diff --git a/README.md b/README.md
index 811e384..06fb823 100644
--- a/README.md
+++ b/README.md
@@ -111,10 +111,12 @@ parallel-cli
 │   ├── list                # List available Parallel skills
 │   ├── reinstall           # Cleanly reinstall managed Parallel skills
 │   └── uninstall           # Remove only CLI-managed Parallel skills
-└── memory                  # Saved Task, Monitor, and FindAll entries
-    ├── retrieve            # Search Memory or list recent entries
-    ├── evict               # Remove one entry from Memory
-    └── clear               # Permanently clear selected Memory
+├── memory                  # Saved Task, Monitor, and FindAll entries
+│   ├── retrieve            # Search Memory or list recent entries
+│   ├── evict               # Remove one entry from Memory
+│   └── clear               # Permanently clear selected Memory
+└── feedback                # Report problems to the Parallel team
+    └── send                # Send a report and show what Parallel received
 ```
 
 Run `parallel-cli skills install` to install or update agent skills. When using
@@ -239,6 +241,26 @@ Omit `--scope-key` / `--memory-scope-key` to use personal Memory when the
 credential and account are eligible. Application credentials require a stable
 scope key containing only letters, digits, underscores, or hyphens.
 
+### 7. Report a Problem to Parallel
+
+Report a problem with Parallel, with the request IDs attached: wrong, stale or
+missing results, an answer the user says was wrong, a result you had to work
+around, an error your request didn't cause, or a benchmark where Parallel
+underperformed. The CLI sends the report and shows exactly what Parallel received.
+
+```bash
+parallel-cli feedback send \
+    -m "Eval runs return the wrong CEO for most companies." \
+    --category results_quality --product task --severity blocking \
+    --example "trun_abc123:Returned John Roe; expected Jane Doe"
+
+# Agents: --json prints the response, including what Parallel received.
+parallel-cli feedback send ... --json
+```
+
+Feedback is open only to Parallel's own organizations for now, and never to
+organizations on zero or partial data retention.
+
 ## Non-Interactive Mode (for AI Agents & Scripts)
 
 All commands support `--json` output and can be fully controlled via CLI arguments.
diff --git a/parallel_web_tools/cli/commands.py b/parallel_web_tools/cli/commands.py
index c48e4e9..f4c7a3f 100644
--- a/parallel_web_tools/cli/commands.py
+++ b/parallel_web_tools/cli/commands.py
@@ -63,6 +63,14 @@
     update_monitor,
     validate_memory_scope_key,
 )
+from parallel_web_tools.core.feedback import (
+    FEEDBACK_CATEGORIES,
+    FEEDBACK_PRODUCTS,
+    FEEDBACK_SEVERITIES,
+    FeedbackInputError,
+    parse_example,
+    send_feedback,
+)
 
 # Standalone CLI (PyInstaller) has limited features to reduce bundle size
 # YAML config and interactive planner require: pip install parallel-web-tools[cli]
@@ -1084,6 +1092,110 @@ def memory_clear(memory_scope_key: str | None, confirm_clear: bool, output_json:
     console.print("[dim]Underlying Task, Monitor, and FindAll resources were not deleted.[/dim]")
 
 
+# =============================================================================
+# Feedback Commands
+# =============================================================================
+
+
+@main.group()
+def feedback():
+    """Report problems with Parallel to the Parallel team."""
+    pass
+
+
+def _example_option(_ctx: click.Context, _param: click.Parameter, values: tuple[str, ...]) -> list[dict[str, str]]:
+    try:
+        return [parse_example(value) for value in values]
+    except FeedbackInputError as e:
+        raise click.BadParameter(str(e)) from e
+
+
+def _render_feedback_result(response: dict[str, Any]) -> None:
+    if response.get("duplicate"):
+        console.print(f"[yellow]Parallel already received this report as {response.get('feedback_id')}.[/yellow]")
+    else:
+        console.print(f"[bold green]Sent to Parallel as {response.get('feedback_id')}.[/bold green]")
+    report = response.get("report", {})
+    context = report.get("context", {})
+    console.print(
+        f"  {report.get('category')} · {report.get('product')} · {report.get('severity')}",
+        markup=False,
+    )
+    console.print(f"  Message: {report.get('message', '')}", markup=False)
+    for label in ("expected", "actual"):
+        if report.get(label):
+            console.print(f"  {label.capitalize()}: {report[label]}", markup=False)
+    for example in report.get("examples", []):
+        note = f" — {example['issue']}" if example.get("issue") else ""
+        console.print(f"  • {example.get('id')}{note}", markup=False)
+    console.print(f"  From: {context.get('sender', 'unknown')}", markup=False)
+    if report.get("redactions"):
+        console.print(f"  [dim]{report['redactions']} secret(s) removed.[/dim]")
+    if report.get("recipient"):
+        console.print(f"  {report['recipient']}", markup=False, style="dim")
+
+
+@feedback.command(name="send")
+@click.option("-m", "--message", required=True, help="What you were doing, what went wrong, and what you expected.")
+@click.option("--category", type=click.Choice(list(FEEDBACK_CATEGORIES)), required=True, help="The kind of problem.")
+@click.option(
+    "--product", type=click.Choice(list(FEEDBACK_PRODUCTS)), required=True, help="The Parallel product involved."
+)
+@click.option(
+    "--severity",
+    type=click.Choice(list(FEEDBACK_SEVERITIES)),
+    default="degraded",
+    show_default=True,
+    help="blocking, degraded (a workaround was needed) or minor (noticeable, but the work went on unchanged).",
+)
+@click.option(
+    "--example",
+    "examples",
+    multiple=True,
+    callback=_example_option,
+    help="A request ID, optionally with a note: 'trun_abc:Returned the wrong CEO'. Repeatable.",
+)
+@click.option("--expected", help="The overall expected behavior.")
+@click.option("--actual", help="The overall observed behavior.")
+@click.option("--json", "output_json", is_flag=True, help="Output JSON to stdout.")
+def feedback_send(
+    message: str,
+    category: str,
+    product: str,
+    severity: str,
+    examples: list[dict[str, str]],
+    expected: str | None,
+    actual: str | None,
+    output_json: bool,
+):
+    """Report a problem, such as wrong, stale or missing results for specific runs.
+
+    Sends the report and shows exactly what Parallel received.
+    """
+    try:
+        response = send_feedback(
+            message=message,
+            category=category,
+            product=product,
+            severity=severity,
+            examples=examples,
+            expected=expected,
+            actual=actual,
+            source="cli",
+        )
+    except FeedbackInputError as e:
+        _handle_error(e, output_json=output_json, exit_code=EXIT_BAD_INPUT, prefix="Invalid feedback")
+        return
+    except Exception as e:
+        _handle_error(e, output_json=output_json, prefix="Feedback API error")
+        return
+
+    if output_json:
+        print(json.dumps(response, indent=2))
+    else:
+        _render_feedback_result(response)
+
+
 # =============================================================================
 # Search Command
 # =============================================================================
diff --git a/parallel_web_tools/core/feedback.py b/parallel_web_tools/core/feedback.py
new file mode 100644
index 0000000..5f9a1ca
--- /dev/null
+++ b/parallel_web_tools/core/feedback.py
@@ -0,0 +1,107 @@
+"""Send feedback about Parallel to the Parallel team."""
+
+from __future__ import annotations
+
+import re
+from typing import Any, Literal
+
+from parallel_web_tools.core.auth import create_client
+from parallel_web_tools.core.user_agent import ClientSource
+
+FeedbackCategory = Literal["results_quality", "error", "latency", "missing_capability", "docs", "billing", "other"]
+FeedbackProduct = Literal[
+    "search", "extract", "task", "findall", "monitor", "responses", "decisions", "mcp", "cli", "other"
+]
+FeedbackSeverity = Literal["blocking", "degraded", "minor"]
+
+FEEDBACK_CATEGORIES: tuple[FeedbackCategory, ...] = (
+    "results_quality",
+    "error",
+    "latency",
+    "missing_capability",
+    "docs",
+    "billing",
+    "other",
+)
+FEEDBACK_PRODUCTS: tuple[FeedbackProduct, ...] = (
+    "search",
+    "extract",
+    "task",
+    "findall",
+    "monitor",
+    "responses",
+    "decisions",
+    "mcp",
+    "cli",
+    "other",
+)
+FEEDBACK_SEVERITIES: tuple[FeedbackSeverity, ...] = ("blocking", "degraded", "minor")
+
+MIN_MESSAGE_CHARS = 20
+MAX_MESSAGE_CHARS = 4000
+MAX_EXAMPLES = 50
+MAX_NOTE_CHARS = 1000
+_EXAMPLE_ID_RE = re.compile(r"^[A-Za-z0-9_-]{1,128}$")
+
+_FEEDBACK_PATH = "/v1beta/feedback"
+
+
+class FeedbackInputError(ValueError):
+    """Raised when a feedback request is invalid before it reaches the API."""
+
+
+def parse_example(raw: str) -> dict[str, str]:
+    """Parse ``ID`` or ``ID:note`` into an example; IDs never contain a colon."""
+    example_id, sep, note = raw.partition(":")
+    example_id = example_id.strip()
+    if not _EXAMPLE_ID_RE.fullmatch(example_id):
+        raise FeedbackInputError(
+            f"example {raw!r} must start with a Parallel ID such as search_..., extract_..., or trun_..."
+        )
+    note = note.strip()
+    if len(note) > MAX_NOTE_CHARS:
+        raise FeedbackInputError(f"example note must be at most {MAX_NOTE_CHARS} characters")
+    return {"id": example_id, "issue": note} if sep and note else {"id": example_id}
+
+
+def _check_choice(value: str, choices: tuple[str, ...], field: str) -> None:
+    if value not in choices:
+        raise FeedbackInputError(f"{field} must be one of: {', '.join(choices)}")
+
+
+def send_feedback(
+    message: str,
+    category: str,
+    product: str,
+    *,
+    severity: str = "degraded",
+    examples: list[dict[str, str]] | None = None,
+    expected: str | None = None,
+    actual: str | None = None,
+    api_key: str | None = None,
+    source: ClientSource = "python",
+) -> dict[str, Any]:
+    """Send a feedback report; the response shows exactly what Parallel received."""
+    if not MIN_MESSAGE_CHARS <= len(message) <= MAX_MESSAGE_CHARS:
+        raise FeedbackInputError(f"message must be {MIN_MESSAGE_CHARS}-{MAX_MESSAGE_CHARS} characters")
+    _check_choice(category, FEEDBACK_CATEGORIES, "category")
+    _check_choice(product, FEEDBACK_PRODUCTS, "product")
+    _check_choice(severity, FEEDBACK_SEVERITIES, "severity")
+    if examples is not None and len(examples) > MAX_EXAMPLES:
+        raise FeedbackInputError(f"examples must be at most {MAX_EXAMPLES}")
+
+    body: dict[str, Any] = {
+        "message": message,
+        "category": category,
+        "product": product,
+        "severity": severity,
+    }
+    if examples:
+        body["examples"] = examples
+    if expected:
+        body["expected"] = expected
+    if actual:
+        body["actual"] = actual
+
+    client = create_client(api_key, source)
+    return client.post(path=_FEEDBACK_PATH, body=body, cast_to=dict)
diff --git a/tests/test_feedback.py b/tests/test_feedback.py
new file mode 100644
index 0000000..21bf6dc
--- /dev/null
+++ b/tests/test_feedback.py
@@ -0,0 +1,140 @@
+"""Tests for sending feedback to Parallel from the CLI."""
+
+from __future__ import annotations
+
+import json
+from typing import Any
+from unittest import mock
+
+import pytest
+from click.testing import CliRunner
+
+from parallel_web_tools.cli.commands import main
+from parallel_web_tools.core.feedback import FeedbackInputError, parse_example, send_feedback
+
+_MESSAGE = "Search results for our eval queries are mostly irrelevant."
+_RECEIVED = {
+    "status": "received",
+    "feedback_id": "feedback_1",
+    "duplicate": False,
+    "report": {
+        "message": _MESSAGE,
+        "category": "results_quality",
+        "product": "search",
+        "severity": "degraded",
+        "examples": [{"id": "search_abc", "issue": "Off topic", "expected": None}],
+        "expected": None,
+        "actual": None,
+        "context": {"org_id": "org", "sender": "user u1 (personal credential)", "surface": "cli"},
+        "redactions": 0,
+        "recipient": "Parallel's team only.",
+    },
+    "next_step": "Sent.",
+}
+
+
+@pytest.fixture
+def runner() -> CliRunner:
+    return CliRunner()
+
+
+@pytest.fixture
+def client():
+    client = mock.MagicMock()
+    with mock.patch("parallel_web_tools.core.feedback.create_client", return_value=client):
+        yield client
+
+
+def _sent_bodies(client) -> list[dict]:
+    return [call.kwargs["body"] for call in client.post.call_args_list]
+
+
+class TestValidation:
+    @pytest.mark.parametrize(
+        ("kwargs", "message"),
+        [
+            ({"message": "too short"}, "message"),
+            ({"category": "praise"}, "category"),
+            ({"product": "spreadsheet"}, "product"),
+            ({"severity": "urgent"}, "severity"),
+            ({"examples": [{"id": f"trun_{n}"} for n in range(51)]}, "examples"),
+        ],
+    )
+    def test_rejects_invalid_reports_before_calling_the_api(self, client, kwargs, message):
+        fields: dict[str, Any] = {"message": _MESSAGE, "category": "results_quality", "product": "search", **kwargs}
+        with pytest.raises(FeedbackInputError, match=message):
+            send_feedback(**fields)
+        client.post.assert_not_called()
+
+    @pytest.mark.parametrize(
+        ("raw", "expected"),
+        [
+            ("trun_abc", {"id": "trun_abc"}),
+            ("search_abc:Returned: nothing useful", {"id": "search_abc", "issue": "Returned: nothing useful"}),
+        ],
+    )
+    def test_parses_examples(self, raw, expected):
+        assert parse_example(raw) == expected
+
+    def test_rejects_an_example_that_is_not_an_id(self):
+        with pytest.raises(FeedbackInputError, match="example"):
+            parse_example("not an id:note")
+
+
+class TestApi:
+    def test_send_posts_only_the_fields_given(self, client):
+        client.post.return_value = _RECEIVED
+        result = send_feedback(
+            message=_MESSAGE,
+            category="results_quality",
+            product="search",
+            examples=[{"id": "search_abc", "issue": "Off topic"}],
+            source="cli",
+        )
+        assert result == _RECEIVED
+        client.post.assert_called_once_with(
+            path="/v1beta/feedback",
+            body={
+                "message": _MESSAGE,
+                "category": "results_quality",
+                "product": "search",
+                "severity": "degraded",
+                "examples": [{"id": "search_abc", "issue": "Off topic"}],
+            },
+            cast_to=dict,
+        )
+
+
+class TestSendCommand:
+    _ARGS = ["feedback", "send", "-m", _MESSAGE, "--category", "results_quality", "--product", "search"]
+
+    def test_json_sends_once_and_prints_the_response(self, runner, client):
+        client.post.return_value = _RECEIVED
+        result = runner.invoke(main, [*self._ARGS, "--example", "search_abc:Off topic", "--json"])
+        assert result.exit_code == 0, result.output
+        assert json.loads(result.output)["feedback_id"] == "feedback_1"
+        assert len(_sent_bodies(client)) == 1
+
+    def test_shows_what_parallel_received(self, runner, client):
+        client.post.return_value = _RECEIVED
+        result = runner.invoke(main, self._ARGS)
+        assert result.exit_code == 0, result.output
+        assert len(_sent_bodies(client)) == 1
+        assert "feedback_1" in result.output
+        assert _MESSAGE in result.output
+        assert "search_abc" in result.output
+
+    def test_a_repeat_says_so(self, runner, client):
+        client.post.return_value = {**_RECEIVED, "duplicate": True}
+        result = runner.invoke(main, self._ARGS)
+        assert result.exit_code == 0, result.output
+        assert "already received" in result.output
+
+    def test_a_bad_example_is_a_usage_error(self, runner, client):
+        result = runner.invoke(main, [*self._ARGS, "--example", "has spaces", "--json"])
+        assert result.exit_code != 0
+        client.post.assert_not_called()
+
+    def test_there_is_no_consent_command(self, runner, client):
+        result = runner.invoke(main, ["feedback", "consent"])
+        assert result.exit_code != 0