Describe the feature or problem you'd like to solve
Add inferred behavioral posture separate from mode selection.
Proposed solution
Proposed solution
Add an inferred, persisted posture layer separate from mode.
Examples of useful postures:
- Talk — discuss only
- Sketch — refine shape without prematurely formalizing
- Plan — produce a durable plan
- Do — execute continuously
- Checkpoint — persist current state deliberately, including explicit incompleteness if needed
- Close — summarize/approve only when genuinely ready
These labels are examples, not required names.
The important implementation point is that posture should be backed internally by explicit behavioral gates, such as:
- execution allowed?
- artifact creation allowed?
- persistence allowed?
- persistence threshold: milestone or completion?
- stop rule: blockers only or frequent check-ins?
- formalization pressure: low or high?
- closure allowed?
This internal structure matters more than the user-facing labels.
Behavioral expectations:
- infer posture from conversational cues by default
- allow the user to override posture explicitly
- persist posture across context compression
- avoid crossing from refinement into formalization, execution, or closure without sufficient evidence
- support deliberate incomplete checkpointing as a first-class behavior
- preserve the current cadence instead of drifting back toward premature plan/closure behavior
Key distinctions that should be preserved:
- “be in a planning/sketching posture” is not the same as “produce a formal plan artifact now”
- “this is complete” is not the same as “this is important enough to checkpoint even though it is incomplete”
- “the user is still refining” is not the same as “the user is now asking for decision, execution, or closure”
Minimal success criterion:
If the user is still refining, the agent should remain in a low-formalization, low-closure posture unless the user explicitly requests an artifact, execution, or wrap-up, or there is strong evidence of readiness.
That inferred posture should survive context compression and should not reset into a more eager planning/closure cadence later in the same session.
This would benefit users by reducing premature formalization, premature execution, premature closure, and the need to repeatedly correct the agent’s working stance.
Example prompts or workflows
Example prompts or workflows
- Exploratory refinement without premature planning
User: “Let’s think this through first. I’m not ready for a plan yet.”
Desired behavior: The agent stays in discussion/sketch posture and does not create a formal plan artifact until explicitly asked.
- Refinement with open design uncertainty
The agent itself identifies an unresolved design choice, but should not switch into approval/closure posture just because the discussion has become more coherent. It should remain in refinement posture until the open issue is actually resolved or the user explicitly asks to
checkpoint incomplete work.
- Semantic refinement before forced decision framing
User is still working through the conceptual model.
Desired behavior: The agent should avoid jumping to preset multiple-choice questions or decision posture before enough shared grounding exists.
- Continuous refinement without re-prompting every step
User is iterating on an idea over multiple exchanges.
Desired behavior: The agent recognizes an ongoing refinement posture and continues appropriately, instead of behaving as if each refinement step needs a fresh push from the user.
- Deliberate incomplete checkpointing
User: “Write down where we are, but this is not done yet.”
Desired behavior: The agent should support explicit checkpoint posture without treating checkpointing as equivalent to completion or approval.
Additional context
- Premature closure despite acknowledged open holes
The agent tried to close plan mode while its own summary still said that one design choice remained open. I had to stop it and explicitly ask to continue refining before committing the plan.
- Premature concretization instead of semantic refinement
The agent asked a preset multiple-choice question before the semantic model had been worked through. I had to push it back into refinement.
- Over-fragmented prompting cadence
I had to explicitly tell the agent to continue without waiting for me to push it after every refinement step.
- Repeated drift back toward eager planning unless corrected
The interaction converged better when kept in a slower, discussion-heavy posture, but the agent repeatedly drifted back toward premature closure or planning unless corrected.
The problem is not just that the agent sometimes asks the wrong question. The deeper problem is that it lacks an explicit model of interaction posture, so it can be locally reasonable while globally mistimed.
Describe the feature or problem you'd like to solve
Add inferred behavioral posture separate from mode selection.
Proposed solution
Proposed solution
Add an inferred, persisted posture layer separate from mode.
Examples of useful postures:
These labels are examples, not required names.
The important implementation point is that posture should be backed internally by explicit behavioral gates, such as:
This internal structure matters more than the user-facing labels.
Behavioral expectations:
Key distinctions that should be preserved:
Minimal success criterion:
If the user is still refining, the agent should remain in a low-formalization, low-closure posture unless the user explicitly requests an artifact, execution, or wrap-up, or there is strong evidence of readiness.
That inferred posture should survive context compression and should not reset into a more eager planning/closure cadence later in the same session.
This would benefit users by reducing premature formalization, premature execution, premature closure, and the need to repeatedly correct the agent’s working stance.
Example prompts or workflows
Example prompts or workflows
User: “Let’s think this through first. I’m not ready for a plan yet.”
Desired behavior: The agent stays in discussion/sketch posture and does not create a formal plan artifact until explicitly asked.
The agent itself identifies an unresolved design choice, but should not switch into approval/closure posture just because the discussion has become more coherent. It should remain in refinement posture until the open issue is actually resolved or the user explicitly asks to
checkpoint incomplete work.
User is still working through the conceptual model.
Desired behavior: The agent should avoid jumping to preset multiple-choice questions or decision posture before enough shared grounding exists.
User is iterating on an idea over multiple exchanges.
Desired behavior: The agent recognizes an ongoing refinement posture and continues appropriately, instead of behaving as if each refinement step needs a fresh push from the user.
User: “Write down where we are, but this is not done yet.”
Desired behavior: The agent should support explicit checkpoint posture without treating checkpointing as equivalent to completion or approval.
Additional context
The agent tried to close plan mode while its own summary still said that one design choice remained open. I had to stop it and explicitly ask to continue refining before committing the plan.
The agent asked a preset multiple-choice question before the semantic model had been worked through. I had to push it back into refinement.
I had to explicitly tell the agent to continue without waiting for me to push it after every refinement step.
The interaction converged better when kept in a slower, discussion-heavy posture, but the agent repeatedly drifted back toward premature closure or planning unless corrected.
The problem is not just that the agent sometimes asks the wrong question. The deeper problem is that it lacks an explicit model of interaction posture, so it can be locally reasonable while globally mistimed.