Make AGENTS.md the single instruction file; add Gemini pointer, drop Cursor/Windsurf duplication (closes #145) (#149)

AGENTS.md is Launchpad's one agent-instruction contract. Projections must never
duplicate it - they only add what a tool needs to read it.

- Gemini CLI: new opt-in `gemini` projection writes a minimal
  .gemini/settings.json that sets context.fileName to AGENTS.md, so Gemini reads
  the canonical file instead of a separate GEMINI.md. Adds a CONFIG file kind
  for vendor config that points at canonical output.
- Cursor and Windsurf: removed their rule-file projections. Both read a root
  AGENTS.md natively (Cursor as always-on rules; Windsurf/Cascade via the same
  rules engine as .windsurf/rules), so the generated .cursor/rules and
  .windsurf/rules files only duplicated the contract and could drift. Nothing is
  emitted for them now. The Cursor MCP-client integration is untouched.

No engine change; projections remain auto-discovered AgentProjection beans.
Updates docs and changelog.

Author: acltabontabon

CHANGELOG.md

@@ -7,6 +7,7 @@ and this project adheres to [Semantic Versioning](https://semver.org/spec/v2.0.0
 
 ## [Unreleased]
 ### Added
+- **Gemini CLI compatibility projection**: An opt-in `gemini` `AgentProjection` points Gemini CLI at the canonical `AGENTS.md` instead of emitting a redundant context file. It writes a minimal `.gemini/settings.json` setting `context.fileName` to `AGENTS.md`, so Gemini reads the single source-of-truth contract with no duplicated content to drift out of sync. An existing hand-authored `.gemini/settings.json` is never clobbered. Adds a `CONFIG` file kind for vendor files that point at canonical output. Enable it via `projections: ["claude", "gemini"]` in the standards manifest or by ticking Gemini CLI in the TUI projection picker; no engine change was required (closes #145).
 - **Anthropic (Claude) cloud preparation provider**: An opt-in, paid synthesis backend on the `PreparationProvider` SPI (`provider=anthropic`), shipped via the Spring AI Anthropic starter. It is explicit-selection only - `autoDetectable()` is `false`, so auto-detection never routes preparation to a paid API even when a key is present; local AI stays the default. Authentication is vendor-specific (`launchpad.ai.anthropic.api-key`, bound from `LAUNCHPAD_ANTHROPIC_API_KEY`, falling back to the shared key), with its own cloud base URL and `anthropic-version`. A missing key or an unreachable endpoint degrades cleanly to deterministic-only output instead of crashing. Adding it was a single new bean plus its config - no router or health-checker changes (toward #144).
 - **Project readiness evaluator**: `ReadinessEvaluator.evaluate(Path)` returns a read-only `ReadinessResult` (status, structured reason lines, recommended action, last-prepared timestamp) derived from a project's prepared artifacts (`AGENTS.md` and the `.launchpad/*.json` sidecars) and its provenance stamp - the deterministic verdict the readiness dashboard renders. `ProvenanceHeader` gains a `parse(...)` inverse of `render()` (closes #122).
 - **Output contract docs**: `docs/output-contract.adoc` documents every file a scan writes - layout, section structure, provenance stamp, write/merge actions, stable-vs-experimental tiers, and versioning - so downstream tools integrate against a contract instead of snapshot-testing prose. Adds a `docs/index.adoc` landing page (closes #91).
@@ -42,6 +43,7 @@ and this project adheres to [Semantic Versioning](https://semver.org/spec/v2.0.0
 - **Manifest now required for a pack**: A standards source resolves only with a `standards-pack.yml` declaring a supported `schemaVersion` (closes #84).
 
 ### Removed
+- **Cursor and Windsurf rule-file projections**: Dropped the `cursor` (`.cursor/rules/*.mdc`) and `windsurf` (`.windsurf/rules/*.md`) projections. Both tools read a root `AGENTS.md` natively (Cursor applies it as always-on rules; Windsurf/Cascade discovers it through the same rules engine as `.windsurf/rules`), so those generated files only duplicated the canonical contract and could drift out of sync. `AGENTS.md` is the single agent-instruction file; nothing is emitted for Cursor or Windsurf. The Cursor MCP-client integration is unaffected.
 - **Redundant Renderers**: Removed `BuildProfilesRenderer` and duplicate synthesis paths.
 - **Legacy Artifacts**: Removed obsolete planning markers and redundant tests.
 - **Legacy flat-file standards mode**: Dropped the manifest-less fallback that read bare `rules.yml` / `skills.yml` from the root of a standards source. Every pack is now a versioned manifest, so there is no second, unversioned load path (closes #84).

docs/architecture.adoc

@@ -22,7 +22,7 @@ This ordering is deliberate. Deterministic analysis is fast, repeatable, and doe
 
 === Vendor-neutral output
 
-Generated context files use a standard shape (`AGENTS.md` + `.ai/`) that any AI assistant can read. Agent-specific formats (Claude Code skills, Cursor rules, Windsurf rules) are produced as optional projections on top of the canonical output, not as replacements for it.
+Generated context files use a standard shape (`AGENTS.md` + `.ai/`) that any AI assistant can read. `AGENTS.md` is the single agent-instruction contract; projections never duplicate it. Tools that read `AGENTS.md` natively (Cursor, Windsurf) get no generated file at all. Optional projections add only what a tool needs on top of the canonical output -- Claude Code slash-command skills, or a `.gemini/settings.json` that points Gemini CLI at `AGENTS.md` -- never a replacement for it.
 
 The standards companions (`.ai/engineering-rules.md`, `.ai/skills.md`, `.ai/checklists.md`) emit one stable-slug heading per rule, skill, and checklist -- each anchor derived from the record's stable id, not its title -- so downstream sandbox indexers chunk the corpus one item at a time. `AGENTS.md` itself stays standards-free.
 

docs/getting-started.adoc

@@ -98,7 +98,6 @@ Press `s` to apply all plans. Launchpad backs up any file it changes to `.launch
 │   ├── checklists.md                  # checklists from your standards pack
 │   └── index.md                       # index of all .ai/ files
 ├── .claude/skills/<id>/SKILL.md       # Claude Code slash-command stubs (one per skill)
-├── .cursor/rules/<id>.mdc             # Cursor rules (if Cursor projection is enabled)
 └── .launchpad/
     ├── scan.json                      # structured scan cache (read by MCP server and audit)
     ├── audit.md                       # audit findings (if standards pack has check: rules)

docs/output-contract.adoc

@@ -28,18 +28,16 @@ tier (see <<Stability tiers>>).
 ----
 <project-root>/
 ├── AGENTS.md                       # primary agent-instructions file        [stable layout / experimental prose]
+├── .gemini/
+│   └── settings.json              # points Gemini CLI at AGENTS.md (opt-in)  [stable]
 ├── .ai/
 │   ├── index.md                    # navigation index + skills list         [stable]
 │   ├── stack.md                    # detected stack and dependencies        [stable]
 │   ├── engineering-rules.md        # rules (only if rules exist)            [stable]
 │   ├── skills.md                   # skills (only if skills exist)          [stable]
 │   └── checklists.md               # checklists (only if checklists exist)  [stable]
 ├── .claude/
-│   └── skills/<skill-id>/SKILL.md  # Claude Code projection (default-on)     [stable]
-├── .cursor/
-│   └── rules/*.mdc                 # Cursor projection (opt-in)              [stable]
-├── .windsurf/
-│   └── rules/*.md                  # Windsurf projection (opt-in)            [stable]
+│   └── skills/<skill-id>/SKILL.md  # Claude Code slash commands (default-on) [stable]
 └── .launchpad/
     ├── project.model.json          # deterministic structure graph          [stable, versioned]
     ├── standards.index.json        # deterministic standards records        [stable, versioned]
@@ -50,9 +48,17 @@ tier (see <<Stability tiers>>).
 
 Conditional companions (`engineering-rules.md`, `skills.md`, `checklists.md`)
 are written only when the resolved standards pack contributes that record type.
-The projection directories appear only for enabled projections - `claude` is
-on by default; `cursor` and `windsurf` are opt-in via the standards manifest's
-`projections` list.
+The projection outputs appear only for enabled projections - `claude` is on by
+default and `gemini` is opt-in, via the standards manifest's `projections` list
+or the TUI projection picker.
+
+`AGENTS.md` is the single agent-instruction contract; projections never
+duplicate it. Cursor and Windsurf read `AGENTS.md` natively (root = always-on
+rules), so Launchpad generates no vendor instruction files for them. Gemini CLI
+defaults to a different file name but is configurable, so the `gemini` projection
+emits only a `.gemini/settings.json` that points `context.fileName` at
+`AGENTS.md`. The `claude` projection emits Claude Code slash commands, which are
+invokable workflows rather than an instruction file.
 
 == Generated files
 
@@ -67,12 +73,43 @@ on by default; `cursor` and `windsurf` are opt-in via the standards manifest's
 | `.ai/skills.md`         | `SKILLS`  | deterministic (from pack) | only if skills exist
 | `.ai/checklists.md`     | `CONTEXT` | deterministic (from pack) | only if checklists exist
 | `.claude/skills/<id>/SKILL.md` | `SKILL` | deterministic projection | per skill (default-on)
-| `.cursor/rules/*.mdc`   | `RULES` / `SKILL` | deterministic projection | opt-in
-| `.windsurf/rules/*.md`  | `RULES` / `SKILL` | deterministic projection | opt-in
+| `.gemini/settings.json` | `CONFIG`  | deterministic projection | opt-in
 |===
 
+=== Cursor and Windsurf
+
+Neither tool gets a generated file. Cursor reads a root `AGENTS.md` natively
+(applied as always-on rules, alongside any `.cursor/rules`), and Windsurf/Cascade
+discovers a root `AGENTS.md` through the same rules engine that powers
+`.windsurf/rules`. Because both already consume the canonical contract, Launchpad
+emits nothing for them - `AGENTS.md` is the single source of truth.
+
+=== Gemini CLI projection
+
+The `gemini` projection does *not* emit a separate context file. Gemini CLI
+loads its project context from a configurable file name (default `GEMINI.md`),
+so rather than duplicate `AGENTS.md` into a `GEMINI.md` that would clutter the
+repo and drift out of sync, the projection writes a minimal
+`.gemini/settings.json` that sets `context.fileName` to `AGENTS.md`:
+
+[source,json]
+----
+{
+  "context": {
+    "fileName": "AGENTS.md"
+  }
+}
+----
+
+Gemini CLI then reads the single source-of-truth contract directly. The settings
+file carries no JSON comment markers, so an existing hand-authored
+`.gemini/settings.json` is classified `SKIP` and never clobbered - developers who
+already maintain that file add `context.fileName: "AGENTS.md"` themselves. Enable
+the projection with `projections: ["claude", "gemini"]` in the standards manifest
+or by ticking Gemini CLI in the TUI projection picker.
+
 `Kind` is the stable `FileKind` discriminator carried on every generated file
-(`CONTEXT`, `RULES`, `SKILL`, `SKILLS`, `INDEX`). "Deterministic" means the
+(`CONTEXT`, `RULES`, `SKILL`, `SKILLS`, `INDEX`, `CONFIG`). "Deterministic" means the
 content is assembled from the scan and standards pack with no model call;
 "AI-synthesized" means a local model authored the prose (with a deterministic
 fallback when synthesis is disabled).
@@ -158,8 +195,8 @@ rely on this decision table:
 *Stable* - safe to integrate against; changes here follow the versioning policy
 below:
 
-* File paths and directory layout (`AGENTS.md`, `.ai/*`, `.claude/`, `.cursor/`,
-  `.windsurf/`, `.launchpad/*`).
+* File paths and directory layout (`AGENTS.md`, `.ai/*`, `.claude/`,
+  `.gemini/settings.json`, `.launchpad/*`).
 * The `launchpad:provenance` and `launchpad:managed:start/end` marker tokens.
 * The `FileKind` discriminator and the write-action decision table.
 * `.ai/*` heading anchors (explicit `{#slug}` anchors derived from each record's

docs/standards-packs.adoc

@@ -71,14 +71,15 @@ includes:
 
 projections:
   - claude
-  - cursor
+  - gemini
 ----
 
-The `projections` field controls which agent-native formats are emitted alongside `AGENTS.md`:
+The `projections` field controls which agent-native outputs are emitted alongside `AGENTS.md`. Projections never duplicate `AGENTS.md`; they only add what a tool needs to consume it:
 
 * `claude` -- `.claude/skills/<id>/SKILL.md` for Claude Code slash-command discovery
-* `cursor` -- `.cursor/rules/<id>.mdc`
-* `windsurf` -- `.windsurfrules`
+* `gemini` -- `.gemini/settings.json` that points Gemini CLI's `context.fileName` at `AGENTS.md`
+
+Cursor and Windsurf are not listed: both read a root `AGENTS.md` natively, so no projection is needed for them.
 
 Omit `projections` to use the default (`claude` only). Set it to `[]` to disable all projections.
 
@@ -305,7 +306,7 @@ When adding a new HTTP route or extending the API surface.
 
 Anchor slugs must be unique: two ids that collapse to the same slug (e.g. `foo.bar` and `foo-bar`) are an authoring conflict and fail the scan, rather than being silently de-duplicated in the output.
 
-Agent projections (`.claude/skills/`, `.cursor/rules/`, `.windsurf/`) are generated from the same source -- editing the YAML updates all outputs on the next scan.
+Agent projections (`.claude/skills/`, `.gemini/settings.json`) are generated from the same source -- editing the YAML updates all outputs on the next scan.
 
 == Machine-readable sidecar
 

src/main/java/com/acltabontabon/launchpad/template/GeneratedFile.java

@@ -11,7 +11,8 @@ public enum FileKind {
         RULES,      // engineering rules files
         SKILL,      // one skill, vendor-native projection file (e.g. .claude/skills/<id>/SKILL.md)
         SKILLS,     // canonical aggregated skills companion (.ai/skills.md)
-        INDEX       // .ai/index.md, directory indexes
+        INDEX,      // .ai/index.md, directory indexes
+        CONFIG      // vendor tool config that points the tool at canonical output (e.g. .gemini/settings.json)
     }
 
     public String filename() {