v1.17.0

[1.17.0] - 2026-09-12

New Features

• Access projects can now ask what a single control actually does and get a structured answer back: which procedure runs on the event, the call paths under it, and the tables it reads or writes along the way — assembled in one read instead of joined together by hand. Available from the library as getBehaviorEvidence and over MCP as codegraph_behavior_evidence. (#299)
• That answer is deliberately careful about what it does not know. A control name that exists on more than one form is refused with the candidates listed rather than answered for the wrong form, branches are kept apart instead of being strung into one sequence, and anything the index could not resolve — or a budget that cut the answer short — is reported alongside it. An empty list of tables means nothing is recorded, never that the code is harmless. (#299)
• A repeatable acceptance check for Access answers: npm run acceptance:vba indexes a fresh copy of a small checked-in Access export and asks it the questions a consumer really asks — which handler runs on this control, which tables it reaches, which form a query affects when nothing binds it to the screen — then compares the answers against expectations written by hand from the source. It fails on a missing answer and equally on an unrelated one, such as a second form's button of the same name. The same command can be pointed at an authorized copy of your own export, locally, without anything leaving your machine. (#301)

Fixes

• The Access guidance agents actually receive — the README and the instructions the MCP server sends on connect — now describes the form and report model the indexer really produces, instead of the older one it outgrew. It also states the parts that were never written down: an event binding is stored in one direction only, a control belongs to the form that contains it (so the same button name on two forms is two different answers), and the most common way of calling a procedure in VBA does not look like a call. (#300)
• The Access documentation now says plainly that CodeGraph reads the exported source tree, not the live Access file: nothing here proves a handler ran, and a missing relationship is missing evidence rather than proof that nothing happens. The worked example in the README is executed as a test against a checked-in fixture, so following it cannot silently stop working. (#300)
• The event-tracing guidance now separates custom VBA events from Access control events, which are wired differently and were easy to confuse. (#300)
• Tracing an Access form control now reaches its event handler and everything that handler calls. Starting a trace at a button used to come back empty, because the trace followed the wiring between a control and its handler in the wrong direction, and because the common way of calling a procedure in VBA — writing its name on a line by itself — was not being counted as a call at all. (#298)
• Impact analysis for a saved query now names the forms and reports affected by it even when nothing binds the query to the screen directly, so a query reached only through code no longer looks harmless. Ownership of a control is taken from the form that contains it, so two forms with a button of the same name no longer answer for each other. (#298)
• A trace no longer pads its answer with relationships that are not steps in the code's execution, such as what contains a procedure or which table it reads, and an unknown starting point now says so instead of returning an empty result that reads like "nothing calls this". (#298)
• Release notes published to GitHub now carry every section of the changelog. A section whose heading was more than one word — such as the list of new features — could be dropped from the published notes without any warning, which is how v1.16.0 shipped with its entire feature list missing. Those entries have been restored to the changelog. (#296)
• Each version heading in the changelog now links to this project's own release instead of a different project's, where the tag does not exist. (#296)
• Publishing a release now refuses to run from anywhere but the project's main branch, so the changelog update it makes can no longer be left behind on a temporary branch and lost. (#296)