generated: '2026-09-19'
method: searched
source: https://raw.githubusercontent.com/ogasurfproject-jpg/horizon-shield/main/workers/hs-ledger/nenrin/a2a-conduct-walk/pyproject.toml
docs:
- https://github.com/ogasurfproject-jpg/horizon-shield/blob/main/skills/conduct-witness/SKILL.md
- https://github.com/ogasurfproject-jpg/horizon-shield#readme
name: a2a-conduct-walk
version: 1.1.0
language: Python (standard library only, >=3.8)
license: Apache-2.0
package: packages/horizonshield-dev-packages.yml
summary: >-
  A first-party CLI that walks an A2A or MCP agent from the user's own machine as an independent witness, checks the
  five Conduct Extension conditions, and files the observation on the public JIDEC ledger under the user's name with
  a SHA-256 anyone can recompute. The same distribution ships a second entry point, conduct-witness-mcp, a stdio MCP
  server so an agent can be a witness in one tool call. Install methods, flags and the reporting contract below are
  taken verbatim from the provider's conduct-witness skill and pyproject; the CLI was not executed by this pipeline.
install:
- method: uvx (from git)
  command: 'uvx --from "git+https://github.com/ogasurfproject-jpg/horizon-shield#subdirectory=workers/hs-ledger/nenrin/a2a-conduct-walk" a2a-conduct-walk --origin <ORIGIN> --mode a2a --submit --witness-name "<NAME>" --vantage "<VANTAGE>"'
- method: single-file fallback (no uv)
  command: 'curl -sSLO https://raw.githubusercontent.com/ogasurfproject-jpg/horizon-shield/main/workers/hs-ledger/nenrin/a2a-conduct-walk/a2a_conduct_walk.py && python3 a2a_conduct_walk.py --origin <ORIGIN> --mode a2a --submit --witness-name "<NAME>" --vantage "<VANTAGE>"'
entry_points:
- {command: a2a-conduct-walk, module: 'a2a_conduct_walk:main', kind: CLI}
- {command: conduct-witness-mcp, module: 'conduct_witness_mcp:main', kind: stdio MCP server}
commands:
- group: walk
  flags:
  - {flag: --origin <https URL>, meaning: 'The agent''s card origin (card at <origin>/.well-known/agent-card.json), not a page on the site.'}
  - {flag: --mode a2a, meaning: Walk over A2A (MCP mode also exists per the README).}
  - {flag: --submit, meaning: 'File the observation on the ledger; omit to preview the assertions first.'}
  - {flag: --witness-name "<NAME>", meaning: 'A person, a project, or "anonymous" — never invented; operators must say so ("Example Corp (operator)").'}
  - {flag: --vantage "<VANTAGE>", meaning: 'Where the machine is, in a few words ("laptop, Osaka", "VPS eu-west", "CI runner").'}
  - {flag: --transport curl, meaning: Rerun with curl if a fetch node shows http 403 at the edge.}
  - {flag: --bind-task, meaning: 'Bind the observation to the real a2a.task.id returned by the walked agent (README "Task-bound conduct").'}
output_contract:
  report_back: [the purpose line (which endpoint), 'PASS n/5 or FAIL n/5', the record sha256, 'the intake''s answer (submitted … http 200, or why not)', 'the assertion list with pass / FAIL / n/a']
  where_it_lands: 'Bundled once a day at 00:30 UTC at https://ledger.horizonshield.dev/ledger; the month''s ring counts the witness by name.'
  rules: ['A FAIL is filed the same way as a PASS.', 'Do not walk an agent whose owner has declined (listing: "decline" in /.well-known/mcp-conduct.json).', 'One walk per agent per day; never submit twice to pad a count.', 'Never call it a rating, a score, or a verification of quality.']
related_tools:
- {name: mcp-conduct, kind: npm library, purpose: consult the gate before an MCP client connects}
- {name: mcp-conduct-action, kind: GitHub Action, purpose: measure a repo's own MCP server in CI}
- {name: Claude Code plugin slash commands, commands: [/audit, /red-flags, /verify], note: Chat commands, not a shell CLI.}