Skip to content

Latest commit

 

History

History
370 lines (262 loc) · 14.1 KB

File metadata and controls

370 lines (262 loc) · 14.1 KB

Token Optimization

Introduced in v2.17.0

GSD 2.17 introduces a coordinated token optimization system that can reduce token usage by 40-60% without sacrificing output quality for most workloads. The system has three pillars: token profiles, context compression, and complexity-based task routing.

Token Profiles

A token profile is a single preference that coordinates model selection, phase skipping, and context compression level. Set it in your preferences:

---
version: 1
token_profile: balanced
---

Three profiles are available:

budget — Maximum Savings (40-60% reduction)

Optimized for cost-sensitive workflows. Uses cheaper models, skips optional phases, and compresses dispatch context to the minimum needed.

Dimension Setting
Planning model Sonnet
Execution model Sonnet
Simple task model Haiku
Completion model Haiku
Subagent model Haiku
Milestone research Skipped
Slice research Skipped
Roadmap reassessment Skipped
Context inline level Minimal — drops decisions, requirements, extra templates

Best for: prototyping, small projects, well-understood codebases, cost-conscious iteration.

balanced — Smart Defaults (default)

The default profile. Keeps the important phases, skips the ones with diminishing returns for most projects, and uses standard context compression.

Dimension Setting
Planning model User's default
Execution model User's default
Simple task model User's default
Completion model User's default
Subagent model Sonnet
Milestone research Runs
Slice research Skipped
Roadmap reassessment Runs
Context inline level Standard — includes key context, drops low-signal extras

Best for: most projects, day-to-day development.

quality — Full Context (no compression)

Every phase runs. Every context artifact is inlined. No shortcuts.

Dimension Setting
All models User's configured defaults
All phases Run
Context inline level Full — everything inlined

Best for: complex architectures, greenfield projects requiring deep research, critical production work.

Context Compression

Each token profile maps to an inline level that controls how much context is pre-loaded into dispatch prompts:

Profile Inline Level What's Included
budget minimal Task plan, essential prior summaries (truncated). Drops decisions register, requirements, UAT template, secrets manifest.
balanced standard Task plan, prior summaries, slice plan, roadmap excerpt. Drops some supplementary templates.
quality full Everything — all plans, summaries, decisions, requirements, templates, and root files.

How Compression Works

Dispatch prompt builders accept an inlineLevel parameter. At each level, specific artifacts are gated:

Minimal level reductions:

  • buildExecuteTaskPrompt — drops the decisions template, truncates prior summaries to the most recent one
  • buildPlanMilestonePrompt — drops PROJECT.md, REQUIREMENTS.md, decisions, and supplementary templates like secrets-manifest
  • buildCompleteSlicePrompt — drops requirements and UAT template inlining
  • buildCompleteMilestonePrompt — drops root GSD file inlining
  • buildReassessRoadmapPrompt — drops project, requirements, and decisions files

These are cumulative — standard drops a subset, minimal drops more. The full level preserves all context (the pre-2.17 behavior).

Overriding Inline Level

The inline level is derived from your token_profile. To control phases independently of the profile, use the phases preference:

---
version: 1
token_profile: budget
phases:
  skip_research: false    # override: run research even on budget
---

Explicit phases settings always override the profile defaults.

Complexity-Based Task Routing

GSD classifies each task by complexity and routes it to an appropriate model tier when dynamic routing is enabled. Simple documentation fixes use cheaper models while complex architectural work gets the reasoning power it needs.

Prerequisite: Dynamic routing requires explicit models in your preferences. Without a models section, routing is skipped and the session's launch model is used for all phases. Token profiles set models automatically.

Ceiling behavior: When dynamic routing is active, the model configured for each phase acts as a ceiling, not a fixed assignment. The router may downgrade to a cheaper model for simpler tasks but never upgrades beyond the configured model.

How Classification Works

Tasks are classified by analyzing the task plan:

Signal Simple Standard Complex
Step count ≤ 3 4-7 ≥ 8
File count ≤ 3 4-7 ≥ 8
Description length < 500 chars 500-2000 > 2000 chars
Code blocks ≥ 5
Signal words None Any present

Signal words that prevent simple classification: research, investigate, refactor, migrate, integrate, complex, architect, redesign, security, performance, concurrent, parallel, distributed, backward compat, migration, architecture, concurrency, compatibility.

Empty or malformed plans default to standard (conservative).

Unit Type Defaults

Non-task units have built-in tier assignments:

Unit Type Default Tier
complete-slice, run-uat Light
research-*, plan-*, execute-task, complete-milestone Standard
replan-slice, reassess-roadmap Heavy
hook/* Light

Model Routing

Each tier maps to a model configuration:

Tier Model Phase Key Typical Model
Light completion Haiku (budget) / user default
Standard execution Sonnet / user default
Heavy execution Opus / user default

Simple tasks use the execution_simple model key when configured. This is set automatically by the budget profile to Haiku.

Budget Pressure

When approaching your budget ceiling, the classifier automatically downgrades tiers:

Budget Used Effect
< 50% No adjustment
50-75% Standard → Light
75-90% Standard → Light
> 90% Everything except Heavy → Light; Heavy → Standard

This graduated approach preserves model quality for the most complex work while progressively reducing cost as the ceiling approaches.

Adaptive Learning (Routing History)

GSD tracks the success and failure of each tier assignment over time and adjusts future classifications accordingly. This is opt-in — it happens automatically and persists in .gsd/routing-history.json.

How It Works

  1. After each unit completes, the outcome (success/failure) is recorded against the unit type and tier used
  2. Outcomes are tracked per-pattern (e.g., execute-task, execute-task:docs) with a rolling window of the last 50 entries
  3. If a tier's failure rate exceeds 20% for a given pattern, future classifications for that pattern are bumped up one tier
  4. The system also accepts tag-specific patterns (e.g., execute-task:test vs execute-task:frontend) for more granular routing

User Feedback

Use /gsd rate to submit feedback on the last completed unit's model tier:

/gsd rate over    # model was overpowered — encourage cheaper next time
/gsd rate ok      # model was appropriate — no adjustment
/gsd rate under   # model was too weak — encourage stronger next time

Feedback signals are weighted 2× compared to automatic outcomes. Requires dynamic routing to be active (the last unit must have tier data).

Data Management

# Routing history is stored per-project
.gsd/routing-history.json

# Clear history to reset adaptive learning
# (happens via the routing-history module API)

The feedback array is capped at 200 entries. Per-pattern outcome counts use a rolling window of 50 to prevent stale data from dominating.

Configuration Examples

Cost-Optimized Setup

---
version: 1
token_profile: budget
budget_ceiling: 25.00
models:
  execution_simple: claude-haiku-4-5-20250414
---

Balanced with Custom Models

---
version: 1
token_profile: balanced
models:
  planning:
    model: claude-opus-4-6
    fallbacks:
      - openrouter/z-ai/glm-5
  execution: claude-sonnet-4-6
---

Full Quality for Critical Work

---
version: 1
token_profile: quality
models:
  planning: claude-opus-4-6
  execution: claude-opus-4-6
---

Per-Phase Overrides

The token_profile sets defaults, but explicit preferences always win:

---
version: 1
token_profile: budget
phases:
  skip_research: false     # override: keep milestone research
models:
  planning: claude-opus-4-6  # override: use Opus for planning despite budget profile
---

How the Pieces Fit Together

PREFERENCES.md
  └─ token_profile: balanced
       ├─ resolveProfileDefaults() → model defaults + phase skip defaults
       ├─ resolveInlineLevel() → standard
       │    └─ prompt builders gate context inclusion by level
       ├─ classifyUnitComplexity() → routes to execution/execution_simple model
       │    ├─ task plan analysis (steps, files, signals)
       │    ├─ unit type defaults
       │    ├─ budget pressure adjustment
       │    ├─ adaptive learning from routing-history.json
       │    └─ capability scoring (when capability_routing: true)
       │         └─ 7-dimension model profiles × task requirement vectors
       └─ context_management
            ├─ observation masking (before_provider_request hook)
            ├─ tool result truncation (tool_result_max_chars)
            └─ phase handoff anchors (injected into prompt builders)

The profile is resolved once and flows through the entire dispatch pipeline. Explicit preferences override profile defaults at every layer.

Observation Masking

Introduced in v2.59.0

During auto-mode sessions, tool results accumulate in the conversation history and consume context window space. Observation masking replaces tool result content older than N user turns with a lightweight placeholder before each LLM call. This reduces token usage with zero LLM overhead — no summarization calls, no latency.

Masking is enabled by default during auto-mode. Configure via preferences:

context_management:
  observation_masking: true     # default: true (set false to disable)
  observation_mask_turns: 8     # keep results from last 8 user turns (range: 1-50)
  tool_result_max_chars: 800    # truncate individual tool results beyond this length

How It Works

  1. Before each provider request, the before_provider_request hook inspects the messages array
  2. Tool results (toolResult, bashExecution) older than the configured turn threshold are replaced with [result masked — within summarized history]
  3. Recent tool results (within the keep window) are preserved in full
  4. All assistant and user messages are always preserved — only tool result content is masked

This pairs with the existing compaction system: masking reduces context pressure between compactions, and compaction handles the full context reset when the window fills.

Tool Result Truncation

Individual tool results that exceed tool_result_max_chars (default: 800) are truncated with a …[truncated] marker. This prevents a single large tool output from dominating the context window.

Phase Handoff Anchors

Introduced in v2.59.0

When auto-mode transitions between phases (research → planning → execution), structured JSON anchors are written to .gsd/milestones/<mid>/anchors/<phase>.json. Downstream prompt builders inject these anchors so the next phase inherits intent, decisions, blockers, and next steps without re-inferring from artifact files.

This reduces context drift — the 65% of enterprise agent failures caused by agents losing track of prior decisions across phase boundaries.

Anchors are written automatically after successful completion of research-milestone, research-slice, plan-milestone, and plan-slice units. No configuration needed.

Prompt Compression

Introduced in v2.29.0

GSD can apply deterministic prompt compression before falling back to section-boundary truncation. This preserves more information when context exceeds the budget.

Compression Strategy

Set via preferences:

---
version: 1
compression_strategy: compress
---

Two strategies are available:

Strategy Behavior Default For
truncate Drop entire sections at boundaries (pre-v2.29 behavior) quality profile
compress Apply heuristic text compression first, then truncate if still over budget budget and balanced profiles

Compression removes redundant whitespace, abbreviates verbose phrases, deduplicates repeated content, and removes low-information boilerplate — all deterministically with no LLM calls.

Context Selection

Controls how files are inlined into prompts:

---
version: 1
context_selection: smart
---
Mode Behavior Default For
full Inline entire files balanced and quality profiles
smart Use TF-IDF semantic chunking for large files (>3KB), including only relevant portions budget profile

Structured Data Compression

At budget and balanced inline levels, decisions and requirements are formatted in a compact notation that saves 30-50% tokens compared to full markdown tables.

Summary Distillation

When a slice has 3+ dependency summaries and the total exceeds the summary budget, GSD extracts essential structured data (provides, requires, key_files, key_decisions) and drops verbose prose sections before falling back to section-boundary truncation.

Cache Hit Rate Tracking

The metrics ledger now tracks cacheHitRate per unit (percentage of input tokens served from cache) and provides aggregateCacheHitRate() for session-wide cache performance.