Compress tool outputs, logs, files, and RAG chunks before they reach the LLM. 20% fewer tokens for coding agents, 60-95% fewer tokens for JSON, same answers. Library, proxy, MCP server.
Python agentaianthropicclaude-codecompressioncontext-engineeringcontext-windowcursorfastapilangchainllmmcpopenaiprompt-engineeringproxypythonragtoken-optimizationtokenstypescript
fix(router): stop protecting passing build/test output as error traces (#1740)
Ashish Patel7ab83c5107b5
diff --git a/.github/pr-images/issue-1696-error-protection-fix.svg b/.github/pr-images/issue-1696-error-protection-fix.svgnew file mode 100644index 000000000..c3d9915fd--- /dev/null+++ b/.github/pr-images/issue-1696-error-protection-fix.svg@@ -0,0 +1,78 @@+<svg xmlns="http://www.w3.org/2000/svg" viewBox="0 0 1200 900" font-family="Segoe UI, Helvetica, Arial, sans-serif">+ <rect x="0" y="0" width="1200" height="900" fill="#ffffff"/>+ <text x="600" y="36" text-anchor="middle" font-size="22" font-weight="700" fill="#111827">Headroom: Error-Output Protection False-Positive Fix (#1696)</text>++ <!-- Box 1: pipeline -->+ <g>+ <rect x="20" y="60" width="560" height="190" rx="10" fill="#eff6ff" stroke="#2563eb"/>+ <text x="36" y="88" font-size="15" font-weight="700" fill="#1d4ed8">1. THE GATE</text>+ <text x="36" y="112" font-size="13" fill="#111827">content_router.py routes every message/content-block through</text>+ <text x="36" y="130" font-size="13" fill="#111827">content_has_strong_error_indicators() before deciding whether</text>+ <text x="36" y="148" font-size="13" fill="#111827">to compress it or protect it (pass through untouched).</text>+ <text x="36" y="176" font-size="13" fill="#111827">Rule: if the text contains 2+ DISTINCT keywords from</text>+ <text x="36" y="194" font-size="13" font-family="Consolas, monospace" fill="#b91c1c">[error, fail, exception, traceback, fatal, panic, crash]</text>+ <text x="36" y="218" font-size="13" fill="#111827">the block is protected — assumed to be a real failure trace.</text>+ </g>++ <!-- Box 2: the bug -->+ <g>+ <rect x="620" y="60" width="560" height="190" rx="10" fill="#fef2f2" stroke="#dc2626"/>+ <text x="636" y="88" font-size="15" font-weight="700" fill="#b91c1c">2. THE BUG</text>+ <text x="636" y="112" font-size="13" fill="#111827">Passing build/test output mentions BOTH words without failing:</text>+ <text x="636" y="136" font-size="13" font-family="Consolas, monospace" fill="#111827">tsc: "Found 0 errors"</text>+ <text x="636" y="156" font-size="13" font-family="Consolas, monospace" fill="#111827">jest: "0 failures, 42 passed"</text>+ <text x="636" y="184" font-size="13" fill="#111827">→ 2 distinct keyword hits ("error" + "fail") on a CLEAN run</text>+ <text x="636" y="202" font-size="13" font-weight="700" fill="#b91c1c">→ wrongly protected, forever, from compression.</text>+ </g>++ <!-- Box 3: impact -->+ <g>+ <rect x="20" y="270" width="1160" height="150" rx="10" fill="#fff7ed" stroke="#ea580c"/>+ <text x="36" y="298" font-size="15" font-weight="700" fill="#c2410c">3. OBSERVED IMPACT (issue #1696 stats.json)</text>+ <text x="36" y="324" font-size="13" fill="#111827">In a long multi-turn JS/TS coding session (KiloCode via headroom proxy), "router:protected:error_output" fired on</text>+ <text x="36" y="344" font-size="13" fill="#111827">nearly every request. Combined with tool-output exclusion and small-block skips, only a sliver of tokens were ever</text>+ <text x="36" y="364" font-size="13" fill="#111827">eligible for compression.</text>+ <rect x="900" y="325" width="60" height="60" fill="#fecaca"/>+ <text x="930" y="360" text-anchor="middle" font-size="16" font-weight="700" fill="#7f1d1d">0.3%</text>+ <text x="930" y="400" text-anchor="middle" font-size="11" fill="#111827">actual</text>+ <rect x="980" y="285" width="60" height="100" fill="#bbf7d0"/>+ <text x="1010" y="340" text-anchor="middle" font-size="14" font-weight="700" fill="#14532d">60-95%</text>+ <text x="1010" y="400" text-anchor="middle" font-size="11" fill="#111827">expected</text>+ </g>++ <!-- Box 4: the fix -->+ <g>+ <rect x="20" y="440" width="560" height="200" rx="10" fill="#ecfdf5" stroke="#059669"/>+ <text x="36" y="468" font-size="15" font-weight="700" fill="#047857">4. THE FIX</text>+ <text x="36" y="494" font-size="13" fill="#111827">Strip common zero-result phrases before the keyword scan:</text>+ <text x="36" y="516" font-size="13" font-family="Consolas, monospace" fill="#111827">"0 error(s)", "no error(s)", "0 failing",</text>+ <text x="36" y="536" font-size="13" font-family="Consolas, monospace" fill="#111827">"0 failure(s)", "no failure(s)"</text>+ <text x="36" y="562" font-size="13" fill="#111827">headroom/transforms/error_detection.py</text>+ <text x="36" y="582" font-size="13" fill="#111827">content_has_strong_error_indicators()</text>+ <text x="36" y="608" font-size="13" font-weight="700" fill="#047857">→ clean tool output no longer trips protection.</text>+ </g>++ <!-- Box 5: still safe -->+ <g>+ <rect x="620" y="440" width="560" height="200" rx="10" fill="#f5f3ff" stroke="#7c3aed"/>+ <text x="636" y="468" font-size="15" font-weight="700" fill="#6d28d9">5. REAL FAILURES STILL PROTECTED</text>+ <text x="636" y="494" font-size="13" font-family="Consolas, monospace" fill="#111827">Traceback (most recent call last):</text>+ <text x="636" y="512" font-size="13" font-family="Consolas, monospace" fill="#111827"> ...</text>+ <text x="636" y="530" font-size="13" font-family="Consolas, monospace" fill="#111827">ValueError: fatal error during load</text>+ <text x="636" y="558" font-size="13" fill="#111827">"traceback" + "fatal" (+ "error") — 2+ distinct hits</text>+ <text x="636" y="576" font-size="13" fill="#111827">outside a stripped zero-result phrase</text>+ <text x="636" y="602" font-size="13" font-weight="700" fill="#6d28d9">→ still protected, correctly.</text>+ </g>++ <!-- Box 6: tests -->+ <g>+ <rect x="20" y="660" width="1160" height="140" rx="10" fill="#f8fafc" stroke="#334155"/>+ <text x="36" y="688" font-size="15" font-weight="700" fill="#1e293b">6. TEST COVERAGE ADDED (tests/test_error_detection.py — none existed before)</text>+ <text x="36" y="714" font-size="13" fill="#111827">✓ real error/traceback text is still flagged ✓ single keyword mention is not flagged</text>+ <text x="36" y="736" font-size="13" fill="#111827">✓ passing tsc summary ("Found 0 errors") is not flagged ✓ passing eslint summary is not flagged</text>+ <text x="36" y="758" font-size="13" fill="#111827">✓ "0 errors" text does not mask a real second indicator elsewhere in the same blob</text>+ <text x="36" y="782" font-size="13" fill="#111827">All 5 new tests + full existing content_router suite (86 tests total) pass.</text>+ </g>++ <text x="600" y="850" text-anchor="middle" font-size="12" fill="#6b7280">Fixes GitHub issue #1696 · headroomlabs-ai/headroom</text>+</svg>diff --git a/headroom/transforms/error_detection.py b/headroom/transforms/error_detection.pyindex 8b6a6cac6..545eef5cf 100644--- a/headroom/transforms/error_detection.py+++ b/headroom/transforms/error_detection.py@@ -145,6 +145,28 @@ def content_has_error_indicators(text: str) -> bool:return bool(_rust_content_has_error_indicators(text))+# Success-summary phrases from common build/test/lint tools that legitimately+# pair two indicator keywords (`error` + `fail`) while reporting a PASS, e.g.+# tsc's "Found 0 errors", jest's "0 failing" / "0 failures" / "0 failed",+# eslint's "0 problems (0 errors, 0 warnings)", or label:value summaries like+# "Failures: 0", "failed: 0", "Errors=0". Stripped before the keyword scan+# below so a clean JS/TS toolchain run doesn't get permanently protected from+# compression for the rest of a long coding session (issue #1696).+#+# `fail(?:ed|ing|ures?)?` covers fail/failed/failing/failure/failures — the+# keyword scan below matches the "fail" substring inside all of them, so the+# scrubber must strip all of them too, not just the forms literally named+# "failing"/"failure(s)".+_ZERO_RESULT_PATTERN = re.compile(+ # "0 errors" / "no failed" — count-first forms.+ r"\b(?:0|no)\s+(?:errors?|fail(?:ed|ing|ures?)?)\b"+ # "Errors: 0" / "failed=0" — label:value / label=value forms used by+ # other CI/test tools' summary lines.+ r"|\b(?:errors?|fail(?:ed|ing|ures?)?)\s*[:=]\s*0\b",+ re.IGNORECASE,+)++def content_has_strong_error_indicators(text: str) -> bool:"""Stricter triage for compression-protection gates.@@ -161,8 +183,16 @@ def content_has_strong_error_indicators(text: str) -> bool:``crash``), while passing mentions rarely do. Misses here aresafe — downstream compressors (LogCompressor) still preserveerror lines.++ Before scanning, ``_ZERO_RESULT_PATTERN`` strips zero-result+ summary phrases (``"0 errors"``, ``"failed: 0"``, ...) so a+ passing build/test/lint run doesn't trip the two-keyword+ threshold just because its PASS summary happens to mention both+ "error" and "fail" at count zero (see issue #1696 — this was+ firing on nearly every request in a long JS/TS coding session+ and defeating compression almost entirely)."""- lowered = text.lower()+ lowered = _ZERO_RESULT_PATTERN.sub(" ", text.lower())hits = 0for keyword in ERROR_INDICATOR_KEYWORDS:if keyword in lowered:diff --git a/tests/test_error_detection.py b/tests/test_error_detection.pynew file mode 100644index 000000000..2238e9244--- /dev/null+++ b/tests/test_error_detection.py@@ -0,0 +1,58 @@+"""Regression tests for the error/importance detection triage helpers."""++from __future__ import annotations++from headroom.transforms.error_detection import content_has_strong_error_indicators+++def test_real_error_output_is_detected() -> None:+ text = "Traceback (most recent call last):\n ...\nValueError: fatal error during load"+ assert content_has_strong_error_indicators(text)+++def test_single_keyword_mention_is_not_flagged() -> None:+ # Only one distinct indicator keyword ("error") — should not trip the+ # two-keyword threshold.+ text = 'Wrote error_handler.py with an "errors": [] field.'+ assert not content_has_strong_error_indicators(text)+++def test_tsc_passing_summary_is_not_flagged() -> None:+ # Regression for issue #1696: a clean `tsc` run mentions both "error"+ # and (via "0 failures" in a paired test run) "fail" while reporting+ # success. Previously this tripped the two-keyword heuristic and got+ # the message permanently protected from compression.+ text = "Found 0 errors. Watching for file changes.\nTests: 0 failures, 42 passed"+ assert not content_has_strong_error_indicators(text)+++def test_eslint_passing_summary_is_not_flagged() -> None:+ text = "0 problems (0 errors, 0 warnings)\nno failing tests"+ assert not content_has_strong_error_indicators(text)+++def test_zero_result_phrase_does_not_mask_a_real_second_error() -> None:+ # "0 errors" is stripped, but a genuine second distinct indicator+ # elsewhere in the same blob must still trigger protection.+ text = "0 errors from linter, but the build crashed with a fatal signal"+ assert content_has_strong_error_indicators(text)+++def test_zero_failed_form_is_not_flagged() -> None:+ # Reviewer regression (PR #1740): "0 failed" wasn't covered by the+ # original pattern (only "failing"/"failure(s)"), so "failed" still+ # contributed a "fail" keyword hit alongside "0 errors" and tripped+ # the false positive this fix targets.+ text = "Found 0 errors\nTests: 0 failed, 42 passed"+ assert not content_has_strong_error_indicators(text)+++def test_label_value_summary_formats_are_not_flagged() -> None:+ # Broader CI summary formats (not just "N word" / "word N"): label:value+ # and label=value pairs, in either error/fail order.+ for text in (+ "Failures: 0, Errors: 0",+ "failed: 0, errors: 0",+ "Errors=0 Failures=0",+ ):+ assert not content_has_strong_error_indicators(text), text