Documentation
¶
Overview ¶
Package compaction summarizes old conversation turns to reduce context size.
The core loop calls Compact when context approaches the model's input limit. Old turns are serialized, sent to the LLM for summarization, and replaced with a single compaction_summary message. Recent turns are kept verbatim.
Index ¶
- func AppendCheckpoint(result *Result, compacted []core.AgentMessage, checkpoint string)
- func FindCutPoint(msgs []core.AgentMessage, contextTokens, contextWindow int, ...) int
- func GenerateSummary(ctx context.Context, provider core.Provider, model core.Model, ...) (string, *core.Usage, error)
- func SerializeForSummary(msgs []core.AgentMessage, maxInput int) string
- type FileOps
- type Result
Constants ¶
This section is empty.
Variables ¶
This section is empty.
Functions ¶
func AppendCheckpoint ¶
func AppendCheckpoint(result *Result, compacted []core.AgentMessage, checkpoint string)
AppendCheckpoint mechanically appends a checkpoint to a summary and rewrites the summary message in place. Both the manual and automatic compaction paths must call this: previously only the manual path did, so an automatic compaction silently discarded a checkpoint the user had already paid for.
func FindCutPoint ¶
func FindCutPoint(msgs []core.AgentMessage, contextTokens, contextWindow int, settings core.CompactionSettings) int
FindCutPoint returns the index of the first message to KEEP (everything before it gets summarized). Returns 0 if nothing needs cutting.
The cut targets contextWindow - reserveTokens - summary overhead, ensuring the result actually fits. Snaps to a valid boundary (user, assistant, or compaction_summary — never mid-tool-result).
func GenerateSummary ¶
func GenerateSummary(ctx context.Context, provider core.Provider, model core.Model, opts core.StreamOptions, msgs []core.AgentMessage, previousSummary, focus string) (string, *core.Usage, error)
GenerateSummary makes an LLM call to summarize conversation messages. Returns the summary text, provider-reported usage (may be nil), or an error. focus, when non-empty, is a caller instruction (from `/compact <focus>`) that tells the summarizer what to keep in the foreground; it is advisory and never replaces the structured format.
func SerializeForSummary ¶
func SerializeForSummary(msgs []core.AgentMessage, maxInput int) string
SerializeForSummary converts messages to a human-readable transcript for the summarization prompt. Truncates at a limit derived from the model's context window (maxInput tokens). Pass 0 for the default (400k chars).
When the transcript exceeds the limit, the OLDEST messages are dropped rather than the newest. The previous implementation appended forward and broke on overflow, discarding the most recent turns — precisely the ones closest to the kept tail and most likely to describe live work.
Types ¶
type FileOps ¶
FileOps tracks file operations found in tool calls.
func ExtractFileOps ¶
func ExtractFileOps(msgs []core.AgentMessage) FileOps
ExtractFileOps scans messages for tool calls that reference files.
type Result ¶
type Result struct {
Summary string
TokensBefore int
TokensAfter int
ReadFiles []string
ModifiedFiles []string
Usage *core.Usage // LLM usage for the summarization call
}
Result holds the outcome of a compaction.
func Compact ¶
func Compact(ctx context.Context, provider core.Provider, model core.Model, opts core.StreamOptions, msgs []core.AgentMessage, contextTokens, contextWindow int, settings core.CompactionSettings, focus string) (*Result, []core.AgentMessage, error)
Compact orchestrates context compaction. Returns nil Result if nothing needs compacting. On LLM failure, returns the error with the original messages unchanged (non-fatal). focus is an optional caller instruction (from `/compact <focus>`) telling the summarizer what to keep in the foreground; empty for automatic compaction.
Source Files