Repository navigation
Expand file tree
/
Copy pathdefaults.yaml
More file actions
101 lines (97 loc) · 9.95 KB
/
Copy pathdefaults.yaml
File metadata and controls
101 lines (97 loc) · 9.95 KB
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
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
# Bundled role map: the bottom layer of the multi-agent cascade.
# Keys and layering: reference/config.md. A consumer layer overrides per key.
#
# model: an alias (opus, sonnet, haiku, fable) or `inherit`. `inherit`
# means a workflow omits opts.model on a generic agent() call
# (reference/config.md).
# effort: low | medium | high | xhigh | max. Always passed explicitly.
# pointer*, as_of, recheck: where the default's basis lives, read by
# /multi-agent:audit-defaults. The resolver never emits them.
# basis_<field>: per role, `upstream` or `judgment` for that role's <field>
# (model or effort). `upstream` names pointer_<field>, the section whose
# statement sets the value or the rule it follows from, and claim_<field>,
# which statement in it. `judgment` names reason_<field>: no upstream
# statement sets the value, and why it is what it is. The resolver emits the
# basis beside each value; the audit checks that an upstream pointer states it.
# A workloads.<w>.<field> value records its basis on the role as
# basis_<w>_<field>, with pointer_<w>_<field> and claim_<w>_<field> or
# reason_<w>_<field>.
# Which roles may route to haiku, and at what effort: reference/routing-rubric.md.
schema: 1
# Workloads a role may set values for (roles.<r>.workloads.<w>). A consumer
# layer declares its own the same way, or withdraws one with false.
workloads:
code: true
research: true
mechanical: true
# Session model aliases treated as frontier by the fan-out guard.
frontier: fable
# The fan-out guard: the resolver names no frontier model for a stage that runs
# more than one agent. Under a frontier session, or when the caller does not name
# the session model, every role's fan-out variant names fanout.model where it
# would otherwise inherit (or name) a frontier model. Single-agent stages keep
# the role's own model. Turning frontier_guard off is an explicit opt-in to
# frontier fan-outs.
fanout:
frontier_guard: true
model: opus
pointer: https://code.claude.com/docs/en/workflows#cost
pointer_cost: https://platform.claude.com/docs/en/about-claude/models/optimizing-for-cost-and-intelligence#orchestrator-strategy-delegate-bulk-work
as_of: 2026-11-09
recheck: a new model named on the cost page, or a workflows Cost section change to how a stage's model is chosen
roles:
orchestrator:
model: inherit
effort: high
basis_model: judgment
reason_model: the orchestrator is the session itself, so it runs on the model the user chose; the cost page's orchestrator strategy puts a frontier model in the loop, which inherit follows only when the session is frontier, and no upstream statement sets the orchestrator's model for other sessions; the owner's decision (session model, Opus by default, Fable for long-horizon work) is docs/adr/0100-run-the-orchestrator-on-the-session-model-opus-by-default.md
basis_effort: judgment
reason_effort: no upstream statement sets this orchestrator's effort; the model guide's task table (https://claude.dev/blog/which-claude-model-should-you-build-with#pick-a-model-by-task) starts a Fable lead of many subagents at medium, then high, while this orchestrator inherits the session model and merges every worker's result, so it takes the second step, which pointer_levels' effort table gives to work where verification matters or edge cases are likely
pointer_levels: https://code.claude.com/docs/en/model-config#choose-an-effort-level
pointer_skill: /workflow-authoring (bundled first-party skill, agent() opts.model and opts.effort)
as_of: 2026-11-09
recheck: the guide's task table changes the starting effort for leading many subagents, model-config's effort table changes the row for verification-heavy work, the cost page's orchestrator-strategy section changes which model holds the loop, the Projects page's coordinator default changes (https://code.claude.com/docs/en/claude-projects#choose-models-and-let-claude-manage-context; the CLI changelog does not track Projects), or a Claude Code release edits the bundled /workflow-authoring skill
worker:
model: inherit
effort: medium
workloads:
mechanical:
model: sonnet
basis_model: judgment
reason_model: the cost page's orchestrator strategy delegates to lower-cost workers, and the fan-out guard applies that under a frontier session; under any other session no cheaper alias is set, because the same page measured delegation paying only in named cases
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 day-to-day engineering work with a clear scope
basis_mechanical_model: judgment
reason_mechanical_model: mechanical work needs no Opus-level reasoning; the model guide's task table (https://claude.dev/blog/which-claude-model-should-you-build-with#pick-a-model-by-task) starts well-scoped coding, such as bug fixes, feature iteration and spec-driven work, on Sonnet, but sits on a host the drift audit does not fetch, so it is recorded as context rather than an upstream pointer. A separate, directional measurement on five cases, taken 2026-11-14 on Claude Code 2.1.325 for the fix-docs lane activity, had Sonnet 5.5 at medium effort score 1.00 on all 15 runs of the ai-slop plugin's five audit-fix eval cases (3 runs each, deterministic graders only). The planned Opus arm was not run, a deviation from the plan made after the Sonnet result, because at that ceiling no Opus score could move the default. The result compares no models and is no pass rate for production fix-docs runs, and none of the measured cases checked whether a fix changed what a sentence means. A second directional measurement, started 2026-11-14 on the same version, effort and graders, added the audit-fix-keeps-referent case and ran both models on the six cases, 3 runs each, with its decision rule written before the first run. Sonnet 5.5 had a mean case score of 0.96 and Opus 5.5 had 1.00, so by that rule the suite sat at its ceiling and the default stays. Sonnet failed the added case on all 3 runs by leaving the figure of speech in place, and Opus passed it on all 3. The paired difference over the six cases was -0.04 with a 95% interval of -0.12 to +0.04, which does not show Sonnet within 0.05 of Opus. A third directional measurement, taken 2026-11-15 on the same version, effort and graders after the ai-slop fix guidance for a figure of speech changed, had Sonnet 5.5 pass that case on 4 of 5 runs against 0 of 5 on the earlier guidance run beside it, pass the other five cases on 25 of 25, and pass a new case that checks a fix leaves plain idioms alone on 5 of 5; its rule, written before the first run, allowed one revision of the guidance, and the result comes from the revised text
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)
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
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)
retrieval:
model: sonnet
effort: medium
basis_model: judgment
reason_model: the model guide's task table (https://claude.dev/blog/which-claude-model-should-you-build-with#pick-a-model-by-task) starts subagents that search files, read logs or summarize on Haiku or Sonnet; Haiku is excluded by owner rule (routing-rubric.md), so sonnet; the guide sits on a host the drift audit does not fetch, so it is recorded as context rather than an upstream pointer
basis_effort: upstream
pointer_effort: https://code.claude.com/docs/en/model-config#choose-an-effort-level
claim_effort: the effort table's medium row names it the default on the Sonnet 5.5 model
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)