agentplane 0.45.0

Durable, replayable agent runtime — the journal is the plan of record
Documentation 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41

# A declarative tool loop — no Rust, and the tools come from an MCP server the
# operator names on the command line:
#
#   agentplane run examples/tool-calling.yaml \
#     --input '{"ticket": "T-1"}' \
#     --mcp "tickets=python3 examples/mcp-server.py"
#
# The manifest grants `tool://tickets/read`. **Which transport reaches
# `tickets` is not in here**, and that is deliberate: an agent's declaration —
# and therefore its digest — must not change when it moves between a laptop and
# a cluster. Grants are reviewed; wiring is deployed.
apiVersion: agentplane.hupe1980.github.io/v1alpha1
kind: Agent
metadata: { name: ticket-desk, version: "1.0.0" }
spec:
  execution: { kind: tool-calling, max_turns: 3 }
  identity:
    role: "Answer a support question using the ticket tool"
    constraints: "One sentence. Cite the ticket id."
  capabilities: { provides: [support.answer] }
  models:
    privileged: { provider: fake, model: m-1 }
  tools:
    # `mutates: false` is a claim the *operator's* catalogue is held to as well:
    # a server calling its own tool read-only cannot make it so, and a catalogue
    # laxer than this grant refuses the build.
    - ref: "tool://tickets/read"
      mutates: false
      # What the model asks for is untrusted and at least `internal`, so a tool
      # that may be shown only `public` data refuses every call a model makes.
      max_sensitivity: internal
      description: "Read a ticket by id"
      arguments:
        type: object
        additionalProperties: false
        required: [id]
        properties: { id: { type: string } }
  # A tool's answer is `internal` too, and it goes back to the model on the
  # next turn: without this, the model sink's `public` default refuses it.
  security: { max_sensitivity_egress: internal }
  budgets: { max_tokens: 1000 }