feat: 扩展 Dyro 只读控制面与安全边界

.github/workflows/pypi-publish.yml

@@ -108,6 +108,33 @@ jobs:
       - name: Build distributions
         run: uv run python -m build
 
+      - name: Reject generated bytecode in distributions
+        run: |
+          uv run python - <<'PY'
+          from pathlib import Path, PurePosixPath
+          import tarfile
+          import zipfile
+
+          def forbidden(name: str) -> bool:
+              path = PurePosixPath(name)
+              return "__pycache__" in path.parts or path.suffix in {".pyc", ".pyo"}
+
+          for artifact in Path("dist").glob("dyro-*"):
+              if artifact.suffix == ".whl":
+                  with zipfile.ZipFile(artifact) as archive:
+                      names = archive.namelist()
+              elif artifact.name.endswith(".tar.gz"):
+                  with tarfile.open(artifact, "r:gz") as archive:
+                      names = archive.getnames()
+              else:
+                  continue
+              leaked = [name for name in names if forbidden(name)]
+              if leaked:
+                  raise SystemExit(
+                      f"generated bytecode found in {artifact.name}: {leaked[:5]}"
+                  )
+          PY
+
       - name: Verify installed wheel and sdist outside checkout
         run: |
           set -euo pipefail

.gitignore

@@ -18,3 +18,7 @@ dist/
 # Optional local Mermaid→PNG exports (docs use GitHub Mermaid; do not commit)
 docs/images/diagrams/*.png
 docs/images/diagrams/zh/*.png
+
+# Local handoffs and point-in-time agent review/evidence exports
+plans/*-handoff-*.md
+docs/superpowers/

CHANGELOG.md

@@ -2,6 +2,34 @@
 
 ## Unreleased
 
+- Expand the packaged `dyro-control-plane` Skill into a host-neutral,
+  intent-routed read-only control surface for workspace health, lines, Change
+  Sets, integrations, and Objective attention/graph/tick/plan observations.
+- Add stable JSON views for `workspace list`, `status`, `doctor`, `next`,
+  `line list`, `changeset list|verify`, `integration status`, and
+  `objective list|status` while preserving existing text output by default.
+- Make `objective list` and `objective status` strictly zero-write by refusing
+  to recover an interrupted Objective transaction during observation.
+- Run control-plane Git observations with `git --no-optional-locks` so status,
+  doctor, and Change Set verification cannot refresh Git index metadata.
+- Bind workspace-local `next --format json` handoffs to their resolved alias or
+  absolute root, and only offer bootstrap when every failure is an absent
+  repository with a configured remote.
+- Return one stable JSON error envelope for machine-facing runtime failures and
+  use deadline-, byte-, record-, and symlink-bounded reads for Profile, line,
+  Change Set, integration, Objective, Task, evidence, and Git observations.
+- Keep machine-facing Objective completion consistent with text views by
+  checking Task integration evidence and Git ancestry inside the same budget;
+  reject unsafe bootstrap targets before `next` can hand off a mutation.
+- Minimize Agent-visible local metadata: workspace and Skill integration JSON
+  and health diagnostics omit absolute paths by default, expose them only through explicit
+  `--include-paths`, and let the Skill skip global discovery when an alias is
+  already known.
+- Let Enter confirm the already-previewed feature worktree plan while keeping
+  `b` as the explicit route back to baseline selection.
+- Exclude generated Python bytecode from source and wheel distributions, even
+  when release tests imported packaged integration assets before the build.
+
 ## 0.6.4 - 2026-08-12
 
 - Guide control-plane Skill install during interactive `dyro setup` (preview in

MANIFEST.in

@@ -7,3 +7,4 @@ include docs/workspace-blueprints.md
 include examples/blueprints/acme-platform.toml
 recursive-include src/dyro/console/assets *
 recursive-include src/dyro/integrations/assets *
+global-exclude *.py[cod]

docs/adr/0006-agent-bridge-phase-0.md

@@ -15,12 +15,11 @@ both of which are brittle and easy to misinterpret.
 
 An earlier Agent Bridge proposal combined read operations, planning, generic
 confirmed apply, Skill delivery, MCP tools, and cross-host Plugin packaging in
-one v1. The adversarial review at
-[`2026-08-06-dyro-agent-bridge-design-adversarial-review-board.md`](../superpowers/reviews/2026-08-06-dyro-agent-bridge-design-adversarial-review-board.md)
-rejected that scope. The subsequent
-[Phase 0 design closure review](../superpowers/reviews/2026-08-06-dyro-agent-bridge-phase-0-design-closure-review.md)
-closed eight implementation-blueprint findings and approved only the S1
-contract/catalog step. Current source disproves two key assumptions:
+one v1. A point-in-time adversarial design review rejected that scope. Its
+follow-up closure resolved eight implementation-blueprint findings and
+approved only the S1 contract/catalog step. The durable decisions from those
+reviews are incorporated into this ADR and the acceptance matrix. Current
+source disproves two key assumptions:
 
 - `task gates` executes configured argv, writes gate logs, and appends the
   ledger; it is not a read operation.

docs/designs/agent-bridge-operation-inventory.md

@@ -4,8 +4,8 @@ Status: Linux Ubuntu 24.04 Mandatory Core Surface promoted at S5
 
 Decision source: [ADR 0006](../adr/0006-agent-bridge-phase-0.md)
 
-Review source:
-[2026-08-06 adversarial review board](../superpowers/reviews/2026-08-06-dyro-agent-bridge-design-adversarial-review-board.md)
+Review outcomes are incorporated into ADR 0006 and the Phase 0 acceptance
+matrix; point-in-time review exports are not part of the product documentation.
 
 ## 1. Purpose