Working With Scouts

Working with Signals scouts

A scout is a scheduled agent that wakes on its own interval, looks at one PostHog project, and either writes a finding into the Signals inbox as a report or closes out empty. Think of the fleet as a team of junior analysts you've hired to watch things for you: they work unattended, they hold a high evidence bar, and — critically — they get better the more you work with them. Every dismissal reason you write, every note you leave, every report you act on feeds back into what they do next.

This skill is the operating manual for that working relationship: how to delegate a watching job, how to act on what comes back, and how to steer the fleet so it converges on what your team actually cares about. Three sibling skills carry the mechanics — reach for them when a workflow below hands off:

First: is the fleet running?

Don't delegate to a fleet that isn't there. Two reads answer it: posthog:scout-metadata-get says whether the project is enrolled to run scouts at all, and posthog:scout-config-list is the roster — one row per scout with its schedule, enabled, emit posture, and description. Check enrollment first, whatever the roster shows — config rows outlive enrollment, so a drained project can carry a roster of enabled scouts that never run (stale last_run_at across the board is the tell). (Scout tools were recently renamed from signals-scout-* to scout-*; if a scout-* name comes back unknown, try the legacy signals-scout-* name.) One access rule covers everything here: scout rows live on the project's canonical parent, so every scout read and write — this roster read included, plus the notes and config steering below — returns 403 for a credential scoped only to a child environment; work from the parent project (or a credential that covers it).

• Not enrolled — point the user at the Signals scout settings / PostHog Desktop onboarding rather than inventing activity.
• Enrolled, empty roster — likely newly enrolled and awaiting the first coordinator tick (configs auto-register then); say so instead of re-sending the user through onboarding.
• Enrolled, rows exist — note each scout's enabled, emit (false = dry-run: it runs but writes nothing), and status / pause_reason. A paused or dry-run scout explains most "scouts aren't doing anything" complaints before any deeper digging. Also check summary.emit_eligibility on posthog:scout-project-profile-get (pass summary_only=true when you only need the gate). When can_emit is false, scout writes cannot reach the inbox. Show its remediation before promising coverage. blocking_reason identifies the gate: ai_processing_not_approved or source_disabled applies to every scout on the team. scout_emit_disabled applies to one scout's dry-run setting. That scout continues its investigation without emitting findings or reports. Outside a run, the tool returns the team-wide answer; pass run_id to check one scout's write eligibility. (For read callers the profile is a cached snapshot built by scout runs, so a 404 means no fresh profile exists — not ineligibility; fall back to checking the signals_scout source config via posthog:inbox-source-configs-list and treat eligibility as unknown rather than blocking on the profile.)

The description on each row says what that scout watches — scan it to answer "which scout covers X?" without loading any skill bodies. (A row with an empty description is usually an orphan whose skill was since deleted — it can't run, so don't count it as coverage.) PostHog ships specialists for most product surfaces (error tracking, logs, web analytics, AI observability, experiments, feature flags, session replay, surveys, revenue, and more) plus a cross-product generalist, and teams add custom scouts beyond that (any skill name works; only the signals-scout- prefix gets a config auto-registered, so a scout named anything else comes in through posthog:scout-create). Each row also carries scout_origin (canonical or custom), which tells a PostHog-seeded scout from a team-authored one. It cannot tell you whether a canonical scout has been edited in place and so diverged from upstream: that row still reads canonical. When divergence matters, compare the row's body with the canonical source in the PostHog repository (products/signals/skills/<skill>/SKILL.md); posthog:skill-get returns the team's live row, which is the edited copy.

The working loop

Everything in this skill is one loop, run continuously:

1. Delegate — decide what the fleet should watch, and get the right scout watching it.
2. Receive — reports land in the inbox, routed to a suggested reviewer when the scout could name one.
3. Act — verify the finding, fix it (or decide not to), and record the outcome on the report.
4. Feed back — the outcome and your written reasons flow back to the scout, along with any notes you leave.
5. Calibrate — periodically review fleet health and promote recurring steers into permanent policy.

The single most important habit: never let a report just sit. Acting on reports — even dismissing them with a reason — is what trains the fleet, and a scout whose inbox reports nobody ever touches is automatically warned and then paused (pause_reason=ignored; Slack-delivered scouts are exempt, since consumption there can't be measured). An untended inbox doesn't just decay; it switches the fleet off.

Delegating a job

When you want something watched, pick the cheapest path that gets it watched — most jobs don't need a new scout:

references/delegation-recipes.md has worked recipes for the common asks — watching a freshly shipped event, a time-boxed funnel watch, a daily digest, an external status page, quieting a noisy fleet, and more.

Two delegation habits that pay off:

• Ground the job in real data first. Before pointing a scout at an event or surface, confirm it exists (posthog:scout-project-profile-get, posthog:read-data-schema) and actually has volume (a quick posthog:execute-sql count over a recent window — the profile and schema tools don't return counts for a new or rare event) — a watch on data the project doesn't capture is dead on arrival.
• State the job in terms of what's worth interrupting a human for. Scouts hold a report bar ("would you own this finding end-to-end?"); a steer like "tell me about anything interesting" produces noise, while "tell me when checkout conversion drops while entrants hold steady" produces signal.

Acting on what comes back

Report triage mechanics live in inbox-exploration; what matters here is how acting doubles as steering.

• Verify before implementing. A scout report is an LLM diagnosis, not ground truth — confirm the cited entities and behavior against the live data or code before fixing. A report that doesn't hold up is a dismissal candidate, and dismissing it well is valuable work (see below).
• Close every report with the honest state: resolved when the work landed (PR-backed fixes resolve themselves on merge — don't resolve at PR-open time), suppressed (dismissed) when it's not real or not worth fixing, potential (snoozed) when it's real but deferred.
• The dismissal note is a steering message. On a dismiss, snooze, or restore, the dismissal_note is forwarded to the scout that filed the report, and every future run reads it as prior context (a wrong_repo dismissal forwards even without a note, since the repositories it names are the feedback). Write it for that reader: name the evidence that settles it ("staging traffic — hosts match *.dev.example.com, ignore this pattern"), not just the verdict. A well-written dismissal is the cheapest scout edit you will ever make; a bare dismissal teaches nothing and the report comes back. Two caveats: forwarding is best-effort and requires the dismisser to hold scout-steering (skill-editor) access on the canonical project, and a dismissal made on a child environment's report never forwards at all; in both cases the note still lands on the report but never reaches the scout, so for a steer that must stick, confirm it arrived (posthog:scout-notes-list, with a credential that also holds task:read, since derived notes are hidden without it) or have someone authorized leave a note directly. The forwarded note also expires after ~30 days — it becomes durable only if the scout folds it into scratchpad memory, so a steer that must outlive that belongs up the ladder as a skill edit.
• Three more inbox actions steer the same way. A question typed into a report's "Discuss" box, a note left with a thumbs rating on a report, and adding or removing a suggested reviewer each also become a scout note (visible in posthog:scout-notes-list with an origin of report_discussion, report_feedback, or report_reviewer_correction). A reviewer correction is the strongest routing evidence the fleet gets: it reaches the scouts that filed or edited the report and the scouts whose reviewer: memory names a removed login (capped at twenty targets per correction, so on a very busy report a few can miss it), so fixing a misrouted report in place teaches the fleet who owns the surface. The forwarded note carries GitHub logins and is written only when the corrector holds skill-editor access on the canonical project, so a correction made without that access, or one that only adds or removes a user_uuid reviewer with no linked GitHub account, stays on the report and is not forwarded. The discussion and rating paths demand the full notes-write authorization (skill-editor access plus the signal_scout:write / llm_skill:write key scopes), so a note typed by someone without it is not forwarded: a discussion question still lives on the report's thread, but a rating note survives only in the analytics event, so if it must reach the scout, have someone authorized leave it as a scout note.
• Resolving a report is not the end of it — a check measures whether the fix held. A check is a follow-up measurement a scout (or the report pipeline) attaches to a report: one expectation, and a time to test it. Resolving the report starts its clock, and after a soak window — 24 hours by default, longer for a fix that reaches users slowly, like a mobile release or a cached client bundle — it re-measures, and the verdict lands on the report. So "did that actually work?" becomes a stored fact instead of something somebody has to remember to go and re-derive. Three things this changes in how you work:
  • Read a resolved report's checks before you call it done. posthog:inbox-report-checks-list shows each check's status and last_outcome. A check still open means a verdict is coming, so there is no need to re-measure by hand. Reading the rows is exploring-scouts' job.
  • A failed check means the fix did not hold, and the relapse is filed as a fresh report linked back to the resolved one rather than reopening it — a resolved report has left the inbox, so nothing there would be read. Treat that new report as the live item.
  • Resolve honestly, and resolve on the merge. A check dated from a premature resolve measures the window before the fix shipped and can fail a fix that worked. Checks are written only by scout runs and by the pipeline; there is no create surface for a person, so if a resolved report should be re-measured and carries no check, the lever is the scout — a note asking it to attach one, or a skill edit via authoring-scouts.
• Reports route to people. A scout that can name a plausible owner sets suggested_reviewers, and the inbox floats those reports to the top of that person's view. A reviewer is a PostHog user: a scout routes by user_uuid (any org member, no GitHub account needed) or by github_login (matched against the member's linked GitHub identity), and is_suggested_reviewer flips for the viewer on either match. If reports for a surface keep landing unrouted or misrouted, that's fixable: correct the reviewers on the report itself (the correction is forwarded as above), leave a fleet-wide routing note (posthog:scout-notes-create with no skill_name, "route billing-adjacent reports to Dana"), or steer the scout (note or skill edit) toward the right owner for the area. A pipeline:report-research note steers only the reports the pipeline builds from clustered signals; a scout that authors reports directly sets suggested_reviewers itself and never reads that audience.

Auditing what a scout changed

A scout that holds write_scopes (granted via authoring-scouts) changes real objects in the project, and each change lands in the project's activity log like any other edit. The row names the scout's acting user, the person whose identity the run mints its token as, and carries a small "via MCP" tag next to the timestamp. The row does not name the scout, the run, the skill, or the scopes it held. A scout editing a dashboard and the same person editing one from an MCP client produce identical rows. So auditing a run is a reconstruction from the run window, not a lookup by scout.

To reconstruct one run's changes:

1. Read the run (posthog:scout-runs-retrieve) for started_at, completed_at, metadata.write_scopes (present only when the run actually held a grant), and the close-out summary. The run prompt asks a granted scout to name every object it changed.
2. Bracket the window (posthog:advanced-activity-logs-list) with start_date and end_date around the run, clients: ["mcp"], and scopes for the objects the grant covers. Raise page_size: the tool defaults to 10 rows and the endpoint takes up to 1000, so the default can hide most of a run behind one page. A non-null next means rows are missing, and the tool cannot send that cursor back, so page through with page instead, which also returns the total count.
3. Read the actor off the returned rows instead of filtering by it up front. users takes user UUIDs, and the acting user is derived rather than configured (the scout skill's earliest known version author, then the config's enabled_by, then its created_by), so it is easier to confirm than to predict. The rows name the actor but carry no UUID, so resolve one from the email they show (posthog:org-members-list, which searches on name and email and needs organization_member:read), then add users to drop other people's rows.
4. Cross-check against the close-out. A row the summary does not mention, or a change the summary claims with no row behind it, is the thing to look at.

Which scopes value each granted scope writes under:

Pass Notebook as well, whatever the grant says: notebooks are the floor write every scout holds, so any scout can leave rows under that scope. Some grants also reach objects they are not named for: both warehouse grants reach the data quality checks on their subject, and llm_skill:write reaches the skill-store install command, which mints or rotates the acting user's marketplace credential. The scope-named objects themselves still log nothing: a skill body edit (including another scout's) and a warehouse table write leave no row, so read the skill's version history instead.

Four caveats change what the answer means:

• The window is not an attribution. The same filter also catches the acting user's own MCP writes in that window, from Claude Code, Cursor, or any other MCP client. Narrow by scope and timestamp, then read the rows.
• The advanced log is gated twice. advanced-activity-logs-list and advanced-activity-logs-filters need activity_log:read on the credential, and on PostHog Cloud they also need the organization's audit-logs entitlement. Self-hosted is never gated on it. Without the entitlement the MCP server drops both tools from the toolset rather than failing them, so they read as tools that do not exist; a call that still reaches the backend gets a 402 asking for a paid plan. Neither outcome means you built the filter wrong. With the entitlement, results are trimmed to the plan's lookback, so an older run can fall outside the window. The plain side-panel feed is gated on neither, but it is a weak substitute: it filters only by user, scope, and item, so it cannot isolate MCP writes. It still shows the "via MCP" tag, so on a short run window you can read an object's own feed and narrow by eye.
• A dry run drops the grant, not the floor. A scout on emit: false never holds the granted scopes, so it writes no rows under the scopes in the table above. It keeps notebook:write, the floor write every scout holds, so a dry run can still create, edit, or delete a notebook. Keep Notebook in the filter for a dry-run window.
• A refused write is not logged, because it never happened. The grant is an upper bound and the acting user's own permissions still apply, so a close-out that reports a refused write will have no matching row. That is the expected pairing, not a discrepancy.

The steering ladder

When you want a scout to behave differently, climb this ladder from cheapest to most permanent — and stop at the lowest rung that does the job:

1. React to its output. Dismiss / snooze with a specific, evidence-bearing note, ask a question in its Discuss box, rate it with a note, or fix its reviewers (each forwarded to the scout automatically). Right for: one wrong report, a known-noise pattern surfacing for the first time, a misrouted report.
2. Steer one run (posthog:scout-run-now with a note). Right for: a one-off focus you want checked now ("look at the checkout regression") without leaving anything the scheduled runs will read later. The note steers that run only (up to 1,000 characters) and is never delivered to later runs as a note, though it stays visible in that run's metadata when a later run reads its history, so phrase it as a dated one-off ("today only: ..."). It spends a run from the daily budget like any manual run, and it needs llm_skill:write on top of signal_scout:write plus skill-editor access on the project (a 403 otherwise). A caller without that access can still run the scout without a note.
3. Leave a note (posthog:scout-notes-create, per-scout or fleet-wide, optionally time-boxed with expires_at). Right for: feedback, pointers, and context with a shelf life — "the spike you keep flagging is known noise", "dig into EU signups this week", "new checkout shipped Tuesday". Notes are advisory: they direct attention but never lower the evidence bar or force a report. Address skill_name: "pipeline:report-research" to steer how the pipeline researches, judges, and routes the reports it builds from clustered signals. Scout-authored reports never pass through that stage, so a routing rule for them goes in a fleet-wide note (no skill_name) or a per-scout note.
4. Tune the config (posthog:scout-config-update). Right for: when and whether it runs, and what it can reach, but not what it investigates; slow a chatty scout (run_interval_minutes; if the config carries a run_cron_schedule, that takes precedence, so update or clear it too), pause one (enabled=false), dry-run a risky one (emit=false), grant external reach (network_access=full), give a code scout a checkout (repositories=["organization/repository"], read-only), exempt a deliberately quiet watchdog from auto-pause (auto_pause_exempt=true), deliver its reports to a Slack channel or DM as well as the inbox (output_destinations.slack), or pin the model it runs on (model).
5. Edit the skill body, or author a new scout (via authoring-scouts). Right for: permanent policy — a disqualifier, a threshold, a scope change, a new surface. For a custom scout this is the strongest steer there is: the skill body is yours, edit it freely — it's where recurring notes and repeated dismissal reasons should end up. For a canonical scout the same edit forks it: your team's copy is marked diverged, PostHog's canonical sync leaves it alone from then on, and you stop receiving upstream improvements to that scout — you maintain it yourself. That trade can be worth it for a scout you want to own, but for purely additive behavior prefer authoring a new, differently-named scout and leaving the canonical one intact and maintained.

Promote steers that repeat. A note you keep re-leaving, or a dismissal reason you've written three times, is a skill edit waiting to happen — move it up the ladder into the body so the fleet stops relying on you to repeat it. The reverse also holds: try a nudge as a note before committing it as an edit. Note hygiene stays with humans — scouts never delete notes, so retire acted-on ones (or set expires_at up front) to keep the channel high-signal.

What the fleet learns without you

Some feedback loops run on their own — knowing they exist changes how you work:

• Scratchpad memory. Scouts write durable per-team memory (baselines, noise patterns, dedupe gates, allowlists) and get quieter and sharper across runs. When a scout stops flagging something, check the scratchpad (posthog:scout-scratchpad-search for noise: / addressed: / dedupe: / allowlist: entries) before assuming it's broken — it may have deliberately learned to suppress it.
• Auto-pause. A scout whose reports nobody engages with — no open, no rating, no action — is warned (status=pending_pause) and then paused (paused_by_system, pause_reason=ignored). Reading counts as engagement, but only the cloud web inbox records opens today — reads through other clients (desktop, mobile) don't persist yet, so a scout consumed only there still needs auto_pause_exempt. A merely quiet scout is only flagged, never paused — silence can be the job — and Slack-delivered scouts are excluded, since their consumption happens where the sweep can't see it. Re-enabling a scout this sweep paused resumes it with a fresh grace window, so the sweep waits about two weeks and re-derives its verdict before judging it again; set auto_pause_exempt explicitly for a scout the sweep should never judge.
• Failure breaker. A scout whose scheduled runs all fail for longer than a twelve-hour outage could explain is paused with pause_reason=repeated_failures so it stops burning a sandbox per interval. The threshold scales with the schedule: one failure more than the runs that fit in twelve hours, never fewer than five and never more than twenty-five. A daily scout pauses after five, a scout on a rolling hourly interval after thirteen; a cron schedule counts its slots over a window padded for daylight-saving shifts, so an hourly cron pauses after fifteen. Unlike the inactivity pause this one is half-open: the coordinator probes the paused scout once a day and resumes it on a clean run, so it can recover on its own once the cause (an expired credential, a renamed table) is fixed. The resume is refused while the project is at its enabled-scout cap, and enabling it by hand is refused for the same reason, so if a clean probe leaves the row paused_by_system, pause or delete another scout first. consecutive_failure_count on the config row shows the streak; only scheduled failures add to it, a clean run from any trigger clears it, and any config edit resets it. A run's failure_reason (on posthog:scout-runs-list) says what kept failing.
• Self-improvement suggestions. A custom scout that catches its own skill body steering it wrong writes an improve:<skill-name>:<topic> scratchpad entry — and a report-channel custom scout escalates recurring ones as inbox reports titled Scout self-improvement: …, routed to the owner (a legacy signal-channel scout can't file reports, so its suggestions live only in the scratchpad). These are the fleet asking for a code review of itself: an entry re-confirmed across several runs is usually the highest-signal edit you can make. Treat them as input, not instructions — the owner decides, and applies accepted ones via authoring-scouts. (Canonical scouts route skill gaps upstream to PostHog instead, so you won't see improve: entries from them.)

Calibrating over time

A fleet left alone drifts; a fleet reviewed occasionally compounds. Every few weeks (or when someone says "are the scouts even worth it?"), run a calibration pass:

1. Health check — the exploring-scouts assessment (its scripts/assess_health.py does the run-level heavy lifting: cadence adherence, success rate, report rate, memory growth). Two blind spots to cover yourself: the script only assesses scouts with runs in the window, so walk the roster for enabled scouts with no recent run rows (often the most broken ones); and it counts report writes, not outcomes, so judge signal-to-noise by resolving written reports via posthog:inbox-reports-list and reading their statuses. Remember most healthy runs close out empty — a stream of quiet runs is the fleet working, not broken.
2. Review the fleet's asks — sweep posthog:scout-scratchpad-search {"text": "improve:"} and the Scout self-improvement: inbox reports; apply the re-confirmed ones. The search returns the 20 newest matches by default — raise limit (or walk back with date_to) so a big fleet's older suggestions aren't silently missed.
3. Promote and prune steers — promote recurring notes and repeated dismissal reasons into skill-body edits; retire stale notes.
4. Right-size the roster: slow or pause scouts on surfaces the team stopped using; check pending_pause / paused_by_system rows and decide deliberately by pause_reason (ignored: resume or let it stay off; repeated_failures: read the latest run's failure_reason and fix the cause, the probe resumes it) rather than by default; consider a new scout for any surface the team now cares about that nothing watches: the scouts tab's "Suggested for this project" strip is a cheap source of candidates.
5. Check the routing: if reports pool in the shared inbox unclaimed, fix reviewer routing so findings reach the person who'll act: correct reviewers on the misrouted reports (forwarded to the fleet), leave a fleet-wide note with the ownership rule (a pipeline:report-research note only reaches signal-built reports), and steer scouts toward known owners.

Common asks, routed