feat(agent): BetterHarness — core skills + slim system prompt

Extract all per-tool documentation from the system prompt into 6
embedded Core Skill files (file-access, memory, plans, tasks,
rules-skills, harness). System prompt shrinks from ~470 to ~240 lines;
tool docs are now served lazily via `skills_read`.

- Add `SkillSourceKind::Core` variant (backend + frontend wire types)
- `list_skills` prepends 6 always-present core skills; `read_skill`
  serves embedded content; `remove_skill` rejects core names
- Skills tab: Core/User sub-tabs; "Install" button hidden in Core view
- SkillCard: "core" badge + remove button suppressed for core skills
- All 13 locale files updated with SrSkillsTabCore/User/SrSourceCore

Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>

Author: Maik Roland Damm

.agents/plans/PLANS.md

@@ -13,3 +13,4 @@ Persistent plans for multi-step work on **blxcode**. Individual plans live as Ma
 | done | [terminal-agent-context-handoff.md](terminal-agent-context-handoff.md) | BLXCode Agent kann aktuellen Kontext gezielt per Toolcall an Terminal-Agent-Sessions uebergeben; Bilder werden als Metadaten plus lokale Pfade exportiert |
 | planned | [plan-manager.md](plan-manager.md) | Plan Manager fuer `.agents/plans`: Manage-Tab wie Memory Files, plan-linked Tasks gruppiert, Agent-Toolcalls, System-Prompt-Update und shared Context fuer BLXCode Agent plus Terminal-Handoff |
 | planned | [kanban-board-view.md](kanban-board-view.md) | Kanban-View im Plans-Panel fuer alle Plan-Tasks eines Workspaces: Status-Spalten, Drag-and-Drop fuer Karten und Spalten, Spalten ein-/ausblenden und Markdown-Writeback |
+| in-progress | [better-harness.md](better-harness.md) | BetterHarness: Shrink system prompt by extracting tool docs into 6 embedded Core Skills; Skills tab gets Core/User sub-tabs |

.agents/plans/better-harness.md

@@ -0,0 +1,89 @@
+# BetterHarness: Core Harness Skills + Slim System Prompt
+
+## Context
+
+The current system prompt (`system_prompt.rs`) is ~470 lines and embeds full documentation for 50+ tools inline. This bloats every API request, consuming tokens and making prompt maintenance fragile. The goal is to extract tool documentation into **Core Skills** — built-in Markdown skill files shipped with the app — and reduce the system prompt to a minimal checklist + compact tool-name index. The Skills panel UI is extended to separate Core (built-in) from User (installed) skills.
+
+---
+
+## Overview of Changes
+
+### 1. Core Harness Skill Files (new embedded assets)
+Create `src-tauri/src/agent/harness_skills/` with 6 Markdown skill files (embedded via `include_str!`):
+
+| File | Covers |
+|---|---|
+| `memory.md` | `memory_*`, `memory_context_*`, `image_context_*` |
+| `plans.md` | `plan_*`, `plan_context_*` |
+| `tasks.md` | `task_*` |
+| `rules-skills.md` | `rules_*`, `skills_*` |
+| `file-access.md` | `list_workspace_files`, `read_workspace_file`, `list_tools` |
+| `harness.md` | `harness.*` client tools |
+
+Each file contains usage guidance, parameter notes, and patterns — exactly as a user-authored `SKILL.md` would.
+
+### 2. Backend: Add `Core` skill source + load harness skills
+
+**`src-tauri/src/skills_rules/types.rs`**
+- Add `Core` variant to `SkillSourceKind` enum
+
+**`src-tauri/src/skills_rules/store.rs`**
+- Add `CORE_SKILLS: &[(&str, &str)]` array — pairs of `(name, content)` using `include_str!`
+- `list_skills()` prepends core skills into the result (enabled state from workspace index, default `true`)
+- `read_skill()` serves core skill content from the embedded array
+- Core skills cannot be removed (error on `remove_skill` for core names)
+- `set_skill_enabled` works for both core and user skills (persisted in workspace index)
+
+**`src-tauri/src/agent/system_prompt.rs`** — major reduction (~470 → ~240 lines):
+- Remove all per-tool documentation sections
+- Keep: workspace scope, security policy, turn checklist
+- Replace tool catalog with compact grouped name list (~25 lines)
+- Add: *"Call `skills_read` with a core skill name to get full usage guidance"*
+
+**`src/skills_rules_wire.rs`**
+- Add `Core` variant to `SkillSourceKind` (mirrors backend type)
+
+### 3. Frontend: Skills Tab — Core / User tabs
+
+**`src/workbench/skills_rules_panel/skills_tab.rs`**
+- Add `SkillsView` signal: `Core | User` (default: `Core`)
+- Render a two-button sub-tab strip below the header
+- Filter displayed skills by view; "Install skill" button only visible in User view
+
+**`src/workbench/skills_rules_panel/skill_card.rs`**
+- Add `"core"` badge with `blx-sr-card__badge--core` CSS class
+- Remove button suppressed for core skills (`is_core` flag)
+
+**I18n strings added** (all 13 locale files):
+- `SrSkillsTabCore` — "Core" / locale equivalents
+- `SrSkillsTabUser` — "User" / locale equivalents
+- `SrSourceCore` — "core"
+
+---
+
+## Critical Files
+
+| File | Change |
+|---|---|
+| `src-tauri/src/skills_rules/types.rs` | Add `Core` to `SkillSourceKind` |
+| `src-tauri/src/skills_rules/store.rs` | Embed + serve core skills, merge with index |
+| `src-tauri/src/skills_rules/install.rs` | Add `Core` arm to install match |
+| `src-tauri/src/agent/system_prompt.rs` | Reduce to ~240 lines; add core skill reference |
+| `src-tauri/src/agent/harness_skills/*.md` | 6 new embedded skill files |
+| `src/skills_rules_wire.rs` | Add `Core` to `SkillSourceKind` |
+| `src/workbench/skills_rules_panel/skills_tab.rs` | Core/User sub-tab strip + filtering |
+| `src/workbench/skills_rules_panel/skill_card.rs` | Core badge + no-remove guard |
+| `src/workbench/skills_rules_panel/install_dialog.rs` | `Core` arm in source kind matches |
+| `src/i18n/locales/*.rs` (all 13 locales) | 3 new I18n keys |
+
+## Tasks
+
+- [x] `create-skill-files` - Create 6 harness skill .md files in src-tauri/src/agent/harness_skills/
+- [x] `extend-source-kind` - Add Core variant to SkillSourceKind (backend + frontend wire)
+- [x] `update-store` - Embed + serve core skills, guard remove, enable/disable
+- [x] `slim-system-prompt` - Replace inline tool docs with compact list + core skill reference
+- [x] `i18n-keys` - Add SrSkillsTabCore, SrSkillsTabUser, SrSourceCore to all locales
+- [x] `skill-card-ui` - Core badge + remove button suppression
+- [x] `skills-tab-ui` - Core/User sub-tab strip with filtering
+- [x] `fix-match-exhaustive` - Add Core arms to install.rs and install_dialog.rs
+- [x] `update-tests` - Fix store tests to account for core skills always being present

src-tauri/src/agent/harness_skills/file-access.md

@@ -0,0 +1,25 @@
+# File Access
+
+Explore and read files inside the active workspace sandbox.
+
+## Tools
+
+### `list_tools`
+Returns the full catalog of every available tool (name, site, description, parameters schema). Call this when you are unsure what tools exist.
+
+### `list_workspace_files { path?, recursive?, maxEntries? }`
+Lists files and directories under the workspace root or a relative subdirectory.
+- `path` — relative subdirectory (omit to list the root)
+- `recursive` — default `false`; set `true` for a deep scan
+- `maxEntries` — cap on returned entries
+
+**Pattern:** Always call `list_workspace_files` before reading files when you do not know the exact path. Do not guess directory names.
+
+### `read_workspace_file { path }`
+Reads a UTF-8 text file under the workspace root. Output is truncated at 4 000 characters. Path is relative to the workspace root.
+
+**Pattern:** After exploring the tree with `list_workspace_files`, read only the files you actually need. Cite the path you read in your reply.
+
+## Sandbox rules
+- All paths are relative to the workspace root. Never use `..` or absolute paths.
+- The tools enforce the boundary; path-escape attempts are rejected.

src-tauri/src/agent/harness_skills/harness.md

@@ -0,0 +1,46 @@
+# Harness Actions
+
+Client-side tools that mutate the BLXCode workbench window. After each call you receive a `role:"tool"` reply describing the result.
+
+## Workspace & terminal management
+
+### `harness.create_workspace { title?, cwd?, terminalCount?, agentSlugs? }`
+Creates and selects a new workspace in the UI.
+- `terminalCount` — 1–16 terminal slots
+- `agentSlugs` — optional per-slot array like `["claude", "claude"]`
+- `cwd` — omit to use the active workspace cwd or the configured harness root
+
+### `harness.open_terminal { count?, agentSlug?, agentSlugs? }`
+Opens one or more terminal slots in the **active** workspace.
+- **Default:** call with no arguments `{}` for a single plain shell.
+- `count` — open multiple at once (max 16). Do NOT call in a loop.
+- `agentSlug` — apply the same CLI agent to every new slot
+- `agentSlugs` — per-slot array (length must equal `count`)
+- Only pass agent slugs when the user explicitly names one of: `claude`, `codex`, `gemini`, `opencode`, `cursor`
+
+Example — open 3 Codex terminals: `{ "count": 3, "agentSlug": "codex" }`
+
+## Inspecting & driving other CLI agents
+
+### `harness.list_terminals`
+Returns `[{ slotId, agentSlug, running }]` for the active workspace. **Always call this first** when you intend to interact with another agent.
+
+### `harness.send_terminal_keys { slotId? | agentSlug?, text, submit? }`
+Types `text` into a slot's PTY.
+- `submit: true` — appends a newline so the command/prompt is executed
+- Address by `slotId` when possible (unique); `agentSlug` picks the first matching slot
+- Use to ask a running CLI agent for status, delegate work, or drive plain shells
+
+### `harness.send_agent_context { slotId? | agentSlug?, instruction?, includeKinds?, submit? }`
+Hands off BLXCode-attached context to a terminal CLI agent. Prefer this over raw `send_terminal_keys` when the other agent needs workspace context (memory/learnings, plans, tasks, images).
+- Image bytes are exported to `<workspace>/.blxcode/agent-context/images/`; base64 is never written into the prompt
+- `includeKinds` defaults to all four: `["memory", "plans", "tasks", "images"]`
+- Call `harness.list_terminals` first when multiple slots could match
+
+### `harness.read_terminal_output { slotId? | agentSlug?, maxBytes? }`
+Non-destructively reads the last bytes from a slot's rolling tail buffer (cap 64 KiB). Use after `send_terminal_keys` to observe the response. Output contains ANSI escapes — focus on the readable text.
+
+## Delegation pattern
+1. `harness.list_terminals` — find the target slot
+2. `harness.send_agent_context` or `harness.send_terminal_keys` — send the prompt
+3. Wait briefly, then `harness.read_terminal_output` — capture the reply (repeat for long-running tasks)

src-tauri/src/agent/harness_skills/memory.md

@@ -0,0 +1,76 @@
+# Workspace Memory
+
+Persist and retrieve notes and learnings scoped to the active workspace.
+
+## Storage roots
+- `.agents/memory/` — general notes (category `memory` + user-created subcategories)
+- `.agents/learnings/` — durable repo learnings (category `learnings`, API path `learnings/<file>.md`)
+- Subdirectories of `.agents/memory/` become their own categories (e.g. `projects/setup.md` → category `projects`).
+
+## Note CRUD (server-side)
+
+### `memory_list`
+Lists every note (up to 200) with size and modified time. Cheap overview call.
+
+### `memory_read { path }`
+Reads one note. `path` is the API path ending in `.md` (e.g. `projects/setup.md` or `learnings/api-notes.md`).
+
+### `memory_search { query }`
+Full-text search; returns up to 50 hits with paths and snippets. Use this before `memory_list` for targeted lookups.
+
+### `memory_create { path, content? }`
+Creates a **new** note (32 KiB max). Path must end in `.md` and must not already exist.
+
+### `memory_write { path, content }`
+Overwrites an existing note. Prefer this over `memory_create` to avoid near-duplicates.
+
+### `memory_delete { path }`
+Deletes one note permanently.
+
+### `memory_rename { oldPath, newPath, rewriteLinks? }`
+Renames or moves a note. Cross-root (`memory` ↔ `learnings`) is rejected. `rewriteLinks` defaults to `true` (updates `[[wikilinks]]` in other notes).
+
+### `memory_graph`
+Returns graph nodes/edges/tags clustered by category. Use for a high-level overview of the knowledge base.
+
+### `memory_backlinks { path }`
+Returns notes that link to the given path via `[[wikilinks]]`.
+
+### `memory_list_categories`
+Lists every category present in the workspace (built-in + user-created folders).
+
+### `memory_create_category { name }`
+Creates an empty category (subfolder under `.agents/memory/`). Use sparingly — prefer creating the first note with a `<category>/<note>.md` path instead.
+
+## Category UI & agent context (client-side)
+
+### `memory_category_list`
+Returns label/color/sidebar/graph flags for every visible category.
+
+### `memory_category_update { category, label?, color?, showInSidebar?, showInGraph? }`
+Updates display settings for a category. `color` as `#rrggbb`.
+
+### `memory_context_list`
+Lists items currently attached to BLXCode Agent context.
+
+### `memory_context_attach { kind, path?, label? }`
+Attaches a memory item. `kind` values: `memory_category`, `learning_category`, `memory_note`, `learning_note` (notes require `path`).
+
+### `memory_context_detach { id }`
+Removes an attached context item by id.
+
+### `image_context_list`
+Lists images attached to the active Agent context.
+
+### `image_context_detach { id }`
+Removes an attached image by id.
+
+## When to read / write
+
+**Read when:** the question is about this repo's conventions, architecture, prior decisions, or pitfalls. Start with `memory_search` (focused query), then `memory_read` on the 1–3 best paths.
+
+**Write when:** a durable convention, decision, API contract, migration step, or non-obvious pitfall emerged from the work.
+
+**Skip when:** trivial questions, single-line fixes, or topics answered entirely from the current message and one file read.
+
+Use `[[wikilinks]]` when linking related notes. Keep notes concise and free of secrets.

src-tauri/src/agent/harness_skills/plans.md

@@ -0,0 +1,60 @@
+# Workspace Plans
+
+Durable Markdown files that capture a multi-step implementation strategy. Lives at `<workspace>/.agents/plans/`. Plans are the long-lived counterpart to the task list and are checked into git.
+
+`PLANS.md` is the protected index — never delete or rename it.
+
+## Tools (server-side)
+
+### `plan_list`
+Overview of every plan with task summary counts. Call before guessing about existing plans.
+
+### `plan_read { path }`
+Reads one plan's Markdown body. `path` is relative to `<workspace>/.agents/plans/` (e.g. `my-plan.md`).
+
+### `plan_create { path, content? }`
+Creates a new plan. Path must end in `.md` and must not exist. Optionally seed with initial content.
+
+### `plan_write { path, content }`
+Overwrites an existing plan. Use `plan_sync_from_tasks` instead when you only need to update the `## Tasks` section.
+
+### `plan_delete { path }`
+Deletes a plan (cannot delete `PLANS.md`).
+
+### `plan_rename { oldPath, newPath }`
+Renames within `.agents/plans/`. Task records pointing at the old path are rewritten automatically.
+
+### `plan_load { path }`
+Syncs the plan's `## Tasks` (or `## Todos`) section into the task manager. Also attaches the plan to BLXCode Agent shared context (`PlanFile`) and sets `activePlanPath` on the snapshot.
+
+**Call `plan_load` whenever you open or start working from a plan** — including plans you just created. Without it, plan tasks are not in the task manager.
+
+### `plan_sync_from_tasks { path }`
+Writes the current plan-linked task state back into the plan Markdown. Use after reordering or batch-status-changing plan tasks via `task_*` tools.
+
+## Agent context (client-side)
+
+### `plan_context_list`
+Lists plans attached to BLXCode Agent shared context.
+
+### `plan_context_attach { path }`
+Attaches a plan (makes it visible to terminal CLI agents via `harness.send_agent_context`).
+
+### `plan_context_detach { id }`
+Detaches a plan from context by id.
+
+## Task line syntax in `## Tasks`
+
+```
+- [ ] `task-id` - Pending task title
+- [>] `task-id` - In-progress task title
+- [!] `task-id` - Blocked task title
+- [x] `task-id` - Completed task title
+- [-] `task-id` - Cancelled task title
+```
+
+## Key semantics
+- `plan_*` manage durable plan files; `task_*` manage live execution state. Both coexist for one plan.
+- `task_update` on plan-linked tasks writes status back into the plan Markdown automatically.
+- Plan files and the task store survive workspace reload, harness restart, and OS exit.
+- Reference plans from `PLANS.md` with Markdown links.

src-tauri/src/agent/harness_skills/rules-skills.md

@@ -0,0 +1,61 @@
+# Rules & Skills
+
+Two customisation layers stored under `<workspace>/.agents/`, each with an `index.json` manifest.
+
+## Rules vs Skills
+
+| | Rules | Skills |
+|---|---|---|
+| Storage | `.agents/rules/rule-*.md` | `.agents/skills/<name>/SKILL.md` |
+| Authority | **Binding** — override your defaults | **Advisory** — apply when relevant |
+| Conflict | Rules win over skills | — |
+| Naming | `rule-<stem>.md` | `[a-z0-9][a-z0-9-_]{0,40}` |
+
+Only entries with `enabled: true` count. Disabled entries do not exist — never apply, cite, or lobby the user to re-enable them.
+
+## Rules tools (server-side)
+
+### `rules_list`
+Returns JSON of every rule with `enabled`, `title`, `summary`, `updatedAt`. Filter by `enabled == true` before applying.
+
+### `rules_read { name }`
+Reads the Markdown body of one rule. Read every active rule whose title/summary is plausibly relevant before starting work.
+
+### `rules_write { name, content }`
+Creates or overwrites a rule. Name must start with `rule-` and end with `.md`. Only on explicit user request.
+
+### `rules_set_enabled { name, enabled }`
+Toggles the manifest flag. Only on explicit user request.
+
+### `rules_remove { name }`
+Deletes a rule and cleans its index entry. Confirm with the user before removing.
+
+## Skills tools (server-side)
+
+### `skills_list`
+Returns JSON of every skill (core and user) with `enabled`, `title`, `summary`, `source`, `updatedAt`.
+
+### `skills_read { name }`
+Reads `SKILL.md` for a user-installed skill, or the embedded content for a core harness skill. Use this to get full guidance for a tool group.
+
+**Core skill names:** `memory`, `plans`, `tasks`, `rules-skills`, `file-access`, `harness`
+
+### `skills_write { name, content }`
+Creates or overwrites a skill's `SKILL.md`. Source is set to `agent-created`. Only on explicit user request.
+
+### `skills_set_enabled { name, enabled }`
+Toggles the enabled flag. Works for both core and user skills.
+
+### `skills_remove { name }`
+Removes a user-installed skill directory and its index entry. Core skills cannot be removed.
+
+### `skills_install { name, source }`
+Installs a new skill. `source.kind` is one of:
+- `git` — `url` (required) + `ref` (optional branch/tag)
+- `npm` — `package` (required) + `version` (optional)
+- `local` — `path` (workspace-relative)
+
+The source must contain `SKILL.md` at the top level; otherwise the install is rejected and rolled back. Only call when the user explicitly asks to install something.
+
+## Index files
+`.agents/rules/index.json` and `.agents/skills/index.json` are managed by the harness. Do not hand-edit them — use the `*_set_enabled` / `*_remove` / `skills_install` tools instead.