pdl-taskmaster 2.4.0__tar.gz → 2.5.1__tar.gz

{pdl_taskmaster-2.4.0 → pdl_taskmaster-2.5.1}/.gitignore

@@ -7,7 +7,9 @@ dist/
 build/
 *.egg-info/
 mlflow.db
+**/mlflow.db
 mlruns/
+**/mlruns/
 runs/
 scratch/
 *.log

pdl_taskmaster-2.5.1/NOTICE.md

@@ -0,0 +1,28 @@
+# Notices and attribution
+
+## Project license
+
+Original project code and documentation are released under the
+[MIT License](LICENSE).
+
+## Pseudocode-standard compatibility
+
+The local PDL compatibility profile was informed by a supplied HTML snapshot
+titled **Pseudocode Standard**, attributed in its source history to J. Dalbey
+and published on the Cal Poly course site at:
+
+<https://users.csc.calpoly.edu/~jdalbey/SWE/pdl_std.html>
+
+The public bundle does not include the supplied HTML snapshot or copy its prose
+or examples. It contains a project-authored description of the functional
+notation and an original set of examples needed for compatibility. That local
+profile is released under MIT with the project's other original contributions.
+The source title, author attribution, and site name are used only for provenance
+and compatibility identification; their rights remain with their respective
+owners.
+
+## Evaluation models and product names
+
+Model, product, and organization names in the evaluation documentation are
+used for identification and do not imply endorsement. Their marks remain the
+property of their respective owners.

{pdl_taskmaster-2.4.0 → pdl_taskmaster-2.5.1}/PKG-INFO

@@ -1,10 +1,11 @@
-Metadata-Version: 2.5
+Metadata-Version: 2.3
 Name: pdl-taskmaster
-Version: 2.4.0
+Version: 2.5.1
 Summary: Deterministic protocol governance and bounded AI agent harness
 Author: PDL-Standard Contributors
-License-Expression: MIT
+License: MIT
 License-File: LICENSE
+License-File: NOTICE.md
 Classifier: Development Status :: 5 - Production/Stable
 Classifier: Intended Audience :: Developers
 Classifier: License :: OSI Approved :: MIT License
@@ -52,6 +53,8 @@ docs/                      documentation
   architecture/            framing.md, whitepaper.md, protocol specs
   operations/              eval-metrics.md, efficiency-report.md
   governance/              roadmap.md, experiment-log.md (full decision history)
+  releases/                v2.5.0.md release notes
+  ABLATIONS.md             substantive correctness and dual-plane empirical ablations
   adr/ trd/                background architecture-decision records (history, not required reading)
 ```
 
@@ -93,7 +96,7 @@ pdlt                        # Launch interactive REPL (default)
 
 ```powershell
 pdlt verify                 # Deterministic baseline verifier
-pytest                      # Run complete test suite (offline, 87 passed)
+pytest                      # Run complete test suite (offline, 174 passed, 2 skipped)
 ```
 
 The verifier checks required runtime/instructional files, imports, the zero-template workspace invariant, fixture hashes, REPL subprocess-script presence, source-repository isolation, a fresh-workspace lifecycle test (prompt → plan → execute → result), and resuming that same workspace. All verifier workspaces are temporary.
@@ -107,22 +110,34 @@ pdlt --worker recorded --case-ids G06 --new-session
 
 Recorded mode is **exact, deterministic replay**: it only responds to the exact interaction sequences captured in the fixture (`G06`: full lifecycle; `A02`: prompt revision). Use `--quit` to exit — there's no `/exit` command.
 
-Commands: `/help`, `/status`, `/session`, `/new`, `/resume`, `/mlflow [on|off]`, `/tokens [on|off]`, `/timeout [seconds]`, `/model [name]`, `/worker [api|codex|recorded]`, `/config` (codex only), `/sandbox` (codex only), `/workdir [path]`, `/transcript [path]`, `/quit`.
+Commands: `/help`, `/status`, `/session`, `/new`, `/resume`, `/dev [on|off]`, `/mlflow [on|off]`, `/tokens [on|off]`, `/timeout [seconds]`, `/model [name]`, `/worker [api|codex|recorded]`, `/config` (codex only), `/sandbox` (codex only), `/workdir [path]`, `/transcript [path]`, `/paste` (or `"""`), `/quit`.
 
 Fast-path review commands: `/confirm`, `/revise <feedback>`, `/stop`.
 
 ## Live model worker (default)
 
 ```powershell
-# Default worker is 'api' with model 'z-ai/glm-4.7'
+# Default worker is 'api' with model 'openai/gpt-oss-120b'
 pdlt --new-session
 # or: python -m pdl_taskmaster.host.repl --candidate-repo . --new-session
 ```
 
-> **System 1 / System 2 Architecture Note**: Production default is `--worker api` (System 2, e.g. `z-ai/glm-4.7`). Local fast System 1 classification models are undergoing contrastive RLCD fine-tuning (ADR-0012) and fail-closed to System 2 via the confidence ladder if threshold ($0.85$) or top-2 margin ($0.40$) are unmet.
+> **System 1 / System 2 Architecture Note**: Production default is `--worker api` (System 2: `openai/gpt-oss-120b` with reasoning effort `low`; benchmarks evaluated on `z-ai/glm-4.7`). Local fast System 1 classification models are undergoing contrastive RLCD fine-tuning (ADR-0012) and fail-closed to System 2 via the confidence ladder if threshold ($0.85$) or top-2 margin ($0.40$) are unmet.
 
 `--worker api` sends the compiled interpretation/plan directly to an OpenAI-compatible `/responses` endpoint (instructions bundled in `src/pdl_taskmaster/runtime/worker-bootstrap.txt`), with no tool definitions, sandbox, or agentic system prompt attached — deliberately; see the "instruction-lightness" finding in the whitepaper (§3).
 
+### REPL Dev Mode & Substantive Verification
+
+```powershell
+pdlt --dev                  # Launch with live telemetry inspector and stage controls
+pdlt --dev --exit-on-close  # Headless test run: exit when CLOSED_SUCCESS or error reached
+```
+
+- **Interactive Telemetry Inspector**: Displays real-time controller stage transitions, entity drops, and verification events.
+- **Substantive Witness Verification & OS Sandbox**: For combinatorial and algorithmic tasks, execution occurs within an OS-confined sandbox (Job Objects on Windows, `rlimit` on POSIX) with deterministic witness verification ([ADR-0013](docs/adr/0013-substantive-correctness-verification.md), [ADR-0015](docs/adr/0015-model-synthesized-verification-and-confinement-boundaries.md)).
+- **Deliverable Formatting**: Automatic ASCII-safe human/agent deliverable cards (`format_friendly_deliverable`) replacing raw Result IR wire blocks in user view.
+- **Empirical Benchmarks & Ablations**: Full evaluation results comparing System 1, ungrounded System 2, and the dual-plane harness: [`docs/ABLATIONS.md`](docs/ABLATIONS.md).
+
 ### Efficiency flags (`--worker api`)
 
 - `--api-reasoning-effort low|medium|high` — reasoning budget override, applied globally.

{pdl_taskmaster-2.4.0 → pdl_taskmaster-2.5.1}/README.md

@@ -26,6 +26,8 @@ docs/                      documentation
   architecture/            framing.md, whitepaper.md, protocol specs
   operations/              eval-metrics.md, efficiency-report.md
   governance/              roadmap.md, experiment-log.md (full decision history)
+  releases/                v2.5.0.md release notes
+  ABLATIONS.md             substantive correctness and dual-plane empirical ablations
   adr/ trd/                background architecture-decision records (history, not required reading)
 ```
 
@@ -67,7 +69,7 @@ pdlt                        # Launch interactive REPL (default)
 
 ```powershell
 pdlt verify                 # Deterministic baseline verifier
-pytest                      # Run complete test suite (offline, 87 passed)
+pytest                      # Run complete test suite (offline, 174 passed, 2 skipped)
 ```
 
 The verifier checks required runtime/instructional files, imports, the zero-template workspace invariant, fixture hashes, REPL subprocess-script presence, source-repository isolation, a fresh-workspace lifecycle test (prompt → plan → execute → result), and resuming that same workspace. All verifier workspaces are temporary.
@@ -81,22 +83,34 @@ pdlt --worker recorded --case-ids G06 --new-session
 
 Recorded mode is **exact, deterministic replay**: it only responds to the exact interaction sequences captured in the fixture (`G06`: full lifecycle; `A02`: prompt revision). Use `--quit` to exit — there's no `/exit` command.
 
-Commands: `/help`, `/status`, `/session`, `/new`, `/resume`, `/mlflow [on|off]`, `/tokens [on|off]`, `/timeout [seconds]`, `/model [name]`, `/worker [api|codex|recorded]`, `/config` (codex only), `/sandbox` (codex only), `/workdir [path]`, `/transcript [path]`, `/quit`.
+Commands: `/help`, `/status`, `/session`, `/new`, `/resume`, `/dev [on|off]`, `/mlflow [on|off]`, `/tokens [on|off]`, `/timeout [seconds]`, `/model [name]`, `/worker [api|codex|recorded]`, `/config` (codex only), `/sandbox` (codex only), `/workdir [path]`, `/transcript [path]`, `/paste` (or `"""`), `/quit`.
 
 Fast-path review commands: `/confirm`, `/revise <feedback>`, `/stop`.
 
 ## Live model worker (default)
 
 ```powershell
-# Default worker is 'api' with model 'z-ai/glm-4.7'
+# Default worker is 'api' with model 'openai/gpt-oss-120b'
 pdlt --new-session
 # or: python -m pdl_taskmaster.host.repl --candidate-repo . --new-session
 ```
 
-> **System 1 / System 2 Architecture Note**: Production default is `--worker api` (System 2, e.g. `z-ai/glm-4.7`). Local fast System 1 classification models are undergoing contrastive RLCD fine-tuning (ADR-0012) and fail-closed to System 2 via the confidence ladder if threshold ($0.85$) or top-2 margin ($0.40$) are unmet.
+> **System 1 / System 2 Architecture Note**: Production default is `--worker api` (System 2: `openai/gpt-oss-120b` with reasoning effort `low`; benchmarks evaluated on `z-ai/glm-4.7`). Local fast System 1 classification models are undergoing contrastive RLCD fine-tuning (ADR-0012) and fail-closed to System 2 via the confidence ladder if threshold ($0.85$) or top-2 margin ($0.40$) are unmet.
 
 `--worker api` sends the compiled interpretation/plan directly to an OpenAI-compatible `/responses` endpoint (instructions bundled in `src/pdl_taskmaster/runtime/worker-bootstrap.txt`), with no tool definitions, sandbox, or agentic system prompt attached — deliberately; see the "instruction-lightness" finding in the whitepaper (§3).
 
+### REPL Dev Mode & Substantive Verification
+
+```powershell
+pdlt --dev                  # Launch with live telemetry inspector and stage controls
+pdlt --dev --exit-on-close  # Headless test run: exit when CLOSED_SUCCESS or error reached
+```
+
+- **Interactive Telemetry Inspector**: Displays real-time controller stage transitions, entity drops, and verification events.
+- **Substantive Witness Verification & OS Sandbox**: For combinatorial and algorithmic tasks, execution occurs within an OS-confined sandbox (Job Objects on Windows, `rlimit` on POSIX) with deterministic witness verification ([ADR-0013](docs/adr/0013-substantive-correctness-verification.md), [ADR-0015](docs/adr/0015-model-synthesized-verification-and-confinement-boundaries.md)).
+- **Deliverable Formatting**: Automatic ASCII-safe human/agent deliverable cards (`format_friendly_deliverable`) replacing raw Result IR wire blocks in user view.
+- **Empirical Benchmarks & Ablations**: Full evaluation results comparing System 1, ungrounded System 2, and the dual-plane harness: [`docs/ABLATIONS.md`](docs/ABLATIONS.md).
+
 ### Efficiency flags (`--worker api`)
 
 - `--api-reasoning-effort low|medium|high` — reasoning budget override, applied globally.

{pdl_taskmaster-2.4.0 → pdl_taskmaster-2.5.1}/docs/adr/README.md

@@ -14,6 +14,11 @@
 | [ADR-0010](0010-pydantic-wire-enforcement.md) | Accepted | Pydantic v2 schema enforcement and automated operator retry corrections for wire contracts. |
 | [ADR-0011](0011-in-memory-vfs-and-microvm-sandboxing.md) | Accepted | Software-defined in-memory VFS context flow and ephemeral MicroVM agent sandboxing via MCP. |
 | [ADR-0012](0012-system-1-decision-models-via-rlcd.md) | Accepted | Realign Track L to System 1 decision models (Laya/Jev) aligned via RLCD contrastive distillation. |
+| [ADR-0013](0013-substantive-correctness-verification.md) | Proposed | Verified execution and bidirectional witness retention for substantive correctness. |
+| [ADR-0014](0014-dual-plane-boundary-and-wire-conformance.md) | Accepted | Dual-plane boundary architecture: positive structural guidance, wire-level Pydantic contract enforcement, and tripartite clause governance. |
+| [ADR-0015](0015-model-synthesized-verification-and-confinement-boundaries.md) | Proposed | Model-synthesized verification and confinement boundaries in ephemeral sandboxes. |
+| [ADR-0016](0016-pydantic-ssot-wire-and-deliverable-boundary-enforcement.md) | Accepted | Pydantic Single Source of Truth (SSOT) for wire and deliverable boundary enforcement. |
+| [ADR-0017](0017-dual-plane-runtime-realignment-and-mrv-solver-governance.md) | Accepted | Dual-Plane Runtime Realignment: System 1 (Jev) Baseline & Constraint-Ordered Solver Governance. |
 
 The records are intentionally separated so UI, context construction, execution
 isolation, output representation, and reasoning policy can evolve without

{pdl_taskmaster-2.4.0 → pdl_taskmaster-2.5.1}/pyproject.toml

@@ -1,5 +1,5 @@
 [build-system]
-requires = ["hatchling"]
+requires = ["hatchling<1.26.0"]
 build-backend = "hatchling.build"
 
 [project]
@@ -8,8 +8,7 @@ dynamic = ["version"]
 description = "Deterministic protocol governance and bounded AI agent harness"
 readme = "README.md"
 requires-python = ">=3.10"
-license = "MIT"
-license-files = ["LICENSE"]
+license = { text = "MIT" }
 authors = [
     { name = "PDL-Standard Contributors" },
 ]

{pdl_taskmaster-2.4.0 → pdl_taskmaster-2.5.1}/src/pdl_taskmaster/__init__.py

@@ -1,6 +1,6 @@
 """PDL-Taskmaster: Deterministic protocol governance and bounded AI agent harnesses."""
 
-__version__ = "2.4.0"
+__version__ = "2.5.1"
 __protocol_version__ = "v2"
 
 __all__ = ["__version__", "__protocol_version__"]

{pdl_taskmaster-2.4.0 → pdl_taskmaster-2.5.1}/src/pdl_taskmaster/controller/schemas/result_ir_repair.schema.json

@@ -16,6 +16,27 @@
         "open_defects"
       ],
       "properties": {
+        "witness": {
+          "type": "object",
+          "description": "Witness certifying substantive execution correctness (ADR-0013 / ADR-0015).",
+          "properties": {
+            "polarity": { "type": "string", "enum": ["positive", "negative"] },
+            "evidence": {
+              "type": "object",
+              "properties": {
+                "path": { "type": "string" },
+                "section": { "type": "string" },
+                "observed": { "type": "string" }
+              },
+              "required": ["path"]
+            },
+            "data": { "type": "object" },
+            "search_exhausted": { "type": "boolean" },
+            "nodes_explored": { "type": "integer" },
+            "method": { "type": "string" }
+          },
+          "required": ["polarity"]
+        },
         "files": {
           "type": "array",
           "items": {

{pdl_taskmaster-2.4.0 → pdl_taskmaster-2.5.1}/src/pdl_taskmaster/host/app.py

@@ -84,6 +84,7 @@ class PDLtHost:
                     higher_priority_constraints=self.higher_priority_constraints,
                     available_execution_tools=[],
                     render_compact=self.render_compact,
+                    sys1_client=getattr(self.worker, "sys1_client", None),
                 )
             except Exception as exc:
                 # Graceful degradation: a session with no committed protocol
@@ -99,6 +100,7 @@ class PDLtHost:
                 available_execution_tools=[],
                 workspace_root=self.workspace_root,
                 render_compact=self.render_compact,
+                sys1_client=getattr(self.worker, "sys1_client", None),
             )
         self.engine = engine
         if self.observation_dir is not None:

{pdl_taskmaster-2.4.0 → pdl_taskmaster-2.5.1}/src/pdl_taskmaster/host/cli.py

@@ -73,7 +73,13 @@ def _cmd_verify(argv: list[str]) -> int:
     return verify_main()
 
 
-def main(argv: list[str] | None = None) -> int:
+def _cmd_viewer(argv: list[str]) -> int:
+    from pdl_taskmaster.tools.viewer_server import main as viewer_main
+
+    return viewer_main(argv)
+
+
+def _main_impl(argv: list[str] | None = None) -> int:
     if argv is None:
         argv = sys.argv[1:]
 
@@ -89,6 +95,8 @@ def main(argv: list[str] | None = None) -> int:
         return _cmd_init(argv[1:])
     if subcmd == "verify":
         return _cmd_verify(argv[1:])
+    if subcmd in {"viewer", "ui", "dashboard"}:
+        return _cmd_viewer(argv[1:])
 
     # Default: route all other invocations to REPL
     from pdl_taskmaster.host.repl import main as repl_main
@@ -96,5 +104,13 @@ def main(argv: list[str] | None = None) -> int:
     return repl_main()
 
 
+def main(argv: list[str] | None = None) -> int:
+    try:
+        return _main_impl(argv)
+    except KeyboardInterrupt:
+        print("\n[session terminated by user]", file=sys.stderr, flush=True)
+        return 130
+
+
 if __name__ == "__main__":
     raise SystemExit(main())