slaude: Instructions file for Claude Code

slaude

Slack-native Claude Code runtime. Onboard AI agent as team member.

North Star

Like NousResearch/hermes-agent, but Slack-only, powered by Claude Code as engine. Agent has:

• Soul — persistent persona/identity file. Defines voice, values, mandate.
• Skills — grow over time. New capability = new skill file.
• Memory — episodic (what happened) + semantic (what learned).
• Autonomy — runs unattended. Asks owner via Telegram only when blocked on important question.

Shell host (folk fork or greenfield) = backend. Multiple sessions, one per agent identity.

Scope

In: Slack integration only. Single chat surface. Multi-agent (each agent = own slack identity). Out: Discord, Teams, web chat, CLI UX. Don't dilute focus.

Owner

Zidni Mubarok [email protected]. Telegram bridge available — use for blocking questions.

Working Rules

• Public repo — no internal/proprietary references. This repo is public. Never commit real people's names, company/employer/org names, internal Slack channel names, or internal service / KB / data-source identifiers in code, tests, comments, docs, commit messages, or PR text. Use generic placeholders instead (e.g. bulk-corpus, org/team-directory, #team-channel, Jane Doe). Findings docs describe the mechanism, never the internal incident specifics or the operator's deployment.
• Pre-commit hygiene — run BEFORE every git add/commit. A leak in a public repo's history is near-impossible to fully undo (forks/clones/cached SHAs survive a rewrite). Catch it before it lands:
  1. Scrub scratch artifacts — never stage runtime/tooling junk: .handoff, .mcp.json (runtime config; commit .mcp.json.example instead), .playwright-cli/, .playwright-mcp/, stray screenshots/PNGs in repo root, *.log. These are gitignored — if one shows in git status, it's a new escapee; gitignore it.
  2. Leak scan the staged diff for internal references. Quick grep over what you're about to commit:

     git diff --cached -U0 | grep -nIiE 'acme|\.acme\.|\.slack\.com|squadrondevel|\b[CUTGW]0[A-Z0-9]{8,}\b|AKIA[0-9A-Z]{16}|xox[baprs]-|ghp_|sk-[A-Za-z0-9]{20,}|-----BEGIN [A-Z ]*PRIVATE KEY|vault|deepseek|real-employee-names'
     

     Any hit that isn't an intentional placeholder or kept branding (the sim-TUI logo asset) → replace with a generic placeholder before committing.
  3. Secrets are values, not names — but treat leaked secret names (Vault keys, env var names tied to a real deployment) as sensitive too; they map to the operator's infra.
  4. If a leak already landed, see the history-rewrite playbook: git filter-repo --invert-paths --path <file> + --replace-text <rules> then force-push; always git bundle create a backup first.
• Granular commits. One logical change per commit.
• No AI co-authorship. Never add a Co-Authored-By: or "Generated with …" trailer naming Claude/Anthropic to commits or PRs, and never commit under an AI author identity. The .githooks/commit-msg hook enforces this (enable once per clone: git config core.hooksPath .githooks).
• Log significant findings/decisions/mistakes as a new docs/site/_content/field-notes/<date>-<slug>.md file and link it from the Findings Log index below (newest first). Keep this file lean — only the index lives here.
• Autonomous by default. Don't ask trivial; ask via Telegram only when:
  • Irreversible action needed
  • Architecture fork-in-the-road
  • Secret/credential required
• Memory — KB only (overrides harness default): All memory writes go to kb_memoize. Never write to memory/ files — the # auto memory harness instructions are superseded by this rule. For reads, kb_search/kb_think are primary; treat injected MEMORY.md context as legacy.
• Releases: every release ships a hand-written docs/site/_content/releases/<tag>.md with decent markdown notes — group by category (Features / Fixes / Docs / Internal), explain the why not just the commit subject, link findings docs when relevant. The release workflow prefers this file over auto-generated git-log dumps.
• Release candidates. Anything touching install.sh, the dist/version layout, the DB schema, or the agent loop goes out as vX.Y.Z-rc.N first. The RC publishes as a GitHub pre-release (auto-detected from the - in the tag) so install.sh never resolves it as latest; install it explicitly with SLAUDE_VERSION=X.Y.Z-rc.N. Soak, then scripts/promote-rc.sh vX.Y.Z-rc.N to bump, tag, and push the stable release. Notes live under the stable name (docs/site/_content/releases/vX.Y.Z.md) from the first RC onward. Tag and package.json must always agree — release.yml fails the build otherwise. Full procedure: .claude/skills/release-prep/SKILL.md.