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.

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, command format, and target directory. 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, .codex/skills, .agent/skills, .cursor/skills).
• Copilot prompt file (.github/prompts/lagune.<phase>.prompt.md): GitHub Copilot's prompt-file format, invoked as /lagune.<phase>.
• Markdown command (<dir>/lagune.<phase>.md): a single markdown file whose name becomes the command.
• Forge (.forge/commands/lagune.<phase>.md): a markdown command that swaps $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 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 prompt file, 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/prompts/, and opencode from .opencode/commands/. All hold the /lagune.* commands the user invokes.

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.

The tracking map

Each detect finding is one tracked item. The later phases carry that same item forward: it is a single entry that plan and harden re-report by name as they act on it, never a separate entry per phase. Its identity is its name (the finding's name, written identically as the section title in detect.md, plan.md, and harden.md), and the map keys on that name and nothing else. The prose artifacts carry only that shared title and never a path. The file paths are the one volatile thing, and they live only in the tracking map. A rename or a moved path is corrected in one place, and every memory stays valid. The tracking map is tracking only: no prose, no note, no separate id, no phase.

• Where it lives: .lagune/tracking.json, a sibling of manifest.json and memory/. It is internal state, committed but never hand-edited and never shown to the user.
• Shape: { name: 'lagune', entries: [{ name, paths }] }, where name is the identity and paths is string[] (charter has no items). There is no cross-entry link: the conveyor lives in the prose (which artifact carries the item's section), not in the map. Types live in the types module under src/types/.
• Two flow acts, plus repair apart: the map is maintained by two acts the phases perform in the normal flow, track (registration) and untrack (stand-down), and by repair (realignment), a maintenance act that runs only via /lagune.repair, never as part of a phase. Track records each finding a phase reports and re-reports an existing item by name as a later phase writes, following a renamed path and never removing. By convention detect is where a finding is born and plan and harden re-report it, but the hook keys on the name alone and does not enforce the order. Untrack stands a proven-closed finding down: verify hands it the name, and the hook removes that finding's section from the prose artifacts and drops its entry from the map in one pass. Repair realigns the whole map against the artifacts and the code, and is the only act that surfaces what is gone. Without tracking there is nothing to repair, so track is what populates the map along the normal flow.
• Who calls them: detect, plan, and harden call the track hook at the end of their write step. Detect registers the findings it wrote, and plan and harden re-report each item by the finding's name (additive, never removing). Only /lagune.verify calls the untrack hook, handing it a proven-closed finding's name, which is why verify alone writes across the other phases' artifacts. Only the /lagune.repair command calls the repair hook, to realign the whole map when something diverged. The phases never repair the tracking themselves: when one notices the tracking is inconsistent (a tracked path that no longer exists), it runs /lagune.repair and continues. There is no user-facing track or untrack command, both are internal plumbing.
• Non-goals: the hooks are mechanical and read nothing but the snippets they are handed. The /lagune.repair command reads the user's source only to learn a renamed file's new path, never to author or judge security. Only untrack edits the .lagune/memory/*.md prose, and only to stand a proven-closed finding down: it injects no IDs or anchors and authors no security content. The one call no hook makes alone (fixed vs renamed) is returned for the agent or user to decide, never guessed.