"Claude and Codex Session Resume — Who Retransmits What?"

Is Session Resume Really Free?

When resuming a previous session with --continue or --resume, it is tempting to assume the server holds the conversation in memory and picks up where it left off at low cost. This assumption applies equally to Claude Code and OpenAI Codex CLI. The UX of a session "continuing" feels like server-side state persistence, but the underlying mechanics are different.

The short answer: neither system is server-stateful. Both tools replay locally stored conversation history and retransmit it to the server on each resume. The difference lies in the prompt cache policy operating on top of that replay, and in how stably the prefix is reconstructed at replay time. In Claude Code v2.1.116, a bug has been observed in which the prefix is not reassembled stably on resume.

Anthropic Prompt Cache Official Spec

Anthropic's prompt caching is opt-in. The default TTL is 5 minutes; a 1-hour TTL option is available at additional cost.

The pricing structure is as follows:

  • cache_write (5 min): 1.25× baseline input token cost
  • cache_write (1 hour): 2× baseline input token cost
  • cache_read: 0.1× baseline input token cost (90% discount)

High hit rates make this very economical, but the write cost is the key variable. cache_read is cheap, but cache_write at the 1-hour TTL costs twice the baseline input rate. For cache_write to be offset by subsequent cache_read savings, repeated requests must arrive with the same prefix within the TTL window.

Cache invalidation is hierarchical: tools → system → messages in order; if an upper block changes, everything below it is invalidated. The lookback window is 20 blocks, and when a cache entry is actually used, its TTL is automatically refreshed at no additional cost.

Source: Anthropic Prompt Caching Official Documentation

OpenAI Automatic Prompt Cache Spec

OpenAI's cache requires no opt-in. It is applied automatically to requests of 1,024 tokens or more, with a 50% discount on cache hits. The standard TTL is 5–10 minutes, and during off-peak hours it may be retained for up to 1 hour. No manual cache-point marking is required; the system operates via automatic prefix matching.

There are two key differences from Anthropic. First, there is no cache_write premium. No additional cost is incurred when a cache entry is constructed, so a low hit rate does not result in a loss. Second, the discount rate is 50%, lower than Anthropic's 90% (on cache_read). In summary, OpenAI offers a conservative discount with no risk, while Anthropic offers a larger discount but requires paying the write cost upfront.

Codex CLI의 resume 동작 — 통념 교정

Codex CLI Resume Behavior — Correcting a Common Misconception

A common belief holds that "Codex resumes because the server remembers the conversation." This misconception originates from the fact that OpenAI's Responses API supports server-side conversation state via store: true + previous_response_id.

However, direct inspection of the codex-rs open-source code reveals different behavior:

  • protocol.rs:2525 — The ResumedHistory { conversation_id, history: Vec<RolloutItem>, rollout_path } struct replays the entire local rollout and spawns it as history. There is no previous_response_id field.
  • thread_manager.rs:531 — The resume_thread_from_rollout function calls RolloutRecorder::get_rollout_history(path) to read a local jsonl file and reconstruct the full conversation.
  • client.rs:998 — previous_response_id is reused only within a WebSocket session inside the same process. When the process is restarted, the connection is severed.

In other words, Codex CLI does not leverage Responses API server state when resuming after a process restart. It operates on a local rollout replay + server prefix caching structure, fundamentally identical to Claude Code.

This structure has an important implication. The number of tokens transmitted on resume grows proportionally with the length of the prior conversation, and if the prefix fails to hit the server cache, the full retransmission cost is borne in full. The belief that "resume is free because the server remembers" does not hold for either system.

Evidence — TTL-matched Measurement (Claude Code v2.1.116)

A comparative measurement was conducted to assess the difference in cache hit rate when using --resume in Claude Code, under matched idle-gap conditions. Same-session consecutive turns (baseline) and cross-session resumed turns (test) were compared under the same idle gap to eliminate the TTL variable mathematically.

Sample: 14 jsonl files, 197 turns, 11 hours 52 minutes of session.

Condition 4-min gap 28-min gap
Same session (baseline) 97–99% hit 99% hit
Cross session (resume) 41.2% hit 0% hit
Δ −56 pp −99 pp

Under identical TTL conditions, the cache hit rate dropped dramatically when a session was resumed. Inspecting the cache types used internally by Claude Code reveals that only ephemeral_1h (1-hour cache) is in use; ephemeral_5m registers at 0. The fact that hit rates are low despite ample TTL headroom indicates this is a prefix reassembly issue, not a TTL issue.

The identified cause is the deferred_tools_delta and attachment reordering introduced in v2.1.69. On resume, the injection order of skill_listing, todo_reminders, and nested_memory changes, breaking the prefix hash. This analysis was confirmed via simpolism's gist.

Impact — Actual Quota Consumption

Figures observed in the session above (197 turns, 11 hours 52 minutes):

  • cache_creation_input_tokens / output_tokens ratio (cc/output ratio): 4.52× (normal range: 1–2×)
  • Within resume segments falling inside the 1-hour TTL window: 40,260 tokens of cache_creation occurred. This is cost that would have been avoided had the prefix been preserved.
  • In observed cases, approximately 20% of Claude Pro/Max subscribers' 5-hour rolling quota was consumed by resume-triggered cache regeneration.

cache_creation at the 1-hour TTL costs 2× the baseline input rate. Compared to cache_read at 0.1×, this is a 20× more expensive path. When the cache fails to hit and must be reconstructed on every turn, caching — which was expected to reduce costs — ends up increasing them.

Response

GitHub Issue #51764 (reproduced on v2.1.116) has been filed, continuing the tracking thread from the previously closed #42338. Some code paths patched in v2.1.90–92 may be covered, but reproduction has been confirmed in v2.1.116 on the code path combining custom agents + MCP + skills + hooks.

How to verify directly: Aggregate cache_read_input_tokens and cache_creation_input_tokens per turn from ~/.claude/projects/<slug>/*.jsonl. If the cc/output ratio begins to exceed 2×, there is a high probability that prefix reassembly is being repeated. ArkNill's claude-code-cache-analysis script makes per-session aggregation straightforward. taekim34 documented a detailed reproduction path in the issue thread via an environment reproduction comment, covering the combination of a custom CLAUDE.md with a multi-stage hook setup.

Practical recommendation: In auto-restart watchdog patterns, consider switching from automatic claude --continue resume to a manual session start after stop. This pays only the CLAUDE.md reload cost and builds the prefix stably from scratch. A predictable cost structure is preferable to repeatedly replaying an unstable prefix.

정리

Summary

Summarizing the resume structure of the two systems:

  • Claude Code: local jsonl replay + server prefix caching (opt-in, hierarchical invalidation)
  • Codex CLI: local rollout replay + server prefix caching (automatic, no write premium)

There is no system in which the server maintains state and resumes "for free." Resume always carries the cost of retransmitting the entire conversation.

Cost structure difference: Anthropic offers a large discount at cache_read 0.1×, but charges a write premium of up to 2×. OpenAI offers a 50% discount with no write premium. Which is cheaper depends on the workload and TTL hit rate; measuring actual hit rates takes precedence over simple comparison.

The resume bug in Claude Code v2.1.116 is a real problem in which cache_write cost is incurred redundantly and unnecessarily. Even under conditions where TTL is sufficient, prefix reassembly failure causes cache_write to repeat instead of cache_read. The issue is currently tracked under Issue #51764, and the direct measurement method described above can be used to verify reproduction in your own environment.

References

  • Anthropic Prompt Caching documentation: https://docs.claude.com/en/docs/build-with-claude/prompt-caching
  • Codex source: https://github.com/openai/codex
  • Issue #51764: https://github.com/anthropics/claude-code/issues/51764
  • Issue #42338 (closed/locked): https://github.com/anthropics/claude-code/issues/42338
  • Issue #34629, #46829 (related)
  • ArkNill cache analysis: https://github.com/ArkNill/claude-code-cache-analysis
  • simpolism gist: https://gist.github.com/simpolism/302621e661f462f3e78684d96bf307ba
  • taekim34 environment reproduction comment: https://github.com/anthropics/claude-code/issues/42338#issuecomment-4174599756

댓글

댓글 쓰기

이 블로그의 인기 게시물

Agent Memory Engine (2/10) — Building an AI Agent Memory System with SQLite Alone

← 1/10 SQLite Memory Engine f… 📚 Series Index 3/10 memcore Completed → A memory engine that runs without daemons, without dependencies, anywhere 핵심 요약 This post covers how to design an AI agent memory engine (memcore) on a single-file SQLite backend. - Structure: 25 modules, ~6,464 lines, 22 DB tables - Principles: no daemon, single-file portability, no required external dependencies, host-agnostic - Core techniques: 4-layer hybrid search (topic · vector · FTS5 · LIKE), U-tag dialectic-based confidence evolution, ingestion-layer bias rejection rules What This Post Covers One approach to building AI agent memory without a server or cloud vector DB. Specifically: (1) the limitations of text-file-based memory, (2) vector DB alternatives compared and selection criteria, (3) how to layer search tiers, (4) how to evolve confidence scores over time, and (5) how to block biased memory accumulation at the ingestion stage. Design Principles: Why Single-File SQLite Text-file-based ...
Read more »

"ML Foundations (9/9) — PyTorch vs TensorFlow, and the Road to Local LLMs"

이미지
← 8/9 Deep Learning Architec… 📚 Series Index (series end) The final part. So far we've focused on models. Now we focus on the tools that actually run them . We'll lay out the philosophical difference between PyTorch and TensorFlow, and trace how Hugging Face transformers , llama.cpp , MLX , and Ollama built the bridge to running large language models on your own machine. By the end you should have the full mental model of "download a pretrained LLM and serve it locally." 0. Learning Objectives Explain the eager-vs-graph difference between PyTorch and TensorFlow. Explain, in graph terms, how autograd automates the backward pass. Use Hugging Face transformers ' AutoModel , AutoTokenizer , and pipeline abstractions. Describe llama.cpp's GGUF quantization, INT4 inference, and CPU-first flow. Describe Apple MLX's use of unified memory and how it differs from PyTorch. Run a local LLM with Ollama and call it through an OpenAI-compatible API. ...
Read more »

"RAG Core Study (14/26) — Evaluation Sets with RAGAS & DeepEval"

이미지
Series overview: Series index 📚 Series Index Without an evaluation set, every RAG improvement is a story, not evidence. RAG systems fail in more than one place: retrieval, context selection, answer generation, and grounding. That means "looks better" is not a metric. Part 14 explains how to build a golden dataset , how RAGAS and DeepEval fit into the workflow, and why teams must separate retrieval quality from answer quality if they want tuning decisions to hold up. 0. Prerequisites Part 13 reranking — system quality now depends on multi-stage ranking. Part 1 grounding — a correct-looking answer is not enough. Part 12 Hybrid — multiple retrieval settings need comparison on fixed data. 1. Learning Objectives Build a small but useful golden dataset for RAG. Distinguish retrieval metrics from answer metrics . Use RAGAS and DeepEval in the right roles. Avoid the common traps in judge-model-based evaluation. 2. 핵심 요약 An evaluation set for RAG is ...
Read more »

"ML Foundations (8/9) — Deep Learning Architectures: CNN, RNN, Attention"

이미지
← 7/9 Deep Learning Training 📚 Series Index 9/9 PyTorch vs TensorFlow → What did we keep adding on top of an MLP? Vision went CNN, sequences went RNN/LSTM, and eventually both converged on Attention and the Transformer. Each architecture is easier to remember if you read it as what it gave up to gain something else . This part is the one-line summary of the decisive inflection points: LeNet → AlexNet → ResNet → LSTM → Attention → Transformer. 0. Learning Objectives Explain why CNN's convolution, pooling, and weight sharing match images so well. Trace what got unblocked from LeNet → AlexNet → VGG → ResNet as depth grew. State BPTT for RNNs and the vanishing/exploding gradient problem. Write the LSTM gate equations (forget, input, output) with intuition. Write the attention formula in query/key/value form and the scaled dot-product variant. State the precise reasons Transformers replaced RNNs (parallelism + long-range dependencies). 1. 핵심 요약 CNN : weight sharin...
Read more »

"ML Foundations (7/9) — Deep Learning Training: Optimizers, Regularization, Initialization"

← 6/9 Neural Networks 📚 Series Index 8/9 Deep Learning Architec… → Building an MLP, as we did in Part 6, is not the same as getting it to train well . The right optimizer, regularization, initialization, and learning rate — when these line up, deep networks converge. When they don't, the network refuses to learn at all. This part is the catalog of those crafts, with formulas, paper citations, and code in one place. 0. Learning Objectives Compare and write the update rules for SGD, Momentum, Nesterov, and Adam. Explain how Dropout, BatchNorm, and LayerNorm work and where they belong in a model. Derive the variance formulas for Xavier and He initialization and match them to activations. Implement step, cosine, and warmup learning-rate schedules in PyTorch. Explain why gradient clipping is effectively required for RNNs and Transformers. Diagnose the most common training failures (NaN, plateau, overfitting) and apply first-line fixes. 1. 핵심 요약 SGD : \(w \leftarro...
Read more »

OpenClaw to Hermes Migration (2/13) — What to Preserve, Partially Port, or Discard

← 1/13 Current Structure Inve… 📚 Series Index 3/13 What Moves to Hermes → A code review record classifying operational scripts into three tiers: preserve, partial-preserve, and discard. What This Post Covers An asset classification framework (preserve / partial-preserve / discard) for migrating a legacy agent system to a successor A concrete case applying the Strangler Fig + Subset Migration strategy to a memory pipeline Classification results evaluated across LOC, external dependencies, and replaceability axes A method for slimming down a bloated single file (1,310 LOC) using the migration as a natural entry point Problem Definition: Port Everything or Cut Deep? The OpenClaw inventory comprises multiple agents, multiple scheduler/daemon processes, and multiple operational scripts. When moving to the successor system, Hermes, the choice is not binary. A full port carries accumulated technical debt along with it; a minimal port discards hard-won domain logic. Code-revie...
Read more »

AI Agents I Built (5/7) — Building an Automated Blogger API Publishing System

← 4/7 My Life Organizer Agent 📚 Series Index 6/7 Dual → Designing a Markdown-Based Auto-Publish Pipeline with Blogger API v3 + OAuth 2.0 Key Summary Blogger API v3 + OAuth 2.0 enables a pipeline that converts Markdown files into publishable HTML posts blogger_publish.py handles the full flow — frontmatter parsing → body cleanup → HTML conversion → CSS injection → API post — in a single script For bulk publishing, quota management (delay, retry, TOC suppression) is the critical design concern Platform Selection Criteria API access is mandatory for AI agent-driven content publishing. A comparison of platforms: Platform API Cost AdSense Decision WordPress.com REST API Paid plan required Paid only Cost overhead Ghost REST API Self-hosted or paid Manual setup Ops overhead Blogger v3 API Free Native integration Adopted Blogger is free, exposes a REST API, and integrates natively with AdSense. For personal blog automation, this combination i...
Read more »