ENG-1390: prefix the rule-retrieval analytics action so the collector relays it

[alecantu7]

Follow-up to ENG-1390 (merged in #330). Two-line behaviour change plus two tests.

The bug

rule_retrieval shipped and produced nothing in PostHog project 355390. The collector lambda relays an action only if it starts with anton_ or ds_connect — a prefix rule, not a list — and drops everything else while still returning HTTP 200. send_event is fire-and-forget behind a bare except, so there is no client-side signal whatsoever.

Measured against the live collector:

action=rule_retrieval        -> posthogResult: false   (silently dropped)
action=anton_rule_retrieval  -> posthogResult: true

Renamed to anton_rule_retrieval. It also matches every other action on this channel — anton_started, anton_query, anton_first_query, ds_connect_* — so the unprefixed name was off-convention as well as broken.

Scope: this fixes ONE of two gates

The lambda also allowlists property names: action, aid, engine, llm_provider, has_mdb_key. So after this, the event arrives but carries none of its data until ENG-1355 lands.

Verified by probing a passing action name with the real payload:

Sent	Arrived
outcome, when_rules, kept_rules, stop_reason, duration_ms	all dropped
llm_provider	✅ (pre-allowlisted)

So this PR is necessary, not sufficient. Merging it does not make ENG-1390's Done-when #2 pass — that needs ENG-1355, and specifically the generic pass-through option rather than adding turn_completed's 28 names, since rule_retrieval's properties aren't among those either.

Tests

Two, deliberately in their own class. TestRuleRetrievalObservability autouse-stubs _emit_rule_retrieval — which is exactly what these must not have stubbed, since they assert what reaches send_event. I hit that while writing them: the first version silently exercised the stub and failed with KeyError: 'action'.

• test_the_action_name_carries_the_prefix_the_collector_requires — pins the constant.
• test_the_emitted_action_is_the_prefixed_one — pins the name that actually goes out, so a refactor reintroducing a literal is caught.

Both mutation-verified against the old name:

AssertionError: 'rule_retrieval' would be silently dropped by the collector
AssertionError: {'action': 'rule_retrieval', 'props': {...}}

Suite: 1745 passed, 28 skipped.

Why a test rather than just a fixed string

The constraint lives in a lambda whose code is in no repo in the workspace — ENG-1355 documents that. So nothing in this codebase can enforce it structurally, and a test with the measured posthogResult values in its docstring is the only place the constraint can be recorded where the next person will trip over it.

Note on turn_completed

turn_completed fails the same gate (posthogResult: false) and would be fixed the same cheap way. Deliberately not touched here: ENG-1355 is in progress and its Done-when queries name the literal string turn_completed, so renaming it unilaterally would invalidate acceptance criteria someone is actively working against. Worth a conversation rather than a surprise — flagged on the ticket.

🤖 Generated with Claude Code