What this is

The synchronous, in-the-loop intake path

Path 2 is the fast path. When Mike is already at the desk and the question is "let's re-quote X" or "log this new RFP", going through the email mailbox + HITL gate cycle is overhead. Path 2 skips it. Mike opens the admin chat at chat.ai-globalfoodsolutions.co (or any chat surface) and either types the customer + intent or drag-drops a bid PDF.

Because Mike is the operator, there is no separate proposed_actions approval gate — every action Mike takes IS the approval. ADR-031's HITL invariant is satisfied by Mike's presence; the gate moves from before the review runs to any NS-bound write inside the skill (which today is none on this path — NS writes are deferred to Path 3).

The dispatch is identical to Path 1 from step 4 onward: customer resolved via name_synonyms.json, SY-version picked from bid_quote_versions, customer + SY-version becomes the trace thread, one of 11 skills runs, 8 D1 bid_* tables UPSERT, .docx regenerates, STATUS.json refreshes, the dashboard at gfs-pricing.pages.dev re-renders. Events fire (bid.chat_review_completed, bid.quote_versioned) and the new SY-version sits in the Bid Center until July 1.

Diagram: ns-bid-center-path-2-direct-chat.html. Sibling paths: Path 1 email · Path 3 Jul 1 rollover. Master: master wiki.

When to use it

Trigger conditions

Mike wants to re-quote a known customer right now (no email round trip needed).
A PDF or pricing sheet is sitting on the desk — drag-drop into chat is faster than emailing it to bids@.
Mike wants to run COMPARE_CUSTOMER_QUOTES on two versions for an existing customer.
Onboarding a brand-new customer via CREATE_NEW_CUSTOMER.
Logging an external bid that arrived via a non-email channel (portal, phone, conversation notes).
Quick admin tasks: SYSTEM_HEALTH_CHECK, BID_PIPELINE review, ARCHIVE_CUSTOMER.

No NS write on this path either

Like Path 1, Path 2 only writes to D1 + folder + dashboard. NetSuite writes are deferred to Path 3 (Jul 1 rollover). The Bid Center is the rolling-state pillar; NS named price lists are the annual snapshot.

★ Same trace thread as Path 1

customer_id + sy_version threads through chat_session, bid_quote_versions, bid_price_snapshots, and STATUS.json. Sample: ACC Distributors / SY2627-v2.5. From the audit trail, a Path 1 row and a Path 2 row are indistinguishable except for the origin field on the chat_session.

Worked example

Mike re-quotes ACC in 4 minutes

Scenario

April 14, 2026, 11:02 AM. Mike just got off a call with ACC Distributors' buyer. They want the cheddar items re-quoted today. Mike has the new pricing PDF from the call notes.

Mike opens chat at chat.ai-globalfoodsolutions.co and drag-drops the 3-page PDF. Types: "re-quote ACC Distributors — cheddar refresh, SY2627."

The Worker calls document_converter on the PDF bytes; Markdown is in context within ~3 seconds. name_synonyms.json resolves "ACC Distributors" to the canonical customer at confidence 1.0 (exact match). bid_quote_versions picks SY2627-v2.4 as the latest. The trace thread ACC Distributors / SY2627-v2.5 is established (auto-incremented).

Chat preloads: bid Markdown, customer.json, bid_quote_versions latest row. The LLM suggests ADD_UPDATE_CUSTOMER_PRICING; Mike confirms with a single "go." The skill runs, snapshots 47 line items, applies the HPS MAP GPO allowance, UPSERTs the 8 bid_* tables, regenerates quotes/SY2627-v2.5.docx, refreshes STATUS.json, and re-renders the dashboard.

Total elapsed: ~4 minutes from drag-drop to "done, v2.5 live on dashboard." No proposed_actions row was created on this path — the chat session row itself is the audit record. Events fire: bid.chat_review_completed, bid.quote_versioned. The new pricing sits in the Bid Center until Path 3 picks it up on July 1.

Step-by-step what happens

Chat opens → intent typed → review runs

01
Mike opens chat
Admin dashboard chat OR direct chat URL. Session row inserted in chat_session with origin='direct_chat'.
02
Mike types intent
Free-form: "re-quote ACC Distributors", "compare ACC v2.4 vs v2.5", "log new external bid from buyer X", …
03
Attach / paste bid content
Drag-drop PDF OR paste bid text. PDF bytes flow through document_converter → Markdown.
04
Worker resolves customer
name_synonyms.json → canonical customer_id. Ambiguous? Worker asks in chat: "did you mean X or Y?"
05
Pick SY-version
Latest is_latest=1 from bid_quote_versions; else mint a new currentSchoolYear() + '-v0.1'.
06
Preload context
Bid Markdown + customer.json + latest bid_quote_versions row attached to the chat session.
07
Decide skill
Mike picks one of 11, or the LLM suggests and Mike confirms.
08
Run 11-skill standardized review
Same dispatcher used by Path 1 and Path 3. router.dispatch(skill, payload).
09
UPSERT D1 bid_*
Migration 136 (Agent M owns). Until landed, writes go to folder + STATUS.json only.
10
Regen .docx
Per-customer per-SY-version .docx.
11
Refresh STATUS.json + dashboard
Freshness metadata updated; gfs-pricing.pages.dev re-renders on next request.
12
Events fired
bid.chat_review_completed, bid.quote_versioned.

Outcomes

What's different after Path 2 runs

HITL gate

Mike IS the gate

D1 tables

8 bid_*

UPSERT inline

NS write

none

deferred to Path 3

Latency

~4 min

drag-drop to done

One new bid_quote_versions row (is_latest=1) per approved chat run; prior version demoted.
chat_session row is the audit record (no separate proposed_actions insert).
Per-customer .docx quote regenerated; STATUS.json freshness updated.
Dashboard at gfs-pricing.pages.dev reflects the new SY-version.
Same downstream outputs as Path 1 — the artifacts are indistinguishable except for the origin field.

Failure modes

What can go wrong

Ambiguous customer match in chat

If name_synonyms is below 0.85, the Worker asks Mike inline: "did you mean X or Y?" Mike picks; flow continues.

PDF drag-drop too large

Hard cap at ~25MB for chat upload. Workaround: email the file to bids@ (Path 1) or split the PDF.

document_converter throws on the PDF

Chat surfaces the error; Mike can paste the bid text inline OR open the source in another tool and paste excerpts.

Skill picked wrong

Mike can abort the chat session and start fresh with the right skill named explicitly. chat_session row stays for audit.

Migration 136 not yet landed

D1 bid_* writes no-op gracefully; folder + STATUS.json remain authoritative.

Adjacent flows + diagrams

For developers

Code paths + invariants

Concern	Where
Chat entrypoint	/api/chat (src/index.ts)
Document converter	src/document_converter.ts
Name synonyms	data/name_synonyms.json
Chat session row	chat_session (origin='direct_chat')
HITL invariant	Mike-in-the-loop (no proposed_actions row)
D1 tables	bid_customers, bid_quote_versions, bid_price_snapshots, …
Dispatcher	router.dispatch(suggested_skill, payload)
Events fired	bid.chat_review_completed, bid.quote_versioned

// Path 2 - chat-initiated, Mike-in-the-loop async function runPath2(env: Env, session: ChatSession, intent: ChatIntent) { const { customer_id } = await resolveCustomer(env, intent.raw_text); const sy_version = await pickOrMintSyVersion(env, customer_id); // trace thread established: customer_id + sy_version await session.preload([intent.bid_markdown, await loadCustomerJson(customer_id), await loadLatestVersion(env, customer_id)]); const skill = await chooseSkill(session, intent); await dispatcher.run(skill, { session, customer_id, sy_version }); await emit(env, 'bid.chat_review_completed', { customer_id, sy_version }); }

Changelog

Dated trail

Date	Round	Change	Touched by
`2026-05-26`	`R597`	Wiki ships alongside Path 2 diagram. Documents the Mike-in-the-loop invariant (no separate HITL gate on this path). Same customer + SY-version trace thread as Path 1.	Mike + Claude (Agent L)

Schema · data contract

The machine-readable spec

workflow_type · bid_center_direct_chat_path · risk_level 2. Parent: bid_center_lifecycle (master).

Customer + SY-version threading (the trace thread)

Identical to Path 1 except the session row replaces the proposed_actions row at the top of the trail.

Record / table	Field carrying customer + SY-version	Sample value
`chat_session`	`context.customer_id` + `context.sy_version`	`"ACC Distributors / SY2627-v2.5"`
`chat_message`	system message with preloaded context	references above pair
`bid_customers`	`customer_id`	`"ACC Distributors"`
`bid_quote_versions`	`customer_id + sy_version`	`"ACC Distributors / SY2627-v2.5"`
`bid_price_snapshots`	`customer_id + sy_version + sku`	`"ACC / SY2627-v2.5 / SKU-419"`
`STATUS.json`	`customers[customer_id].latest_sy_version`	`"SY2627-v2.5"`

D1 tables written (Path 2)

Table	Operation	Purpose
`chat_session`	INSERT	session origin + context
`chat_message`	INSERT (per turn)	conversation log
`bid_customers`	UPSERT	customer normalize
`bid_quote_versions`	INSERT + demote prior	new SY-version
`bid_price_snapshots`	INSERT-many	line-level snapshot
`bid_customer_prices`	UPSERT	workshop layer
`events`	INSERT	append-only audit

Endpoints

Method	Path	Purpose
`POST`	`/api/chat`	main chat entrypoint (carries intent + attachments)
`POST`	`/api/chat/upload`	PDF / XLSX upload for chat
`GET`	`/quote/<slug>/<version>`	customer-facing rendered quote

Events fired

event_type	When
`bid.chat_session_opened`	chat_session row inserted
`bid.chat_review_completed`	skill dispatcher finishes
`bid.quote_versioned`	new `bid_quote_versions` row written

Runbook · when it breaks

It broke — what now

Scenario · Chat upload of PDF fails silently

Drag-drop seems to work but the bid Markdown never lands in the session.

Check upload endpoint: network tab on /api/chat/upload — 200 OK?
Check file size: > 25MB? Split or use Path 1 email instead.
Check document_converter: npx wrangler tail filter on document_converter
Workaround: paste the bid text inline in chat

Scenario · Wrong skill ran

Dispatcher picked LOG_EXTERNAL_BID but it should have been ADD_UPDATE_CUSTOMER_PRICING.

Abort cleanly: chat command /abort-skill rolls back the in-flight UPSERTs
Roll back manually: UPDATE bid_quote_versions SET is_latest=0 WHERE customer_id=? AND sy_version=? on the just-written row; restore prior is_latest=1
Restart: open fresh chat, name the skill explicitly, re-run

Scenario · Customer match was wrong (typo, alias)

Skill ran against the wrong customer; need to undo + redo.

Identify the bad rows: SELECT * FROM bid_quote_versions WHERE customer_id='<wrong>' ORDER BY created_at DESC LIMIT 1
Demote / delete: set is_latest=0 on the bad row; restore prior
Add alias: update name_synonyms.json to prevent future mismatches
Re-run: open fresh chat, explicit customer this time

Logs to check

chat_session · origin='direct_chat'
chat_message · turn-by-turn log
events · bid.chat_* trail
workflow_run_log · per-skill dispatcher audit
npx wrangler tail · live Worker logs

Kill switches

kill:bid_chat_dispatch · pauses Path 2 skill dispatch (chat sessions still open)
kill:bid_writes · stops all bid_* UPSERTs

Backlog · open questions

What's not done · what's uncertain

STUB
Migration 136 (8 bid_* tables)
Agent M owns. Until landed, Path 2 writes go to folder + STATUS.json only.
STUB
Skill abort + rollback semantics
Today /abort-skill is best-effort. Need clean transactional rollback for any partial UPSERT batch.
OPEN
Voice intake
Chat could accept dictation. POST /api/chat/voice → Whisper → intent. Useful when Mike is on the phone with the customer.
OPEN
Multi-customer batch in one session
"Re-quote ACC + BCD + CDE all on SY2627 with new cheddar pricing" — today three separate sessions. Could batch in one.
DECISION
Auto-skill confidence threshold
Today the LLM suggests + Mike confirms. With baked confidence (> 0.97), should we auto-run for trivial intents like "show ACC's current quote"?

Path 2 — direct chat

The synchronous, in-the-loop intake path

Trigger conditions

Mike re-quotes ACC in 4 minutes

Chat opens → intent typed → review runs

Mike opens chat

Mike types intent

Attach / paste bid content

Worker resolves customer

Pick SY-version

Preload context

Decide skill

Run 11-skill standardized review

UPSERT D1 `bid_*`

Regen `.docx`

Refresh STATUS.json + dashboard

Events fired

What's different after Path 2 runs

What can go wrong

Adjacent flows + diagrams

Code paths + invariants

Dated trail

The machine-readable spec

Customer + SY-version threading (the trace thread)

D1 tables written (Path 2)

Endpoints

Events fired

It broke — what now

Scenario · Chat upload of PDF fails silently

Scenario · Wrong skill ran

Scenario · Customer match was wrong (typo, alias)

Logs to check

Kill switches

What's not done · what's uncertain

The synchronous, in-the-loop intake path

Trigger conditions

Mike re-quotes ACC in 4 minutes

Chat opens → intent typed → review runs

Mike opens chat

Mike types intent

Attach / paste bid content

Worker resolves customer

Pick SY-version

Preload context

Decide skill

Run 11-skill standardized review

UPSERT D1 bid_*

Regen .docx

Refresh STATUS.json + dashboard

Events fired

What's different after Path 2 runs

What can go wrong

Adjacent flows + diagrams

Code paths + invariants

Dated trail

The machine-readable spec

Customer + SY-version threading (the trace thread)

D1 tables written (Path 2)

Endpoints

Events fired

It broke — what now

Scenario · Chat upload of PDF fails silently

Scenario · Wrong skill ran

Scenario · Customer match was wrong (typo, alias)

Logs to check

Kill switches

What's not done · what's uncertain

UPSERT D1 `bid_*`

Regen `.docx`