hermes: recover incomplete routed tool calls

services/hermes/agent-deployment.yaml

@@ -25,7 +25,7 @@ spec:
         ai.bstein.dev/execution: Hermes Kanban with durable direct Codex and Claude Code CLI workers
         ai.bstein.dev/model-policy: Jetson-assisted AUTO routing, low through xhigh, cross-provider fallback
         ai.bstein.dev/placement: rpi5 preferred; Jetson deferred until state storage is available
-        ai.bstein.dev/config-rev: "20260812-provider-status-refreshable"
+        ai.bstein.dev/config-rev: "20260812-codex-incomplete-failover"
         vault.hashicorp.com/agent-inject: "true"
         vault.hashicorp.com/role: hermes-agent
         vault.hashicorp.com/agent-inject-secret-anthropic-token: kv/data/atlas/hermes/agent-tokens

services/hermes/chat-configmap.yaml

@@ -128,10 +128,12 @@ data:
     When a user asks to create or edit an image, use an image generation tool.
     Treat natural follow-ups such as "edit this", "turn it into", "make this
     image", or a reference to the subject in the most recent generated image
-    as image-edit requests. For those turns, reuse the newest `MEDIA:` image
+    as image-edit requests. For those turns use `image_edit_latest`,
-    path in the conversation as `image_url`; do not answer with instructions,
+    `image_edit_latest_local`, or `image_edit_latest_hosted`; these compact tools
-    route the request as ordinary text, or require the user to upload the image
+    resolve the newest private image on the server, so never reproduce its long
-    again. Preserve the most recently selected image lane for an edit unless
+    `MEDIA:` path in tool arguments. Do not answer with instructions, route the
+    request as ordinary text, or require the user to upload the image again.
+    Preserve the most recently selected image lane for an edit unless
     the user explicitly requests local, OpenAI/hosted, or AUTO instead.
     Use `image_generate_local` when the request says local, private, on my
     hardware, or FLUX. Use `image_generate_hosted` when the request says
@@ -163,7 +165,8 @@ data:
     present a single final answer.
     Use the image generation tool for natural-language image creation and
     editing requests; generated images remain in this tenant's private cache.
-    Natural follow-ups that refer to the latest generated image must edit its
+    Natural follow-ups that refer to the latest generated image must use a
-    newest `MEDIA:` path rather than starting an unrelated text-only answer.
+    compact `image_edit_latest*` tool rather than starting an unrelated
+    text-only answer or copying a long `MEDIA:` path into tool arguments.
     Do not claim access to Kubernetes, Vault, Gitea, Brad's projects, other
     users, the agent coordinator, or automated triage.

services/hermes/chat-statefulset.yaml

@@ -29,7 +29,7 @@ spec:
         ai.bstein.dev/router-wire-contract: ollama-numeric-keepalive
         ai.bstein.dev/isolation: one Hermes process and PVC per Keycloak subject
         ai.bstein.dev/model-policy: uniform automatic policy with per-user overrides
-        ai.bstein.dev/config-rev: "20260812-image-reference-normalization"
+        ai.bstein.dev/config-rev: "20260812-compact-image-edits"
         vault.hashicorp.com/agent-inject: "true"
         vault.hashicorp.com/role: hermes-chat
         vault.hashicorp.com/agent-inject-secret-anthropic-token: kv/data/atlas/hermes/agent-tokens

services/hermes/plugins/image-gen-broker/__init__.py

@@ -74,6 +74,23 @@ IMAGE_GENERATE_PARAMETERS = {
     "required": ["prompt"],
 }
 
+IMAGE_EDIT_PARAMETERS = {
+    "type": "object",
+    "properties": {
+        "prompt": {
+            "type": "string",
+            "description": "Detailed description of the requested edit.",
+        },
+        "aspect_ratio": {
+            "type": "string",
+            "enum": ["landscape", "square", "portrait"],
+            "default": DEFAULT_ASPECT_RATIO,
+            "description": "Requested output aspect ratio.",
+        },
+    },
+    "required": ["prompt"],
+}
+
 LOCAL_IMAGE_SCHEMA = {
     "name": "image_generate_local",
     "description": (
@@ -104,6 +121,42 @@ HOSTED_IMAGE_SCHEMA = {
     "parameters": IMAGE_GENERATE_PARAMETERS,
 }
 
+AUTO_EDIT_SCHEMA = {
+    "name": "image_edit_latest",
+    "description": (
+        "Edit the newest generated image in this private conversation using "
+        "the automatic image route. Use this compact tool for natural "
+        "follow-ups such as turn this cat into a clown, change it, edit the "
+        "image, or make the pictured subject different when the user does not "
+        "name a provider. The server resolves the source image; do not copy a "
+        "MEDIA path into the tool call. Hosted quality is tried first and local "
+        "FLUX is the fallback."
+    ),
+    "parameters": IMAGE_EDIT_PARAMETERS,
+}
+
+LOCAL_EDIT_SCHEMA = {
+    "name": "image_edit_latest_local",
+    "description": (
+        "Edit the newest generated image using only private local FLUX. Use "
+        "when an edit follow-up says local, private, FLUX, or on my hardware. "
+        "The server resolves the source image; pass only the edit prompt and "
+        "optional aspect ratio."
+    ),
+    "parameters": IMAGE_EDIT_PARAMETERS,
+}
+
+HOSTED_EDIT_SCHEMA = {
+    "name": "image_edit_latest_hosted",
+    "description": (
+        "Edit the newest generated image using only hosted OpenAI GPT Image at "
+        "highest quality. Use when an edit follow-up says OpenAI, GPT Image, "
+        "hosted, or highest hosted quality. The server resolves the source "
+        "image; pass only the edit prompt and optional aspect ratio."
+    ),
+    "parameters": IMAGE_EDIT_PARAMETERS,
+}
+
 
 def _broker_key() -> str:
     """Load the internal relay key from process env or the private .env."""
@@ -184,6 +237,29 @@ def _local_image_data_url(value: str) -> str:
     return f"data:{mime};base64,{base64.b64encode(raw).decode('ascii')}"
 
 
+def _latest_generated_image() -> str:
+    """Return the newest generated artifact from this tenant's private cache."""
+    home = Path(os.environ.get("HERMES_HOME", "/opt/data")).resolve()
+    cache = (home / "cache" / "images").resolve()
+    if not cache.is_dir():
+        raise ValueError("no generated image is available to edit")
+    candidates: list[Path] = []
+    for path in cache.iterdir():
+        try:
+            resolved = path.resolve(strict=True)
+            resolved.relative_to(cache)
+        except (OSError, ValueError):
+            continue
+        if (
+            resolved.is_file()
+            and resolved.suffix.lower() in {".png", ".jpg", ".jpeg", ".gif", ".webp"}
+        ):
+            candidates.append(resolved)
+    if not candidates:
+        raise ValueError("no generated image is available to edit")
+    return str(max(candidates, key=lambda item: item.stat().st_mtime_ns))
+
+
 class AtlasBrokerImageProvider(ImageGenProvider):
     """High-quality GPT Image generation through the owner-only broker."""
 
@@ -343,13 +419,45 @@ def _handle_hosted_image(args: dict[str, Any], **_kwargs: Any) -> str:
     return _handle_image_generate(args, "hosted")
 
 
+def _handle_latest_edit(args: dict[str, Any], route: str) -> str:
+    """Edit the newest tenant artifact without model-generated path arguments."""
+    try:
+        image_url = _latest_generated_image()
+    except ValueError as exc:
+        return json.dumps(
+            {
+                "success": False,
+                "image": None,
+                "error": str(exc),
+                "error_type": "missing_reference",
+                "requested_route": route,
+            }
+        )
+    return _handle_image_generate({**args, "image_url": image_url}, route)
+
+
+def _handle_auto_edit(args: dict[str, Any], **_kwargs: Any) -> str:
+    """Edit the newest image with hosted-first automatic failover."""
+    return _handle_latest_edit(args, "auto")
+
+
+def _handle_local_edit(args: dict[str, Any], **_kwargs: Any) -> str:
+    """Edit the newest image without hosted provider substitution."""
+    return _handle_latest_edit(args, "local")
+
+
+def _handle_hosted_edit(args: dict[str, Any], **_kwargs: Any) -> str:
+    """Edit the newest image only with hosted GPT Image."""
+    return _handle_latest_edit(args, "hosted")
+
+
 def _image_tool_available() -> bool:
     """Expose the tool when at least one broker route is healthy."""
     return AtlasBrokerImageProvider().is_available()
 
 
 def register(ctx: Any) -> None:
-    """Register the broker and explicit local/hosted image tools."""
+    """Register generation routes and compact latest-image edit tools."""
     ctx.register_image_gen_provider(AtlasBrokerImageProvider())
     ctx.register_tool(
         name="image_generate_local",
@@ -373,3 +481,19 @@ def register(ctx: Any) -> None:
         description=HOSTED_IMAGE_SCHEMA["description"],
         emoji="🎨",
     )
+    for name, schema, handler in (
+        ("image_edit_latest", AUTO_EDIT_SCHEMA, _handle_auto_edit),
+        ("image_edit_latest_local", LOCAL_EDIT_SCHEMA, _handle_local_edit),
+        ("image_edit_latest_hosted", HOSTED_EDIT_SCHEMA, _handle_hosted_edit),
+    ):
+        ctx.register_tool(
+            name=name,
+            toolset="image_gen",
+            schema=schema,
+            handler=handler,
+            check_fn=_image_tool_available,
+            requires_env=[],
+            is_async=False,
+            description=schema["description"],
+            emoji="🎨",
+        )

services/hermes/scripts/codex_broker.py

@@ -26,6 +26,9 @@ UPSTREAM = os.environ.get(
     "https://chatgpt.com/backend-api/codex",
 ).rstrip("/")
 MAX_BODY_BYTES = int(os.environ.get("HERMES_CODEX_BROKER_MAX_BODY", str(64 << 20)))
+MAX_RESPONSE_BYTES = int(
+    os.environ.get("HERMES_CODEX_BROKER_MAX_RESPONSE", str(64 << 20))
+)
 READ_TIMEOUT_SECONDS = float(os.environ.get("HERMES_CODEX_BROKER_READ_TIMEOUT", "900"))
 FALLBACK_ALLOWED_MODELS = {
     value.strip()
@@ -129,12 +132,24 @@ def _validate_payload(payload: Any) -> dict[str, Any]:
     # selected provider request.  The first-party subscription Codex endpoint
     # does not accept any of the public API token-budget aliases.  Let Codex use
     # its own response budget instead of turning a healthy fallback into a 400.
-    for token_budget_key in (
+    for unsupported_key in (
         "max_output_tokens",
         "max_completion_tokens",
         "max_tokens",
+        # The subscription Codex backend owns sampling.  OpenAI-compatible
+        # clients may add these public API fields during a retry; forwarding
+        # them turns an otherwise healthy fallback into HTTP 400.
+        "temperature",
+        "top_p",
+        "frequency_penalty",
+        "presence_penalty",
+        "logprobs",
+        "top_logprobs",
+        "seed",
+        "n",
+        "stop",
     ):
-        payload.pop(token_budget_key, None)
+        payload.pop(unsupported_key, None)
     # Tenant conversations must not enter the owner's server-side history.
     payload["store"] = False
     payload["stream"] = True
@@ -178,6 +193,18 @@ def _completed_response(lines: Iterable[str]) -> dict[str, Any]:
             else:
                 upstream_error = str(error or "")
     if terminal_response is not None:
+        status = str(terminal_response.get("status") or "").lower()
+        if status != "completed":
+            details = terminal_response.get("incomplete_details") or {}
+            reason = (
+                str(details.get("reason") or "")
+                if isinstance(details, dict)
+                else ""
+            )
+            raise RuntimeError(
+                "Codex returned a retryable incomplete response"
+                + (f": {reason}" if reason else "")
+            )
         # The subscription Codex endpoint streams complete output items but
         # currently leaves the terminal response's output array empty.  Public
         # Responses clients, including Switchyard, expect those items there.
@@ -273,22 +300,34 @@ class Handler(BaseHTTPRequestHandler):
                         self.wfile.write(body)
                         return
 
+                    # Validate the terminal Responses event before committing
+                    # HTTP 200 downstream.  The Codex subscription endpoint can
+                    # end a tool call with ``response.incomplete`` while still
+                    # returning HTTP 200.  If that is streamed through, Hermes
+                    # sees malformed JSON and eventually exposes "Response
+                    # truncated" to the user.  Buffering one model boundary
+                    # lets Switchyard receive a retryable 502 and choose another
+                    # provider/model instead.  Tool execution remains streamed
+                    # by Hermes after this short model boundary completes.
+                    body = response.read()
+                    if len(body) > MAX_RESPONSE_BYTES:
+                        raise RuntimeError("Codex response exceeded broker limit")
+                    completed = _completed_response(
+                        body.decode("utf-8", errors="replace").splitlines()
+                    )
                     if not requested_stream:
-                        self._json(200, _completed_response(response.iter_lines()))
+                        self._json(200, completed)
                         return
 
-                    # HTTP/1.0 close-delimited streaming avoids buffering a
-                    # potentially long tool-calling turn in the broker.
                     self.send_response(200)
                     self.send_header("Content-Type", "text/event-stream")
+                    self.send_header("Content-Length", str(len(body)))
                     self.send_header("Cache-Control", "no-store")
                     self.send_header("X-Accel-Buffering", "no")
                     self.end_headers()
                     response_started = True
-                    for chunk in response.iter_raw():
+                    self.wfile.write(body)
-                        if chunk:
+                    self.wfile.flush()
-                            self.wfile.write(chunk)
-                            self.wfile.flush()
         except ValueError as exc:
             self._json(400, {"error": {"message": str(exc), "type": "invalid_request_error"}})
         except (BrokenPipeError, ConnectionResetError):

testing/tests/test_hermes_chat_quality.py

@@ -352,7 +352,8 @@ def test_chat_image_generation_uses_private_owner_broker():
     assert "Use `image_generate_local`" in configmap["data"]["SOUL.md"]
     assert "Use `image_generate_hosted`" in configmap["data"]["SOUL.md"]
     assert "ComfyUI endpoint" in configmap["data"]["SOUL.md"]
-    assert "newest `MEDIA:` image" in configmap["data"]["SOUL.md"]
+    assert "`MEDIA:` path" in configmap["data"]["SOUL.md"]
+    assert "`image_edit_latest`" in configmap["data"]["SOUL.md"]
     assert "Preserve the most recently selected image lane" in configmap["data"]["SOUL.md"]
     config = yaml.safe_load(configmap["data"]["config.yaml"])
     assert config["image_gen"] == {
@@ -378,6 +379,10 @@ def test_chat_image_generation_uses_private_owner_broker():
     assert '"hosted": "gpt-image-2-high"' in plugin
     assert 'name="image_generate_local"' in plugin
     assert 'name="image_generate_hosted"' in plugin
+    assert '"name": "image_edit_latest"' in plugin
+    assert '"name": "image_edit_latest_local"' in plugin
+    assert '"name": "image_edit_latest_hosted"' in plugin
+    assert "def _latest_generated_image" in plugin
     assert "newest MEDIA: path from the conversation" in plugin
     assert 'candidate.upper().startswith("MEDIA:")' in plugin
     assert "override=True" not in plugin
@@ -421,6 +426,56 @@ def test_chat_image_generation_uses_private_owner_broker():
     )
 
 
+def test_compact_image_edit_resolves_latest_tenant_artifact(tmp_path, monkeypatch):
+    """Follow-up edits resolve the source server-side and keep tool JSON small."""
+    provider_module = SimpleNamespace(
+        DEFAULT_ASPECT_RATIO="square",
+        ImageGenProvider=object,
+        error_response=lambda **value: value,
+        normalize_reference_images=lambda value: value,
+        resolve_aspect_ratio=lambda value: value,
+        save_b64_image=lambda *_args, **_kwargs: tmp_path / "saved.png",
+        success_response=lambda **value: value,
+    )
+    monkeypatch.setitem(sys.modules, "agent", SimpleNamespace())
+    monkeypatch.setitem(sys.modules, "agent.image_gen_provider", provider_module)
+    spec = importlib.util.spec_from_file_location(
+        "hermes_image_plugin",
+        HERMES / "plugins" / "image-gen-broker" / "__init__.py",
+    )
+    assert spec and spec.loader
+    module = importlib.util.module_from_spec(spec)
+    spec.loader.exec_module(module)
+
+    image_dir = tmp_path / "cache" / "images"
+    image_dir.mkdir(parents=True)
+    older = image_dir / "atlas_flux-old.png"
+    newest = image_dir / "atlas_gpt-image-new.png"
+    older.write_bytes(b"older")
+    newest.write_bytes(b"newest")
+    older.touch()
+    time.sleep(0.001)
+    newest.touch()
+    monkeypatch.setenv("HERMES_HOME", str(tmp_path))
+
+    calls = []
+    monkeypatch.setattr(
+        module,
+        "_handle_image_generate",
+        lambda args, route: calls.append((args, route)) or "ok",
+    )
+    assert module._handle_hosted_edit({"prompt": "make it a clown"}) == "ok"
+    assert calls == [
+        (
+            {
+                "prompt": "make it a clown",
+                "image_url": str(newest.resolve()),
+            },
+            "hosted",
+        )
+    ]
+
+
 def test_chat_reasoning_uses_switchyard_without_owner_credentials():
     """Family pods use AUTO/manual routes without mounting owner credentials."""
     configmap = _documents(HERMES / "chat-configmap.yaml")[0]
@@ -485,7 +540,7 @@ def test_chat_reasoning_uses_switchyard_without_owner_credentials():
     statefulset = _documents(HERMES / "chat-statefulset.yaml")[0]
     assert statefulset["spec"]["template"]["metadata"]["annotations"][
         "ai.bstein.dev/config-rev"
-    ] == "20260812-image-reference-normalization"
+    ] == "20260812-compact-image-edits"
     hermes = next(
         item
         for item in statefulset["spec"]["template"]["spec"]["containers"]
@@ -529,6 +584,8 @@ def test_codex_broker_auth_and_request_contract(tmp_path: Path, monkeypatch):
             "max_output_tokens": 96,
             "max_completion_tokens": 96,
             "max_tokens": 96,
+            "temperature": 0.7,
+            "top_p": 0.9,
         }
     )
     assert payload["store"] is False
@@ -536,6 +593,8 @@ def test_codex_broker_auth_and_request_contract(tmp_path: Path, monkeypatch):
     assert "max_output_tokens" not in payload
     assert "max_completion_tokens" not in payload
     assert "max_tokens" not in payload
+    assert "temperature" not in payload
+    assert "top_p" not in payload
     assert payload["input"] == [
         {
             "type": "message",
@@ -601,6 +660,15 @@ def test_codex_broker_auth_and_request_contract(tmp_path: Path, monkeypatch):
             "data: [DONE]",
         ]
     )["output"] == [completed_item]
+    with pytest.raises(RuntimeError, match="retryable incomplete response"):
+        module._completed_response(
+            [
+                "event: response.incomplete",
+                'data: {"type":"response.incomplete","response":'
+                '{"status":"incomplete","incomplete_details":'
+                '{"reason":"max_output_tokens"}}}',
+            ]
+        )
     with pytest.raises(RuntimeError, match="provider unavailable"):
         module._completed_response(
             [