Intent-level conflict detection

A diff tells you what changed. It cannot tell you that the change broke a contract another module relied on. By the time a diff exists, the work is done and the conflict is expensive.

Intent-level conflict detection moves the check earlier. Each agent declares what it is about to do. The master spots the overlap, the dependency break, and the contradicting assumption before either agent writes a line.

Git was built for humans typing slowly. It shows line changes. It cannot say this change broke the contract that module X relied on. Two agents can touch completely different files and still collide at the concept level.

Detection at merge time means the cost has already been paid: both agents finished, both diffs exist, and someone has to unwind one of them. The bottleneck moved from writing code to reconciling it.

The shim coaxes a plan out of the agent's natural turn-taking and submits it to the master as an intent: a natural-language summary plus steps. The agent never enumerates paths or symbols. It describes the work in sentences, and the master derives the concrete claims.

When the intent arrives, the master resolves concepts, derives the files and symbols the agent will likely touch, detects conflicts against all active intents, and returns a verdict plus the relevant world slice. The richer the intent, the smarter the verdict.

The master is LLM-driven from day one for one job: matching. It walks the linked-file memory to pull candidate concepts connected to the intent's area, reasons over the intent and the candidates, and decides matched-existing, propose-new, or refuse.

Every decision mutates the registry. A match adds realizations, refines names, and bumps contract checksums. Lexical similarity alone cannot read a natural-language intent, which is why the lookup plus reasoning loop is there from the first version, not bolted on later.

Plans drift. During execution the shim resolves the agent's file writes to concepts, a semantic diff rather than a line diff. When a write lands on a concept outside the plan's matched set, the shim asks the agent to explain, then submits that explanation as a rescope intent on the same code path.

Planned and unplanned intents run the same matching and conflict pipeline. The master sees one uniform primitive whether the work was declared up front or discovered mid-flight.

See the semantic concept catalogue for the registry the matching loop reads from, or why a coordination layer beats more guardrails for the argument behind catching conflicts at the plan stage.