Test prompt building

[nrodd]

Chunk 3 of the test plan. src/prompt/build.test.ts, 49 tests, plus one source change.

Based on #42 (the harness). GitHub will retarget this to main when that merges. Independent of #43 and #44 — different files.

Source change

AGENT_STEPS and METADATA_FIELDS are now exported from telemetry-marker.ts, for the contract test below. Nothing else moved.

Why this chunk matters

These prompts are the product. Everything the wizard does ends with one of them being handed to an agent that auto-accepts file edits. So the two invariants are worth more than any amount of wording coverage:

1. No placeholder survives rendering

expect(prompt).not.toMatch(/\{\{/) across all six variants. A template that gains a placeholder nobody fills currently ships a literal {{PLAN_GATE}} to the agent, and nothing catches it.

2. The prompt matches what the parser will accept

The prompt tells the agent which steps and metadata fields to emit; telemetry-marker.ts decides which ones survive sanitization. Nothing linked those two lists — a field added on one side alone is dropped silently and the funnel just loses it. The comments in both files ask for exactly this check.

The test parses the rendered stdout table and checks every step against AGENT_STEPS and every field against METADATA_FIELDS, so it verifies the contract as delivered to the agent rather than as declared in a const.

The rest

• Per-phase report filenames — both headless runs write a report into the same directory, so a shared name would have the enrich pass overwrite the install report the outro points at.
• Gates — headless blanks the plan gate and swaps the identity/privacy/explain wording; interactive keeps "wait for approval".
• stdout transport — names the marker prefix, forbids network calls for telemetry, and never asks for start/complete, which the parser rejects by design.
• MCP transport — points at the telemetry-event tool, doesn't re-ask for consent, and carries the complete row only for the snippet phase (the enrich phase is never driven over MCP).
• Integrations — the detect-it-yourself fallback, package and window.* hints, the free-text "Other" path, and the three-space indent that keeps linkage examples inside their fence.

A note on snapshots

The plan called for full-prompt snapshots. I skipped them: six prompts is ~1000 lines of snapshot that would bury the 49 assertions and get approved unread. Structure and contract are asserted explicitly instead. Easy to add later if wording drift turns out to be a real problem.

Verification

49 tests passing (102 across the suite), typecheck clean. Mutation-checked both invariants — appended a {{BOGUS_PLACEHOLDER}} to a template and added an unlisted metadata field to STEP_META — and confirmed exactly the right tests fail, then reverted.

---

Note

Low Risk
Test-only coverage plus exporting two existing constants; no runtime behavior change in prompt building or marker parsing.

Overview
Adds src/prompt/build.test.ts (~49 tests) that lock in wizard prompt output across all six snippet/enrich × headless/interactive × telemetry variants.

Exports AGENT_STEPS and METADATA_FIELDS from telemetry-marker.ts so tests can assert the stdout telemetry table in rendered prompts only names steps and metadata fields the marker parser will accept—closing a silent drift risk between build.ts and sanitization.

Coverage highlights: no unfilled {{ placeholders; formatting/newlines; phase-specific steps and report paths (subtext-setup-report.md vs subtext-enrich-report.md); headless vs interactive approval gates; stdout vs MCP telemetry wording (prefix, no start/complete on stdout, MCP complete only on snippet); enrich integrations and indented linkage examples.

^{Reviewed by Cursor Bugbot for commit f97c773. Bugbot is set up for automated code reviews on this repo. Configure here.}