Agentify · Personalize models, context, and opt-in features

01 · starter

Change the policy, not every default.

Agentify writes .agentify.yaml during install. Configuration is deep-merged, so the useful mental model is: keep the generated file, edit the few values your evidence supports.

focused example

This example chooses the balanced profile, adds rolling spend caps, keeps deterministic summaries, and leaves pre-push review off until you opt in.

models:
  profile: balanced
  budget:
    dailyUsd: 5
    monthlyUsd: 50
    onLimit: block
  routes:
    review:
      provider: codex
      model: null
      maxBudgetUsd: 0.75

context:
  injection: relevant
  sessionSummaries: extractive

hooks:
  prePush: false

null model

model: null lets the Codex CLI use its configured default. Pin a full model ID only when reproducibility matters more than automatic model updates.

hard limits

Per-route ceilings remain in force across fallbacks. Rolling daily and monthly caps use locally recorded spend; onLimit: block fails before starting another provider process.

02 · models

Pick an objective, then inspect the route.

Profiles select within hard ceilings and local eval evidence. They never rewrite your config, and a higher price is not treated as proof of higher quality.

cost

Spend less with a quality floor

Chooses the cheapest evaluated route that meets the configured floor. Without sufficient evidence, it does not silently downgrade.

balanced · default

Optimize cost per passing task

Uses the lowest measured cost per pass. With no eval evidence, it behaves like the explicit route table.

performance

Prefer measured pass rate

Escalates only when local evaluations show a gain, while staying inside the same budget and tier bounds.

one-off override

Keep experiments temporary

--profile, --provider, and --model let you test a route without changing the repo policy.

agentify delegate quick "summarize this diff" --provider codex --dry-run

Inspect the resolved profile, tier, limits, fallback chain, evidence, and alias-drift warning before any paid run.

agentify route explain "fix the checkout race" --profile balanced Usage guide

03 · routes

Route by work, not by habit.

Every default route has a dollar, turn, and wall-clock ceiling. Override a route only when your own evals or operational needs justify it.

quick · Claude Haiku

Small, low-impact work

Mechanical edits and fast questions. Default ceiling: $0.10, 4 turns, 120 seconds.

implement · Claude Sonnet

Normal feature work

Multi-file implementation and refactors. Default ceiling: $1.00, 30 turns, 600 seconds.

heavy · Claude Opus

Architecture and deep debugging

High-risk reasoning where more capability is justified. Default ceiling: $2.50, 40 turns, 600 seconds.

review · Codex default

Independent second opinion

Uses a different vendor by default. Default ceiling: $0.75, 20 turns, 600 seconds.

# override only one route in .agentify.yaml
models:
  routes:
    quick:
      provider: claude
      model: haiku
      maxBudgetUsd: 0.10
      maxTurns: 4
      timeoutSeconds: 120
    review:
      provider: codex
      model: null
      maxBudgetUsd: 0.75
      maxTurns: 20
      timeoutSeconds: 600

04 · opt in

Features that stay off until you choose them.

These switches add cost, external tooling, or broader automation. Enable them one at a time and verify the resolved behavior.

cross-vendor review

Review before every push

Advisory findings print without blocking the push. Enable the setting, then install the managed hook.

hooks:
  prePush: true

agentify hooks install

budgeted LLM handoff

Refine larger summaries

The quick route receives only the extractive summary, stays under its own cap, and falls back to deterministic text.

context:
  sessionSummaries: llm
  summary:
    llmMinEvents: 20
    maxBudgetUsd: 0.03

opt-in providers

Add providers after evaluation

Gemini and OpenCode stay out of default routes and fallbacks until the repo explicitly enables them.

models:
  providers:
    gemini:
      enabled: true
    opencode:
      enabled: true

context budget · advanced

Pin only with evidence

Keep the default relevant-injection policy unless an eval shows a better token budget for this repo.

context:
  injection: relevant
  maxInjectedTokens: 1200
  reserve:
    decisions: 250
    failures: 250

external search · optional

Enable Zoekt only when needed

Agentify's built-in index is the default. Opt into the external search toolchain for repositories where it adds measurable value.

toolchain:
  zoekt: true

05 · verify

Know what will run before it runs.

Agentify exposes the resolved policy instead of asking you to infer it from YAML.

availability

Models and controls

Shows routes, installed provider CLIs, fallback targets, limits, and which ceilings each CLI can enforce natively.

agentify models

decision trace

Route explanation

Shows the selected profile and model, evidence source, budget, fallback chain, and alias warnings without invoking a provider.

agentify route explain "design the migration" --profile performance

delegation plan

Dry run

Resolves the exact command and sandbox policy while keeping the provider process stopped.

agentify delegate review --diff origin/main --dry-run

environment

Doctor

Checks local toolchain readiness after config changes and before relying on optional providers or search tools.

agentify doctor