Support triage kit

AF-CS-0110 files, 30,250 BMIT licence

Sorts new tickets by intent and urgency. Drafts replies from your help center. Hands refunds over your limit to a person.

Install the kitIncluded in Builder, $19 a month

To install, sign in, or open a free account and add Builder.

Install the kitBuilder, $19 a month

Every part, in assembly order

Bars show how full each size budget is. MODELS.yaml is at revision B from October 22: its Claude Code block now sets effort for each step.

Order	Part	What it does	Size	Test	Revision
1	`README.md`AF-CS-01-RM	Setup, what success looks like and the recovery steps	3,170 B	Setup replays in a fresh project	A, Oct 18
2	`DEPENDENCIES.yaml`AF-CS-01-DP	MCP servers and libraries at exact versions	1,130 B	Every pin resolves exactly	A, Oct 18
3	`MODELS.yaml`AF-CS-01-MD	Model, effort and output limit for each harness	920 B	One block for each harness	B, Oct 22
4	`AGENTS.md`AF-CS-01-AG	Role, boundaries and stop conditions	3,180 B of 4,096 B	Loads in all four harnesses	A, Oct 18
5	`RULES.md`AF-CS-01-RU	Numbered rules, checked on every turn	2,240 B of 3,072 B	Checked on every turn	A, Oct 18
6	`SKILL.md`AF-CS-01-SK	Checked methods, formulas and snippets	5,620 B of 6,144 B	Rubric sorts the 24 recorded tickets	A, Oct 18
7	`TOOLS.json`AF-CS-01-TL	Typed tool schemas with path limits	3,410 B of 4,096 B	Every schema validates on load	A, Oct 18
8	`WORKFLOW.yaml`AF-CS-01-WF	Ordered steps, a call budget and recovery	1,520 B of 2,048 B	Runs inside 12 calls	A, Oct 18
9	`contract_card.json`AF-CS-01-CC	Typed ports, slot ID, refusal codes and checksums	1,690 B of 2,048 B	Every SHA-256 checksum matches	A, Oct 18
10	`verifier.py`AF-CS-01-VF	Independent checks, run apart from the worker	7,370 B	Replays 24 recorded tickets	A, Oct 18

What it works with.

verifier.py replays 24 recorded tickets on every harness and model: 12 routine questions, 6 refunds over your limit and 6 look-alikes that mention a refund and need a normal reply.

Harness	Model	Checks passed	Files it reads
Claude Code	`claude-sonnet-5-5`	24 of 24	CLAUDE.md, .mcp.json, .claude/settings.json
Claude Code	`claude-haiku-5-5`	23 of 24	CLAUDE.md, .mcp.json, .claude/settings.json
Codex	GPT-6.1 Sol	24 of 24	AGENTS.md, ~/.codex/config.toml
OpenCode	`qwen3-coder:30b`, on your own machine	22 of 24	AGENTS.md, opencode.json
Pi	`gemma4:31b`, on Ollama Cloud	22 of 24	AGENTS.md, .pi/extensions/support-triage.ts

Every row hands all six refunds to a person. The misses on the smaller models are routine drafts a person edits.

Calls, dependencies and model settings.

Three files make the kit run the same way in every harness. Secrets appear by name; your harness supplies the values.

WORKFLOW.yaml

# WORKFLOW.yaml: steps, call budget, recovery
call_budget: 12
steps:
  - read_ticket     # helpdesk.read_ticket
  - classify        # intent and urgency
  - lookup_order    # orders.lookup
  - draft_reply     # helpdesk.draft_reply
  - hand_off        # refunds over the limit
recovery: retry_once_then_hand_off

DEPENDENCIES.yaml

# DEPENDENCIES.yaml: exact versions
mcp_servers:
  helpdesk:
    url: https://helpdesk.example.com/mcp
    scopes: [tickets.read, drafts.write]
  orders:
    url: https://orders.example.com/mcp
    scopes: [orders.read]
libraries:
  python: "3.12"
  pydantic: "2.11.7"
  httpx: "0.28.1"
secrets: [HELPDESK_TOKEN]   # by name

MODELS.yaml

# MODELS.yaml, revision B: one block per harness
claude-code:
  model: claude-sonnet-5-5
  effort: high
  max_output_tokens: 8000
codex:
  model: default   # GPT-6.1 Sol, the Codex default
opencode:
  model: ollama-cloud/muse-spark-open
  num_ctx: 32768
pi:
  model: gemma4:31b   # Ollama Cloud
  num_ctx: 32768

Set it up in your harness.

The same ten files, mounted four ways. Your models, permissions and project stay yours. Or hand the job to your agent with this prompt.

Installed kit files carry the MIT licence: use, change and share them in your own projects. Builder installs kits over MCP, with setup for each harness and every new revision.

Agent prompt: check, then install

Read https://getagentfoundry.si/llms.txt. Get kit AF-CS-01, the support triage kit. Check each file against the checksums in contract_card.json, run verifier.py on its recorded cases and show me the result. Set it up in this harness after I say yes.

Claude Code

CLAUDE.md imports AGENTS.md and RULES.md, and a hook runs verifier.py after each edit.

# CLAUDE.md
@AGENTS.md @RULES.md
# .claude/settings.json
"model": "claude-sonnet-5-5"
# PostToolUse hook, Edit|Write
python3 verifier.py

Codex

Codex reads AGENTS.md at the project root, and a WORKFLOW.yaml gate runs verifier.py.

# ~/.codex/config.toml
[mcp_servers.agentfoundry]
url = "https://getagentfoundry.si/mcp"
# project root
AGENTS.md  WORKFLOW.yaml

OpenCode

OpenCode reads AGENTS.md. opencode.json adds the server and the model from MODELS.yaml.

// opencode.json
"model": "ollama-cloud/muse-spark-open",
"mcp": { "agentfoundry": {
  "type": "remote",
  "url": "https://getagentfoundry.si/mcp" } }

Pi

Pi reads AGENTS.md. The kit’s tools load through one extension file, and the model runs on Ollama Cloud.

# extension
.pi/extensions/support-triage.ts
# model, on Ollama Cloud
gemma4:31b