Context budget: conversation history replay is expensive in long sessions

[JustAGhosT]

Problem

Long sessions involving debugging, code archaeology, implementation, and follow-up accumulate expensive conversation history. Claude Code compresses automatically near context limits, but by that point the session is already degraded:

• Responses slow down as context grows
• Earlier context gets compressed/lost at the worst time
• Cost per turn increases linearly with history

Proposed Solutions

1. Session segmentation guidance — retort docs should recommend session boundaries (e.g., "start a new session after completing a feature, not mid-debug"). The /handoff command already exists but isn't enforced.
2. Context usage indicator — retort healthcheck or a session hook could warn when estimated context usage exceeds 50% of the window ("Consider starting a fresh session").
3. Subagent offloading patterns — document when to use Agent tool to offload expensive research to subagents (isolated context) vs doing it inline. retort's agent team specs could include guidance on this.
4. History-aware /handoff — enhance handoff to produce a minimal continuation prompt that captures decisions + state without replaying the full conversation.

[blocksorg]

Mention Blocks like a regular teammate with your question or request:

@blocks can you add technical details to this issue?
@blocks make the following changes ...
@blocks create an issue from what was mentioned in the following comment ...
@blocks explain the following code ...
@blocks are there any security or performance concerns?

Run @blocks /help for more information.

Workspace settings | Disable this message

[Camj78]

This looks like it could be tied to how the prompt/context is being constructed under the hood.

I’ve seen similar issues where:
• system + user prompts get re-injected every call
• context/history grows silently between requests
• small prompt changes lead to big jumps in token usage

That combination can cause both weird behavior and unexpectedly high cost.

I ran into this in production and ended up building a small CLI to analyze prompts before they’re deployed — it flags token growth patterns and estimates cost impact.

You can test it locally:

npx @camj78/costguardai analyze prompt.txt

Might be useful to run this against the prompt chain here to see what’s expanding.