Adding a Self-driving Inbox source

The Self-driving Inbox connects a PostHog data warehouse source and syncs one table of actionable records. After each sync, the signals pipeline reads the new rows, filters them with an LLM actionability gate, and emits signals. The pipeline is driven by a static registry keyed on (ExternalDataSourceType, table), so each table needs an emitter config.

All work happens in this repository:

Do not edit the desktop app (products/desktop/) or packages/agent/packages/agent-contracts/src/inbox-types.ts for a new source. A desktop change in the same PR as a new SignalSourceProduct value fails the desktop/backend coupling check, and the inbox does not need it.

The data warehouse source itself must already exist and be implemented.

Find candidates

1. List implemented sources: iterate SourceRegistry.get_all_sources() and keep a source when type(src).get_schemas is not _BaseSource.get_schemas and get_source_config.unreleasedSource is false. Run it on a fresh origin/master worktree.
2. Call get_documented_tables() for each source. It gives table names and incremental fields without credentials. get_canonical_descriptions() gives column names.
3. Keep tables that hold one actionable record per row: issues, tickets, conversations, incidents, alerts, findings, feedback, reviews, survey responses. Skip CRM tasks and notes, message logs, blog posts and event tickets.
4. Check that each table has a human-readable text column, a stable id, and a creation timestamp. Tables with text only in nested JSON need a bespoke emitter. Tables with one row per occurrence need grouping (see below).
5. Compare against _register_all_emitters() in products/signals/backend/emission/registry.py and open PRs (gh pr list --search "inbox source in:title").

Backend checklist (per source)

Code lives in products/signals/backend/. Read emission/CLAUDE.md first.

1. Enum: add a SignalSourceProduct value and a SIGNAL_SOURCE_PRODUCT_LABELS entry in enums.py. The model's choices are callables, so a new value needs no migration. A new record kind needs a SignalSourceType value, the matching models.py SourceType choice, and a prompt pair in emission/_prompts.py.
2. Emitter: add emission/<product>_<table>.py exporting a SignalSourceTableConfig.
   • Flat tables use make_flat_emitter from emission/_common.py (see pagerduty_incidents.py).
   • Nested JSON: extract in HogQL with JSONExtractString(col, 'key') AS new_alias, using an alias that differs from the column name, or parse in Python (see jira_issues.py, intercom_conversations.py, emission/_survey.py).
   • Warehouse columns are snake_case. The import renames API fields (createdAt becomes created_at). Check the source's canonical_descriptions.py and settings.py.
   • partition_field is the creation timestamp. Set partition_field_is_datetime_string=True for string columns. The fetcher wraps the column in parseDateTimeBestEffort(toString(...)), so a native timestamp also works. Epoch columns use an expression such as fromUnixTimestamp(intDiv(toInt(col), 1000)).
   • Tables with one row per occurrence (no grouped issue id) use fetchers/grouped_warehouse.py with a GROUP BY, so one issue emits once instead of flooding the inbox.
   • Use where_clause to skip resolved, closed or empty-text rows only when the status values are certain. A filter on a column that a vendor sometimes omits makes the whole query fail.
   • Do not select PII (emails, names, user objects, IPs). Use ids, urls, statuses and severities.
   • Reuse the shared prompt pair for the record kind. Siblings use max_records=200, first_sync_lookback_days=1 and description_summarization_threshold_chars=2000.
3. Register: register_signal_source(ExternalDataSourceType.X, "<table>", CONFIG) in _register_all_emitters(). The table name must match the schema name the source returns, including casing (Issues, Cases).
4. Contract: in contracts.py add <Name><Kind>SignalExtra (one str | None per extra field, list for list fields) and <Name><Kind>SignalInput, then add the Input to the SignalInput union and the SIGNAL_INPUT_VARIANTS tuple.
5. Tests: add a row to TIER1_SOURCES in emission/tests/test_tier1_registry.py. It checks registration, the contract and the prompts. A bespoke emitter also gets a small parameterized test file for its parsing edge cases.
6. Generated types: a new enum value changes products/signals/frontend/generated/ and the MCP generated files. Run hogli build:openapi, or let the CI bot commit the regenerated types to the branch.

Check that the fetch query parses: build the SELECT the fetcher builds and run parse_select from posthog.hogql.parser on it.

Frontend checklist (per source)

1. products/signals/frontend/inbox/externalSignalSources.ts: add an EXTERNAL_SIGNAL_SOURCES entry with source, label, category, watches, dwSourceType (the exact ExternalDataSourceType string), requiredTables and sourceType (the record kind). Write watches in sentence case and follow /writing-user-facing-copy. WAREHOUSE_SOURCE_SETUP derives from this list, so it needs no edit.
2. products/signals/frontend/inbox/signalSourcesLogic.ts: add the product to every hand-written externalSourceStates: Record<...> union, in alphabetical order. There are several, and they must all equal the EXTERNAL_SIGNAL_SOURCES list.

The self-driving wizard (PostHog/context-mill), optional

The wizard's connected-tools list is hardcoded, so a new source is not offered during npx @posthog/wizard self-driving until you add it. Edit context/skills/self-driving/references/5-connected-tools.md (plus the source lists in 4-sources.md, 2-read-context.md and description.md):

1. Add the tool to the step-5 wizard_ask options.
2. Add its responder mapping: the same source_product and record kind as the emitter.
3. Add its warehouse source_type to the "already connected" detection list.
4. Credential sources are recorded as a follow-up, with no connector file. One-click OAuth sources get a connector reference (clone 5b-linear.md).

Run node scripts/build.js once to confirm the skill bundles.

created_via attribution

ExternalDataSource.created_via is set server-side from the request transport (get_event_source in posthog/event_usage.py). Callers cannot set it, and a new source needs no wiring. The upgrade from mcp to self_driving lives in _create_external_data_source in products/warehouse_sources/backend/presentation/views/external_data_source.py.

PR conventions

• Title: feat(data-warehouse): add <Source or group> as self-driving inbox sources.
• Group related sources (incidents, security, trackers, support, feedback) in one PR per group.
• PRs that edit the same lists (registry.py, contracts.py, test_tier1_registry.py, externalSignalSources.ts, signalSourcesLogic.ts) conflict with each other after one merges. Keep the entries from both sides when you resolve the conflict, and keep every signalSourcesLogic.ts union complete.

Gotchas

• source_type on the warehouse source ("Jira"), source_product ("jira") and the record kind ("issue") are three different fields.
• dwSourceType must equal the real ExternalDataSourceType string.
• A column the vendor omits for some accounts (priority, date_closed, archived_at) may be missing from the synced table. Do not select or filter on it unless it is always present.
• Every canonical actionability prompt keeps the When in doubt, classify as ACTIONABLE line, because steering injection anchors on it.