Move agent persona definitions from agents/ to prompts/personas/.
Add plan templates (bug-fix, code-review, feature-port, new-feature,
refactor) from the PHP PlanTemplateService into prompts/templates/.
Co-Authored-By: Virgil <virgil@lethean.io>
Tuned OpenBrain recall queries to use type-filtered searches (decision,
plan, convention) instead of generic keyword matching. Increased content
truncation from 300 to 600 chars. Added roadmap/priorities query.
Renamed MCP server from openbrain to core in mcp.json.
Co-Authored-By: Virgil <virgil@lethean.io>
- Add source field to brain_memories (manual, ingest:claude-md, etc.)
- Pass Qdrant similarity scores through to API response
- Minimum 50-char content length filter during ingest
- Content hash deduplication prevents duplicate memories on re-ingest
- Update VALID_TYPES to include all 13 memory types
- Include score and source in toMcpContext response
Co-Authored-By: Virgil <virgil@lethean.io>
Resolve stale forge.lthn.ai/core/cli v0.1.0 references (tag never existed,
earliest is v0.0.1) and regenerate go.sum via workspace-aware go mod tidy.
Co-Authored-By: Virgil <virgil@lethean.io>
Prevents flooding OpenBrain with near-identical context memories
during long sessions. Uses per-project timestamp file for debounce.
Also includes recent commits in saved context.
Co-Authored-By: Virgil <virgil@lethean.io>
SessionStart: query OpenBrain for recent activity + project context,
inject into Claude's context window on every new session.
Stop: async save session context to OpenBrain when git changes exist.
PreCompact: save working state to OpenBrain before context compaction.
API key read from ~/.claude/brain.key or CORE_BRAIN_KEY env var.
Co-Authored-By: Virgil <virgil@lethean.io>
- Allow go mod tidy, go work sync, go get, go run (no core wrapper)
- Only block sed -i on local files (allow via ssh/docker)
- grep -l only blocked when piped to destructive commands
- Remove sed > and awk > blocks (too aggressive)
- check-debug: only flag fmt.Print* not log.Println (go-log is valid)
Co-Authored-By: Virgil <virgil@lethean.io>
Remove claude/qa/ and claude/ci/ plugins (replaced by go-build and
go-devops repo-native plugins). Strip duplicate qa command from code
plugin. Marketplace now references plugins at their source repos.
Co-Authored-By: Virgil <virgil@lethean.io>
Go cmd_prep.go mirrors PHP PrepWorkspaceCommand — pulls wiki KB, copies
specs, generates TODO from issue, recalls context from OpenBrain. PHP
output dir changed from ./workspace/ to ./.core/ with lowercase filenames.
Co-Authored-By: Virgil <virgil@lethean.io>
- Update CI skill commands (status, fix, run, workflow) to detect Forge vs GitHub
and use Forge API with curl + FORGE_TOKEN instead of gh CLI
- Add detect-forge.sh script for CI provider detection from git remote
- Add hooks.json with PreToolUse, PostToolUse, PreCompact, SessionStart hooks
- Fix prefer-core.sh false positives: strip heredoc content before checking commands,
tighten wildcard matching for mv/cp to only block bare wildcards
- Update plugin.json: rename to "code", bump to v0.1.3, point URLs to Forge
- Update block-docs.sh with improved blocking rules
Co-Authored-By: Virgil <virgil@lethean.io>
Replaces 4 AltumCode Docker cron containers with a single scheduled
action. Inbound webhooks stored as append-only log, no inline processing.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
15 categories: design, engineering, game-development, marketing,
paid-media, product, project-management, sales, spatial-computing,
specialized, strategy, support, testing, examples, integrations.
Each agent has frontmatter (name, description, color, emoji, vibe)
and a detailed system prompt. Source: msitarzewski/agency-agents (MIT).
These feed into our content pipeline: agent personas drive strategy
and content generation through MixPost Enterprise, Bio.Host, and
the wider Host UK toolkit via MCP.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
Reorganise as a marketplace with multiple focused plugins:
- claude/code: Core development (hooks, scripts, data collection)
- claude/review: Code review automation
- claude/verify: Work verification
- claude/qa: Quality assurance loops
- claude/ci: CI/CD integration
Structure:
- .claude-plugin/marketplace.json lists all plugins
- Each plugin has its own .claude-plugin/plugin.json
- Commands namespaced: /code:*, /review:*, /qa:*, etc.
Install individual plugins or all via marketplace:
claude plugin add host-uk/core-agent
claude plugin add host-uk/core-agent/claude/code
Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>
- Add .claude-plugin/plugin.json manifest for auto-updates
- Move claude/ contents to root level (commands/, hooks/, scripts/, skills/)
- Update hooks.json to use ${CLAUDE_PLUGIN_ROOT} for portability
- Add .gitignore for IDE files
- Update CLAUDE.md with new structure and installation instructions
Plugin can now be installed via:
claude plugin add host-uk/core-agent
Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>
Scripts:
- auto-approve.sh: PermissionRequest hook that allows all tools
- ensure-commit.sh: Stop hook that blocks with uncommitted changes
- qa-filter.sh: PostToolUse hook that extracts QA failures only
- qa-verify.sh: Stop hook that runs QA and blocks on failures
Updated skill files to reference script paths.
Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>
Add skill files for autonomous workflows:
- yes.md: Auto-approve mode with commit requirement
- qa.md: Iterative QA fix loop for Go/PHP projects
Both skills use skill-scoped hooks to modify Claude's behavior
during execution.
Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>
Skill that enables "just do it" mode:
- Auto-approves all permission requests via PermissionRequest hook
- Ensures work completes with commit via Stop hook
- Skill-scoped hooks (only active during skill)
- Still respects safety blocks from prefer-core.sh
Usage: /core:yes fix the failing test and commit
Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>
6 new issues for Claude Code hook improvements:
- 012: Test output filtering (reduce noise, show failures only)
- 013: Stop verification (verify work complete before stopping)
- 014: Auto-test on edit (async tests after code changes)
- 015: Session context injection (git/issues/CI on startup)
- 016: Silent auto-formatting (suppress formatter output)
- 017: Expose/hide policy (define what to show vs suppress)
Based on Claude Code hooks documentation:
- PostToolUse with suppressOutput for noise reduction
- Stop hooks with agent verification
- Async hooks for background testing
- SessionStart for context injection
- additionalContext for exposing important info
Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>
Document the approach of using `core` CLI instead of raw commands:
- Command mapping table (go/php/git → core equivalents)
- Key commands reference for dev, go, php, build, ai, qa
- Missing features → raise issues on host-uk/core
- TDD approach: code as if functionality exists
Every agent using core tests the framework, making it bulletproof.
Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>
Consolidates the repository structure:
- Move skills/ to claude/skills/
- Move collection hooks to claude/collection/
- Update CLAUDE.md to reflect unified structure
Public version remains at core-claude; this is the advanced in-house version.
Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>
