superqode 2.4.7__tar.gz → 2.4.9__tar.gz

{superqode-2.4.7 → superqode-2.4.9}/CHANGELOG.md

@@ -5,6 +5,40 @@ All notable changes to SuperQode will be documented in this file.
 The format is based on [Keep a Changelog](https://keepachangelog.com/en/1.0.0/),
 and this project adheres to [Semantic Versioning](https://semver.org/spec/v2.0.0.html).
 
+## [Unreleased]
+
+## [2.4.9] - 2026-09-20
+
+### Changed
+- Install banner and welcome screen call out Smart Decisions with Jev and SystemOne Harness beside ACP / A2A / UHP.
+- Shortcut bar adds `🧭 :systemone` next to connect, hub, home, and help.
+
+## [2.4.8] - 2026-09-20
+
+### 🧠 SystemOne Tune
+
+- 🎯 Fully unlabeled datasets now use an active-learning loop: Jev ranks the
+  development pool by uncertainty, adds a random audit example and a random
+  sealed example, and asks for only 2 to 20 judgments per round.
+- 🔁 Tune sessions persist accepted and rejected candidates across rounds. A
+  pending decision can be resumed without repeating GEPA or Jev calls.
+- 🧪 Developers can explicitly accept a small candidate as experimental while the
+  existing evidence threshold remains required for verified adoption.
+- 💾 Pool predictions are cached for rejected rounds, and each round keeps its own
+  GEPA output, evaluation journal, held-out comparison, and report.
+- 🎬 A checked-in unlabeled routing pool and demo guide exercise the complete
+  acquire, label, optimize, accept/reject, and resume workflow.
+
+### 🛠️ Fixed
+
+- 🏠 SystemOne Tune preferences are best-effort, so a read-only user directory no
+  longer blocks preparing, labeling, or starting an experiment.
+- 🔓 Custom no-auth SystemOne endpoints no longer incorrectly require
+  `TYPESAFE_API_KEY`, including in the TUI preflight.
+- 📊 The Tune pilot gate counts both search and validation rows as
+  optimizer-visible development evidence. The checked-in 35-development / 35-test
+  routing dataset is no longer misclassified as having only 27 development rows.
+
 ## [2.4.7] - 2026-09-20
 
 ### Fixed

{superqode-2.4.7 → superqode-2.4.9}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: superqode
-Version: 2.4.7
+Version: 2.4.9
 Summary: The harness layer for coding agents
 Author-email: Shashi Jagtap <info@super-agentic.ai>
 Maintainer-email: Shashi Jagtap <info@super-agentic.ai>
@@ -238,6 +238,10 @@ Teach SystemOne how your team routes or judges work from reviewed examples, then
 superqode harness tune --setup          # once: install pinned GEPA support
 superqode harness tune                  # guided labeling + experiment
 superqode harness tune --demo --live    # small synthetic routing demo
+
+# From a repository checkout: uncertainty-ranked, multi-round demo
+superqode harness tune --data examples/tune/factory-route-active.csv \
+  --batch-size 5 --max-evals 30 --max-reflection-cost 0.50
 ```
 
 In the TUI, run `:systemone tune` or open **Improve decisions** on SystemOne in Harness Hub. See the [SystemOne Tune guide](https://docs.superqode.dev/advanced/systemone/#improve-decisions-with-systemone-tune).

{superqode-2.4.7 → superqode-2.4.9}/README.md

@@ -83,6 +83,10 @@ Teach SystemOne how your team routes or judges work from reviewed examples, then
 superqode harness tune --setup          # once: install pinned GEPA support
 superqode harness tune                  # guided labeling + experiment
 superqode harness tune --demo --live    # small synthetic routing demo
+
+# From a repository checkout: uncertainty-ranked, multi-round demo
+superqode harness tune --data examples/tune/factory-route-active.csv \
+  --batch-size 5 --max-evals 30 --max-reflection-cost 0.50
 ```
 
 In the TUI, run `:systemone tune` or open **Improve decisions** on SystemOne in Harness Hub. See the [SystemOne Tune guide](https://docs.superqode.dev/advanced/systemone/#improve-decisions-with-systemone-tune).

superqode-2.4.9/examples/tune/README.md

@@ -0,0 +1,59 @@
+# Free SystemOne Tune demo data
+
+Original synthetic routing tickets for the `factory_route` pack.
+No third-party corpus. Free to use in demos and tests.
+
+- `factory-route-free.csv`: 70 labeled examples (35 development / 35 test)
+- `factory-route-active.csv`: the same 70 inputs with blank labels and
+  rationales, for the uncertainty-ranked, multi-round workflow
+- Tune deterministically reserves 8 of the 35 development rows for validation,
+  leaving 27 search rows. Both subsets are optimizer-visible; the 35 test rows
+  remain sealed until the final comparison.
+- Columns: `id`, `state`, `label`, `split`, `rationale`
+- Labels: `private`, `local`, `cheap`, `best`, `review`, `long-context`, `no-subscription`
+
+## Active-learning demo
+
+From the repository root, configure Jev and one reflection provider, then run:
+
+```sh
+export TYPESAFE_API_KEY="..."
+export OPENAI_API_KEY="..." # or Anthropic, Gemini, or another LiteLLM provider
+
+superqode harness tune --setup
+superqode harness tune \
+  --data examples/tune/factory-route-active.csv \
+  --batch-size 5 \
+  --max-evals 30 \
+  --max-reflection-cost 0.50
+```
+
+The first round scores 35 development inputs, then asks for five development
+judgments and one randomly reserved test judgment. Each card says whether it
+was selected for uncertainty, random audit, or sealed evaluation. Add an
+optional rationale when a routing boundary matters.
+
+After GEPA finishes, review the metric change and question diff. Choose
+`accept` to keep the proposal as an experimental version, `reject` to keep the
+current pack, or `later` to leave the decision pending. The terminal prints the
+saved run directory.
+
+Resume the next round with:
+
+```sh
+superqode harness tune --resume .superqode/tuning/<run-id>
+```
+
+For a non-interactive pending decision:
+
+```sh
+superqode harness tune --resume .superqode/tuning/<run-id> \
+  --accept --experimental --json
+
+superqode harness tune --resume .superqode/tuning/<run-id> \
+  --reject --json
+```
+
+This demo is intentionally a pilot. A candidate becomes verified only after it
+passes the regression gate with at least 30 reviewed development and 30 sealed
+test examples.

{superqode-2.4.7 → superqode-2.4.9}/pyproject.toml

@@ -1,6 +1,6 @@
 [project]
 name = "superqode"
-version = "2.4.7"
+version = "2.4.9"
 description = "The harness layer for coding agents"
 readme = "README.md"
 license = "Apache-2.0"

{superqode-2.4.7 → superqode-2.4.9}/src/superqode/__init__.py

@@ -38,7 +38,7 @@ __all__ = [
     "sidebar",
 ]
 
-__version__ = "2.4.7"
+__version__ = "2.4.9"
 
 # Stable, lightweight public extension surface.  Importing this package does
 # not discover or execute third-party extensions; discovery happens only when

{superqode-2.4.7 → superqode-2.4.9}/src/superqode/app/welcome.py

@@ -230,7 +230,7 @@ def render_welcome(
     desc_text.append("Terminal-first · Any agent or model\n", style=f"bold {THEME['purple']}")
     if not narrow:
         desc_text.append("\n", style="")
-    interoperability = "Local · ACP · MCP · A2A · UHP · BYOK · System One Models"
+    interoperability = "Local · ACP · MCP · A2A · UHP · BYOK · Jev SystemOne Harness"
     if narrow:
         desc_text.append(interoperability, style=THEME["muted"])
     else:

{superqode-2.4.7 → superqode-2.4.9}/src/superqode/app/widgets.py

@@ -1084,6 +1084,7 @@ class HintsBar(Static):
             hints = [
                 ("⏏", ":disconnect", THEME["pink"]),
                 ("⚓", ":hub", THEME["link"]),
+                ("🧭", ":systemone", THEME["cyan"]),
                 ("🧠", ":memory", THEME["link"]),
                 ("📊", ":eval", THEME["link"]),
                 ("⚡", ":skills", THEME["link"]),
@@ -1094,6 +1095,7 @@ class HintsBar(Static):
             hints = [
                 ("🔌", ":connect", THEME["pink"]),
                 ("⚓", ":hub", THEME["link"]),
+                ("🧭", ":systemone", THEME["cyan"]),
                 ("🏠", ":home", THEME["link"]),
                 ("?", ":help", THEME["link"]),
             ]

{superqode-2.4.7 → superqode-2.4.9}/src/superqode/commands/harness/tune.py

@@ -54,6 +54,20 @@ from superqode.systemone import tune
     show_default=True,
 )
 @click.option("--seed", type=int, default=0)
+@click.option(
+    "--batch-size",
+    type=click.IntRange(min=2, max=20),
+    default=5,
+    show_default=True,
+    help="Examples to review per active-learning round.",
+)
+@click.option("--accept", "accept_candidate", is_flag=True, help="Accept a pending candidate.")
+@click.option("--reject", "reject_candidate", is_flag=True, help="Reject a pending candidate.")
+@click.option(
+    "--experimental",
+    is_flag=True,
+    help="Allow --accept before the candidate has enough evidence to be verified.",
+)
 @click.option(
     "--live", is_flag=True, help="Start model calls without the interactive start prompt."
 )
@@ -77,6 +91,10 @@ def harness_tune(
     max_evals,
     max_reflection_cost,
     seed,
+    batch_size,
+    accept_candidate,
+    reject_candidate,
+    experimental,
     live,
     json_output,
 ):
@@ -87,6 +105,10 @@ def harness_tune(
     """
     interactive = sys.stdin.isatty() and not json_output
     try:
+        if accept_candidate and reject_candidate:
+            raise ValueError("Choose either --accept or --reject, not both.")
+        if (accept_candidate or reject_candidate) and not resume:
+            raise ValueError("Use --accept or --reject together with --resume.")
         if setup:
             message = tune.install_support()
             click.echo(
@@ -109,6 +131,7 @@ def harness_tune(
                 for name, value in (
                     ("max_evals", max_evals),
                     ("max_reflection_cost", max_reflection_cost),
+                    ("batch_size", batch_size),
                 ):
                     if (
                         click.get_current_context().get_parameter_source(name)
@@ -152,6 +175,7 @@ def harness_tune(
                 max_evals=max_evals,
                 max_reflection_cost=max_reflection_cost,
                 seed=seed,
+                batch_size=batch_size,
             )
             rows = (
                 tune.demo_examples()
@@ -164,15 +188,52 @@ def harness_tune(
             report = json.loads((output / "report.json").read_text())
             click.echo(json.dumps(report) if json_output else tune.render_report(report))
             return
+        if manifest["status"] == "decision":
+            report = json.loads((output / "report.json").read_text())
+            decision = "accept" if accept_candidate else "reject" if reject_candidate else ""
+            if not decision and interactive:
+                decision = click.prompt(
+                    tune.render_report(report) + "\nChoose",
+                    type=click.Choice(["accept", "reject", "later"]),
+                    default="later",
+                )
+            if decision in {"accept", "reject"}:
+                report = tune.decide_candidate(
+                    output,
+                    decision,
+                    allow_experimental=experimental,
+                )
+            click.echo(json.dumps(report) if json_output else tune.render_report(report))
+            return
         if manifest["status"] != "review":
             raise ValueError(
                 f"Experiment is {manifest['status']}. Evidence is in {output}; start a new experiment for a new budget."
             )
+        if manifest.get("workflow") == "active" and not tune.review_batch(manifest):
+            if not live:
+                if not interactive:
+                    raise ValueError(
+                        f"Unlabeled pool saved. Add --live to --resume {output} to select a review batch."
+                    )
+                if not click.confirm(
+                    "Use Jev to find uncertain examples plus a random audit sample?", default=True
+                ):
+                    return
+            tune.acquire_review_batch(
+                output,
+                progress=(lambda message: click.echo(message, err=True))
+                if not json_output
+                else lambda _: None,
+            )
+            manifest = tune.load_run(output)
         question = next(iter(manifest["pack"]["questions"].values()))
         labels = list(question["criteria"])
-        for row in manifest["examples"]:
-            if row.get("label") is not None:
-                continue
+        rows_to_review = (
+            tune.review_batch(manifest)
+            if manifest.get("workflow") == "active"
+            else [row for row in manifest["examples"] if row.get("label") is None]
+        )
+        for row in rows_to_review:
             if not interactive:
                 raise ValueError(
                     f"Unlabeled examples saved. Resume in a terminal: superqode harness tune --resume {output}"
@@ -213,6 +274,18 @@ def harness_tune(
             if not json_output
             else lambda _: None,
         )
+        if report.get("status") == "pending_decision" and interactive:
+            decision = click.prompt(
+                tune.render_report(report) + "\nChoose",
+                type=click.Choice(["accept", "reject", "later"]),
+                default="later",
+            )
+            if decision in {"accept", "reject"}:
+                report = tune.decide_candidate(
+                    output,
+                    decision,
+                    allow_experimental=(decision == "accept" and not report["eligible"]),
+                )
         click.echo(json.dumps(report) if json_output else tune.render_report(report))
     except KeyboardInterrupt:
         raise click.ClickException(f"Stopped. Saved work: {output}") from None