signature

Every developer using a coding agent has faced this situation: the agent makes a change that breaks the project, and returning to the last stable state becomes difficult without losing valuable work. Or you switch agents, start a new session, and have to explain the entire project again—including its architecture, constraints, previous failures, and the reasoning behind important decisions. Meanwhile, the new agent scans large parts of the codebase to rebuild that understanding, consuming time, context space, and tokens. Project Signature is here to help. Project Signature is a repository-native memory and recovery layer for AI coding agents. It keeps important architectural decisions, discovered constraints, failed approaches, and verified solutions with the project rather than inside a single conversation. This knowledge remains available when a session ends, when the context becomes too long, or when the developer moves between IBM Bob, Codex, Claude Code, and other compatible agents. Through MCP, every connected agent can access the same durable project knowledge. Before beginning a task, Project Signature ranks existing records using the current objective, affected files, tags, project scope, and error fingerprints. It returns only the most relevant information within a defined token budget, helping the agent understand the task without loading the entire project history or repeatedly analyzing the whole codebase. When an agent reaches an important conclusion, it can suggest saving it as a project decision. For example, after discovering that a dataset is strongly imbalanced, it might propose using stratified cross-validation. The proposal includes the reasoning, evidence, affected scope, constraints, and rejected alternatives. The developer always remains in control. A proposed decision can be reviewed, edited, approved, rejected, or ignored. Nothing becomes permanent project memory without explicit approval. The same workflow records recurring failures, attempted fixes, verified root causes, tested remedies, and the evidence showing that a solution works. Project Signature uses an append-only ledger, so approved knowledge cannot be silently rewritten. When a decision evolves, a new record supersedes the previous one while preserving its history. Future agents can understand not only the current decision, but also why it was made, which alternatives were rejected, and what changed over time. The project also provides session-aware recovery for agent-generated changes. Before modifying the repository, an agent starts a session and creates a checkpoint. During implementation, it records the files it changes and the important commands it executes. If the work fails, Project Signature creates a read-only rollback preview that clearly distinguishes agent-created changes from work that existed before the session or was made outside it. Rollback is deliberately conservative and always requires explicit user approval. It restores only files that can be safely attributed to the agent session, removes only files created by that session, and protects pre-existing or unrelated developer work. Conflicts, stale previews, changed Git commits, staged session changes, and unsafe paths automatically block the rollback. The result is a portable “project signature” that follows the codebase across agents and sessions. It helps coding agents remember why the project was built a certain way, avoid repeating known mistakes, retrieve focused context with fewer tokens, and safely recover when an automated change goes wrong.