@lucagiftzek/dsh-subagent-tiering

dsh plugin --profile web add @lucagiftzek/dsh-subagent-tiering

dsh-subagent-tiering


Subagents run on a cheaper sibling of the parent's model unless the main agent

names a route itself.

Install

Install from npm: dsh plugin --profile web add @lucagiftzek/dsh-subagent-tiering, or from GitHub: dsh plugin --profile web add github:lucagiftzek/dsh-subagent-tiering.

dsh plugin --profile web add @lucagiftzek/dsh-subagent-tiering
# or from GitHub: dsh plugin --profile web add github:lucagiftzek/dsh-subagent-tiering
# then restart `dsh web`

The package declares dsh.bundle, so dsh plugin add lists it in dsh.profile.bundles and

its own patch inserts the subagent-tiering entry with the built-in rules. A local checkout

works too: dsh plugin --profile web add link:/absolute/path/to/dsh-subagent-tiering.

Requires Node 22.18 or newer; used on DSH 0.1.5 and 0.2.1. Read "Needs" below: without a

subagent model-selection policy the plugin changes nothing.

Parent route · Child default

anthropic Opus / Fable / Sonnet · anthropic claude-sonnet-5-5 @ medium (falls back to claude-sonnet-5)

alibaba Qwen max/plus · qwen3.8-flash @ high

alibaba / byteplus DeepSeek pro · the provider's DeepSeek flash @ high

byteplus GLM / Seed · glm-5-3-flash-260828 @ high

deepseek-official · deepseek-flash @ high

mistral large · mistral-medium-latest

any other mistral, alibaba GLM, hetzner · same model, effort capped at high (never raised)

anything else · inherits the parent, as stock dsh does

vision-toolkit-<x> parents keep their wrapper provider. Rules live in

src/rules.js and can be overridden with a - id: subagent-tiering row in the

profile's cordis.patch.yml (config.rules, same shape).

A rule's child.fallbackModels (optional list of model ids, same provider and

effort) is tried in order when the primary model is not in the chat's captured

allow-list or does not resolve. 2026-09-29: the Anthropic default moved from

Sonnet 5 to Sonnet 5.5; chats created before that day have no Sonnet 5.5 in

their allow-list, so they keep getting Sonnet 5 instead of inheriting Opus.

Why (2026-09-27)

Over 14 days of session logs on one busy install, 70% of dsh's Anthropic calls were

subagents. The stock ptc preset gives subagent no child defaults, so every

child ran on the parent's route and effort (usually Opus 5.5 at max). Per human

prompt, dsh cost 2.1x what Claude Code did on the same box, while per call the

two were about equal. tool-subagent's own agentOptions is one static route

and cannot express "Opus parent → Sonnet child, Qwen parent → Qwen-flash child".

How it works

A tools/pre-execute handler. When a subagent call names no provider,

model or reasoning_effort, it fills them in from the first matching rule, as

if the model had asked. The stock tool then runs its normal checks. It leaves the

call untouched, so the child inherits the parent, when:

• the main agent passed any of the three fields itself (its choice always wins);
• the chat has no model-selection policy. dsh records one only for chats created

while subagent-model-selection.enabled is true, so chats that predate

2026-09-27 are not affected;

• the target route is not in that policy's allowedModels;
• llm.resolveCallConfig rejects the route. If it rejects only the effort, the

effort is dropped and the route kept;

• core dsh ever freezes the execution record. This is logged once.

subagent_fork is never touched. A fork must stay on the parent route to reuse

its cached prefix.

A short system-prompt section (subagent-tiering, about 90 tokens) tells the

model about the default and how to override it.

Trade-off

A child on a different model cannot reuse the parent's prompt cache, so its

first call pays a full cache write (about 48k tokens on Anthropic after the GenUI

removal). A one-call subagent can therefore cost slightly more than before. The

median subagent here makes 28 calls, so it comes out well ahead.

Needs

• settings subagent-model-selection: { enabled: true, allowedModels: [...] }.

On DSH 0.2.x this harness setting lives under the entry

subagent-model-selection-settings and is not imported from an old settings.yaml

(see [docs/TROUBLESHOOTING.md](docs/TROUBLESHOOTING.md#t-after-update)).

• the ptc / ptc-lean preset's tool-subagent row with

modelSelectionSettings: true (stock).

• adaptive-reasoning must NOT list anthropic/claude-sonnet-5-5 (or

claude-sonnet-5, the fallback) among its routes,

or it would override the child's medium effort on every step.

Verify

node --test test/*.test.js runs the unit tests. End to end: start a new chat on

Opus 5.5 and have it call subagent with no model. The child session's

request/header shows claude-sonnet-5-5 / medium (was claude-sonnet-5

until 2026-09-29). Checked on 2026-09-27 on

a production install (Opus → Sonnet) and in a sandbox (Qwen max → Qwen flash; an explicit

qwen3.8-max @ low was left as asked).

When something goes wrong

• Ask your agent to load the subagent-tiering-doctor skill (copy

skills/subagent-tiering-doctor into ~/.dsh/skills/ once,

[how](docs/TROUBLESHOOTING.md#skill)).

• Symptom-first guide: [docs/TROUBLESHOOTING.md](docs/TROUBLESHOOTING.md).
• Log lines and start-up messages: [docs/ERRORS.md](docs/ERRORS.md).

---

Built by Luca Tzekos, who also makes Job Opportunities API: employer-direct job data with per-field provenance you can verify yourself, free to read and available as an API. If this plugin is useful to you, a star on this repo or a look at jobopportunitiesapi.org is the best thanks.