The daily cap is now provider-aware; the seeded chain resolves to a
user-egress provider, so the refusal must be seeded at USER_PLAN_CAP
rather than the smaller operator cap to trip the gate.
The daily token cap was read before the provider call and recorded only
after a successful reply, so concurrent requests could all read
`spent < cap` and all hit the provider, and an empty/sentinel reply
recorded zero spend despite burning upstream tokens.
Reserve the budget atomically before the call: a single upsert-increment
returns the new total, so concurrent requests serialise on the row and
cannot both pass the cap; over-cap refunds the reservation and surfaces
`coach.budget.exceeded` as an SSE error frame. Reconcile the reservation
to the actual token count immediately after the provider returns —
before any empty/sentinel short-circuit — so burned tokens are always
recorded. An integration test proves two concurrent requests cannot both
exceed the cap and that an empty reply still records its cost.
When the model emits only a `---KEYVALUES---` block with no leading
prose (or a malformed block missing the closing marker), the
post-strip prose is empty. The previous fallback chained back to
`rawReply`, surfacing raw sentinel markers to the user. Replace with
a structured `coach.provider.empty` error frame so the UI renders
the localised provider-error copy instead of leaking the markers.
Add a regression integration test pinning that the streamed body
never contains `---KEYVALUES---` for a sentinel-only reply, and that
no assistant message is persisted in this branch.
Closes Code-H1.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Add a text-marker sentinel parser (`---KEYVALUES---` … `---END---`)
that strips the Coach's evidence block out of the prose before it
streams to the client, validates each line against a Zod schema, and
attaches the parsed entries to `provenance.keyValues`. The persisted
envelope round-trips so the disclosure renders on conversation
reload. Hard-capped at 8 lines and 1 KB to guard against
prompt-injection padding; malformed blocks log a wide-event and pass
the prose through unchanged.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
The route gates on requireAuth (cookie or bearer), enforces the daily
token budget, runs the prompt-injection / off-topic refusal scanner,
and only then walks the user's provider chain.
The conversation is created via withIdempotency() when no
conversationId is supplied, so a retry under the same Idempotency-Key
replays the original assistant reply rather than spinning up a second
thread. Existing-thread turns skip the wrap.
Replies stream as SSE: one "token" frame per word-sized chunk, then a
"provenance" frame describing the windows + metrics the assistant
could see, then a "done" frame carrying the persisted message id.
Provider failure cascades to a single "error" frame and the assistant
message is never persisted. Conversation history is capped at 20 turns
in the prompt window via a synthetic summary placeholder so cost stays
bounded.
`buildCoachSnapshot()` reuses the analytics features pipeline so the
Coach narrates the same numbers the dashboard tiles render.
`getCoachSystemPrompt()` extends the strict-insight ground rules with
conversational framing while keeping the zero-hallucination rule.
5 integration tests pin the round-trip, refusal short-circuit, budget
429, foreign-conversation 404, and validation 422 paths against the
postgres testcontainer.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Marc-André Bombeck