Architecture

fullstack-expert is a static Cordis plugin mounted into an existing Harness composition. It is not a second agent loop.

Planes

• Host-native capabilities: filesystem, shell, jobs, sandbox, approval, session persistence, model routing, and browser support remain owned by Harness rows.
• Agent/preset contribution: this plugin adds one scoped methodology section, one dynamic state context, embedded skills, two tools, and an optional monotonic guard.
• Private runtime seam: FullstackExpertRuntime owns deterministic plan/check state and keeps only scalar projections. It does not serialize live Services, Sessions, Agents, Tool executions, or credentials.

Interfaces

• inspectProject(fs, options) — bounded read-only ProjectModel builder over ctx.fs.
• createPlan(input) — pure request-to-vertical-slice planner.
• checkProject(fs, input) — bounded read-only checks with pass/warn/unknown results.
• applyEvidence(state, evidence) — immutable evidence transition; VERIFIED requires a passing verification record.
• fullstack_plan / fullstack_check — small model-facing interfaces; existing native tools perform edits, commands, browser actions, and approvals.

The plugin observes session/event and tools/result for compact facts, but the native session log remains the durable source of truth. There is no competing event bus or persistence format.

Composition

The supplied cordis.patch.yml inserts the plugin row after the profile's base bundle. End users can install that layer with dsh plugin --profile <name> add @deepseek-ai/fullstack-expert; see the quickstart. It assumes dsh-tools, dsh-skill, dsh-system-prompt, and an optional dsh-fs implementation already exist. If the filesystem capability is absent, the plugin still mounts but plan/check tools fail closed with an actionable error.