diff --git a/agent/decider.py b/agent/decider.py
index 59582cce..6a27c359 100644
--- a/agent/decider.py
+++ b/agent/decider.py
@@ -20,6 +20,7 @@
 import re
 
 from agent.context import context_for_agent
+from agent.philosophy import serialize_philosophy_for_prompt
 from agent.planner import _release_cadence_signal, _release_timing_state
 
 logger = logging.getLogger(__name__)
@@ -290,7 +291,7 @@ def _run_lens(name: str, context: dict, philosophy: dict, request: str, llm) ->
     system = _LENS_SYSTEMS[name]
     if name == "direction":
         user = (
-            f"Repository philosophy:\n{json.dumps(philosophy, indent=1)[:3000]}\n\n"
+            f"Repository philosophy:\n{serialize_philosophy_for_prompt(philosophy, 3000)}\n\n"
             f"Decision request: {request}\n\n"
             'Return JSON: {"verdict": "one short sentence", "reasoning": "why"}'
         )
@@ -314,7 +315,7 @@ def decide(context: dict, philosophy: dict, request: str, llm) -> dict:
         for name, verdict in lenses.items()
     )
     user = (
-        f"Repository philosophy:\n{json.dumps(philosophy, indent=1)[:3000]}\n\n"
+        f"Repository philosophy:\n{serialize_philosophy_for_prompt(philosophy, 3000)}\n\n"
         f"Repository state:\n{_render(context)}\n"
         f"{_release_context_note(context)}"
         f"{_planning_version_bump_note(context, request)}"
diff --git a/agent/philosophy.py b/agent/philosophy.py
index 38886f37..bcfe00de 100644
--- a/agent/philosophy.py
+++ b/agent/philosophy.py
@@ -133,3 +133,28 @@ def _render(context: dict) -> str:
         "labels", "milestones", "releases", "readme_excerpt",
     )}
     return json.dumps(keep, indent=1)[:12000]
+
+
+def serialize_philosophy_for_prompt(philosophy: dict, budget: int) -> str:
+    """Serialize ``philosophy`` for agent prompts without breaking JSON syntax.
+
+    When the full ``json.dumps(..., indent=1)`` fits ``budget``, the result is
+    byte-identical. Otherwise whole trailing ``evidence`` entries are dropped so
+    ``summary``/``values``/``merge_bar``/``direction`` survive as valid JSON. A raw
+    string slice is used only when the object still cannot fit with ``evidence``
+    empty — the same last-resort behavior as before this helper existed.
+    """
+    full = json.dumps(philosophy, indent=1)
+    if len(full) <= budget:
+        return full
+
+    if isinstance(philosophy, dict) and isinstance(philosophy.get("evidence"), list):
+        evidence = list(philosophy["evidence"])
+        while evidence:
+            evidence.pop()
+            trimmed = {**philosophy, "evidence": evidence}
+            candidate = json.dumps(trimmed, indent=1)
+            if len(candidate) <= budget:
+                return candidate
+
+    return full[:budget]
diff --git a/agent/planner.py b/agent/planner.py
index b14d7e00..1b19393d 100644
--- a/agent/planner.py
+++ b/agent/planner.py
@@ -12,6 +12,7 @@
 from datetime import datetime, timezone
 
 from agent.context import context_for_agent
+from agent.philosophy import serialize_philosophy_for_prompt
 
 logger = logging.getLogger(__name__)
 
@@ -1098,7 +1099,7 @@ def plan_next_actions(context: dict, philosophy: dict, n: int, llm) -> list:
     if not isinstance(context, dict):
         return _offline_plan_stub({}, n)
     user = (
-        f"Repository philosophy:\n{json.dumps(philosophy, indent=1)[:4000]}\n\n"
+        f"Repository philosophy:\n{serialize_philosophy_for_prompt(philosophy, 4000)}\n\n"
         f"Repository state:\n{_render(context)}\n"
         f"{_repo_layout_note(context)}"
         f"{_recent_kinds_note(context)}"
diff --git a/tests/test_philosophy.py b/tests/test_philosophy.py
index 43b50fd1..47d8be61 100644
--- a/tests/test_philosophy.py
+++ b/tests/test_philosophy.py
@@ -21,6 +21,7 @@
     _normalize_string_list,
     _normalize_text,
     infer_philosophy,
+    serialize_philosophy_for_prompt,
 )
 
 EXPECTED_KEYS = {"summary", "values", "merge_bar", "direction", "evidence"}
@@ -137,3 +138,64 @@ def test_infer_philosophy_non_dict_context_returns_fresh_copy():
     b = infer_philosophy(None, llm)
     a["summary"] = "mutated"
     assert b["summary"] != "mutated"
+
+
+def _oversized_philosophy(num_evidence: int = 13) -> dict:
+    """Shape sized like the #1706 instrumented run (4,237 chars full serialization)."""
+    return {
+        "summary": "A mature library prioritizing stability and a small dependency surface.",
+        "values": ["conservative", "stability-over-features", "docs-first"],
+        "merge_bar": (
+            "Merges well-justified fixes and incremental hardening; rejects churn, new "
+            "dependencies without payoff, and breaking changes outside a deprecation window."
+        ),
+        "direction": "Incremental hardening on the current major line with careful API evolution.",
+        "evidence": [f"signal {i}: " + ("x" * 300) for i in range(num_evidence)],
+    }
+
+
+def test_serialize_philosophy_for_prompt_is_byte_identical_when_it_fits():
+    philosophy = {"summary": "ship fast", "values": ["feature-first"], "merge_bar": "low",
+                  "direction": "expand", "evidence": ["recent feature flags"]}
+    rendered = serialize_philosophy_for_prompt(philosophy, 4000)
+    assert rendered == json.dumps(philosophy, indent=1)
+
+
+def test_serialize_philosophy_for_prompt_drops_trailing_evidence_to_stay_valid_json():
+    philosophy = _oversized_philosophy()
+    full = json.dumps(philosophy, indent=1)
+    assert len(full) > 4000
+
+    for budget in (4000, 3000):
+        rendered = serialize_philosophy_for_prompt(philosophy, budget)
+        assert len(rendered) <= budget
+        decoded = json.loads(rendered)
+        assert decoded["summary"] == philosophy["summary"]
+        assert decoded["values"] == philosophy["values"]
+        assert decoded["merge_bar"] == philosophy["merge_bar"]
+        assert decoded["direction"] == philosophy["direction"]
+        assert isinstance(decoded["evidence"], list)
+        assert decoded["evidence"] == philosophy["evidence"][: len(decoded["evidence"])]
+        if len(full) > budget:
+            assert len(decoded["evidence"]) < len(philosophy["evidence"])
+
+
+def test_serialize_philosophy_for_prompt_hard_slices_when_core_fields_exceed_budget():
+    philosophy = {
+        "summary": "p" * 5000,
+        "values": [],
+        "merge_bar": "high",
+        "direction": "forward",
+        "evidence": [],
+    }
+    full = json.dumps(philosophy, indent=1)
+    rendered = serialize_philosophy_for_prompt(philosophy, 3000)
+    assert rendered == full[:3000]
+    assert len(rendered) == 3000
+
+
+def test_serialize_philosophy_for_prompt_does_not_mutate_input():
+    philosophy = _oversized_philosophy()
+    original_evidence = list(philosophy["evidence"])
+    serialize_philosophy_for_prompt(philosophy, 3000)
+    assert philosophy["evidence"] == original_evidence