fix: tolerate string-encoded skills param for LLM provider compatibility by jinglun010-cpu · Pull Request #771 · usestrix/strix

jinglun010-cpu · 2026-07-15T06:08:46Z

Summary

Some LLM providers (notably qwen3.7-max via Alibaba DashScope OpenAI-compatible endpoint) serialize array-typed function call parameters as JSON-encoded strings ('["sql_injection"]') instead of native JSON arrays (["sql_injection"]). This causes Pydantic validation to reject the skills parameter before the function body executes, making every create_agent and load_skill call fail.

This completely breaks multi-agent orchestration — the agent cannot spawn specialist child agents and falls back to serial execution with severely reduced test coverage.

Root Cause

The skills parameter is typed as list[str] in both create_agent and load_skill. When an LLM provider passes it as a string, Pydantic's strict validation rejects it:

Invalid JSON input for tool create_agent: 1 validation error for create_agent_args
skills
  Input should be a valid list [type=list_type, input_value='["sql_injection"]', input_type=str]

Fix

Three changes per tool (create_agent + load_skill):

Widen type annotation: list[str] → str | list[str] — allows Pydantic to accept string input
Add strict_mode=False to @function_tool — relaxes schema validation

Parse string in function body:

if isinstance(skills, str):
    try:
        skills = json.loads(skills)
    except json.JSONDecodeError:
        skills = [s.strip() for s in skills.split(",") if s.strip()]

The fallback comma-split handles edge cases where the model passes a bare comma-separated string like "xss,sqli".

Testing

Before fix: 6 consecutive create_agent calls all failed with Pydantic validation error. Agent could not spawn any child agents.
After fix: create_agent successfully spawns child agents. Verified with a quick-mode scan where the child agent executed 23 exec_command calls (curl + Python SQL injection scripts) and completed normally.

No regressions for providers that correctly pass arrays (OpenAI, Anthropic) — isinstance(skills, str) is False for native lists, so the tolerance code is skipped.

Files Changed

strix/tools/agents_graph/tools.py — create_agent function
strix/tools/load_skill/tool.py — load_skill function

Closes #770

Some LLM providers (e.g. qwen3.7-max via DashScope OpenAI-compatible endpoint) serialize array-typed function call parameters as JSON-encoded strings instead of native JSON arrays. This causes Pydantic validation to reject the skills parameter before the function body executes, making every create_agent and load_skill call fail. Changes: - Widen skills type annotation: list[str] -> str | list[str] - Add strict_mode=False to @function_tool decorator - Parse string inputs in function body (json.loads with comma-split fallback) Tested with qwen3.7-max: create_agent successfully spawns child agents that execute 23+ exec_command calls for SQL injection testing. Closes usestrix#770

greptile-apps · 2026-07-15T06:11:42Z

Greptile Summary

This PR makes skill arguments compatible with providers that encode arrays as strings. The main changes are:

Widen skills to accept strings in create_agent and load_skill.
Disable strict tool validation for both functions.
Decode JSON strings with a comma-separated fallback.

Confidence Score: 4/5

The decoded skill values need shape validation before merging.

JSON-encoded string arrays follow the intended path.
Valid JSON scalars, objects, and mixed arrays can violate the downstream list[str] contract.
Both changed tools can misinterpret input or fail unexpectedly.

strix/tools/agents_graph/tools.py and strix/tools/load_skill/tool.py

Important Files Changed

Filename	Overview
strix/tools/agents_graph/tools.py	Adds string decoding for `create_agent`, but decoded values are not checked against the required string-list shape.
strix/tools/load_skill/tool.py	Adds matching compatibility parsing for `load_skill`, with the same missing decoded-value validation.

Prompt To Fix All With AI

Fix the following 2 code review issues. Work through them one at a time, proposing concise fixes.

---

### Issue 1 of 2
strix/tools/agents_graph/tools.py:423-427
**Decoded JSON Bypasses List Shape**

When a provider sends valid JSON that is not a string array, `json.loads()` still succeeds. A JSON string is expanded into individual characters by `list(skills)`, an object is treated as its keys, and a mixed array can raise inside `validate_requested_skills()` while formatting heterogeneous values, so `create_agent` can misapply skills or fail instead of returning a controlled validation error.

```suggestion
    if isinstance(skills, str):
        original_skills = skills
        try:
            decoded_skills = json.loads(skills)
        except json.JSONDecodeError:
            skills = [s.strip() for s in skills.split(",") if s.strip()]
        else:
            if isinstance(decoded_skills, str):
                skills = [decoded_skills]
            elif isinstance(decoded_skills, list) and all(
                isinstance(skill, str) for skill in decoded_skills
            ):
                skills = decoded_skills
            else:
                skills = [original_skills]
```

### Issue 2 of 2
strix/tools/load_skill/tool.py:31-35
**Decoded JSON Bypasses List Shape**

When `skills` contains valid non-array JSON, decoding succeeds without enforcing the downstream `list[str]` contract. A JSON string is split into characters, an object silently becomes its keys, and mixed element types can make validation raise, so `load_skill` can load the wrong selection or terminate the tool call unexpectedly.

```suggestion
    if isinstance(skills, str):
        original_skills = skills
        try:
            decoded_skills = json.loads(skills)
        except json.JSONDecodeError:
            skills = [s.strip() for s in skills.split(",") if s.strip()]
        else:
            if isinstance(decoded_skills, str):
                skills = [decoded_skills]
            elif isinstance(decoded_skills, list) and all(
                isinstance(skill, str) for skill in decoded_skills
            ):
                skills = decoded_skills
            else:
                skills = [original_skills]
```

_{Reviews (1): Last reviewed commit: "fix: tolerate string-encoded skills para..." | Re-trigger Greptile}

greptile-apps · 2026-07-15T06:11:45Z

+    if isinstance(skills, str):
+        try:
+            skills = json.loads(skills)
+        except json.JSONDecodeError:
+            skills = [s.strip() for s in skills.split(",") if s.strip()]


Decoded JSON Bypasses List Shape

When a provider sends valid JSON that is not a string array, json.loads() still succeeds. A JSON string is expanded into individual characters by list(skills), an object is treated as its keys, and a mixed array can raise inside validate_requested_skills() while formatting heterogeneous values, so create_agent can misapply skills or fail instead of returning a controlled validation error.

Suggested change

if isinstance(skills, str):

try:

skills = json.loads(skills)

except json.JSONDecodeError:

skills = [s.strip() for s in skills.split(",") if s.strip()]

if isinstance(skills, str):

original_skills = skills

try:

decoded_skills = json.loads(skills)

except json.JSONDecodeError:

skills = [s.strip() for s in skills.split(",") if s.strip()]

else:

if isinstance(decoded_skills, str):

skills = [decoded_skills]

elif isinstance(decoded_skills, list) and all(

isinstance(skill, str) for skill in decoded_skills

):

skills = decoded_skills

else:

skills = [original_skills]

Prompt To Fix With AI

This is a comment left during a code review. Path: strix/tools/agents_graph/tools.py Line: 423-427 Comment: **Decoded JSON Bypasses List Shape** When a provider sends valid JSON that is not a string array, `json.loads()` still succeeds. A JSON string is expanded into individual characters by `list(skills)`, an object is treated as its keys, and a mixed array can raise inside `validate_requested_skills()` while formatting heterogeneous values, so `create_agent` can misapply skills or fail instead of returning a controlled validation error. ```suggestion if isinstance(skills, str): original_skills = skills try: decoded_skills = json.loads(skills) except json.JSONDecodeError: skills = [s.strip() for s in skills.split(",") if s.strip()] else: if isinstance(decoded_skills, str): skills = [decoded_skills] elif isinstance(decoded_skills, list) and all( isinstance(skill, str) for skill in decoded_skills ): skills = decoded_skills else: skills = [original_skills] ``` How can I resolve this? If you propose a fix, please make it concise.

greptile-apps · 2026-07-15T06:11:46Z

+    if isinstance(skills, str):
+        try:
+            skills = json.loads(skills)
+        except json.JSONDecodeError:
+            skills = [s.strip() for s in skills.split(",") if s.strip()]


Decoded JSON Bypasses List Shape

When skills contains valid non-array JSON, decoding succeeds without enforcing the downstream list[str] contract. A JSON string is split into characters, an object silently becomes its keys, and mixed element types can make validation raise, so load_skill can load the wrong selection or terminate the tool call unexpectedly.

Suggested change

if isinstance(skills, str):

try:

skills = json.loads(skills)

except json.JSONDecodeError:

skills = [s.strip() for s in skills.split(",") if s.strip()]

if isinstance(skills, str):

original_skills = skills

try:

decoded_skills = json.loads(skills)

except json.JSONDecodeError:

skills = [s.strip() for s in skills.split(",") if s.strip()]

else:

if isinstance(decoded_skills, str):

skills = [decoded_skills]

elif isinstance(decoded_skills, list) and all(

isinstance(skill, str) for skill in decoded_skills

):

skills = decoded_skills

else:

skills = [original_skills]

Prompt To Fix With AI

This is a comment left during a code review. Path: strix/tools/load_skill/tool.py Line: 31-35 Comment: **Decoded JSON Bypasses List Shape** When `skills` contains valid non-array JSON, decoding succeeds without enforcing the downstream `list[str]` contract. A JSON string is split into characters, an object silently becomes its keys, and mixed element types can make validation raise, so `load_skill` can load the wrong selection or terminate the tool call unexpectedly. ```suggestion if isinstance(skills, str): original_skills = skills try: decoded_skills = json.loads(skills) except json.JSONDecodeError: skills = [s.strip() for s in skills.split(",") if s.strip()] else: if isinstance(decoded_skills, str): skills = [decoded_skills] elif isinstance(decoded_skills, list) and all( isinstance(skill, str) for skill in decoded_skills ): skills = decoded_skills else: skills = [original_skills] ``` How can I resolve this? If you propose a fix, please make it concise.

Greptile review on usestrix#771 flagged that the previous fix accepted any valid JSON, not just string arrays. A JSON object would be silently converted to its keys, an int array would trip downstream validation, and a JSON scalar would be split into characters. Apply the suggested shape-validation fallback: after json.loads, only accept str (wrap in list) or list[str]; otherwise fall back to treating the original string as a single skill name. This keeps the qwen-style JSON-encoded array path working while hardening the other shapes. Cases now handled: '["a", "b"]' -> ['a', 'b'] (target case, unchanged) 'a, b' -> ['a', 'b'] (comma fallback, unchanged) '"a"' -> ['a'] (JSON string scalar) '{"a": 1}' -> ['{"a": 1}'] (was: ['a'] — keys leaked) '[1, 2, 3]' -> ['[1, 2, 3]'] (was: ints passed downstream) '["a", 1, "b"]' -> ['["a", 1, "b"]'] (was: mixed types downstream)

jinglun010-cpu · 2026-07-19T14:04:56Z

Thanks @greptile-apps[bot] for the thorough review — the three flagged regressions (object → keys leak, int array, mixed array) were real.

Applied the suggested shape-validation fallback in c1a04fe:

After json.loads, only accept:

str → wrap in list ('"a"' → ['a'])
list[str] with all-string elements → use as-is

Otherwise fall back to treating the original string as a single skill name. This preserves the qwen/DashScope JSON-encoded-array path while hardening the other shapes.

Verified locally against the flagged cases:

input	before	after
`'{"a": 1}'`	`['a']` (keys leaked)	`['{"a": 1}']`
`'[1, 2, 3]'`	`[1, 2, 3]` (ints downstream)	`['[1, 2, 3]']`
`'["a", 1, "b"]'`	mixed downstream	`['["a", 1, "b"]']`
`'["a", "b"]'`	`['a', 'b']`	`['a', 'b']` (unchanged)
`'a, b'`	`['a', 'b']`	`['a', 'b']` (unchanged)

Both create_agent and load_skill patched identically. Ready for another look.

greptile-apps Bot reviewed Jul 15, 2026

View reviewed changes

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

Uh oh!

fix: tolerate string-encoded skills param for LLM provider compatibility#771

fix: tolerate string-encoded skills param for LLM provider compatibility#771
jinglun010-cpu wants to merge 2 commits into
usestrix:mainfrom
jinglun010-cpu:fix/create-agent-skills-string-tolerance

jinglun010-cpu commented Jul 15, 2026

Uh oh!

greptile-apps Bot commented Jul 15, 2026

Uh oh!

greptile-apps Bot Jul 15, 2026

Uh oh!

greptile-apps Bot Jul 15, 2026

Uh oh!

jinglun010-cpu commented Jul 19, 2026

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

1 participant

Uh oh!

Conversation

jinglun010-cpu commented Jul 15, 2026

Summary

Root Cause

Fix

Testing

Files Changed

Uh oh!

greptile-apps Bot commented Jul 15, 2026

Greptile Summary

Confidence Score: 4/5

Important Files Changed

Uh oh!

greptile-apps Bot Jul 15, 2026

Choose a reason for hiding this comment

Uh oh!

greptile-apps Bot Jul 15, 2026

Choose a reason for hiding this comment

Uh oh!

jinglun010-cpu commented Jul 19, 2026

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

1 participant