Status: retired. The preset directory that implemented it has been removed. This document is the record — what was built, what the host actually provides, what the measurements said, and what would have to change for the idea to come back.
It is kept because a good deal of it was learned by reading the installed DSH rather than its documentation, and that knowledge is worth more than the code it produced.
The two-tier loop does not answer one question: which files does a unit need, and do two units agree? The worker cannot answer it — it has no repository read. The architect could, but its window is the resource the whole plugin exists to protect.
So a third role was proposed: a local thinking model with repository access that reads the code, works out the contract for each unit, and hands it back. The architect stays out of the code; the lead sits between.
The roles were meant to mirror an engineering organisation — architecting, engineering, writing — with the lead as the engineer who turns an architectural intent into a concrete code-level change. It never decides what should exist; it decides how that lands.
Learned from the installed packages, not from documentation:
- Presets are directories holding
preset.yml(name, description, order) andagent.cordis.yml, an agent-plane composition. The composition decides the tool schemas and prompt sections an agent sees. DSH shipsminimal,standard,ptcandcordisas working references. - Discovery scans configured
roots, each with apathandtrust.includeShippedRootprepends the bundled presets;includeUserRootappends<dshHome>/.agent-presetsas auserroot. So installing one is a directory copy. The preset id is the directory name and must match[a-z0-9][a-z0-9-]*. - Discovery owns health. A directory whose composition is missing or unloadable becomes a broken roster row with a reason, not a silent skip.
- The model route is host-plane. A preset does not choose a provider or model; the session's selection
does. That is what made
leadTier/leadProvidersthe load-bearing plugin setting. dsh-tool-subagentinstances can set childpersona,toolFilter,agentOptions(provider, model, reasoningEffort, maxTokens) andmaxDepth— and the in-process spawn provider declares support for all of them:capabilities = { agentOptions: true, toolFilter: true, persona: true }. So an architect can spawn a child pinned to a local model, carrying a chosen persona, restricted to chosen tools, and forbidden from delegating further. That is the shape a lead tier should have been built as — a tool the architect holds, not a session the operator opens.
That last point is the correction worth remembering: the tier was first built as a session preset the user selects, which is two configured agents. The host already had the primitive for one configured agent that spawns the rest.
- The window does not fit. Re-measured; this ground no longer holds for a single module. At
retirement the lead's configured context was 32,768 tokens and
plugin/src/index.tswas 147 KB — roughly 40,900 tokens: one file, larger than the whole window, withplugin/srctogether at ~46,700. The refactor inrefactor.mdhas since split it. Measured now:index.tsis ~13,400 tokens, the largest module (delegation.ts) is ~7,300, andplugin/srctogether is ~45,700. Any one module therefore fits the lead's window with room to spare, and only the whole ofsrcdoes not. The objection as originally written — "the plugin cannot be engineered by its own lead" — is dead. - The reasoning budget is not ours to set. Unchanged. LM Studio's per-model setting is
authoritative. Four variants of one request — server default,
enable_thinking: false,true, andreasoning_effort: high— returned byte-identical results: 64 completion tokens, the same 183-character reasoning field, the same answer. Nothing the plugin or DSH sends changes it. Reproduce withscripts/probe-thinking.mjs. - The saving was on the wrong side. Re-measured; weaker, still true. Offloading typing is worth ~159,000 tokens across this project's entire history. At retirement the arithmetic was one 41,000-token file carried for fifty turns: ~2,046,000 input tokens. With the split, the largest module is ~7,300 tokens, so the same fifty turns cost ~365,000 — an order of magnitude less, and still more than twice the entire typing history. Reading still dominates typing; the margin is simply smaller than recorded.
- Nothing was ever proven. Unchanged. No lead-authored contract was dispatched and verified. The failure mode is an under-specified contract producing a confident wrong patch — worse than the frontier model writing the code itself.
The decision stands, and this is what it now rests on. Grounds 2 and 4 are unchanged, and each is sufficient on its own: the reasoning budget is not the plugin's to set, and nothing was ever proven. Grounds 1 and 3 were re-measured after the refactor and both weakened — a module can now be read by a lead, but the reading still costs more than the typing it saves.
- A codebase that fits. Largely satisfied. The tier is only viable if a unit's relevant code fits
in roughly 20K tokens. That is a property of the repository, not the model:
index.tsaccumulated because the architect had been editing it directly. Split it and the arithmetic changes. It has now been split — every module is under ~13,400 tokens, and the largest is ~7,300 — so this is the one precondition that has been met since the retirement was recorded. - A way to control reasoning where the model is served, or a separate non-reasoning model for the worker so the two tiers do not contend for one server setting.
- Evidence that a local model writes a discriminating contract. The shape is now known and written down; the question is whether a 27B can fill it well enough to be worth a tier.
leadTier / leadProviders, PROFILES.LEAD, delegateReadPolicy, rule 8 and the durable registry are
all still built, tested and available. An operator may still run a local thinking agent alongside the
architect — the plugin will not repin it to the cloud, and will not tell it that it is the architect.