ace-pm b29c12d5e5 refactor(native): rename gsd_parser.rs to forge_parser.rs

Final rebrand: rename remaining Rust source file to complete the gsd → forge
transition. All parser references already use forge_parser after earlier commits.

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

2026-04-15 14:58:21 +02:00

3.1 KiB

Raw Blame History

Dynamic Model Routing

Dynamic model routing automatically selects cheaper models for simple work and reserves expensive models for complex tasks. This reduces cost by 20-50% without sacrificing quality where it matters.

Enabling

dynamic_routing:
  enabled: true

How It Works

Each unit passes through two stages:

Complexity classification — classifies work as light, standard, or heavy
Capability scoring — within the tier, ranks models by how well they match the task

Key rule: Your configured model is always the ceiling — routing never upgrades beyond what you've set.

Tier	Typical Work	Model Level
Light	Slice completion, UAT, hooks	Haiku-class
Standard	Research, planning, execution	Sonnet-class
Heavy	Replanning, roadmap reassessment	Opus-class

Configuration

dynamic_routing:
  enabled: true
  tier_models:                    # optional: explicit model per tier
    light: claude-haiku-4-5
    standard: claude-sonnet-4-6
    heavy: claude-opus-4-6
  escalate_on_failure: true       # bump tier on failure (default)
  budget_pressure: true           # auto-downgrade near budget ceiling (default)
  cross_provider: true            # consider models from other providers (default)
  capability_routing: true        # score models by task fit (default)

Escalate on Failure

When a task fails at a given tier, the router escalates to the next tier on retry: Light → Standard → Heavy. This prevents cheap models from burning retries on work that needs more reasoning.

Budget Pressure

When approaching the budget ceiling, the router progressively downgrades:

Budget Used	Effect
< 50%	No adjustment
50-75%	Standard → Light
75-90%	More aggressive
> 90%	Nearly everything → Light

Cross-Provider

When enabled, the router may select models from providers other than your primary, using the built-in cost table to find the cheapest model at each tier.

Capability Routing

Models are scored across 7 dimensions: coding, debugging, research, reasoning, speed, long context handling, and instruction following. Different task types weight these dimensions differently — a research task prioritizes research and reasoning, while an execution task prioritizes coding and instruction following.

Set capability_routing: false to revert to simple cheapest-in-tier selection.

Interaction with Token Profiles

Dynamic routing and token profiles work together:

Token profiles control phase skipping and context compression
Dynamic routing controls per-unit model selection

The budget profile + dynamic routing provides maximum cost savings.

Adaptive Learning

SF tracks routing outcomes in .sf/routing-history.json. If a tier's failure rate exceeds 20% for a given task type, future classifications are bumped up.

Use /sf rate to submit feedback:

/sf rate over    # too powerful — use cheaper next time
/sf rate ok      # just right
/sf rate under   # too weak — use stronger next time

Feedback is weighted 2x compared to automatic outcomes.

3.1 KiB Raw Blame History