claude-code-generator 0.6.2__tar.gz → 0.6.3__tar.gz

{claude_code_generator-0.6.2/src/claude_code_generator.egg-info → claude_code_generator-0.6.3}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: claude-code-generator
-Version: 0.6.2
+Version: 0.6.3
 Summary: Orchestrator CLI that drives Claude Code end-to-end to generate whole projects from a requirements.md file.
 Author: Silvio Baratto
 License: MIT

{claude_code_generator-0.6.2 → claude_code_generator-0.6.3}/pyproject.toml

@@ -4,7 +4,7 @@ build-backend = "setuptools.build_meta"
 
 [project]
 name = "claude-code-generator"
-version = "0.6.2"
+version = "0.6.3"
 description = "Orchestrator CLI that drives Claude Code end-to-end to generate whole projects from a requirements.md file."
 readme = "README.md"
 license = { text = "MIT" }

{claude_code_generator-0.6.2 → claude_code_generator-0.6.3/src/claude_code_generator.egg-info}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: claude-code-generator
-Version: 0.6.2
+Version: 0.6.3
 Summary: Orchestrator CLI that drives Claude Code end-to-end to generate whole projects from a requirements.md file.
 Author: Silvio Baratto
 License: MIT

{claude_code_generator-0.6.2 → claude_code_generator-0.6.3}/src/code_generator/__init__.py

@@ -1,3 +1,3 @@
 """code-generator: orchestrator CLI for end-to-end project generation."""
 
-__version__ = "0.6.2"
+__version__ = "0.6.3"

{claude_code_generator-0.6.2 → claude_code_generator-0.6.3}/src/code_generator/orchestrator/phase6_5_verify.py

@@ -739,6 +739,7 @@ async def run(
             max_turns=max_turns,
             attempt=attempt,
             target=target,
+            port=verify_port,
         )
 
     raise VerificationFailed(
@@ -1023,8 +1024,15 @@ async def _run_repair_pass(
     max_turns: int | None,
     attempt: int,
     target: Any,
+    port: int,
 ) -> None:
-    """Dispatch the Sonnet 4.6 repair pass via the per-cycle shared client (#275)."""
+    """Dispatch the Sonnet 4.6 repair pass via the per-cycle shared client (#275).
+
+    *port* is the ephemeral verify port. It is injected into the prompt as
+    ``VERIFY_BASE_URL`` so the repair agent binds the app to the exact
+    host:port the orchestrator's authoritative T2 probe hits — the agent
+    inventing its own port was the entire failure mode this guards against.
+    """
     del runner_module  # rate_limit signal is replayed via run_with_shared_client
     requirements_path = project_dir / ".code-generator" / "requirements.md"
     spec_path = project_dir / ".code-generator" / "phase6_5_spec.json"
@@ -1042,6 +1050,7 @@ async def _run_repair_pass(
         ORACLE_REPORT=_load_oracle_report_text(requirements_path),
         TIERS_ACTIVE=tiers_active,
         TIERS_SKIPPED=tiers_skipped,
+        VERIFY_BASE_URL=f"http://127.0.0.1:{port}",
     )
     effort_level = _effort.resolve_effort(
         phase=66, complexity="", issue_labels=[], retry_count=attempt

{claude_code_generator-0.6.2 → claude_code_generator-0.6.3}/src/code_generator/prompts/__init__.py

@@ -68,6 +68,9 @@ PLACEHOLDER_SPEC: dict[str, frozenset[str]] = {
             "AGENT_SPEC_PATH",
             "TIERS_ACTIVE",
             "TIERS_SKIPPED",
+            # Ephemeral verify-port isolation: the exact host:port the
+            # orchestrator probes; the repair agent must bind precisely this.
+            "VERIFY_BASE_URL",
         }
     ),
     "prompt-phase-7-commit.md": frozenset(

{claude_code_generator-0.6.2 → claude_code_generator-0.6.3}/src/code_generator/prompts/prompt-phase-6-5-verify.md

@@ -14,6 +14,8 @@ Verification runs in five tiers, lightest first; each is a hard gate on the next
 - **T0 — Build.** Run the canonical build command for the detected stack. Compile errors here block T1.
 - **T1 — Boot smoke.** Start the application or its main entry point. Watch for crash-on-boot, missing dependencies, unreachable migrations, or import-time errors. Time-bound: a successful boot is detected by either an explicit "ready" log line or the process surviving for `BOOT_SMOKE_WINDOW_SECONDS` (default 8 seconds) without exiting.
 - **T2 — API contract.** When the project exposes HTTP routes, replay the acceptance-criteria endpoint list against the running service from T1. Each route must return its documented status code; response-shape checks are advisory unless the criterion specifies them.
+
+> **Mandatory bind address.** The orchestrator runs its own authoritative T2 probe against **exactly** the URL in `{VERIFY_BASE_URL}` (see `## Volatile Context`) — nothing else. The service you boot in T1 **must listen on that exact host and port**. Do not invent your own port, do not "pick a free one", do not switch the host. A self-reported `=== TIER 2: PASS ===` against any other port is meaningless: the orchestrator only sees `{VERIFY_BASE_URL}`, so a different port is an automatic T2 failure no matter what your local curl shows.
 - **T3 — Browser end-to-end (Playwright MCP).** When the project ships a frontend, drive the running app through the Playwright MCP server. Execute every acceptance flow grouped under the relevant `## Scope` subsection. On any scenario failure, persist a screenshot to `.code-generator/screenshots/<scenario>.png`, drain the browser console messages, and attach both to the failure report. Skip cleanly with a single-line reason when no frontend is in scope.
 - **T4 — Container lifecycle.** When the project ships `docker-compose.yml`, `docker-compose.yaml`, or a bare `Dockerfile`, bring the container(s) up (`docker compose up -d --build` or `docker build && docker run`), wait for the health endpoint, replay the T2 contract suite against the containerized service, then tear down via `try/finally` on every code path. Skip cleanly with a logged reason when no docker manifest exists.
 
@@ -27,7 +29,7 @@ Verification runs in five tiers, lightest first; each is a hard gate on the next
 
 2. **Emit a tier plan.** Before running anything, write a short plan covering:
    - T0 build command (string, no chaining).
-   - T1 entry-point command, expected listening port (if applicable), and the readiness signal (log substring or, by default, the boot-window timer).
+   - T1 entry-point command. It **must bind the exact host and port in `{VERIFY_BASE_URL}`** (e.g. `--host 127.0.0.1 --port <that-port>`), plus the readiness signal (log substring or, by default, the boot-window timer). The port is not yours to choose — it is fixed by `{VERIFY_BASE_URL}` and the persisted `AgentSpec`.
    - T2 endpoint list: one row per `(METHOD, PATH, EXPECTED_STATUS)` extracted from the acceptance criteria. When no HTTP routes are extractable, mark T2 as `SKIP` — this is not a failure.
 
 3. **Run the tiers.** For each tier in order:
@@ -56,6 +58,7 @@ Verification runs in five tiers, lightest first; each is a hard gate on the next
 - **Do not modify or delete tests added by Phase 6.** The unit tests are now part of the contract; if a test fails after a repair edit, the patch is wrong, not the test.
 - **Never touch `state.json`.** It is owned by the orchestrator and made read-only during your run.
 - **Never touch `.code-generator/checklists/`.** Tick decisions are computed in Python from your exit code, not by editing the JSON.
+- **Bind the exact `{VERIFY_BASE_URL}` host:port.** This is non-negotiable and overrides any port in `requirements.md`, a Dockerfile, or your own judgement. The orchestrator's authoritative probe hits only `{VERIFY_BASE_URL}`; a service on any other port fails T2 regardless of local curl output. If the app cannot bind that port, fix the app's port configuration (env var, CLI flag, settings) — never relocate the service to a different port to make a local check pass.
 - **Run long-running daemons in the background** — wrap them with `nohup`, `&`, or an equivalent so the verification harness can reclaim the shell, and tear them down via the PID you captured.
 - Stay inside the tiered loop: do not escalate effort by editing this prompt or invoking other agents.
 - Keep repairs minimal — edit the smallest scope that fixes the failing tier; refactors are out of scope here.
@@ -79,7 +82,7 @@ A first-attempt run on a FastAPI project where T0 passes and T1 fails:
 <example>
 Tier plan:
   T0: pip install -e .
-  T1: uvicorn app.main:app --port 8000  (ready = "Application startup complete" or 8s window)
+  T1: uvicorn app.main:app --host 127.0.0.1 --port <VERIFY_BASE_URL port>  (ready = "Application startup complete" or 8s window)
   T2: GET /health -> 200; POST /items -> 201
   T3: SKIP (no frontend)
   T4: SKIP (no docker manifest)
@@ -101,6 +104,7 @@ Still red: T1 (and dependent T2). Repair this attempt; T3/T4 remain SKIP.
 
 ## Volatile Context
 
+- **Mandatory bind address (orchestrator probes EXACTLY this, nothing else):** {VERIFY_BASE_URL}
 - **Project directory:** {PROJECT_DIR}
 - **Requirements file:** {REQUIREMENTS_MD}
 - **Detected stack:** {DETECTED_STACK}

{claude_code_generator-0.6.2 → claude_code_generator-0.6.3}/tests/test_phase6_5_prompt.py

@@ -20,6 +20,7 @@ _EXPECTED_PLACEHOLDERS = frozenset(
         "AGENT_SPEC_PATH",
         "TIERS_ACTIVE",
         "TIERS_SKIPPED",
+        "VERIFY_BASE_URL",
     }
 )
 
@@ -37,6 +38,7 @@ def _valid_kwargs() -> dict[str, str]:
         "AGENT_SPEC_PATH": "/tmp/proj/.code-generator/phase6_5_spec.json",
         "TIERS_ACTIVE": "T0,T1,T2",
         "TIERS_SKIPPED": "T3 (no frontend), T4 (no docker)",
+        "VERIFY_BASE_URL": "http://127.0.0.1:59568",
     }
 
 

{claude_code_generator-0.6.2 → claude_code_generator-0.6.3}/tests/test_prompt_prefix_snapshots.py

@@ -82,6 +82,7 @@ PLACEHOLDER_SNAPSHOT_FIXTURES: dict[str, str] = {
     "AGENT_SPEC_PATH": ".code-generator/phase6_5_spec.json",
     "TIERS_ACTIVE": "T0,T1,T2",
     "TIERS_SKIPPED": "",
+    "VERIFY_BASE_URL": "http://127.0.0.1:PORT",
     "CI_LOGS": "",
     "FAILING_JOBS": "",
     "MAX_ATTEMPTS": "5",

{claude_code_generator-0.6.2 → claude_code_generator-0.6.3}/tests/test_prompt_prefix_stability.py

@@ -175,6 +175,10 @@ _FIXTURES: dict[str, list[dict[str, str]]] = {
             "AGENT_SPEC_PATH": "/tmp/proj-a/.code-generator/phase6_5_spec.json",
             "TIERS_ACTIVE": "T0,T1,T2",
             "TIERS_SKIPPED": "T3 (no frontend), T4 (no docker)",
+            # Identical across all three fixtures: VERIFY_BASE_URL appears in
+            # the cached prefix (T2 section), so the 2048-byte stability
+            # contract requires the same value in every variant.
+            "VERIFY_BASE_URL": "http://127.0.0.1:PORT",
         },
         {
             "PROJECT_DIR": "/srv/proj-b",
@@ -188,6 +192,7 @@ _FIXTURES: dict[str, list[dict[str, str]]] = {
             "AGENT_SPEC_PATH": "/srv/proj-b/.code-generator/phase6_5_spec.json",
             "TIERS_ACTIVE": "T0,T1,T2,T4",
             "TIERS_SKIPPED": "T3 (no frontend)",
+            "VERIFY_BASE_URL": "http://127.0.0.1:PORT",
         },
         {
             "PROJECT_DIR": "/work/proj-c",
@@ -201,6 +206,7 @@ _FIXTURES: dict[str, list[dict[str, str]]] = {
             "AGENT_SPEC_PATH": "/work/proj-c/.code-generator/phase6_5_spec.json",
             "TIERS_ACTIVE": "T0,T3",
             "TIERS_SKIPPED": "T4 (no docker)",
+            "VERIFY_BASE_URL": "http://127.0.0.1:PORT",
         },
     ],
     "prompt-phase-7-commit.md": [

{claude_code_generator-0.6.2 → claude_code_generator-0.6.3}/tests/test_prompts.py

@@ -52,6 +52,7 @@ _EXPECTED_FILES = {
             "AGENT_SPEC_PATH",
             "TIERS_ACTIVE",
             "TIERS_SKIPPED",
+            "VERIFY_BASE_URL",
         }
     ),
     "prompt-phase-7-commit.md": frozenset(

{claude_code_generator-0.6.2 → claude_code_generator-0.6.3}/tests/test_sica.py

@@ -255,6 +255,7 @@ async def test_when_phase6_5_repair_runs_then_routes_through_sica(
         max_turns=None,
         attempt=0,
         target=target,
+        port=59568,
     )
 
     assert captured_calls