Problem / motivation
In multi-turn agent sessions, costs come not only from generation but also from repeatedly carrying prior content into later requests. It would be useful to attribute these costs to specific context sources and content blocks.
1. Cost Attribution by Context Source
Break down tokens and costs by source:
- Harness / agent runtime
- Prompts: system, schema, skills, user, and runtime injections
- Shell commands: command, arguments, and output
- Drill down by command and subcommand, such as
git, git diff, and git log
- Tool calls: arguments, results, reads, and writes
- Bidirectional tools and agent-to-agent interactions
- Attachments: images, documents, code, and other inputs
- Reminders and automation context
- Assistant-generated content
2. Generation vs. Repeated-Carry Cost
For each content block, distinguish between:
- Its initial generation cost
- Its cost when later used as input
- How many subsequent requests carried it
- Its cumulative repeated-carry cost
- Its total lifetime cost
This would reveal content that was inexpensive to generate but became costly because it remained in the context for many turns.
3. Content-Level Drill-Down and Traceability
Allow users to open a cost item and inspect:
- The original content and its source
- Where it first appeared
- Its token size
- Which later requests carried it
- Generation, repeated-carry, and lifetime costs
- Whether it is still present in the current context
A stable identifier for each context block could help track the same content across requests.
4. Visualization
Possible views include:
- A sortable table for exact attribution
- Pie or donut charts for cost proportions
- Petal or sunburst charts for hierarchical breakdowns
- Timelines or stacked charts for context growth across turns
- Cumulative curves for repeated-carry costs
- Sankey diagrams showing content flow from generation to reuse
Problem / motivation
In multi-turn agent sessions, costs come not only from generation but also from repeatedly carrying prior content into later requests. It would be useful to attribute these costs to specific context sources and content blocks.
1. Cost Attribution by Context Source
Break down tokens and costs by source:
git,git diff, andgit log2. Generation vs. Repeated-Carry Cost
For each content block, distinguish between:
This would reveal content that was inexpensive to generate but became costly because it remained in the context for many turns.
3. Content-Level Drill-Down and Traceability
Allow users to open a cost item and inspect:
A stable identifier for each context block could help track the same content across requests.
4. Visualization
Possible views include: