The commercial chat cutover

What /w/:slug/chat became, why each piece moved, and what is still owed.

The routing half of this document is superseded by the chat manager. classifyChatRoute is deleted: a regex ladder held the authority over meaning and failed by a conjugation, so one typed manager decision replaced both it and the research layer’s second semantic router on the Chat path. Everything below about EXECUTION — the Work Session, the activity trail, the report artifact, the customer-safe projection — is unchanged and still current.

The request path, end to end

/w/:slug/chat  or  /w/:slug/chat/:threadId
  └─ ResearchPage → ResearchChat
       ├─ hydrate, before the composer opens
       │    GET  …/chat/threads/:id/messages        the persisted transcript
       │    GET  …/chat/threads/:id/work-session    the Work Session snapshot
       │    └─ subscribe SSE from snapshot.lastSeq  the tail, never a replay
       └─ POST …/chat/threads/:id/messages
            └─ ChatService.postMessage
                 ├─ withTurnLock                          one turn per thread → 409
                 ├─ decideChatTurn(...)                   ONE semantic decision
                 ├─ destination === 'market_research' ────┐
                 │                                        │
                 │  (director surface, when enabled, reads first and may decline)
                 │                                        ▼
                 │                             runMarketResearchTurn
                 │                               ├─ WorkSession RUN_STARTED
                 │                               ├─ MarketResearchRunner
                 │                               │    └─ semantic market-data tools
                 │                               │       search_suppliers /
                 │                               │       search_opportunities /
                 │                               │       search_awards / …
                 │                               ├─ onEvent → research-progress.ts
                 │                               │    TOOL_CALL_START/END/RESULT
                 │                               │    with customer-safe copy
                 │                               ├─ ARTIFACT_CREATED (ResearchReport)
                 │                               └─ RUN_FINISHED
                 └─ otherwise: the general belt, unchanged

The browser renders the Work Session projection and nothing else. It never sees an operation name, a semantic code, a row count phrased as a fraction, or a statement.

Root cause of each observed incident

#	Observed	Root cause	Fixed by
1	`/research` crashed on `reading 'filters'`	`filtersFor` read `capabilities.filters` without checking the shape	already fixed in staging (`useMarketData.ts:208`); the surface that called it is gone
2	`409 CONFLICT: A turn is already running`	the client posted a second turn and rendered the refusal as a generic failure with a retry	the composer refuses the duplicate before it is sent; a server 409 becomes an active-run state that joins the live run
3	50–120 s with no useful activity	`sendTurn`’s contract was mis-stated — it resolves when the TURN COMPLETES, not when it is accepted — and nothing filled the wait	the stream is subscribed from hydration; each lookup renders as a step as it starts
4	technical wall: `7 filtros disponibles en este modo`, `source_slugs`, `observed_kind`	`ResearchPage` rendered `/capabilities` as a disclosure for the customer	removed; `/capabilities` is read by the agent through its typed tools
5	`13/13`, `listo`, internal names, UUIDs	`TOOL_CALL_START.summary` was the raw operation name; the trace header printed a bare fraction	`research-progress.ts` composes every visible string server-side; the header says “13 de 13 pasos”
6	tried `discover_market_capabilities`, fell back to Topics/Collections	the turn classified as `general`, which carries the full ~20-tool belt including the legacy market catalogue	the ladder now classifies the errand as `market_opportunity`, which dispatches to the runner and builds no belt
7	never called the new semantic operations	`mode: "research"` was the only door to `MarketResearchRunner`, and no client ever sent it	route-based dispatch in `chat.service.ts`

Incidents 6 and 7 are one cause. The router existed and was already deterministic; it was missing the shapes people use — an offer framing (“vendo”, “fabricamos”), the demand side (“quién compra”, “quiénes han ganado”), and a supply-side noun list that omitted fabricantes and proveedores. contact_request also sat above the market ladder, so the bare word contacto in a supplier search sent the errand to the CRM.

Routing, before and after

Question	Before	After
”Vendo estructuras metálicas y quiero saber quién compra en Nuevo León”	`general` → full tool belt	`market_opportunity` → `MarketResearchRunner`
”Busca fabricantes de bombas en Jalisco con contacto publicado.”	`contact_request` → CRM	`market_opportunity` → `MarketResearchRunner`
”¿Qué licitaciones accionables hay para tratamiento de agua?”	`market_opportunity` → legacy Opportunity Flow	`market_opportunity` → `MarketResearchRunner`
”¿Quiénes han ganado contratos de alumbrado público y por cuánto?”	`general` → full tool belt	`market_opportunity` → `MarketResearchRunner`
”qué hay en mi pipeline”	`pipeline`	`pipeline` (unchanged)
“qué sabemos de la arquitectura”	`topic_context`	`topic_context` (unchanged)
“¿a quién le hablo?”	`contact_request`	`contact_request` (unchanged)
“Avísame cada lunes si aparecen licitaciones”	`create_monitor`	`create_monitor` (unchanged)

Deletion ledger

Every deletion below was proven unreachable by reference before it was made.

Deleted

Path	Lines	Why it was safe
`redesign/ChatThreadView.tsx` (+2 specs)	~1,356	no route in `App.tsx`; every remaining mention in the tree is a comment
`redesign/chat/` (13 files)	~1,536	`OperateChatSurface` imported only by `ChatThreadView`
`redesign/experience/` (7 files)	~1,200	the projection layer for the Operate surface; after its deletion the only importers were its own specs
`redesign/ResearchArtifactPanel.tsx` (+spec)	~182	single consumer was `ChatThreadView`
`redesign/AgenticWorkbench.tsx`	~90	zero references anywhere in the tree
`redesign/spike/` (2 files)	~301	self-referencing spike
`research/ResearchPage` mode nav, search form, capability disclosure	~180	replaced by the conversation; see incidents 4 and 6
`research.css` rules for the above	~86	swept after the markup that used them

Total: 7,864 lines removed, 30 added, in the consolidation commit alone.

Kept, because the experience rests on them

redesign/chatThreads.ts (thread creation, message POST, regenerate, checkpoint resolution), redesign/agentic-events.ts (A1 parser and reducer), the SSE subscription in api.ts, the Work Session snapshot route, ResearchReport and its artifacts, stop and retry.

Kept, with the deletion condition named

Piece	Still reachable?	Deletion condition
the legacy Opportunity Flow in `chat.service.ts`	yes — only in a deployment where the market-research dependencies are absent. Its three tools are no longer on any belt (see Capability isolation) and its instruction is out of the general manual	when the market-research dependencies stop being `@Optional()`, i.e. when no supported deployment can lack them
`useSupplierSearch`, `useAwardSearch`, `useOpportunitySearch`, `useCapabilities`, `filtersFor`, `RecordsTable`, `supplierColumns`, `awardColumns`, `groupAwards`	no — nothing in the shipped tree reaches them	when `ResearchReport` publishes `suppliers[]` and `historicalRecords[]` (the `suppliers-in-report` and `history-rows-in-report` gaps in `phases.ts`), these render them; if those gaps are closed differently, delete these instead

The second row is the honest reading of “quiero que todo el UI se utilice”: the nine-route client and its three domain hooks are built and correct, and exactly one of the three domains — opportunities — currently has rows in the report to render. The other two are one backend field away, and are named here rather than deleted quietly or left to look used.

Capability isolation

The classifier can be wrong. Safety does not rest on it being right. A commercial capability used to be reachable from the general runtime through two doors, both open by default:

the belt. chat-tools.ts registered discover_market_capabilities, query_market_signals and get_signal_evidence behind includeMarketTools, computed as “not a market route” — so the tools built for commercial work were withheld from commercial turns and handed to every other one.
the manual. chat.skill.ts opened its market section by instructing the model to call the first two in sequence.

Both are closed. marketIntelligence, runMarketIntelligence and includeMarketTools are gone from ChatToolDeps; there is no dependency to inject and no flag to get wrong. The manual carries a boundary instead, naming Collections, Topics and Projects explicitly as things that must not be substituted for market data. apps/api/src/chat/commercial-isolation.spec.ts pins all eight invariants:

#	Invariant
1–4	supplier / opportunity / history / mixed intents all route to `market_opportunity`
5	an ambiguous commercial question degrades to `general` — and `general` has no market tools, so the degradation is safe
6	a non-commercial question keeps Topics, Collections, Projects, Docs and connections
7	the general belt cannot invoke `discover_market_capabilities`; calling it is an unknown tool
8	the runner reaches no Radar, Collections, Topics or CRM — grepped, because a mock proves one scripted run and this must hold for all of them

Plus two the eight imply: the SEMANTIC operations are not on the general belt either (replacing one leak with a newer one is the same defect), and a market-research turn builds no chat tool belt at all.

The activity trail

The runner has always reported every stage through step_started / step_finished. The turn pipeline dropped all of them and forwarded only tool calls — which is why the wait began after the search was already prepared. The channel is the A1 contract’s own ACTIVITY_DELTA / ActivityEntry, already frozen, already reduced into snapshot.activity, already served by the Work Session endpoint. No second stream was built, and the trail survives a reload because it is the same durable projection everything else here reads.

Moment	Source	What a person reads
turn accepted	emitted before any work	”Recibí tu solicitud”
route selected	`step_started: route_research`	”Entendiendo tu solicitud”
plan ready	`step_started: load_capabilities`	”Preparando la búsqueda comercial”
tool started	`tool_call`	”Buscando compradores en adjudicaciones”
tool completed	`tool_result`	same row, updated: “8 adjudicaciones encontradas”
synthesis started	`step_started: finalize`	”Preparando la respuesta”
answer ready	terminal	”Respuesta lista”
turn failed	terminal	”No pude terminar la investigación”
turn cancelled	terminal	”Investigación detenida”

investigate is deliberately not narrated: it is the tool loop and its children already name themselves. Allowed in an entry: a human sentence (summary), a domain (name), a status, ok, and a safe count or sanitized reason (preview). Never: an operation name, a semantic code, a UUID as a label, SQL, a prompt, reasoning, a capability catalogue, warehouse rows, or a provider — providerId is always null. Ids are stable KEYS (phase:route_research, tool:q_0), never labels. The prefix also keeps them non-numeric, so object key order stays insertion order and the trail rehydrates in the order it happened.

Known residual exposure

A commercial question that slips the ladder still classifies as general. That degradation is now bounded rather than dangerous: the general belt holds no market capability, so the worst case is an honest “I can’t answer that from here” instead of a tour of Topics and Collections. The legacy Opportunity Flow still runs where the market-research dependencies are absent. Its deletion condition is unchanged: those dependencies ceasing to be @Optional().

What is NOT proven

No live end-to-end run against staging, and no screenshots. The container has no API on localhost:3000, no Docker, no database URL, no model credential and no Clerk session; Playwright is not resolvable from apps/dashboard, and /chat sits behind an authenticated workspace route, so a browser would reach a sign-in screen rather than any of the states above. Every claim here rests on hermetic tests and on reading the code.

The live E2E checklist, for whoever has staging

# 1. Point the harness at staging and confirm it answers.
export STAGING_BASE_URL=https://api-staging.driftless.icu
export STAGING_API_KEY=…            # do not paste this into a PR
bash scripts/harness/smoke.sh staging

# 2. Confirm the semantic layer is deployed — the gate the cutover depends on.
#    Absent, commercial turns fall to the legacy Opportunity Flow by design.
bash scripts/harness/research-staging.sh

Then, in /w/<slug>/chat, with the network tab open on the SSE stream:

#	Do this	Expect
1	Ask `Vendo estructuras metálicas y quiero saber quién compra en Nuevo León` in a NEW thread	”Recibí tu solicitud” within a second, then the phases in order
2	Watch the stream frames	`ACTIVITY_DELTA` entries only; no `discover_market_capabilities`, no `query_market_signals`
3	Reload mid-run	the trail and the running state come back; the stream resubscribes from `lastSeq`
4	Send a second question while the first runs	no POST is made; “Hay una investigación en curso” plus a Stop
5	Press Stop	”Investigación detenida”; no error banner
6	Let one finish	the answer lands in the SAME thread, with the report, its table and a row that opens its record
7	Ask `qué hay en mi pipeline`	answers from Collections; unchanged
8	Ask `Necesito vender más`	an honest short answer — NOT a tour of Topics/Collections

Repeat 1 for each domain: Busca fabricantes de bombas en Jalisco con contacto publicado. · ¿Qué licitaciones accionables hay para tratamiento de agua? · ¿Quiénes han ganado contratos de alumbrado público y por cuánto, separados por moneda y ámbito?

Get started

Connect

Concepts

Work & collaborate

Guides

Agents

MCP reference

CLI reference

API reference

Platform

Chat market intelligence cutover

The commercial chat cutover

The request path, end to end

Root cause of each observed incident

Routing, before and after

Deletion ledger

Deleted

Kept, because the experience rests on them

Kept, with the deletion condition named

Capability isolation

The activity trail

Known residual exposure

What is NOT proven

The live E2E checklist, for whoever has staging

​The commercial chat cutover

​The request path, end to end

​Root cause of each observed incident

​Routing, before and after

​Deletion ledger

​Deleted

​Kept, because the experience rests on them

​Kept, with the deletion condition named

​Capability isolation

​The activity trail

​Known residual exposure

​What is NOT proven

​The live E2E checklist, for whoever has staging

The commercial chat cutover

The request path, end to end

Root cause of each observed incident

Routing, before and after

Deletion ledger

Deleted

Kept, because the experience rests on them

Kept, with the deletion condition named

Capability isolation

The activity trail

Known residual exposure

What is NOT proven

The live E2E checklist, for whoever has staging