ADR 0082 — Chat RAG: its own hybrid-search tab, without renaming "Chat"

Status

Accepted.

Context

ADR 0080 (PROGRAM 28, Wave 4, front G2) left the indexing pipeline and the three Chat RAG HTTP routes ready — POST .../rag/search, POST .../rag/reindex, GET .../rag/coverage — but explicitly declared the SCREEN as Wave 5 work. Meanwhile, ADR 0078 (RN-202) had decided the sessions tab would keep the label "Chat," never "Chat RAG" as the more recent handoff asks — because, at that moment, "Chat RAG" described a capability the product didn't have (no pipeline, no citation UI). This front (Wave 5, G3) is the moment that capability starts to genuinely exist, and the question it needs to answer is: now that RAG exists, where does the "Chat RAG" label go?

Two options: rename sessions to "Chat RAG" (what the handoff literally suggests), or give "Chat RAG" its OWN tab. The two screens answer structurally different questions — sessions is conversation with an ACTIVATED agent, which spends the owner's key per turn (RN-058) and can write to the backlog, to the code, to any tool the agent has; RAG is search over an already-built INDEX, with no agent in the middle, and read-only by nature (the three ADR 0080 routes are viewer/viewer/ maintainer, none of them invokes a conversational LLM). Merging the two into the same tab would force the UI to express "this is a citation from an indexed chunk" and "this is an agent speaking" in the same thread, when the user never asked for both at once.

Decision

"Chat RAG" becomes its own tab, key: 'rag', without touching the sessions label. RN-202 remains valid as it was — sessions was never "Chat RAG," and continues not being it — but its reasoning shifts from "the capability doesn't exist" to "the capability exists somewhere ELSE." The tab enters apps/web/src/routes/project-tabs.ts (ordem: 28, right after code and before backlog) — the same neighborhood of "look at what has already been produced/indexed" that the Code tab occupies, and before Backlog, which is where new production is born.

The screen consumes the three ADR 0080 contracts without guessing their shape (apps/web/src/lib/api-client.ts/api-types.ts gain searchRag/ getRagCoverage/reindexRag and the types mirrored 1:1 from the DTO — RagSearchHit, RagChunkOrigin as a union discriminated by kind, RagCoverage). The UI has three pieces:

1. RagCoveragePanel — REAL count of indexed docs/adr/sessions against the real total, and chunksTotal/chunksWithoutVector. No "reindexed Xmin ago": the backend response has not carried that data since ADR 0080, and this screen isn't the one to invent it (RN-252).
2. Search with a scope filter (docs/adr/session pills, absence = all) and a warning when vectorAvailable: false — "keyword-only search — embedding unavailable," with the reason when the backend provides it (RN-252, the same honest degradation as RN-233 reaching the UI).
3. RagCitationCard — the citation with combined score and the two separate signals (null when the signal didn't find the chunk, never 0%, preserving ADR 0080's distinction). file origin shows path/ headingPath as text; session origin navigates to the exact EVENT via useNavigate + search: { highlightEvent } — the SAME route and parameter the Psychologist's evidence chips already use (HypothesisCard.tsx, Phase 4b) — reuse, not a second navigation path (RN-253).

The "Reindex now" button only appears for owner/maintainer, mirroring on the client the rule the route already enforces (RN-238) — same pattern as useCurrentWorkspaceWithRole that ProjectSettingsTab/ ProjectApprovalsTab already use for other maintainer gates (RN-254). Whoever doesn't have the role simply doesn't see the button, rather than seeing it disabled: reindexing triggers N calls to the project's repository and to the embedding provider, the same "changes what the product spends without asking" rule as the area parallelism cap (RN-083).

Consequences

The promise RN-202 had deferred arrives, but not the way the handoff designed it. The handoff renames an existing tab; the product opens a new one. It's the more defensible decision given the two screens never answered the same question — and the cost of being wrong (one extra tab in the registry) is lower than the cost of mixing conversation-with-agent and search-over-index in the same surface.

No deep-link by path to the Code tab. A file-origin citation shows sourcePath/headingPath as plain text, without navigating — the Code tab (PHASE 26/26b) has no mechanism today to open a specific file via route/search, and building that mechanism is out of scope for this front. It stays declared, not half-implemented: the text shows the real path, it just isn't clickable.

The maintainer gate on the client is UX only. What guarantees reindex doesn't run without the role remains the api (RequireRole, RN-238) — the screen only avoids the click the api would refuse anyway. This follows the same pattern as every other UI gate in the product (auto mode, parallelism cap, area model): none of them replaces the server-side check.