colibri-audit-memory

分类: 工具与效率 | 上传者: LastEld | 下载: 0 | 版本: v1.0（最新）

HERITAGE 技能描述了审计跟踪和内存框架之间的捐赠者桥梁（memory_pack、memory_bundle、memory_verify、context_create、ams_session_resume）。不要在阶段 0 中调用——内存_*系列被推迟，并且在阶段 0 中审计和内存之间没有“桥梁”。 ze（决策跟踪）+ η（证明存储）一半通过audit_session_start /thought_record/audit_verify_chain/merkle_finalize/merkle_root/audit_session_end（6个工具）工作。有关有效的第 0 阶段证明级别工作，请参阅 colibri-mcp-server/references/workflow.md §工作流程 3 中的 colibri-audit-proof 技能和第 0 阶段证明级别工作流程。

更新日志: Source: GitHub https://github.com/LastEld/AMS

目录结构

当前层级: tree/main/.agents/skills/_heritage/colibri-audit-memory/

📁 agents/
- 📄 openai.yaml 299 B
📁 references/
- 📄 tools.md 1.8 KB
- 📄 workflows.md 1.4 KB
📄 SKILL.md 6.5 KB

SKILL.md

---
name: colibri-audit-memory
description: HERITAGE skill describing the donor bridge between audit trails and memory frames (memory_pack, memory_bundle, memory_verify, context_create, ams_session_resume). DO NOT invoke in Phase 0 — the memory_* family is deferred and there is no "bridge" between audit and memory in Phase 0. The ζ (Decision Trail) + η (Proof Store) half works via audit_session_start / thought_record / audit_verify_chain / merkle_finalize / merkle_root / audit_session_end (6 tools). For active Phase 0 proof-grade work, see the colibri-audit-proof skill and the Phase 0 proof-grade workflow in colibri-mcp-server/references/workflow.md §Workflow 3.
status: heritage
round: R74.5
updated: 2026-04-10
---

# Colibri Audit Memory — HERITAGE (not in Phase 0)

> ⚠ **Heritage skill. Not in Phase 0. Not in Phase 1.** This file describes a donor-era bridge between two systems that no longer coexist in Phase 0: the ζ/η audit trail (which **does** exist in Phase 0 as 6 tools) and the `memory_*` family (which **does not**). The bridge tools themselves — `memory_pack`, `memory_bundle`, `memory_verify`, `context_create`, `ams_session_list`, `ams_session_resume` — are all donor-era and are not registered in Phase 0.
>
> **Phase 0 reality**: there is no "memory" surface. The entire `memory_*` family (`memory_pack`, `memory_bundle`, `memory_map`, `memory_verify`, `memory_retention`, `memory_gc`) is deferred. The `context_*` family (`context_create`, `context_verify`) is deferred. The `ams_session_*` and `ams_autonomous_*` families are deferred.
>
> **What remains of proof-grade work in Phase 0** is the 6-tool ζ/η surface:
>
> | Phase 0 tool | Role |
> |--------------|------|
> | `audit_session_start` | Open a proof-grade session, returns `session_id` |
> | `thought_record` | Append hash-chained reasoning (`plan`, `analysis`, `decision`, `reflection`) |
> | `audit_verify_chain` | Verify the hash chain integrity |
> | `merkle_finalize` | Build the Merkle tree, returns root |
> | `merkle_root` | Fetch the stored root |
> | `audit_session_end` | Close the session after the Merkle root is finalized |
>
> **The load-bearing ordering rule** (CLAUDE.md §7) survives into Phase 0 **unchanged**: the final `thought_record` MUST come **before** `merkle_finalize`. Otherwise the reflection is not anchored in the Merkle root and the session is not proof-grade.
>
> **For active Phase 0 proof-grade work**, use:
> - [`colibri-audit-proof`](../colibri-audit-proof/SKILL.md) — the active ζ/η skill
> - [`colibri-mcp-server/references/workflow.md`](../colibri-mcp-server/references/workflow.md) §Workflow 3 — the canonical Phase 0 proof-grade workflow
> - [`docs/agents/writeback-protocol.md`](../../../docs/agents/writeback-protocol.md) — the T3 writeback protocol
>
> **Forbidden donor references you will see below** — none exist in Phase 0:
> - `unified_init`, `unified_vitals`, `unified_help`
> - `memory_pack`, `memory_bundle`, `memory_verify`
> - `context_create`, `context_verify`
> - `ams_session_list`, `ams_session_resume`
> - `include_rag`, `include_audit_trail` params (no RAG surface in Phase 0)
> - "memory frames", "audit_refs", "memory_hash", "frame graph" — these data structures are not in the Phase 0 schema
> - `src/server.js` → replaced by `src/server.ts` target (P0.2.1)

<details>
<summary>Heritage content (read-only, future phase reference)</summary>

Use this skill to integrate audit trails with memory frames for comprehensive session recording and archival.

## Quick Start

Run this bootstrap sequence first:

1. `unified_init`
2. `unified_vitals`
3. `unified_help`
4. `audit_session_start` with descriptive session_id
5. `memory_pack` to gather session data

Treat `src/server.js` as the active server and `archive/` as reference only.

## Core Concepts

### Audit Trail vs. Memory

| Aspect | Audit Trail | Memory |
|--------|-------------|--------|
| Purpose | Immutable action log | Context and reasoning cache |
| Structure | Hash chain | Frame graph with links |
| Proof | Merkle tree | Vector indexes |
| Access | Query by session/action | Semantic search, LLM retrieval |
| Lifecycle | Forever (immutable) | Tiered (hot/warm/cold) |

### Integration Points

1. **Session Start**: Create audit session → Create context snapshot → Pack initial memory
2. **During Work**: Record actions → Add thoughts → Update memory with incremental changes
3. **Session End**: Finalize audit chain → Merkle finalization → Bundle memory with proofs

## Core Tools (donor, deferred)

- `audit_session_start` — Start audit session *(Phase 0: kept)*
- `context_create` — Create immutable context *(donor, deferred)*
- `memory_pack` — Pack context, actions, thoughts into memory frames *(donor, deferred)*
- `thought_record` — Record reasoning *(Phase 0: kept)*
- `audit_verify_chain` — Verify audit chain integrity *(Phase 0: kept)*
- `memory_verify` — Verify memory frame checksums *(donor, deferred)*
- `merkle_finalize` — Finalize Merkle proof *(Phase 0: kept)*
- `memory_bundle` — Create prompt-ready bundle *(donor, deferred)*

## Session Lifecycle (donor version)

```
1. audit_session_start
2. context_create (optional but recommended)
3. memory_pack (initial)
4. [Work: thought_record, audit actions, code commits, etc.]
5. memory_pack (incremental updates)
6. audit_verify_chain
7. memory_verify
8. merkle_finalize
9. memory_bundle (final prompt-ready archive)
10. audit_session_end
```

## Operating Rules (donor version)

1. **Always start with session**: `audit_session_start` before any work
2. **Pack incrementally**: Call `memory_pack` at major checkpoints
3. **Record reasoning**: Use `thought_record` for both audit trail and memory
4. **Verify before finalizing**: Run both `audit_verify_chain` and `memory_verify`
5. **Finalize atomically**: `merkle_finalize` builds proofs for both audit and memory
6. **Bundle for archival**: `memory_bundle` creates the final prompt-ready record
7. **Cross-reference**: Use audit_refs in memory frames to link to audit trail

## Recovery from Interrupted Work

If a session is interrupted:

```
1. ams_session_list to find session_id
2. ams_session_resume to continue
3. memory_bundle to rebuild prompt context
4. Continue work (new actions still audit-linked)
5. memory_pack again at completion
6. merkle_finalize to rebuild proofs
```

## References

- Tool contracts: `data/tool-contracts/`

</details>

<!-- End of heritage content. For Phase 0 proof-grade work, use colibri-audit-proof + colibri-mcp-server/references/workflow.md §Workflow 3. -->

登录后下载/点赞/收藏 ❤ 10 | ★ 0

colibri-audit-memory

目录结构

SKILL.md

举报内容

提示