Class: Rubino::Context::PromptAssembler

Inherits:
Object
  • Object
show all
Defined in:
lib/rubino/context/prompt_assembler.rb

Overview

Assembles the complete prompt from all context sources. Returns the message array (system + summary + history) for LLM submission.

Constant Summary collapse

TOOL_USE_ENFORCEMENT_MODELS =

Model-name substrings that trigger tool-use-enforcement steering under the "auto" default. Ported from Hermes' TOOL_USE_ENFORCEMENT_MODELS, plus "minimax" — rubino's default targets it and its field notes document the same text-leaked-tool-call / narrate-instead-of-act failures.

%w[
  gpt codex gemini gemma grok glm qwen deepseek minimax
].freeze

Class Method Summary collapse

Instance Method Summary collapse

Constructor Details

#initialize(session:, memory_context:, config:, agent_definition: nil, ignore_rules: false) ⇒ PromptAssembler

Returns a new instance of PromptAssembler.



54
55
56
57
58
59
60
61
62
63
64
65
66
 
# File 'lib/rubino/context/prompt_assembler.rb', line 54

def initialize(session:, memory_context:, config:, agent_definition: nil,
               ignore_rules: false)
  @session = session
  @memory_context = memory_context
  @config = config
  @agent_definition = agent_definition
  # --ignore-rules suppresses project-context discovery
  # (AGENTS.md/CLAUDE.md/.rubino.md/.cursorrules). The flag is threaded
  # from Lifecycle so the CLI option genuinely skips discovery (#47), not
  # just the trust gate.
  @ignore_rules = ignore_rules
  @message_store = Session::Store.new
end

Class Method Details

.reset_all_snapshots! ⇒ Object

Test/teardown hook. Not part of the public API.



41
42
43
 
# File 'lib/rubino/context/prompt_assembler.rb', line 41

def reset_all_snapshots!
  @snapshots_mutex.synchronize { @snapshots.clear }
end

.reset_snapshot!(session_id) ⇒ Object

Drops the cached snapshot for a session so the next assembly captures fresh memory state. Use this when a tool call must influence the very next turn (trade-off: the freeze stops protecting against same-turn poisoning).



36
37
38
 
# File 'lib/rubino/context/prompt_assembler.rb', line 36

def reset_snapshot!(session_id)
  @snapshots_mutex.synchronize { @snapshots.delete(session_id) }
end

.snapshot_for(session_id) ⇒ Object

Returns the cached snapshot for a session, computing it via the supplied block on first access. The block receives no args and must return the memory-context hash to freeze.



26
27
28
29
30
 
# File 'lib/rubino/context/prompt_assembler.rb', line 26

def snapshot_for(session_id)
  @snapshots_mutex.synchronize do
    @snapshots[session_id] ||= yield
  end
end

Instance Method Details

#build ⇒ Object

Builds and returns the full message array for LLM submission



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
 
# File 'lib/rubino/context/prompt_assembler.rb', line 69

def build
  messages = []

  # Single system message. The content is split into a STABLE PREFIX
  # (identity / product / env / user-profile snapshot / skills / project
  # context) and a VOLATILE TAIL (freshly-retrieved relevant-memories +
  # the post-compaction session-summary). The prefix carries the prompt-
  # cache breakpoint (#311); the tail — which can change turn-to-turn —
  # sits AFTER it so the cached bytes stay byte-stable. Both regions live
  # in ONE role:"system" entry (#253), built as a Content::Raw array of
  # text blocks when caching is on, or a plain joined String otherwise.
  messages << { role: "system", content: system_content }

  # Conversation history. Repair tool pairing across the FULL list before
  # mapping to wire format — this is the defensive "net" that recovers
  # sessions already corrupted by the historical metadata-dropping bug in
  # compaction/fork (those rows exist in prod). Mirrors Claude Code's
  # pre-call sanitization: never emit an orphan tool block that 400s a
  # strict provider. Conservative by design — when in doubt, keep.
  #
  # Drop `[harness control]` messages FIRST: they are ephemeral within-turn
  # runtime control (Loop's iteration-cap summary nudge, the resume /
  # truncation / blocked-tool reminders, #75), written on the user's behalf
  # for ONE model call — not durable conversation. They stay in the DB for
  # audit/scrollback (and are already excluded from the rewind picker,
  # chat_command#rewindable_message?), but must never re-enter the model's
  # context on a LATER turn. Re-feeding the synthetic "you've reached the
  # maximum tool-calling iterations" user turn made the model believe it
  # was perpetually capped and parrot a "start a new session" refusal on
  # every subsequent turn instead of acting. This is the single chokepoint
  # every surface (CLI + web/API) rebuilds context through, so filtering
  # here fixes them all.
  durable = @message_store.for_session(@session[:id]).reject { |m| harness_control?(m) }
  history = repair_tool_pairs(durable)
  history.each do |msg|
    messages << msg.to_context
  end

  messages
end