llm-obs-session-classify

Backend

Detection — At the start of every invocation, before taking any action, determine which backend to use:

1. If the user passed --backend pup anywhere in their invocation → use pup mode immediately, regardless of whether MCP tools are present. Skip steps 2–4.
2. Check whether MCP tools are present in your active tool list. The canonical signal is whether mcp__datadog-llmo-mcp__search_llmobs_spans appears in your available tools.
3. If MCP tools are present → use MCP mode throughout. Call MCP tools exactly as named in this skill's workflow sections.
4. If MCP tools are absent → check whether pup is executable: run pup --version via Bash. A JSON response containing "version" confirms pup is available.
5. If pup responds → use pup mode throughout. Translate every MCP tool call to its pup equivalent using the Tool Reference appendix at the bottom of this file.
6. If neither is available → stop and tell the user:

   "Neither the Datadog MCP server nor the pup CLI is available. Connect the MCP server (claude mcp add --scope user --transport http datadog-llmo-mcp 'https://mcp.datadoghq.com/api/unstable/mcp-server/mcp?toolsets=llmobs,rum') or install pup."

--backend pup is accepted anywhere in the invocation arguments and is stripped before passing remaining args to the skill logic.

pup invocation rules:

• Invoke via Bash: pup llm-obs <subcommand> [flags]
• pup always outputs JSON. Parse directly — no content-block unwrapping (unlike MCP results, which may wrap JSON in [{"type": "text", "text": "<json>"}]).
• If pup returns an auth error, tell the user to run pup auth login and stop.
• Parallelization: issue multiple Bash tool calls in a single message (one pup command per call).
• Time flags: pup accepts bare duration strings (1h, 7d, 30m) and RFC3339 timestamps. Do not use now--prefixed strings — strip the prefix when converting from a skill --timeframe argument: now-7d → 7d, now-24h → 24h, now-30d → 30d.
• --summary on pup llm-obs spans search strips payload fields to essential metadata only. Use it in bulk/search phases where content is not needed.

pup mode notes by entry mode:

• session_id mode: Steps 1–3 and Step 5 work fully. Step 4 (RUM) uses pup rum aggregate --user-email EMAIL instead of analyze_rum_events — see Tool Reference. Step 4b (audit trail) is pup-native and queries the active user's own org via OAuth.
• trace_id mode: Full parity with MCP mode.
• ml_app mode: Option A (aggregate_spans) is unavailable in pup — skip it and proceed directly to Option B.

Invocation ID: At the very start of each invocation, before any MCP tool call, generate an 8-character hex invocation ID (e.g., 3a9f1c2b). Keep it constant for the entire invocation.

Intent tagging: On every MCP tool call, prefix telemetry.intent with skill:llm-obs-session-classify[<inv_id>] — followed by a description of why the tool is being called. On the first MCP tool call only, use skill:llm-obs-session-classify:start[<inv_id>] — instead (note the :start suffix). Example first call: skill:llm-obs-session-classify:start[3a9f1c2b] — Step 1: enumerate turn root spans for session abc-123

Skill: eval-session-classify

Classification skill for Datadog LLM Obs sessions and traces. Produces a satisfaction verdict (yes / partial / no) with a brief reasoning string. Designed to be called standalone or by eval-pipeline.

Inputs

If none of session_id, trace_id, or ml_app is provided → stop immediately and return:

{
  "error": "missing_input",
  "detail": "Provide one of: session_id, trace_id, or ml_app."
}

Phase 0 — Mode Detection

• session_id provided → session_id mode → go to Session Mode
• trace_id provided (no session_id) → trace_id mode → go to Trace Mode
• ml_app provided (no session_id, no trace_id) → ml_app mode → go to ML App Mode
• Nothing provided → emit error (see above)

Output Format

Default (compact) — used in all modes unless the caller requests verbose:

verdict: yes | partial | no
reason: <one sentence>

Verbose — full markdown report (see end of each mode section). Request verbose by including verbose: true in input or asking for a detailed report.

In ml_app mode, a summary table is always appended after the per-trace compact blocks.

Content Retrieval Cascade

Reading conversation content follows this cascade across all modes. Run the cascade for every turn root span identified in Step 1 (session mode) or Step T1 (trace mode) before forming a verdict.

Completeness gate: do not proceed to RUM (session mode) or classification (any mode) until the cascade has run for every turn. Previews from search_llmobs_spans (input.preview / output.preview, truncated to ~200 chars) do NOT satisfy this gate — they identify turn boundaries, nothing more. Only actual content from get_llmobs_agent_loop, expand_llmobs_spans + get_llmobs_span_content, or get_llmobs_span_content(field="input"/"output") counts.

Short sessions (≤ 20 turns): run C1/C2 for every turn — no selection, no "key turns" heuristic.

Long sessions (> 20 turns): run C1/C2 for the first 5 and last 5 turns, plus any turn whose input.preview or output.preview flags an anomaly (error, correction loop, repeated intent, unexpected tool). Scan all remaining turns' previews for anomalies before skipping them.

Parallelism: issue all N get_llmobs_agent_loop calls in a single message — one call per turn, all in the same batch. For an N-turn session this means N simultaneous C1 calls. Do not process turns sequentially and do not form partial verdicts before all results are in hand. Move to C2 for a turn only when C1 returns iterations: [] for that turn.

C1 — get_llmobs_agent_loop(trace_id, agent_span_id)

The richest source: full system prompt, user message, tool call arguments + results, assistant response, and token economics per iteration. Attempt this first for every agent span.

get_llmobs_agent_loop(
  trace_id           = "<TRACE_ID>",
  span_id            = "<AGENT_SPAN_ID>",
  from               = "now-90d",
  to                 = "now",
  max_content_length = 2000
)

• iterations: [] and timeline: null → the app's LLM spans go through an intermediate workflow layer (e.g. get_answer_from_model_step) rather than as direct children of the agent span. Fall through to C2.
• 404 → span ID does not resolve in the trace store. Fall through to C2.
• Content fields contain <REDACTED_INPUT> or <MASKED_STREAMING_RESPONSE> → IO tracing is disabled by a feature flag. Structure (iteration count, tool names, token usage) is still useful — note it. Fall through to C2 for actual content.

When a C1 result is too large for context and gets persisted to disk (common on long sessions): do NOT try to Read the full file — it will exceed the token limit. Run a targeted Python extraction to build a structured per-turn summary:

A large-but-parseable C1 result is real content; only fall through to C2 if C1 returned iterations: [].

C2 — get_llmobs_span_content(field="messages") on LLM child spans

When C1 returns iterations: [], the LLM spans typically sit 2 levels below the root agent span — under a workflow wrapper (e.g. get_answer_from_model_step) that the agent-loop API does not descend into. The concrete tree per turn:

Use expand_llmobs_spans to navigate to the LLM span IDs — not get_llmobs_trace, which only returns depth-1 children and cannot reach LLM spans nested under a workflow:

From the returned tree, collect all nodes with span_kind=llm and has_input=true (commonly named anthropic.request, openai.request, chat_completion-call, messages-call). Call get_llmobs_span_content(field="messages") on each.

JSONPath + truncation trap: path is applied after max_tokens truncation. Each LLM-call span typically starts with a multi-KB system prompt, so a low max_tokens means the JSONPath filter operates on system-prompt-only content and silently returns the wrong messages.

To extract the meaningful tail (user query, tool calls, final answer):

1. Call once without path to read total_tokens_approx from the response.
2. Re-call with path = "$.[-5:]" and max_tokens = total_tokens_approx + 500.

This returns the last 5 messages: typically [user context+query, reasoning, assistant text, assistant tool_call, tool result] — enough to understand what the turn did.

• Messages are "REDACTED" → IO tracing disabled. Fall through to C3.
• content_info map does not include messages → not a chat span. Fall through to C3.

C3 — get_llmobs_span_content(field="input") and field="output" on the root span

The root span often carries synthetic summaries written by the app (e.g. the raw user query as input, the final response as output, or "Investigate error for issue: <id>" / "Investigation completed with status: completed"). Minimal signal, but enough to confirm what the task was and whether it completed.

C4 — Structural signals only

When all content is inaccessible, classify from span metadata alone:

• status (ok / error), stop_reason, response_truncated
• Child span names (tool names, workflow step names like classify, generate-summary, suggest-action)
• iteration tag count (total LLM rounds)
• Duration and token counts if available

Drill-down — targeted extraction for suspicious tool results

After the summary pass, drill into any turn whose tool list includes a client-rendered tool, a write operation, or an iteration flagged with error_message. These checks surface silent failures the summary loop cannot detect.

Triggers:

• Client-rendered tools (browser-handled, e.g. show_content, create_datadog_notebook, edit_datadog_notebook) — always mandatory when present. The server fires an event; the client handles display. The result tells you whether rendering was delegated or confirmed.

  • "This tool is handled by the client application..." → delegated to browser, no server-side confirmation the user saw anything. Check for a client_tool_response in the next turn's input. If absent, or if the next turn shows user confusion, treat as invisible output — strong negative signal.
  • Any other result → executed server-side; output was in the response stream.

• Write operations (e.g. call_datadog_api, upsert_datadog_dashboard, upsert_widget, monitor/notebook create-or-update) — the assistant often narrates "I updated X" regardless of the actual HTTP outcome. Check the result for:

  • 403 / "not allowlisted" / "PUT not allowlisted" → write blocked; the change was never applied.
  • 404 → asset not found; editing something that doesn't exist.
  • 500 / "internal server error" → write uncertain, may not have landed.
  • "Session not found" → resource expired before the tool ran.
  • 200 / 201 with payload → confirmed success.

• Data queries returning empty (e.g. search_datadog_metrics, get_datadog_metric, search_datadog_logs, search_datadog_spans) — when the summary shows many iterations (> 5) of the same query tool, check whether results were empty. An assistant reasoning from NO_DATA / empty arrays may produce plausible-sounding but fabricated analysis (hallucination failure mode).

• Error iterations — any iteration where error_message is set: confirm the exact error, which tool triggered it, and whether the assistant retried successfully or gave up. A transient retry is neutral; a permanent API restriction or repeated identical failure is negative.

Interpretation cheat sheet:

The tool-name lists above are examples calibrated for Datadog assistant apps. For other apps, look at the tool names that appear in C1's tool_calls and reason about their semantics (client rendering vs server side-effect vs read-only query).

Session Mode

Classifies a single Datadog CMD+I assistant session. Uses RUM behavioral signals to confirm the trace-based verdict.

Step 1 — Enumerate turn root spans

pup mode: pup llm-obs spans search --query "@session.id:<SESSION_ID>" --root-spans-only --limit 500 --from 90d

Use tags={...} (not the query string form). The free-form query="session_id:<id>" triggers full-text behavior that returns the entire trace bundle for every turn (5× the spans, structural filters silently ignored) and exhausts the row budget on long sessions. tags={"session_id": "<id>"} hits the indexed tag filter and respects the other parameters.

root_spans_only=True returns one span per user turn — typically the root agent span — rather than every workflow/LLM/tool descendant. Paginate via next_cursor if the session has more than 500 turns.

For each returned root span, record:

• trace_id, span_id
• start_ms, duration_ms
• Tags: user_handle, user_id, org_id, product_area, message_id
• name and span_kind of the root (usually agent / assistant, but app-dependent)
• input.preview, output.preview (for turn-boundary identification only — see warning below)

Sort the turns ascending by start_ms. From the first turn's tags, capture session-level identity: user_handle, user_id, org_id, product_area, session start_ms.

Preview-truncation warning: input.preview and output.preview are truncated to ~200 characters. They identify turn boundaries and surface obvious anomalies — they are NOT sufficient for verdict formation. Never skip the Content Retrieval Cascade because the previews already "look conclusive."

If search_llmobs_spans returns no results → stop, return error llmobs_not_found.

Per-turn structural details (iteration count, tool names, stop_reason, mcp flag, response_truncated) are recovered later from each turn's get_llmobs_agent_loop result in Step 3 (C1).

Step 2 — Get evaluations and metadata

From the evaluations map: iterate all keys, note .value, .reasoning, .tags for each judge. From content_info.metadata: query_string, referrer_path, referrer_url, entities_json, user_info_json.

If get_llmobs_span_details fails or returns empty → skip silently, proceed to step 3.

If content_info shows metadata is present, fetch it via get_llmobs_span_content(field="metadata") to get the actual query_string and referrer_path — these give the user's question and the page they were on without needing the agent loop.

Step 3 — Read the full conversation

For every turn root span enumerated in Step 1, follow the Content Retrieval Cascade. Issue all per-turn C1 calls (get_llmobs_agent_loop) in a single parallel batch per the cascade's parallelism rule, then fall through to C2 for any turns where C1 returns iterations: [].

For apps where LLM spans sit under a workflow wrapper (e.g. assistant_api, where anthropic.request is a grandchild via get_answer_from_model_step), C1 will routinely return empty iterations — this is expected and the cascade handles it by routing those turns through C2's expand_llmobs_spans navigation. C3 (metadata + output on the root span) is the last resort when no LLM-call content is available.

Step 4a — Resolve RUM session ID(s)

Before the main RUM queries, run a narrow ±2min lookup around the first turn's start_ms to find the exact browser session(s) the user was on. This isolates the RUM session_id(s) and avoids cross-session bleed that occurs when filtering only by @usr.email over a wide time window (a user with multiple tabs across the day will return events from all of them).

Extract all distinct session_id values → store as rum_session_ids. The RUM filter for all subsequent queries (<rum_filter> below) is determined by the result:

• 0 results → <rum_filter> = @usr.email:<user_handle> (fallback to email + time window).
• 1 result → <rum_filter> = @session.id:<rum_session_id>.
• Multiple results → <rum_filter> = @session.id:(<id1> OR <id2>).

pup mode: pup rum aggregate --user-email <user_handle> --query "@action.type:custom" --from <start_ms - 120000> --to <start_ms + 120000> --compute count --group-by @session.id. Confirm the --group-by flag name with pup rum aggregate --help.

Step 4 — Get RUM behavioral signals

pup mode: Replace analyze_rum_events calls with pup rum aggregate calls (see Tool Reference). Page views: pup rum aggregate --user-email USER_HANDLE --from START_MS --to END_MS --compute count --group-by @session.id. Custom actions: pup rum aggregate --user-email USER_HANDLE --query "@action.type:custom" --from START_MS --to END_MS --compute count --group-by @evt.name. Confirm event-type filtering flag with pup rum aggregate --help if the API rejects the query.

With start_ms and <rum_filter> (from Step 4a), define the window:

• pre: [start_ms − 30min, start_ms]
• during: [start_ms, start_ms + session_duration_ms]
• post: [start_ms + session_duration_ms, start_ms + session_duration_ms + 60min]

Run in parallel:

RUM Query A — Page view timeline:

@view.time_spent is in nanoseconds.

RUM Query B — Custom actions (product-area narrowed):

The LIKE filters above are a Datadog assistant example — narrow to whatever action-name prefix is meaningful for the app under classification, or drop the filter to read all custom actions on the session.

If either query returns 0 rows or hits a permission/auth error, do NOT abort the classification. Missing or partial RUM is a coverage gap, not behavioral evidence — treating it as negative is a documented failure mode.

Hard rule: never use the absence of RUM as evidence for a negative verdict. If trace and audit signals point to satisfaction, the absence of RUM does not override them; if trace signals are ambiguous, default to the trace-only verdict rather than penalizing for missing behavioral data.

The rum_unavailable error code is reserved for the unrecoverable case where RUM was required by the classification protocol (e.g., the trace is so content-redacted that RUM is the only remaining signal) AND no RUM exists. Otherwise, proceed in degraded mode.

Signal interpretation framework

RUM actions are not equal evidence. Stratify them into three tiers and let the tier control how much weight each carries in Step 5. The categories below are agent-agnostic — map your specific app's RUM action names onto them. Treat unrecognized actions as Ambiguous until you have evidence for how they correlate with satisfaction.

Tier-A — sufficient on its own (when present, can flip or confirm the trace verdict)

Tier-B — corroborating, never decisive on its own

Ambiguous — DO NOT use as sole evidence

For your specific agent: before classifying, build a mapping from your app's RUM action names onto these categories. A reference mapping for the Datadog CMD+I assistant is in @rum-actions-bits-assistant.md — use it as a template for what to look up in your own app's RUM event taxonomy. Unrecognized action names default to Ambiguous; promote to Tier-B or Tier-A only after you have evidence (from labeled sessions or product knowledge) for how the action correlates with satisfaction.

For interpreting Datadog assistant action names specifically, consult @rum-actions-bits-assistant.md.

Step 4b — Get audit-trail signals (when an asset was created or edited)

The audit trail surfaces server-confirmed effects of the assistant's actions (writes that landed, modifications that stuck, deletions, follow-up edits) and post-session user behavior on the same asset. For sessions where the assistant created or edited a Datadog asset (dashboard, notebook, monitor, SLO, etc.), this is the most authoritative behavioral signal — more direct than RUM clicks.

Primary tool: pup audit-logs search, which queries the active user's own org via OAuth (pup auth login).

When to run this step:

• The trace's tool calls include any write op (upsert_*, call_datadog_api with PUT/POST, monitor/notebook create-or-update) → run it.
• The session has a known asset ID (from a tool-call argument, the dashboard_id tag, the session's referrer_path, or the user query) → run it.
• The trace is pure read-only (no writes, no asset ID) → skip; nothing for audit to add over Step 4.

Query priority (most signal first, least noise last):

1. @asset.id — modification/deletion history with full diff. Most authoritative. Required when an asset ID is known.
2. Q3 (targeted HTTP path) — write count and timing for the same asset. Required when an asset ID is known.
3. Q1 / Q2 (broad-signal HTTP audit) — fallback only. Noisy on active users.

@asset.id — modification history

Each row's body lives at row["attributes"]["attributes"]. Fields:

• action — created / modified / deleted / accessed
• usr.email
• asset.type (dashboard, notebook, monitor, …), asset.id, asset.name
• asset.new_value / asset.prev_value — full before/after state as nested dicts (use directly; no JSON parsing needed)
• timestamp — ISO8601 string

Pup returns descending order by default. Sort ascending for the chronological session arc:

Signals:

Do not over-interpret audit signals. Audit measures backend correctness, not user satisfaction — a missing write or absent event is not by itself proof of dissatisfaction. The exception is explicit user actions: a deleted event, a manual modified with widget/cell removal, or a manual write that overrides what the agent produced. Those are direct user intent and can flip a verdict. Backend-symptom signals (zero writes despite acceptance, missing PUTs, etc.) should only corroborate trace + RUM, never flip them alone.

For dashboards, the most useful diff is widget count:

Q3 — targeted HTTP path (write count + timing)

While @asset.id shows the diff, Q3 shows the HTTP event count for writes to the resource — useful for cross-checking the trace's tool calls and finding manual post-session edits.

Phase-bucket the events to see what happened during vs after the session:

Each PUT is one write. 200/201 = confirmed success. status_code: 0 means the browser sent the request but navigated away before the response landed — server almost certainly processed it; count it alongside the 200s.

Interpretation:

• during PUT 200 count should match the trace's upsert tool-call count.
• Writes timestamped 10–30s after session_end and whose count matches the last turn's expected writes are async tool execution lag — not user-initiated.
• Writes beyond 30s after session_end are user manual edits. Count them post-session: a small clustered burst (e.g., 5–10 PUTs in ~1 min) = fine-tuning (mildly positive — user kept the result and improved it); zero post-session PUTs = clean acceptance.
• Confirmed cancelled sessions have zero during PUT events.

Q1 / Q2 — broad-signal HTTP audit (fallback)

Use only when no asset ID is known, or to look for retry / support-ticket signals after the targeted queries.

Noise dominates on active users. The 200-row cap is real and there is no pagination cursor. Other high-volume paths to strip when needed: /api/ui/frontend_telemetry/metrics, /api/v1/logs-analytics/aggregate, /api/v1/logs-analytics/list, /api/v1/logs-analytics/facet_info, /api/v2/watchdog/insights/search/*, /api/v2/notifications/notify/*, /api/v2/notifications/presence/*. If you still hit the cap after stripping, narrow the time window.

Look for: support ticket creation (POST /api/v2/cases), retry assistant calls (POST /api/v2/assistant), navigation to docs (visible in RUM Step 4, not here).

Response-extraction helper

The doubly-nested envelope (row["attributes"]["attributes"]) is verbose. Normalize once per script:

Graceful degradation

If a query returns 0 rows and the session is in an external customer org, do not error — flag audit_unavailable in the verbose output and continue with trace + RUM. Some external customer audit trails (notably AP1 / SOC2-isolated orgs) are not reachable via pup. Trace + RUM alone is still sufficient for a verdict.

Step 4b — Get audit-trail signals (when an asset was created or edited)

The audit trail surfaces server-confirmed effects of the assistant's actions (writes that landed, modifications that stuck, deletions, follow-up edits) and post-session user behavior on the same asset. For sessions where the assistant created or edited a Datadog asset (dashboard, notebook, monitor, SLO, etc.), this is the most authoritative behavioral signal — more direct than RUM clicks.

Primary tool: pup audit-logs search, which queries the active user's own org via OAuth (pup auth login).

When to run this step:

• The trace's tool calls include any write op (upsert_*, call_datadog_api with PUT/POST, monitor/notebook create-or-update) → run it.
• The session has a known asset ID (from a tool-call argument, the dashboard_id tag, the session's referrer_path, or the user query) → run it.
• The trace is pure read-only (no writes, no asset ID) → skip; nothing for audit to add over Step 4.

Query priority (most signal first, least noise last):

1. @asset.id — modification/deletion history with full diff. Most authoritative. Required when an asset ID is known.
2. Q3 (targeted HTTP path) — write count and timing for the same asset. Required when an asset ID is known.
3. Q1 / Q2 (broad-signal HTTP audit) — fallback only. Noisy on active users.

@asset.id — modification history

Each row's body lives at row["attributes"]["attributes"]. Fields:

• action — created / modified / deleted / accessed
• usr.email
• asset.type (dashboard, notebook, monitor, …), asset.id, asset.name
• asset.new_value / asset.prev_value — full before/after state as nested dicts (use directly; no JSON parsing needed)
• timestamp — ISO8601 string

Pup returns descending order by default. Sort ascending for the chronological session arc:

Signals:

For dashboards, the most useful diff is widget count:

Q3 — targeted HTTP path (write count + timing)

While @asset.id shows the diff, Q3 shows the HTTP event count for writes to the resource — useful for cross-checking the trace's tool calls and finding manual post-session edits.

Phase-bucket the events to see what happened during vs after the session:

Each PUT is one write. 200/201 = confirmed success. status_code: 0 means the browser sent the request but navigated away before the response landed — server almost certainly processed it; count it alongside the 200s.

Interpretation:

• during PUT 200 count should match the trace's upsert tool-call count.
• Writes timestamped 10–30s after session_end and whose count matches the last turn's expected writes are async tool execution lag — not user-initiated.
• Writes beyond 30s after session_end are user manual edits. Count them post-session: a small clustered burst (e.g., 5–10 PUTs in ~1 min) = fine-tuning (mildly positive — user kept the result and improved it); zero post-session PUTs = clean acceptance.
• Confirmed cancelled sessions have zero during PUT events.

Q1 / Q2 — broad-signal HTTP audit (fallback)

Use only when no asset ID is known, or to look for retry / support-ticket signals after the targeted queries.

Noise dominates on active users. The 200-row cap is real and there is no pagination cursor. Other high-volume paths to strip when needed: /api/ui/frontend_telemetry/metrics, /api/v1/logs-analytics/aggregate, /api/v1/logs-analytics/list, /api/v1/logs-analytics/facet_info, /api/v2/watchdog/insights/search/*, /api/v2/notifications/notify/*, /api/v2/notifications/presence/*. If you still hit the cap after stripping, narrow the time window.

Look for: support ticket creation (POST /api/v2/cases), retry assistant calls (POST /api/v2/assistant), navigation to docs (visible in RUM Step 4, not here).

Response-extraction helper

The doubly-nested envelope (row["attributes"]["attributes"]) is verbose. Normalize once per script:

Graceful degradation

If a query returns 0 rows and the session is in an external customer org, do not error — flag audit_unavailable in the verbose output and continue with trace + RUM. Some external customer audit trails (notably AP1 / SOC2-isolated orgs) are not reachable via pup. Trace + RUM alone is still sufficient for a verdict.

Step 5 — Classify

Using the conversation (step 3), evaluations (step 2), RUM signals (step 4), and audit signals (step 4b, when run):

Satisfaction verdict:

• yes: Final response directly answers user's intent, no negative feedback, no abandon signals
• partial: Partially correct, or user got unblocked only through additional effort
• no: Negative feedback given, user abandoned, or core intent structurally unachievable

Failure mode codes (omit if yes):

Compact output (default):

verdict: yes | partial | no
reason: <one sentence>

Verbose output (only when requested):

Trace Mode

Classifies a single LLM Obs trace by trace_id. No RUM. Works for any LLM app.

Step T1 — Get span structure

From the results, map the full span hierarchy:

• Identify the root span (the one with parent_id: undefined)
• Note the root span's span_kind — it may be agent, workflow, llm, or other
• Extract all span kinds present: agent, llm, tool, workflow, retrieval, embedding
• Extract tags: ml_app, service, session_id, investigation.id, issue_id, user_handle, org_id, matched_model_name, stop_reason, current_turn, iteration

Determine app type from span structure:

Find the agent span (if any): The agent span is not always the root. Search all spans for span_kind=agent — it may be a child or grandchild of a workflow root (e.g. investigate → Agent workflow → generate-summary). Use the deepest / most meaningful agent span for the agent loop call.

If no spans found → stop, return error llmobs_not_found.

Unclassifiable guard: If all spans are span_kind=embedding with no messages content, stop and return:

Step T2 — Get span details and evaluations

Call get_llmobs_span_details on the root span (and on the agent span if different from root). Extract:

• children_ids → confirms the hierarchy
• evaluations map → any judge verdicts (iterate all keys; treat WARN-status evals as not applicable)
• content_info → which fields exist and their size

Step T3 — Read content

Follow the Content Retrieval Cascade using the agent span (C1), then LLM child spans (C2), then the root span (C3), then structural signals (C4).

Notes on specific patterns observed:

• service:ai_gateway LLM spans (bits-copilot, assistant_api): standalone chat_completion-call or messages-call spans from the AI Gateway proxy — get_llmobs_span_content(messages) works reliably on these.
• generate-summary agent spans (error-tracking-investigator): agent loop works but content is often REDACTED when IO tracing is gated by a feature flag. Fall back to structural signals: child span names (classify, suggest-action, add-markdown-links), tool calls, turn count from current_turn tag, overall status.
• assistant agent spans (assistant_api): agent loop returns empty iterations because LLM spans are grandchildren via get_answer_from_model_step workflow. Use get_llmobs_span_content(metadata) for query_string and get_llmobs_span_content(output) for the response.

Step T4 — Classify

Using content and evaluations:

Satisfaction criteria:

• yes: Output directly and completely addresses the task. No errors, truncation, or refusals.
• partial: Addresses part of the task, correct but incomplete, or shows degraded quality.
• no: Fails to address the task — hallucinations, errors, empty response, wrong tool use, structurally unachievable.
• When content is fully REDACTED, classify from structural signals: status:error → no; clean completion with expected child span names → lean yes; note low confidence.

Failure mode codes: same taxonomy as session mode (see above).

Compact output (default):

Verbose output (when requested):

ML App Mode

Samples and classifies multiple sessions or traces from a given LLM app. Handles diverse instrumentation patterns across teams.

Step M1 — Sample spans and understand the app

If no spans found → stop, return:

Unclassifiable guard: If all root spans are span_kind=embedding → stop and return:

Sanity check: Skim the sampled root spans before classifying. Confirm they represent the app you expect — check service:, span names, and tags. Discard outliers that look unrelated before proceeding.

Root span kind analysis:

LLM Experiments traces: If root spans have span_kind: experiment and carry input, output, and expected_output structured fields, you are looking at a Datadog LLM Experiments trace. Each experiment span represents one dataset record run. Content is in the root span's structured I/O fields — use get_llmobs_span_content(field="input"/"output"/"expected_output") on the root span rather than following the LLM sub-span message cascade (C1/C2). LLM sub-spans in experiment traces may contain stub or placeholder content.

Grouping granularity — session vs trace:

First, determine what business ID tag to group by. Use this cascade:

Option A — aggregate_spans (preferred): If the app emits APM spans (most Go/Java/Python services do), use APM aggregation to get all unique IDs in one call. In pup mode, aggregate_spans is not available — skip Option A and proceed directly to Option B.

Try @session_id first. If buckets are empty, try the business ID tag you observed on the LLM Obs spans (@investigation.id, @chatid, @correlation_id, etc.). This returns up to 1000 unique IDs with a count per ID in a single call — far more efficient than paginating LLM Obs search results.

Option B — search_llmobs_spans fallback: If the app has no APM instrumentation or aggregate_spans returns nothing, paginate LLM Obs root spans and deduplicate the business ID tag client-side:

Collect unique values of session_id, investigation.id, chatid, or whichever tag is consistent across spans.

Granularity decision (after collecting IDs):

1. session_id (UUID format) present → session granularity, group by session_id.
2. A consistent business ID (investigation.id, chatid, correlation_id, etc.) present → use as session equivalent.
3. Neither → trace granularity: one root span = one classification unit.

Sample up to sample_limit unique IDs from the collected set. The fallback cascade is automatic — no user input needed.

Step M2 — Classify each unit (parallel)

Issue all content retrieval calls in a single message. For each sampled unit:

Session granularity: For each session_id, run the Session Mode steps 1–5. Include RUM only if user_handle is available; skip RUM gracefully if not.

Trace granularity: For each trace, run the Trace Mode steps T1–T4 including the content retrieval cascade.

In both cases, the content retrieval cascade (C1→C4) applies. Emit results as they complete — do not wait for all units.

Compact block per unit:

Step M3 — Emit summary

After all units are classified, emit the summary. The # Session Classification Summary header is the detection sentinel for downstream skills (eval-trace-rca, eval-pipeline) — emit it exactly as shown:

• ID column: display as <first8>...<last8> (16 visible chars), linked to the full trace/session URL.
  • Session link: https://app.datadoghq.com/llm/traces?query=session_id:<full_id>
  • Trace link: https://app.datadoghq.com/llm/traces?query=trace_id:<full_id>
• Reason column: the one-sentence classification rationale from the compact block. Use none for yes verdicts.

Error Output Schema

Error codes:

Tool Reference

This appendix applies only in pup mode. In MCP mode, use the tool names in the workflow sections directly.

Spans and traces

RUM

--user-email EMAIL prepends @usr.email:EMAIL to the query, composing cleanly with any --query already set. Confirm event-type filtering flag with pup rum aggregate --help if needed.

Audit trail (Step 4b)

Audit trail is pup-native in both backends — the MCP server does not expose an analyze_audit_events equivalent. Always shell out to pup audit-logs search via Bash, regardless of MCP availability.

Response shape: data[].attributes.attributes.{action,asset,http,usr,timestamp,...} — doubly nested under attributes. timestamp is ISO8601. asset.new_value / asset.prev_value are nested dicts. Pup returns descending order by default — sorted() ascending in code. No pagination cursor; cap is --limit (default 100, set 200 for broad queries).

Tools with no pup equivalent

Operating Rules

• MCP result parsing safety: Before writing any script (Python, jq, etc.) that iterates over or accesses fields in an MCP tool result, inspect the raw structure first — check type(result), top-level keys, and whether the payload is nested inside a content block (e.g. [{'type': 'text', 'text': '<json>'}]). Extract and json.loads() the inner payload if needed before parsing. Never assume MCP results are bare dicts or lists.