Class: Rubino::Context::Compressor
- Inherits:
-
Object
- Object
- Rubino::Context::Compressor
- Defined in:
- lib/rubino/context/compressor.rb
Overview
Orchestrates context compaction: flush memory, split messages into head/middle/tail, generate summary, create child session.
Constant Summary collapse
- INEFFECTIVE_SAVINGS_PCT =
Anti-thrashing back-off (#415a, ported from Hermes context_compressor.py should_compress). A session hovering right at the threshold re-pays a summary call every turn even though each pass only shaves a message or two. If the two most recent compactions in this session’s lineage each saved less than INEFFECTIVE_SAVINGS_PCT of their original tokens, skip auto-compaction until genuinely new work pushes savings back up (the user can still force /compact). Returns true when compaction should be SKIPPED.
0.10- INEFFECTIVE_STREAK =
2
Instance Method Summary collapse
-
#compact! ⇒ Object
Performs full compaction and returns metadata.
-
#initialize(session_id:, config: nil, db: nil) ⇒ Compressor
constructor
A new instance of Compressor.
- #thrashing? ⇒ Boolean
Constructor Details
#initialize(session_id:, config: nil, db: nil) ⇒ Compressor
Returns a new instance of Compressor.
10 11 12 13 14 15 16 17 |
# File 'lib/rubino/context/compressor.rb', line 10 def initialize(session_id:, config: nil, db: nil) @session_id = session_id @config = config || Rubino.configuration @db = db || Rubino.database.db @message_store = Session::Store.new(db: @db) @session_repo = Session::Repository.new(db: @db) @summary_store = Session::SummaryStore.new(db: @db) end |
Instance Method Details
#compact! ⇒ Object
Performs full compaction and returns metadata
38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 |
# File 'lib/rubino/context/compressor.rb', line 38 def compact! session = @session_repo.find(@session_id) raise CompactionError, "Session not found: #{@session_id}" unless session # Resolve a SHORT id to the FULL session id before any message lookup # (#352): #find prefix-matches "5aebd8ce" to the row, but # `for_session(short_id)` matches messages by EXACT session_id and so # returned 0 rows — compaction then short-circuited to no_op_result and # the CLI printed a fake "compacted · saved 0 tok" success. Pin every # downstream lookup (messages, summaries, lineage) to the resolved id. @session_id = session[:id] @session_model = session[:model] = @message_store.for_session(@session_id) return no_op_result(:too_few_messages) if .size < # Below the token threshold a summary call COSTS more than it saves: the # generated summary (budgeted up to compression.max_summary_tokens) is # routinely larger than the handful of middle messages it replaces, so a # forced compaction on a small session GROWS context instead of shrinking # it (the QA "/compact on a 872-tok session → 2736 tok" bug). Both paths # must clear the SAME gate the auto path uses (TokenBudget#needs_compaction?) # — without it the manual path silently summarized, inflated, AND forked. # Industry norm (Claude Code / Codex): /compact below threshold is a no-op. return no_op_result(:below_threshold) unless needs_compaction?() # 1. Flush memory before compaction flush_memory! # 2. Split messages into head / middle / tail boundary = MessageBoundary.new(messages: , config: @config) head = boundary.head middle = boundary.middle tail = boundary.tail return no_op_result if middle.empty? # 3. Sanitize tool pairs in middle if @config.compression_preserve_tool_pairs? sanitizer = ToolPairSanitizer.new middle = sanitizer.sanitize(middle) end # saved_tokens reports what leaves the LIVE transcript, so measure it # on the pre-prune middle (the pruned copy feeds only the summarizer). middle_tokens = estimate_tokens(middle) # 3b. Cheap LLM-free pre-pass (#415d): dedupe + summarize old tool # results in the middle BEFORE the paid summary call, so raw tool # noise (file reads, terminal dumps) doesn't inflate the summarizer # prompt. The middle is summarized then discarded, so a lossy # representation here is safe. middle = ToolResultPruner.new.prune(middle) # 4. Load previous summary (capture id now, before the insert below # overwrites "latest" — the lineage link must point at the prior row) previous = @summary_store.latest(@session_id) previous_summary = previous&.dig(:content) previous_summary_id = previous&.dig(:id) # 5. Generate new summary summary_builder = SummaryBuilder.new(session_id: @session_id) new_summary = summary_builder.build( messages: middle, previous_summary: previous_summary ) # Steps 6-8 are the irreversible state mutation; commit them atomically. summary_id, child_session = commit_compaction!( session: session, head: head, tail: tail, messages: , new_summary: new_summary, previous_summary_id: previous_summary_id ) { source_session_id: @session_id, target_session_id: child_session[:id], original_messages: .size, compacted_messages: head.size + tail.size + 1, # +1 for summary saved_tokens: middle_tokens, summary_id: summary_id } end |
#thrashing? ⇒ Boolean
30 31 32 33 34 35 |
# File 'lib/rubino/context/compressor.rb', line 30 def thrashing? rows = recent_lineage_compactions(INEFFECTIVE_STREAK) return false if rows.size < INEFFECTIVE_STREAK rows.all? { |r| ineffective?(r) } end |