Harden M9: Ed25519 attestation, full SHA-256 integrity, locked-by-default, prompt improvements, registry verification

README.md

@@ -123,6 +123,7 @@ Opens the graphical interface where you can create and manage tools visually. Fe
 - **My Tools** - Browse, create, edit, and delete tools organized by category
 - **Registry** - Search and install community tools from the CmdForge registry
 - **Providers** - Manage AI provider configurations
+- **Reuse & Discovery** - Compare similar tools and preview safe extraction of reusable steps
 
 ### CLI Mode
 
@@ -178,10 +179,26 @@ cmdforge settings mytool diff         # Show changes from defaults
 cmdforge config show                  # Show current config
 cmdforge config connect username      # Connect to registry account
 cmdforge config disconnect            # Disconnect from registry
+
+# Private, local pipeline suggestions (disabled by default)
+cmdforge usage enable                 # Begin recording tool names in shell pipes
+cmdforge usage suggestions            # Show frequent pipelines
+cmdforge usage clear                  # Delete all local usage history
+
+# Model Context Protocol (optional: pip install -e ".[mcp]")
+cmdforge mcp add local --command npx --arg=-y --arg @scope/server
+cmdforge mcp add remote --transport streamable-http --url https://example.com/mcp
+cmdforge mcp connect remote
+cmdforge mcp serve                    # Expose approved tools over stdio
+cmdforge mcp serve --transport streamable-http  # Loopback-only HTTP by default
 ```
 
 `cf` searches the public registry when no local tool matches. Registry results
 show available relevance and quality evidence and install on selection.
+When local pipeline discovery is explicitly enabled, `cf` also suggests
+frequently repeated pipe pairs. Only tool names, anonymous pipe identifiers,
+counts, and timestamps are kept in `~/.cmdforge/usage.json`; command arguments
+and input/output content are never recorded or transmitted.
 
 ### Running Tools
 
@@ -664,6 +681,9 @@ cmdforge registry search --owner official --min-downloads 100
 # Connect your account (opens browser for authentication)
 cmdforge config connect yourusername
 
+# Create an Ed25519 release key and register its public half
+cmdforge registry signing-key init
+
 # Publish a tool
 cmdforge registry publish mytool
 
@@ -678,6 +698,51 @@ cmdforge registry status --sync
 ```
 
 Published tools go through moderation before appearing publicly. You'll receive feedback if changes are requested.
+Once a signing key is registered, publishing signs the transitive content
+identity automatically. The registry rejects unsigned or invalid releases and
+clients verify the signature before installing. The private key remains in
+`~/.cmdforge/release-signing-key.json` with `0600` permissions.
+
+### Prompt Optimization
+
+Generate local deterministic prompt variations without contacting a provider:
+
+```bash
+cmdforge optimize summarize -n 5
+```
+
+Structural contract checks cannot rank prompt meaning, so CmdForge will not
+claim a winning variation unless explicit behavioral cases are supplied:
+
+```json
+[
+  {"input": "A long input", "contains": "summary"},
+  {"input": "Another input", "expected": {"result": "expected value"}}
+]
+```
+
+```bash
+cmdforge optimize summarize --behavior-tests behavior.json \
+  --test-provider opencode-pickle
+```
+
+Behavior tests execute every candidate and may incur provider cost. Use
+`--provider NAME` separately when you explicitly want a provider to generate a
+rephrased candidate.
+
+### Community Improvements
+
+```bash
+cmdforge registry improve owner/tool 1.0.0 0 proposed-prompt.txt \
+  --rationale "Clearer output requirement"
+cmdforge registry review-improvement 42 approve --notes "Tests pass"
+# Apply the approved text, bump the tool version, then publish it:
+cmdforge registry publish mytool --improvement-id 42
+```
+
+Submissions are auto-tested before review. Contributor credit and the
+`optimized`/`community-reviewed` badges are awarded only when a published
+version actually contains the approved change.
 
 ### Tool Documentation
 
@@ -750,6 +815,10 @@ The graphical interface provides a modern desktop experience:
 
 ### Tool Builder
 - Visual form for creating and editing tools
+- Reuse & Discovery panel with asynchronous registry similarity results,
+  exact local reuse evidence, and opt-in pipeline suggestions
+- Extraction creates a separate tool only after showing a unified diff; the
+  current draft is never rewritten automatically
 - Add arguments with flags and default values
 - Add prompt steps (AI calls) with profile selection
 - Add code steps with **AI-assisted code generation**:

pyproject.toml

@@ -36,6 +36,7 @@ dependencies = [
     "NodeGraphQt>=0.6.0",
     "setuptools",  # Required for distutils compatibility (Python 3.12+)
     "jsonschema>=4.0",  # JSON schema validation for structured output
+    "cryptography>=41.0",  # Ed25519 release attestations
 ]
 
 [project.optional-dependencies]

src/cmdforge/attestation.py

@@ -1,28 +1,52 @@
-"""Supply chain attestation (M9.5).
+"""Ed25519 supply-chain attestations for CmdForge releases."""
 
-Tool publishers sign releases. Registry verifies signatures before
+from __future__ import annotations
-accepting publish. Clients verify signatures before installing.
-"""
 
-import hashlib
+import base64
-import hmac
 import json
-from dataclasses import dataclass, field
+from dataclasses import dataclass
 from datetime import datetime, timezone
-from typing import Any, Dict, Optional
+from typing import Dict, Optional, Tuple
+
+from cryptography.exceptions import InvalidSignature
+from cryptography.hazmat.primitives import serialization
+from cryptography.hazmat.primitives.asymmetric.ed25519 import (
+    Ed25519PrivateKey,
+    Ed25519PublicKey,
+)
 
 
-@dataclass
+def _canonical_payload(
+    tool_name: str,
+    version: str,
+    content_hash: str,
+    signer: str,
+    signed_at: str,
+    algorithm: str,
+) -> bytes:
+    return json.dumps(
+        {
+            "algorithm": algorithm,
+            "content_hash": content_hash,
+            "signed_at": signed_at,
+            "signer": signer,
+            "tool_name": tool_name,
+            "version": version,
+        },
+        sort_keys=True,
+        separators=(",", ":"),
+    ).encode("utf-8")
+
+
+@dataclass(frozen=True)
 class Attestation:
-    """A signed attestation for a tool release."""
-
     tool_name: str
     version: str
     content_hash: str
-    signer: str  # publisher identity
+    signer: str
     signature: str
-    signed_at: str = ""
+    signed_at: str
-    algorithm: str = "hmac-sha256"
+    algorithm: str = "ed25519"
 
     def to_dict(self) -> dict:
         return {
@@ -35,76 +59,89 @@ class Attestation:
             "algorithm": self.algorithm,
         }
 
+    @classmethod
+    def from_dict(cls, value: dict) -> "Attestation":
+        return cls(**{key: value[key] for key in (
+            "tool_name", "version", "content_hash", "signer", "signature",
+            "signed_at", "algorithm",
+        )})
+
+
+def generate_keypair() -> Tuple[str, str]:
+    """Generate base64-encoded raw Ed25519 private/public keys."""
+    private = Ed25519PrivateKey.generate()
+    private_bytes = private.private_bytes(
+        serialization.Encoding.Raw,
+        serialization.PrivateFormat.Raw,
+        serialization.NoEncryption(),
+    )
+    public_bytes = private.public_key().public_bytes(
+        serialization.Encoding.Raw,
+        serialization.PublicFormat.Raw,
+    )
+    return (
+        base64.b64encode(private_bytes).decode("ascii"),
+        base64.b64encode(public_bytes).decode("ascii"),
+    )
+
 
 def sign_tool(
     tool_name: str,
     version: str,
     content_hash: str,
     signer: str,
-    secret_key: str,
+    private_key: str,
+    *,
+    signed_at: Optional[str] = None,
 ) -> Attestation:
-    """Sign a tool release with HMAC-SHA256.
+    """Sign release identity using a publisher's private Ed25519 key."""
-
+    timestamp = signed_at or datetime.now(timezone.utc).isoformat()
-    Args:
+    algorithm = "ed25519"
-        tool_name: Tool name
+    payload = _canonical_payload(
-        version: Tool version
+        tool_name, version, content_hash, signer, timestamp, algorithm
-        content_hash: Content hash from integrity module
+    )
-        signer: Publisher identity (username)
+    key = Ed25519PrivateKey.from_private_bytes(base64.b64decode(private_key))
-        secret_key: Signing key (from config or keyring)
+    signature = base64.b64encode(key.sign(payload)).decode("ascii")
-
-    Returns:
-        Attestation with signature
-    """
-    payload = f"{tool_name}:{version}:{content_hash}:{signer}"
-    signature = hmac.new(
-        secret_key.encode(),
-        payload.encode(),
-        hashlib.sha256,
-    ).hexdigest()
-
     return Attestation(
-        tool_name=tool_name,
+        tool_name, version, content_hash, signer, signature, timestamp, algorithm
-        version=version,
-        content_hash=content_hash,
-        signer=signer,
-        signature=signature,
-        signed_at=datetime.now(timezone.utc).isoformat(),
     )
 
 
-def verify_attestation(attestation: Attestation, secret_key: str) -> bool:
+def verify_attestation(attestation: Attestation, public_key: str) -> bool:
-    """Verify a tool attestation signature.
+    """Verify every signed field using the publisher's trusted public key."""
+    if attestation.algorithm != "ed25519":
+        return False
+    try:
+        key = Ed25519PublicKey.from_public_bytes(base64.b64decode(public_key))
+        signature = base64.b64decode(attestation.signature, validate=True)
+        payload = _canonical_payload(
+            attestation.tool_name,
+            attestation.version,
+            attestation.content_hash,
+            attestation.signer,
+            attestation.signed_at,
+            attestation.algorithm,
+        )
+        key.verify(signature, payload)
+        return True
+    except (ValueError, TypeError, InvalidSignature):
+        return False
 
-    Args:
-        attestation: The attestation to verify
-        secret_key: The signing key to verify against
 
-    Returns:
+def verify_trusted_attestation(
-        True if signature is valid
+    attestation: Attestation, trusted_publishers: Dict[str, str]
-    """
+) -> bool:
-    payload = f"{attestation.tool_name}:{attestation.version}:{attestation.content_hash}:{attestation.signer}"
+    """Anchor an attestation to an explicitly trusted publisher key."""
-    expected = hmac.new(
+    public_key = trusted_publishers.get(attestation.signer)
-        secret_key.encode(),
+    return bool(public_key) and verify_attestation(attestation, public_key)
-        payload.encode(),
-        hashlib.sha256,
-    ).hexdigest()
-    return hmac.compare_digest(expected, attestation.signature)
 
 
 def verify_content_hash(content_hash: str, tool_dict: dict) -> bool:
-    """Verify that a content hash matches the tool definition.
+    """Verify a definition hash against a serialized Tool definition."""
-
-    Args:
-        content_hash: The hash to verify
-        tool_dict: Tool dictionary (from to_dict())
-
-    Returns:
-        True if hash matches
-    """
     from .integrity import compute_tool_hash
-    # Reconstruct what the hash should be
+    from .tool import Tool
-    tool_dict_copy = dict(tool_dict)
+
-    tool_dict_copy.pop("path", None)
+    try:
-    content = json.dumps(tool_dict_copy, sort_keys=True)
+        return compute_tool_hash(Tool.from_dict(tool_dict)) == content_hash
-    expected = hashlib.sha256(content.encode()).hexdigest()[:16]
+    except (KeyError, TypeError, ValueError):
-    return expected == content_hash
+        return False

src/cmdforge/cli/__init__.py

@@ -17,6 +17,7 @@ from .config_commands import cmd_config
 from .settings_commands import cmd_settings
 from .system_deps_commands import cmd_system_deps
 from .mcp_commands import cmd_mcp
+from .usage_commands import cmd_usage
 
 
 def main():
@@ -107,6 +108,17 @@ def main():
     p_optimize = subparsers.add_parser("optimize", help="Generate and test prompt variations for a tool")
     p_optimize.add_argument("name", help="Tool name")
     p_optimize.add_argument("-n", "--count", type=int, default=5, help="Number of variations")
+    p_optimize.add_argument(
+        "--provider", default="mock",
+        help="Provider used to generate a variation (default: local deterministic mock)",
+    )
+    p_optimize.add_argument(
+        "--behavior-tests", metavar="FILE",
+        help="JSON cases used to score semantic behavior (executes the tool)",
+    )
+    p_optimize.add_argument(
+        "--test-provider", help="Provider override used while running behavior tests",
+    )
     p_optimize.set_defaults(func=cmd_optimize)
 
     # 'check' command
@@ -229,8 +241,40 @@ def main():
     p_reg_publish.add_argument("--dry-run", action="store_true", help="Validate without publishing")
     p_reg_publish.add_argument("-f", "--force", action="store_true", help="Skip confirmation prompts")
     p_reg_publish.add_argument("--owner", default="", help="Owner override (admin only, e.g. 'official')")
+    p_reg_publish.add_argument(
+        "--improvement-id", type=int,
+        help="Credit an approved improvement contained in this version",
+    )
     p_reg_publish.set_defaults(func=cmd_registry)
 
+    p_reg_signing = registry_sub.add_parser(
+        "signing-key", help="Initialize or inspect the release signing key"
+    )
+    p_reg_signing.add_argument(
+        "action", nargs="?", choices=["init", "status"], default="status"
+    )
+    p_reg_signing.set_defaults(func=cmd_registry)
+
+    p_reg_improve = registry_sub.add_parser(
+        "improve", help="Submit an auto-tested prompt/code improvement"
+    )
+    p_reg_improve.add_argument("tool", help="Tool as owner/name")
+    p_reg_improve.add_argument("version", help="Exact tool version")
+    p_reg_improve.add_argument("step_index", type=int, help="Step to replace")
+    p_reg_improve.add_argument("file", help="File containing proposed prompt/code")
+    p_reg_improve.add_argument("--rationale", default="")
+    p_reg_improve.set_defaults(func=cmd_registry)
+
+    p_reg_review_improvement = registry_sub.add_parser(
+        "review-improvement", help="Review a tested community improvement"
+    )
+    p_reg_review_improvement.add_argument("id", type=int)
+    p_reg_review_improvement.add_argument(
+        "decision", choices=["approve", "reject", "request_changes"]
+    )
+    p_reg_review_improvement.add_argument("--notes", default="")
+    p_reg_review_improvement.set_defaults(func=cmd_registry)
+
     # registry update-readme
     p_reg_update_readme = registry_sub.add_parser("update-readme", help="Update README for a published tool")
     p_reg_update_readme.add_argument("tool", nargs="?", default="", help="Tool name (local name, will resolve owner)")
@@ -477,8 +521,21 @@ def main():
     # mcp serve
     p_mcp_serve = mcp_sub.add_parser("serve", help="Start CmdForge as an MCP server")
     p_mcp_serve.add_argument(
-        "--transport", choices=["stdio"], default="stdio",
+        "--transport", choices=["stdio", "streamable-http"], default="stdio",
-        help="Transport (M7.3 supports stdio)",
+        help="MCP server transport",
+    )
+    p_mcp_serve.add_argument("--host", default="127.0.0.1")
+    p_mcp_serve.add_argument("--port", type=int, default=8000)
+    p_mcp_serve.add_argument(
+        "--allowed-origin", action="append", default=[],
+        help="Allowed browser Origin; repeat as needed",
+    )
+    p_mcp_serve.add_argument(
+        "--auth-token", help="Bearer token (or ${ENV_NAME})",
+    )
+    p_mcp_serve.add_argument(
+        "--external-url",
+        help="Public HTTPS URL when binding beyond localhost (TLS proxy required)",
     )
     p_mcp_serve.set_defaults(func=cmd_mcp)
 
@@ -495,10 +552,15 @@ def main():
     p_mcp_add = mcp_sub.add_parser("add", help="Add an MCP server")
     p_mcp_add.add_argument("name", help="Server name")
     p_mcp_add.add_argument(
-        "--transport", choices=["stdio"], default="stdio",
+        "--transport", choices=["stdio", "streamable-http"], default="stdio",
-        help="Transport type (M7.2 supports stdio)",
+        help="Transport type",
     )
     p_mcp_add.add_argument("--command", help="Executable for stdio servers")
+    p_mcp_add.add_argument("--url", help="Endpoint for Streamable HTTP servers")
+    p_mcp_add.add_argument(
+        "--header", action="append", default=[], metavar="NAME=VALUE",
+        help="HTTP header; repeat as needed (supports ${NAME} references)",
+    )
     p_mcp_add.add_argument(
         "--arg", dest="server_args", action="append", default=[], metavar="VALUE",
         help="Command argument; repeat for each argument (use --arg=-y for leading dashes)",
@@ -525,6 +587,20 @@ def main():
     # Default for mcp with no subcommand (list)
     p_mcp.set_defaults(func=cmd_mcp, mcp_cmd="list")
 
+    p_usage = subparsers.add_parser(
+        "usage", help="Manage opt-in local pipeline discovery"
+    )
+    usage_sub = p_usage.add_subparsers(dest="usage_cmd")
+    for action in ("enable", "disable", "status", "clear"):
+        command = usage_sub.add_parser(action)
+        command.set_defaults(func=cmd_usage)
+    p_usage_suggestions = usage_sub.add_parser(
+        "suggestions", help="Show frequently composed tool pipelines"
+    )
+    p_usage_suggestions.add_argument("--threshold", type=int, default=3)
+    p_usage_suggestions.set_defaults(func=cmd_usage)
+    p_usage.set_defaults(func=cmd_usage, usage_cmd="status")
+
     args = parser.parse_args()
 
     # If no command, launch UI
@@ -682,7 +758,7 @@ def cmd_inspect(args):
 
 def cmd_optimize(args):
     """Generate and test prompt variations for a tool."""
-    from ..prompt_optimizer import optimize_tool
+    from ..prompt_optimizer import load_behavioral_evaluator, optimize_tool
     from ..tool import load_tool
 
     tool = load_tool(args.name)
@@ -693,7 +769,21 @@ def cmd_optimize(args):
     print(f"Optimizing '{tool.name}' — generating {args.count} prompt variations...")
     print()
 
-    result = optimize_tool(tool, count=args.count)
+    if args.count < 1:
+        print("Error: --count must be at least 1.", file=sys.stderr)
+        return 2
+    evaluator = None
+    if args.behavior_tests:
+        try:
+            evaluator = load_behavioral_evaluator(
+                args.behavior_tests, args.test_provider
+            )
+        except (OSError, ValueError) as exc:
+            print(f"Error loading behavior tests: {exc}", file=sys.stderr)
+            return 2
+    result = optimize_tool(
+        tool, count=args.count, provider=args.provider, evaluator=evaluator
+    )
 
     print(f"Baseline score: {result.baseline_score}")
     print(f"Variations tested: {len(result.variations)}")
@@ -708,7 +798,7 @@ def cmd_optimize(args):
         print()
         print("To apply this variation, edit the tool's config.yaml and update the prompt step.")
     else:
-        print("No improvements found. The original prompt may already be optimal.")
+        print(result.note or "No improvements found.")
 
     return 0
 

src/cmdforge/cli/mcp_commands.py

@@ -51,6 +51,8 @@ def _cmd_mcp_list(args):
             print(f"      Command: {_sanitize(command)}")
             if s.cwd:
                 print(f"      Working dir: {s.cwd}")
+        else:
+            print(f"      URL: {_sanitize(s.url or '')}")
         print(f"      Approved: {'yes' if s.approved else 'no'}")
         print(f"      Timeout: {s.timeout}s")
         print()
@@ -98,6 +100,7 @@ def _cmd_mcp_add(args):
 
     try:
         server_env = _parse_env_entries(getattr(args, "env", []))
+        headers = _parse_header_entries(getattr(args, "header", []))
     except ValueError as exc:
         print(f"Error: {exc}")
         return 1
@@ -111,6 +114,8 @@ def _cmd_mcp_add(args):
         name=name,
         transport=getattr(args, "transport", "stdio"),
         command=getattr(args, "command", None),
+        url=getattr(args, "url", None),
+        headers=headers,
         args=getattr(args, "server_args", []) or [],
         cwd=getattr(args, "cwd", None),
         env=server_env,
@@ -171,13 +176,34 @@ def _parse_env_entries(entries):
     return result
 
 
+def _parse_header_entries(entries):
+    result = {}
+    for entry in entries:
+        if "=" not in entry:
+            raise ValueError(f"invalid --header value '{entry}'; expected NAME=VALUE")
+        name, value = entry.split("=", 1)
+        if not name.strip() or any(char in name for char in "\r\n"):
+            raise ValueError(f"invalid HTTP header name '{name}'")
+        if any(char in value for char in "\r\n"):
+            raise ValueError(f"invalid HTTP header value for '{name}'")
+        result[name.strip()] = value
+    return result
+
+
 def _cmd_mcp_serve(args):
     """Start CmdForge as an MCP server."""
     from ..mcp_server import serve as start_server
 
     transport = getattr(args, "transport", "stdio")
     try:
-        start_server(transport=transport)
+        start_server(
+            transport=transport,
+            host=getattr(args, "host", "127.0.0.1"),
+            port=getattr(args, "port", 8000),
+            allowed_origins=getattr(args, "allowed_origin", []),
+            auth_token=getattr(args, "auth_token", None),
+            external_url=getattr(args, "external_url", None),
+        )
     except ImportError as exc:
         print(f"Error: {exc}")
         return 1

src/cmdforge/cli/picker.py

@@ -479,6 +479,19 @@ def main():
     # Use stderr for UI if stdout is piped, so tool output stays clean
     _ui_out = sys.stderr if not sys.stdout.isatty() else sys.stdout
 
+    try:
+        from ..usage import get_suggestions
+        suggestions = get_suggestions()[:3]
+        if suggestions:
+            _write(f"{GREEN}Frequent pipelines you could save as tools:{RESET}\n")
+            for item in suggestions:
+                _write(
+                    f"  {' | '.join(item['tools'])} "
+                    f"{DIM}({item['count']} uses){RESET}\n"
+                )
+    except (OSError, ValueError):
+        pass
+
     try:
         with TTYInput() as tty_input:
             result = run_picker(tty_input)

src/cmdforge/cli/registry_commands.py

@@ -29,6 +29,12 @@ def cmd_registry(args):
         return _cmd_registry_update(args)
     elif args.registry_cmd == "publish":
         return _cmd_registry_publish(args)
+    elif args.registry_cmd == "signing-key":
+        return _cmd_registry_signing_key(args)
+    elif args.registry_cmd == "improve":
+        return _cmd_registry_improve(args)
+    elif args.registry_cmd == "review-improvement":
+        return _cmd_registry_review_improvement(args)
     elif args.registry_cmd == "update-readme":
         return _cmd_registry_update_readme(args)
     elif args.registry_cmd == "my-tools":
@@ -51,6 +57,8 @@ def cmd_registry(args):
         print("  info <tool>        Show tool information")
         print("  update             Update local index cache")
         print("  publish [path]     Publish a tool")
+        print("  signing-key        Manage Ed25519 release signing")
+        print("  improve            Submit a community improvement")
         print("  update-readme      Update README for published tool(s)")
         print("  my-tools           List your published tools")
         print("  status <tool>      Check moderation status of a tool")
@@ -598,6 +606,7 @@ def _cmd_registry_publish(args):
     from ..tool import load_tool, Tool, ToolStep
 
     dep_result = None
+    my_owner = ""
     try:
         client = get_client()
 
@@ -737,6 +746,7 @@ def _cmd_registry_publish(args):
     _print_quality_summary(preflight_result.get("quality"))
     if remote_report.get("errors"):
         return 1
+    release_content_hash = preflight_result.get("content_hash", "")
     if sys.stdin.isatty() and not getattr(args, "force", False):
         try:
             if input("Publish this validated version? [y/N] ").strip().lower() != "y":
@@ -751,7 +761,26 @@ def _cmd_registry_publish(args):
     try:
         client = get_client()
         owner = getattr(args, "owner", "")
-        result = client.publish_tool(config_yaml, readme, defaults, owner=owner)
+        attestation = None
+        from ..signing import load_signing_key
+        signing_key = load_signing_key()
+        if signing_key:
+            if not release_content_hash:
+                print(
+                    "Registry did not return a content identity; refusing to sign.",
+                    file=sys.stderr,
+                )
+                return 1
+            from ..attestation import sign_tool
+            signer = owner or my_owner
+            attestation = sign_tool(
+                name, version, release_content_hash, signer, signing_key[0]
+            ).to_dict()
+        result = client.publish_tool(
+            config_yaml, readme, defaults, owner=owner,
+            attestation=attestation,
+            improvement_id=getattr(args, "improvement_id", None),
+        )
 
         pr_url = result.get("pr_url", "")
         status = result.get("status", "")
@@ -826,6 +855,64 @@ def _cmd_registry_publish(args):
     return 0
 
 
+def _cmd_registry_signing_key(args):
+    """Initialize a local key and register its public half with the registry."""
+    from ..signing import SIGNING_KEY_FILE, initialize_signing_key, load_signing_key
+    from ..registry_client import RegistryError, get_client
+
+    if args.action == "status":
+        key = load_signing_key()
+        print(
+            f"Signing key: {SIGNING_KEY_FILE}"
+            if key else "No release signing key configured."
+        )
+        return 0 if key else 1
+    try:
+        _, public_key = initialize_signing_key()
+        get_client().set_signing_public_key(public_key)
+    except (OSError, ValueError, RegistryError) as exc:
+        print(f"Could not initialize signing key: {exc}", file=sys.stderr)
+        return 1
+    print(f"Release signing key initialized: {SIGNING_KEY_FILE}")
+    print("The private key is stored locally with mode 0600; back it up securely.")
+    return 0
+
+
+def _cmd_registry_improve(args):
+    from ..registry_client import RegistryError, get_client
+    if "/" not in args.tool:
+        print("Tool must be specified as owner/name.", file=sys.stderr)
+        return 2
+    owner, name = args.tool.split("/", 1)
+    try:
+        proposed = Path(args.file).read_text(encoding="utf-8")
+        result = get_client().submit_improvement(
+            owner, name, args.version, args.step_index, proposed,
+            args.rationale,
+        )
+    except (OSError, RegistryError) as exc:
+        print(f"Could not submit improvement: {exc}", file=sys.stderr)
+        return 1
+    print(f"Improvement {result.get('id')} tested and submitted for review.")
+    return 0
+
+
+def _cmd_registry_review_improvement(args):
+    from ..registry_client import RegistryError, get_client
+    try:
+        result = get_client().review_improvement(args.id, args.decision, args.notes)
+    except RegistryError as exc:
+        print(f"Could not review improvement: {exc}", file=sys.stderr)
+        return 1
+    print(f"Improvement {args.id}: {result.get('status')}")
+    if result.get("ready_to_apply"):
+        print(
+            "Publish the updated version with "
+            f"--improvement-id {args.id} to apply credit and badges."
+        )
+    return 0
+
+
 def _cmd_registry_update_readme(args):
     """Update README for published tool(s) on the registry."""
     from ..registry_client import RegistryError, get_client

src/cmdforge/cli/usage_commands.py

@@ -0,0 +1,29 @@
+"""CLI for opt-in local usage discovery."""
+
+from ..usage import clear_usage, get_suggestions, is_enabled, set_enabled
+
+
+def cmd_usage(args):
+    action = getattr(args, "usage_cmd", None) or "status"
+    if action == "enable":
+        set_enabled(True)
+        print("Local pipeline discovery enabled. No input, output, arguments, or telemetry is recorded.")
+        return 0
+    if action == "disable":
+        set_enabled(False)
+        print("Local pipeline discovery disabled.")
+        return 0
+    if action == "clear":
+        clear_usage()
+        print("Local usage history cleared.")
+        return 0
+    if action == "suggestions":
+        suggestions = get_suggestions(getattr(args, "threshold", 3))
+        if not suggestions:
+            print("No frequent pipeline suggestions yet.")
+            return 0
+        for index, item in enumerate(suggestions, start=1):
+            print(f"{index}. {' | '.join(item['tools'])} ({item['count']} uses)")
+        return 0
+    print(f"Local pipeline discovery: {'enabled' if is_enabled() else 'disabled'}")
+    return 0

src/cmdforge/community.py

@@ -98,6 +98,9 @@ def run_submission_tests(submission: ImprovementSubmission, tool: Tool) -> dict:
 
     from .contract_testing import run_contract_tests
 
+    if submission.tool_name != tool.name or submission.tool_version != (tool.version or ""):
+        raise ValueError("Submission targets a different tool or version")
+
     # Baseline
     baseline = run_contract_tests(tool)
     baseline_passed = sum(
@@ -110,9 +113,27 @@ def run_submission_tests(submission: ImprovementSubmission, tool: Tool) -> dict:
         return {"error": "Step index out of range", "baseline_passed": baseline_passed}
 
     step = tool_copy.steps[submission.step_index]
+    current = step.prompt if isinstance(step, PromptStep) else (
+        step.code if isinstance(step, CodeStep) else None
+    )
+    if current != submission.original:
+        return {
+            "error": "Original step has changed since submission",
+            "baseline_passed": baseline_passed,
+        }
     if isinstance(step, PromptStep) and submission.step_type == "prompt":
         step.prompt = submission.proposed
     elif isinstance(step, CodeStep) and submission.step_type == "code":
+        try:
+            compile(submission.proposed, f"<{tool.name}:step-{submission.step_index}>", "exec")
+        except SyntaxError as exc:
+            result = {
+                "error": f"Proposed code has invalid syntax: {exc.msg}",
+                "baseline_passed": baseline_passed,
+                "passed_for_review": False,
+            }
+            submission.test_result = result
+            return result
         step.code = submission.proposed
     else:
         return {"error": "Step type mismatch", "baseline_passed": baseline_passed}
@@ -128,6 +149,18 @@ def run_submission_tests(submission: ImprovementSubmission, tool: Tool) -> dict:
         "proposed_passed": proposed_passed,
         "improvement": proposed_passed - baseline_passed,
         "regressed": proposed_passed < baseline_passed,
+        "baseline_outcome": baseline.outcome,
+        "proposed_outcome": proposed_result.outcome,
+        "passed_for_review": (
+            (
+                proposed_result.outcome == "passed"
+                and proposed_passed >= baseline_passed
+            )
+            or (
+                submission.step_type == "code"
+                and proposed_result.outcome == "unsupported"
+            )
+        ),
         "details": [r.to_dict() for r in proposed_result.results],
     }
 
@@ -146,10 +179,12 @@ def review_submission(
     if decision not in ("approve", "reject", "request_changes"):
         raise ValueError("Decision must be: approve, reject, or request_changes")
 
-    if submission.status == "pending":
+    if submission.status != "tested" or not (
+        submission.test_result or {}
+    ).get("passed_for_review"):
         raise ValueError("Submission must be tested before review")
 
-    submission.status = decision
+    submission.status = "approved" if decision == "approve" else decision
     return SubmissionReview(
         submission=submission,
         decision=decision,

src/cmdforge/gui/pages/tool_builder_page.py

@@ -1,5 +1,9 @@
 """Tool builder page - create and edit tools."""
 
+import copy
+import difflib
+import re
+
 import yaml
 
 from PySide6.QtWidgets import (
@@ -7,9 +11,9 @@ from PySide6.QtWidgets import (
     QLineEdit, QTextEdit, QPlainTextEdit, QComboBox, QPushButton,
     QGroupBox, QListWidget, QListWidgetItem, QLabel,
     QMessageBox, QSplitter, QFrame, QStackedWidget,
-    QButtonGroup
+    QButtonGroup, QInputDialog
 )
-from PySide6.QtCore import Qt
+from PySide6.QtCore import Qt, QThread, Signal, QTimer
 
 from ...tool import (
     Tool, ToolArgument, PromptStep, CodeStep, ToolStep,
@@ -19,6 +23,84 @@ from ...tool import (
 from ..widgets.icons import get_prompt_icon, get_code_icon, get_tool_icon
 
 
+class SimilarToolsWorker(QThread):
+    """Search the registry without blocking the tool builder."""
+
+    results_ready = Signal(str, list)
+
+    def __init__(self, query: str, parent=None):
+        super().__init__(parent)
+        self.query = query
+
+    def run(self):
+        try:
+            from ...registry_client import RegistryClient, RegistryError
+
+            try:
+                result = RegistryClient().search_tools(self.query, per_page=5)
+            except RegistryError:
+                result = None
+            self.results_ready.emit(self.query, list(result.data) if result else [])
+        except Exception:
+            # Guidance is advisory; an unavailable registry must never prevent
+            # local authoring.
+            self.results_ready.emit(self.query, [])
+
+
+def build_extracted_tool(tool: Tool, opportunity: dict, name: str) -> Tool:
+    """Build, but do not save, a reusable tool from an evidenced sequence."""
+    if opportunity.get("type") != "repeated_sequence":
+        raise ValueError("Only an internal repeated sequence can be extracted")
+    locations = opportunity.get("locations") or []
+    length = int(opportunity.get("length", 0))
+    if not locations or length < 2:
+        raise ValueError("Reuse evidence does not identify a valid sequence")
+    start = int(locations[0]) - 1
+    selected = tool.steps[start:start + length]
+    if len(selected) != length:
+        raise ValueError("Reuse evidence falls outside the current draft")
+    available = {"input", "settings"} | {
+        argument.variable for argument in tool.arguments
+    }
+    for step in selected:
+        if not isinstance(step, PromptStep):
+            raise ValueError("Only prompt-step sequences can be extracted safely")
+        references = set(re.findall(r"(?<!\{)\{([A-Za-z_]\w*)\}(?!\})", step.prompt or ""))
+        missing = references - available
+        if missing:
+            raise ValueError(
+                "The sequence depends on values produced outside its boundary: "
+                + ", ".join(sorted(missing))
+            )
+        if step.output_var:
+            available.update(
+                value.strip() for value in step.output_var.split(",") if value.strip()
+            )
+    output_var = getattr(selected[-1], "output_var", "result") or "result"
+    output_schema = getattr(selected[-1], "output_schema", None)
+    return Tool(
+        name=name,
+        description=f"Extracted reusable sequence from {tool.name}",
+        category=tool.category,
+        arguments=copy.deepcopy(tool.arguments),
+        steps=copy.deepcopy(selected),
+        output=f"{{{output_var}}}",
+        dependencies=[],
+        system_dependencies=copy.deepcopy(tool.system_dependencies),
+        visibility=tool.visibility,
+        input_schema=copy.deepcopy(tool.input_schema),
+        output_schema=copy.deepcopy(output_schema),
+    )
+
+
+def extraction_diff(tool: Tool) -> str:
+    """Return a unified-diff preview for a proposed new tool."""
+    rendered = yaml.safe_dump(tool.to_dict(), sort_keys=False).splitlines(True)
+    return "".join(difflib.unified_diff(
+        [], rendered, fromfile="/dev/null", tofile=f"{tool.name}/config.yaml",
+    ))
+
+
 def _switch_to_existing_tool(main_window, name: str) -> None:
     """Remove the creation page before opening an existing tool editor."""
     main_window.close_tool_builder()
@@ -35,6 +117,7 @@ class ToolBuilderPage(QWidget):
         self.original_name = tool_name
         self._tool = None
         self._flow_widget = None  # Lazy-loaded
+        self._guidance_worker = None
 
         self._setup_ui()
 
@@ -104,6 +187,44 @@ class ToolBuilderPage(QWidget):
 
         left_layout.addWidget(info_box)
 
+        # Reuse guidance is advisory and never changes a draft automatically.
+        guidance_box = QGroupBox("Reuse & Discovery")
+        guidance_layout = QVBoxLayout(guidance_box)
+        self.guidance_summary = QLabel(
+            "CmdForge can check for similar tools, reusable step sequences, "
+            "and frequent local pipelines."
+        )
+        self.guidance_summary.setWordWrap(True)
+        guidance_layout.addWidget(self.guidance_summary)
+        self.guidance_list = QListWidget()
+        self.guidance_list.setMaximumHeight(130)
+        self.guidance_list.currentItemChanged.connect(
+            self._guidance_selection_changed
+        )
+        guidance_layout.addWidget(self.guidance_list)
+        guidance_buttons = QHBoxLayout()
+        self.btn_refresh_guidance = QPushButton("Refresh")
+        self.btn_refresh_guidance.clicked.connect(self._refresh_guidance)
+        guidance_buttons.addWidget(self.btn_refresh_guidance)
+        self.btn_open_guidance = QPushButton("Open Existing")
+        self.btn_open_guidance.setObjectName("secondary")
+        self.btn_open_guidance.clicked.connect(self._open_guidance_tool)
+        self.btn_open_guidance.setEnabled(False)
+        guidance_buttons.addWidget(self.btn_open_guidance)
+        self.btn_extract_guidance = QPushButton("Extract as New Tool")
+        self.btn_extract_guidance.clicked.connect(self._extract_guidance)
+        self.btn_extract_guidance.setEnabled(False)
+        guidance_buttons.addWidget(self.btn_extract_guidance)
+        guidance_layout.addLayout(guidance_buttons)
+        left_layout.addWidget(guidance_box)
+
+        self._guidance_timer = QTimer(self)
+        self._guidance_timer.setSingleShot(True)
+        self._guidance_timer.setInterval(700)
+        self._guidance_timer.timeout.connect(self._refresh_guidance)
+        self.name_input.textChanged.connect(self._schedule_guidance)
+        self.desc_input.textChanged.connect(self._schedule_guidance)
+
         # Arguments group
         args_box = QGroupBox()
         args_layout = QVBoxLayout(args_box)
@@ -404,6 +525,175 @@ class ToolBuilderPage(QWidget):
 
         layout.addWidget(splitter, 1)
 
+    def _current_draft(self) -> Tool:
+        """Materialize the current form without saving it."""
+        current = self._tool
+        return Tool(
+            name=self.name_input.text().strip() or "untitled-tool",
+            description=self.desc_input.text().strip(),
+            category=self.category_combo.currentText() or "Other",
+            arguments=copy.deepcopy(current.arguments) if current else [],
+            steps=copy.deepcopy(current.steps) if current else [],
+            output=self.output_input.toPlainText().strip() or "{response}",
+            dependencies=list(current.dependencies) if current else [],
+            system_dependencies=(
+                copy.deepcopy(current.system_dependencies) if current else []
+            ),
+            visibility=current.visibility if current else "public",
+            input_schema=copy.deepcopy(current.input_schema) if current else None,
+            output_schema=copy.deepcopy(current.output_schema) if current else None,
+        )
+
+    def _schedule_guidance(self):
+        self._guidance_timer.start()
+
+    def _add_guidance_item(self, text: str, data: dict) -> None:
+        item = QListWidgetItem(text)
+        item.setData(Qt.UserRole, data)
+        self.guidance_list.addItem(item)
+
+    def _refresh_guidance(self):
+        """Refresh deterministic local guidance, then search remotely."""
+        from ...preflight import analyze_tool
+        from ...usage import get_suggestions
+
+        draft = self._current_draft()
+        self.guidance_list.clear()
+        report = analyze_tool(
+            draft,
+            check_local_dependencies=False,
+            include_contract_tests=False,
+            check_reuse=True,
+        )
+        for opportunity in report.reuse_opportunities:
+            self._add_guidance_item(
+                f"Reusable: {opportunity.get('detail', 'exact sequence')}",
+                opportunity,
+            )
+        for suggestion in get_suggestions():
+            tools = suggestion["tools"]
+            self._add_guidance_item(
+                f"Frequent pipeline ({suggestion['count']}x): " + " | ".join(tools),
+                {"type": "usage_pipeline", **suggestion},
+            )
+
+        query = (draft.description or draft.name).strip()
+        if not query or query == "untitled-tool":
+            self._finish_guidance(query, [])
+            return
+        if self._guidance_worker and self._guidance_worker.isRunning():
+            # The pending result is ignored if the query has since changed.
+            self.guidance_summary.setText("Checking registry…")
+            return
+        self.guidance_summary.setText("Checking registry…")
+        self.btn_refresh_guidance.setEnabled(False)
+        self._guidance_worker = SimilarToolsWorker(query, self)
+        self._guidance_worker.results_ready.connect(self._finish_guidance)
+        self._guidance_worker.start()
+
+    def _finish_guidance(self, query: str, results: list):
+        current_query = (
+            self.desc_input.text().strip() or self.name_input.text().strip()
+        )
+        if query == current_query:
+            seen = set()
+            for result in results:
+                name = result.get("name", "")
+                owner = result.get("owner", "")
+                reference = f"{owner}/{name}" if owner else name
+                if reference and reference not in seen:
+                    seen.add(reference)
+                    score = result.get("quality_score")
+                    quality = f" [{score}/100]" if score is not None else ""
+                    self._add_guidance_item(
+                        f"Registry: {reference}{quality} — {result.get('description', '')}",
+                        {"type": "registry_similar", "tool": reference},
+                    )
+        self.btn_refresh_guidance.setEnabled(True)
+        count = self.guidance_list.count()
+        self.guidance_summary.setText(
+            f"{count} advisory suggestion{'s' if count != 1 else ''}. "
+            "Nothing changes until you confirm an action."
+            if count else "No evidenced reuse opportunities found."
+        )
+        self._guidance_selection_changed(self.guidance_list.currentItem())
+        if query != current_query:
+            self._guidance_timer.start(0)
+
+    def _guidance_selection_changed(self, current, previous=None):
+        data = current.data(Qt.UserRole) if current else {}
+        kind = data.get("type")
+        local_name = data.get("tool", "")
+        self.btn_open_guidance.setEnabled(
+            kind == "duplicate_sequence" and bool(local_name)
+        )
+        self.btn_extract_guidance.setEnabled(
+            kind in {"repeated_sequence", "usage_pipeline"}
+        )
+
+    def _open_guidance_tool(self):
+        item = self.guidance_list.currentItem()
+        data = item.data(Qt.UserRole) if item else {}
+        if data.get("type") == "duplicate_sequence" and data.get("tool"):
+            _switch_to_existing_tool(self.main_window, data["tool"])
+
+    def _extract_guidance(self):
+        """Preview and explicitly save an evidenced reusable tool."""
+        item = self.guidance_list.currentItem()
+        data = item.data(Qt.UserRole) if item else {}
+        kind = data.get("type")
+        if kind not in {"repeated_sequence", "usage_pipeline"}:
+            return
+        suggested = (
+            "-".join(data.get("tools", []))
+            if kind == "usage_pipeline" else f"{self.name_input.text().strip()}-shared"
+        )
+        name, accepted = QInputDialog.getText(
+            self, "Extract Reusable Tool", "New tool name:", text=suggested
+        )
+        name = name.strip()
+        if not accepted or not name:
+            return
+        valid, error = validate_tool_name(name)
+        if not valid:
+            QMessageBox.warning(self, "Validation", error)
+            return
+        if tool_exists(name):
+            QMessageBox.warning(self, "Tool Already Exists", f"'{name}' already exists.")
+            return
+        try:
+            if kind == "usage_pipeline":
+                from ...usage import build_composite_tool
+                extracted = build_composite_tool(name, data["tools"])
+            else:
+                extracted = build_extracted_tool(self._current_draft(), data, name)
+        except (KeyError, TypeError, ValueError) as exc:
+            QMessageBox.warning(self, "Cannot Extract", str(exc))
+            return
+
+        preview = QMessageBox(self)
+        preview.setIcon(QMessageBox.Information)
+        preview.setWindowTitle("Review Extraction")
+        preview.setText(
+            f"Create '{name}' as a separate tool? Your current draft will not be modified."
+        )
+        preview.setDetailedText(extraction_diff(extracted))
+        create_button = preview.addButton("Create Tool", QMessageBox.AcceptRole)
+        preview.addButton(QMessageBox.Cancel)
+        preview.exec()
+        if preview.clickedButton() is not create_button:
+            return
+        try:
+            save_tool(extracted)
+        except (OSError, ValueError) as exc:
+            QMessageBox.critical(self, "Extraction Failed", str(exc))
+            return
+        QMessageBox.information(
+            self, "Tool Created", f"Created '{name}'. The current draft is unchanged."
+        )
+        self._populate_deps_combo()
+        self._refresh_guidance()
+
     def _set_view_mode(self, mode: int):
         """Switch between list (0) and flow (1) views."""
         if mode == 1 and self._flow_widget is None:

src/cmdforge/improvement.py

@@ -63,7 +63,9 @@ class ImprovementReport:
         }
 
 
-def generate_improvements(tool: Tool) -> ImprovementReport:
+def generate_improvements(
+    tool: Tool, scrutiny_report: Optional[dict] = None
+) -> ImprovementReport:
     """Analyze a tool and generate improvement suggestions.
 
     Categories:
@@ -77,6 +79,35 @@ def generate_improvements(tool: Tool) -> ImprovementReport:
     _check_efficiency(tool, report)
     _check_transparency(tool, report)
 
+    # Preserve the registry scrutiny engine's evidence rather than running a
+    # disconnected approximation. Only actionable warnings/failures become
+    # suggestions; passing checks are deliberately omitted.
+    for finding in (scrutiny_report or {}).get("findings", []):
+        result = finding.get("result")
+        if result not in ("warning", "fail"):
+            continue
+        check = str(finding.get("check", "scrutiny")).lower()
+        category = next(
+            (value for value in ("honesty", "efficiency", "transparency")
+             if value in check),
+            "transparency",
+        )
+        suggestion = finding.get("suggestion") or (
+            "Review this finding and make the behavior explicit."
+        )
+        candidate = ImprovementSuggestion(
+            category=category,
+            severity="high" if result == "fail" else "medium",
+            title=f"Scrutiny: {finding.get('check', 'review required')}",
+            description=finding.get("message", "Scrutiny identified an issue."),
+            location=finding.get("location") or "tool",
+            suggested=suggestion,
+        )
+        marker = (candidate.category, candidate.title, candidate.location)
+        existing = {(s.category, s.title, s.location) for s in report.suggestions}
+        if marker not in existing:
+            report.suggestions.append(candidate)
+
     return report
 
 

src/cmdforge/integrity.py

@@ -1,24 +1,27 @@
-"""Transitive integrity verification (M9.4).
+"""Content-addressable, transitive integrity verification for CmdForge tools."""
 
-Content-addressable tool identity: tool = hash(tool definition + all dep hashes).
+from __future__ import annotations
-Extends lockfile to include transitive integrity chain.
-"""
 
 import hashlib
 import json
 from dataclasses import dataclass, field
-from pathlib import Path
+from typing import Dict, List, Optional, Set
-from typing import Any, Dict, List, Optional
 
-from .tool import Tool, load_tool
+from .tool import Tool, ToolStep, load_tool
+
+
+def _sha256(value: object) -> str:
+    encoded = json.dumps(
+        value, sort_keys=True, separators=(",", ":"), ensure_ascii=False
+    ).encode("utf-8")
+    return f"sha256:{hashlib.sha256(encoded).hexdigest()}"
 
 
 @dataclass
 class IntegrityNode:
-    """One node in the integrity chain."""
-
     name: str
     version: str = ""
+    definition_hash: str = ""
     content_hash: str = ""
     dependencies: List[str] = field(default_factory=list)
     dependency_hashes: Dict[str, str] = field(default_factory=dict)
@@ -27,6 +30,7 @@ class IntegrityNode:
         return {
             "name": self.name,
             "version": self.version,
+            "definition_hash": self.definition_hash,
             "hash": self.content_hash,
             "dependencies": self.dependencies,
             "dependency_hashes": self.dependency_hashes,
@@ -35,99 +39,142 @@ class IntegrityNode:
 
 @dataclass
 class IntegrityChain:
-    """Full transitive integrity chain for a tool."""
-
     root: IntegrityNode
     nodes: Dict[str, IntegrityNode] = field(default_factory=dict)
+    errors: List[str] = field(default_factory=list)
 
     @property
     def is_valid(self) -> bool:
-        """Check all dependency hashes match."""
+        if self.errors or self.root.name not in self.nodes:
+            return False
         for name, node in self.nodes.items():
+            if not node.definition_hash or not node.content_hash:
+                return False
+            if sorted(node.dependencies) != sorted(node.dependency_hashes):
+                return False
             for dep_name in node.dependencies:
-                expected_hash = node.dependency_hashes.get(dep_name)
+                dep = self.nodes.get(dep_name)
-                if expected_hash == "unresolved":
+                if dep is None or node.dependency_hashes[dep_name] != dep.content_hash:
-                    # Unresolved dependencies don't invalidate the chain
-                    continue
-                if dep_name not in self.nodes:
-                    return False
-                dep_node = self.nodes[dep_name]
-                if expected_hash and expected_hash != dep_node.content_hash:
                     return False
+            expected = _identity_hash(node.definition_hash, node.dependency_hashes)
+            if not _constant_time_equal(expected, node.content_hash):
+                return False
         return True
 
     def to_dict(self) -> dict:
         return {
             "root": self.root.to_dict(),
-            "nodes": {k: v.to_dict() for k, v in self.nodes.items()},
+            "nodes": {key: value.to_dict() for key, value in self.nodes.items()},
+            "errors": list(self.errors),
             "valid": self.is_valid,
         }
 
 
+def _constant_time_equal(left: str, right: str) -> bool:
+    import hmac
+    return hmac.compare_digest(left, right)
+
+
 def compute_tool_hash(tool: Tool) -> str:
-    """Compute a content hash for a tool definition."""
+    """Return the full SHA-256 hash of the canonical tool definition."""
-    tool_dict = tool.to_dict()
+    data = tool.to_dict()
-    # Remove path (not part of content identity)
+    data.pop("path", None)
-    tool_dict.pop("path", None)
+    # Registry bookkeeping is not executable tool content.
-    content = json.dumps(tool_dict, sort_keys=True)
+    for key in ("registry_hash", "registry_status", "registry_feedback"):
-    return hashlib.sha256(content.encode()).hexdigest()[:16]
+        data.pop(key, None)
+    return _sha256(data)
 
 
-def build_integrity_chain(tool: Tool, max_depth: int = 10) -> IntegrityChain:
+def _identity_hash(definition_hash: str, dependency_hashes: Dict[str, str]) -> str:
-    """Build a transitive integrity chain for a tool.
+    """Hash a definition together with the identities of every direct dep."""
+    return _sha256({
+        "definition": definition_hash,
+        "dependencies": dict(sorted(dependency_hashes.items())),
+    })
 
-    Traverses all ToolStep dependencies and computes hashes for each.
+
-    """
+def compute_content_identity(
-    chain = IntegrityChain(root=_build_node(tool))
+    definition_hash: str, dependency_hashes: Dict[str, str]
-    chain.nodes[tool.name] = chain.root
+) -> str:
-    _traverse_deps(tool, chain, depth=0, max_depth=max_depth)
+    """Return the canonical identity used by integrity chains and lockfiles."""
+    return _identity_hash(definition_hash, dependency_hashes)
+
+
+def _dependency_names(tool: Tool) -> List[str]:
+    names = list(tool.dependencies)
+    names.extend(step.tool for step in tool.steps if isinstance(step, ToolStep))
+    return sorted(set(names))
+
+
+def build_integrity_chain(tool: Tool, max_depth: int = 50) -> IntegrityChain:
+    """Resolve dependencies bottom-up and compute a Merkle-style identity."""
+    placeholder = IntegrityNode(tool.name, tool.version or "")
+    chain = IntegrityChain(root=placeholder)
+    visiting: Set[str] = set()
+
+    def visit(current: Tool, depth: int) -> Optional[IntegrityNode]:
+        name = current.name
+        if name in visiting:
+            chain.errors.append(f"Dependency cycle detected at {name}")
+            return None
+        if name in chain.nodes:
+            return chain.nodes[name]
+        if depth > max_depth:
+            chain.errors.append(f"Maximum dependency depth exceeded at {name}")
+            return None
+
+        visiting.add(name)
+        dep_hashes: Dict[str, str] = {}
+        dependencies = _dependency_names(current)
+        for dep_name in dependencies:
+            from .resolver import ToolSpec
+            from .semver import matches_constraint
+            dep_spec = ToolSpec.parse(dep_name)
+            dep_tool = load_tool(dep_spec.full_name)
+            if dep_tool is None:
+                chain.errors.append(f"Unresolved dependency: {dep_name}")
+                continue
+            if dep_spec.version and not matches_constraint(
+                dep_tool.version or "", dep_spec.version
+            ):
+                chain.errors.append(
+                    f"Dependency version mismatch: {dep_name} "
+                    f"resolved to {dep_tool.version or 'unknown'}"
+                )
+                continue
+            dep_node = visit(dep_tool, depth + 1)
+            if dep_node is not None:
+                dep_hashes[dep_name] = dep_node.content_hash
+        visiting.remove(name)
+
+        definition_hash = compute_tool_hash(current)
+        node = IntegrityNode(
+            name=name,
+            version=current.version or "",
+            definition_hash=definition_hash,
+            content_hash=_identity_hash(definition_hash, dep_hashes),
+            dependencies=dependencies,
+            dependency_hashes=dep_hashes,
+        )
+        chain.nodes[name] = node
+        return node
+
+    root = visit(tool, 0)
+    if root is not None:
+        chain.root = root
     return chain
 
 
-def _build_node(tool: Tool) -> IntegrityNode:
+def verify_integrity(chain: IntegrityChain, *, verify_installed: bool = False) -> bool:
-    """Build an integrity node from a tool."""
+    """Verify chain structure, optionally rebuilding it from installed tools."""
-    deps = []
+    if not chain.is_valid:
-    dep_hashes = {}
+        return False
-
+    if not verify_installed:
-    for step in tool.steps:
+        return True
-        if hasattr(step, "tool") and step.tool:
+    root_tool = load_tool(chain.root.name)
-            deps.append(step.tool)
+    if root_tool is None:
-
+        return False
-    # Resolve dependency hashes
+    rebuilt = build_integrity_chain(root_tool)
-    for dep_name in deps:
+    return rebuilt.is_valid and _constant_time_equal(
-        dep_tool = load_tool(dep_name)
+        rebuilt.root.content_hash, chain.root.content_hash
-        if dep_tool:
-            dep_hashes[dep_name] = compute_tool_hash(dep_tool)
-        else:
-            dep_hashes[dep_name] = "unresolved"
-
-    return IntegrityNode(
-        name=tool.name,
-        version=tool.version or "",
-        content_hash=compute_tool_hash(tool),
-        dependencies=deps,
-        dependency_hashes=dep_hashes,
     )
-
-
-def _traverse_deps(tool: Tool, chain: IntegrityChain, depth: int, max_depth: int):
-    """Recursively traverse dependencies and add to chain."""
-    if depth >= max_depth:
-        return
-
-    for step in tool.steps:
-        if hasattr(step, "tool") and step.tool:
-            dep_name = step.tool
-            if dep_name in chain.nodes:
-                continue
-            dep_tool = load_tool(dep_name)
-            if not dep_tool:
-                continue
-            node = _build_node(dep_tool)
-            chain.nodes[dep_name] = node
-            _traverse_deps(dep_tool, chain, depth + 1, max_depth)
-
-
-def verify_integrity(chain: IntegrityChain) -> bool:
-    """Verify that all hashes in the chain are consistent."""
-    return chain.is_valid

src/cmdforge/lockfile.py

@@ -26,6 +26,8 @@ class LockedPackage:
     direct: bool                         # True if in manifest
     required_by: List[str] = field(default_factory=list)  # Parent packages
     path: Optional[str] = None           # Relative path for local tools
+    content_hash: str = ""               # Definition plus dependency identities
+    dependency_hashes: Dict[str, str] = field(default_factory=dict)
 
     @property
     def owner(self) -> str:
@@ -145,7 +147,9 @@ class Lockfile:
                 source=pkg_data.get("source", "registry"),
                 direct=pkg_data.get("direct", False),
                 required_by=pkg_data.get("required_by", []),
-                path=pkg_data.get("path")
+                path=pkg_data.get("path"),
+                content_hash=pkg_data.get("content_hash", ""),
+                dependency_hashes=pkg_data.get("dependency_hashes", {})
             )
 
         return cls(
@@ -180,6 +184,10 @@ class Lockfile:
                 pkg_dict["required_by"] = pkg.required_by
             if pkg.path:
                 pkg_dict["path"] = pkg.path
+            if pkg.content_hash:
+                pkg_dict["content_hash"] = pkg.content_hash
+            if pkg.dependency_hashes:
+                pkg_dict["dependency_hashes"] = dict(sorted(pkg.dependency_hashes.items()))
             d["packages"][name] = pkg_dict
 
         return d
@@ -259,6 +267,34 @@ def generate_lockfile(
         )
         lock.packages[qualified_name] = pkg
 
+    from .integrity import compute_content_identity
+    visiting = set()
+
+    def identity_for(package_name: str) -> str:
+        pkg = lock.packages.get(package_name)
+        node = graph.nodes.get(package_name)
+        if pkg is None or node is None or not pkg.integrity:
+            return ""
+        if pkg.content_hash:
+            return pkg.content_hash
+        if package_name in visiting:
+            return ""
+        visiting.add(package_name)
+        dependencies = {}
+        for child_name in sorted(node.children):
+            child_hash = identity_for(child_name)
+            if not child_hash:
+                visiting.remove(package_name)
+                return ""
+            dependencies[child_name] = child_hash
+        visiting.remove(package_name)
+        pkg.dependency_hashes = dependencies
+        pkg.content_hash = compute_content_identity(pkg.integrity, dependencies)
+        return pkg.content_hash
+
+    for package_name in sorted(lock.packages):
+        identity_for(package_name)
+
     return lock
 
 
@@ -368,4 +404,30 @@ def verify_lockfile(
             except Exception as e:
                 errors.append(f"{name}: could not verify integrity ({e})")
 
+        if locked.content_hash:
+            from .integrity import compute_content_identity
+            actual_dependencies = {}
+            missing_dependencies = []
+            for dep_name, expected_hash in locked.dependency_hashes.items():
+                dep = lock.packages.get(dep_name)
+                if dep is None or not dep.content_hash:
+                    missing_dependencies.append(dep_name)
+                elif dep.content_hash != expected_hash:
+                    errors.append(
+                        f"{name}: dependency identity mismatch ({dep_name})"
+                    )
+                else:
+                    actual_dependencies[dep_name] = dep.content_hash
+            if missing_dependencies:
+                errors.append(
+                    f"{name}: unresolved integrity dependencies: "
+                    + ", ".join(missing_dependencies)
+                )
+            elif locked.integrity:
+                actual_identity = compute_content_identity(
+                    locked.integrity, actual_dependencies
+                )
+                if actual_identity != locked.content_hash:
+                    errors.append(f"{name}: transitive content identity mismatch")
+
     return errors

src/cmdforge/mcp_client.py

@@ -13,6 +13,7 @@ from dataclasses import dataclass, field
 from datetime import timedelta
 from pathlib import Path
 from typing import Any, Awaitable, Callable, Dict, List, Optional, TypeVar
+from urllib.parse import urlparse
 
 import yaml
 
@@ -20,7 +21,7 @@ MCP_CONFIG_FILE = Path.home() / ".cmdforge" / "mcp.yaml"
 MCP_CONFIG_VERSION = 1
 MCP_DEPTH_ENV = "CMDFORGE_MCP_DEPTH"
 RESULT_MODES = ("auto", "structured", "content", "text")
-SUPPORTED_TRANSPORTS = ("stdio",)
+SUPPORTED_TRANSPORTS = ("stdio", "streamable-http")
 DEFAULT_INHERITED_ENV = (
     "PATH",
     "HOME",
@@ -95,13 +96,15 @@ def _glob_match(pattern: str, name: str) -> bool:
 class McpServerConfig:
     """Configuration for one MCP server.
 
-    M7.2 intentionally supports stdio only. ``approved`` records explicit
+    ``approved`` records explicit user consent to execute a local command or
-    user consent to execute the configured local command.
+    connect to the configured HTTP endpoint.
     """
 
     name: str
     transport: str = "stdio"
     command: Optional[str] = None
+    url: Optional[str] = None
+    headers: Dict[str, str] = field(default_factory=dict)
     args: List[str] = field(default_factory=list)
     cwd: Optional[str] = None
     env: Dict[str, str] = field(default_factory=dict)
@@ -116,10 +119,25 @@ class McpServerConfig:
         if self.transport not in SUPPORTED_TRANSPORTS:
             raise ValueError(
                 f"Unsupported MCP transport '{self.transport}'. "
-                f"Supported in M7.2: {', '.join(SUPPORTED_TRANSPORTS)}"
+                f"Supported transports: {', '.join(SUPPORTED_TRANSPORTS)}"
             )
-        if not isinstance(self.command, str) or not self.command.strip():
+        if self.transport == "stdio":
-            raise ValueError(f"MCP server '{self.name}' requires a command")
+            if not isinstance(self.command, str) or not self.command.strip():
+                raise ValueError(f"MCP server '{self.name}' requires a command")
+            if self.url is not None:
+                raise ValueError(f"MCP stdio server '{self.name}' cannot define url")
+            if self.headers:
+                raise ValueError(f"MCP stdio server '{self.name}' cannot define headers")
+        else:
+            if self.command is not None:
+                raise ValueError(
+                    f"MCP streamable-http server '{self.name}' cannot define command"
+                )
+            _validate_remote_url(self.name, self.url)
+            if self.args or self.cwd or self.env:
+                raise ValueError(
+                    f"MCP streamable-http server '{self.name}' cannot define args, cwd, or env"
+                )
         if not isinstance(self.args, list) or not all(isinstance(arg, str) for arg in self.args):
             raise ValueError(f"MCP server '{self.name}' args must be a list of strings")
         if not isinstance(self.env, dict) or not all(
@@ -131,6 +149,15 @@ class McpServerConfig:
             raise ValueError(
                 f"MCP server '{self.name}' env must map valid variable names to strings"
             )
+        if not isinstance(self.headers, dict) or not all(
+            isinstance(key, str) and key.strip()
+            and "\n" not in key and "\r" not in key
+            and isinstance(value, str) and "\n" not in value and "\r" not in value
+            for key, value in self.headers.items()
+        ):
+            raise ValueError(
+                f"MCP server '{self.name}' headers must map safe names to strings"
+            )
         if not isinstance(self.inherit_env, list) or not all(
             isinstance(name, str) and _ENV_NAME.fullmatch(name)
             for name in self.inherit_env
@@ -155,6 +182,8 @@ def _fingerprint(cfg: McpServerConfig) -> str:
         [
             cfg.transport,
             cfg.command,
+            cfg.url,
+            cfg.headers,
             cfg.args,
             cfg.cwd,
             cfg.env,
@@ -203,6 +232,36 @@ def _build_server_env(cfg: McpServerConfig) -> Dict[str, str]:
     return environment
 
 
+def _is_loopback_host(hostname: Optional[str]) -> bool:
+    if not hostname:
+        return False
+    if hostname.lower() == "localhost":
+        return True
+    try:
+        import ipaddress
+        return ipaddress.ip_address(hostname).is_loopback
+    except ValueError:
+        return False
+
+
+def _validate_remote_url(name: str, url: Optional[str]) -> None:
+    if not isinstance(url, str) or not url.strip():
+        raise ValueError(f"MCP streamable-http server '{name}' requires a url")
+    parsed = urlparse(url)
+    if parsed.scheme not in ("http", "https") or not parsed.hostname:
+        raise ValueError(f"MCP server '{name}' has an invalid HTTP URL")
+    if parsed.username or parsed.password:
+        raise ValueError("MCP URLs must not contain credentials; use headers")
+    if parsed.scheme != "https" and not _is_loopback_host(parsed.hostname):
+        raise ValueError(
+            f"Remote MCP server '{name}' must use HTTPS; HTTP is allowed only for localhost"
+        )
+
+
+def _build_http_headers(cfg: McpServerConfig) -> Dict[str, str]:
+    return {name: _expand_env_value(value) for name, value in cfg.headers.items()}
+
+
 def _load_mcp_document() -> dict:
     if not MCP_CONFIG_FILE.exists():
         return {}
@@ -229,6 +288,8 @@ def load_mcp_config() -> List[McpServerConfig]:
             name=name,
             transport=raw.get("transport", "stdio"),
             command=raw.get("command"),
+            url=raw.get("url"),
+            headers=raw.get("headers", {}),
             args=raw.get("args", []),
             cwd=raw.get("cwd"),
             env=raw.get("env", {}),
@@ -276,9 +337,14 @@ def save_mcp_config(servers: List[McpServerConfig]) -> None:
 def _server_to_dict(server: McpServerConfig) -> dict:
     data: dict = {
         "transport": server.transport,
-        "command": server.command,
         "approved": server.approved,
     }
+    if server.command:
+        data["command"] = server.command
+    if server.url:
+        data["url"] = server.url
+    if server.headers:
+        data["headers"] = server.headers
     if server.args:
         data["args"] = server.args
     if server.cwd:
@@ -329,8 +395,8 @@ def _serialize_tool(tool: Any) -> Dict[str, Any]:
 class McpClientManager:
     """Invocation-scoped MCP configuration and schema manager.
 
-    SDK sessions are intentionally scoped to each discovery or tool operation
+    SDK sessions are intentionally scoped to each discovery or tool operation.
-    in M7.2. This guarantees subprocess cleanup while the manager reuses loaded
+    This guarantees transport cleanup while the manager reuses loaded
     configuration and discovered schemas across all MCP steps in one tool run.
     """
 
@@ -368,7 +434,10 @@ class McpClientManager:
             raise KeyError(f"MCP server '{server_name}' not configured. Add it to {MCP_CONFIG_FILE}")
         cfg.validate()
         if not cfg.approved:
-            command = " ".join([cfg.command or ""] + cfg.args)
+            command = (
+                " ".join([cfg.command or ""] + cfg.args)
+                if cfg.transport == "stdio" else cfg.url or ""
+            )
             raise PermissionError(
                 f"MCP server '{server_name}' is not approved to execute: {_sanitize(command)}. "
                 "Re-add it with 'cmdforge mcp add ...' or set approved: true after review."
@@ -401,7 +470,7 @@ class McpClientManager:
                 seen_cursors.add(cursor)
             return tools
 
-        tools = _run_stdio_operation(cfg, list_all_tools, "discover tools")
+        tools = _run_operation(cfg, list_all_tools, "discover tools")
         self._schemas[server_name] = {tool["name"]: tool for tool in tools}
         self._schema_fingerprints[server_name] = fingerprint
         return tools
@@ -440,7 +509,57 @@ class McpClientManager:
             )
             return _normalize_result(result, result_mode)
 
-        return _run_stdio_operation(cfg, call, f"call {tool_name}")
+        return _run_operation(cfg, call, f"call {tool_name}")
+
+
+def _run_operation(
+    cfg: McpServerConfig,
+    operation: Callable[[Any], Awaitable[T]],
+    operation_name: str,
+) -> T:
+    if cfg.transport == "stdio":
+        return _run_stdio_operation(cfg, operation, operation_name)
+    return _run_http_operation(cfg, operation, operation_name)
+
+
+def _run_http_operation(
+    cfg: McpServerConfig,
+    operation: Callable[[Any], Awaitable[T]],
+    operation_name: str,
+) -> T:
+    from mcp import ClientSession
+    from mcp.client.streamable_http import streamable_http_client
+    import httpx
+
+    resolved_headers = _build_http_headers(cfg)
+
+    async def run() -> T:
+        timeout = httpx.Timeout(cfg.timeout)
+        async with httpx.AsyncClient(
+            headers=resolved_headers, timeout=timeout, follow_redirects=False
+        ) as client:
+            async with streamable_http_client(
+                cfg.url, http_client=client
+            ) as (read, write, _):
+                async with ClientSession(
+                    read, write,
+                    read_timeout_seconds=timedelta(seconds=cfg.timeout),
+                ) as session:
+                    await session.initialize()
+                    return await operation(session)
+
+    try:
+        return asyncio.run(asyncio.wait_for(run(), timeout=cfg.timeout))
+    except (asyncio.TimeoutError, TimeoutError) as exc:
+        raise TimeoutError(
+            f"MCP {operation_name} on '{cfg.name}' timed out after {cfg.timeout}s"
+        ) from exc
+    except Exception as exc:
+        secrets = list(resolved_headers.values())
+        raise RuntimeError(
+            f"MCP {operation_name} on '{cfg.name}' failed: "
+            f"{_sanitize(str(exc), secrets)}"
+        ) from exc
 
 
 def _run_stdio_operation(

src/cmdforge/mcp_server.py

@@ -1,14 +1,18 @@
 """MCP server support: expose CmdForge tools as MCP tools."""
 
 import inspect
+import hmac
 import os
 import sys
-from typing import Annotated, Any, Literal
+from typing import Annotated, Any, Literal, Optional
+from urllib.parse import urlparse
 
 from .mcp_client import (
     MCP_DEPTH_ENV,
     McpServeConfig,
     _MCP_CALL_DEPTH,
+    _expand_env_value,
+    _is_loopback_host,
     _require_mcp_sdk,
     load_mcp_serve_config,
 )
@@ -87,16 +91,70 @@ def _build_tool_schema(tool) -> dict:
     }
 
 
-def serve(transport: str = "stdio") -> None:
+class _StaticTokenVerifier:
+    def __init__(self, token: str):
+        self._token = token
+
+    async def verify_token(self, token: str):
+        if not hmac.compare_digest(token, self._token):
+            return None
+        from mcp.server.auth.provider import AccessToken
+        return AccessToken(
+            token=token, client_id="cmdforge-mcp-client", scopes=[],
+            subject="cmdforge-mcp",
+        )
+
+
+def _validate_http_server_options(
+    host: str,
+    port: int,
+    allowed_origins: list[str],
+    auth_token: Optional[str],
+    external_url: Optional[str],
+) -> tuple[list[str], Optional[str]]:
+    if isinstance(port, bool) or not isinstance(port, int) or not 1 <= port <= 65535:
+        raise ValueError("MCP server port must be between 1 and 65535")
+    local = _is_loopback_host(host)
+    resolved_token = _expand_env_value(auth_token) if auth_token else None
+    if not local:
+        parsed = urlparse(external_url or "")
+        if parsed.scheme != "https" or not parsed.hostname:
+            raise ValueError(
+                "Non-local MCP serving requires --external-url https://... "
+                "behind a TLS-terminating proxy"
+            )
+        if not resolved_token:
+            raise ValueError("Non-local MCP serving requires --auth-token")
+    origins = allowed_origins or (
+        [f"http://localhost:{port}", f"http://127.0.0.1:{port}"]
+        if local else [f"{urlparse(external_url).scheme}://{urlparse(external_url).netloc}"]
+    )
+    for origin in origins:
+        parsed = urlparse(origin)
+        if parsed.scheme not in ("http", "https") or not parsed.netloc or parsed.path not in ("", "/"):
+            raise ValueError(f"Invalid allowed origin: {origin}")
+        if not local and parsed.scheme != "https":
+            raise ValueError("Non-local MCP origins must use HTTPS")
+    return origins, resolved_token
+
+
+def serve(
+    transport: str = "stdio",
+    *,
+    host: str = "127.0.0.1",
+    port: int = 8000,
+    allowed_origins: Optional[list[str]] = None,
+    auth_token: Optional[str] = None,
+    external_url: Optional[str] = None,
+) -> None:
     """Start CmdForge as an MCP server on the given transport.
 
-    Currently only stdio transport is supported (M7.3).
+    HTTP defaults to loopback and enables SDK DNS-rebinding/origin protection.
+    Non-local binding requires bearer auth and an explicit public HTTPS URL;
+    TLS is expected to terminate at the reverse proxy represented by that URL.
     """
-    if transport != "stdio":
+    if transport not in ("stdio", "streamable-http"):
-        raise NotImplementedError(
+        raise ValueError(f"Unsupported MCP server transport '{transport}'")
-            f"Unsupported MCP server transport '{transport}'. "
-            f"M7.3 supports stdio only."
-        )
 
     _require_mcp_sdk()
     from mcp.server.fastmcp import FastMCP
@@ -104,7 +162,37 @@ def serve(transport: str = "stdio") -> None:
     serve_config = load_mcp_serve_config()
 
     from .tool import list_tools, load_tool
-    server = FastMCP("CmdForge")
+    server_options = {}
+    if transport == "streamable-http":
+        origins, resolved_token = _validate_http_server_options(
+            host, port, allowed_origins or [], auth_token, external_url
+        )
+        from mcp.server.transport_security import TransportSecuritySettings
+        allowed_hosts = [host, f"{host}:{port}", "localhost", f"localhost:{port}"]
+        if external_url:
+            external_host = urlparse(external_url).netloc
+            allowed_hosts.extend([external_host, urlparse(external_url).hostname or ""])
+        server_options.update({
+            "host": host,
+            "port": port,
+            "transport_security": TransportSecuritySettings(
+                enable_dns_rebinding_protection=True,
+                allowed_hosts=[value for value in allowed_hosts if value],
+                allowed_origins=origins,
+            ),
+        })
+        if resolved_token:
+            from mcp.server.auth.settings import AuthSettings
+            issuer = external_url or f"http://{host}:{port}"
+            server_options.update({
+                "token_verifier": _StaticTokenVerifier(resolved_token),
+                "auth": AuthSettings(
+                    issuer_url=issuer,
+                    resource_server_url=f"{issuer.rstrip('/')}/mcp",
+                    required_scopes=[],
+                ),
+            })
+    server = FastMCP("CmdForge", **server_options)
 
     tool_names = list_tools()
     exposed_count = 0

src/cmdforge/prompt_optimizer.py

@@ -5,7 +5,7 @@ contract conformance tests, and surfaces the highest-scoring variation.
 """
 
 from dataclasses import dataclass, field
-from typing import Any, Dict, List, Optional
+from typing import Callable, Dict, List, Optional
 
 from .tool import Tool, PromptStep
 
@@ -37,6 +37,8 @@ class OptimizationResult:
     scores: Dict[int, int] = field(default_factory=dict)  # variation_index -> score
     best_index: Optional[int] = None
     baseline_score: int = 0
+    evaluation_kind: str = "structural"
+    note: str = ""
 
     @property
     def best(self) -> Optional[PromptVariation]:
@@ -55,6 +57,8 @@ class OptimizationResult:
         return {
             "tool": self.tool_name,
             "baseline_score": self.baseline_score,
+            "evaluation_kind": self.evaluation_kind,
+            "note": self.note,
             "best_index": self.best_index,
             "improvement_pct": round(self.improvement, 1),
             "best": self.best.to_dict() if self.best else None,
@@ -93,7 +97,7 @@ def generate_variations(
 
         original = step.prompt
 
-        strategies = _get_strategies(original, count)
+        strategies = _get_strategies(original, count, provider=provider)
 
         for strategy, variation_text in strategies:
             if variation_text and variation_text != original:
@@ -107,16 +111,25 @@ def generate_variations(
     return variations
 
 
-def _get_strategies(original: str, count: int) -> List[tuple]:
+def _get_strategies(
+    original: str, count: int, provider: str = "mock"
+) -> List[tuple]:
     """Generate variations using different strategies."""
     strategies: List[tuple] = []
 
-    # Attempt AI-driven rephrase if a real provider is available
+    if count <= 0:
-    rephrased = _call_provider_for_variation(
+        return []
-        f"Rephrase this instruction while keeping the same meaning. Output only the rephrased text:\n\n{original}"
+
-    )
+    # Provider calls are explicit.  The default/mock path is deterministic and
-    if rephrased and rephrased != original:
+    # never sends a user's prompt to an external process or service.
-        strategies.append(("rephrase", rephrased))
+    if provider != "mock":
+        rephrased = _call_provider_for_variation(
+            provider,
+            "Rephrase this instruction while keeping the same meaning. "
+            f"Output only the rephrased text:\n\n{original}",
+        )
+        if rephrased and rephrased != original:
+            strategies.append(("rephrase", rephrased))
 
     # Deterministic mock variations
     mock_strategies = [
@@ -141,17 +154,15 @@ def _truncate(text: str, max_len: int) -> str:
     return text[:max_len - 3].rsplit(" ", 1)[0] + "..."
 
 
-def _call_provider_for_variation(prompt: str) -> Optional[str]:
+def _call_provider_for_variation(provider: str, prompt: str) -> Optional[str]:
-    """Try to call a provider for variation generation."""
+    """Call the provider explicitly selected by the user."""
-    try:
+    from .providers import call_provider
-        from .providers import call_provider
+
-        result = call_provider("opencode-pickle", prompt, timeout=15)
+    result = call_provider(provider, prompt, timeout=15)
-        if result.success and result.text:
+    if result.success and result.text:
-            text = result.text.strip()
+        text = result.text.strip()
-            if text and len(text) >= 10:
+        if text and len(text) >= 10:
-                return text
+            return text
-    except Exception:
-        pass
     return None
 
 
@@ -159,6 +170,7 @@ def optimize_tool(
     tool: Tool,
     count: int = 5,
     provider: str = "mock",
+    evaluator: Optional[Callable[[Tool], float]] = None,
 ) -> OptimizationResult:
     """Run prompt optimization for a tool.
 
@@ -175,11 +187,17 @@ def optimize_tool(
     """
     from .contract_testing import run_contract_tests
 
-    result = OptimizationResult(tool_name=tool.name)
+    result = OptimizationResult(
+        tool_name=tool.name,
+        evaluation_kind="behavioral" if evaluator else "structural",
+    )
 
     # Baseline score
-    baseline = run_contract_tests(tool)
+    if evaluator:
-    result.baseline_score = _count_passed(baseline)
+        result.baseline_score = evaluator(tool)
+    else:
+        baseline = run_contract_tests(tool)
+        result.baseline_score = _count_passed(baseline)
 
     # Generate variations
     result.variations = generate_variations(tool, count=count, provider=provider)
@@ -198,14 +216,26 @@ def optimize_tool(
                 step.prompt = variation.variation
 
         try:
-            test_result = run_contract_tests(tool_copy)
+            if evaluator:
-            result.scores[idx] = _count_passed(test_result)
+                result.scores[idx] = evaluator(tool_copy)
+            else:
+                test_result = run_contract_tests(tool_copy)
+                result.scores[idx] = _count_passed(test_result)
         except Exception:
             result.scores[idx] = 0
 
     # Find best
-    if result.scores:
+    if evaluator and result.scores:
-        result.best_index = max(result.scores, key=lambda k: result.scores[k])
+        candidate = max(result.scores, key=lambda k: result.scores[k])
+        if result.scores[candidate] > result.baseline_score:
+            result.best_index = candidate
+        else:
+            result.note = "No variation improved on the behavioral baseline."
+    elif result.scores:
+        result.note = (
+            "Structural conformance cannot compare prompt semantics; "
+            "no best variation was selected."
+        )
 
     return result
 
@@ -214,3 +244,57 @@ def _count_passed(test_result) -> int:
     """Count passed tests in a ConformanceReport."""
     return sum(1 for r in getattr(test_result, "results", [])
                if getattr(r, "state", "") == "passed")
+
+
+def load_behavioral_evaluator(path, provider_override: Optional[str] = None):
+    """Load explicit behavioral cases and return a semantic score function.
+
+    The JSON file must contain a list of cases. Each case accepts ``input`` and
+    optional ``args``, plus exactly one of ``expected`` (exact value) or
+    ``contains`` (substring). Supplying this file is explicit authorization to
+    execute the candidate tools and their configured providers.
+    """
+    import json
+    from pathlib import Path
+
+    cases = json.loads(Path(path).read_text(encoding="utf-8"))
+    if not isinstance(cases, list) or not cases:
+        raise ValueError("Behavior test file must contain a non-empty JSON list")
+    for index, case in enumerate(cases):
+        if not isinstance(case, dict):
+            raise ValueError(f"Behavior case {index} must be an object")
+        assertions = [key for key in ("expected", "contains") if key in case]
+        if len(assertions) != 1:
+            raise ValueError(
+                f"Behavior case {index} needs exactly one of expected or contains"
+            )
+        if not isinstance(case.get("args", {}), dict):
+            raise ValueError(f"Behavior case {index} args must be an object")
+
+    def evaluate(tool: Tool) -> float:
+        from .runner import run_tool
+
+        passed = 0
+        for case in cases:
+            output, exit_code = run_tool(
+                tool,
+                input_text=str(case.get("input", "")),
+                custom_args=case.get("args", {}),
+                provider_override=provider_override,
+                dry_run=False,
+                verbose=False,
+            )
+            if exit_code != 0:
+                continue
+            if "expected" in case:
+                try:
+                    actual = json.loads(output)
+                except (TypeError, json.JSONDecodeError):
+                    actual = output
+                if actual == case["expected"]:
+                    passed += 1
+            elif str(case["contains"]) in output:
+                passed += 1
+        return passed
+
+    return evaluate

src/cmdforge/registry/app.py

@@ -1140,6 +1140,25 @@ def create_app() -> Flask:
             except (AttributeError, TypeError, ValueError):
                 audit_stale = True
 
+        contributor_rows = query_all(
+            g.db,
+            """SELECT p.slug, c.contribution FROM tool_contributors c
+               JOIN publishers p ON p.id = c.publisher_id
+               WHERE c.tool_id = ? ORDER BY p.slug""",
+            [row["id"]],
+        )
+        approved_improvement = query_one(
+            g.db,
+            """SELECT id FROM improvement_submissions
+               WHERE applied_tool_id = ? LIMIT 1""",
+            [row["id"]],
+        )
+        badges = []
+        if row.get("attestation_json"):
+            badges.append("verified")
+        if approved_improvement:
+            badges.extend(["optimized", "community-reviewed"])
+
         payload = {
             "owner": row["owner"],
             "name": row["name"],
@@ -1164,6 +1183,11 @@ def create_app() -> Flask:
             "quality": json.loads(latest_audit["quality_json"]) if latest_audit else None,
             "audit_evaluated_at": latest_audit["evaluated_at"] if latest_audit else None,
             "audit_stale": audit_stale,
+            "contributors": [
+                {"name": item["slug"], "contribution": item["contribution"]}
+                for item in contributor_rows
+            ],
+            "badges": badges,
         }
         response = jsonify({"data": payload})
         response.headers["Cache-Control"] = "max-age=60"
@@ -1313,6 +1337,18 @@ def create_app() -> Flask:
             except Exception:
                 g.db.rollback()
 
+        publisher = query_one(
+            g.db, "SELECT signing_public_key FROM publishers WHERE id = ?",
+            [row["publisher_id"]],
+        )
+        attestation = None
+        if row.get("attestation_json"):
+            try:
+                attestation = json.loads(row["attestation_json"])
+            except (TypeError, json.JSONDecodeError):
+                return error_response(
+                    "INVALID_ATTESTATION", "Stored release attestation is corrupt", 500
+                )
         response = jsonify({
             "data": {
                 "owner": row["owner"],
@@ -1321,12 +1357,172 @@ def create_app() -> Flask:
                 "config": row["config_yaml"],
                 "readme": row["readme"] or "",
                 "config_hash": row.get("config_hash") or "",
+                "content_hash": row.get("content_hash") or "",
+                "dependency_hashes": json.loads(
+                    row.get("dependency_hashes_json") or "{}"
+                ),
                 "defaults": row.get("defaults") or "",
+                "attestation": attestation,
+                "signing_public_key": (
+                    publisher["signing_public_key"] if publisher else ""
+                ) or "",
             }
         })
         response.headers["Cache-Control"] = "max-age=3600, immutable"
         return response
 
+    @app.route("/api/v1/tools/by-content-hash/<path:content_hash>", methods=["GET"])
+    def tool_by_content_hash(content_hash: str) -> Response:
+        """Resolve an approved public release by its transitive identity."""
+        if not re.fullmatch(r"sha256:[0-9a-f]{64}", content_hash):
+            return error_response("VALIDATION_ERROR", "Invalid content hash", 400)
+        row = query_one(
+            g.db,
+            """SELECT owner, name, version, description, category, tags,
+                      downloads, published_at
+               FROM tools WHERE content_hash = ? AND visibility = 'public'
+                 AND moderation_status = 'approved'
+               ORDER BY published_at DESC LIMIT 1""",
+            [content_hash],
+        )
+        if not row:
+            return error_response("TOOL_NOT_FOUND", "Content hash not found", 404)
+        result = dict(row)
+        result["tags"] = json.loads(result.get("tags") or "[]")
+        return jsonify({"data": result})
+
+    @app.route(
+        "/api/v1/tools/<owner>/<name>/<version>/improvements", methods=["POST"]
+    )
+    @require_token
+    def submit_improvement(owner: str, name: str, version: str) -> Response:
+        """Submit and automatically test a prompt/code improvement."""
+        row = query_one(
+            g.db,
+            """SELECT id, config_yaml, visibility, owner FROM tools
+               WHERE owner = ? AND name = ? AND version = ?
+                 AND moderation_status = 'approved'""",
+            [owner, name, version],
+        )
+        if not row:
+            return error_response("TOOL_NOT_FOUND", "Tool release not found", 404)
+        if (
+            row["visibility"] != "public"
+            and g.current_publisher["slug"] != row["owner"]
+            and g.current_publisher.get("role") not in ("moderator", "admin")
+        ):
+            return error_response("TOOL_NOT_FOUND", "Tool release not found", 404)
+        payload = request.get_json(silent=True) or {}
+        try:
+            from ..community import create_submission, run_submission_tests
+            from ..tool import Tool
+            tool = Tool.from_dict(yaml.safe_load(row["config_yaml"]) or {})
+            submission = create_submission(
+                tool,
+                int(payload.get("step_index", -1)),
+                str(payload.get("proposed", "")),
+                g.current_publisher["slug"],
+                str(payload.get("rationale", "")),
+            )
+            if not submission.proposed.strip():
+                raise ValueError("Proposed content must not be empty")
+            test_result = run_submission_tests(submission, tool)
+        except (KeyError, TypeError, ValueError) as exc:
+            return error_response("VALIDATION_ERROR", str(exc), 400)
+        if not test_result.get("passed_for_review"):
+            return error_response(
+                "IMPROVEMENT_TEST_FAILED",
+                test_result.get("error") or "Proposed change did not pass automated tests",
+                400,
+                details={"test_result": test_result},
+            )
+        cursor = g.db.execute(
+            """INSERT INTO improvement_submissions (
+                   tool_id, submitter_id, step_index, step_type, original,
+                   proposed, rationale, status, test_result_json, submitted_at
+               ) VALUES (?, ?, ?, ?, ?, ?, ?, 'tested', ?, ?)""",
+            [
+                row["id"], g.current_publisher["id"], submission.step_index,
+                submission.step_type, submission.original, submission.proposed,
+                submission.rationale, json.dumps(test_result, sort_keys=True),
+                submission.submitted_at,
+            ],
+        )
+        g.db.commit()
+        return jsonify({
+            "data": {"id": cursor.lastrowid, **submission.to_dict()}
+        }), 201
+
+    @app.route("/api/v1/improvements/<int:submission_id>", methods=["PATCH"])
+    @require_token
+    def review_improvement(submission_id: int) -> Response:
+        """Allow the tool owner/moderator to review a tested submission."""
+        row = query_one(
+            g.db,
+            """SELECT s.*, t.owner FROM improvement_submissions s
+               JOIN tools t ON t.id = s.tool_id WHERE s.id = ?""",
+            [submission_id],
+        )
+        if not row:
+            return error_response("NOT_FOUND", "Improvement not found", 404)
+        if (
+            g.current_publisher["slug"] != row["owner"]
+            and g.current_publisher.get("role") not in ("moderator", "admin")
+        ):
+            return error_response("FORBIDDEN", "Owner or moderator required", 403)
+        decision = (request.get_json(silent=True) or {}).get("decision")
+        if decision not in ("approve", "reject", "request_changes"):
+            return error_response("VALIDATION_ERROR", "Invalid decision", 400)
+        status = "approved" if decision == "approve" else decision
+        now = datetime.now(timezone.utc).isoformat()
+        notes = (request.get_json(silent=True) or {}).get("notes", "")
+        g.db.execute(
+            """UPDATE improvement_submissions SET status = ?, reviewer_id = ?,
+                      review_notes = ?, reviewed_at = ? WHERE id = ?""",
+            [status, g.current_publisher["id"], notes, now, submission_id],
+        )
+        g.db.commit()
+        return jsonify({
+            "data": {
+                "id": submission_id, "status": status,
+                "ready_to_apply": status == "approved",
+            }
+        })
+
+    @app.route("/api/v1/me/signing-key", methods=["PUT"])
+    @require_token
+    def set_signing_key() -> Response:
+        """Register an Ed25519 public key as the publisher trust anchor."""
+        public_key = (request.get_json(silent=True) or {}).get("public_key", "")
+        try:
+            import base64
+            raw = base64.b64decode(public_key, validate=True)
+            if len(raw) != 32:
+                raise ValueError
+        except (ValueError, TypeError):
+            return error_response(
+                "VALIDATION_ERROR", "public_key must be a base64 Ed25519 key", 400
+            )
+        existing_key = query_one(
+            g.db, "SELECT signing_public_key FROM publishers WHERE id = ?",
+            [g.current_publisher["id"]],
+        )
+        current_key = (
+            existing_key["signing_public_key"] if existing_key else ""
+        ) or ""
+        if current_key and current_key != public_key:
+            return error_response(
+                "SIGNING_KEY_EXISTS",
+                "Signing key rotation requires administrator recovery",
+                409,
+            )
+        g.db.execute(
+            "UPDATE publishers SET signing_public_key = ? WHERE id = ?",
+            [public_key, g.current_publisher["id"]],
+        )
+        g.db.commit()
+        return jsonify({"data": {"signing_public_key": public_key}})
+
     @app.route("/api/v1/categories", methods=["GET"])
     def list_categories() -> Response:
         page, per_page, sort, order, error = parse_pagination("/categories", "name")
@@ -2470,6 +2666,8 @@ def create_app() -> Flask:
         readme = payload.get("readme") or ""
         defaults = payload.get("defaults") or ""
         dry_run = bool(payload.get("dry_run"))
+        attestation_data = payload.get("attestation")
+        improvement_id = payload.get("improvement_id")
 
         size_resp = validate_payload_size("config", config_text, MAX_CONFIG_BYTES)
         if size_resp:
@@ -2638,10 +2836,64 @@ def create_app() -> Flask:
 
         # Compute config hash early for idempotency check
         config_hash = compute_yaml_hash(config_text)
+        from ..integrity import compute_content_identity
+        from ..tool import ToolStep
+        dependency_names = set(published_tool.dependencies)
+        dependency_names.update(
+            step.tool for step in published_tool.steps if isinstance(step, ToolStep)
+        )
+        dependency_hashes = {}
+        for dependency_name in sorted(dependency_names):
+            from ..resolver import ToolSpec
+            dep_spec = ToolSpec.parse(dependency_name)
+            dep_owner = dep_spec.owner or owner
+            dep_row = resolve_tool(dep_owner, dep_spec.name, dep_spec.version)
+            if dep_row:
+                dep_identity = dep_row.get("content_hash") or dep_row.get("config_hash")
+                if dep_identity:
+                    dependency_hashes[dependency_name] = dep_identity
+        content_hash = compute_content_identity(config_hash, dependency_hashes)
+
+        publisher_record = query_one(
+            g.db, "SELECT signing_public_key FROM publishers WHERE id = ?",
+            [g.current_publisher["id"]],
+        )
+        signing_public_key = (
+            publisher_record["signing_public_key"] if publisher_record else ""
+        ) or ""
+        attestation_json = None
+        if signing_public_key and not dry_run and not attestation_data:
+            return error_response(
+                "ATTESTATION_REQUIRED",
+                "This publisher has a signing key; releases must be signed",
+                400,
+            )
+        if attestation_data:
+            try:
+                from ..attestation import Attestation, verify_attestation
+                attestation = Attestation.from_dict(attestation_data)
+            except (KeyError, TypeError, ValueError):
+                return error_response(
+                    "INVALID_ATTESTATION", "Malformed release attestation", 400
+                )
+            if (
+                attestation.tool_name != name
+                or attestation.version != version
+                or attestation.content_hash != content_hash
+                or attestation.signer != owner
+                or not signing_public_key
+                or not verify_attestation(attestation, signing_public_key)
+            ):
+                return error_response(
+                    "INVALID_ATTESTATION",
+                    "Release attestation does not match this publisher and content",
+                    400,
+                )
+            attestation_json = json.dumps(attestation.to_dict(), sort_keys=True)
 
         existing = query_one(
             g.db,
-            "SELECT published_at, config_hash, moderation_status, visibility FROM tools WHERE owner = ? AND name = ? AND version = ?",
+            "SELECT published_at, config_hash, content_hash, moderation_status, visibility FROM tools WHERE owner = ? AND name = ? AND version = ?",
             [owner, name, version],
         )
         if existing:
@@ -2653,6 +2905,7 @@ def create_app() -> Flask:
                         "name": name,
                         "version": version,
                         "config_hash": config_hash,
+                        "content_hash": existing.get("content_hash") or content_hash,
                         "pr_url": "",
                         "status": existing["moderation_status"],
                         "visibility": existing["visibility"],
@@ -2667,6 +2920,43 @@ def create_app() -> Flask:
                 details={"published_at": existing["published_at"]},
             )
 
+        applied_improvement = None
+        if improvement_id is not None:
+            try:
+                improvement_id = int(improvement_id)
+            except (TypeError, ValueError):
+                return error_response("VALIDATION_ERROR", "Invalid improvement_id", 400)
+            applied_improvement = query_one(
+                g.db,
+                """SELECT s.*, t.owner, t.name FROM improvement_submissions s
+                   JOIN tools t ON t.id = s.tool_id
+                   WHERE s.id = ? AND s.status = 'approved'
+                     AND s.applied_tool_id IS NULL""",
+                [improvement_id],
+            )
+            if (
+                not applied_improvement
+                or applied_improvement["owner"] != owner
+                or applied_improvement["name"] != name
+                or applied_improvement["step_index"] >= len(published_tool.steps)
+            ):
+                return error_response(
+                    "VALIDATION_ERROR", "Improvement cannot be applied to this release", 400
+                )
+            candidate_step = published_tool.steps[applied_improvement["step_index"]]
+            candidate_value = (
+                candidate_step.prompt if applied_improvement["step_type"] == "prompt"
+                and hasattr(candidate_step, "prompt") else
+                candidate_step.code if applied_improvement["step_type"] == "code"
+                and hasattr(candidate_step, "code") else None
+            )
+            if candidate_value != applied_improvement["proposed"]:
+                return error_response(
+                    "VALIDATION_ERROR",
+                    "Published config does not contain the approved improvement",
+                    400,
+                )
+
         suggestions = {"category": None, "similar_tools": []}
         try:
             from .categorize import suggest_categories
@@ -2714,6 +3004,10 @@ def create_app() -> Flask:
         # Check scrutiny decision
         if scrutiny_report:
             suggestions["scrutiny"] = scrutiny_report
+            from ..improvement import generate_improvements
+            suggestions["improvements"] = generate_improvements(
+                published_tool, scrutiny_report
+            ).to_dict()
             if scrutiny_report.get("decision") == "reject":
                 # Find the failing check for error message
                 fail_findings = [f for f in scrutiny_report.get("findings", []) if f.get("result") == "fail"]
@@ -2749,6 +3043,7 @@ def create_app() -> Flask:
                     "suggestions": suggestions,
                     "preflight": preflight_report.to_dict(),
                     "quality": quality_report,
+                    "content_hash": content_hash,
                 }
             })
 
@@ -2780,9 +3075,10 @@ def create_app() -> Flask:
                 owner, name, version, description, category, tags, config_yaml, readme,
                 defaults, publisher_id, deprecated, deprecated_message, replacement, downloads,
                 scrutiny_status, scrutiny_report, source, source_url, source_json,
-                config_hash, visibility, moderation_status, forked_from, forked_version,
+                config_hash, content_hash, dependency_hashes_json, visibility,
-                published_at
+                moderation_status, forked_from, forked_version, published_at,
-            ) VALUES (?, ?, ?, ?, ?, ?, ?, ?, ?, ?, ?, ?, ?, ?, ?, ?, ?, ?, ?, ?, ?, ?, ?, ?, ?)
+                attestation_json
+            ) VALUES (?, ?, ?, ?, ?, ?, ?, ?, ?, ?, ?, ?, ?, ?, ?, ?, ?, ?, ?, ?, ?, ?, ?, ?, ?, ?, ?, ?)
             """,
             [
                 owner,
@@ -2805,14 +3101,28 @@ def create_app() -> Flask:
                 source_url,
                 source_json,
                 config_hash,
+                content_hash,
+                json.dumps(dependency_hashes, sort_keys=True),
                 visibility,
                 moderation_status,
                 forked_from,
                 forked_version,
                 datetime.now(timezone.utc).isoformat(),
+                attestation_json,
             ],
         )
         tool_id = insert_cursor.lastrowid
+        if applied_improvement:
+            g.db.execute(
+                "UPDATE improvement_submissions SET applied_tool_id = ? WHERE id = ?",
+                [tool_id, applied_improvement["id"]],
+            )
+            g.db.execute(
+                """INSERT OR IGNORE INTO tool_contributors
+                   (tool_id, publisher_id, contribution)
+                   VALUES (?, ?, 'improvement')""",
+                [tool_id, applied_improvement["submitter_id"]],
+            )
         audit = preflight_report.audit_evidence or {}
         g.db.execute(
             """
@@ -2866,6 +3176,7 @@ def create_app() -> Flask:
                 "name": name,
                 "version": version,
                 "config_hash": config_hash,
+                "content_hash": content_hash,
                 "pr_url": "",
                 "status": moderation_status,
                 "visibility": visibility,

src/cmdforge/registry/auditing.py

@@ -26,22 +26,32 @@ def registry_dependency_findings(conn, tool: Tool, owner: str) -> Dict[str, List
     missing: List[str] = []
     deprecated: List[str] = []
     for reference in sorted(references):
-        if reference == tool.name or reference == f"{owner}/{tool.name}":
+        from ..resolver import ToolSpec
+        from ..semver import matches_constraint
+        spec = ToolSpec.parse(reference)
+        if spec.name == tool.name and (spec.owner in (None, owner)):
             continue
-        if "/" in reference:
+        if spec.owner:
-            dep_owner, dep_name = reference.split("/", 1)
+            candidates = [(spec.owner, spec.name)]
-            candidates = [(dep_owner, dep_name)]
         else:
-            candidates = [(owner, reference), ("official", reference)]
+            candidates = [(owner, spec.name), ("official", spec.name)]
         row = None
         for dep_owner, dep_name in candidates:
-            row = conn.execute(
+            rows = conn.execute(
                 """
-                SELECT deprecated, replacement FROM tools
+                SELECT version, deprecated, replacement FROM tools
-                WHERE owner = ? AND name = ? ORDER BY id DESC LIMIT 1
+                WHERE owner = ? AND name = ? ORDER BY id DESC
                 """,
                 [dep_owner, dep_name],
-            ).fetchone()
+            ).fetchall()
+            row = next(
+                (
+                    candidate for candidate in rows
+                    if not spec.version
+                    or matches_constraint(candidate["version"], spec.version)
+                ),
+                None,
+            )
             if row is not None:
                 break
         if row is None:

src/cmdforge/registry/db.py

@@ -27,6 +27,7 @@ CREATE TABLE IF NOT EXISTS publishers (
     ban_reason TEXT,
     created_at TIMESTAMP DEFAULT CURRENT_TIMESTAMP,
     updated_at TIMESTAMP DEFAULT CURRENT_TIMESTAMP
+    ,signing_public_key TEXT
 );
 
 CREATE TABLE IF NOT EXISTS api_tokens (
@@ -61,6 +62,8 @@ CREATE TABLE IF NOT EXISTS tools (
     source_url TEXT,
     source_json TEXT,
     config_hash TEXT,
+    content_hash TEXT,
+    dependency_hashes_json TEXT,
     visibility TEXT DEFAULT 'public',
     moderation_status TEXT DEFAULT 'pending',
     moderation_note TEXT,
@@ -69,6 +72,7 @@ CREATE TABLE IF NOT EXISTS tools (
     forked_from TEXT,
     forked_version TEXT,
     published_at TIMESTAMP DEFAULT CURRENT_TIMESTAMP,
+    attestation_json TEXT,
     UNIQUE(owner, name, version)
 );
 
@@ -94,6 +98,31 @@ CREATE TABLE IF NOT EXISTS tool_audits (
 CREATE INDEX IF NOT EXISTS idx_tool_audits_tool_time
 ON tool_audits(tool_id, evaluated_at DESC, id DESC);
 
+CREATE TABLE IF NOT EXISTS improvement_submissions (
+    id INTEGER PRIMARY KEY AUTOINCREMENT,
+    tool_id INTEGER NOT NULL REFERENCES tools(id) ON DELETE CASCADE,
+    submitter_id INTEGER NOT NULL REFERENCES publishers(id),
+    step_index INTEGER NOT NULL,
+    step_type TEXT NOT NULL,
+    original TEXT NOT NULL,
+    proposed TEXT NOT NULL,
+    rationale TEXT,
+    status TEXT NOT NULL DEFAULT 'tested',
+    test_result_json TEXT NOT NULL,
+    reviewer_id INTEGER REFERENCES publishers(id),
+    review_notes TEXT,
+    submitted_at TIMESTAMP NOT NULL,
+    reviewed_at TIMESTAMP
+    ,applied_tool_id INTEGER REFERENCES tools(id)
+);
+
+CREATE TABLE IF NOT EXISTS tool_contributors (
+    tool_id INTEGER NOT NULL REFERENCES tools(id) ON DELETE CASCADE,
+    publisher_id INTEGER NOT NULL REFERENCES publishers(id),
+    contribution TEXT NOT NULL,
+    PRIMARY KEY (tool_id, publisher_id, contribution)
+);
+
 CREATE VIRTUAL TABLE IF NOT EXISTS tools_fts USING fts5(
     name, description, tags, readme,
     content='tools',
@@ -489,6 +518,8 @@ def migrate_db(conn: sqlite3.Connection) -> None:
         ("source_url", "TEXT", "NULL"),
         ("source_json", "TEXT", "NULL"),
         ("config_hash", "TEXT", "NULL"),
+        ("content_hash", "TEXT", "NULL"),
+        ("dependency_hashes_json", "TEXT", "NULL"),
         ("visibility", "TEXT", "'public'"),
         ("moderation_status", "TEXT", "'pending'"),
         ("moderation_note", "TEXT", "NULL"),
@@ -497,6 +528,7 @@ def migrate_db(conn: sqlite3.Connection) -> None:
         ("forked_from", "TEXT", "NULL"),
         ("forked_version", "TEXT", "NULL"),
         ("defaults", "TEXT", "NULL"),
+        ("attestation_json", "TEXT", "NULL"),
     ]
 
     for col_name, col_type, default in tools_migrations:
@@ -514,6 +546,7 @@ def migrate_db(conn: sqlite3.Connection) -> None:
         ("banned_at", "TIMESTAMP", "NULL"),
         ("banned_by", "TEXT", "NULL"),
         ("ban_reason", "TEXT", "NULL"),
+        ("signing_public_key", "TEXT", "NULL"),
     ]
 
     for col_name, col_type, default in publishers_migrations:
@@ -540,6 +573,18 @@ def migrate_db(conn: sqlite3.Connection) -> None:
             except sqlite3.OperationalError:
                 pass
 
+    cursor = conn.execute("PRAGMA table_info(improvement_submissions)")
+    improvement_cols = {row[1] for row in cursor.fetchall()}
+    if improvement_cols and "applied_tool_id" not in improvement_cols:
+        try:
+            conn.execute(
+                "ALTER TABLE improvement_submissions "
+                "ADD COLUMN applied_tool_id INTEGER REFERENCES tools(id)"
+            )
+            conn.commit()
+        except sqlite3.OperationalError:
+            pass
+
     # Grandfather existing tools: set moderation_status to 'approved' for tools that have NULL
     # This ensures existing tools remain visible after migration (one-time migration)
     # Note: Only applies to NULL, NOT to 'pending' - pending tools need manual review
@@ -573,6 +618,7 @@ def migrate_db(conn: sqlite3.Connection) -> None:
         conn.execute("CREATE INDEX IF NOT EXISTS idx_tools_owner ON tools(owner)")
         conn.execute("CREATE INDEX IF NOT EXISTS idx_tools_moderation ON tools(moderation_status, visibility)")
         conn.execute("CREATE INDEX IF NOT EXISTS idx_tools_hash ON tools(config_hash)")
+        conn.execute("CREATE INDEX IF NOT EXISTS idx_tools_content_hash ON tools(content_hash)")
         conn.execute("CREATE INDEX IF NOT EXISTS idx_publishers_role ON publishers(role)")
         conn.execute("CREATE INDEX IF NOT EXISTS idx_publishers_banned ON publishers(banned)")
         conn.execute("CREATE INDEX IF NOT EXISTS idx_audit_log_target ON audit_log(target_type, target_id)")

src/cmdforge/registry_client.py

@@ -116,6 +116,10 @@ class DownloadResult:
     readme: str = ""
     config_hash: str = ""  # Registry hash for integrity verification
     defaults: str = ""  # Default settings YAML content
+    attestation: Optional[Dict[str, Any]] = None
+    signing_public_key: str = ""
+    content_hash: str = ""
+    dependency_hashes: Dict[str, str] = field(default_factory=dict)
 
 
 class RegistryClient:
@@ -562,8 +566,56 @@ class RegistryClient:
             readme=data.get("readme", ""),
             config_hash=data.get("config_hash", ""),
             defaults=data.get("defaults", "")
+            ,attestation=data.get("attestation")
+            ,signing_public_key=data.get("signing_public_key", "")
+            ,content_hash=data.get("content_hash", "")
+            ,dependency_hashes=data.get("dependency_hashes", {})
         )
 
+    def get_tool_by_content_hash(self, content_hash: str) -> ToolInfo:
+        response = self._request("GET", f"/tools/by-content-hash/{content_hash}")
+        if response.status_code != 200:
+            self._handle_error_response(response)
+        return ToolInfo.from_dict(response.json().get("data", {}))
+
+    def set_signing_public_key(self, public_key: str) -> Dict[str, Any]:
+        """Register the public key used to verify this publisher's releases."""
+        response = self._request(
+            "PUT", "/me/signing-key",
+            json_data={"public_key": public_key}, require_auth=True,
+        )
+        if response.status_code != 200:
+            self._handle_error_response(response)
+        return response.json().get("data", {})
+
+    def submit_improvement(
+        self, owner: str, name: str, version: str, step_index: int,
+        proposed: str, rationale: str = "",
+    ) -> Dict[str, Any]:
+        response = self._request(
+            "POST", f"/tools/{owner}/{name}/{version}/improvements",
+            json_data={
+                "step_index": step_index, "proposed": proposed,
+                "rationale": rationale,
+            },
+            require_auth=True,
+        )
+        if response.status_code != 201:
+            self._handle_error_response(response)
+        return response.json().get("data", {})
+
+    def review_improvement(
+        self, submission_id: int, decision: str, notes: str = ""
+    ) -> Dict[str, Any]:
+        response = self._request(
+            "PATCH", f"/improvements/{submission_id}",
+            json_data={"decision": decision, "notes": notes},
+            require_auth=True,
+        )
+        if response.status_code != 200:
+            self._handle_error_response(response)
+        return response.json().get("data", {})
+
     def get_categories(self) -> List[Dict[str, Any]]:
         """
         Get list of tool categories.
@@ -623,7 +675,9 @@ class RegistryClient:
         defaults: str = "",
         dry_run: bool = False,
         visibility: str = "public",
-        owner: str = ""
+        owner: str = "",
+        attestation: Optional[Dict[str, Any]] = None,
+        improvement_id: Optional[int] = None,
     ) -> Dict[str, Any]:
         """
         Publish a tool to the registry.
@@ -649,6 +703,10 @@ class RegistryClient:
             payload["defaults"] = defaults
         if owner:
             payload["owner"] = owner
+        if attestation:
+            payload["attestation"] = attestation
+        if improvement_id is not None:
+            payload["improvement_id"] = improvement_id
 
         response = self._request(
             "POST",

src/cmdforge/resolver.py

@@ -354,7 +354,10 @@ class ToolResolver:
                 config_yaml=result.config_yaml,
                 readme=result.readme,
                 config_hash=result.config_hash,
-                defaults=result.defaults
+                defaults=result.defaults,
+                attestation=result.attestation,
+                signing_public_key=result.signing_public_key,
+                content_hash=result.content_hash,
             )
 
             if self.verbose:
@@ -383,7 +386,10 @@ class ToolResolver:
         config_yaml: str,
         readme: str = "",
         config_hash: str = "",
-        defaults: str = ""
+        defaults: str = "",
+        attestation: Optional[dict] = None,
+        signing_public_key: str = "",
+        content_hash: str = "",
     ) -> ResolvedTool:
         """Install a tool fetched from registry to global directory."""
         # Verify hash if provided
@@ -396,6 +402,24 @@ class ToolResolver:
                     f"got {computed_hash[:20]}... - content may have been tampered with"
                 )
 
+        if attestation:
+            from .attestation import Attestation, verify_attestation
+            try:
+                signed = Attestation.from_dict(attestation)
+            except (KeyError, TypeError, ValueError) as exc:
+                raise RuntimeError("Malformed release attestation") from exc
+            if (
+                not signing_public_key
+                or signed.tool_name != name
+                or signed.version != version
+                or signed.content_hash != content_hash
+                or signed.signer != owner
+                or not verify_attestation(signed, signing_public_key)
+            ):
+                raise RuntimeError(
+                    f"Invalid release attestation for {owner}/{name}@{version}"
+                )
+
         # Create directory structure
         tool_dir = TOOLS_DIR / owner / name
         tool_dir.mkdir(parents=True, exist_ok=True)
@@ -628,7 +652,10 @@ def install_from_registry(spec: str, version: Optional[str] = None) -> ResolvedT
         config_yaml=result.config_yaml,
         readme=result.readme,
         config_hash=result.config_hash,
-        defaults=result.defaults
+        defaults=result.defaults,
+        attestation=result.attestation,
+        signing_public_key=result.signing_public_key,
+        content_hash=result.content_hash,
     )
 
 

src/cmdforge/runner.py

@@ -1234,6 +1234,14 @@ def main():
         else:
             print(output)
 
+    if exit_code == 0 and not args.dry_run:
+        try:
+            from .usage import record_invocation
+            record_invocation(resolved.full_name)
+        except (OSError, ValueError):
+            # Usage discovery is optional and must never break tool execution.
+            pass
+
     sys.exit(exit_code)
 
 

src/cmdforge/signing.py

@@ -0,0 +1,43 @@
+"""Local publisher signing-key storage."""
+
+from __future__ import annotations
+
+import json
+import os
+from pathlib import Path
+from typing import Optional, Tuple
+
+from .attestation import generate_keypair
+from .config import CONFIG_DIR
+
+SIGNING_KEY_FILE = CONFIG_DIR / "release-signing-key.json"
+
+
+def load_signing_key(path: Path = SIGNING_KEY_FILE) -> Optional[Tuple[str, str]]:
+    if not path.exists():
+        return None
+    data = json.loads(path.read_text(encoding="utf-8"))
+    private_key = data.get("private_key", "")
+    public_key = data.get("public_key", "")
+    if not private_key or not public_key:
+        raise ValueError("Signing key file is incomplete")
+    return private_key, public_key
+
+
+def initialize_signing_key(
+    path: Path = SIGNING_KEY_FILE, *, overwrite: bool = False
+) -> Tuple[str, str]:
+    if path.exists() and not overwrite:
+        existing = load_signing_key(path)
+        if existing is None:
+            raise ValueError("Signing key file could not be loaded")
+        return existing
+    private_key, public_key = generate_keypair()
+    path.parent.mkdir(parents=True, exist_ok=True, mode=0o700)
+    path.write_text(json.dumps({
+        "algorithm": "ed25519",
+        "private_key": private_key,
+        "public_key": public_key,
+    }, indent=2) + "\n", encoding="utf-8")
+    os.chmod(path, 0o600)
+    return private_key, public_key

src/cmdforge/usage.py

@@ -0,0 +1,202 @@
+"""Opt-in, local-only pipeline usage discovery.
+
+Only tool names, anonymous pipe inode numbers, counts, and timestamps are
+stored. Inputs, outputs, arguments, environment, and working directories are
+never recorded.
+"""
+
+from __future__ import annotations
+
+import json
+import os
+import stat
+import tempfile
+from contextlib import contextmanager
+from datetime import datetime, timezone
+from pathlib import Path
+from typing import Dict, Iterator, List, Optional
+
+from .config import CONFIG_DIR
+
+USAGE_FILE = CONFIG_DIR / "usage.json"
+USAGE_LOCK_FILE = CONFIG_DIR / ".usage.lock"
+USAGE_VERSION = 1
+EVENT_TTL_SECONDS = 120
+MAX_EVENTS = 200
+DEFAULT_THRESHOLD = 3
+
+
+def _empty_document(enabled: bool = False) -> dict:
+    return {
+        "version": USAGE_VERSION,
+        "enabled": enabled,
+        "events": [],
+        "pipelines": {},
+    }
+
+
+@contextmanager
+def _locked() -> Iterator[None]:
+    CONFIG_DIR.mkdir(parents=True, exist_ok=True, mode=0o700)
+    with open(USAGE_LOCK_FILE, "a+", encoding="utf-8") as lock:
+        os.chmod(USAGE_LOCK_FILE, 0o600)
+        try:
+            import fcntl
+            fcntl.flock(lock.fileno(), fcntl.LOCK_EX)
+        except ImportError:
+            pass
+        try:
+            yield
+        finally:
+            try:
+                import fcntl
+                fcntl.flock(lock.fileno(), fcntl.LOCK_UN)
+            except ImportError:
+                pass
+
+
+def _load_unlocked() -> dict:
+    if not USAGE_FILE.exists():
+        return _empty_document()
+    try:
+        data = json.loads(USAGE_FILE.read_text(encoding="utf-8"))
+    except (OSError, json.JSONDecodeError):
+        return _empty_document()
+    if not isinstance(data, dict) or data.get("version") != USAGE_VERSION:
+        return _empty_document()
+    return data
+
+
+def _save_unlocked(data: dict) -> None:
+    CONFIG_DIR.mkdir(parents=True, exist_ok=True, mode=0o700)
+    temp_path = None
+    try:
+        with tempfile.NamedTemporaryFile(
+            "w", encoding="utf-8", dir=CONFIG_DIR,
+            prefix=".usage-", suffix=".tmp", delete=False,
+        ) as handle:
+            temp_path = Path(handle.name)
+            os.chmod(temp_path, 0o600)
+            json.dump(data, handle, indent=2, sort_keys=True)
+            handle.write("\n")
+            handle.flush()
+            os.fsync(handle.fileno())
+        os.replace(temp_path, USAGE_FILE)
+        os.chmod(USAGE_FILE, 0o600)
+    finally:
+        if temp_path and temp_path.exists():
+            temp_path.unlink()
+
+
+def is_enabled() -> bool:
+    with _locked():
+        return bool(_load_unlocked().get("enabled", False))
+
+
+def set_enabled(enabled: bool) -> None:
+    with _locked():
+        data = _load_unlocked()
+        data["enabled"] = bool(enabled)
+        if not enabled:
+            data["events"] = []
+        _save_unlocked(data)
+
+
+def clear_usage() -> None:
+    with _locked():
+        enabled = bool(_load_unlocked().get("enabled", False))
+        _save_unlocked(_empty_document(enabled=enabled))
+
+
+def _pipe_inode(stream) -> Optional[int]:
+    try:
+        info = os.fstat(stream.fileno())
+    except (AttributeError, OSError, ValueError):
+        return None
+    return info.st_ino if stat.S_ISFIFO(info.st_mode) else None
+
+
+def record_invocation(tool_name: str, stdin=None, stdout=None) -> None:
+    """Record a successful invocation when it participates in a shell pipe."""
+    import sys
+    stdin = stdin or sys.stdin
+    stdout = stdout or sys.stdout
+    input_pipe = _pipe_inode(stdin)
+    output_pipe = _pipe_inode(stdout)
+    if input_pipe is None and output_pipe is None:
+        return
+
+    now = datetime.now(timezone.utc)
+    timestamp = now.timestamp()
+    with _locked():
+        data = _load_unlocked()
+        if not data.get("enabled", False):
+            return
+        events = [
+            event for event in data.get("events", [])
+            if timestamp - float(event.get("time", 0)) <= EVENT_TTL_SECONDS
+        ]
+        discovered = set()
+        for event in events:
+            if input_pipe is not None and event.get("stdout_pipe") == input_pipe:
+                discovered.add((event["tool"], tool_name))
+            if output_pipe is not None and event.get("stdin_pipe") == output_pipe:
+                discovered.add((tool_name, event["tool"]))
+        pipelines = data.setdefault("pipelines", {})
+        for first, second in discovered:
+            if first == second:
+                continue
+            key = json.dumps([first, second], separators=(",", ":"))
+            current = pipelines.setdefault(key, {"tools": [first, second], "count": 0})
+            current["count"] = int(current.get("count", 0)) + 1
+            current["last_seen"] = now.isoformat()
+        events.append({
+            "tool": tool_name,
+            "time": timestamp,
+            "stdin_pipe": input_pipe,
+            "stdout_pipe": output_pipe,
+        })
+        data["events"] = events[-MAX_EVENTS:]
+        _save_unlocked(data)
+
+
+def get_suggestions(threshold: int = DEFAULT_THRESHOLD) -> List[Dict]:
+    with _locked():
+        data = _load_unlocked()
+    if not data.get("enabled", False):
+        return []
+    suggestions = [
+        {
+            "tools": list(item.get("tools", [])),
+            "count": int(item.get("count", 0)),
+            "last_seen": item.get("last_seen", ""),
+        }
+        for item in data.get("pipelines", {}).values()
+        if int(item.get("count", 0)) >= threshold
+        and len(item.get("tools", [])) >= 2
+    ]
+    return sorted(suggestions, key=lambda item: (-item["count"], item["tools"]))
+
+
+def build_composite_tool(name: str, tools: List[str]):
+    """Build, but do not save, a Tool that composes the suggested pipeline."""
+    from .tool import Tool, ToolStep
+    if len(tools) < 2:
+        raise ValueError("A composite requires at least two tools")
+    steps = []
+    previous = "input"
+    for index, tool_name in enumerate(tools, start=1):
+        output_var = f"pipeline_{index}"
+        steps.append(ToolStep(
+            tool=tool_name,
+            input_template="{input}" if previous == "input" else f"{{{previous}}}",
+            output_var=output_var,
+        ))
+        previous = output_var
+    return Tool(
+        name=name,
+        description="Composite pipeline: " + " | ".join(tools),
+        dependencies=list(dict.fromkeys(tools)),
+        steps=steps,
+        output=f"{{{previous}}}",
+    )

tests/test_attestation.py

@@ -2,44 +2,70 @@
 
 from cmdforge.attestation import (
     Attestation,
+    generate_keypair,
     sign_tool,
     verify_attestation,
     verify_content_hash,
+    verify_trusted_attestation,
 )
 from cmdforge.tool import Tool
 
 
 class TestSignTool:
     def test_creates_attestation(self):
-        att = sign_tool("mytool", "1.0.0", "abc123", "alice", "secret-key")
+        private, public = generate_keypair()
+        att = sign_tool("mytool", "1.0.0", "abc123", "alice", private)
         assert att.tool_name == "mytool"
         assert att.version == "1.0.0"
         assert att.content_hash == "abc123"
         assert att.signer == "alice"
-        assert len(att.signature) == 64  # SHA256 hex
+        assert len(att.signature) == 88
-        assert att.algorithm == "hmac-sha256"
+        assert att.algorithm == "ed25519"
+        assert verify_attestation(att, public)
 
     def test_different_keys_different_signatures(self):
-        att1 = sign_tool("tool", "1.0.0", "hash", "alice", "key1")
+        key1, _ = generate_keypair()
-        att2 = sign_tool("tool", "1.0.0", "hash", "alice", "key2")
+        key2, _ = generate_keypair()
+        att1 = sign_tool("tool", "1.0.0", "hash", "alice", key1)
+        att2 = sign_tool("tool", "1.0.0", "hash", "alice", key2)
         assert att1.signature != att2.signature
 
 
 class TestVerifyAttestation:
     def test_valid_signature(self):
-        att = sign_tool("mytool", "1.0.0", "abc123", "alice", "secret-key")
+        private, public = generate_keypair()
-        assert verify_attestation(att, "secret-key")
+        att = sign_tool("mytool", "1.0.0", "abc123", "alice", private)
+        assert verify_attestation(att, public)
 
     def test_wrong_key_fails(self):
-        att = sign_tool("mytool", "1.0.0", "abc123", "alice", "secret-key")
+        private, _ = generate_keypair()
-        assert not verify_attestation(att, "wrong-key")
+        _, wrong_public = generate_keypair()
+        att = sign_tool("mytool", "1.0.0", "abc123", "alice", private)
+        assert not verify_attestation(att, wrong_public)
 
     def test_tampered_content_fails(self):
-        att = sign_tool("mytool", "1.0.0", "abc123", "alice", "secret-key")
+        from dataclasses import replace
-        att.content_hash = "tampered"
+        private, public = generate_keypair()
-        assert not verify_attestation(att, "secret-key")
+        att = sign_tool("mytool", "1.0.0", "abc123", "alice", private)
+        att = replace(att, content_hash="tampered")
+        assert not verify_attestation(att, public)
 
     def test_tampered_signer_fails(self):
-        att = sign_tool("mytool", "1.0.0", "abc123", "alice", "secret-key")
+        from dataclasses import replace
-        att.signer = "eve"
+        private, public = generate_keypair()
-        assert not verify_attestation(att, "secret-key")
+        att = sign_tool("mytool", "1.0.0", "abc123", "alice", private)
+        att = replace(att, signer="eve")
+        assert not verify_attestation(att, public)
+
+    def test_timestamp_and_algorithm_are_signed(self):
+        from dataclasses import replace
+        private, public = generate_keypair()
+        att = sign_tool("tool", "1.0.0", "hash", "alice", private)
+        assert not verify_attestation(replace(att, signed_at="tomorrow"), public)
+        assert not verify_attestation(replace(att, algorithm="hmac-sha256"), public)
+
+    def test_requires_trusted_publisher_mapping(self):
+        private, public = generate_keypair()
+        att = sign_tool("tool", "1.0.0", "hash", "alice", private)
+        assert verify_trusted_attestation(att, {"alice": public})
+        assert not verify_trusted_attestation(att, {"mallory": public})

tests/test_cli.py

@@ -650,3 +650,76 @@ def test_switch_to_existing_tool_closes_creation_page_first():
         call.close_tool_builder(),
         call.open_tool_builder("existing"),
     ]
+
+
+def test_guided_extraction_builds_new_tool_without_mutating_draft():
+    pytest.importorskip("PySide6")
+    from cmdforge.gui.pages.tool_builder_page import (
+        build_extracted_tool,
+        extraction_diff,
+    )
+
+    schema = {"type": "string"}
+    repeated = [
+        PromptStep(
+            prompt="First {input}", provider="mock", output_var="first",
+            output_schema=schema,
+        ),
+        PromptStep(
+            prompt="Second {first}", provider="mock", output_var="second",
+            output_schema=schema,
+        ),
+    ]
+    draft = Tool(
+        name="draft", steps=repeated + [
+            PromptStep(
+                prompt="First {input}", provider="mock", output_var="first_copy",
+                output_schema=schema,
+            ),
+            PromptStep(
+                prompt="Second {first}", provider="mock", output_var="second_copy",
+                output_schema=schema,
+            ),
+        ], input_schema=schema,
+    )
+    before = draft.to_dict()
+
+    extracted = build_extracted_tool(
+        draft,
+        {"type": "repeated_sequence", "locations": [1, 3], "length": 2},
+        "draft-shared",
+    )
+
+    assert len(extracted.steps) == 2
+    assert extracted.output == "{second}"
+    assert extracted.output_schema == schema
+    assert draft.to_dict() == before
+    assert "+++ draft-shared/config.yaml" in extraction_diff(extracted)
+
+
+def test_guided_extraction_rejects_hidden_external_step_dependency():
+    pytest.importorskip("PySide6")
+    from cmdforge.gui.pages.tool_builder_page import build_extracted_tool
+
+    schema = {"type": "string"}
+    draft = Tool(name="draft", steps=[
+        PromptStep(
+            prompt="Produce", provider="mock", output_var="earlier",
+            output_schema=schema,
+        ),
+        PromptStep(
+            prompt="Use {earlier}", provider="mock", output_var="one",
+            output_schema=schema,
+        ),
+        PromptStep(
+            prompt="Finish {one}", provider="mock", output_var="two",
+            output_schema=schema,
+        ),
+    ])
+
+    with pytest.raises(ValueError, match="outside its boundary"):
+        build_extracted_tool(
+            draft,
+            {"type": "repeated_sequence", "locations": [2, 4], "length": 2},
+            "draft-shared",
+        )

tests/test_collection_api.py

@@ -339,6 +339,146 @@ class TestPublishPreflightEndpoint:
         assert response.status_code == 200
 
 
+@flask_required
+class TestM9RegistryTrustAndCommunity:
+    def test_signed_publish_download_and_content_lookup(
+        self, client, auth_headers
+    ):
+        from cmdforge.attestation import (
+            Attestation, generate_keypair, sign_tool, verify_attestation,
+        )
+        from cmdforge.registry.db import connect_db
+
+        private_key, public_key = generate_keypair()
+        response = client.put(
+            "/api/v1/me/signing-key", headers=auth_headers,
+            json={"public_key": public_key},
+        )
+        assert response.status_code == 200
+
+        config = "name: signed-tool\nversion: 1.0.0\noutput: stable\n"
+        preflight = client.post(
+            "/api/v1/tools", headers=auth_headers,
+            json={"config": config, "dry_run": True},
+        )
+        assert preflight.status_code == 200
+        content_hash = preflight.get_json()["data"]["content_hash"]
+        attestation = sign_tool(
+            "signed-tool", "1.0.0", content_hash, "testuser", private_key
+        )
+        published = client.post(
+            "/api/v1/tools", headers=auth_headers,
+            json={"config": config, "attestation": attestation.to_dict()},
+        )
+        assert published.status_code == 201
+        assert published.get_json()["data"]["content_hash"] == content_hash
+
+        conn = connect_db()
+        try:
+            conn.execute(
+                "UPDATE tools SET moderation_status = 'approved' "
+                "WHERE owner = 'testuser' AND name = 'signed-tool'"
+            )
+            conn.commit()
+        finally:
+            conn.close()
+
+        downloaded = client.get(
+            "/api/v1/tools/testuser/signed-tool/download?install=false"
+        )
+        assert downloaded.status_code == 200
+        data = downloaded.get_json()["data"]
+        assert data["content_hash"] == content_hash
+        assert verify_attestation(Attestation.from_dict(data["attestation"]), public_key)
+        lookup = client.get(f"/api/v1/tools/by-content-hash/{content_hash}")
+        assert lookup.status_code == 200
+        assert lookup.get_json()["data"]["name"] == "signed-tool"
+
+    def test_registered_key_requires_valid_signature(self, client, auth_headers):
+        from cmdforge.attestation import generate_keypair
+        _, public_key = generate_keypair()
+        assert client.put(
+            "/api/v1/me/signing-key", headers=auth_headers,
+            json={"public_key": public_key},
+        ).status_code == 200
+        response = client.post(
+            "/api/v1/tools", headers=auth_headers,
+            json={"config": "name: unsigned-tool\nversion: 1.0.0\n"},
+        )
+        assert response.status_code == 400
+        assert response.get_json()["error"]["code"] == "ATTESTATION_REQUIRED"
+
+    def test_improvement_is_tested_reviewed_and_credited(
+        self, client, auth_headers
+    ):
+        from cmdforge.registry.db import connect_db
+        config = (
+            "name: improvable\nversion: 1.0.0\noutput: '{result}'\n"
+            "input_schema:\n  type: string\noutput_schema:\n  type: string\n"
+            "steps:\n  - type: prompt\n    prompt: 'Summarize: {input}'\n"
+            "    provider: mock\n    output_var: result\n"
+        )
+        published = client.post(
+            "/api/v1/tools", headers=auth_headers, json={"config": config}
+        )
+        assert published.status_code == 201
+        conn = connect_db()
+        try:
+            conn.execute(
+                "UPDATE tools SET moderation_status = 'approved' "
+                "WHERE owner = 'testuser' AND name = 'improvable'"
+            )
+            conn.commit()
+        finally:
+            conn.close()
+
+        submitted = client.post(
+            "/api/v1/tools/testuser/improvable/1.0.0/improvements",
+            headers=auth_headers,
+            json={
+                "step_index": 0,
+                "proposed": "Summarize the input accurately and concisely: {input}",
+                "rationale": "Clearer expected behavior",
+            },
+        )
+        assert submitted.status_code == 201
+        submission_id = submitted.get_json()["data"]["id"]
+        reviewed = client.patch(
+            f"/api/v1/improvements/{submission_id}", headers=auth_headers,
+            json={"decision": "approve", "notes": "Validated"},
+        )
+        assert reviewed.status_code == 200
+        assert reviewed.get_json()["data"]["ready_to_apply"] is True
+        improved_config = config.replace("version: 1.0.0", "version: 1.0.1").replace(
+            "Summarize: {input}",
+            "Summarize the input accurately and concisely: {input}",
+        )
+        applied = client.post(
+            "/api/v1/tools", headers=auth_headers,
+            json={"config": improved_config, "improvement_id": submission_id},
+        )
+        assert applied.status_code == 201
+        conn = connect_db()
+        try:
+            conn.execute(
+                "UPDATE tools SET moderation_status = 'approved' "
+                "WHERE owner = 'testuser' AND name = 'improvable' "
+                "AND version = '1.0.1'"
+            )
+            conn.commit()
+            credited = conn.execute(
+                "SELECT COUNT(*) AS count FROM tool_contributors"
+            ).fetchone()["count"]
+        finally:
+            conn.close()
+        assert credited == 1
+        detail = client.get("/api/v1/tools/testuser/improvable?version=1.0.1")
+        assert detail.status_code == 200
+        assert detail.get_json()["data"]["badges"] == [
+            "optimized", "community-reviewed"
+        ]
+
+
 @flask_required
 class TestPostCollectionsEndpoint:
     """Tests for POST /api/v1/collections endpoint."""

tests/test_community.py

@@ -49,11 +49,11 @@ class TestReviewSubmission:
         sub = ImprovementSubmission(
             tool_name="test", tool_version="1.0.0", submitter="alice",
             step_index=0, step_type="prompt", original="a", proposed="b",
-            status="tested",
+            status="tested", test_result={"passed_for_review": True},
         )
         review = review_submission(sub, "approve", "admin", "Good improvement")
         assert review.decision == "approve"
-        assert sub.status == "approve"
+        assert sub.status == "approved"
 
     def test_review_pending_rejected(self):
         sub = ImprovementSubmission(

tests/test_integrity.py

@@ -44,12 +44,14 @@ class TestIntegrityChain:
         )
         chain = build_integrity_chain(tool)
         assert "missing-dep" not in chain.nodes
-        # Chain is still valid because the dependency hash is "unresolved"
+        assert not chain.is_valid
-        # and there's no node to compare against
+        assert chain.errors == ["Unresolved dependency: missing-dep"]
-        assert chain.is_valid
 
     def test_chain_valid_with_resolved_dep(self, tmp_path):
-        with patch("cmdforge.tool.TOOLS_DIR", tmp_path / ".cmdforge"):
+        with (
+            patch("cmdforge.tool.TOOLS_DIR", tmp_path / ".cmdforge"),
+            patch("cmdforge.tool.BIN_DIR", tmp_path / "bin"),
+        ):
             from cmdforge.tool import save_tool
 
             child = Tool(name="child", version="1.0.0")
@@ -67,13 +69,35 @@ class TestIntegrityChain:
             assert "child" in chain.nodes
             assert chain.is_valid
 
+    def test_child_change_changes_root_identity(self, tmp_path):
+        with (
+            patch("cmdforge.tool.TOOLS_DIR", tmp_path / ".cmdforge"),
+            patch("cmdforge.tool.BIN_DIR", tmp_path / "bin"),
+        ):
+            from cmdforge.tool import save_tool
+            child = Tool(name="child", version="1.0.0", output="one")
+            save_tool(child)
+            parent = Tool(
+                name="parent", version="1.0.0",
+                steps=[ToolStep(tool="child", output_var="x")], output="{x}",
+            )
+            first = build_integrity_chain(parent).root.content_hash
+            child.output = "two"
+            save_tool(child)
+            second = build_integrity_chain(parent).root.content_hash
+            assert first != second
+
 
 class TestVerifyIntegrity:
     def test_valid_chain(self):
-        chain = IntegrityChain(
+        from cmdforge.integrity import compute_content_identity
-            root=IntegrityNode(name="root", content_hash="abc"),
+        definition_hash = "sha256:" + "a" * 64
-            nodes={"root": IntegrityNode(name="root", content_hash="abc")},
+        content_hash = compute_content_identity(definition_hash, {})
+        node = IntegrityNode(
+            name="root", definition_hash=definition_hash,
+            content_hash=content_hash,
         )
+        chain = IntegrityChain(root=node, nodes={"root": node})
         assert verify_integrity(chain)
 
     def test_tampered_chain(self):

tests/test_lockfile.py

@@ -259,6 +259,12 @@ class TestGenerateLockfile:
         assert "official/text-utils" in lock.packages
         assert lock.packages["official/summarize"].direct is True
         assert lock.packages["official/text-utils"].direct is False
+        child_identity = lock.packages["official/text-utils"].content_hash
+        assert child_identity.startswith("sha256:")
+        assert lock.packages["official/summarize"].dependency_hashes == {
+            "official/text-utils": child_identity
+        }
+        assert lock.packages["official/summarize"].content_hash.startswith("sha256:")
 
 
 class TestVerifyLockfile:
@@ -319,7 +325,9 @@ class TestLockfileRoundTrip:
                     integrity="sha256:abc123",
                     source="registry",
                     direct=True,
-                    required_by=[]
+                    required_by=[],
+                    content_hash="sha256:root",
+                    dependency_hashes={"official/text-utils": "sha256:child"},
                 ),
                 "official/text-utils": LockedPackage(
                     name="official/text-utils",
@@ -364,6 +372,10 @@ class TestLockfileRoundTrip:
         assert summarize.integrity == "sha256:abc123"
         assert summarize.source == "registry"
         assert summarize.direct is True
+        assert summarize.content_hash == "sha256:root"
+        assert summarize.dependency_hashes == {
+            "official/text-utils": "sha256:child"
+        }
 
         text_utils = loaded.packages["official/text-utils"]
         assert text_utils.direct is False

tests/test_mcp.py

@@ -12,6 +12,7 @@ from cmdforge.mcp_client import (
     McpServerConfig,
     McpClientManager,
     _build_server_env,
+    _build_http_headers,
     _normalize_result,
     _sanitize,
     _fingerprint,
@@ -34,14 +35,15 @@ class TestMcpServerConfig:
     @pytest.mark.parametrize(
         "change, message",
         [
-            ({"transport": "streamable-http"}, "Unsupported MCP transport"),
+            ({"transport": "streamable-http"}, "requires a url"),
             ({"args": "-y"}, "list of strings"),
             ({"timeout": 0}, "greater than 0"),
             ({"timeout": True}, "must be a number"),
         ],
     )
     def test_validation_rejects_invalid_config(self, change, message):
-        cfg = McpServerConfig(name="test", command="server", **change)
+        command = None if change.get("transport") == "streamable-http" else "server"
+        cfg = McpServerConfig(name="test", command=command, **change)
         with pytest.raises(ValueError, match=message):
             cfg.validate()
 
@@ -55,6 +57,31 @@ class TestMcpServerConfig:
         b = McpServerConfig(name="b", command="cmd", args=["-v"])
         assert _fingerprint(a) == _fingerprint(b)
 
+    def test_streamable_http_accepts_https(self):
+        cfg = McpServerConfig(
+            name="remote", transport="streamable-http",
+            url="https://example.com/mcp", approved=True,
+        )
+        cfg.validate()
+
+    def test_streamable_http_allows_loopback_http_only(self):
+        McpServerConfig(
+            name="local", transport="streamable-http",
+            url="http://127.0.0.1:8000/mcp",
+        ).validate()
+        with pytest.raises(ValueError, match="must use HTTPS"):
+            McpServerConfig(
+                name="remote", transport="streamable-http",
+                url="http://example.com/mcp",
+            ).validate()
+
+    def test_streamable_http_rejects_url_credentials(self):
+        with pytest.raises(ValueError, match="must not contain credentials"):
+            McpServerConfig(
+                name="remote", transport="streamable-http",
+                url="https://user:pass@example.com/mcp",
+            ).validate()
+
 
 class TestSanitize:
     def test_redacts_bearer(self):
@@ -85,6 +112,15 @@ class TestEnvironmentIsolation:
         )
         assert _build_server_env(cfg)["SERVER_TOKEN"] == "allowed"
 
+    def test_http_header_reference_is_resolved(self, monkeypatch):
+        monkeypatch.setenv("MCP_TOKEN", "secret")
+        cfg = McpServerConfig(
+            name="remote", transport="streamable-http",
+            url="https://example.com/mcp",
+            headers={"Authorization": "Bearer ${MCP_TOKEN}"},
+        )
+        assert _build_http_headers(cfg)["Authorization"] == "Bearer secret"
+
     def test_missing_environment_reference_fails(self, monkeypatch):
         monkeypatch.delenv("CMDFORGE_MISSING", raising=False)
         cfg = McpServerConfig(
@@ -122,6 +158,19 @@ class TestMcpConfigPersistence:
     def test_load_no_file(self, temp_mcp_file):
         assert load_mcp_config() == []
 
+    def test_streamable_http_round_trip(self, temp_mcp_file):
+        save_mcp_config([McpServerConfig(
+            name="remote", transport="streamable-http",
+            url="https://example.com/mcp",
+            headers={"Authorization": "Bearer ${MCP_TOKEN}"},
+            approved=True,
+        )])
+        loaded = load_mcp_config()[0]
+        assert loaded.transport == "streamable-http"
+        assert loaded.command is None
+        assert loaded.url == "https://example.com/mcp"
+        assert loaded.headers == {"Authorization": "Bearer ${MCP_TOKEN}"}
+
     def test_save_creates_0600_permissions(self, temp_mcp_file):
         save_mcp_config([McpServerConfig(name="test", command="echo")])
         perms = oct(temp_mcp_file.stat().st_mode & 0o777)
@@ -651,6 +700,45 @@ class TestRegisteredMcpTool:
                 serve()
 
 
+class TestStreamableHttpServerSafety:
+    def test_defaults_are_loopback_with_origin_allowlist(self):
+        from cmdforge.mcp_server import _validate_http_server_options
+        origins, token = _validate_http_server_options(
+            "127.0.0.1", 8000, [], None, None
+        )
+        assert token is None
+        assert origins == [
+            "http://localhost:8000", "http://127.0.0.1:8000"
+        ]
+
+    def test_nonlocal_requires_https_and_auth(self):
+        from cmdforge.mcp_server import _validate_http_server_options
+        with pytest.raises(ValueError, match="external-url"):
+            _validate_http_server_options("0.0.0.0", 8000, [], "secret", None)
+        with pytest.raises(ValueError, match="auth-token"):
+            _validate_http_server_options(
+                "0.0.0.0", 8000, [], None, "https://mcp.example.com"
+            )
+
+    def test_nonlocal_infers_https_origin_and_expands_token(self, monkeypatch):
+        from cmdforge.mcp_server import _validate_http_server_options
+        monkeypatch.setenv("MCP_AUTH", "secret-value")
+        origins, token = _validate_http_server_options(
+            "0.0.0.0", 8000, [], "${MCP_AUTH}",
+            "https://mcp.example.com",
+        )
+        assert origins == ["https://mcp.example.com"]
+        assert token == "secret-value"
+
+    def test_nonlocal_rejects_insecure_origin(self):
+        from cmdforge.mcp_server import _validate_http_server_options
+        with pytest.raises(ValueError, match="origins must use HTTPS"):
+            _validate_http_server_options(
+                "0.0.0.0", 8000, ["http://example.com"], "secret",
+                "https://mcp.example.com",
+            )
+
+
 class TestCmdForgeMcpServerEndToEnd:
     def test_stdio_discovery_and_invocation(self, tmp_path, monkeypatch):
         pytest.importorskip("mcp")

tests/test_prompt_optimizer.py

@@ -23,6 +23,21 @@ class TestTruncate:
 
 
 class TestGenerateVariations:
+    def test_mock_provider_never_calls_external_provider(self):
+        with patch("cmdforge.providers.call_provider") as call:
+            _get_strategies("Summarize the text", 2, provider="mock")
+        call.assert_not_called()
+
+    def test_selected_provider_is_honored(self):
+        from cmdforge.providers import ProviderResult
+        with patch(
+            "cmdforge.providers.call_provider",
+            return_value=ProviderResult(text="A materially clearer instruction", success=True),
+        ) as call:
+            result = _get_strategies("Summarize the text", 1, provider="chosen")
+        assert result[0][1] == "A materially clearer instruction"
+        assert call.call_args.args[0] == "chosen"
+
     def test_generates_for_prompt_steps(self, tmp_path):
         from cmdforge.tool import Tool, PromptStep
 
@@ -75,6 +90,29 @@ class TestGetStrategies:
 
 
 class TestOptimizeTool:
+    def test_structural_results_do_not_claim_a_best_prompt(self):
+        from cmdforge.tool import Tool, PromptStep
+        tool = Tool(
+            name="structural", steps=[PromptStep("Summarize", "mock", "out")],
+            output="{out}", input_schema={"type": "string"},
+            output_schema={"type": "string"},
+        )
+        result = optimize_tool(tool, count=2)
+        assert result.best is None
+        assert "cannot compare prompt semantics" in result.note
+
+    def test_behavioral_evaluator_selects_only_an_improvement(self):
+        from cmdforge.tool import Tool, PromptStep
+        tool = Tool(
+            name="behavioral", steps=[PromptStep("Do it", "mock", "out")],
+            output="{out}",
+        )
+        result = optimize_tool(
+            tool, count=3,
+            evaluator=lambda candidate: len(candidate.steps[0].prompt),
+        )
+        assert result.best is not None
+        assert result.scores[result.best_index] > result.baseline_score
     def test_baseline_score_recorded(self, tmp_path):
         from cmdforge.tool import Tool, PromptStep
 

tests/test_signing.py

@@ -0,0 +1,11 @@
+import stat
+
+from cmdforge.signing import initialize_signing_key, load_signing_key
+
+
+def test_signing_key_is_persistent_and_private(tmp_path):
+    path = tmp_path / "release-key.json"
+    created = initialize_signing_key(path)
+    assert load_signing_key(path) == created
+    assert stat.S_IMODE(path.stat().st_mode) == 0o600
+    assert initialize_signing_key(path) == created

tests/test_usage.py

@@ -0,0 +1,67 @@
+import json
+import stat
+
+from cmdforge.usage import (
+    build_composite_tool,
+    clear_usage,
+    get_suggestions,
+    is_enabled,
+    record_invocation,
+    set_enabled,
+)
+
+
+class Stream:
+    def __init__(self, inode=None):
+        self.inode = inode
+
+
+def configure_paths(tmp_path, monkeypatch):
+    monkeypatch.setattr("cmdforge.usage.CONFIG_DIR", tmp_path)
+    monkeypatch.setattr("cmdforge.usage.USAGE_FILE", tmp_path / "usage.json")
+    monkeypatch.setattr("cmdforge.usage.USAGE_LOCK_FILE", tmp_path / ".usage.lock")
+
+
+def test_tracking_is_disabled_by_default(tmp_path, monkeypatch):
+    configure_paths(tmp_path, monkeypatch)
+    assert not is_enabled()
+    monkeypatch.setattr("cmdforge.usage._pipe_inode", lambda stream: stream.inode)
+    record_invocation("one", Stream(), Stream())
+    assert not (tmp_path / "usage.json").exists()
+
+
+def test_detects_only_pipe_link_and_stores_no_content(tmp_path, monkeypatch):
+    configure_paths(tmp_path, monkeypatch)
+    set_enabled(True)
+    monkeypatch.setattr("cmdforge.usage._pipe_inode", lambda stream: stream.inode)
+    for inode in range(44, 47):
+        record_invocation("tool-a", Stream(), Stream(inode))
+        record_invocation("tool-b", Stream(inode), Stream())
+
+    assert get_suggestions() == [{
+        "tools": ["tool-a", "tool-b"],
+        "count": 3,
+        "last_seen": get_suggestions()[0]["last_seen"],
+    }]
+    data = json.loads((tmp_path / "usage.json").read_text())
+    serialized = json.dumps(data)
+    for forbidden in ("input_text", "output", "arguments", "cwd", "environment"):
+        assert forbidden not in serialized
+    assert stat.S_IMODE((tmp_path / "usage.json").stat().st_mode) == 0o600
+
+
+def test_disable_and_clear_preserve_consent_state(tmp_path, monkeypatch):
+    configure_paths(tmp_path, monkeypatch)
+    set_enabled(True)
+    clear_usage()
+    assert is_enabled()
+    set_enabled(False)
+    assert not is_enabled()
+
+
+def test_build_composite_tool_preserves_pipeline_order():
+    tool = build_composite_tool("combined", ["tool-a", "tool-b"])
+    assert tool.dependencies == ["tool-a", "tool-b"]
+    assert tool.steps[0].input_template == "{input}"
+    assert tool.steps[1].input_template == "{pipeline_1}"
+    assert tool.output == "{pipeline_2}"