fix(agent): heal stringified payloads + enable prompt caching + lower step cap

[gianniskotsas]

Description

Investment plan run hit two issues at once:

• Validation failed twice: `monthlyAction` was emitted as a JSON-encoded string instead of a nested object → Pydantic rejected → run marked failed.
• $1.40 cost for one failed run — way too high.

This PR fixes both and adds a step-count guardrail.

Fixes

1. `_normalize_payload` — heal stringified nested fields

Sonnet sometimes returns nested object/array values in tool-input as JSON-encoded strings rather than native objects when the tool schema is deeply nested (the `InvestmentPlanOutput` schema is). Example we just saw:

```json
{ "monthlyAction": "{"proposedBuys":[{"symbol":"VUAA",...}]}", ... }
```

Pre-walk the payload before Pydantic validation; `json.loads` any string starting with `{` or `[`. Applied identically in `routine_runner._validate_or_downgrade` and `investment_plan_runner._validate`.

2. Anthropic prompt caching on system + tools

The system prompt (3-5k tokens) and tool definitions including the `InvestmentPlanOutput` JSON schema (5-15k tokens) are static across every iteration of an agent loop. We were re-paying full input cost ($3/M for Sonnet) on these tokens every step.

Marking the last block of `system` and `tools` with `cache_control: { type: "ephemeral" }` caches them for 5 minutes — easily long enough to cover the entire loop. Cached tokens cost 0.1× = $0.30/M instead of $3/M.

For a 10-step run that was costing $1.40, this drops to roughly $0.50-0.70. Bigger wins on longer loops.

3. `MAX_AGENT_STEPS` 30 → 15

Each step that uses `web_search` adds 5-20k tokens of search-result content. That content is re-included in every subsequent iteration's input → cost grows quadratically. Capping the loop bounds worst-case spend.

Type of Change

• Bug fix
• New feature
• Breaking change
• Documentation update

Tests

`tests/services/test_routine_runner.py` and `tests/services/test_investment_plan_runner.py` (6/6) still pass — they patch `call_agent_step` so caching internals don't affect tests, and the validation-test payload is well-formed so `_normalize_payload` is a no-op.

---

Summary by cubic

Fixes validation failures from stringified nested tool payloads and reduces agent run cost by enabling prompt caching in anthropic and lowering the step cap. MAX_AGENT_STEPS is now 15, and runs that previously failed on fields like monthlyAction now pass; typical loops cost ~50% less.

• Bug Fixes
  • Pre-parse JSON-like strings in tool payloads before Pydantic validation in both routine and investment plan runners.

^{Written for commit e031a20. Summary will update on new commits.}

[vercel]

The latest updates on your projects. Learn more about Vercel for GitHub.

Project	Deployment	Actions	Updated (UTC)
syllogic	Ready	Preview, Comment	May 4, 2026 10:11am

[cubic-dev-ai]

2 issues found across 3 files

Prompt for AI agents (unresolved issues)

Check if these issues are valid — if so, understand the root cause of each and fix them. If appropriate, use sub-agents to investigate and fix each issue separately.


<file name="backend/app/services/investment_plan_runner.py">

<violation number="1" location="backend/app/services/investment_plan_runner.py:149">
P1: `_normalize_payload` only normalizes top-level keys, so nested stringified JSON fields are still left as strings and can continue to fail validation.</violation>
</file>

<file name="backend/app/services/routine_runner.py">

<violation number="1" location="backend/app/services/routine_runner.py:162">
P1: `_normalize_payload` only heals top-level fields, so nested stringified JSON can still bypass normalization and trigger validation failures.</violation>
</file>

_{Reply with feedback, questions, or to request a fix. Tag @cubic-dev-ai to re-run a review.}

[cubic-dev-ai]

P1: _normalize_payload only normalizes top-level keys, so nested stringified JSON fields are still left as strings and can continue to fail validation.

Prompt for AI agents

Check if this issue is valid — if so, understand the root cause and fix it. At backend/app/services/investment_plan_runner.py, line 149:

<comment>`_normalize_payload` only normalizes top-level keys, so nested stringified JSON fields are still left as strings and can continue to fail validation.</comment>

<file context>
@@ -136,7 +136,29 @@ def _agent_loop(plan: InvestmentPlan, transcript: list[dict], usage_totals: dict
+    if not isinstance(payload, dict):
+        return payload
+    healed: dict = {}
+    for k, v in payload.items():
+        if isinstance(v, str) and v and v[0] in ("{", "["):
+            try:
</file context>

[cubic-dev-ai]

P1: _normalize_payload only heals top-level fields, so nested stringified JSON can still bypass normalization and trigger validation failures.

Prompt for AI agents

Check if this issue is valid — if so, understand the root cause and fix it. At backend/app/services/routine_runner.py, line 162:

<comment>`_normalize_payload` only heals top-level fields, so nested stringified JSON can still bypass normalization and trigger validation failures.</comment>

<file context>
@@ -149,8 +149,30 @@ def _agent_loop(routine: Routine, transcript: list[dict], usage_totals: dict) ->
+    if not isinstance(payload, dict):
+        return payload
+    healed: dict = {}
+    for k, v in payload.items():
+        if isinstance(v, str) and v and v[0] in ("{", "["):
+            try:
</file context>

Suggested change

	for k, v in payload.items():
	if isinstance(v, str) and v and v[0] in ("{", "["):
	try:
	healed[k] = json.loads(v)
	continue
	except json.JSONDecodeError:
	pass
	healed[k] = v
	return healed
	def _heal(value: Any) -> Any:
	if isinstance(value, dict):
	return {kk: _heal(vv) for kk, vv in value.items()}
	if isinstance(value, list):
	return [_heal(item) for item in value]
	if isinstance(value, str):
	stripped = value.lstrip()
	if stripped and stripped[0] in ("{", "["):
	try:
	return _heal(json.loads(stripped))
	except json.JSONDecodeError:
	return value
	return value
	for k, v in payload.items():
	healed[k] = _heal(v)
	return healed