Harness conformance checklist
Status: complete as of 2026-11-17, on the owner's request "finish the harness checklist". It is
a checklist, not a standard: it has no schema and no vocabulary terms. It is published at
standards.bundu.org/harness-checklist/ (owner,
2026-11-19), so its register entry (harness_checklist) is proposed, awaiting the owner's
adoption. Tracked in #50.
A harness is the loop around a model: it builds the context, calls the model, runs tools, asks
for permission, keeps a log and may run code in a sandbox. Following AGENTS.md ("adopt the global
standard; where none exists, write one"), this checklist does not define a harness. It says what a
conforming Bundu harness must expose, names the existing standard to use for each part, and states
the gaps. Claims are labelled measured (read at source on the date and URL given) or
inferred (my reading; verify before relying on it).
1. Parts that already have a standard: use it
Every row below is a standard in the register (register/register.yaml). The version and licence are
the register's, each read at source on the date it records.
| Part | A conforming harness must… | Use (register id, status) | Version / licence |
|---|---|---|---|
| Tools and context | call tools and read context through the Model Context Protocol | MCP (mcp, adopted) | revision 2026-07-28 |
| Instructions | read project instructions from AGENTS.md | AGENTS.md (agents_md, adopted) | unversioned; MIT |
| Skills | load reusable skills as SKILL.md folders | Agent Skills (agent_skills, adopted) | unversioned (no tags); Apache-2.0 code, CC-BY-4.0 docs |
| Model wire shape | exchange one request and response as Open Responses items | Open Responses (open_responses, adopted) | pin the dated OpenAPI file (2026-04-24), not openapi.json; Apache-2.0 |
| Agent to agent | delegate to other agents through A2A | A2A (a2a, adopted) | 1.0.0 |
| Sandbox lifecycle | create, run in and destroy a sandbox through a standard protocol | OpenSandbox (opensandbox, adopted 2026-11-17) | Lifecycle API 0.1.0 (pre-v1), Execd API 1.0.0; Apache-2.0; an AAIF project. Measured 2026-11-17 at commit f1e4ac0 |
| Editor to agent | if it talks to an editor, use ACP | Agent Client Protocol (agent_client_protocol, adopted 2026-11-15) | protocol v1; Apache-2.0; interim Zed and JetBrains governance, no foundation host |
| Stored conversations | export and import chats, projects, saved answers | BCI (bci, adopted 2026-11-15) | 0.1.0; first implementation nyuchi/nyuchi-servers-api#50; vocabulary terms stable in vocab.bundu.org |
| Session log | write each session as an ordered trajectory | ATIF (atif, adopted 2026-11-14) | ATIF-v1.8 (RFC 0001, Active, April 2026); Apache-2.0 |
ACP and BCI were adopted in #70 (merged 2026-11-17), and OpenSandbox by the owner on 2026-11-17. Every row's standard is now adopted. OpenSandbox's Lifecycle API is pre-v1, so a harness pins the OpenAPI files it builds against. Bundu runs no sandbox today, so that row applies only once a Bundu harness runs code.
The Agentic AI Foundation's projects
The AAIF lists ten projects (aaif.io/projects, read 2026-11-17). Measured.
-
In the table above (6): MCP, AGENTS.md, Agent Skills, Open Responses, A2A and OpenSandbox.
-
OpenEnv: adopted (
openenv, owner 2026-11-17). These are environments for training agents with reinforcement learning, from Hugging Face, under BSD-3-Clause. Its environment spec (RFC 002) is still In Review, so a trainer pins the release. It becomes a harness row when Bundu trains agents. -
Implementations, not standards (3), adopted into the build (owner, 2026-11-17: "the software we should be adopting as well … The agent router will [be] our routing"):
- Agent Router, a gateway on Envoy for models and MCP tools (v1.2.1, formerly Envoy AI Gateway): Shamwari's model routing;
- agentgateway, a gateway for agents, models and tools (v1.7.0): MCP and A2A traffic;
- goose, an agent (see the dialect table in section 3): a client that can use Shamwari.
All three are Apache-2.0. The register lists standards, so they are noted here and not there. Where each sits in Shamwari, always behind its own Rule 1 scope gate, is in
shamwari-ai/shamwaridocs/aaif-runtime.md(shamwari-ai/shamwari#325).
2. Parts a second gap check found partly covered
Read at source 2026-11-11. ATIF: harbor-framework/harbor, rfcs/0001-trajectory-format.md (Apache-2.0, last push 2026-11-11, current
format ATIF-v1.8, no spec release other than the RFC's version history). OpenTelemetry GenAI:
open-telemetry/semantic-conventions-genai docs/gen-ai/ (Apache-2.0, last push 2026-11-11, every
document status Development, no release cut).
| Part | What exists | Evidence |
|---|---|---|
| Session log | ATIF-v1.8 is a session-log interchange format: ordered steps with source (system, user, agent), message, tool_calls, observation, metrics, embedded subagent trajectories and audio. Measured. It is Harbor's own format, written for evaluation and training runs. | ATIF RFC, sections II and version history |
| Context management | ATIF section VII is a convention (step.extra.context_management) that records compaction, pruning and injection, with a normative replace boundary rule, plus is_copied_context. Measured. It records what happened; it does not say what a harness should keep. | ATIF RFC, section VII |
| Telemetry | OTel GenAI defines spans for inference, embeddings, retrieval, execute_tool, create_agent, invoke_agent, memory operations (gen_ai.memory.client) and agent-skill loading, with gen_ai.conversation.id. Measured. Development status: names may change. | README.md, gen-ai-spans.md, execute-tool.md, gen-ai-agent-spans.md |
| Permissions | ACP v1 has session/request_permission (coordinator, 2026-11-11). No permission field appears in the ATIF RFC or in six OTel GenAI documents searched by keyword (permission, approval). Measured for those files only. | keyword search of the files named |
3. Gaps, checked at source 2026-11-17
Each gap below was checked at source on 2026-11-17. Repositories were read at these commits: Codex 05ae1b0, Gemini CLI 4fa64e7, OpenHands software-agent-sdk 06f643c, goose eb4f412, MCP 05aeb07, A2A b30ab16, OpenTelemetry semantic-conventions-genai e6658f1, Harbor 110c71b, Open Responses dce9ddb and LangGraph 02f2c2e.
-
Context policy: partly covered. This is what a harness must keep, drop or summarise, and who may see it.
- Open Responses 2026-04-24 has
truncation("auto": the server "MAY truncate earlier parts" and "SHOULD apply sensible heuristics"),previous_response_id(the server "MUST preserve the semantic order") and a/responses/compactendpoint. That is provider-side policy, not harness policy. - ATIF section VII reconstructs the context after the fact.
- OTel GenAI says instrumentations "SHOULD NOT capture" inputs and outputs by default, which covers telemetry only.
- No rule for what a harness keeps, or for who may see it, was found in MCP. Measured.
- Open Responses 2026-04-24 has
-
Permission model: partly covered.
- MCP 2026-07-28 (
server/tools.mdx): "there SHOULD always be a human in the loop with the ability to deny tool invocations". - MCP elicitation (
client/elicitation.mdx): clients MUST "provide clear decline and cancel options". - MCP tool annotations (
destructiveHintand the others) are hints that clients MUST treat as untrusted unless the server is trusted. - A2A 1.0.0 section 7.6: an agent that needs human approval before a destructive action
MUST move to
TASK_STATE_AUTH_REQUIRED. Section 7.6.4 leaves "scope, representation, validity, or revocation" undefined. - ACP has
session/request_permission, between an editor and an agent only. - Measured. No standard defines remembered grants, grant scope or how an approval is recorded.
- MCP 2026-07-28 (
-
Cross-harness tool-call dialects: still a gap. Four wire dialects are in use: Chat Completions
tool_calls, the Responses API'sfunction_callitems, Anthropictool_useand GeminifunctionCall. None of the products below claims Open Responses conformance. The Codex CLI accepts only the Responses wire, which Open Responses is based on. Measured for the repositories and vendor pages named. The Qwen Responses page was read from search excerpts only.Product Tool-call wire Sessions on disk MCP AGENTS.md Skills ACP DeepSeek Chat Completions; Anthropic Messages; Responses (no store)n/a n/a n/a n/a n/a Qwen Chat Completions; Responses (OpenAI-compatible); DashScope native n/a n/a n/a n/a n/a GLM (z.ai) Chat Completions; Anthropic Messages n/a n/a n/a n/a n/a Codex CLI Responses only ( wire_api = "chat"is refused)JSONL rollout files yes yes yes through an adapter Gemini CLI Gemini functionCallJSONL chat files yes if configured yes yes, as an agent OpenHands Chat Completions or Responses, through LiteLLM a state file and one JSON file per event yes yes yes yes, as a client goose its own message type, translated per provider SQLite yes yes yes yes, as an agent So the harnesses agree on MCP, AGENTS.md, Agent Skills and, increasingly, ACP. They disagree on the model wire and on how a session is stored. BCI (stored conversations) and ATIF (session log) are the register's answer to the second disagreement. Nothing in the register answers the first beyond adopting Open Responses.
-
Portable session log: decided. The owner adopted ATIF on 2026-11-14 (register id
atif). BCI's orderedMessagestream maps onto ATIF steps rather than defining its own log; the mapping is an open item on #48. ATIF is Harbor-hosted and evaluation-oriented, with no foundation host, so the adoption is of the format, not an endorsement of its steward.- ISO/IEC 24970 "AI system logging" (JTC 1/SC 42) is at stage 60.00, under publication. It is the nearest formal standard and has not been read here. Measured status only.
Checked and found not to be harness standards:
- IEEE P3119 is about AI procurement. Measured.
- IEEE P3394 "Large Language Model Agent Interface" is an active PAR from 2023-09-21, not published, and its page does not mention tool calls, logs or context. Measured.
- ISO/IEC 42105 (human oversight) and ISO/IEC 42001 (a management system) are not harness formats. Measured for 42105's stage, inferred for the rest.
- ISO/IEC AWI TR 27258 (multi-agent interactions) is an early technical report. Measured.
- LangGraph checkpoints (
v: 1) are a versioned internal persistence format, serialised as Python types, not an interchange format. Inferred from the source.
4. Bundu's own rules (not an external standard)
These are Bundu's rules for a conforming harness, taken from Shamwari. They are requirements here, not part of any standard above.
- Personal-scope data stays on the person's device and under their keys. It is not sent to a third-party inference service.
- Only
open_weightprovenance may train. Storing a conversation is not consent to train on it. - No memory entry without the person having stated or confirmed it.
5. How to use this
A harness passes when, for every row of section 1 that applies to it, it uses the named standard,
and it satisfies section 4. Section 3 lists what it cannot yet be checked against. Where section 3
names a partial standard (MCP's human-in-the-loop rules, A2A's AUTH_REQUIRED), a harness follows
it.
What is left is the owner's, not the checklist's:
- Whether to write a Bundu profile for the two partly covered gaps: context policy, and permission
grants.
AGENTS.md's rule says to write one where none exists. The checklist does not start one. - Running this checklist against Shamwari's harness (
shamwari-ai/shamwari). The data-plane audit is nyuchi/nyuchi-servers-api#48.