Context Resume
What it is
When you start a new coding session, the agent has no memory of what happened in previous sessions. The context resume feature solves this by injecting a structured summary of past conversations — what you asked, what the agent did, which files it touched — directly into the new session's context window.
This lets you continue work naturally. The agent knows the history without you having to re-explain the project, the decisions already made, or the state of in-progress tasks.
How it works
The ContextManager maintains a list of registered context providers. When a session resumes, OpenACP queries the appropriate provider for the relevant history, receives a formatted markdown document, and prepends it to the first prompt.
Providers are pluggable. The current built-in provider is EntireProvider, which reads conversation history stored by the Entire.io git checkpoint system.
Results are cached on disk at ~/.openacp/cache/entire/ so that repeated requests for the same query (same branch, same commit range) do not re-read and re-parse transcript files.
Entire.io provider
Entire.io saves git checkpoints during agent sessions. Each checkpoint stores:
The agent name and git branch
A JSONL transcript of the conversation (user messages and assistant responses, including file edits and writes)
The list of files touched
The EntireProvider is available when the repository contains an entire git branch. It supports several query types for resolving which sessions to include:
branch
main
All sessions on a specific branch
commit
abc12345
Sessions associated with a specific commit
pr
https://github.com/.../pull/42
Sessions for a pull request
checkpoint
chk_abc123
A single specific checkpoint
session
sess_abc123
A single session by ID
latest
5
The N most recent sessions (default 5)
Adaptive rendering modes
Conversation history can be long. To avoid consuming the entire context window, the provider automatically selects a rendering mode based on total turn count:
full
up to 10
Full text of all user messages and assistant responses; full diffs and file writes
balanced
11 to 25
User messages in full; diffs truncated to 12 lines; file writes truncated to 15 lines
compact
more than 25
User messages in full; edits shown as a one-line summary with line counts; writes shown as filename and line count only
If the resulting markdown still exceeds the configured maxTokens limit, the provider automatically downgrades to compact mode. If it is still too large, the oldest sessions are dropped until it fits.
A disclaimer is always appended to the context document reminding the agent that the history may be outdated and should not be taken as ground truth for current file contents.
/resume command
Send /resume in any session topic to attach context from recent sessions to your next prompt. The command accepts the same query types as the provider:
The injected context appears as a collapsible block above your message so you can verify what was included before the agent processes it.
Token estimation
Token counts are estimated as floor(characters / 4), which approximates GPT-style tokenization. This estimate is used both to select the rendering mode and to decide whether to truncate. The estimate is intentionally conservative — actual token counts vary by model and content type.
To see how many tokens a context block would consume before sending, use:
This reports the session count, turn count, rendering mode, and estimated token count without attaching the context.
Last updated
Was this helpful?