diff --git a/openapi.yml b/openapi.yml
index 909544c..03b5000 100644
--- a/openapi.yml
+++ b/openapi.yml
@@ -12,6 +12,16 @@ info:
     - **Evaluate Task**: automatic feedback and grading for student submissions. <br>
     - **Chat**: conversational interactions around tasks, submissions, or general
       learning questions.
+    <br><br>
+    **Progress streaming.** Any operation MAY support an opt-in Server-Sent
+    Events progress stream, selected per request with the
+    `Accept: text/event-stream` header. The stream is a sequence of
+    `SseProgressStep` frames, optional keep-alive comments, and exactly one
+    terminal frame carrying the operation's normal `200` body (on success) or an
+    `ErrorResponse` (on failure); the HTTP status stays `200` throughout.
+    Services advertise support via `supportsStreaming` in their `/health`
+    capabilities. This is distinct from `configuration.llm.stream`, which
+    governs token-level streaming from the LLM provider.
 
 tags:
   - name: evaluate
diff --git a/paths/chat/chat-health.yml b/paths/chat/chat-health.yml
index 4d3ec90..8401390 100644
--- a/paths/chat/chat-health.yml
+++ b/paths/chat/chat-health.yml
@@ -37,6 +37,12 @@ get:
                   supportsChat: true
                   supportsUserPreferences: true
                   supportsStreaming: true
+                  supportedProgressStages:
+                    - "preparing"
+                    - "starting"
+                    - "thinking"
+                    - "completed"
+                    - "failed"
                   "supportsDataPolicy": "NOT_SUPPORTED"
                   supportedLanguages:
                     - "en"
diff --git a/paths/chat/chat.yml b/paths/chat/chat.yml
index ff3dd0c..a475199 100644
--- a/paths/chat/chat.yml
+++ b/paths/chat/chat.yml
@@ -139,7 +139,35 @@ post:
                   temperature: 0.5
   responses:
     "200":
-      description: Successful chat response.
+      description: >
+        Successful chat response.
+
+
+        ### Streaming variant (opt-in)
+
+
+        If the request sends `Accept: text/event-stream`, the response is a
+        Server-Sent Events progress stream instead of a single JSON body. The
+        HTTP status stays `200` for the entire stream, including failures. It
+        consists of:
+
+
+        - zero or more progress frames — the SSE `event:` field is the stage
+          name and `data:` is an `SseProgressStep` JSON object;
+
+        - optional keep-alive comment lines (each starting with `:`);
+
+        - exactly one terminal frame — `event: completed` with `data:` an
+          `SseChatTerminalFrame` (the `200` body as `output` / `metadata`, plus
+          a `steps` replay), or `event: failed` with `data:` an
+          `SseChatTerminalFrame` whose `error` is an `ErrorResponse`.
+
+
+        Failures are always delivered as a terminal `failed` frame, never as an
+        HTTP error status. `X-Request-Id` and `X-Api-Version` are sent once, as
+        response headers, when the stream opens. This is independent of
+        `configuration.llm.stream`, which governs token-level streaming from the
+        LLM provider.
       headers:
         X-Request-Id:
           description: Request id for tracing this request across services.
@@ -213,6 +241,40 @@ post:
                     model: "gpt-5.2"
                     temperature: 0.5
                     outputTokens: 143
+        text/event-stream:
+          schema:
+            $ref: "./schemas/SseChatTerminalFrame.yml"
+          examples:
+            completedFrame:
+              summary: >
+                Terminal `event: completed` frame `data` (preceded on the wire by
+                `event: preparing` / `event: thinking` progress frames whose
+                `data` is an SseProgressStep, and optional `:` keep-alive lines).
+              value:
+                output:
+                  role: ASSISTANT
+                  content: "Polymorphism lets one interface represent many underlying forms."
+                metadata:
+                  responseTimeMs: 1800
+                steps:
+                  - stage: "preparing"
+                    message: "Preparing worker"
+                    timestamp: "2025-12-10T11:15:00Z"
+                  - stage: "thinking"
+                    message: "Composing response"
+                    timestamp: "2025-12-10T11:15:02Z"
+            failedFrame:
+              summary: "Terminal `event: failed` frame `data` (HTTP status is still 200)."
+              value:
+                output: null
+                error:
+                  title: "Chat failed"
+                  message: "The LLM provider connection failed."
+                  code: "LLM_PROVIDER_ERROR"
+                steps:
+                  - stage: "preparing"
+                    message: "Preparing worker"
+                    timestamp: "2025-12-10T11:15:00Z"
     "400":
       $ref: "./responses/400-BadRequest.yml"
     "403":
diff --git a/paths/chat/schemas/ChatCapabilities.yml b/paths/chat/schemas/ChatCapabilities.yml
index 08cf660..d8ffbb7 100644
--- a/paths/chat/schemas/ChatCapabilities.yml
+++ b/paths/chat/schemas/ChatCapabilities.yml
@@ -4,6 +4,8 @@ required:
   - supportsChat
   - supportsDataPolicy
 additionalProperties: true
+allOf:
+  - $ref: "../../shared/schemas/StreamingCapabilities.yml"
 properties:
   supportsChat:
     type: boolean
@@ -11,9 +13,6 @@ properties:
   supportsUserPreferences:
     type: boolean
     description: Indicates whether the service supports adapting to user preferences.
-  supportsStreaming:
-    type: boolean
-    description: Indicates whether the service supports streaming responses.
   supportsDataPolicy:
     $ref: "../../shared/enums/DataPolicySupport.yml"
   supportedLanguages:
diff --git a/paths/chat/schemas/SseChatTerminalFrame.yml b/paths/chat/schemas/SseChatTerminalFrame.yml
new file mode 100644
index 0000000..a852b6a
--- /dev/null
+++ b/paths/chat/schemas/SseChatTerminalFrame.yml
@@ -0,0 +1,36 @@
+type: object
+description: >
+  Data payload of the single terminal frame of a `POST /chat` SSE progress
+  stream.
+
+  On `event: completed`, `output` / `metadata` hold exactly the endpoint's
+  normal `200 application/json` body (`ChatResponse`) and `error` is absent.
+
+  On `event: failed`, `output` is null and `error` holds an `ErrorResponse` with
+  the same `title` / `code` / `details` shape clients already handle for `400` /
+  `500` responses. The HTTP status stays `200` regardless.
+
+  `steps` is always present: the ordered replay of every progress step emitted
+  during the stream.
+allOf:
+  - $ref: "../../shared/schemas/SseTerminalSteps.yml"
+  - type: object
+    properties:
+      output:
+        description: The generated assistant response; null on a `failed` frame.
+        anyOf:
+          - $ref: "./Message.yml"
+          - type: "null"
+      metadata:
+        type:
+          - object
+          - "null"
+        description: Optional metadata about response generation.
+        additionalProperties: true
+      error:
+        type:
+          - object
+          - "null"
+        description: Present only on a `failed` frame.
+        allOf:
+          - $ref: "../../shared/schemas/ErrorResponse.yml"
diff --git a/paths/evaluate/evaluate-health.yml b/paths/evaluate/evaluate-health.yml
index 36cacdc..e224d37 100644
--- a/paths/evaluate/evaluate-health.yml
+++ b/paths/evaluate/evaluate-health.yml
@@ -40,6 +40,13 @@ get:
                   supportsFormativeFeedback: true
                   supportsSummativeFeedback: true
                   supportsDataPolicy: "PARTIAL"
+                  supportsStreaming: true
+                  supportedProgressStages:
+                    - "preparing"
+                    - "starting"
+                    - "evaluating"
+                    - "completed"
+                    - "failed"
                   supportedArtefactProfiles:
                     - type: "TEXT"
                       supportedFormats: ["plain", "markdown"]
diff --git a/paths/evaluate/evaluate.yml b/paths/evaluate/evaluate.yml
index 77818a0..57d2ff6 100644
--- a/paths/evaluate/evaluate.yml
+++ b/paths/evaluate/evaluate.yml
@@ -339,7 +339,35 @@ post:
                 version: 1
   responses:
     "200":
-      description: Successfully generated feedback.
+      description: >
+        Successfully generated feedback.
+
+
+        ### Streaming variant (opt-in)
+
+
+        If the request sends `Accept: text/event-stream`, the response is a
+        Server-Sent Events progress stream instead of a single JSON body. The
+        HTTP status stays `200` for the entire stream, including failures. It
+        consists of:
+
+
+        - zero or more progress frames — the SSE `event:` field is the stage
+          name and `data:` is an `SseProgressStep` JSON object;
+
+        - optional keep-alive comment lines (each starting with `:`);
+
+        - exactly one terminal frame — `event: completed` with `data:` an
+          `SseEvaluateTerminalFrame` (the `200` body under `feedback`, plus a
+          `steps` replay), or `event: failed` with `data:` an
+          `SseEvaluateTerminalFrame` whose `error` is an `ErrorResponse`.
+
+
+        Failures are always delivered as a terminal `failed` frame, never as an
+        HTTP error status. `X-Request-Id` and `X-Api-Version` are sent once, as
+        response headers, when the stream opens. This is independent of
+        `configuration.llm.stream`, which governs token-level streaming from the
+        LLM provider.
       headers:
         X-Request-Id:
           description: Request id for tracing this request across services.
@@ -352,9 +380,7 @@ post:
       content:
         application/json:
           schema:
-            type: array
-            items:
-              $ref: "./schemas/Feedback.yml"
+            $ref: "./schemas/EvaluateResponse.yml"
           examples:
             exampleResponse:
               summary: Example feedback response
@@ -379,6 +405,40 @@ post:
                 - feedbackId: "fb-2"
                   title: "Overall structure"
                   message: "The overall structure of your answer is clear and easy to follow."
+        text/event-stream:
+          schema:
+            $ref: "./schemas/SseEvaluateTerminalFrame.yml"
+          examples:
+            completedFrame:
+              summary: >
+                Terminal `event: completed` frame `data` (preceded on the wire by
+                `event: preparing` / `event: evaluating` progress frames whose
+                `data` is an SseProgressStep, and optional `:` keep-alive lines).
+              value:
+                feedback:
+                  - feedbackId: "fb-1"
+                    title: "Clarify your definition"
+                    message: "Distinguish subtype from parametric polymorphism."
+                    awardedPoints: 2.5
+                steps:
+                  - stage: "preparing"
+                    message: "Preparing worker"
+                    timestamp: "2025-12-16T09:30:01Z"
+                  - stage: "evaluating"
+                    message: "Ran 3/10 checks"
+                    timestamp: "2025-12-16T09:30:04Z"
+            failedFrame:
+              summary: "Terminal `event: failed` frame `data` (HTTP status is still 200)."
+              value:
+                feedback: null
+                error:
+                  title: "Evaluation failed"
+                  message: "The worker exited before returning feedback."
+                  code: "LLM_PROVIDER_ERROR"
+                steps:
+                  - stage: "preparing"
+                    message: "Preparing worker"
+                    timestamp: "2025-12-16T09:30:01Z"
     "202":
       $ref: "./responses/202-Accepted.yml"
     "400":
diff --git a/paths/evaluate/schemas/EvaluateCapabilities.yml b/paths/evaluate/schemas/EvaluateCapabilities.yml
index 92796e9..cb7d3b5 100644
--- a/paths/evaluate/schemas/EvaluateCapabilities.yml
+++ b/paths/evaluate/schemas/EvaluateCapabilities.yml
@@ -7,6 +7,8 @@ required:
   - supportsSummativeFeedback
   - supportsDataPolicy
 additionalProperties: true
+allOf:
+  - $ref: "../../shared/schemas/StreamingCapabilities.yml"
 properties:
   supportsEvaluate:
     type: boolean
diff --git a/paths/evaluate/schemas/EvaluateResponse.yml b/paths/evaluate/schemas/EvaluateResponse.yml
new file mode 100644
index 0000000..0303b1a
--- /dev/null
+++ b/paths/evaluate/schemas/EvaluateResponse.yml
@@ -0,0 +1,4 @@
+type: array
+description: Ordered list of feedback items generated for the submission.
+items:
+  $ref: "./Feedback.yml"
diff --git a/paths/evaluate/schemas/SseEvaluateTerminalFrame.yml b/paths/evaluate/schemas/SseEvaluateTerminalFrame.yml
new file mode 100644
index 0000000..8319adc
--- /dev/null
+++ b/paths/evaluate/schemas/SseEvaluateTerminalFrame.yml
@@ -0,0 +1,32 @@
+type: object
+description: >
+  Data payload of the single terminal frame of a `POST /evaluate` SSE progress
+  stream.
+
+  On `event: completed`, `feedback` holds exactly the endpoint's normal
+  `200 application/json` body and `error` is absent.
+
+  On `event: failed`, `feedback` is null and `error` holds an `ErrorResponse`
+  with the same `title` / `code` / `details` shape clients already handle for
+  `400` / `500` responses. The HTTP status stays `200` regardless.
+
+  `steps` is always present: the ordered replay of every progress step emitted
+  during the stream.
+allOf:
+  - $ref: "../../shared/schemas/SseTerminalSteps.yml"
+  - type: object
+    properties:
+      feedback:
+        type:
+          - array
+          - "null"
+        description: The generated feedback items; null on a `failed` frame.
+        items:
+          $ref: "./Feedback.yml"
+      error:
+        type:
+          - object
+          - "null"
+        description: Present only on a `failed` frame.
+        allOf:
+          - $ref: "../../shared/schemas/ErrorResponse.yml"
diff --git a/paths/shared/schemas/LLMConfiguration.yml b/paths/shared/schemas/LLMConfiguration.yml
index 7829e1e..0d07dea 100644
--- a/paths/shared/schemas/LLMConfiguration.yml
+++ b/paths/shared/schemas/LLMConfiguration.yml
@@ -23,7 +23,12 @@ properties:
     type:
       - boolean
       - "null"
-    description: Optional flag indicating whether streaming responses are requested.
+    description: >
+      Optional flag requesting token-level streaming from the underlying LLM
+      provider. This is distinct from API-layer progress streaming, which is
+      opt-in per request via the `Accept: text/event-stream` header and is
+      advertised by `supportsStreaming` in a service's `/health` capabilities.
+      A service MAY support either, both, or neither.
   credentials:
     type:
       - object
diff --git a/paths/shared/schemas/SseProgressStep.yml b/paths/shared/schemas/SseProgressStep.yml
new file mode 100644
index 0000000..6acce1b
--- /dev/null
+++ b/paths/shared/schemas/SseProgressStep.yml
@@ -0,0 +1,30 @@
+type: object
+description: >
+  A single progress step emitted while an operation runs. Carried as the SSE
+  `data:` payload of an intermediate progress frame (the SSE `event:` field
+  carries the stage name), and replayed in the terminal frame's `steps` array.
+additionalProperties: true
+required:
+  - stage
+  - timestamp
+properties:
+  stage:
+    type: string
+    description: >
+      Lifecycle stage this step reports. Informative, not a fixed enum:
+      implementations MAY introduce new stages without a spec change, and clients
+      MUST tolerate values they do not recognise. Each service advertises the
+      stage values it emits via `supportedProgressStages` in its `/health`
+      capabilities. Common values: "preparing", "starting", "evaluating"
+      (evaluate), "thinking" (chat), and the terminal "completed" / "failed".
+  message:
+    type:
+      - string
+      - "null"
+    description: >
+      Short, human-readable description of the step, suitable for display to a
+      learner or teacher.
+  timestamp:
+    type: string
+    format: date-time
+    description: Time this step was generated (RFC 3339 / ISO 8601).
diff --git a/paths/shared/schemas/SseTerminalSteps.yml b/paths/shared/schemas/SseTerminalSteps.yml
new file mode 100644
index 0000000..9a9240d
--- /dev/null
+++ b/paths/shared/schemas/SseTerminalSteps.yml
@@ -0,0 +1,13 @@
+type: object
+description: >
+  Shared fragment for the terminal SSE frame: the ordered replay of every
+  progress step emitted during the stream, so a client that connected late or
+  dropped intermediate frames still receives the full trace.
+required:
+  - steps
+properties:
+  steps:
+    type: array
+    description: Ordered list of every SseProgressStep emitted during the stream.
+    items:
+      $ref: "./SseProgressStep.yml"
diff --git a/paths/shared/schemas/StreamingCapabilities.yml b/paths/shared/schemas/StreamingCapabilities.yml
new file mode 100644
index 0000000..54422fd
--- /dev/null
+++ b/paths/shared/schemas/StreamingCapabilities.yml
@@ -0,0 +1,26 @@
+type: object
+description: >
+  Shared capability fragment describing an operation's support for opt-in
+  Server-Sent Events (SSE) progress streaming. Composed via `allOf` into each
+  service's capabilities schema so every streaming-capable endpoint advertises
+  streaming the same way.
+additionalProperties: true
+properties:
+  supportsStreaming:
+    type: boolean
+    description: >
+      Indicates whether this operation supports opt-in SSE progress streaming,
+      selected per request with the `Accept: text/event-stream` header. This is
+      distinct from `configuration.llm.stream`, which governs token-level
+      streaming from the underlying LLM provider. A service MAY support either,
+      both, or neither.
+  supportedProgressStages:
+    type:
+      - array
+      - "null"
+    description: >
+      Optional list of `SseProgressStep.stage` values this service may emit
+      (e.g. ["preparing", "starting", "evaluating", "completed", "failed"]).
+      Informative only; clients MUST tolerate stages that are not listed.
+    items:
+      type: string