[Bug] Dev agent fails to update story status due to conflicting instructions in dev.md

[ichunlai]

Describe the bug
The dev agent does not automatically update the story status to "Review" upon completion of the *develop-story task, even though the completion instructions in dev.md explicitly require it. This is caused by a direct contradiction in the agent's critical operating instructions.

One rule under story-file-updates-ONLY authorizes the agent to edit the Status, while the very next rule explicitly forbids it. This conflict leads the AI model to conservatively follow the prohibitive instruction, thus failing to update the status.

Steps to Reproduce

1. Ensure a story file (e.g., 1.2.story.md) has its status set to Approved.
2. Activate the dev agent in an agentic IDE (e.g., Cursor).
3. Execute the command: @dev *develop-story docs/stories/1.2.story.md.
4. Allow the dev agent to complete all tasks and validations.
5. Observe the agent's final output. It will indicate that the work is complete and ready for review.
6. Check the story file (1.2.story.md). The Status field remains Approved instead of being updated to Review.

PR
I do not have a PR to submit at this time, but the fix seems straightforward. A clear solution is outlined below.

Expected behavior
After successfully completing all tasks and validations for a story, the dev agent should automatically edit the story file to change its status from Approved (or InProgress) to Review, as per the completion step in its instructions: set story status: 'Ready for Review'.

Please be Specific if relevant

• Model(s) Used: GPT-5 Mini
• Agentic IDE Used: Any IDE capable of running BMad agents (e.g., Cursor, etc.). The issue is not IDE-specific but instruction-based.
• WebSite Used: N/A
• Project Language: C (with CMake)
• BMad Method version: 4.40

Screenshots or Links
N/A - The issue is clearly reproducible from the agent's instruction file.

Additional context
The core of the issue lies in these conflicting lines within .bmad-core/agents/dev.md under the develop-story command:

1. The Authorization Rule:

   - CRITICAL: You are ONLY authorized to edit these specific sections of story files - Tasks / Subtasks Checkboxes, Dev Agent Record section..., Change Log, **Status**

2. The Contradictory Prohibition Rule:

   - CRITICAL: DO NOT modify **Status**, Story, Acceptance Criteria, Dev Notes, Testing sections, or any other sections not listed above

The presence of Status in the "DO NOT modify" list appears to override the authorization and the final step of the completion workflow. Removing Status from the prohibitive rule should resolve the bug. This will improve the workflow's automation and align the agent's behavior with the documented process.

[dracic]

So obviously GPT-5 is way more strict at following the rules. I've never stumbled upon this with Sonnet.

[ichunlai]

Hi team,

I'd like to provide an update on this issue.

Initially, I believed the problem was caused by a logical contradiction in the prompt, as detailed in my original post. However, after correcting the conflicting instructions, the dev agent—running on GPT-5 Mini within GitHub Copilot's custom dev chat mode—still frequently fails to update the story's status to "Ready for Review" upon completing its development tasks.

I have to repeatedly prompt the agent, asking it to check the project's state, in order to compel it to update the Status field in the story. This persistence of the issue, even after the prompt fix, leads me to suspect that the root cause may be related to the behavior of the GPT-5 Mini model itself.

At this point, I'm not sure what the best next step is. I am willing to conduct some further experiments to better understand which prompts or contexts are influencing the model's decision-making process. Please let me know if you would prefer that I keep this issue open while I investigate further, or if I should close it for now. I will defer to your judgment.

Any insights or suggestions from the team would be greatly appreciated.

[joyshmitz]

@ichunlai Regarding GPT-5 Pro, not GPT-5 mini.
Some time ago, I researched the topic and read other researchers' work on why GPT-5 Pro is both the smartest and, in some ways, the most frustrating artificial intelligence model today.

Its parallel reasoning architecture improves accuracy by mimicking a group of experts, making it exceptional for multi-perspective analysis in fields such as science, finance, and law.

However, the same design weakens sequential performance, personality consistency, and security. To succeed with GPT-5 Pro, you need well-structured, multi-layered data — something that most organizations lack.

Strategically, this underscores the shift toward AI specialization, where deep reasoning systems, conversational AI, and specialized tools will coexist, rather than one model dominating all tasks.

[bmadcode]

@joyshmitz thank you for this explanation!

[joyshmitz]

@bmadcode I will leave it here, these observations formed the conclusion above

1. Parallel reasoning architecture: The main innovation of GPT-5 Pro is the parallel execution of multiple chains of reasoning, synthesizing them into a single answer, which improves the accuracy of solving complex, multifaceted problems.

2. Smarter, but worse in terms of experience: The same architecture that improves reasoning accuracy has drawbacks such as slower response times, loss of individuality, and vulnerability to hostile prompts.

3. High-value use cases: Excels in areas where accuracy is important and multifaceted analysis is required—scientific research, financial modeling, legal analysis, and technical architecture design.

4. Weaknesses in sequential tasks: It struggles with linear, personality-oriented, or narrative tasks such as coding implementation, creative writing, and real-time conversation.

5. Data structure requirements: For optimal performance, GPT-5 Pro requires multidimensional, structured datasets with clear perspectives, temporal references, and cross-references.

6. Strategic implications for AI developers: Signals the dawn of an era of architectural specialization, with different labs leaning toward deep reasoning, coding optimization, or personality-oriented conversational AI.

7. Intelligence ≠ usefulness: The smartest model is not always the most useful—value depends on matching architecture to task and data readiness.

What stuck with me:

• We are entering an era of architectural specialization, where different models will dominate different domains.

• Parallel reasoning makes GPT-5 Pro smarter, but also expands the attack surface and undermines personality.

• Intelligence is not the same as usefulness — what matters is whether the model's architecture fits your problem.