Full graph at rate card: 39 code/integration x $0.025 = $0.975; 8 std x $0.10 = $0.80; 2 mid x $0.20 = $0.40; parse $0.20; total $2.38 ceiling. Costliest single lane: 39 nodes = 31 code ($0.775) + 5 std ($0.50) + 2 mid ($0.40) + parse ($0.20) = $1.88. Measured figure (the catalog's CX anchor, logged Aug 2026): 13 credits = $1.30 for a full run, ~3 credits ($0.30) on early exit, run on a mid-tier model (03-projects/27-cx/01-planning/overview.md '校准' section). Reconciliation: measured $1.30 sits below my $1.88 max-lane arithmetic because a real run does not touch every parallel side-write; against the $0.88 standard-tier menu ('aftersales + intake') the tier assumption that closes the gap is MID — the menu's own top-tier figure $2.68 implies $0.60 of standard-tier judgment ($2.68-$0.88 = 3x$0.60), so mid doubles it to a predicted $1.48 vs $1.30 measured (within ~12%, a run skipping 1-2 judgment steps). Caveat: in the 2026-08-28 snapshot only N4 Classification and N12 ResponseAssembly sit on GEMINI_3_1_PRO; the other 8 prompt nodes are Flash — so the measured mid-tier billing reflects run-time tier, not today's per-node preferredModel. Code 39 = 24 tool/code nodes (10 CodeExecutor, 3 Airtable formula lookups, 2 RecordCreate, 2 DAB QueryData, 3 Outlook tags, 2 MessageSend, 1 MessageReply, 1 S3) + 14 rule_based conditions + 1 inert no-tool execution node (empty objective).

Measured: 13 credits = $1.30 for a full logged run (~3 credits on early exit), measured August 2026 on a mid-tier model — documented as the 'CX aftersales full run' anchor; standard-tier menu equivalent is $0.88, so the gap is model tier (credit-rail measured)

Models seen in the graph: preferredModel: GEMINI_3_FLASH x33, GEMINI_3_1_PRO x2 (N4 Classification and N12 ResponseAssembly — N12's fallback chain is BEDROCK_CLAUDE_OPUS_4_7, GPT5_4); all 14 conditions rule_based (llmModel GEMINI_3_FLASH x6, GPT4_1_MINI x8)

Graph read: 04-agents/w-nsche/tst-wunsche-aftersales-agent/agent.json (50 nodes). Note: 04-agents/README.md marks w-nsche/ as superseded by wunsche/ (slug-normalization fix), but the TST production graph snapshot exists ONLY under w-nsche/ — the catalog's source path is correct.

• Classifier-router topology: N4 classifies into a canonical category with confidence on Gemini 3.1 Pro (the largest prompt in the CX fragment library), per-category route nodes N6-N10 (repair/replacement by warranty status and claim value, spare parts, user manual, non-service, SC delegation), then exactly two convergence nodes — N12 assembles every automated reply, N13 packages every manual-review handoff.
• 10 CodeExecutor nodes hold the deterministic spine: Airtable formula builders, case-payload builders, DAB query builder + response handlers, subject tokenizer, and — critically — reply templates and the signature are held in code, so styling cannot drift with the model (27-cx overview calls this the most deterministic graph on the CX line; UNiDAYS has 0 code nodes).
• All 14 conditions are rule_based, routing on flags set by upstream prompt nodes; intake N1 screens escalation keywords and applies a hard automated-notification veto (false positive = silently abandoning a customer) before anything else runs.
• Lookups are three distinct Airtable tables (case history, service-center contacts, e-mail-target contacts) plus the DAB product database via custom QueryData API with code-built queries — the reply is grounded in records, and every outcome writes a case record back to Airtable before send/tag.

• The same graph bills 3x depending on model tier: $0.88 standard menu vs $1.30 measured mid vs $2.68 top — quote tier explicitly in any offer; also, the API totalCost field is credits/100, NOT dollars (27-cx/01-planning/overview.md, flagged as a known trap).
• Full run vs early exit is a 4x cost spread (13 vs ~3 credits) — per-task averages are meaningless without the exit-mix.
• This is the docs' clearest named production CX case, but volume is only 10 tasks/30d (20 in the prior 14-day window) — cite it as production wiring and pattern proof, not scale proof.
• Workspace hygiene: read the graph from w-nsche/ even though the folder is marked superseded — the wunsche/ replacement folder never received the TST agent, and the DEP_/copy variants in the probe all show 0 tasks.

• cxWünsche 售后(生产) · 提示词03-projects/27-cx/02-resources/run-steps/01-intake-接件与解析.md#1
• cxWünsche 售后(生产) · 代码03-projects/27-cx/02-resources/run-steps/01-intake-接件与解析.md#2
• cxWünsche 售后(生产) · 代码03-projects/27-cx/02-resources/run-steps/01-intake-接件与解析.md#3
• cxWünsche 售后(生产) · 提示词03-projects/27-cx/02-resources/run-steps/03-classify-定类与抽取.md#1
• cxWünsche 售后(生产) · 提示词03-projects/27-cx/02-resources/run-steps/03-classify-定类与抽取.md#2
• cxWünsche 售后(生产) · 代码03-projects/27-cx/02-resources/run-steps/05-lookup-查库与取证.md#1
• cxWünsche 售后(生产) · 代码03-projects/27-cx/02-resources/run-steps/05-lookup-查库与取证.md#2
• cxWünsche 售后(生产) · 代码03-projects/27-cx/02-resources/run-steps/05-lookup-查库与取证.md#3
• cxWünsche 售后(生产) · 代码03-projects/27-cx/02-resources/run-steps/05-lookup-查库与取证.md#4
• cxWünsche 售后(生产) · 代码03-projects/27-cx/02-resources/run-steps/05-lookup-查库与取证.md#5
• cxWünsche 售后(生产) · 提示词03-projects/27-cx/02-resources/run-steps/06-route-定路径与判定.md#1
• cxWünsche 售后(生产) · 提示词03-projects/27-cx/02-resources/run-steps/06-route-定路径与判定.md#2
• cxWünsche 售后(生产) · 提示词03-projects/27-cx/02-resources/run-steps/06-route-定路径与判定.md#3
• cxWünsche 售后(生产) · 提示词03-projects/27-cx/02-resources/run-steps/07-compose-写回复.md#1
• cxWünsche 售后(生产) · 提示词03-projects/27-cx/02-resources/run-steps/07-compose-写回复.md#2
• cxWünsche 售后(生产) · 代码03-projects/27-cx/02-resources/run-steps/07-compose-写回复.md#3
• cxWünsche 售后(生产) · 代码03-projects/27-cx/02-resources/run-steps/08-commit-写回与打标.md#1
• cxWünsche 售后(生产) · 代码03-projects/27-cx/02-resources/run-steps/08-commit-写回与打标.md#2
• cxWünsche 售后(生产) · 提示词03-projects/27-cx/02-resources/run-steps/09-handover-交人打包.md#1
• cxWünsche 售后(生产) · 提示词03-projects/27-cx/02-resources/run-steps/09-handover-交人打包.md#2