squeezr-ai 1.81.2 → 1.99.2

package/squeezr.toml

@@ -11,7 +11,12 @@ disabled = false
 # API key + backend = "gpt-mini" / "gemini-flash" / "local".
 ai_compression = false
 compress_system_prompt = true   # only runs when ai_compression = true (it makes a Haiku call)
-compress_conversation = true    # deterministic clean on old user/assistant messages + tool_use inputs (pure regex, no AI)
+compress_conversation = true    # deterministic clean on OLD user/assistant PROSE (pure regex, no AI)
+# compress_tool_inputs = false   # DEFAULT FALSE. Lossy-clean OLD Write/Edit/Bash tool INPUTS.
+                                 # OFF porque corrompía código en disco (v1.82.0): Write.content se
+                                 # escribe verbatim y Edit.old_string debe casar byte a byte. No lo
+                                 # enciendas salvo que aceptes ese riesgo; el bloat de turnos viejos
+                                 # ya lo recupera stale_turns de forma segura.
 keep_recent_assistant = 3       # leave last N assistant turns untouched
 assistant_threshold = 300       # min chars before applying prose cleanup
 stale_turns = true               # collapse old assistant turns when conversation > threshold (default 40)
@@ -31,6 +36,15 @@ tool_desc_expand = true          # store full spec in expand store — Claude ll
 # Per-command: add "# squeezr:skip" anywhere in a Bash command to skip that result:
 #   git diff HEAD~3  # squeezr:skip
 
+[output]
+# Output-side token reduction — trims what the model WRITES BACK (output costs
+# ~5x input on Opus). DEFAULT OFF (opt-in). Enable here or with SQUEEZR_OUTPUT_SHAPER=1.
+enabled = false
+verbosity_steering = true        # append a byte-stable "be terse" block to the system tail (cache-safe)
+level = 2                        # 1 (light) .. 4 (caveman). 2 = no ceremony + don't restate context
+effort_routing = true            # on mechanical turns (clean tool_result) lower an EXISTING thinking budget
+mechanical_thinking_floor = 1024 # clamp thinking.budget_tokens down to this on mechanical turns
+
 [cache]
 enabled = true
 max_entries = 1000       # LRU cap for cached compressions