Skip to content

Feat/v2 #9

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Merged
mdvcode merged 3 commits into from
Aug 12, 2026
Merged

Feat/v2 #9

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
3 commits
Select commit Hold shift + click to select a range
535fcef
Add task execution modes to dashboard
mdvcode Aug 7, 2026
d4be2a2
Add structured dashboard answers and prevent repeated questions
mdvcode Aug 11, 2026
7a03fca
Add checkpointed goal mode and improve task monitoring
mdvcode Aug 12, 2026
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
9 changes: 7 additions & 2 deletions .agent-workflows.yaml
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -5,18 +5,23 @@ workflows:
budgets:
max_roles: 40
max_repair_iterations: 12
max_duration_seconds: 7200
max_duration_seconds: 3600
max_tokens: 1500000
mode_budgets:
fast:
max_roles: 12
max_repair_iterations: 1
max_duration_seconds: 900
max_tokens: 300000
goal:
max_roles: 80
max_repair_iterations: 24
max_duration_seconds: 14400
max_tokens: 3000000
runtime_provider: codex-cli
executor: "python3 scripts/agent_role_runner.py --workflow full_agent_workflow --run-id {run_id} --task-id {task_id} --goal {goal} --project {project} --repo {repository} --branch {branch} --base-branch {base_branch} --artifacts-dir {artifacts_dir} --create-worktree --runtime-provider {runtime_provider}"
max_iterations: 1
timeout_seconds: 7200
timeout_seconds: 14400
retry:
max_retries: 0
backoff_seconds: 0
Expand Down
16 changes: 9 additions & 7 deletions README.md
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -70,15 +70,17 @@ The default mode is `auto`:
agent task "Fix a typo in the settings page"
agent task --mode fast "Apply a small local styling change"
agent task --mode full "Refactor the authentication architecture"
agent task --mode goal "Complete a checkpointed multi-hour objective"
```

| Mode | Behavior |
| --- | --- |
| `auto` | Uses the guarded fast workflow for ordinary work and selects the full workflow when the goal names sensitive or broad changes. |
| `fast` | Requests the short workflow with implementation and review as the only model-backed roles. Context, quality, security, and verdict stages are deterministic. |
| `full` | Runs the complete specialist workflow intentionally. |
| `auto` | Uses the guarded fast workflow for ordinary work and selects the full workflow when the goal names sensitive or broad changes. It never selects `goal`. |
| `fast` | Runs the short workflow for at most 15 minutes, with implementation and review as the only model-backed roles. Context, quality, security, and verdict stages are deterministic. |
| `full` | Runs the complete specialist workflow for at most 60 minutes. |
| `goal` | Explicitly runs a checkpointed long objective for at most 4 hours. Use it only when the success condition genuinely needs multiple hours. |

Fast mode has a 15-minute workflow budget. It automatically escalates to the full workflow before publication if the patch touches protected areas, changes more than five files, exceeds 200 changed lines, or reports increased risk. Required checks and approval gates are never bypassed.
Fast mode automatically escalates to the full workflow before publication if the patch touches protected areas, changes more than five files, exceeds 200 changed lines, or reports increased risk. Required checks and approval gates are never bypassed. The 30-minute role timeout is an emergency limit for one model executor, not the duration of the whole task; workflow, recovery, iteration, and human-attention limits are tracked separately.

## Branch and workspace modes

Expand Down Expand Up @@ -161,7 +163,7 @@ agent init [--repo PATH] [--project-id ID]
```sh
agent task [--repo PATH] [--task-id ID] [--branch BRANCH]
[--current-branch | --worktree] [--keep-paused]
[--mode auto|fast|full] [--priority -100..100]
[--mode auto|fast|full|goal] [--priority -100..100]
[--max-retries 0..10] [--dry-run] [--json]
"TASK DESCRIPTION"
```
Expand Down Expand Up @@ -192,7 +194,7 @@ agent worker stop [--json]
agent dashboard [--repo PATH] [--port PORT] [--no-open]
```

The dashboard binds to loopback, opens in the default browser, and provides task launch, status, answer, approval, retry, and abort controls. `--no-open` starts the server without opening a browser. `Ctrl+C` stops the dashboard server but does not stop the worker service.
The dashboard binds to loopback, opens in the default browser, and provides task launch, execution-mode (`auto`, `fast`, `full`, or explicit `goal`) and Git-workspace selection, status, structured answer choices with a custom-answer fallback, approval, retry, and abort controls. Answered questions are fingerprinted so the same question cannot silently reopen in a loop. `--no-open` starts the server without opening a browser. `Ctrl+C` stops the dashboard server but does not stop the worker service.

### Status and monitoring

Expand All @@ -204,7 +206,7 @@ agent watch [--repo PATH] [--task-id ID] [--run-id ID]
```

- `status` shows compact project, queue, run, and worker state without raw model transcripts.
- `watch` follows state transitions until completion or human attention. A timeout of `0` waits indefinitely.
- `watch` follows state transitions until completion or human attention and returns after 30 minutes by default. Use `--timeout 0` only when an intentionally unbounded terminal wait is desired; if the worker service is down, `watch` returns immediately with the start command.

### Failure inspection

Expand Down
99 changes: 82 additions & 17 deletions ai_harness/cli.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -570,6 +570,7 @@ def project_tasks(db_path: Path, repository: Path) -> list[dict[str, Any]]:
if "cancellation_requested_at" in row.keys()
else 0.0
),
"created_at": float(row["created_at"]),
"updated_at": float(row["updated_at"]),
}
)
Expand All @@ -586,10 +587,20 @@ def workflow_attention(workflow: dict[str, Any]) -> dict[str, Any]:
"details": [str(item) for item in details] if isinstance(details, list) else [],
"role": str(raw.get("role", workflow.get("current_role", ""))),
"action": str(raw.get("action", "")),
"question": raw.get("question", {}) if isinstance(raw.get("question"), dict) else {},
"repeated_question": raw.get("repeated_question") is True,
}
status = str(workflow.get("execution_status", ""))
if status not in {"awaiting_approval", "blocked", "dead_letter", "failed"}:
return {"required": False, "summary": "", "details": [], "role": "", "action": ""}
return {
"required": False,
"summary": "",
"details": [],
"role": "",
"action": "",
"question": {},
"repeated_question": False,
}
values: list[str] = []
blockers = workflow.get("blockers", [])
if isinstance(blockers, list):
Expand All @@ -616,6 +627,8 @@ def workflow_attention(workflow: dict[str, Any]) -> dict[str, Any]:
# an unknown approval gate conservatively: an informational answer must
# never stand in for a risk or security decision.
"action": "approve" if status == "awaiting_approval" else "fix_then_retry",
"question": {},
"repeated_question": False,
}


Expand Down Expand Up @@ -771,7 +784,14 @@ def handle_approve(args: argparse.Namespace) -> int:
return 0


def record_human_input(run_dir: Path, *, run_id: str, actor: str, response: str) -> Path:
def record_human_input(
run_dir: Path,
*,
run_id: str,
actor: str,
response: str,
attention: dict[str, Any] | None = None,
) -> Path:
path = run_dir / "human-input.json"
temporary = path.with_suffix(".json.tmp")
if path.is_symlink() or temporary.is_symlink():
Expand All @@ -780,13 +800,17 @@ def record_human_input(run_dir: Path, *, run_id: str, actor: str, response: str)
entries = existing.get("entries", []) if isinstance(existing, dict) else []
if not isinstance(entries, list):
entries = []
entries.append(
{
"time": datetime.now(timezone.utc).isoformat(),
"actor": sanitized_message(actor, limit=128),
"response": sanitized_message(response, limit=10_000),
}
)
entry = {
"time": datetime.now(timezone.utc).isoformat(),
"actor": sanitized_message(actor, limit=128),
"response": sanitized_message(response, limit=10_000),
}
if isinstance(attention, dict):
entry["question_fingerprint"] = str(attention.get("fingerprint", ""))[:100]
question = attention.get("question", {})
if isinstance(question, dict):
entry["question_id"] = sanitized_message(str(question.get("id", "")), limit=80)
entries.append(entry)
payload = {"version": 1, "run_id": run_id, "entries": entries[-50:]}
with temporary.open("w", encoding="utf-8") as handle:
handle.write(json.dumps(payload, indent=2, ensure_ascii=False) + "\n")
Expand All @@ -805,7 +829,7 @@ def resolve_answer_attention(run_dir: Path) -> None:
raise CLIError("refusing to update workflow state through a symbolic link")
workflow = read_json_object(path)
attention = workflow.get("attention")
if not isinstance(attention, dict) or attention.get("action") != "answer_or_approve":
if not isinstance(attention, dict) or attention.get("action") not in {"answer", "answer_or_approve"}:
raise CLIError("the workflow question changed before the answer could be applied")
attention_details = attention.get("details", [])
active_values = {str(attention.get("summary", "")).strip()}
Expand Down Expand Up @@ -872,7 +896,7 @@ def handle_answer(args: argparse.Namespace) -> int:
run_dir = root / ".agent-runs" / args.run_id
workflow = read_json_object(run_dir / "workflow.json")
attention = workflow.get("attention", {})
if not isinstance(attention, dict) or attention.get("action") != "answer_or_approve":
if not isinstance(attention, dict) or attention.get("action") not in {"answer", "answer_or_approve"}:
raise CLIError(
"this gate requires an explicit approval decision, not an informational answer; use `agent approve`"
)
Expand All @@ -881,6 +905,7 @@ def handle_answer(args: argparse.Namespace) -> int:
run_id=args.run_id,
actor=args.actor,
response=response,
attention=attention,
)
approval_lifecycle = load_harness_module(root, "approval_lifecycle")
task_queue = load_harness_module(root, "task_queue")
Expand Down Expand Up @@ -951,6 +976,12 @@ def attention_items(
"summary": summary or "The task requires attention.",
"details": list(attention.get("details", [])) if isinstance(attention, dict) else [],
"action": str(attention.get("action", "")) if isinstance(attention, dict) else "",
"question": dict(attention.get("question", {}))
if isinstance(attention, dict) and isinstance(attention.get("question"), dict)
else {},
"repeated_question": bool(
isinstance(attention, dict) and attention.get("repeated_question") is True
),
"approval_pending": bool(
isinstance(run, dict) and run.get("approval", {}).get("status") == "pending"
),
Expand All @@ -973,6 +1004,10 @@ def attention_items(
"summary": attention.get("summary", "The task requires attention."),
"details": list(attention.get("details", [])),
"action": str(attention.get("action", "")),
"question": dict(attention.get("question", {}))
if isinstance(attention.get("question"), dict)
else {},
"repeated_question": attention.get("repeated_question") is True,
"approval_pending": run.get("approval", {}).get("status") == "pending",
}
)
Expand All @@ -990,7 +1025,17 @@ def attention_output_lines(items: list[dict[str, Any]], repository: Path) -> lis
if str(detail).strip() and str(detail).strip() != item["summary"]:
lines.append(f" needed: {detail}")
run_id = str(item["run_id"])
if item["approval_pending"] and run_id and item.get("action") == "answer_or_approve":
question = item.get("question", {})
options = question.get("options", []) if isinstance(question, dict) else []
if isinstance(options, list):
for index, option in enumerate(options, start=1):
if not isinstance(option, dict):
continue
recommended = " (recommended)" if option.get("recommended") is True else ""
description = str(option.get("description", "")).strip()
suffix = f" — {description}" if description else ""
lines.append(f" option {index}: {option.get('label', '')}{recommended}{suffix}")
if item["approval_pending"] and run_id and item.get("action") in {"answer", "answer_or_approve"}:
lines.append(f' answer: agent answer --repo {shlex.quote(str(repository))} {run_id} "Your answer"')
elif item["approval_pending"] and run_id:
lines.append(f" approve if acceptable: agent approve --repo {shlex.quote(str(repository))} --run-id {run_id}")
Expand Down Expand Up @@ -1036,9 +1081,10 @@ def handle_status(args: argparse.Namespace) -> int:
lines.extend(attention_output_lines(attention, repository))
for item in tasks[: args.limit]:
marker = " !" if item["requires_human"] else ""
age_seconds = max(0, int(time.time() - item["created_at"]))
lines.append(
f" task {item['queue_task_id']}: {item['task_id']} [{item['status']}]{marker} "
f"workspace={item['workspace_mode']}"
f"workspace={item['workspace_mode']} age={age_seconds}s"
)
if item["failure_kind"]:
lines.append(
Expand Down Expand Up @@ -1076,7 +1122,7 @@ def handle_status(args: argparse.Namespace) -> int:
f"(gate: {run['approval']['gate']})"
)
run_attention = run.get("attention", {})
if isinstance(run_attention, dict) and run_attention.get("action") == "answer_or_approve":
if isinstance(run_attention, dict) and run_attention.get("action") in {"answer", "answer_or_approve"}:
lines.append(" answer the question shown above; approval alone cannot supply missing information")
elif run["approval"]["registration_required"]:
lines.append(" publication requires trusted repository registration")
Expand Down Expand Up @@ -1153,6 +1199,17 @@ def handle_watch(args: argparse.Namespace) -> int:
lines=attention_output_lines(attention, repository),
)
return 0
service = worker_service_status(root)
if task_status not in {"completed", "cancelled"} and run_status != "completed" and not service["alive"]:
emit(
{**snapshot, "worker_service": service},
as_json=args.json,
lines=(
f"Task {task_id} is waiting because the worker service is not running.",
f"Start it with: agent start --repo {repository}",
),
)
return 0
terminal = task_status in {"completed", "cancelled"} or run_status == "completed"
if terminal:
emit(
Expand Down Expand Up @@ -1964,9 +2021,12 @@ def build_parser() -> argparse.ArgumentParser:
)
task_parser.add_argument(
"--mode",
choices=("auto", "fast", "full"),
choices=("auto", "fast", "full", "goal"),
default="auto",
help="select guarded fast routing automatically, require fast routing, or preserve the full role chain",
help=(
"select routing automatically, request the 15-minute fast path, run the full role chain "
"for up to 60 minutes, or explicitly run a checkpointed goal for up to 4 hours"
),
)
task_parser.add_argument("--priority", type=int, choices=range(-100, 101), default=0)
task_parser.add_argument("--max-retries", type=int, choices=range(0, 11), default=2)
Expand Down Expand Up @@ -2004,7 +2064,12 @@ def build_parser() -> argparse.ArgumentParser:
watch_parser.add_argument("--task-id", default="")
watch_parser.add_argument("--run-id", default="")
watch_parser.add_argument("--interval", type=float, default=2.0)
watch_parser.add_argument("--timeout", type=float, default=0.0, help="seconds; zero waits indefinitely")
watch_parser.add_argument(
"--timeout",
type=float,
default=1800.0,
help="seconds (default: 1800); zero waits indefinitely",
)
watch_parser.add_argument("--json", action="store_true")
watch_parser.set_defaults(handler=handle_watch)

Expand Down
Loading
Loading