Bug Description
Component: Skill - ce-doc-review
Summary: During the per-finding walk-through, user responses to questions are overridden by AI logic. When the user gives a directive in response to a finding, the AI will sometimes ignore some or all of the directive and apply its own logic instead.
This is exacerbated by using non-Claude models such as deepseek-v4-pro
Environment
- Plugin Version: 3.11.2 (from cache path)
- Agent Platform: Claude Code
- Agent Version: 2.1.170
- OS: Linux ubuntu-server 6.8.0-117-generic
What Happened
- User invoked /ce-doc-review on a plan document
- Walk-through presented Finding 1 — whether to keep or drop a CSV file
- User gave a clear directive: 'Keep the CSV. Remove duplicate data from Python scripts and STRATEGY.md. The CSV is the single source of truth.'
- AI recorded the following and updated the plan file to include conflicting information: drop the CSV file but also keep the CSV file
User answered Claude's questions:
· Drop the CSV — replace with a shared Python data module (scripts/curriculum_data.py) that all scripts import? → Keep the CSV. Remove the duplicate hard-coded values from both the python scripts and STRATEGY.md. The strategy document should reference the CSV file without duplicating the data. The python shared utility must read the CSV file. To create the CSV, extract the information from the strategy document, NOT the python script. Moving forward, the CSV file will be the source of truth for planned articles. This is a directive that MUST NOT be ignored. Make sure the plan is updated on disk
- Immediately after, Finding 2 was presented
- User had to explicitly say 'go back and redo Finding 1' because the AI's recording didn't match what the user said
Expected Behavior
When the user gives a directive in response to a walk-through finding, that directive should:
- Be recorded accurately (exact user words, not AI summary)
- Cascade to any remaining findings that depend on or contradict it
- Not be re-litigated by subsequent findings
- Be recorded directly in the plan file
Steps to Reproduce
- Create a plan document with a controversial architectural decision
- Run /ce-doc-review on it
- Choose 'Review each finding one by one'
- When a finding presents two options, choose the option the reviewers didn't recommend
- Observe the AI's response when it moves on
Error Messages
None — silent logic error, not a crash.
Additional Context
The synthesis pipeline (Phase 3.5c) has premise-dependency chain linking that could be used here — if Finding 1 is a root and Finding 2-5 are dependents, accepting the root's fix should cascade. But this doesn't appear to fire when the user gives a custom directive (as opposed to picking Apply on the suggested fix).
Reported via /ce-report-bug skill
Bug Description
Component: Skill - ce-doc-review
Summary: During the per-finding walk-through, user responses to questions are overridden by AI logic. When the user gives a directive in response to a finding, the AI will sometimes ignore some or all of the directive and apply its own logic instead.
This is exacerbated by using non-Claude models such as
deepseek-v4-proEnvironment
What Happened
Expected Behavior
When the user gives a directive in response to a walk-through finding, that directive should:
Steps to Reproduce
Error Messages
None — silent logic error, not a crash.
Additional Context
The synthesis pipeline (Phase 3.5c) has premise-dependency chain linking that could be used here — if Finding 1 is a root and Finding 2-5 are dependents, accepting the root's fix should cascade. But this doesn't appear to fire when the user gives a custom directive (as opposed to picking Apply on the suggested fix).
Reported via /ce-report-bug skill