holocron/docs/adr/0016-apm-agent-primitive-drops-provider-specific-fields.md

Plugin-scope agent-author omits tools: and all Claude-only fields from .apm/agents/*.agent.md

This ADR is a narrower, downstream consequence discovered while designing issue #89's implementation under ADR-0015's broader direction (Microsoft APM replaces hand-authored plugin/marketplace authoring). It does not restate ADR-0015's rationale — see that ADR for the parent decision.

Context

APM's agent primitive (.apm/agents/<name>.agent.md) has no per-target integrator in apm compile — confirmed via APM's own Python source (integration/targets.py and related files, cited in plugins/kyberforge/docs/research/docs/microsoft-apm/agent-primitive-schema.md). Compilation does a naive verbatim copy of the whole frontmatter and body to both the Claude Code and Copilot CLI targets. This is unlike:

• The skill primitive, which is also a straight copy (confirmed in the same research doc) but has no field semantics to conflict — SKILL.md's content is target-agnostic already.
• The prompt, instructions, and hooks primitives, which each get real per-target reconstruction through a dedicated integrator (field allowlisting, key renaming, dropped-field warnings).

Because the agent primitive ships the same frontmatter unchanged to both harnesses, two concrete incompatibilities surface:

1. tools: — Claude Code expects tool names drawn from its own vocabulary, as a comma-separated string or a YAML list (agent-definition.md:37); Copilot CLI expects a list drawn from a different alias vocabulary (execute/read/edit/search/agent/web). The incompatibility is the vocabulary, not the punctuation: a value correct for one harness names tools the other does not have.
2. Claude-only knobs with no Copilot equivalent — isolation, maxTurns, effort, memory, permissionMode. Writing any of these means Copilot's copy carries frontmatter keys it doesn't recognize at all. Whether Copilot's agent loader ignores unknown keys or errors on them is unconfirmed by research. (Still unconfirmed as of the 2026-08-14 amendment below, which admits disallowedTools as an explicitly accepted risk rather than by resolving this question.)

Decision

At plugin scope only (destination package has an apm.yml at its root — an APM producer package compiled via apm compile), .apm/agents/<name>.agent.md carries only name, description, model, and the prose body. No tools: field, no Claude-only fields, at all.

(Narrowed by the 2026-08-14 amendment below: disallowedTools is admitted as a fifth allowed field. tools: and every other Claude-only knob remain excluded on the reasoning given here.)

Absent tools: means inherit-all-tools on both harnesses — the one value that is never wrong on either target, unlike a present, harness-specific value that is guaranteed wrong on at least one of them.

agent-audit, at plugin scope, is intended to flag — as a SUGGESTION, not a FAIL, since this is an upstream schema limitation rather than an authoring mistake — any agent whose description or body implies a need for tool restriction or a Claude-only behavior the frontmatter can no longer express. This would give visibility into the gap without pretending the schema can do something it can't. Not yet implemented: check_apm_agent_file() in validate.sh currently validates only the field allowlist, name, description, and body-emptiness/length — it has no heuristic for this case. Tracked as follow-up work.

Scope boundary

This decision applies to plugin-scope agent-author only. Project scope (.claude/agents/

• .github/agents/) and user scope (~/.claude/agents/ + ~/.copilot/agents/) are not APM packages — neither goes through apm compile — so both keep today's dual-file Claude+Copilot pair model exactly as ADR-0005 and ADR-0008 already describe. Those two ADRs remain fully authoritative for project and user scope; only their plugin-scope clauses are affected by this ADR (see the update notes appended to each).

Considered options

Pick one harness's vocabulary and accept breakage on the other (rejected). E.g. always write Claude's space-separated tools: string. Rejected because it ships a value that is silently wrong (or possibly a hard error) on Copilot, and which harness "wins" would be an arbitrary, undocumented asymmetry.

Same as above, but agent-audit flags the cross-harness breakage as a tracked finding (rejected). Rejected for the same core reason — it still ships a wrong value to a real harness. Tracking the breakage doesn't prevent it, and the chosen decision already gets equivalent visibility (a SUGGESTION finding) without ever shipping the wrong value in the first place.

Amendment (2026-08-14): the write fence comes back as a denylist

The decision above generalised from tools: to "no tool restriction at all". That over-reached. The unportability argument is specific to the allowlist: Claude Code reads tools: as a delimited string of its own tool names, Copilot CLI reads it as a list drawn from its alias vocabulary (execute/read/edit/search/agent/web), so one value is wrong on one harness. That reasoning stands, and tools: stays out of every plugin-scope agent.

A denylist has no such conflict. The evidence for that splits three ways, and this amendment states which part is which rather than asserting the whole as settled.

Confirmed — Claude Code honours it for plugin subagents. plugins/kyberforge/docs/research/docs/claude-code-plugins/agent-definition.md:39 documents disallowedTools as a "Denylist applied before tools… Takes precedence over tools", and — the part that matters here — it is not in that document's plugin-subagent ignore list. Line 99 names exactly three fields plugin agents silently ignore: hooks, mcpServers, permissionMode. disallowedTools is absent from that list. Claude Code is also the harness where the fence is actually wanted, so the field earns its place on this evidence alone.

Inferred — the field is very likely inert on Copilot CLI, but by analogy, not by documentation. plugins/kyberforge/docs/research/docs/github-copilot-plugins/troubleshooting.md:50 and :53 record Copilot silently ignoring two agent frontmatter fields it does not process (mcp-servers and metadata outside the cloud runtime) rather than erroring on them. That is a documented tolerance for known-but-unprocessed keys, which is adjacent to, not identical to, tolerance for an unknown key. No stronger evidence exists: a sweep of the vendored Copilot corpus (agent-definition.md, api-reference.md, troubleshooting.md, configuration.md) documents unknown-key handling nowhere.

Unverified — Copilot's loader behaviour on an unrecognised key. Context item 2 above says this is unconfirmed by research and that remains true; nothing found since changes it. An earlier revision of this amendment claimed "an unrecognised frontmatter key is inert" as settled fact and attributed it to apm's verbatim-copy behaviour. That attribution was a non-sequitur — verbatim copy describes what apm does at compile time and says nothing about what Copilot does at load time — and the claim contradicted this ADR's own Context section.

So this is an accepted risk, stated as one. Blast radius if the inference is wrong and Copilot errors on the key: the three affected plugin-scope agents fail to load under Copilot CLI. It is loud, not silent; it is confined to three agents in three plugins; no other primitive and no Claude Code path is affected; and the remedy is a one-line frontmatter deletion. What the denylist shape does rule out categorically — independent of loader behaviour — is the failure mode that motivated dropping tools: in the first place: a denied name the other harness does not recognise denies nothing, so a mis-shaped value can never grant or misroute a capability. The risk is a load failure, never a silent over-grant. That asymmetry is why the same verbatim copy that makes tools: unshippable makes disallowedTools worth shipping.

So the read-only orchestrator agents regain their write fence: gitea-orchestrate, apm-orchestrate and lint-runner each carry disallowedTools: Edit, Write, NotebookEdit plus explicit prose in the body stating the agent does not edit files. git-orchestrate is deliberately excluded — it legitimately declared edit before the conversion and still needs to write.

Residual — the fence is partial, and the prose is doing more of the work than the field is. disallowedTools: Edit, Write, NotebookEdit denies exactly those three tools. It does not deny Bash, and at plugin scope these agents carry no tools: and therefore inherit it, so bash -c 'echo … > f' remains unfenced by frontmatter. Only the body prose covers that path. This is not a regression introduced here — the pre-conversion tools: allowlists also granted Bash, so the shell route was open then too — but the ADR should not credit the mechanism with more than it delivers. Closing it would need a disallowedTools entry for Bash, which these agents cannot take because they legitimately shell out.

Net position: the allowlist stays dropped for the reason originally given, and the denylist is admitted as the portable-by-construction half of what was lost. It restores a real, Claude-Code- confirmed write fence against the tool-call path, not a complete write sandbox. The consequence below is narrowed accordingly.

Enforcement follows the decision: agent-audit's plugin-scope validator reads its allowlist as data from the apm-agent-allowlist section of plugins/kyberforge/.apm/skills/agent-audit/references/field-inventory.md (now factory-audit/references/agent-field-inventory.md, see ADR-0025), and that line now reads name description model source_keys disallowedTools. disallowedTools also stays in that file's claude-code-only-fields list, which is not a contradiction — that list governs whether a field may cross the CC/Copilot boundary in a real project/user-scope pair, a different question from whether a field is safe under verbatim copy in a single vendor-neutral file.

Consequences

• Every plugin-scope APM agent loses per-agent tool allowlisting and any Claude-only capability (isolation, maxTurns, effort, memory, permissionMode) until APM ships a real per-target integrator for the agent primitive. This is a known, accepted regression, not an oversight. Tool denial is not part of that loss — see the 2026-08-14 amendment above.
• ADR-0005 is partially superseded — its plugin-scope clause ("directory containing plugin.json is plugin scope → both files land in <root>/agents/") no longer applies. Plugin scope is now "directory containing apm.yml → single vendor-neutral file lands in <root>/.apm/agents/." Project and user scope, and the rest of ADR-0005, are unaffected.
• ADR-0008 is partially superseded — its counterpart-derivation/pair-validation mechanism no longer applies at plugin scope; agent-audit takes the single file directly there. Project and user scope, where a real pair still exists, are unaffected.
• ADR-0009 is not superseded. The mechanism it established — agent-audit reading field lists from references/field-inventory.md (now factory-audit/references/agent-field-inventory.md, see ADR-0025) rather than hardcoding them, with a source_keys provenance chain — survives and is reused. Only the content shape changes for plugin scope: field-inventory.md shifts from two side-by-side CC-only/Copilot-only blocklists to one vendor-neutral allowlist for plugin-scope agents, while continuing to serve its original two-blocklist role for project/user-scope validation. That file's apm-agent-allowlist section is the authoritative list and is read as data by validate.sh; as amended on 2026-08-14 it holds name/description/model/source_keys/disallowedTools — source_keys for provenance tracking, validated separately by validate-provenance.sh against sources.md rather than being a provider-specific field, and disallowedTools per the amendment above.