Investigating Replay Agent Skill

Investigating a session recording

When a user asks "what happened in this session?" or provides a recording/session ID

to investigate, gather all relevant context in parallel rather than making them

ask for each piece.

Available tools

| Tool | Purpose |

| ------------------------------------------ | -------------------------------------------------------- |

| posthog:session-recording-get | Recording metadata (duration, counts, status) |

| posthog:persons-retrieve | Person profile (properties, distinct IDs) |

| posthog:execute-sql | Query events, errors, and page views in session |

| posthog:query-error-tracking-issues-list | Find error tracking issues linked to the session |

| posthog:vision-observations-list | Check for an existing Replay Vision AI summary |

| posthog:vision-scanners-list | Find summarizer scanners (scanner_type=summarizer) |

| posthog:vision-scanners-scan-session | Run a summarizer scanner on the session (slow, optional) |

| posthog:vision-scanners-create | Create a temporary summarizer scanner (ask first) |

| posthog:vision-scanners-delete | Delete a temporary scanner after summarizing |

Workflow

Step 1 — Get recording metadata and person profile

Start with the recording to get metadata and the person's distinct ID:

posthog:session-recording-get
{
  "id": "<recording_id>"
}

The response includes distinct_id, person, duration, interaction counts,

console error counts, and viewing status. Use the distinct_id to fetch

the full person profile:

posthog:persons-retrieve
{
  "id": "<person_uuid_from_recording>"
}

Step 2 — Query same-session events

Get the timeline of what the user did during the session:

posthog:execute-sql
SELECT
    timestamp,
    event,
    properties.$current_url AS url,
    properties.$browser AS browser,
    properties.$os AS os,
    properties.$device_type AS device_type,
    properties.$screen_width AS screen_width
FROM events
WHERE $session_id = '<session_id>'
ORDER BY timestamp ASC
LIMIT 200

For sessions with many events, focus on the most informative ones:

posthog:execute-sql
SELECT
    timestamp,
    event,
    properties.$current_url AS url,
    if(event = '$exception', properties.$exception_message, null) AS exception_message,
    if(event = '$exception', properties.$exception_type, null) AS exception_type
FROM events
WHERE $session_id = '<session_id>'
    AND event IN ('$pageview', '$pageleave', '$autocapture', '$exception', '$rageclick')
ORDER BY timestamp ASC
LIMIT 100

Step 3 — Check for linked error tracking issues

If the recording has console errors or exceptions, find related error tracking issues:

posthog:execute-sql
SELECT DISTINCT
    properties.$exception_fingerprint AS fingerprint,
    properties.$exception_type AS type,
    properties.$exception_message AS message,
    count() AS occurrences
FROM events
WHERE $session_id = '<session_id>'
    AND event = '$exception'
GROUP BY fingerprint, type, message
ORDER BY occurrences DESC
LIMIT 10

If fingerprints are found, search for the corresponding error tracking issues

to provide links and status:

posthog:query-error-tracking-issues-list
{
  "searchQuery": "<exception_type or message>"
}

Step 4 — Synthesize the investigation

Present the findings as a coherent narrative:

• Who — person properties (name, email, country, plan, etc.)
• What — sequence of pages visited and key actions taken
• Problems — exceptions, console errors, rage clicks, and their frequency
• Related issues — linked error tracking issues with their status (active/resolved)
• Context — session duration, device/browser, activity score

Optional: AI summary via Replay Vision

If the user wants a deeper analysis without reading through events manually,

offer a Replay Vision summary. Follow "check-then-scan" — don't scan blindly,

a scanner can only observe a given session once.

• Check for an existing summary. A scheduled scanner may already have one:

   posthog:vision-observations-list
   {
     "session_id": "<session_id>"
   }

Look for an observation where scanner_snapshot.scanner_type is summarizer

and status is succeeded. If found, read scanner_result.model_output

(title, summary, intent, outcome, friction_points, keywords) — done,

no new scan needed.

• Find a summarizer scanner if none exists yet:

   posthog:vision-scanners-list
   {
     "scanner_type": "summarizer"
   }

• Exactly one → use it.
• More than one → show the user the scanners (name + prompt) and ask which to use.
• None → no summarizer scanner exists. See

No summarizer scanner? Run a temporary one below.

• Scan the session with the chosen scanner. Warn this is async and takes

several minutes (rasterize + LLM):

   posthog:vision-scanners-scan-session
   {
     "id": "<scanner_id>",
     "session_id": "<session_id>"
   }

• Retrieve the result by polling vision-observations-list (step 1) until

the new observation reaches succeeded.

No summarizer scanner? Run a temporary one

If the project has no summarizer scanner, you can still produce a one-off summary

with a throwaway scanner — but ask the user's permission before creating anything.

• Ask permission to create a temporary summarizer scanner just to summarize

this one session.

• Create it disabled so it never sweeps on a schedule — a disabled scanner

only runs when you trigger it on demand, so it won't touch other sessions or

burn quota in the background:

   posthog:vision-scanners-create
   {
     "name": "Temporary on-demand summary",
     "scanner_type": "summarizer",
     "scanner_config": {
       "prompt": "Summarize what the user was trying to do, whether they succeeded, and any friction they hit."
     },
     "query": { "kind": "RecordingsQuery" },
     "model": "gemini-3-flash-preview",
     "enabled": false
   }

• Scan this session on demand with the new scanner, then poll for the result:

   posthog:vision-scanners-scan-session
   {
     "id": "<new_scanner_id>",
     "session_id": "<session_id>"
   }

Poll vision-observations-list until the observation reaches succeeded and

read scanner_result.model_output.

• Ask whether to keep or delete the scanner. Once you have the observation,

ask the user if they want to keep the temporary scanner or delete it with

vision-scanners-delete. Deleting is safe: the summary you just read is also

emitted as an event that persists after the scanner is gone, so cleaning up the

temporary scanner does not lose the result.

Tips

• Run steps 1-3 in parallel when possible — they're independent queries.
• If the recording has very few events, the session was likely very short.

Note this rather than suggesting something is broken.

• Console error count from the recording metadata is a good signal for whether

to dig into exceptions. If it's 0, skip step 3.

• The start_url from the recording tells you where the user's journey began —

use this to frame the narrative.

• If person is null on the recording, the user was anonymous.

Person properties won't be available, but events still are.