Architecture

Lagune architecture

This skill is the specialized, authoritative description of how Lagune is structured: repository layout, how commands and templates are organized, the core/adapter boundary, what Lagune scaffolds into a target project, and the tracking-map model. Consult it before changing source, before adding an agent, and whenever a decision depends on the shape of the codebase.

The product mission and workflow philosophy live in CLAUDE.md. The toolchain, code conventions, build path, and how the tracking hooks are implemented live in the engineering skill. The language conventions and the boundaries on what prose carries and who may author it live in the writer skill. This skill covers the shape, not the build.

Repository layout

Top-level directories. Their internal structure is defined in the sections below.

Build & ship

• Build tool: esbuild bundles src/ into self-contained JavaScript in lib/, so the published package carries no runtime node_modules for the end user.
• Output: lib/ holds the shipped JavaScript. It is generated, never edited by hand.
• Entry point: the package.json bin field maps the lagune command to a file in lib/ (with a node shebang), so npx lagune ... runs the bundle directly.
• End-user install: none. The bundle is self-contained, so running via npx needs no dependency install on the user's machine.

Website

The documentation site under website/ is an npm workspace of the root package, so one npm ci installs both and one lockfile covers them. The site never redeclares what the package already knows: a single bridge module in its data folder is the only website file that imports from src/, re-exporting the agent registry, the skill groups, the catalog, and the dashboard's shared presentation metadata, and every other website module reads from that bridge. Logos live once, in the dashboard's public assets folder, which the site serves as a second static directory. Counts in the docs are the build-time tokens {{agents}} and {{categories}}, substituted before MDX compiles and again when the plain-Markdown mirrors are written, and the supported-agents table is rendered from the registry in both outputs. The README holds the one hand-typed count, and a test keeps it equal to the registry.

Command & template anatomy

A phase of the workflow is a pair: a command and one or more templates.

• A command instructs. It is the agent's reading: how to think about the phase, what to look for, what questions to resolve. Commands live in spec/commands/, as markdown (.md) files with frontmatter, the format Claude Code consumes natively.
• A template structures. It is the mould for the phase's output: the security artifact the command produces, with its sections and the fields to fill in. Templates live in spec/templates/.

A phase command points to its template. The command carries the reasoning, the template carries the shape of the result. Splitting them keeps the "how to think" reusable and the "what to produce" consistent.

A command that fills no template has none, and the type layer encodes that: TemplateKey is CommandKey minus repair and verify. Verify records its verdicts on the harden record, so it fills no template of its own. specialize is the exception that is not a phase yet still pairs with a template (specialize-template.md): its artifact is a sub-skill file under .lagune/skills/, not a memory/ phase artifact.

Sub-skills

spec/skills/ holds the built-in sub-skills: agent-agnostic knowledge modules that load only on demand. They are not commands. Detect consumes them inline: it reads the catalog from .lagune/specializations.md (which lists each sub-skill as name: tags, suffixing [required] on any the catalog flags), then reads and follows the matching .lagune/skills/<name>.md directly. A user can import one into any prompt with @.lagune/skills/<name>.md. Adding a built-in is one .md plus one row in the baked catalog (src/hooks/skills/catalog.ts), no new command. A user grows the same set at runtime with /lagune.specialize, which writes a sub-skill into .lagune/skills/ and its row into .lagune/skills.json. .lagune/specializations.md is an internal, gitignored listing rebuilt from the catalog (by renderSpecializations, exposed as the specializations hook and rerun on every scaffold, pull, update, and skill change), merging the user's .lagune/skills.json over the baked catalog, the user's entry winning a name collision.

Agent-agnostic core vs. agent adapters

The codebase splits into two layers, and the split is what lets new agents be added later without rewriting the workflow.

• The core is agent-agnostic. Templates, the per-phase content, and the context-detection logic know nothing about which agent runs them. This is where the security value lives, and it is written once.
• An adapter is thin. Its only job is to translate the core's commands into the format and location a given agent expects. It carries no security logic of its own.

Adapters are data, not code. Each supported agent is a single entry in the agent registry under src/providers/, declaring its key, display name, homepage, logo, command format, target directory, and, when it applies, its deprecation (a successor or an end-of-life date). A factory turns each entry into a provider, and a format dispatcher under src/transform/ renders the same core command into that agent's packaging. Adding an agent is adding a row, not writing a module. The registry is the single source of truth for which agents are supported.

The command formats are:

• Skill (<dir>/lagune.<phase>/SKILL.md): a directory per command with name / description / argument-hint / user-invocable frontmatter. Skill directories use each agent's current, project-scoped location (for example .claude/skills, .agents/skills, .github/skills, .cursor/skills).
• Continue prompt (.continue/prompts/lagune.<phase>.md): a markdown prompt with name, description, and invokable: true frontmatter, the shape Continue lists under /.
• Markdown command (<dir>/lagune.<phase>.md): a single markdown file whose name becomes the command.
• Forge (.forge/commands/lagune.<phase>.md): a markdown command with name added to its frontmatter, swapping $ARGUMENTS for Forge's {{parameters}} placeholder.
• TOML (<dir>/lagune.<phase>.toml): description plus a multi-line prompt, with $ARGUMENTS rendered as {{args}}.
• Goose recipe (.goose/recipes/lagune.<phase>.yaml): a YAML recipe with the required version / title / description / instructions / prompt fields, with $ARGUMENTS rendered as {{ args }}.

Which agent uses which format, and its exact directory, lives in the registry (src/providers/specs.ts), the single source of truth. Consult it rather than duplicating the mapping here.

A future agent is one more row, leaving the core untouched. Keep adapters thin and the core neutral. Any logic that an adapter would duplicate belongs in the core instead.

What Lagune scaffolds into a target project

Running Lagune in a user's project creates two things:

• .lagune/ holds Lagune's state in that project: the filled-in charter and the artifacts each phase produces (the detect map, the defense plan, and so on), plus hooks/ (the compiled helper scripts the agent runs, see The tracking map and Memory validation) and skills/ (the non-invocable sub-skills the phases load on demand, both the built-ins copied at init and any the user adds with /lagune.specialize). Init also seeds skills.json, the runtime catalog of user sub-skills (empty until specialize writes to it), a sibling of tracking.json. The state is committed and reviewable like the user's code: the charter, the phase artifacts, the tracking, skills.json, and the manifest. The generated material (templates/, hooks/, the built-in skills/*, the agent commands) is gitignored and restored from the manifest by pull. A user sub-skill shares the skills/ directory but stays versioned: specialize re-includes it past the /.lagune/skills/* rule.
• Agent commands are written by the adapter into the agent's own location, in that agent's native format (a skill directory, a markdown command, a TOML command, or a Goose recipe, per the chosen agent's registry entry). For example, Claude Code reads them from .claude/skills/, GitHub Copilot from .github/skills/, and opencode from .opencode/commands/. All hold the /lagune.* commands the user invokes. Across agents the commands are reused, not duplicated: whichever agent came first holds the real file, and every later agent whose rendered command is identical gets a relative symlink to it, while a different rendering keeps its own file. Where a symlink cannot be created, the install falls back to a real copy, and update and pull re-establish the layout. A recorded command that the current version no longer produces is removed on update, so a changed format leaves nothing behind.

The split mirrors the repository's own core/adapter boundary: .lagune/ is the agent-agnostic state, and the agent's own directory is where the adapter places what that specific agent reads.