exploring-scouts

A scout is a scheduled agent that wakes on its own interval, looks at one PostHog project, decides what's genuinely worth surfacing, and either writes it into the Signals inbox as a report or closes out empty (a real, valid outcome). PostHog ships a fleet of canonical scouts — a cross-product generalist (signals-scout-general) plus per-surface specialists (error tracking, logs, AI observability, experiments, feature flags, session replay, web analytics, surveys, and more). A project may also have custom scouts beyond the canonical fleet — any signals-scout-* skill a team authored (e.g. -brand-mentions, -mcp-feedback) shows up here too, so don't assume a fixed roster: scout-config-list is the authoritative roster for a project. (One caveat: a just-authored scout has no config row until the coordinator's next tick auto-registers one — or until someone registers it via the write-side scout-config-create — so a brand-new scout may briefly be missing from the list.)

This skill helps you understand and explore what a project's scouts are doing and how they're performing — entirely through read-only MCP tools. It is the observability counterpart to the authoring-scouts skill (which teaches writing and tuning) and to the inbox-exploration skill (which covers the inbox reports scouts feed into). (The scout tools were recently renamed from signals-scout-* to scout-*; if a scout-* name comes back unknown, the server may still expose it under the legacy signals-scout-* name — search the tool catalog and call whichever name it returns.)

When the question asks for scout failure rate or cost per run, call posthog:metric-list before the scout tools and look for scout_run_fail_pct or scout_cost_per_run. Run an approved, non-drifted match with posthog:data-catalog-metric-run for the canonical headline. If the user also asks which scouts, runs, or failure modes drive the result, answer the headline first, then use the workflows below for a noncanonical breakdown. If no governed metric matches, say so and label the derived measure noncanonical.

A scout's output is inbox reports, written 1:1. Scouts list emit_report / edit_report in their allowed_tools and author or edit inbox reports directly; a run's output shows up as emitted_report_ids (reports it authored) and edited_report_ids (reports it updated). The run rows also carry emitted_count / emitted_finding_ids — legacy fields from the deprecated signal-emitting channel (weak emit_signal findings a pipeline consolidated). On a report-channel scout they stay 0 / empty even on a productive run; a non-zero tally means the run came from a scout still on the legacy channel (an old custom scout, or a canonical scout not yet ported) — real output for that run, not noise. When unsure of a scout's channel, check its allowed_tools via skill-get. Never read emitted_count: 0 as "did nothing" — check the report columns and the run summary first. A scout whose config carries a structured_output_schema has a third output channel next to reports: schema-validated measurement records, recorded as $scout_structured_output events in the project (only scalar top-level payload keys flatten to output_<key> properties — object and array fields live solely inside the full output property, so a missing output_<key> is not a missing value; subject names the judged entity) rather than as run-row columns. The events are the ground truth — metadata.derived.has_structured_output says the run had at least one batch accepted, which is a fast per-run screen but not delivery confirmation (a rare capture failure after acceptance leaves it true with fewer or no events behind it), so count the events when the number of records matters. See references/scout-data-model.md for the event shape. Each run also carries a metadata map. Top-level: the provenance set harness_prompt_version / report_channel (none, emit, edit, or both) / skill_origin / github_guidance, saying which instructions the run was given; plus routing keys (model / runtime_adapter / reasoning_effort) only when a gate or pin overrode the default. Nested under metadata.derived: booleans the harness computes at the end of the run (has_emit_report, has_edit_report, has_self_improvement, has_chart, has_self_validation, has_structured_output). When comparing runs (before/after a prompt change, one model against another), segment on all four provenance values first: runs differing on any of harness_prompt_version, report_channel, skill_origin, or github_guidance were given different instructions and aren't a like-for-like population. Runs predating this field have none of them, so treat missing provenance as unknown and exclude those runs from a comparison rather than pooling them. For "what kind of run was this?" questions — did it author a self-improvement report, did it validate its follow-up queue — read derived rather than parsing the prose summary. It's computed server-side from what the run actually did, so it can't disagree with the run's own output — with one exception: has_structured_output tracks batches the run had accepted, not events delivered, so it alone can be true with fewer or no records behind it (count the events, as above). No derived map at all means unknown, not "all false" — the run predates the field, failed before finishing, or its stamp failed. Most runs from before this shipped have no map, so don't read their absence as a finding.

There are six things you can observe about the fleet, each with its own tool:

scout-config-list takes a tags parameter (comma-separated) to narrow the roster to the scouts carrying any of the given labels — useful on a large fleet when the question is scoped to one area, e.g. tags=revenue.

The orienting tool is scout-project-profile-get — the deterministic snapshot of "what's true about this project" that every scout cold-starts from. When a scout found nothing, this is usually why.

Two of these tools — scout-runs-list and especially tasks-runs-session-logs-retrieve — routinely return payloads that overflow an MCP client's token budget and get spilled to a file. This is the normal path, not an error. Plan for it up front rather than discovering it after a failed call:

• Keep limit small on scout-runs-list (~10–15). Each row carries a long prose summary, and runs come back newest-first across the whole fleet, so even a modest page is large.
• Session logs are large by nature. A single run's log is hundreds of KB to a few MB. Fetch it with call --json (so the saved file is real JSON, not the pretty text format — jq-able) and read the saved file with jq / a script rather than inline.
• Don't hand-parse the session log. The bundled scripts/ do the reconstruction for you — see below.

Don't assume the project has scouts. The fleet only runs on teams enrolled via the signals-scout feature flag, and a project may have no configs, all-disabled scouts, or scouts stuck in dry-run. Run this first whenever a user asks about their scouts for the first time in a session.

scout-config-list

Read the result against three cases:

The config list is unpaginated — it comes back as { results: [...] } (a bare array), with no count field. Read the result against three cases:

• Empty (results: []) — no scouts are registered. The project isn't enrolled in the scout fleet (or hasn't ticked yet). Say so plainly; don't go fishing for runs. Point the user at the Signals scout settings / PostHog Desktop onboarding rather than inventing activity.

• Configs exist but all enabled: false — the fleet is registered but paused. Nothing is running. Tell the user which scouts exist and that they're all off — and say who switched each one off, which status carries: paused_by_user means a person (or a launch seed posture) turned it off, paused_by_system means an automatic pause with its cause in pause_reason (no_output / ignored / repeated_failures). Either kind resumes with enabled: true via scout-config-update.

• At least one enabled: true — the fleet is registered and that scout is allowed to run. For each enabled scout note its run_interval_minutes (cadence), emit (false = dry-run, runs but writes nothing to the inbox), and last_run_at. A status of pending_pause means the scout still runs but the system has flagged it to pause soon (cause in pause_reason); any config edit clears the warning. One caveat before reporting "it's live": runs are gated by the signals-scout feature flag, not by enabled. A project that was enrolled and later drained from the flag keeps its enabled: true rows, but the coordinator no longer plans runs for it — so a stale or null last_run_at on an enabled scout usually means the project is no longer enrolled, not that the scout is idle.

  last_run_at is a dispatch stamp, not proof a run executed. The coordinator advances it the moment it enqueues a child workflow for a due scout — before any worker picks the run up. Child dispatch is fire-and-forget, so if workers are saturated or down the children just queue and no run ever materializes, yet last_run_at keeps marching forward each tick. So a recent last_run_at means "dispatched this tick," not "a run is genuinely happening." The authoritative liveness signal is the newest actual run row in scout-runs-list, not the config stamp. Cross-check them: if last_run_at is fresh (minutes ago) but no run row has appeared for that scout in well over its run_interval_minutes, the fleet is dispatching but not running — workers backed up / down, or runs stranded — a real reliability problem, not a live scout. Don't report "it's running" off last_run_at alone.

A scout that is enabled: true but emit: false is the most common source of "my scout isn't doing anything" confusion: it is running and reasoning every tick, it just isn't allowed to post reports yet. Always surface the emit posture when reporting on a scout.

See references/scout-data-model.md for every field on a config, run, and scratchpad entry, the run status values, and how the pieces link together.

"What scouts do I have / what are they doing?" — lead with config-list, then enrich with the most recent run per scout so the user sees liveness, not just configuration.

1. scout-config-list — the roster.
2. For each enabled scout, scout-runs-list and pick the newest run with a matching skill_name (runs come back newest-first across the whole fleet, so a single call usually covers everyone). Report status and how long ago it ran.

Present it as a table the user can scan — scout, cadence, posture, last run, last outcome — and call out anything anomalous (never run, last run errored, stuck in dry-run for a long time).

"How does my error-tracking scout work / how is it doing?"

1. Read its config — find the row in config-list for signals-scout-error-tracking: schedule, posture, last run.
2. Read its body — posthog:skill-get {"skill_name": "signals-scout-error-tracking"} returns the team's actual instruction set (which may be a canonical default or a diverged, hand-edited row). This is what the agent is told to do every run — its signal-vs-noise discriminator, explore patterns, and disqualifiers. To understand why a scout behaves the way it does, read its body.
3. Read its recent runs — runs-list with text set to the skill's domain, or just scan the newest runs and filter to its skill_name. The end-of-run summary on each run is the scout's own account of what it looked at and decided.
4. Read what it remembered — scratchpad-search (see below). The memory entries a scout wrote reveal the baselines and noise it has internalized about this project.
5. Read what it was told — scout-notes-list {"skill_name": "signals-scout-error-tracking"} returns the steering notes humans left for this scout plus the general fleet-wide ones — exactly what its runs read as prior context. A behavior change that doesn't trace to a skill edit or a scratchpad entry often traces to a note. When asked to steer a scout with a note (rather than observe), hand off to the authoring-scouts skill, which covers the notes channel's write side.

scout-runs-list returns the most recent runs across the whole fleet, newest first (capped at 100). Use it to answer "what happened lately?"

• Scope to a window with date_from / date_to (ISO-8601; inclusive lower, exclusive upper on created_at). Walk backwards by passing an earlier date_to.
• Search summaries with text — a case-insensitive substring match on each run's end-of-run summary. This is how the headless scout dedupes, and it's how you find "did any run already look at the checkout error spike?"
• Filter by output with emitted — emitted=true returns only runs that authored at least one report (or, on legacy runs, emitted a finding), emitted=false only the runs that authored nothing. This is the direct way to answer "which runs actually wrote something?" without parsing prose. One caveat: a run that only edited an existing report doesn't count as emitted=true — check edited_report_ids before calling such a run quiet.

Each summary row carries run_id, skill_name, skill_version, status, started_at, completed_at, emitted_report_ids / edited_report_ids (the reports the run wrote or edited — its output), emitted_count / emitted_finding_ids (the legacy signal-channel tally — 0 / empty on current scouts), task_url (a deep-link into the Tasks UI for the full transcript), and the summary prose. Lead with the summary when narrating to the user — it's the scout's own plain-language close-out — and always offer the task_url for the full reasoning.

When the user wants the full story of one run (or pastes a run id / Tasks URL):

scout-runs-retrieve
{ "id": "<uuid>" }

Note the field name flip: runs-list returns each run's id as run_id, but runs-retrieve takes it as id. Pass the run_id value through as id.

Returns the full run: status, started_at / completed_at (compute duration from these), skill_name / skill_version (what ran, at what body version), the end-of-run summary, emitted_report_ids / edited_report_ids, and task_url. The transcript — the actual tool calls and reasoning — lives in the Tasks UI behind task_url, not in this payload; hand the user that link when they want to see every step. A failed run returns an empty summary and no error field — the payload looks the same as the list row, so to learn why it failed you need the transcript.

You don't have to open the UI for that: tasks-runs-session-logs-retrieve returns the run's session log (every tool call, message, and reasoning step) as data — handy when you're diagnosing a failure or want to trace exactly what a run did without leaving the conversation. Pass the run's task_run_id as id and its task_id (both are on the run row).

The raw stream is large (hundreds of KB to a few MB) and will overflow inline, so fetch it with call --json and let it spill to a file, then run it through scripts/render_run_report.py rather than parsing it by hand.

⚠️ Do not reach for exclude_types: "tool_call_update,…" to slim it down. It is tempting — the stream is dominated by incremental tool_call_update chunks — but each tool's actual input lives only in those chunks: the base tool_call event carries an empty rawInput, and the streamed updates build the input (and the final rawOutput) token by token. Excluding them leaves you with tool names but no idea what the scout actually queried. Fetch the full log and let the script reassemble each call (it groups by toolCallId, keeps the richest rawInput, and attaches the completion's rawOutput/status).

Whether a run wrote anything is a first-class field: emitted_report_ids / edited_report_ids. A non-empty emitted_report_ids lists the reports the run authored via emit_report, in order; edited_report_ids lists the reports it mutated via edit_report (which can target any inbox report, not just ones a scout authored). A productive run typically has one id there and a summary like Report authored: <id>; resolve any id via inbox-reports-retrieve to read the report itself. Don't parse the prose summary for output — a phrase like "already reported P1 … did not re-file" describes a prior run, so substring-matching the summary is unreliable; the id columns are the authoritative tally. One scout shape breaks the equivalence between "no report ids" and "wrote nothing": a measurement scout files no report on a normal run because its output is the record stream, so check metadata.derived.has_structured_output (and the events themselves) before calling such a run empty.

Legacy runs: emitted_count / emitted_finding_ids. Runs from the deprecated signal-emitting channel (a scout without the allowed_tools opt-in — an old custom scout, or a canonical scout not yet ported) tally their output as emitted_count weak findings instead; each finding_id maps to a Signal with source_id = run:<run_id>:finding:<finding_id>. For those runs only, scout-runs-emission-reports (pass the run_id) maps each emitted finding to the inbox report its signal grouped into (or null if it never surfaced). On report-channel scouts these fields are always 0 / empty — don't diagnose off them. See references/scout-data-model.md for the full field reference.

A run with status complete and an empty-handed summary ("surface at baseline, nothing to report") is a healthy outcome, not a failure — most runs should close out empty. Treat a stream of empty close-outs as the fleet doing its job, not as the fleet being broken.

The scratchpad is the fleet's durable, per-team memory — prose entries scouts write so future runs get smarter and quieter. Reading it tells you what the fleet believes about this project.

scout-scratchpad-search
{ "text": "error_tracking" }

Returns entries newest-first (capped at 100); text matches content and key case-insensitively. Omit text to browse everything. Each entry's key carries a category prefix that tells you what kind of learning it is:

This is the common vocabulary, not a closed set — scouts coin their own prefixes and <domain> labels as needed (the live fleet uses watch: heavily, for example), so treat an unfamiliar prefix as just another category. Entries cross-reference each other with [[key]] wikilinks. Keys follow <prefix>:<domain>:<entity> (e.g. dedupe:error_tracking:019e8375-…).

When a user asks "why isn't my scout flagging X anymore?", search the scratchpad for noise:, addressed:, dedupe:, and allowlist: entries — the fleet may have deliberately learned to suppress it. The canonical prefix vocabulary and the four-state dedupe classifier the fleet reasons in terms of are documented in the authoring-scouts skill (references/dedupe-and-memory.md).

Custom scouts self-report skill improvements. A custom (team-authored) scout is invited by the harness to write an improve:<skill-name>:<topic> entry when a run produces concrete evidence its own skill body steered it wrong — the suggested change, the evidence, and a dated observed line, re-confirmed in place on later runs. A custom scout on the report channel escalates recurring or material suggestions further: it files an inbox report about itself (titled Scout self-improvement: <skill-name> – <topic>) and stashes the report_id in the improve: entry — so the strongest suggestions reach the owner through the inbox like any other report, not only via the scratchpad. When assessing a custom scout, search {"text": "improve:"} and surface these to the user: an entry re-confirmed across several runs is the highest-signal edit the owner can make. Reviewing and applying them is a write operation — hand off to the authoring-scouts skill. Canonical scouts never write improve: entries (their skill bodies are synced from PostHog's fleet), so an improve: entry under a canonical scout's domain is itself worth flagging. Instead, a canonical scout routes skill-content gaps upstream to the PostHog team via the agent-feedback MCP tool (feedback_type: "scout"), generalized so no project data travels, and keeps a reported:<skill-name>:<topic> entry as its local record of what it already submitted — so a reported: entry tells you a gap has been raised with PostHog, not with this team.

Scout output reaches the user as inbox reports. Filter the inbox to the scout source:

inbox-reports-list
{ "source_product": "signals_scout", "limit": 20 }

This is the direct way to find the reports scouts authored. Every report a scout authors carries backing signals tagged source_product="signals_scout", and the inbox filter keeps any report whose contributing signals include that tag — so the result is the set of reports the fleet has authored. It does not capture edit-only work: a scout that edits an existing non-scout report (appending a note to a pipeline report, say) adds no signals_scout signal, so that report won't match the filter — trace edits through the run rows' edited_report_ids instead.

An empty result means the fleet hasn't authored any reports (yet), not that the filter is broken. Scouts hold a high bar — most runs close out without writing — so on a quiet or newly enrolled project zero scout reports is the normal, expected state. Note the inbox only shows surfaced reports: a report the safety judge suppressed (or one filed as not_actionable) persists with status SUPPRESSED but doesn't appear in the default inbox view.

For the per-run view, work from the runs instead: scout-runs-list?emitted=true lists every run that authored a report, and each run's emitted_report_ids / edited_report_ids name exactly which reports it wrote or updated — resolve them via inbox-reports-retrieve. The flip side matters when explaining a gap: a run can narrate "authored a report" in its summary yet have the write silently dropped by a preflight gate (dry-run at the time, the org hasn't approved AI processing, or the signals_scout source is disabled) — those leave emitted_report_ids empty, so a claimed-but-absent report is itself a diagnostic. To browse the inbox more broadly, use the inbox-exploration skill (statuses, suggested reviewers, drilling into a report's underlying signals). The report contract behind each report — the report bar, evidence, actionability, reviewer routing — is documented in the authoring-scouts skill (references/report-contract.md).

"Is my scout actually working / earning its cost?" There's no single metric — judge a scout over a window of runs. Pull the runs (runs-list with a date_from), then reason across the dimensions below. The full playbook, including how to read each signal and the common failure modes, is in references/assessing-performance.md.

• Cadence adherence — are runs landing roughly every run_interval_minutes? Large gaps mean the coordinator is skipping it (disabled, drained from the flag, or capped out on busy ticks) — or it's dispatching but the runs aren't materializing. Tell the two apart with last_run_at: if the config's last_run_at is also stale, the coordinator stopped planning it; if last_run_at is fresh but the newest run row is hours old, it's the dispatch-vs-execution divergence above (workers backed up / down, or runs stranded), which runs-list alone hides.
• Success rate — how many runs reach a clean status vs. error out? A run of errors is a broken scout, not a quiet one.
• Report rate — what fraction of runs wrote or edited a report vs. closed out empty. Read it straight off emitted_report_ids / edited_report_ids per run (or split the window with runs-list?emitted=true / ?emitted=false, remembering edit-only runs read as not-emitted). Near-zero over a long window on a live surface can mean the discriminator is too strict (or the surface really is quiet); near-100% usually means it's too noisy. Don't score a measurement scout on this at all — a near-zero report rate is its designed behavior, so read its record stream instead. Most healthy scouts write rarely.
• Signal-to-noise — of what it wrote, how much surfaced as actionable inbox reports vs. got suppressed or dismissed? Resolve each run's emitted_report_ids via inbox-reports-retrieve and read the report statuses — across a window, the share of authored reports that are live and non-suppressed is the scout's hit rate.
• Memory growth — a healthy scout accumulates pattern: / noise: / dedupe: entries over time. A scout with an empty scratchpad after many runs isn't learning.

The skill bundles three pure formatters under scripts/ for the most common asks. They do no network I/O — they are the back half of an "agent fetches, script formats" split. The pattern is always the same:

1. Fetch each payload with the MCP using call --json (raw JSON, not the pretty text format) and save it to a file. For the big ones (runs-list, tasks-runs-session-logs-retrieve) this is mandatory anyway — they overflow inline and spill to a file you can point the script at.
2. Run the script over those files.

All three are stdlib-only Python 3.11+ and print plain text to stdout (or --out) — designed to read well in a terminal, so save them as .txt.

• Always surface the emit posture. "Running but in dry-run" is the single most common reason a user thinks a scout is broken when it isn't.
• An empty close-out is success. Most runs should find nothing. Don't report a wall of clean, empty runs as a problem.
• What a run wrote is a first-class run field. Read emitted_report_ids / edited_report_ids per run (or filter with runs-list?emitted=true) to find what was written, without parsing the prose summary. The source_product: "signals_scout" inbox filter lists the reports the fleet surfaced; an empty result there means it hasn't written anything yet (scouts hold a high bar), not that the filter is broken.
• emitted_count: 0 does not mean "did nothing". emitted_count / emitted_finding_ids are legacy signal-channel fields — they stay 0 / empty on report-channel scouts, productive or not. Judge output by the report columns; a non-zero legacy tally means the run came from a scout still on the legacy channel, and is that run's real output.
• A ~30-min run that failed is usually a timeout, not a broken scout. Completed runs finish in a couple of minutes. Most often the scout over-investigated and ran the full budget (the fleet self-corrects by writing "tight-run recipe" scratchpad entries) — but some are false timeouts where the scout actually finished in a few minutes and the run then hung on a dropped close-out. The session log (above) tells them apart: real over-investigation shows tool calls right up to the wall; a false timeout goes silent long before it. Don't assume over-investigation from duration alone.
• Lead with the run summary, then offer task_url for the full transcript — don't dump raw run rows at the user.
• last_run_at: null means a scout has never fired — check it's enabled and the project is enrolled before digging further.
• To explain a quiet scout, read the project profile. scout-project-profile-get shows whether the surface it watches is even in use — a logs scout on a project with no logs has nothing to do.
• This skill is read-only. To change a scout's schedule, posture, or body, hand off to the authoring-scouts skill — it covers scout-config-update and the skills-store edit path.