# Tara service discovery, saved items and launch · delivered increment · 8 September 2026

> Rewritten on 11 September 2026 as _Tara · find a practice_ chapter four. The
> subjects below are unchanged; the treatments, titles and prose are not. See
> [tara-discovery-chapter.md](tara-discovery-chapter.md).

This package continues the [full V1 program](v1-completion.md) after the
delivered [local web-search guide](v1-tara-web-search.md). Baseline:
`6830fb7b036c002a6453589593ac825645049538`; primary source pin:
`1ceae52c7896a59ac74a4429d410f3218f0b91c6`.

Visual thesis: follow a concrete service-backed recommendation through a search
result, saved-item readback and resolved launch descriptor. Native owner
exchanges, pool/record diagrams, filter and reconciliation branches, page walks
and exact query specimens should expose the mechanism without inventing a UI.

Content plan: distinguish direct recommendations from canonical search; trace
BFF admission and role selection; explain configured search pools, order,
filtering and canonical objects; inspect recommendation/course/favorite card
semantics and freshness; follow explicit add/remove intent and scoped favorite
reconciliation; resolve action, path, query and links; close with the actual
consumer and service evidence still required. The existing catalog, web-player
and session guides retain their own mechanisms. Full selection policy and
native/application consumers remain required in the broader program.

Interaction thesis: keep selectable native diagrams, source and explanation
reveals and the existing reduced-motion navigation. All relationships remain
complete in PDF. No new shared renderer or decorative animation is needed.

## Current source findings

Read the complete canonical adapter, domain adapter, card builder, role adapter,
BFF role wrapper, dedicated Tara routes, favorite reconciliation module and
launch-action/deep-link helpers. Prior session-launch review is being reconciled
with the actual navigation owner. The service transport and API/database
consumers must remain identified by actual calls, not similarly named modules.

The canonical search requests at least its configured recommendation pool size,
then optionally reads courses and favorites only when searchUserId is
configured. It builds recommendation/course/favorite results in that order,
filters by intensity, duration and AND substring tokens, then limits the result.
Course cards lack duration and intensity; favorites lack intensity. Empty tokens
match all otherwise eligible records. ritualMoment is declared on the search
input but is not used in this method. Bridge recommendations have a separate
intensity score table; a zero-scored high-intensity sleep item is not explicitly
excluded.

The BFF wrapper creates the registry with only apiAdapter, so canonical search
has no configured searchUserId. The authenticated route user is not passed into
that search call. Direct /recommended does forward userId/category/intensity to
the BFF service adapter. The domain library HTTP client's recommendation method
forwards only limit; these two clients are distinct implementations.

Role capability arrays are declarations, not runtime object filtering. The
registry returns the same adapter instance for shell, admin and assistant. The
assistant capability array includes saved_items while its TypeScript Pick omits
getSavedItems. The BFF accepts a role selector after Tara-domain admission; that
selector is not itself an administrator authorization check. The actual
abuse/auth prehandlers and downstream service must be described with their own
scope; a controlled route-guard fixture will not represent token verification.

Favorite reconciliation distinguishes complete and partial remote views. A
partial view may add/update but preserves uncovered local items. Fetching uses
200-item pages, stops on a short page, and caps at 25 pages. Duplicate IDs
across the page walk keep the first copy; direct reconciliation itself does not
remove all duplicate input rows. An exception rejects before a reconciled
snapshot is returned. A completed offset walk is not an atomic remote snapshot.

Launch resolution prefers an explicit action, otherwise a mapped card, otherwise
a supplied target or /meditate. Explicit query fields are spread last and can
override action/context metadata. Generated native/web links and parsed intent
are returned values; this method does not navigate or start playback.

## Verification and delivery plan

Execute complete pinned modules with real owning navigation/registry/contract
helpers and declared data adapters. Characterize the actual source route with
Fastify injection and explicit authentication/abuse seams, retaining calls and
outputs. Use source cases for pool exclusion, filtered ordering, object kinds,
role identity, ID/query precedence, complete/partial reconciliation, page
limits, duplicates and failure propagation. Do not label synthetic route
admission as a live authentication or database result.

After source cases settle, author a coherent guide or linked guides based on
teaching load. Use the existing standalone, embedded, accessibility, PDF and
audio avenues for all affected slides. Format all new canonical/draft/report
JSON before recording final input hashes; the prior web-search receipt documents
why normal hook formatting must be settled before publication. Preserve all 722
previous slides and recordings. Publish passing increments with normal hooks and
retain the full V1 objective as active until its complete scope is verified.

## Settled manuscript and sequence

Audience: product, design and engineering readers who have completed the Tara
overview and catalog guide. Outcome: follow service records, saved-item scope
and launch identifiers through their actual owning boundaries.

The twenty-four-slide guide sits after local web search and before web practice.
It requires the Tara overview and practice-catalog guide. Slides 1–4 orient to
the member case, distinct callers, BFF admission and capability declarations;
5–11 explain pools, filtering, result identity, object kinds, freshness and
bridge ordering; 12–19 follow favorite intent, authenticated requests, compact
records, scope, worked reconciliation, pagination, duplicates and failure; 20–24
resolve action mapping, precedence, query metadata and parsed intent, then
return to the member case. This order introduces each record before using it in
the next mechanism.

The manuscript has 3,461 note words, with three explanation paragraphs and three
reader questions per slide. Visible copy remains focused on one mechanism.
Native conditional pools, seven branch diagrams, an owner exchange and three
record specimens accompany worked comparisons and the opening/closing case.
Comparison headers name the actual columns. Peer rows have explicit null
emphasis; the owner exchange emphasizes the BFF because it supplies identity.
There is no positional third-row highlight. Existing source/explanation reveals
and reduced-motion navigation remain available.

The first draft passed all 48 desktop/320-pixel fit cases. All 24 desktop
renders were inspected in six contact sheets, with the duplicate comparison also
opened at full size. Narrow owner-exchange and conditional-pool renders were
inspected directly. Text, edges, sequence and peer emphasis were readable; final
embedded and PDF review is retained separately in the delivery receipt. No
raster UI was invented for these library mechanisms.

## Source exercise and consumer boundary

The final fixture passes 26 cases through twenty complete pinned modules,
including the actual dedicated BFF routes with Fastify injection. Twenty-two
cited or loaded paths are byte-identical between the source pin and the
implementation baseline. The additional BFF service-adapter Tara branch and
local web-search component are inspected/cross-referenced here, not claimed as
executed by this new fixture. The source report records data, middleware, facade
and fixed-clock seams explicitly.

The launch parser was read completely, including path normalization, course URI
decoding, undecoded meditation/ritual segments, query removal and hub fallback.
The owning navigation builder uses URLSearchParams and only adds a leading slash
to its path; parser normalization does not rewrite every field returned by the
generic resolver. Malformed course encoding rejects in the executed fixture.

Pinned symbol searches across apps and libs located favorite reconciliation and
action helpers in their implementations and owning tests, with an additional
workbench index-events documentation reference. They did not locate an
application invocation of fetchAndReconcileTaraFavorites in that scope. This
supports the stated unresolved consumer boundary, not a universal absence claim.
Actual application storage, retry, destination content, service/database
readback and playback remain required in the full program.

Initial harness attempts encountered an unavailable direct Fastify symlink and
cross-VM array prototype comparisons. The installed Fastify 5.8.5 package is
selected through an explicit environment override; comparisons serialize at the
declared fixture boundary, while loaded source bytes still compare exactly.
ESLint then identified an unused result assignment, which was removed. Final
lint and source cases use the settled fixture. These corrections change the test
harness only; product behavior remains the inspected source behavior.

After first-draft review, the parser comparison was tightened to show both
complete /meditate/session/{id} and /meditate/ritual/{id} patterns. This visible
label correction leaves the narration text unchanged and is covered by the final
browser and PDF review rather than reusing the draft parser image.

## Delivery evidence

The [receipt](../verification/v1-tara-discovery-2026-09-08/README.md) retains
174 final browser checks, 96 captures, 72 embedded accessibility runs, six
enlarged-text cases and three rejected missing-content controls. Focused center
review covers all 24 slides in two embedded contexts, two audio boundaries and
24 PDF pages, with all 67 prior guide HTML files and four shared UI files exact.
The settled presentation suite passes 118 cases. New tracks add 1897.670 seconds
and receive scoped decode/pause checks; all 722 earlier scripts, manifest
entries and recording hashes remain exact. No human comprehension, listening,
live member, database or full application acceptance is claimed.
