cogit
An open-source AI programming assistance tool. It captures interaction messages between developers and agents to extract decisions and rejected options, links them to code commits, and injects historical rejection notes before the agent edits files to avoid repeated exploration.
Timeline (last 90 days)
A developer built an open-source tool called cogit to record the decisions and rejected options in messages sent to AI coding agents and associate them with code commits.
cogit captures each message through Claude Code hooks, uses a small model to generate summaries, and writes a line in the commit message pointing back to the corresponding message record.
In how cogit actually takes effect, before the agent is about to edit a file, it receives a brief prompt listing the options previously rejected for that file.
The developer spot-checked 20 message summaries, 6 of which contained errors, with the rejected-options portion being the weakest link in the summaries.
cogit currently provides historical rejection prompts only for a single file, and related files do not receive the same prompt; this problem remains unresolved.
The developer believes the information most worth bringing back to a new session is 'what was rejected' rather than 'what was decided,' because decisions are usually already reflected in the code, whereas rejected options leave no trace.
Most current context management approaches focus on preserving 'what was done,' while cogit's practice shows that 'what was not done' is likewise high-value information.
The developer's attitude toward using cogit summaries is: treat summaries as clues, not as facts; tool output requires human judgment.
All Facts (8)
A developer built an open-source tool called cogit to record the decisions and rejected options in messages sent to AI coding agents and associate them with code commits.
50%Unverifiedcogit captures each message through Claude Code hooks, uses a small model to generate summaries, and writes a line in the commit message pointing back to the corresponding message record.
50%UnverifiedIn how cogit actually takes effect, before the agent is about to edit a file, it receives a brief prompt listing the options previously rejected for that file.
50%UnverifiedThe developer spot-checked 20 message summaries, 6 of which contained errors, with the rejected-options portion being the weakest link in the summaries.
50%Unverifiedcogit currently provides historical rejection prompts only for a single file, and related files do not receive the same prompt; this problem remains unresolved.
50%UnverifiedThe developer believes the information most worth bringing back to a new session is 'what was rejected' rather than 'what was decided,' because decisions are usually already reflected in the code, whereas rejected options leave no trace.
50%UnverifiedMost current context management approaches focus on preserving 'what was done,' while cogit's practice shows that 'what was not done' is likewise high-value information.
50%UnverifiedThe developer's attitude toward using cogit summaries is: treat summaries as clues, not as facts; tool output requires human judgment.
50%