENCIES — NO DEAD CODE`` section
+  now appears before the architecture context in the generation prompt,
+  reducing unused ``terraform_remote_state`` data sources.
+
+Agent-level service filtering
+~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+* **Agent governance checks now filter by service namespace** — added
+  ``stage_services`` field to ``AgentContext``, populated by
+  ``_agent_build_context()``.  ``_apply_governance_check()`` now passes
+  stage services to ``validate_response()``, reducing false positive
+  anti-pattern warnings for irrelevant service namespaces.
+
+ReDoS fix in transform handlers
+~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+* **Replaced nested-quantifier regex with brace counting** — extracted
+  shared ``_find_azapi_blocks()`` helper and rewrote
+  ``_add_response_export_values``, ``_add_resource_group_parent_id``,
+  and ``_remove_private_endpoint_resources`` to use it.  Eliminates
+  potential exponential backtracking on pathological input.
+
+Test suite consolidation
+~~~~~~~~~~~~~~~~~~~~~~~~~~
+* **Consolidated and enhanced unit test coverage** — migrated flat test
+  files to a mirrored directory structure (1:1 test-to-source mapping),
+  merged split test files, and removed ~114 duplicate tests across 10
+  files.  Test suite reduced from 3,644 to 3,530 tests with zero loss
+  of unique coverage.
+
+QA review continuation for large stages
+~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+* **QA review collects complete response before evaluating** — when the
+  QA review response is truncated (``finish_reason=length``), the build
+  session now continues requesting until the full review is received,
+  then evaluates the concatenated result.  Uses the existing
+  ``_execute_with_continuation()`` pattern with a review-specific
+  continuation prompt that prevents the QA agent from generating code
+  in the continuation.  Conversation history is saved and restored
+  around QA calls to prevent review messages from contaminating
+  subsequent stage generation.
+
 0.2.1b6
 +++++++
 

azext_prototype/agents/base.py

@@ -81,6 +81,7 @@ class AgentContext:
     artifacts: dict[str, Any] = field(default_factory=dict)
     shared_state: dict[str, Any] = field(default_factory=dict)
     mcp_manager: Any = None  # MCPManager | None — typed as Any to avoid circular import
+    stage_services: list[str] | None = None  # ARM namespaces for service filtering
 
     def add_artifact(self, key: str, value: Any):
         """Store an artifact for other agents to reference."""
@@ -299,7 +300,7 @@ def _apply_governance_check(self, response: AIResponse, context: AgentContext) -
         avoid duplicating the governance warning block.
         """
         iac_tool = context.project_config.get("project", {}).get("iac_tool") if context.project_config else None
-        warnings = self.validate_response(response.content, iac_tool=iac_tool, services=None)
+        warnings = self.validate_response(response.content, iac_tool=iac_tool, services=context.stage_services)
         if warnings:
             for w in warnings:
                 logger.warning("Governance: %s", w)

azext_prototype/agents/builtin/qa_engineer.py

@@ -233,6 +233,8 @@ def _encode_image(path: str) -> str:
 - [ ] deploy.sh includes error handling (set -euo pipefail, trap)
 - [ ] deploy.sh exports outputs to JSON file for downstream stages
 - [ ] deploy.sh includes Azure login verification
+- [ ] deploy.sh does NOT use `terraform output -state=` — this flag was removed
+      in Terraform 1.9. Use `jq` on the state file or `cd` into the stage directory
 
 ### 4. Output Completeness
 - [ ] outputs.tf exports resource group name(s)
@@ -251,8 +253,8 @@ def _encode_image(path: str) -> str:
 - [ ] All referenced variables are defined in variables.tf
 - [ ] All referenced locals are defined in locals.tf
 - [ ] Application code includes all referenced classes/models/DTOs
-- [ ] Every azapi_resource whose `.output.properties` is referenced in
-      outputs.tf MUST have `response_export_values = ["*"]` declared
+- [ ] EVERY `azapi_resource` block MUST have `response_export_values = ["*"]`
+      declared — no exceptions, even if outputs.tf does not reference its properties
 - [ ] No .tf file is empty or contains only comments (dead files)
 
 ### 7. Terraform File Structure
@@ -314,6 +316,8 @@ def _encode_image(path: str) -> str:
       **NOT** string interpolation on the storage account ID
 - [ ] RBAC assignments for the worker identity (Stage 1) are **unconditional**
       (no `count`). The worker identity exists before any service stage runs.
+- [ ] UUID values in role assignment names contain only valid hex characters
+      `[0-9a-f]` — letters `g`-`z` are invalid and ARM rejects with `InvalidName`
 
 ### 13. Application Code (app stages only)
 - [ ] Application source code is syntactically correct and complete

azext_prototype/ai/token_tracker.py

@@ -44,30 +44,35 @@
 # GitHub Copilot Premium Request Unit (PRU) multipliers.
 # Each API call costs (1 × multiplier) PRUs.  Only applies to the
 # Copilot provider — models not in this table produce 0 PRUs.
-# Source: https://docs.github.com/en/copilot/concepts/billing/copilot-requests
+# Source: https://docs.github.com/en/copilot/managing-copilot/monitoring-usage-and-entitlements/about-premium-requests
+# Last updated: 2026-04-08
 _PRU_MULTIPLIERS: dict[str, float] = {
     # Included with paid plans (0 PRUs)
     "gpt-5-mini": 0,
     "gpt-4.1": 0,
     "gpt-4o": 0,
+    "raptor-mini": 0,
     # Low-cost (0.25–0.33 PRUs per request)
     "grok-code-fast-1": 0.25,
     "claude-haiku-4.5": 0.33,
     "gemini-3-flash": 0.33,
-    "gpt-5.1-codex-mini": 0.33,
     "gpt-5.4-mini": 0.33,
     # Standard (1 PRU per request)
     "claude-sonnet-4": 1,
     "claude-sonnet-4.5": 1,
     "claude-sonnet-4.6": 1,
+    "gemini-2.5-pro": 1,
     "gemini-3-pro": 1,
-    "gemini-3-pro-1.5": 1,
+    "gemini-3.1-pro": 1,
     "gpt-5.1": 1,
     "gpt-5.2": 1,
+    "gpt-5.2-codex": 1,
+    "gpt-5.3-codex": 1,
     "gpt-5.4": 1,
     # Premium (3+ PRUs per request)
     "claude-opus-4.5": 3,
     "claude-opus-4.6": 3,
+    "claude-opus-4.6-fast": 30,
 }
 
 

azext_prototype/azext_metadata.json

@@ -2,7 +2,7 @@
   "azext.isPreview": true,
   "azext.minCliCoreVersion": "2.50.0",
   "name": "prototype",
-  "version": "0.2.1b6",
+  "version": "0.2.1b7",
   "azext.summary": "Azure CLI extension for building rapid prototypes with GitHub Copilot.",
   "license": "MIT",
   "classifiers": [