fix(replay-vision): trim repeated fields from scanner observation rows

products/replay_vision/mcp/tools.yaml

@@ -450,15 +450,34 @@ tools:
             failed — useful for triaging why a scan produced no result), `triggered_by` (`schedule` or `on_demand`),
             `workflow_id`, the `scanner_result` (the LLM output validated against the scanner's response schema for that
             scanner_type), and the `session_id` of the recording it analysed. Use this to inspect what a scanner has
-            found over time; to see every scanner's observations for one session, use `vision-observations-list`.
+            found over time; to see every scanner's observations for one session, use `vision-observations-list`. Rows
+            are trimmed to keep a page small. `scanner_snapshot.scanner_config.prompt` and `scanner_snapshot.query`
+            repeat on every row of one scanner version, so neither is on a row. The citation segments restate the flat
+            text they were split from and are dropped too: `model_output.summary_segments` on a summarizer observation,
+            `model_output.reasoning_segments` on a monitor, classifier, or scorer one. Editing a scanner's prompt or
+            query bumps `scanner_snapshot.scanner_version`, and a page can hold rows from either side of that edit. Call
+            `vision-scanners-observations-get` for one observation's full prompt, query, and citation segments. When a
+            page spans more than one `scanner_version`, read one observation per version before you compare findings.
+            Call `vision-scanners-get` for the scanner's current prompt.
         agent_note: >
-            Each observation's `_posthogUrl` opens the recording it analysed.
-            `scanner_result.model_output.reasoning_segments` interleaves prose with `chip` segments, and a chip's
-            `timestamp_ms` is the recording-relative offset of the moment being cited — append `?t=<seconds>`
-            (`timestamp_ms` / 1000, rounded down) to that URL to seek straight to it. When you report a finding to
-            someone, deep-link the one or two moments it turns on rather than only describing them.
+            Each observation's `_posthogUrl` opens the recording it analysed. When you report a finding to someone,
+            deep-link the one or two moments it turns on rather than only describing them: read that observation with
+            `vision-scanners-observations-get`, take a `chip` segment's `timestamp_ms` from its citation segments, and
+            append `?t=<seconds>` (`timestamp_ms` / 1000, rounded down) to the recording URL. The citation segments are
+            `scanner_result.model_output.summary_segments` on a summarizer observation, and
+            `scanner_result.model_output.reasoning_segments` on a monitor, classifier, or scorer one.
+        param_overrides:
+            limit:
+                description: Number of observations to return per page (default 20).
+                default: 20
         ui_app: vision-observation-list
         url_prefix: /replay
+        response:
+            exclude:
+                - scanner_snapshot.scanner_config.prompt
+                - scanner_snapshot.query
+                - scanner_result.model_output.reasoning_segments
+                - scanner_result.model_output.summary_segments
     vision-scanners-observations-retry-create:
         operation: vision_scanners_observations_retry_create
         enabled: false

products/replay_vision/skills/exploring-replay-vision-observations/SKILL.md

@@ -99,10 +99,14 @@ and, like all observations, runs at most once per `(scanner, session)`.
 
 ### Cite moments, not just sessions
 
-`scanner_result.model_output.reasoning_segments` is the same prose as `reasoning`, pre-split into `text` segments and `chip` segments.
+An observation's citation segments are the same prose as its flat text, pre-split into `text` segments and `chip` segments.
+The field is `scanner_result.model_output.summary_segments` on a summarizer, and `scanner_result.model_output.reasoning_segments` on a monitor, classifier, or scorer.
 Each chip carries a `timestamp_ms`: the recording-relative offset of the moment the model is pointing at.
 That's what makes a finding checkable — it turns "the user hit a paywall" into a link that opens on the paywall.
 
+`vision-scanners-observations-list` leaves the segments off its rows to keep a page small, so read the observation you want to cite with `vision-scanners-observations-get` first.
+`vision-observations-list` and `vision-observations-retrieve` still return them, so a session-scoped page needs no extra call.
+
 The observation's `_posthogUrl` is its recording; append `?t=<seconds>` (`timestamp_ms` / 1000, rounded down) to seek there.
 
 ```text

services/mcp/schema/tool-definitions-all.json

@@ -12301,7 +12301,7 @@
         }
     },
     "vision-scanners-observations-list": {
-        "description": "List the observations recorded by a single scanner. Each observation is one scan of one session and includes its `status` (pending/running/succeeded/failed/ineligible), `error_reason` (set on terminal non-success statuses, e.g. `too_short`/`no_recording` for ineligible or `provider_rejected`/`validation_failed` for failed — useful for triaging why a scan produced no result), `triggered_by` (`schedule` or `on_demand`), `workflow_id`, the `scanner_result` (the LLM output validated against the scanner's response schema for that scanner_type), and the `session_id` of the recording it analysed. Use this to inspect what a scanner has found over time; to see every scanner's observations for one session, use `vision-observations-list`.",
+        "description": "List the observations recorded by a single scanner. Each observation is one scan of one session and includes its `status` (pending/running/succeeded/failed/ineligible), `error_reason` (set on terminal non-success statuses, e.g. `too_short`/`no_recording` for ineligible or `provider_rejected`/`validation_failed` for failed — useful for triaging why a scan produced no result), `triggered_by` (`schedule` or `on_demand`), `workflow_id`, the `scanner_result` (the LLM output validated against the scanner's response schema for that scanner_type), and the `session_id` of the recording it analysed. Use this to inspect what a scanner has found over time; to see every scanner's observations for one session, use `vision-observations-list`. Rows are trimmed to keep a page small. `scanner_snapshot.scanner_config.prompt` and `scanner_snapshot.query` repeat on every row of one scanner version, so neither is on a row. The citation segments restate the flat text they were split from and are dropped too: `model_output.summary_segments` on a summarizer observation, `model_output.reasoning_segments` on a monitor, classifier, or scorer one. Editing a scanner's prompt or query bumps `scanner_snapshot.scanner_version`, and a page can hold rows from either side of that edit. Call `vision-scanners-observations-get` for one observation's full prompt, query, and citation segments. When a page spans more than one `scanner_version`, read one observation per version before you compare findings. Call `vision-scanners-get` for the scanner's current prompt.",
         "category": "Replay vision",
         "feature": "replay_vision",
         "summary": "List scanner observations",

services/mcp/tests/unit/__snapshots__/tool-schemas/vision-scanners-observations-list.json

@@ -18,7 +18,8 @@
             "type": "boolean"
         },
         "limit": {
-            "description": "Number of results to return per page.",
+            "default": 20,
+            "description": "Number of observations to return per page (default 20).",
             "type": "number"
         },
         "max_score": {

services/mcp/tests/unit/vision-scanner-observations-list-trimming.test.ts

@@ -0,0 +1,110 @@
+import { describe, expect, it } from 'vitest'
+
+import { GENERATED_TOOLS } from '@/tools/generated/replay_vision'
+import type { Context, ToolBase, ZodObjectAny } from '@/tools/types'
+
+// One observation row carrying every field the tool's `response.exclude` drops, alongside the
+// siblings it must keep, so a broken path (silent no-op -> full payload back) is visible.
+const FULL_ROW = {
+    id: 'obs-1',
+    session_id: 'sess-1',
+    status: 'succeeded',
+    scanner_snapshot: {
+        name: 'Checkout drop-offs',
+        scanner_type: 'monitor',
+        scanner_version: 3,
+        // Identical on every row of a scanner's page, and the largest field on it.
+        scanner_config: { prompt: 'Did the shopper abandon the cart?', allow_inconclusive: true },
+        query: { kind: 'RecordingsQuery', events: [{ id: '$pageview' }] },
+    },
+    scanner_result: {
+        model_output: {
+            scanner_type: 'monitor',
+            verdict: 'yes',
+            reasoning: 'The shopper opened the cart and left without paying.',
+            // Restates `reasoning` interleaved with timestamp chips.
+            reasoning_segments: [
+                { kind: 'text', value: 'The shopper opened the cart and left without paying.' },
+                { kind: 'chip', timestamp_ms: 42000 },
+            ],
+        },
+        signals_count: 0,
+    },
+}
+
+// A summarizer row: same duplication, under `summary_segments`, because segments are written as
+// `{citation_field}_segments` and summarizer cites `summary` rather than `reasoning`.
+const SUMMARIZER_ROW = {
+    id: 'obs-2',
+    session_id: 'sess-2',
+    status: 'succeeded',
+    scanner_snapshot: {
+        name: 'Session summaries',
+        scanner_type: 'summarizer',
+        scanner_version: 1,
+        scanner_config: { prompt: 'Summarize what the person did.', length: 'long' },
+        query: { kind: 'RecordingsQuery', events: [{ id: '$pageview' }] },
+    },
+    scanner_result: {
+        model_output: {
+            scanner_type: 'summarizer',
+            title: 'Renamed a project, then opened billing',
+            summary: 'The person renamed a project and then opened the billing page.',
+            summary_segments: [
+                { kind: 'text', value: 'The person renamed a project and then opened the billing page.' },
+                { kind: 'chip', timestamp_ms: 91000 },
+            ],
+        },
+        signals_count: 0,
+    },
+}
+
+function mockContext(row: object = FULL_ROW): Context {
+    return {
+        stateManager: { getProjectId: async () => 1 },
+        api: {
+            getProjectBaseUrl: () => 'https://us.posthog.com/project/1',
+            request: async () => ({ count: 1, next: null, previous: null, results: [structuredClone(row)] }),
+        },
+    } as unknown as Context
+}
+
+describe('vision-scanners-observations-list row trimming', () => {
+    const listTool = GENERATED_TOOLS['vision-scanners-observations-list'] as () => ToolBase<ZodObjectAny>
+
+    it('drops the repeated snapshot fields and the duplicated reasoning segments', async () => {
+        const result: any = await listTool().handler(mockContext(), { scanner_id: 'scanner-1' })
+        const [row] = result.results
+
+        expect(row.scanner_snapshot.scanner_config).not.toHaveProperty('prompt')
+        expect(row.scanner_snapshot).not.toHaveProperty('query')
+        expect(row.scanner_result.model_output).not.toHaveProperty('reasoning_segments')
+    })
+
+    it('keeps the per-row output a caller reads, and the version tag beside it', async () => {
+        const result: any = await listTool().handler(mockContext(), { scanner_id: 'scanner-1' })
+        const [row] = result.results
+
+        expect(row.scanner_result.model_output.reasoning).toBe('The shopper opened the cart and left without paying.')
+        expect(row.scanner_result.model_output.verdict).toBe('yes')
+        expect(row.scanner_snapshot.scanner_version).toBe(3)
+        // Trimming one config key must not take the rest of the config with it.
+        expect(row.scanner_snapshot.scanner_config.allow_inconclusive).toBe(true)
+        // Enrichment runs on the trimmed rows, so each row still deep-links its recording.
+        expect(row._posthogUrl).toBe('https://us.posthog.com/project/1/replay/sess-1')
+    })
+
+    it('drops the duplicated summary segments on a summarizer row, and keeps the summary', async () => {
+        const result: any = await listTool().handler(mockContext(SUMMARIZER_ROW), { scanner_id: 'scanner-2' })
+        const [row] = result.results
+
+        expect(row.scanner_result.model_output).not.toHaveProperty('summary_segments')
+        expect(row.scanner_result.model_output.summary).toBe(
+            'The person renamed a project and then opened the billing page.'
+        )
+    })
+
+    it('defaults to a page a caller can read without naming a limit', () => {
+        expect(listTool().schema.parse({ scanner_id: 'scanner-1' }).limit).toBe(20)
+    })
+})