@julioborges/gantry 1.0.1 → 1.0.4

package/.agents/skills/gantry/scripts/execution.py

@@ -0,0 +1,797 @@
+#!/usr/bin/env python3
+"""Role execution resolution, validation, and preflight across harnesses."""
+from __future__ import annotations
+
+import argparse
+import copy
+import json
+import os
+import shutil
+import datetime
+import re
+import subprocess
+import sys
+from pathlib import Path
+from typing import Any, Callable
+
+sys.path.insert(0, str(Path(__file__).parent))
+
+from common import repo_root, resolve_policy
+import result  # noqa: E402
+import runlog  # noqa: E402
+
+SUPPORTED_HARNESSES = {"antigravity", "claude-code", "codex", "opencode"}
+
+MIN_SUPPORTED_VERSIONS = {
+    "antigravity": "1.0.0",
+    "claude-code": "1.0.0",
+    "opencode": "0.1.0",
+    "codex": "0.1.0",
+}
+
+CLI_NAMES = {
+    "antigravity": "agy",
+    "claude-code": "claude",
+    "opencode": "opencode",
+    "codex": "codex",
+}
+
+ROLE_TO_SCHEMA = {
+    "plan": "planner",
+    "planner": "planner",
+    "implement": "implementer",
+    "implementer": "implementer",
+    "review": "reviewer",
+    "reviewer": "reviewer",
+    "critic": "critic",
+    "requirement-critic": "requirement-critic",
+    "plan-critic": "plan-critic",
+    "research": None,
+    "learner": "learner",
+}
+
+
+class UnsupportedHarnessVersionError(Exception):
+    """Raised when an installed harness version is below the supported threshold or invalid."""
+
+
+class ModelFallbackError(Exception):
+    """Raised when native or configured automatic model fallback is detected."""
+
+
+class ProtocolFailureError(Exception):
+    """Raised when an external role result is missing, undecodable or fails the schema contract."""
+
+
+def parse_semver(version_str: str) -> tuple[int, int, int]:
+    match = re.search(r"(\d+)\.(\d+)(?:\.(\d+))?", version_str)
+    if not match:
+        raise ValueError(f"Cannot parse version string: {version_str!r}")
+    major = int(match.group(1))
+    minor = int(match.group(2))
+    patch = int(match.group(3) or 0)
+    return (major, minor, patch)
+
+
+def validate_harness_version(
+    harness: str,
+    runner: Callable[..., Any] | None = None,
+) -> dict[str, Any]:
+    cli_name = CLI_NAMES.get(harness, harness)
+    try:
+        if runner:
+            try:
+                proc = runner([cli_name, "--version"])
+            except TypeError:
+                proc = runner([cli_name, "--version"], capture_output=True, text=True, check=False)
+        else:
+            proc = subprocess.run([cli_name, "--version"], capture_output=True, text=True, check=False)
+        if proc.returncode != 0:
+            return {
+                "valid": False,
+                "version": None,
+                "error": f"Harness CLI {cli_name} authentication/execution validation failed: return code {proc.returncode}",
+            }
+        stdout = (proc.stdout or "").strip() or (proc.stderr or "").strip()
+        version_tuple = parse_semver(stdout)
+        min_ver_str = MIN_SUPPORTED_VERSIONS.get(harness, "0.0.0")
+        min_tuple = parse_semver(min_ver_str)
+        if version_tuple < min_tuple:
+            ver_formatted = f"{version_tuple[0]}.{version_tuple[1]}.{version_tuple[2]}"
+            return {
+                "valid": False,
+                "version": ver_formatted,
+                "error": f"Unsupported installed version {ver_formatted} for {harness}. Minimum supported is {min_ver_str}.",
+            }
+        ver_formatted = f"{version_tuple[0]}.{version_tuple[1]}.{version_tuple[2]}"
+        return {"valid": True, "version": ver_formatted}
+    except Exception as exc:
+        return {
+            "valid": False,
+            "version": None,
+            "error": f"Harness CLI {cli_name} version check failed: {exc}",
+        }
+
+
+BASE_ROLES = ("plan", "implement", "review", "critic")
+DERIVED_ROLES = {
+    "requirement-critic": "critic",
+    "plan-critic": "critic",
+    "research": "plan",
+    "learner": "critic",
+}
+ALL_ROLES = (*BASE_ROLES, *DERIVED_ROLES.keys())
+
+
+def validate_effort(
+    model_id: str,
+    effort: str | None,
+    supported_efforts: list[str] | None = None,
+) -> dict[str, Any]:
+    """Validate that the requested reasoning effort is supported by the model."""
+    if not effort:
+        return {"valid": True}
+    if not supported_efforts:
+        return {
+            "valid": False,
+            "error": f"Model {model_id} does not declare supported reasoning effort values.",
+        }
+    normalized_supported = [e.lower() for e in supported_efforts]
+    if effort.lower() not in normalized_supported:
+        return {
+            "valid": False,
+            "error": f"Unsupported effort value {effort!r} for model {model_id}. Supported: {supported_efforts}",
+        }
+    return {"valid": True}
+
+HARNESS_DEFAULT_MODELS = {
+    "antigravity": "gemini-3.8-flash-medium",
+    "claude-code": "claude-3-7-sonnet-20250219",
+    "codex": "gpt-5.2-codex",
+    "opencode": "claude-3-7-sonnet-20250219",
+}
+
+HARNESS_SUPPORTED_EFFORTS = {
+    "antigravity": ["low", "medium", "high"],
+    "codex": ["low", "medium", "high"],
+}
+
+ENVIRONMENT_DEFAULTS = {
+    "plan": {"harness": "claude-code", "model": "claude-3-7-sonnet-20250219"},
+    "implement": {"harness": "claude-code", "model": "claude-3-7-sonnet-20250219"},
+    "review": {"harness": "claude-code", "model": "claude-3-7-sonnet-20250219"},
+    "critic": {"harness": "claude-code", "model": "claude-3-7-sonnet-20250219"},
+}
+
+
+def _fill_defaults(selection: dict[str, Any]) -> dict[str, Any]:
+    res = copy.deepcopy(selection)
+    harness = res.get("harness", "claude-code")
+    res["harness"] = harness
+    if "model" not in res or not res["model"]:
+        res["model"] = HARNESS_DEFAULT_MODELS.get(harness, "claude-3-7-sonnet-20250219")
+    return res
+
+
+def resolve_single_role(
+    role: str,
+    policy: dict | None = None,
+    run_overrides: dict | None = None,
+    issue_overrides: dict | None = None,
+    environment_defaults: dict | None = None,
+) -> dict[str, Any]:
+    """Resolve the effective execution selection for a single role.
+
+    Resolution order:
+    1. Issue-role override
+    2. Run-role override
+    3. Repository role default (policy["execution"]["roles"])
+    4. Explicitly confirmed environment default
+
+    Derived roles inherit from their parent role at each level unless explicitly set.
+    """
+    parent = DERIVED_ROLES.get(role)
+
+    # 1. Issue-role override
+    if issue_overrides:
+        if role in issue_overrides:
+            return _fill_defaults(issue_overrides[role])
+        if parent and parent in issue_overrides:
+            return _fill_defaults(issue_overrides[parent])
+
+    # 2. Run-role override
+    if run_overrides:
+        if role in run_overrides:
+            return _fill_defaults(run_overrides[role])
+        if parent and parent in run_overrides:
+            return _fill_defaults(run_overrides[parent])
+
+    # 3. Repository role default
+    roles_policy = (policy or {}).get("execution", {}).get("roles", {})
+    if role in roles_policy:
+        return _fill_defaults(roles_policy[role])
+    if parent and parent in roles_policy:
+        return _fill_defaults(roles_policy[parent])
+
+    # 4. Environment default
+    env_defs = environment_defaults or ENVIRONMENT_DEFAULTS
+    if role in env_defs:
+        return _fill_defaults(env_defs[role])
+    if parent and parent in env_defs:
+        return _fill_defaults(env_defs[parent])
+
+    return _fill_defaults(ENVIRONMENT_DEFAULTS.get("plan", {"harness": "claude-code", "model": "claude-3-7-sonnet-20250219"}))
+
+
+def resolve_roles(
+    policy: dict | None = None,
+    run_overrides: dict | None = None,
+    issue_overrides: dict | None = None,
+    environment_defaults: dict | None = None,
+) -> dict[str, dict[str, Any]]:
+    """Resolve all execution roles."""
+    return {
+        r: resolve_single_role(
+            r,
+            policy=policy,
+            run_overrides=run_overrides,
+            issue_overrides=issue_overrides,
+            environment_defaults=environment_defaults,
+        )
+        for r in ALL_ROLES
+    }
+
+
+def assess_model_strength(roles: dict[str, dict[str, Any]]) -> list[str]:
+    """Model strength guidance is advisory; returns non-blocking observations."""
+    notes = []
+    plan_model = roles.get("plan", {}).get("model", "")
+    critic_model = roles.get("critic", {}).get("model", "")
+    impl_model = roles.get("implement", {}).get("model", "")
+
+    plan_harness = roles.get("plan", {}).get("harness", "")
+    critic_harness = roles.get("critic", {}).get("harness", "")
+    impl_harness = roles.get("implement", {}).get("harness", "")
+
+    if critic_harness and impl_harness and critic_harness != impl_harness:
+        notes.append(f"Advisory: Critic uses external harness {critic_harness!r} differing from Implementer {impl_harness!r}.")
+    if "flash" in critic_model.lower() and "pro" in impl_model.lower():
+        notes.append(f"Advisory: Critic model {critic_model!r} may have lower reasoning capacity than Implementer {impl_model!r}.")
+    return notes
+
+
+def validate_selection(
+    selection: dict[str, Any],
+    role: str = "unknown",
+    root: Path | None = None,
+    check_auth: bool = True,
+    runner: Callable[..., Any] | None = None,
+) -> dict[str, Any]:
+    """Validate a single role execution selection."""
+    harness = selection.get("harness", "")
+    if harness not in SUPPORTED_HARNESSES:
+        return {
+            "valid": False,
+            "role": role,
+            "error": f"Unsupported harness {harness!r} for role {role}. Supported: {sorted(SUPPORTED_HARNESSES)}",
+        }
+
+    model = selection.get("model", "")
+    if not model:
+        return {
+            "valid": False,
+            "role": role,
+            "error": f"No model specified for role {role} on harness {harness}.",
+        }
+
+    effort = selection.get("effort")
+    if effort:
+        cap_file = Path(__file__).resolve().parents[1] / "capabilities" / f"{harness}.json"
+        supported_efforts = None
+        if cap_file.exists():
+            try:
+                cap_data = json.loads(cap_file.read_text(encoding="utf-8"))
+                model_meta = cap_data.get("models", {}).get(model, {})
+                supported_efforts = model_meta.get("supportedEfforts")
+            except Exception:
+                pass
+        if supported_efforts is None:
+            supported_efforts = HARNESS_SUPPORTED_EFFORTS.get(harness)
+        eff_check = validate_effort(model, effort, supported_efforts=supported_efforts)
+        if not eff_check["valid"]:
+            return {
+                "valid": False,
+                "role": role,
+                "error": eff_check.get("error", f"Invalid effort {effort} for model {model}"),
+            }
+
+    if root is not None and not root.is_dir():
+        return {
+            "valid": False,
+            "role": role,
+            "error": f"Working directory {root} does not exist",
+        }
+
+    if check_auth:
+        ver_check = validate_harness_version(harness, runner=runner)
+        if not ver_check["valid"]:
+            return {
+                "valid": False,
+                "role": role,
+                "error": ver_check["error"],
+            }
+
+    return {"valid": True, "role": role}
+
+
+def build_dispatch_command(
+    harness: str,
+    model: str,
+    prompt: str,
+    effort: str | None = None,
+    output_format: str = "json",
+) -> list[str]:
+    """Build the bounded native CLI dispatch command for a role invocation."""
+    h = (harness or "").lower()
+    if h == "antigravity":
+        cmd = ["agy", "--print", "--model", model]
+        if effort:
+            cmd.extend(["--effort", effort])
+        cmd.extend(["--output-format", output_format, "--dangerously-skip-permissions", prompt])
+        return cmd
+    elif h == "claude-code":
+        return ["claude", "-p", prompt, "--model", model, "--dangerously-skip-permissions"]
+    elif h == "opencode":
+        return ["opencode", "run", prompt, "--model", model]
+    elif h == "codex":
+        return ["codex", "exec", prompt, "--model", model]
+    else:
+        raise ValueError(f"Unsupported harness: {harness}")
+
+
+def check_model_fallback(requested_model: str, output_data: dict[str, Any] | str) -> None:
+    """Detect if the harness silently replaced the requested model with a fallback model."""
+    if isinstance(output_data, str):
+        if "fallback" in output_data.lower() and "model" in output_data.lower():
+            fb_match = re.search(r"falling back to (\S+)|fallback model:?\s*(\S+)", output_data, re.IGNORECASE)
+            if fb_match:
+                reported = fb_match.group(1) or fb_match.group(2)
+                raise ModelFallbackError(
+                    f"Automatic model fallback detected: requested {requested_model!r} but fell back to {reported!r}"
+                )
+        try:
+            parsed = json.loads(output_data)
+            if isinstance(parsed, dict):
+                output_data = parsed
+        except Exception:
+            pass
+
+    if isinstance(output_data, dict):
+        reported = output_data.get("model") or output_data.get("effective_model") or output_data.get("actual_model")
+        if reported and str(reported).strip().lower() != requested_model.strip().lower():
+            raise ModelFallbackError(
+                f"Automatic model fallback detected: requested {requested_model!r} but harness used {reported!r}"
+            )
+        if output_data.get("fallback_occurred") or output_data.get("model_fallback"):
+            raise ModelFallbackError(
+                f"Automatic model fallback detected for model {requested_model!r}"
+            )
+
+
+def parse_and_validate_result(
+    role: str,
+    raw_output: str,
+    schema: dict[str, Any] | None = None,
+) -> dict[str, Any]:
+    """Parse output and validate against the Result Contract; protocol failure on error."""
+    if not raw_output or not raw_output.strip():
+        raise ProtocolFailureError(f"Protocol failure: empty output returned for role {role}")
+
+    text = raw_output.strip()
+    if text.startswith("```json") and text.endswith("```"):
+        text = text[7:-3].strip()
+    elif text.startswith("```") and text.endswith("```"):
+        text = text[3:-3].strip()
+
+    try:
+        data = json.loads(text)
+    except json.JSONDecodeError as exc:
+        raise ProtocolFailureError(f"Protocol failure: invalid JSON output for role {role}: {exc}") from exc
+
+    if not isinstance(data, dict):
+        raise ProtocolFailureError(f"Protocol failure: expected JSON object for role {role}, got {type(data).__name__}")
+
+    schema_role = ROLE_TO_SCHEMA.get(role, role)
+    if schema_role:
+        if schema is None:
+            try:
+                schema = result.load_schema(schema_role)
+            except Exception as exc:
+                raise ProtocolFailureError(f"Protocol failure: failed to load schema for {schema_role}: {exc}") from exc
+        errors = result.validate(data, schema)
+        if errors:
+            raise ProtocolFailureError(f"Protocol failure: schema validation failed for {role}: {errors}")
+
+    return data
+
+
+def verify_critic_result(critic_result: dict[str, Any]) -> None:
+    """Verify that Critic independently verified criteria and gates without accepting a summary.
+
+    Permission or tool limitations must fail visibly rather than accepting a summary.
+    """
+    if not isinstance(critic_result, dict):
+        raise ProtocolFailureError("Critic result must be a JSON object")
+
+    if critic_result.get("permission_error") or critic_result.get("tool_limitation"):
+        err = critic_result.get("error") or "Critic encountered permission or tool limitation"
+        raise RuntimeError(f"Critic execution failed visibly: {err}")
+
+    if critic_result.get("complete") is True:
+        gate_result = critic_result.get("gateResult")
+        if not gate_result or not isinstance(gate_result, dict):
+            raise ProtocolFailureError("Critic claimed complete=true without required gateResult")
+        if gate_result.get("verdict") != "pass":
+            raise ProtocolFailureError(f"Critic claimed complete=true but gateResult verdict is {gate_result.get('verdict')!r}")
+
+        criteria = critic_result.get("criteria")
+        if not isinstance(criteria, list) or len(criteria) == 0:
+            raise ProtocolFailureError("Critic claimed complete=true with empty criteria list")
+        for c in criteria:
+            if not isinstance(c, dict) or not c.get("met") or not str(c.get("evidence", "")).strip():
+                raise ProtocolFailureError(f"Critic claimed complete=true but criterion lacks verified evidence: {c}")
+
+
+def dispatch_role(
+    role: str,
+    prompt: str,
+    cwd: Path | str,
+    selection: dict[str, Any] | None = None,
+    policy: dict | None = None,
+    run_overrides: dict | None = None,
+    issue_overrides: dict | None = None,
+    environment_defaults: dict | None = None,
+    runner: Callable[..., Any] | None = None,
+    run_id: str | None = None,
+    unit_id: str | None = None,
+    state_root: str | None = None,
+    retry_on_invalid: bool = True,
+) -> dict[str, Any]:
+    """Execute a bounded role invocation in the selected native harness."""
+    cwd_path = Path(cwd).resolve()
+    if not cwd_path.is_dir():
+        raise ValueError(f"Working directory {cwd_path} does not exist")
+
+    eff_selection = selection or resolve_single_role(
+        role,
+        policy=policy,
+        run_overrides=run_overrides,
+        issue_overrides=issue_overrides,
+        environment_defaults=environment_defaults,
+    )
+    harness = eff_selection.get("harness", "")
+    model = eff_selection.get("model", "")
+    effort = eff_selection.get("effort")
+
+    ver_check = validate_harness_version(harness, runner=runner)
+    if not ver_check["valid"]:
+        raise UnsupportedHarnessVersionError(ver_check["error"])
+
+    eff_val = validate_selection(eff_selection, role=role, root=cwd_path, check_auth=False)
+    if not eff_val["valid"]:
+        raise ValueError(eff_val["error"])
+
+    if run_id and unit_id:
+        try:
+            root = runlog.state_root(state_root)
+            log_path = runlog.run_log_path(root, unit_id, run_id)
+            if log_path.exists():
+                event_data = {
+                    "role": role,
+                    "harness": harness,
+                    "model": model,
+                    "cwd": str(cwd_path),
+                }
+                if effort:
+                    event_data["effort"] = effort
+                runlog.append_event(
+                    log_path,
+                    runlog.validate_event({
+                        "ts": datetime.datetime.now(datetime.timezone.utc).strftime("%Y-%m-%dT%H:%M:%SZ"),
+                        "run": run_id,
+                        "event": "subagent.started",
+                        "data": event_data,
+                    }),
+                )
+        except Exception:
+            pass
+
+    cmd = build_dispatch_command(harness, model, prompt, effort=effort)
+    run_fn = runner or subprocess.run
+    proc = run_fn(cmd, cwd=str(cwd_path), capture_output=True, text=True, check=False)
+    if proc.returncode != 0:
+        err = (proc.stderr or "").strip() or (proc.stdout or "").strip()
+        raise RuntimeError(f"Harness {harness} execution failed (exit {proc.returncode}): {err}")
+
+    check_model_fallback(model, proc.stdout)
+
+    try:
+        data = parse_and_validate_result(role, proc.stdout)
+        if role in ("critic", "requirement-critic", "plan-critic"):
+            verify_critic_result(data)
+    except ProtocolFailureError as exc:
+        if retry_on_invalid:
+            retry_prompt = f"{prompt}\n\nYour prior result was invalid: {exc}\nReturn the complete {role} result contract."
+            retry_cmd = build_dispatch_command(harness, model, retry_prompt, effort=effort)
+            proc2 = run_fn(retry_cmd, cwd=str(cwd_path), capture_output=True, text=True, check=False)
+            if proc2.returncode != 0:
+                raise RuntimeError(f"Harness {harness} retry failed (exit {proc2.returncode}): {(proc2.stderr or '').strip()}")
+            check_model_fallback(model, proc2.stdout)
+            data = parse_and_validate_result(role, proc2.stdout)
+            if role in ("critic", "requirement-critic", "plan-critic"):
+                verify_critic_result(data)
+        else:
+            raise
+
+    if run_id and unit_id:
+        try:
+            root = runlog.state_root(state_root)
+            log_path = runlog.run_log_path(root, unit_id, run_id)
+            if log_path.exists():
+                event_data = {
+                    "role": role,
+                    "harness": harness,
+                    "model": model,
+                    "result": data,
+                }
+                if effort:
+                    event_data["effort"] = effort
+                runlog.append_event(
+                    log_path,
+                    runlog.validate_event({
+                        "ts": datetime.datetime.now(datetime.timezone.utc).strftime("%Y-%m-%dT%H:%M:%SZ"),
+                        "run": run_id,
+                        "event": "subagent.stopped",
+                        "data": event_data,
+                    }),
+                )
+        except Exception:
+            pass
+
+    return data
+
+
+def preflight_validate(
+    policy: dict | None = None,
+    run_overrides: dict | None = None,
+    issue_overrides: dict | None = None,
+    environment_defaults: dict | None = None,
+    root: Path | None = None,
+    check_auth: bool = True,
+    runner: Callable[..., Any] | None = None,
+) -> dict[str, Any]:
+    """Preflight check: resolve effective selections and validate all roles."""
+    roles = resolve_roles(
+        policy=policy,
+        run_overrides=run_overrides,
+        issue_overrides=issue_overrides,
+        environment_defaults=environment_defaults,
+    )
+    errors = []
+    for r, sel in roles.items():
+        val = validate_selection(sel, role=r, root=root, check_auth=check_auth, runner=runner)
+        if not val["valid"]:
+            errors.append(val["error"])
+
+    guidance = assess_model_strength(roles)
+    valid = len(errors) == 0
+
+    return {
+        "valid": valid,
+        "roles": roles,
+        "errors": errors,
+        "guidance": guidance,
+    }
+
+
+def validate_role_replacement(
+    role: str,
+    selection: dict[str, Any],
+    issue_ref: str,
+    root: Path | None = None,
+    check_auth: bool = False,
+    runner: Callable[..., Any] | None = None,
+) -> dict[str, Any]:
+    """Validate an explicit role replacement for an Issue without mutating defaults."""
+    if not isinstance(selection, dict):
+        return {"valid": False, "error": "Replacement selection must be an object"}
+    val = validate_selection(selection, role=role, root=root, check_auth=check_auth, runner=runner)
+    if not val["valid"]:
+        return {"valid": False, "error": val["error"]}
+    return {
+        "valid": True,
+        "role": role,
+        "issue": issue_ref,
+        "selection": selection,
+    }
+
+
+def check_unresolved_failures(
+    unit_id: str,
+    run_id: str,
+    state_root_path: Path | str | None = None,
+) -> list[dict[str, Any]]:
+    """Derive unresolved paused executions from the Run log.
+
+    An issue execution is paused if an `issue.paused` event was recorded.
+    It is considered resolved if a later `role.changed` event was recorded for that same issue and role.
+    """
+    root = runlog.state_root(str(state_root_path) if state_root_path else None)
+    log_path = runlog.run_log_path(root, unit_id, run_id)
+    if not log_path.is_file():
+        return []
+
+    events = runlog.read_valid_events(log_path)
+    paused_map: dict[tuple[str, str], dict[str, Any]] = {}
+
+    for event in events:
+        ev_type = event.get("event")
+        issue = event.get("issue")
+        data = event.get("data", {})
+        if ev_type == "issue.paused" and issue:
+            role = data.get("role", "critic")
+            paused_map[(issue, role)] = {
+                "issue": issue,
+                "role": role,
+                "worktree": data.get("worktree"),
+                "reason": data.get("reason", "execution_unavailable"),
+                "error": data.get("error", "execution unavailable"),
+            }
+        elif ev_type == "role.changed" and issue:
+            role = data.get("role", "critic")
+            paused_map.pop((issue, role), None)
+        elif ev_type in ("issue.done", "issue.blocked") and issue:
+            for key in list(paused_map.keys()):
+                if key[0] == issue:
+                    paused_map.pop(key, None)
+
+    return list(paused_map.values())
+
+
+def main() -> int:
+    parser = argparse.ArgumentParser(description=__doc__)
+    subparsers = parser.add_subparsers(dest="command")
+
+    resolve_parser = subparsers.add_parser("resolve", help="resolve effective role executions")
+    resolve_parser.add_argument("--role", help="specific role to resolve")
+    resolve_parser.add_argument("--run-overrides", help="JSON string of run-level overrides")
+    resolve_parser.add_argument("--issue-overrides", help="JSON string of issue-level overrides")
+    resolve_parser.add_argument("--cwd", default=".", help="repository root path")
+    resolve_parser.add_argument("--json", action="store_true", help="output JSON")
+
+    preflight_parser = subparsers.add_parser("preflight", help="run preflight validation across roles")
+    preflight_parser.add_argument("--run-overrides", help="JSON string of run-level overrides")
+    preflight_parser.add_argument("--issue-overrides", help="JSON string of issue-level overrides")
+    preflight_parser.add_argument("--cwd", default=".", help="repository root path")
+    preflight_parser.add_argument("--json", action="store_true", help="output JSON")
+
+    dispatch_parser = subparsers.add_parser("dispatch", help="dispatch bounded role invocation")
+    dispatch_parser.add_argument("--role", required=True, help="role to execute")
+    dispatch_parser.add_argument("--prompt", help="prompt text (or via stdin)")
+    dispatch_parser.add_argument("--cwd", default=".", help="working directory")
+    dispatch_parser.add_argument("--selection", help="JSON selection object with harness, model, effort")
+    dispatch_parser.add_argument("--run-overrides", help="JSON string of run-level overrides")
+    dispatch_parser.add_argument("--issue-overrides", help="JSON string of issue-level overrides")
+    dispatch_parser.add_argument("--run-id", help="Run ID for logging")
+    dispatch_parser.add_argument("--unit-id", help="Unit ID for logging")
+    dispatch_parser.add_argument("--state-root", help="State root directory")
+    dispatch_parser.add_argument("--json", action="store_true", help="output JSON")
+
+    failures_parser = subparsers.add_parser("unresolved-failures", help="check for unresolved paused executions")
+    failures_parser.add_argument("unit_id", help="repository unit ID")
+    failures_parser.add_argument("run_id", help="run ID")
+    failures_parser.add_argument("--state-root", help="override state root")
+    failures_parser.add_argument("--json", action="store_true", help="output JSON")
+
+    validate_rep_parser = subparsers.add_parser("validate-replacement", help="validate role replacement")
+    validate_rep_parser.add_argument("--issue", required=True, help="issue ref")
+    validate_rep_parser.add_argument("--role", required=True, help="role name")
+    validate_rep_parser.add_argument("--selection", required=True, help="selection JSON")
+    validate_rep_parser.add_argument("--cwd", default=".", help="repository root path")
+    validate_rep_parser.add_argument("--json", action="store_true", help="output JSON")
+
+    args = parser.parse_args()
+    root = Path(args.cwd).resolve() if getattr(args, "cwd", None) else Path(".").resolve()
+    policy = resolve_policy(root)
+
+    run_ov = json.loads(args.run_overrides) if getattr(args, "run_overrides", None) else None
+    issue_ov = json.loads(args.issue_overrides) if getattr(args, "issue_overrides", None) else None
+
+    if args.command == "resolve":
+        if args.role:
+            res = resolve_single_role(args.role, policy=policy, run_overrides=run_ov, issue_overrides=issue_ov)
+        else:
+            res = resolve_roles(policy=policy, run_overrides=run_ov, issue_overrides=issue_ov)
+        if args.json:
+            print(json.dumps(res, indent=2))
+        else:
+            print(res)
+        return 0
+
+    elif args.command == "preflight":
+        res = preflight_validate(policy=policy, run_overrides=run_ov, issue_overrides=issue_ov, root=root)
+        if args.json:
+            print(json.dumps(res, indent=2))
+        else:
+            if res["valid"]:
+                print("Preflight validation passed.")
+                for g in res["guidance"]:
+                    print(f"  {g}")
+            else:
+                print("Preflight validation FAILED:")
+                for err in res["errors"]:
+                    print(f"  - {err}", file=sys.stderr)
+        return 0 if res["valid"] else 1
+
+    elif args.command == "dispatch":
+        prompt = args.prompt
+        if not prompt:
+            prompt = sys.stdin.read()
+        selection = json.loads(args.selection) if getattr(args, "selection", None) else None
+        try:
+            res = dispatch_role(
+                role=args.role,
+                prompt=prompt,
+                cwd=root,
+                selection=selection,
+                policy=policy,
+                run_overrides=run_ov,
+                issue_overrides=issue_ov,
+                run_id=getattr(args, "run_id", None),
+                unit_id=getattr(args, "unit_id", None),
+                state_root=getattr(args, "state_root", None),
+            )
+            print(json.dumps(res, indent=2))
+            return 0
+        except ProtocolFailureError as exc:
+            print(f"protocol failure: {exc}", file=sys.stderr)
+            return 2
+        except Exception as exc:
+            print(f"execution failure: {exc}", file=sys.stderr)
+            return 1
+
+    elif args.command == "unresolved-failures":
+        failures = check_unresolved_failures(
+            unit_id=args.unit_id,
+            run_id=args.run_id,
+            state_root_path=getattr(args, "state_root", None),
+        )
+        if args.json:
+            print(json.dumps(failures, indent=2))
+        else:
+            print(f"Unresolved failures: {len(failures)}")
+            for f in failures:
+                print(f"  - {f['issue']} ({f['role']}): {f['reason']}")
+        return 0
+
+    elif args.command == "validate-replacement":
+        sel = json.loads(args.selection)
+        res = validate_role_replacement(
+            role=args.role,
+            selection=sel,
+            issue_ref=args.issue,
+            root=root,
+        )
+        if args.json:
+            print(json.dumps(res, indent=2))
+        else:
+            print("Valid" if res["valid"] else f"Invalid: {res.get('error')}")
+        return 0 if res["valid"] else 1
+
+    else:
+        parser.print_help()
+        return 0
+
+
+if __name__ == "__main__":
+    sys.exit(main())