Built @core/postman, a read-only Postman integration with personal-API-key auth. Runs an incremental sync every 15 minutes across workspaces, collections, environments, APIs and monitor runs (deduped via lastSyncTime + seen-run IDs), and exposes 14 read-only MCP tools over the Postman REST API.

Exposed user-defined CORE skills as MCP tools so coding agents — Claude Code, Codex, Cursor — can discover and load skill instructions on demand instead of pasting them in. Added list_skills (paginated) and get_skill (by id or slugified name) in a new app/utils/mcp/skill.ts module that mirrors the existing MCP modules.

Made self-hosted local Ollama work end-to-end. Fixed four defects: the /api/chat base URL, provider inference short-circuiting to Ollama under CHAT_PROVIDER=ollama (including gpt-prefixed ids), a direct model build for the main conversation agent, and JSON-Schema injection so structured extraction validates on the first pass. Hosted OpenAI / Anthropic / Google paths were left untouched.

Closed an audit gap: integration actions triggered over the API now record who triggered them. POST /api/v1/integration_account/:id/action accepts an optional source (max 128 chars), threaded through executeIntegrationAction into every IntegrationCallLog row, and custom MCP executions now write a call-log row on success and failure like first-party integrations already did.

Fixed blank chat messages on multi-step agent turns. message.parts.find(type === 'text') was grabbing the empty text part emitted for the tool-use step; replaced it with filter + map + join to concatenate every non-empty text part before handing it to the TipTap editor.

Fixed API-key deletion always removing the last key. useTokensColumns shared one open/setOpen state across every row dialog, so all dialogs opened at once and portal stacking left only the last visible — always deleting the last token. Moved visibility into each row's own dialog and switched to DialogClose.

One-line fix to Ghost publishing: append ?source=html to the Admin API request so Ghost accepts HTML-generated post content. Reproduced the failure locally and verified the fix against the Ghost Admin API docs.

Aligned the CLI gateway banners with the redesigned UI. Updated gateway register, env-bootstrap, the Railway and Docker setup flows, and the Tauri app to point users at Sidebar → Gateways → New Gateway.

Added 62 unit tests across three files for previously untested pure-logic agent helpers — parseActionPlan / normalizeActionPlan / createFallbackPlan (30), truncateString / truncateToolResult (20), and prepareHistoryParts (12). No DB or network, so they stay fast and deterministic.

Closed in favour of #73 — "all the commits are in that PR". All five of my commits (e0e7f3f…b6d43d0) are the first five of #73's six, merged to main on 27 Jul 2026.

Let a signed-in visitor lend an NPC their own CORE integrations for a conversation, so the NPC acts on the visitor's account instead of the owner's — without the owner ever seeing the visitor's credentials. Adds an in-chat consent popover, a soft-deletable VisitorIntegrationGrant table, and per-person token routing inside buildNpcTools.

Open. The maintainer clarified mid-review that the issue meant visitor-granted access, not owner-granted — so I built that instead as #72. This stayed open as an optional owner-side convenience panel.

An owner-only ACCESS panel that grants specific CORE integrations — and optionally specific actions within them — to individual NPCs, so characters can act on the resident's connected tools during chat. Ships with an owner_only guardrail, a TownActivity audit row per executed action, and aura pricing.

Added Ollama as a third LLM provider alongside Anthropic and OpenAI, via Ollama Cloud's OpenAI-compatible endpoint, so Town can run on a free or lower-cost model without a paid key. Provider priority stays anthropic → openai → ollama, so existing deployments are unaffected.

The companion to #37 on the CLI side: pickModel() gets an ollama branch, and town test set-key ollama <key> persists to ~/.town/config.json alongside the existing keys — so contributors can test NPC prompts without a paid key.

One line: set includeUsage: true on the Ollama provider. OpenAI-compatible APIs only report token usage on streamed responses when stream_options.include_usage is sent, so without it every streamed chat on an Ollama deploy logged 0/0 and debited no aura.

Open — rebased on main at the maintainer's request; tests and typecheck green.

Added a test runner to a repo that had none, plus 61 unit tests over three pure, security-relevant modules: prompt sanitisation, town-code parsing, and HMAC-signed event envelopes. The setup deliberately mirrors RedPlanetHQ/core's, package versions included, so both repos stay consistent.

The setup guide told contributors to run cp .env.example .env, but the file didn't exist — the repo's .env* ignore rule also excluded .env.example. Fixed .gitignore and added a root and a hosting/ example covering every variable the app and docker-compose actually read.