From 20e7a67e85ed8ae8b8ebffe7f16a60bd26fee192 Mon Sep 17 00:00:00 2001 From: Kyle Sexton <153232337+kyle-sexton@users.noreply.github.com> Date: Mon, 16 Nov 2026 14:46:35 -0500 Subject: [PATCH] docs(multi-agent): recheck role defaults against subagent effort and completion-check guidance Two recheck triggers fired. Claude Code 2.1.316 added CLAUDE_CODE_SUBAGENT_EFFORT for subagents no other source names a level for; every routed dispatch passes effort explicitly, so no value changes. The Opus 5.5 prompting guide suggests a smaller model that checks an unattended conversation against a completion condition; that is a loop-continuation check, not a judge of another stage's output, so the verifier keeps inherit at high. The worker, verifier and retrieval entries move to as_of 2026-11-16 with rearmed recheck conditions, and the verifier records the guide section and why it does not apply. Co-Authored-By: Claude Opus 5.5 Claude-Session: https://claude.ai/code/session_01MWmASPFjTu4anSp6BowqVC --- ...ulti-agent-subagent-effort-recheck-3f9a1c27.md | 7 +++++++ plugins/multi-agent/reference/defaults.yaml | 15 ++++++++------- plugins/multi-agent/scripts/resolve-roles.test.sh | 2 +- 3 files changed, 16 insertions(+), 8 deletions(-) create mode 100644 .changes/multi-agent/docs-multi-agent-subagent-effort-recheck-3f9a1c27.md diff --git a/.changes/multi-agent/docs-multi-agent-subagent-effort-recheck-3f9a1c27.md b/.changes/multi-agent/docs-multi-agent-subagent-effort-recheck-3f9a1c27.md new file mode 100644 index 00000000000..cdfc825a83c --- /dev/null +++ b/.changes/multi-agent/docs-multi-agent-subagent-effort-recheck-3f9a1c27.md @@ -0,0 +1,7 @@ +--- +bump: patch +--- + +### Changed + +- The worker, verifier and retrieval defaults are rechecked against `CLAUDE_CODE_SUBAGENT_EFFORT` and the Opus 5.5 end-of-turn completion check, with no value changed. Their recheck conditions now fire when the docs say where that variable ranks, and the verifier's model reason records why a smaller completion-check model is not a verifier. diff --git a/plugins/multi-agent/reference/defaults.yaml b/plugins/multi-agent/reference/defaults.yaml index a01ea9d9ad3..bbc6cfba89c 100644 --- a/plugins/multi-agent/reference/defaults.yaml +++ b/plugins/multi-agent/reference/defaults.yaml @@ -71,21 +71,22 @@ roles: pointer_skill: /workflow-authoring (bundled first-party skill, agent() opts.model and opts.effort) pointer_spawn: https://code.claude.com/docs/en/sub-agents#choose-an-effort-level pointer_subagent_effort: https://code.claude.com/docs/en/model-config#effort-level-of-subagents-and-teammates - as_of: 2026-11-09 - recheck: model-config's effort table changes the row for clear-scope engineering work, the cost page's orchestrator strategy changes its worker-model advice, a new model is named on the cost page, the guide's task table changes its row for well-scoped coding, the Agent tool's per-spawn effort parameter changes, a role is proposed for haiku (see routing-rubric.md), the default subagent effort setting changes (pointer_subagent_effort), or an ai-slop audit-fix eval case fails a majority of its runs on Sonnet (the mechanical-model measurement, which carries its own date) + as_of: 2026-11-16 + recheck: model-config's effort table changes the row for clear-scope engineering work, the cost page's orchestrator strategy changes its worker-model advice, a new model is named on the cost page, the guide's task table changes its row for well-scoped coding, the Agent tool's per-spawn effort parameter changes, a role is proposed for haiku (see routing-rubric.md), pointer_subagent_effort or the sub-agents page documents CLAUDE_CODE_SUBAGENT_EFFORT (added in Claude Code 2.1.316, https://code.claude.com/docs/en/changelog, for subagents no other source names a level for) or a release lets it outrank a definition's or a call's effort, or an ai-slop audit-fix eval case fails a majority of its runs on Sonnet (the mechanical-model measurement, which carries its own date) verifier: model: inherit effort: high basis_model: judgment - reason_model: no upstream statement sets a verifier model; a verifier judges the other stages' work, so it is not one of the stages that can drop to a smaller model + reason_model: no upstream statement sets a verifier model; a verifier judges the other stages' work, so it is not one of the stages that can drop to a smaller model. pointer_completion_check suggests a separate, smaller model that checks an unattended conversation against a stated completion condition at each end of turn and sends the reason back as the next user message; that check decides whether the loop sends another turn, capped at two or three automatic continuations, and does not judge whether another stage's output is correct, so it is not this role + pointer_completion_check: https://platform.claude.com/docs/en/build-with-claude/prompt-engineering/prompting-claude-opus-5-5#unattended-agentic-runs basis_effort: upstream pointer_effort: https://code.claude.com/docs/en/model-config#choose-an-effort-level claim_effort: the effort table's row for work where verification matters or edge cases are likely pointer_skill: /workflow-authoring (bundled first-party skill, agent() opts.model and opts.effort) pointer_spawn: https://code.claude.com/docs/en/sub-agents#choose-an-effort-level pointer_subagent_effort: https://code.claude.com/docs/en/model-config#effort-level-of-subagents-and-teammates - as_of: 2026-11-09 - recheck: model-config's effort table changes the row for verification-heavy work, an upstream page states a model for a verifying or judging agent, the Agent tool's per-spawn effort parameter changes, or the default subagent effort setting changes (pointer_subagent_effort) + as_of: 2026-11-16 + recheck: model-config's effort table changes the row for verification-heavy work, an upstream page states a model for an agent that judges whether another stage's output is correct, the Agent tool's per-spawn effort parameter changes, or pointer_subagent_effort or the sub-agents page documents CLAUDE_CODE_SUBAGENT_EFFORT (added in Claude Code 2.1.316, https://code.claude.com/docs/en/changelog, for subagents no other source names a level for) or a release lets it outrank a definition's or a call's effort retrieval: model: sonnet effort: medium @@ -97,5 +98,5 @@ roles: pointer_skill: /workflow-authoring (bundled first-party skill, agent() opts.model and opts.effort) pointer_spawn: https://code.claude.com/docs/en/sub-agents#choose-an-effort-level pointer_subagent_effort: https://code.claude.com/docs/en/model-config#effort-level-of-subagents-and-teammates - as_of: 2026-11-09 - recheck: the guide's task table changes its row for search, log and summary subagents, the workflows Cost section changes its smaller-model advice, a model-config change to the sonnet alias or its default effort, the Agent tool's per-spawn effort parameter changes, a role is proposed for haiku (see routing-rubric.md), or the default subagent effort setting changes (pointer_subagent_effort) + as_of: 2026-11-16 + recheck: the guide's task table changes its row for search, log and summary subagents, the workflows Cost section changes its smaller-model advice, a model-config change to the sonnet alias or its default effort, the Agent tool's per-spawn effort parameter changes, a role is proposed for haiku (see routing-rubric.md), or pointer_subagent_effort or the sub-agents page documents CLAUDE_CODE_SUBAGENT_EFFORT (added in Claude Code 2.1.316, https://code.claude.com/docs/en/changelog, for subagents no other source names a level for) or a release lets it outrank a definition's or a call's effort diff --git a/plugins/multi-agent/scripts/resolve-roles.test.sh b/plugins/multi-agent/scripts/resolve-roles.test.sh index 21ad0d6c388..be81d672edf 100755 --- a/plugins/multi-agent/scripts/resolve-roles.test.sh +++ b/plugins/multi-agent/scripts/resolve-roles.test.sh @@ -292,7 +292,7 @@ out="$(run plain all --session-model opus)" assert_lacks "no haiku anywhere: no haiku note" "$out" 'resolves to haiku' out="$("$SUT" pointers)" -assert_contains "pointers lists each role's as_of" "$out" $'worker\tas_of\t2026-11-09' +assert_contains "pointers lists each role's as_of" "$out" $'worker\tas_of\t2026-11-16' assert_contains "pointers lists the fan-out guard basis" "$out" $'fanout\tpointer\thttps://code.claude.com/docs/en/workflows#cost' assert_contains "pointers lists a role's per-value basis" "$out" $'verifier\tbasis_effort\tupstream'