Complexity weights and model choice

Downshift by Tiago Vilas Boas. Updated . Source on GitHub

How Downshift scores a task and turns that score into a model. The weights are harness-agnostic. The harness only chooses which model id represents the tier, and whether the rewrite is applied.

Source of truth: internal/classify/signals.go, internal/classify/classifier.go, internal/classify/brief_task.go, internal/core/policy.go, internal/classify/risk.go, internal/core/models.go. Session lists: session-models.md. What each harness does with the rewrite: harness-matrix.md.

Scoring

Every matching pattern adds its weight to one class: TRIVIAL, SIMPLE, MEDIUM, or COMPLEX. The class with the highest sum wins. A tie goes to the higher class, so the task is not under-powered.

Confidence is a margin, not a fifth class. The winner is confident only when it leads the runner-up by 2 or more. Below that, a downshift or a weak upshift is not applied. The spawn stays on the model it already had.

No pattern at all scores MEDIUM, and that result is not confident.

Extra points, applied after the regex table:

Condition Points
No regex matched, the prompt is add Name( with one call and at most 25 words +2 SIMPLE
The prompt contains implement and at least three exported calls Name( +2 COMPLEX
Some signal already matched and the prompt has at most 6 words +1 TRIVIAL
Some signal already matched, the prompt has at least 80 words and is not a bounded brief +1 COMPLEX
Bounded brief: at least 80 words and two of three limits (scope limit such as read-only or edit only, a test command such as go test, a reply budget such as reply <= 150 words) TRIVIAL votes dropped; a single COMPLEX row (other than review or retry wording) dropped; +2 MEDIUM
Long brief (80+ words) that is not bounded A TRIVIAL keyword does not vote when it sits inside a backtick command quote, is negated (do not git commit, never push), is a git branch label, or is a path segment such as mech-rename
Any prompt: a negated scope word A COMPLEX match on architecture, re-architect, redesign, rewrite, migrate or entire codebase/system does not vote when its own clause negates it: do not touch the architecture, don't redesign, without migrating, não mexa na arquitetura, sem migrar. The negation may be followed only by a change verb (touch, change, modify, edit, alter, mexer, mudar, alterar, tocar) and at most three filler words. A phenomenon (race condition, deadlock, leak) is never negated: make sure it does not deadlock still asks for hard work
Long brief (80+ words) that is not bounded and states its task, or a shorter brief with an explicit Task: marker Task focus, below

Task focus (briefs with a stated task)

A long parent brief wraps a short task in context: what the project is, what was tried, what must not change. Context words such as architecture or consensus describe the project, not the work, yet each scores COMPLEX 3. On real use (2026-10-19) this kept mechanical subagent work unconfident, so it stayed on the expensive model. Task focus scores the task with the rest of the brief weighed as context (internal/classify/brief_task.go).

The task is:

  1. the section after an explicit Task: / Your task: / Tarefa: marker (a heading or bold form works too) at the start of a line or of a sentence (… none of that is in scope here. Task: rename …). It runs to the end of the brief or to the next context section (Background:, Context:, Notes:, History:, Reply: and Portuguese forms). Requirements, constraints and "done means" sections stay in the task, because they shape the work. The marker is the author saying the rest is context, so it counts on a brief of any length.
  2. otherwise, on a long brief (80+ words) only, every sentence that opens with a work verb (rename, fix, add, migrate, implement, review, … and Portuguese imperatives). Reading, running checks, replying and negated sentences (Do not …) are not the task. A shorter prompt's other sentences may be the work, so it gets no imperative extraction.
  3. otherwise, nothing: the brief is scored as before.

The focused tally, used only when the task itself matched a signal:

Signal Inside the task Outside the task
TRIVIAL, SIMPLE full weight does not vote (side steps and constraints)
MEDIUM full weight full weight
COMPLEX, review or retry wording full weight full weight
COMPLEX, any other row full weight 1 per row; under an explicit Task: section all of them together count 1
Length nudge (+1 COMPLEX) applies when the task itself has 80+ words, or when the task was found by imperative sentences (the other sentences may describe the work)

The focused result replaces the whole-brief result only in the safe direction. It never raises the class, and never makes a COMPLEX result confident. An unconfident lower class does not lower anything: it only clears confidence on the whole-brief class, so the spawn keeps its model. The risk floor still reads the whole brief, context included. A rename under a Context: that mentions JWTs or refunds stays MEDIUM and unconfident. Bounded briefs keep their own rule above. A decision shaped by the focus carries the reason code TASK_FOCUS.

Examples (80+ word briefs unless the word count says otherwise):

Brief Before After
Background names the architecture and a consensus protocol; Task: rename cnt to count in summary.go COMPLEX, confident TRIVIAL, confident (downshift applied)
Same background; Task: add a verbose flag to the stats command COMPLEX, confident COMPLEX, unconfident (held)
Background names JWTs and refresh tokens; Task: rename fmtTok to formatToken COMPLEX, unconfident MEDIUM, unconfident (risk floor; no write below the requested model)
Task: rewrite the replication layer and migrate the slots COMPLEX, confident unchanged
No marker; one sentence names the architecture, then Rename the field amount to amountCents … COMPLEX, unconfident unchanged (imperative extraction keeps the nudge: 3 against 2)
65 words on one line: background names the architecture, a redesign, the migration plan and concurrency; … in scope here. Task: rename the variable totalAmt to totalAmount … COMPLEX, confident TRIVIAL, confident (downshift applied)
40 words: context names a distributed design and a rewrite; Task: fix the typo 'occured' to 'occurred' in docs/ingest.md COMPLEX, confident TRIVIAL, confident (downshift applied)

Measured on 2026-10-19 against every labelled set in the maintainers' eval repository (the five frozen benchmark splits, M3 multilingual and two development sets, 1074 items): no item changed class or confidence. None of them is a long brief with a stated task. Re-measured on 2026-10-19 after the marker started counting on short briefs and at a sentence start, together with the widened typo row (fix the typo, fix typos): the only item that changed is fix the typo in the README, labelled TRIVIAL, which moved from MEDIUM unconfident to TRIVIAL confident. The gain is pinned by the table-driven tests in internal/classify/classifier_edge_test.go and internal/core/task_focus_test.go.

A TRIVIAL or SIMPLE task that touches risk (secret, auth, crypto, payment, vulnerability, PII, destructive command, prompt injection, or concurrency) is lifted to MEDIUM. MEDIUM and COMPLEX are not lifted further by risk. The risk table is internal/classify/risk.go. It sets a floor. It does not add to the vote.

Input cap. The scoring regex table and the structural scoring rules read at most the first 32 KB of the prompt (cut on a rune boundary), so cost stays flat for the 1 MiB the hook accepts. A signal past byte 32 768 is ignored by design. Risk and language checks inspect the full input, so a risky or unsupported-language tail cannot bypass them. The word-count rules (6 and 80 words, the add Name( stub, the doc-edit damper) still count words over the full prompt.

Task focus and negated scope words only remove votes; they are part of the regex scoring, not a later step. An optional later step can only raise the class: the in-process word-hash semantic boost. It never lowers a class, is ignored for no-signal prompts, and cannot create COMPLEX without regex evidence. Turn it off with DOWNSHIFT_MINILM=0.

Language coverage

English and Portuguese are covered by the signal and risk rules. Portuguese risk terms include permission, privilege, security, injection and personal-data terms; the risk floor is shared by every harness.

The language guard (internal/core/lang_guard.go) detects unsupported-language indications before a confident routing change can be applied. It looks for distinct Spanish, French or German function words, selected foreign Latin letters, or at least four non-Latin letters. Mixed English helper briefs (fix|rename|format|move|add this helper; plus foreign prose) set UnsupportedLanguage so a confident trivial downshift cannot slip through (M3 / H14). Other mixed-language instructions can trigger the same guard even when an English verb matches a cheap-task signal. Detection drops confidence without changing the scored class; uncertain changes preserve the requested model. Existing deterministic safety floors can still raise an undersized model. A prompt with no matching signal also abstains.

Risk abstention (H14): when risk categories match and the class stays Medium (or Simple), applyRiskFloor may set RiskAbstain and clear confidence without changing complexity: long bounded delegation briefs (scope + verify cues such as “stay within these modules” and “run the relevant tests”), and narrow high-stakes phrases from the M3 parallel bucket (for example idempotent payment/refund wording or “Revise as permissões…”). RiskAbstain counts as a deterministic lift for guardrails R7/R9. Generic short “implement … auth” tasks stay confident Medium when the pattern does not match.

This is a conservative heuristic, not a complete language detector. Short foreign phrases may escape detection, while identifiers or quoted text may trigger it. Scoring inspects a rune-aligned prefix of at most 32 KB; risk and language checks inspect the full input. Do not interpret English and Portuguese coverage as calibrated accuracy or support for all multilingual prompts.

English and Portuguese coverage is the shipped language contract. Multilingual E5 and external MiniLM were removed on 2026-10-17; they do not extend coverage. The archived initiative note is multilingual semantic intelligence.

Weights

Weight is the integer added when the pattern matches. Patterns are regular expressions, case-insensitive. Several matches in the same class add up.

COMPLEX

Weight Matches
3 architect, re-architect, migrate, redesign, security audit/review, code review, cross-system, etcd/raft/consensus, race condition, deadlock, multi-system/service/region/tenant, entire codebase/system/platform, design doc, RFC, service mesh, event sourcing, CQRS, chaos engineering, end-to-end encryption, data lineage, cross-cluster, memory leak, cascading failure, streaming pipeline, model serving, key management, recursive-descent, longest common subsequence, operator precedence, postfix operator, backtracking, cron, fn func(, previous attempt / tentativa anterior, compilation failed, subagent failed / tests failed, fix the build
2 rewrite, distributed, trade-off, think through, why … fail/crash/hang/break/regress, directed graph, binary heap, concurrent/goroutine, a fixed-size grid such as [9][9]

MEDIUM

Weight Matches
3 zero-downtime / online schema change
2 refactor, implement, integrate, feature, across N files, debug, revisar/revisão
1 endpoint, module

SIMPLE

Weight Matches
3 rotating cube, three.js scene, hello world, landing page, one file, one function, one component
2 add field/param/flag/property/column/argument/method/function, fix the bug, write a function/test/helper, explain (only at the start), single file/function
1 what is / what does / what are

TRIVIAL

Weight Matches
3 rename, renomear, format, indent, prettier, eslint, git commit/push/pull/status/add/stash/log/diff/checkout/branch, fix a/the/some typo(s), add a comment, move the file, delete the file, bump the version, listar arquivos
2 lint, boilerplate, remove unused/dead

From class to tier

The tier does not depend on the harness.

Class Tier Effort
TRIVIAL small low
SIMPLE small low
MEDIUM mid mid
COMPLEX frontier high
anything else mid mid

Effort is written only on harnesses that accept it (Codex reasoning_effort, Grok config). It is the same mapping everywhere.

How the harness picks the model

The weights stop at the tier. The model id comes from the operator's session list for that harness, ordered from least to most capable (~/.downshift/session-models.json). Downshift picks the listed model whose catalog tier matches the decision. It does not invent an id that is not on the list, so an expensive frontier model stays out until it is appended.

What changes per harness is the wire format and whether the executor honors the rewrite, not the score:

Harness What the hook can do
Claude Code Writes the family name (haiku, sonnet, opus, fable) into updatedInput.model. A full id is rejected.
Codex Writes the model id and reasoning_effort. A held decision writes neither; an unconfident one changes effort only together with a model rewrite.
Cursor Writes updated_input.model. Free and legacy Pro plans discard it. Usage-based plans are not yet proven.
Antigravity Overwrites the subagent model. Observed on desktop 2.21.1 only.
KiroCrew No rewrite channel. Exit 0 allows, exit 2 blocks and asks for the cheaper model.
Grok CLI No hook rewrite. Effort is pinned in config.toml.

Local tier memory (internal/adapt) runs after this score. It does not add weight. It can move the tier only when a reviewed outcome already recorded a hit or a miss for the same task shape. No memory file means the table above is the whole decision.

More docs