Prompt templates: reusable briefs for every loop type

[scgopi]

Implements PROMPT_TEMPLATES.md (New Designs v4), checked against GraphCode New Loop v4.dc.html, README.md §3 and MAIN_LOOP.md, and folds in the fixes from a design review of the first draft.

What it adds

• Templates button + ⌘T picker in the New loop dialog — search-first, replacing the dialog body, grouped This project above All projects.
• Applied state that never gains a shape silently: a removable chip, a plain-words strip, a "from template" dot on every field it set (nothing locked), Undo the shape and ✕.
• ⇥ walks the unfilled tokens across brief, done check, metric and branch — the canvas prints "One token left to fill · ⇥ to jump to it", so the key had to actually do it.
• Storage: one template is one markdown file. ~/.graphcode/templates is where the app writes; <project>/.graphcode/templates is also read and wins on a filename collision. Sharing is a git action — no sync layer, no .gitignore touching.
• Save as template from the dialog, a loop's context menu, and a composite's menu. Values typed into {token} positions come back as tokens.
• Follow vs snapshot: timed and composite loops re-read their template on the next run; main, goal and turn loops snapshot at creation. Detach sits beside the Follows chip on the card and in its context menu.
• Template editor in Settings (the picker's "Manage…" destination) — name, shape, brief and settings, listing home and project templates, with the design's load-bearing usage line above Save.

Review fixes included

Data loss / correctness

Was	Now
move wrote the copy then deleted "the original" at the same path when a project relocation fell back to home — destroying the only copy	The move that isn't a move stops before writing
A save silently overwrote a different template that slugged onto the same filename	The second takes the next free name; same guard on move
Clicking a picker row applied whatever the keyboard was on	Applies the row that was clicked
"Save as Template…" from a card menu set state nothing presented	The canvas and the overview host the sheet, with a quiet notice bar for the path
A following composite re-identified its whole sub-graph on every pilot, orphaning the previous pass's child sessions (node ids are zmx session names) and stranding their memory	Compares what a human authored; replaces only on a real change, tearing down the outgoing children when it does

Fidelity to the spec

• Asking whether a project can take a template no longer creates .graphcode/templates in the checkout — nothing lands there by default.
• The use count counts uses, not saves.
• A resolve that flips a follow to missing broadcasts, so the warning reaches the card; an unchanged resolve doesn't, so the timer-driven sweeps don't persist the graph every tick.
• Templates carry an agent only when they name one, so a shared template doesn't pin an agent a teammate may not have installed.
• The relocation notice reports where the file actually landed, not where it was asked to go.
• A rename in the save sheet reaches the filename; a rename in the editor moves the file and keeps the id, so followers stay attached.
• The directory watch re-arms, so a git pull that creates .graphcode/templates is picked up.
• An unwritable project is dropped from the save sheet rather than an ineffectual .disabled on a Picker tag.
• Tokens left in a done check, metric or branch block Start too; the save notice is dismissible; ✕ restores the pre-template draft after two applies.
• The applied strip uses the canvas's own copy — one shared verb ("sets a done check and a worktree"), finishing with "Undo the shape to keep just the prompt."

The repo-wide cyclomatic_complexity relaxation the first draft added is dropped — the reducer is split into picker / apply / save switches instead, which is what the complexity was actually reporting.

The usage line

§ Follow vs snapshot calls this load-bearing rather than decorative, and it is: a committed edit to a project template changes what runs on a teammate's machine. TemplateUsage counts followers across every known project and at any depth inside composites, and the editor puts the line directly above Save. The other half is stated too — loops that snapshotted keep their brief — so "3 loops use this" is never read as "and the other four change too". Deleting a followed template says what actually happens rather than letting someone assume it stops a nightly run.

Note on the mock's field labels

The canvas's applied-state artboard labels the Goal fields "Prompt" and "Done when". README.md §3 specifies "What does done look like?" and "Done check", and PROMPT_TEMPLATES.md says the fields are "exactly as specced in README.md § 3" — so those are read as designer shorthand, not a rename, and the labels are unchanged. Worth a second opinion.

Verification

• xcodebuild -workspace graphcode.xcworkspace -scheme graphcode build — succeeds
• test — 1356 tests in 147 suites pass (up from 1328/144 on the first draft; 28 new)
• swiftlint lint — 0 errors, with the original .swiftlint.yml
• swift format lint --strict — clean on every file this branch touches

Not verified visually — this environment has no GUI automation, so the SwiftUI layout of the picker, the applied strip and the editor wants a human eye before merge.

🤖 Generated with Claude Code

https://claude.ai/code/session_01MJR8Wuwc4qnaSBUMcYf7Af