@miller-tech/uap 1.179.8 → 1.179.9

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@miller-tech/uap",
-  "version": "1.179.8",
+  "version": "1.179.9",
   "description": "Autonomous AI agent memory system with CLAUDE.md protocol enforcement",
   "type": "module",
   "main": "dist/index.js",
@@ -21,7 +21,7 @@
     "start": "node dist/bin/cli.js",
     "test": "vitest",
     "test:ci": "vitest run",
-    "test:enforcers": "UAP_PROXY_ENV_AUTOLOAD=0 python3 -m unittest tools.agents.tests.test_enforcement_self_protect tools.agents.tests.test_schema_diff_gate tools.agents.tests.test_stream_telemetry tools.agents.tests.test_project_telemetry_events tools.agents.tests.test_workdir_scope_enforcer tools.agents.tests.test_path_containment tools.agents.tests.test_path_normalizer_hardened tools.agents.tests.test_sandbox_tool_strip tools.agents.tests.test_proxy_env_loader tools.agents.tests.test_stream_required_tool tools.agents.tests.test_tool_call_wrapper_profiles tools.agents.tests.test_tool_convert_cache tools.agents.tests.test_doubling_break tools.agents.tests.test_error_loop_ignores_correctives tools.agents.tests.test_attractor_detection tools.agents.tests.test_client_disconnect tools.agents.tests.test_confidence_escalation tools.agents.tests.test_coordination_ban tools.agents.tests.test_coordination_early_ban tools.agents.tests.test_cycle_break_exploration tools.agents.tests.test_deferral_break tools.agents.tests.test_deliver_autoroute tools.agents.tests.test_delivery_enforcement_all_langs tools.agents.tests.test_delivery_enforcement_exemptions tools.agents.tests.test_delivery_enforcement_filepath tools.agents.tests.test_delivery_enforcement_web_and_bash tools.agents.tests.test_disconnect_watcher tools.agents.tests.test_empty_maxtokens_recovery tools.agents.tests.test_empty_tool_loop_break tools.agents.tests.test_enforcer_escape_hatches tools.agents.tests.test_error_loop_break tools.agents.tests.test_finalize_suppression tools.agents.tests.test_malformed_unclosed_think tools.agents.tests.test_mandate_beats_recon tools.agents.tests.test_mandate_deliver tools.agents.tests.test_overflow_truncate_count_tokens tools.agents.tests.test_passthrough_oauth tools.agents.tests.test_project_telemetry tools.agents.tests.test_proxy_auth_headers tools.agents.tests.test_prune_preserve_force_write tools.agents.tests.test_recon_deliver_gate tools.agents.tests.test_session_admission tools.agents.tests.test_stream_heartbeat tools.agents.tests.test_stuck_break_reattach tools.agents.tests.test_turn_count_breaker_periodic tools.agents.tests.test_upstream_chokepoint tools.agents.tests.test_vision_passthrough tools.agents.tests.test_worktree_required tools.agents.tests.test_enforcer_suite_coverage tools.agents.tests.test_validate_plan_gate tools.agents.tests.test_validate_plan_inside_project tools.agents.tests.test_anthropic_proxy_streaming tools.agents.tests.test_delivery_enforcement_worktree",
+    "test:enforcers": "UAP_PROXY_ENV_AUTOLOAD=0 python3 -m unittest tools.agents.tests.test_enforcement_self_protect tools.agents.tests.test_schema_diff_gate tools.agents.tests.test_stream_telemetry tools.agents.tests.test_project_telemetry_events tools.agents.tests.test_workdir_scope_enforcer tools.agents.tests.test_path_containment tools.agents.tests.test_path_normalizer_hardened tools.agents.tests.test_sandbox_tool_strip tools.agents.tests.test_proxy_env_loader tools.agents.tests.test_stream_required_tool tools.agents.tests.test_tool_call_wrapper_profiles tools.agents.tests.test_tool_convert_cache tools.agents.tests.test_doubling_break tools.agents.tests.test_error_loop_ignores_correctives tools.agents.tests.test_attractor_detection tools.agents.tests.test_client_disconnect tools.agents.tests.test_confidence_escalation tools.agents.tests.test_coordination_ban tools.agents.tests.test_coordination_early_ban tools.agents.tests.test_cycle_break_exploration tools.agents.tests.test_deferral_break tools.agents.tests.test_deliver_autoroute tools.agents.tests.test_delivery_enforcement_all_langs tools.agents.tests.test_delivery_enforcement_exemptions tools.agents.tests.test_delivery_enforcement_filepath tools.agents.tests.test_delivery_enforcement_web_and_bash tools.agents.tests.test_disconnect_watcher tools.agents.tests.test_empty_maxtokens_recovery tools.agents.tests.test_empty_tool_loop_break tools.agents.tests.test_enforcer_escape_hatches tools.agents.tests.test_error_loop_break tools.agents.tests.test_finalize_suppression tools.agents.tests.test_malformed_unclosed_think tools.agents.tests.test_mandate_beats_recon tools.agents.tests.test_mandate_deliver tools.agents.tests.test_overflow_truncate_count_tokens tools.agents.tests.test_passthrough_oauth tools.agents.tests.test_project_telemetry tools.agents.tests.test_proxy_auth_headers tools.agents.tests.test_prune_preserve_force_write tools.agents.tests.test_recon_deliver_gate tools.agents.tests.test_session_admission tools.agents.tests.test_stream_heartbeat tools.agents.tests.test_stuck_break_reattach tools.agents.tests.test_turn_count_breaker_periodic tools.agents.tests.test_upstream_chokepoint tools.agents.tests.test_vision_passthrough tools.agents.tests.test_worktree_required tools.agents.tests.test_enforcer_suite_coverage tools.agents.tests.test_validate_plan_gate tools.agents.tests.test_validate_plan_inside_project tools.agents.tests.test_anthropic_proxy_streaming tools.agents.tests.test_delivery_enforcement_worktree tools.agents.tests.test_output_token_ceilings",
     "test:coverage": "vitest --coverage",
     "bench": "vitest --config vitest.bench.config.ts",
     "lint": "eslint src --ext .ts",

package/tools/agents/scripts/anthropic_proxy.py

@@ -42,11 +42,14 @@ Configuration (Environment Variables)
     PROXY_READ_TIMEOUT   Read timeout in seconds for upstream LLM streaming
                          Default: 600 (10 minutes)
 
+    PROXY_DEFAULT_MAX_TOKENS   max_tokens applied when the client sends none
+                               Default: 16384
+
     PROXY_TOOL_TURN_MAX_TOKENS   Max tokens for tool-call turns (0 to disable)
-                                Default: 8192
+                                Default: 32768
 
     PROXY_TOOL_TURN_MAX_TOKENS_GARBLED   Max tokens after garbled/malformed output
-                                         Default: 4096
+                                         Default: 16384
 
     PROXY_MAX_CONNECTIONS   Max concurrent connections to upstream
                             Default: 20
@@ -1011,6 +1014,14 @@ PROXY_STREAM_REASONING_MAX_CHARS = int(
     os.environ.get("PROXY_STREAM_REASONING_MAX_CHARS", "240")
 )
 PROXY_MAX_TOKENS_FLOOR = int(os.environ.get("PROXY_MAX_TOKENS_FLOOR", "16384"))
+# What a client gets when it sends no max_tokens at all — replaces a hardcoded
+# 4096 buried in openai_to_anthropic_request. Deliberately modest: for TOOL
+# turns this value barely matters, because _resolve_max_tokens_request already
+# lifts anything above 1024 to PROXY_MAX_TOKENS_FLOOR. It governs NO-TOOL turns
+# — judges, critics, acceptance verdicts — where a large budget buys nothing
+# and gives a thinking-runaway room to burn, so it is aligned with
+# THINKING_MIN_NO_TOOLS rather than raised to match a file-write ceiling.
+PROXY_DEFAULT_MAX_TOKENS = int(os.environ.get("PROXY_DEFAULT_MAX_TOKENS", "8192"))
 PROXY_TOOL_TURN_MAX_TOKENS = int(os.environ.get("PROXY_TOOL_TURN_MAX_TOKENS", "8192"))
 PROXY_TOOL_TURN_MAX_TOKENS_GARBLED = int(
     os.environ.get("PROXY_TOOL_TURN_MAX_TOKENS_GARBLED", "4096")
@@ -4848,7 +4859,17 @@ def openai_to_anthropic_request(openai_body: dict) -> dict:
     anthropic_body: dict = {
         "model": openai_body.get("model", "default"),
         "messages": anthropic_messages,
-        "max_tokens": int(openai_body.get("max_tokens", 4096) or 4096),
+        # A client that omits max_tokens is not asking for a small answer — it
+        # has no opinion. Defaulting to 4096 (~16KB of code) silently truncated
+        # every large write: the agentic executor sends no max_tokens, so 52 of
+        # 56 requests in a measured build ran at 4096 and a 36KB game.js came
+        # back cut off mid-file, which the model then spent turns "repairing".
+        # Neither the tool-turn cap (8192) nor llama's --n-predict (16384) was
+        # ever reached, so nothing logged a truncation — it just looked like a
+        # model that could not finish a file.
+        "max_tokens": int(
+            openai_body.get("max_tokens") or PROXY_DEFAULT_MAX_TOKENS
+        ),
     }
     if system_text_parts:
         anthropic_body["system"] = "\n\n".join(p for p in system_text_parts if p)
@@ -9249,6 +9270,7 @@ def _build_malformed_retry_body(
     attempt: int = 1,
     total_attempts: int = 1,
     is_garbled: bool = False,
+    is_truncated: bool = False,
     exclude_tools: list[str] | None = None,
 ) -> dict:
     retry_body = dict(openai_body)
@@ -9299,6 +9321,18 @@ def _build_malformed_retry_body(
             PROXY_TOOL_TURN_MAX_TOKENS_GARBLED,
             attempt,
         )
+    elif is_truncated:
+        # A write that ran out of tokens is the one case where retrying SMALLER
+        # is guaranteed to fail again — the attempt was cut off for lack of room,
+        # so clamping the retry to PROXY_MALFORMED_TOOL_RETRY_MAX_TOKENS (8192
+        # deployed, ~32KB) below the budget that already proved insufficient
+        # just reproduces the truncation. Keep the original budget instead.
+        # Observed: a 36KB module truncated, then every retry re-truncated.
+        logger.info(
+            "RETRY TRUNCATED: keeping max_tokens=%s (retrying smaller cannot fit "
+            "what was already cut off)",
+            retry_body.get("max_tokens"),
+        )
     elif PROXY_MALFORMED_TOOL_RETRY_MAX_TOKENS > 0:
         current_max = int(
             retry_body.get("max_tokens", PROXY_MALFORMED_TOOL_RETRY_MAX_TOKENS)
@@ -9632,6 +9666,7 @@ async def _apply_malformed_tool_guardrail(
             attempt=attempt + 1,
             total_attempts=attempts,
             is_garbled=current_issue.kind == "invalid_tool_args",
+            is_truncated=current_issue.kind == "truncated_tool_args",
             exclude_tools=exclude,
         )
         retry_resp = await client.post(

package/tools/agents/tests/test_output_token_ceilings.py

@@ -0,0 +1,155 @@
+#!/usr/bin/env python3
+"""Retrying a TRUNCATED write must not ask for less room than already failed.
+
+Measured on the Octopus Invaders build (2026-08-01): a 36KB game.js came back
+cut off mid-file and every retry re-truncated.
+
+The first diagnosis of this was wrong and is worth recording, because the wrong
+version is the intuitive one:
+
+  - The `max_tokens=4096` visible in the proxy's REQ: log is NOT the value sent
+    upstream. It is logged on the converted body, BEFORE
+    `_resolve_max_tokens_request` applies PROXY_MAX_TOKENS_FLOOR — which
+    returns max(requested, floor) and so raises it. With the deployed
+    FLOOR=32768 the default was never binding.
+  - A truncated write is reclassified to `truncated_tool_args` and therefore
+    never sets `last_response_garbled`, so PROXY_TOOL_TURN_MAX_TOKENS_GARBLED
+    does not govern its retry either.
+  - The actual clamp is PROXY_MALFORMED_TOOL_RETRY_MAX_TOKENS (8192 deployed),
+    applied via min() on the retry — i.e. the retry of a file that was cut off
+    for lack of room was given LESS room than the attempt that failed.
+
+So these tests pin the retry-budget branching, which is where the bug was, and
+the ordering properties of the ceiling chain. They deliberately do not assert
+absolute token literals: the deployed EnvironmentFile overrides most of these
+constants, so a test on a code default would pass while production disagreed.
+"""
+
+import importlib.util
+import os
+import unittest
+from pathlib import Path
+
+
+def _load_proxy(env=None):
+    prev = {}
+    env = dict(env or {})
+    env.setdefault("UAP_PROXY_ENV_AUTOLOAD", "0")
+    for k, v in env.items():
+        prev[k] = os.environ.get(k)
+        os.environ[k] = v
+    try:
+        path = Path(__file__).resolve().parents[1] / "scripts" / "anthropic_proxy.py"
+        spec = importlib.util.spec_from_file_location("anthropic_proxy_ceilings", path)
+        m = importlib.util.module_from_spec(spec)
+        spec.loader.exec_module(m)
+        return m
+    finally:
+        for k, v in prev.items():
+            if v is None:
+                os.environ.pop(k, None)
+            else:
+                os.environ[k] = v
+
+
+proxy = _load_proxy({"PROXY_MALFORMED_TOOL_RETRY_MAX_TOKENS": "8192"})
+
+BIG = 32768  # the budget a whole-module write actually runs with
+
+
+def _retry(**kw):
+    body = {"model": "m", "messages": [{"role": "user", "content": "hi"}], "max_tokens": BIG}
+    return proxy._build_malformed_retry_body(body, {"tools": []}, **kw)["max_tokens"]
+
+
+class TruncatedRetryBudgetTest(unittest.TestCase):
+    def test_truncated_retry_keeps_the_budget_that_was_cut_off(self):
+        # The whole bug: clamping here guarantees the retry truncates too.
+        self.assertEqual(_retry(is_truncated=True), BIG)
+
+    def test_garbled_retry_still_tightens(self):
+        # Genuine malformed args SHOULD get less room — that guard is intact.
+        self.assertEqual(
+            _retry(is_garbled=True), proxy.PROXY_TOOL_TURN_MAX_TOKENS_GARBLED
+        )
+
+    def test_plain_malformed_retry_still_clamps(self):
+        self.assertEqual(_retry(), proxy.PROXY_MALFORMED_TOOL_RETRY_MAX_TOKENS)
+
+    def test_truncation_wins_over_the_generic_clamp_but_not_over_garbled(self):
+        # Ordering matters: garbled is checked first by design, since args that
+        # are both malformed AND long are a degeneration risk, not a big file.
+        self.assertEqual(
+            _retry(is_garbled=True, is_truncated=True),
+            proxy.PROXY_TOOL_TURN_MAX_TOKENS_GARBLED,
+        )
+
+
+class TruncationSignalIsWiredTest(unittest.TestCase):
+    """The helper is only correct if the call site actually tells it.
+
+    Asserting the branch through the real call site would mean standing up the
+    whole async malformed-retry handler with a mocked upstream; the cheaper and
+    still-effective guard is that the kwarg is derived from the issue kind
+    rather than hardcoded. Without this, deleting the derivation leaves every
+    other test in this file green while the bug returns — which is exactly what
+    mutation testing showed.
+    """
+
+    SRC = (
+        Path(__file__).resolve().parents[1] / "scripts" / "anthropic_proxy.py"
+    ).read_text()
+
+    def test_call_site_derives_is_truncated_from_the_issue_kind(self):
+        self.assertIn(
+            'is_truncated=current_issue.kind == "truncated_tool_args"',
+            self.SRC,
+            "the retry no longer learns that the previous attempt was truncated",
+        )
+
+    def test_the_kind_it_keys_on_is_the_one_the_classifier_produces(self):
+        # Guards a rename on one side only.
+        self.assertIn('kind="truncated_tool_args"', self.SRC)
+
+
+class CeilingChainOrderingTest(unittest.TestCase):
+    def test_floor_raises_a_small_request_rather_than_capping_it(self):
+        # This is what made the 4096 default a red herring.
+        m = _load_proxy({"PROXY_MAX_TOKENS_FLOOR": "32768"})
+        self.assertEqual(m._resolve_max_tokens_request(4096), 32768)
+
+    def test_a_zero_floor_leaves_the_request_alone(self):
+        m = _load_proxy({"PROXY_MAX_TOKENS_FLOOR": "0"})
+        self.assertEqual(m._resolve_max_tokens_request(4096), 4096)
+
+    def test_garbled_cap_is_a_tightening_of_the_tool_turn_cap(self):
+        self.assertLessEqual(
+            proxy.PROXY_TOOL_TURN_MAX_TOKENS_GARBLED, proxy.PROXY_TOOL_TURN_MAX_TOKENS
+        )
+
+    def test_default_is_overridable_and_replaces_the_hardcoded_fallback(self):
+        m = _load_proxy({"PROXY_DEFAULT_MAX_TOKENS": "12345"})
+        out = m.openai_to_anthropic_request(
+            {"model": "m", "messages": [{"role": "user", "content": "hi"}]}
+        )
+        self.assertEqual(out["max_tokens"], 12345)
+
+    def test_an_explicit_client_value_survives_the_conversion(self):
+        # NB: only the conversion. build_openai_request may still raise it via
+        # the thinking floor — asserted here at the conversion layer only.
+        out = proxy.openai_to_anthropic_request(
+            {"model": "m", "messages": [{"role": "user", "content": "hi"}], "max_tokens": 512}
+        )
+        self.assertEqual(out["max_tokens"], 512)
+
+    def test_zero_or_null_means_no_opinion_not_emit_nothing(self):
+        for val in (0, None):
+            with self.subTest(val=val):
+                out = proxy.openai_to_anthropic_request(
+                    {"model": "m", "messages": [{"role": "user", "content": "hi"}], "max_tokens": val}
+                )
+                self.assertEqual(out["max_tokens"], proxy.PROXY_DEFAULT_MAX_TOKENS)
+
+
+if __name__ == "__main__":
+    unittest.main()