From Gate to Test in Gantry's Prose
A documentation pass that says test where the words are for people and keeps gate where they name a thing.
0 milestones 4 sprints
2 plan 5 execute 5 review
1h 28m total 12m 35s per sprint
Gantry checks its own work after an agent changes the tree, and the prose around that mechanism used to ask one word to do too much. This job moved the reader-facing language toward tests, the test suite, and checks while preserving literal names such as the script path, prompt filenames, stage labels, and serialized keys.
The useful part of the run is the surface-based decomposition. Public prose set the wording, core prompts carried it into worker instructions, design prompts handled the special case where some validation is structural rather than executable, and the remaining documentation pass applied the policy across the broader live tree. Later work has reorganized many files and reintroduced some shorthand, but the live prompts still show the distinction the job established.
How this walkthrough is structured
Feature
What did the run build and ship?
The tree still separates many human instructions into tests and checks while retaining `bin/gate` and gate-shaped config keys as literal protocol names.
Build
How did Gantry structure the work?
The run treated the rename as a prose-boundary problem, letting each fresh worker handle one reader surface while a shared preserve-list kept machine names stable.
Public Overview
2 agents 8m 51s wall time
The first pass changed the overview and nearby public docs from gate-centered explanation to test-centered explanation. At current `HEAD`, the README still says Gantry runs tests itself and shows the implement, test, fix, review flow, but later prose also says Gantry trusts its gate and uses gate as a shorthand again. The feature therefore survives as a partial public wording change, not as a durable ban on the old word in overview prose.
This sprint carried the highest-register wording first, so the later prompt and documentation workers had an example instead of a blank terminology choice. The brief kept the task narrow: rewrite explanatory prose, preserve code spans and paths, and keep the retained script path literal while changing the surrounding description. That is a good memoryless slice because the worker could read the overview, apply one disambiguation rule, and stop before touching role prompts or broad reference material. Its boundary held without repair.
Core Prompts
2 agents 12m 35s wall time
The old core prompt files have since been reorganized into `config/agent-prompts`, but their descendants still carry the rename. The test-builder prompt now says it authors tests and an executable test script at `{{ORCH}}/bin/gate`; execute and fix prompts say the pass or fail decision comes from `{{GATE_PATH}}`; review text treats tests as green but incomplete with respect to scope. The literal path and placeholder names remain stable, so the prompt contract changed for readers without renaming the protocol the engine invokes.
This was the coupled prompt slice, where future agents learn what Gantry expects from planning, execution, repair, review, and test authoring. The brief singled out the prompt that writes the retained script path, which made the hard part explicit: describe the concept as tests while keeping the path and placeholders untouched. A fresh worker could hold it because the scope was one prompt family with one vocabulary policy. The run facts show no retry or red outcome here.
Design Prompts
2 agents 12m 10s wall time
The current design prompt tree keeps that distinction. The briefs pass asks for a structural check that verifies coverage and required sections, while the pages pass asks for tests that run the project and prove new routes render. Both still write the executable script at `{{ORCH}}/bin/gate`, so design mode exposes the same preserved machine path with prose matched to what each pass actually validates.
The design prompt slice had a different vocabulary problem than the core prompts. Some design-mode passes produce prose briefs, so calling their validator a test suite would be inaccurate; other passes build pages and should be described as tests. The brief gave the worker that split directly, which kept the work small enough to handle without reinterpreting design mode from scratch. It landed cleanly on the same verification path as the earlier prompt work.
Live Docs
2 agents 27m 48s wall time
The documentation pass reached reference docs, handovers, plans, screencast material, scripts, and source-adjacent prose. Many of the one-off plan and handover files it touched are no longer present, while durable areas such as screencast runbooks, map-generation comments, and harness references still contain test and check wording from this line of work. Some current docs again use gate in ordinary prose, so this sprint's lasting contribution is best described as a broad terminology refresh whose results were later mixed with new shorthand rather than as an enforced repository-wide convention.
This was the broad sweep after the narrower wording patterns were established. Its brief covered live documentation outside archived history and repeated the conservative preserve-list: examples, identifiers, filenames, paths, serialized keys, and historical quotations were not to move. That made the sprint larger than the earlier ones but less conceptually open; the worker was applying a settled policy across many files. The run facts show extra successful test runs around this boundary, and the ledger records the sprint as green.