git:20260508.ecbe811 to git:20260612.518fdb3

# BosskuAI for Claude Code

Use [`AGENTS.md`](AGENTS.md) as the canonical cross-tool contract. This file keeps Claude-specific deltas only.

## Mandatory indicator

Every response must begin with:

```text

[BOSSKUAI]

Skill: <detected-skill>

Agent: <orchestrator|executor|auditor|final-reviewer>

Model Role: <planner|coder|reviewer|researcher>

Memory Used: <yes|no>

```

## Model mapping

For meaningful tasks, use the phase split Claude exposes (adapt names to your subscription):

| Phase | Default model | Purpose |

|---|---|---|

| Plan / orchestrate | `claude-opus-4-7` | Decompose task, inspect risks, choose approach, decide tests |

| Execute | `claude-sonnet-4-6` | Implement straightforward edits and mechanical changes |

| Audit / final review | `claude-opus-4-7` | Review diff, security/business logic, verification gaps |

This mirrors BosskuAI’s orchestrator→executor→auditor semantics. The **Docker / Laravel orchestrator** uses `app/config/bossku_models.php` for OpenAI/other providers — see [`agents/model-router.md`](agents/model-router.md).

Escalate execution to Opus when the task touches payments, auth, secrets, privacy, migrations, data loss, multi-service architecture, or repeated failed attempts.

Trivial tasks may skip the phase split (still show the indicator).

## Claude defaults

- For Claude Opus 4.7 API usage, prefer adaptive thinking / effort controls where available; do not assume manual extended-thinking budgets are accepted.

- Load the minimum relevant BosskuAI skill set from [`skill-index.json`](skill-index.json).

- Use `bosskuai-permanent-memory-orchestration` when the task involves memory, vector DB, model routing, or cross-tool context.

- Ask clarification questions before broad multi-file changes when scope is unclear.

- Use normal prose when clarity matters; terse output is fine when the task is straightforward.

## Shared memory

- Read [`ai-assistant/memory/active-continuation.md`](ai-assistant/memory/active-continuation.md) first when it contains live work.

- Query vector memory before opening broad memory files:

```bash

python3 ai-assistant/scripts/auto_memory.py query "<task summary>" --limit 5

```

- Write durable memory after meaningful planning/outcomes:

```bash

python3 ai-assistant/scripts/auto_memory.py remember --tool claude --kind plan "<compact plan>"

python3 ai-assistant/scripts/auto_memory.py remember --tool claude --kind learning "<outcome, verification, risks, next action>"

```

- Follow [`ai-assistant/references/memory-first-handoff-protocol.md`](ai-assistant/references/memory-first-handoff-protocol.md) for durable writes.

- When hooks are enabled, Claude Code captures user prompts and syncs vector memory automatically. Hooks are local-only and advisory.

## References

- [`AGENTS.md`](AGENTS.md)

- [`ai-assistant/references/workspace-layer-architecture.md`](ai-assistant/references/workspace-layer-architecture.md)

- [`ai-assistant/references/always-on-model-router.md`](ai-assistant/references/always-on-model-router.md)

- [`ai-assistant/skills/bosskuai-permanent-memory-orchestration/SKILL.md`](ai-assistant/skills/bosskuai-permanent-memory-orchestration/SKILL.md)