Problem. Projection data had loose or inconsistent matrix contracts: renderer uniforms, custom layers, and terrain render-to-texture paths did not clearly distinguish 32-bit and 64-bit backing arrays. The custom-layer documentation also referenced a non-existent property.

Solution. Added explicit Mat4f32 / Mat4f64 aliases, made projection data generic, split renderer and custom-layer contracts, initialized the terrain RTT tile matrix as Float32Array, and corrected the docs reference.

Result. A 16-file change with typecheck, targeted projection/WebGL/terrain tests, and git diff --check validation. The public rendering API now communicates matrix precision instead of leaving it implicit.

Problem. Google Translate could wrap React-managed text nodes in <font> elements. During API-key creation, React then attempted to remove a node that the translator had already replaced, causing NotFoundError: Failed to execute 'removeChild' on 'Node' and a false 500 page while the backend still returned HTTP 200.

Solution. Added notranslate metadata to both frontend entry points and marked the React root with translate="no" / notranslate. The project already has i18next, so browser translation is not required for the app shell.

Result. The reproduced failure path completed successfully with the guard in place; the change stayed focused to the two HTML entry points.

Problem. CodeIsland could not install or manage extensions for the pi and Oh My Pi coding-agent runtimes, leaving those agent sessions outside its normal lifecycle controls.

Solution. Added both runtimes as installer targets, installed bundled TypeScript extensions into their global directories, wired status/enable/disable/uninstall/repair flows, added version markers for stale installs, and covered the installer behavior with tests.

Result. A first-class integration rather than a one-off copy step. The PR also included typechecks against both agent runtimes and a live smoke test confirming a recorded source: "pi" session.

Problem. Multi-agent teams need predictable model selection across different execution paths, but routing decisions were not represented as one explicit, reusable policy.

Solution. Added an opt-in routing policy that can match phase, agent, task role, priority, leaf status, and dependency status without mutating team or agent configuration. Worker, delegated, short-circuit, coordinator, and synthesis calls use the same policy shape.

Result. Deterministic routing became a testable orchestration concern instead of scattered call-site behavior; the PR added focused orchestrator, agent-pool, and task utility coverage.

Problem. A plan-only run could describe a team execution, but there was no durable artifact that could be inspected and replayed without asking the coordinator to plan again.

Solution. Added a serializable plan artifact and a replay path that preserves task IDs, descriptions, assignees, dependencies, and memory scope, reusing the explicit-task execution path.

Result. Team execution can be reviewed, persisted, and reproduced with deterministic inputs; the change included orchestrator regression coverage.

Problem. The public SharedMemory boundary accepted strings only, forcing agents to flatten structured handoffs even though coordination data often contains objects and arrays.

Solution. Accepted JSON-serializable values at the public API, kept the storage boundary string-based with JSON metadata, parsed structured values on read/list, rendered them clearly in summaries, and added optional Zod validation.

Result. Agent handoffs can carry structured data without weakening the persistence boundary; the PR added dedicated shared-memory tests.

Problem. The MCP memory graph initially showed only a subset of documents, but the UI did not make the subset/total relationship clear or provide a way to explore the remaining graph.

Solution. Added pagination metadata, changed the summary to showing N of total documents, added a Load more action, and reused the existing app-only graph fetch tool. The initial page remains capped at 10 documents.

Result. Large memory graphs load incrementally instead of blocking the first view; structured-content and summary wording received regression coverage.

Problem. Research Hub could publish resources, but it lacked a shared workflow for creating, editing, browsing, and viewing knowledge graphs.

Solution. Added the graph data model and Supabase schema, CRUD API routes, profile publishing flow, list/detail/editor pages, a Transformer demo graph, and focused service/form tests.

Result. The product gained an end-to-end knowledge-graph surface rather than a static demo; the PR passed npm test, npm run lint, and npm run build.