dr-context-pipeline-x-2

原始内容

---
name: dr-context-pipeline
description: "Deterministic memory/context pipeline with semantic retrieval checks, compression/lint, safe fallback, and memory watchdog."
---

# DR Context Pipeline

Use this skill to standardize how an agent loads memory into its prompt for correctness.

## Prerequisites
- A file-based memory layout with `memory/always_on.md` and topic files under `memory/topics/`.
- Recommended: install `dr-memory-foundation` or an equivalent memory layout.
- For semantic retrieval assurance, OpenClaw memory search should be configured and visible through `openclaw memory status --deep --json`.

## Quick Install
These commands work from anywhere because they set `WORKSPACE`:

```bash
export WORKSPACE=${WORKSPACE:-~/.openclaw/workspace}
cd "$WORKSPACE"
clawhub install dr-context-pipeline --version 2.1.1 --dir skills --force
python3 ./skills/dr-context-pipeline/scripts/install_pipeline.py --target "$WORKSPACE/context_pipeline"
ls -1 "$WORKSPACE/context_pipeline"
git -C "$WORKSPACE" diff -U20 AGENTS.md | cat
python3 ./skills/dr-context-pipeline/scripts/validate_pipeline.py --context-root "$WORKSPACE/context_pipeline"
python3 ./skills/dr-context-pipeline/scripts/memory_watchdog.py --freshness-minutes 240 --min-bytes 200 --require-semantic --probe-query "context pipeline embeddings semantic retrieval" --min-search-results 1
git -C "$WORKSPACE" status -sb context_pipeline AGENTS.md
echo "CONTEXT PIPELINE APPLY COMPLETE"
```

Paste every command's output. If any step fails or the watchdog reports `GAP`, stop and reply `NOT EXECUTED: <reason>`.

## Apply To This Workspace
When the user asks to apply this skill, follow this order:

1. Lay down the files:
   `python3 ./skills/dr-context-pipeline/scripts/install_pipeline.py --target context_pipeline`
2. Show the tree:
   `ls -1 context_pipeline`
3. Patch `AGENTS.md` with the Context Pipeline instructions, preserving unrelated content.
4. Validate:
   `python3 ./skills/dr-context-pipeline/scripts/validate_pipeline.py --context-root context_pipeline`
5. Run the semantic memory watchdog:
   `python3 ./skills/dr-context-pipeline/scripts/memory_watchdog.py --freshness-minutes 240 --min-bytes 200 --require-semantic --probe-query "context pipeline embeddings semantic retrieval" --min-search-results 1`
6. Show final state:
   `git status -sb context_pipeline AGENTS.md`
7. Print the success banner:
   `echo "CONTEXT PIPELINE APPLY COMPLETE"`

This flow must be idempotent. If validation or the watchdog fails, stop with `NOT EXECUTED` and quote the issue.

## Memory Commit / Continue Workflow
- When Daniel says "memorize this" or similar, run `references/MEMORY_COMMIT.md`.
- When he says "let's continue" after a reset, reload `memory/now.md`, `memory/open-loops.md`, and the relevant topic files before acting.

## Runtime Modes
Every task must follow `references/RUNTIME_CHECKLIST.md`.

Modes:
- `normal` - compact day-to-day operation
- `debug` - full Retrieval Bundle and Context Pack artifacts
- `audit` - debug artifacts plus Receipt Ledger

Default to `normal` unless the user explicitly asks for debug, audit, tracing, receipts, or a full evidence dump.

## Semantic Memory Contract
When persistent memory is relevant, semantic or hybrid memory search is the preferred retrieval source.

The agent must record retrieval mode internally as one of:
- `semantic`
- `hybrid`
- `file_fallback`
- `not_needed`

If semantic memory is unavailable, not attempted, or degraded to file/keyword retrieval:
- normal mode must say so compactly in the footer
- debug/audit mode must include the reason
- operational checks that depend on semantic recall must stop with `NOT EXECUTED: memory/embedding gap`

Normal footer shape:

```text
Context: task type <type>; snippets <ids>; retrieval <semantic|hybrid|file_fallback|not_needed>; compression <succeeded|fallback>.
```

## Notification Safety
The watchdog can emit a dry-run notification payload on `GAP`, but this skill must not enable live delivery by itself.

Live memory/embedding alerts require a separate approval gate covering:
- schedule/frequency
- Discord DM target
- cooldown/suppression behavior
- exact message format
- retry behavior

Until approved, the watchdog prints JSON only.

## Operating Procedure
1. Load `memory/always_on.md`.
2. Route the message deterministically using `references/router.yml`.
3. Retrieve relevant snippets from memory; prefer semantic/hybrid memory search when persistent memory is relevant.
4. Build a Retrieval Bundle object that matches the schema and records retrieval mode.
5. Compress to a Context Pack using `references/compressor_prompt.txt`.
6. Lint the Context Pack. If lint fails, fall back to raw retrieved snippets.
7. Call the main reasoning model with always-on policy + Context Pack + user message.
8. For `debug` and `audit`, persist artifacts described in `references/RUNTIME_ARTIFACTS.md`.
9. Emit evidence according to the active runtime mode.

## What To Read / Use
- Router + caps: `references/router.yml`
- Compressor prompt: `references/compressor_prompt.txt`
- Retrieval Bundle schema: `references/schemas/retrieval_bundle.schema.json`
- Context Pack schema: `references/schemas/context_pack.schema.json`
- Receipt Ledger schema: `references/schemas/receipt_ledger.schema.json`
- Runtime artifact layout: `references/RUNTIME_ARTIFACTS.md`
- Runtime checklist: `references/RUNTIME_CHECKLIST.md`
- Semantic memory checklist: `references/SEMANTIC_MEMORY_ASSURANCE.md`
- Starter example file: `references/tests/golden.json`
- Positive-path harness fixtures: `references/tests/harness_cases.json`
- Anti-fabrication fixtures: `references/tests/anti_fabrication_cases.json`
- Installer/validator/watchdog scripts: `scripts/install_pipeline.py`, `scripts/validate_pipeline.py`, `scripts/memory_watchdog.py`
- Behavioral test runner: `scripts/run_pipeline_tests.py`

## Notes
- Keep `memory/always_on.md` tiny; put detail behind retrieval.
- Use snippet IDs only (`S1`, `S2`, etc.) in Context Pack `sources`.
- Follow `references/deterministic_ids.md` when deterministic snippet IDs matter.
- Run `scripts/run_pipeline_tests.py` after changing routing, retrieval caps, snippet IDs, runtime artifacts, derived queries, daily-log triggers, or transcript evidence.
- Run `scripts/memory_watchdog.py --require-semantic` before claiming semantic memory is healthy.

name: dr-context-pipeline description: "Deterministic memory/context pipeline with semantic retrieval checks, compression/lint, safe fallback, and memory watchdog."

DR Context Pipeline

Use this skill to standardize how an agent loads memory into its prompt for correctness.

Prerequisites

A file-based memory layout with memory/always_on.md and topic files under memory/topics/.
Recommended: install dr-memory-foundation or an equivalent memory layout.
For semantic retrieval assurance, OpenClaw memory search should be configured and visible through openclaw memory status --deep --json.

Quick Install

These commands work from anywhere because they set WORKSPACE:

export WORKSPACE=${WORKSPACE:-~/.openclaw/workspace}
cd "$WORKSPACE"
clawhub install dr-context-pipeline --version 2.1.1 --dir skills --force
python3 ./skills/dr-context-pipeline/scripts/install_pipeline.py --target "$WORKSPACE/context_pipeline"
ls -1 "$WORKSPACE/context_pipeline"
git -C "$WORKSPACE" diff -U20 AGENTS.md | cat
python3 ./skills/dr-context-pipeline/scripts/validate_pipeline.py --context-root "$WORKSPACE/context_pipeline"
python3 ./skills/dr-context-pipeline/scripts/memory_watchdog.py --freshness-minutes 240 --min-bytes 200 --require-semantic --probe-query "context pipeline embeddings semantic retrieval" --min-search-results 1
git -C "$WORKSPACE" status -sb context_pipeline AGENTS.md
echo "CONTEXT PIPELINE APPLY COMPLETE"

Paste every command's output. If any step fails or the watchdog reports GAP, stop and reply NOT EXECUTED: <reason>.

Apply To This Workspace

When the user asks to apply this skill, follow this order:

Lay down the files: python3 ./skills/dr-context-pipeline/scripts/install_pipeline.py --target context_pipeline
Show the tree: ls -1 context_pipeline
Patch AGENTS.md with the Context Pipeline instructions, preserving unrelated content.
Validate: python3 ./skills/dr-context-pipeline/scripts/validate_pipeline.py --context-root context_pipeline
Run the semantic memory watchdog: python3 ./skills/dr-context-pipeline/scripts/memory_watchdog.py --freshness-minutes 240 --min-bytes 200 --require-semantic --probe-query "context pipeline embeddings semantic retrieval" --min-search-results 1
Show final state: git status -sb context_pipeline AGENTS.md
Print the success banner: echo "CONTEXT PIPELINE APPLY COMPLETE"

This flow must be idempotent. If validation or the watchdog fails, stop with NOT EXECUTED and quote the issue.

Memory Commit / Continue Workflow

When Daniel says "memorize this" or similar, run references/MEMORY_COMMIT.md.
When he says "let's continue" after a reset, reload memory/now.md, memory/open-loops.md, and the relevant topic files before acting.

Runtime Modes

Every task must follow references/RUNTIME_CHECKLIST.md.

Modes:

normal - compact day-to-day operation
debug - full Retrieval Bundle and Context Pack artifacts
audit - debug artifacts plus Receipt Ledger

Default to normal unless the user explicitly asks for debug, audit, tracing, receipts, or a full evidence dump.

Semantic Memory Contract

When persistent memory is relevant, semantic or hybrid memory search is the preferred retrieval source.

The agent must record retrieval mode internally as one of:

semantic
hybrid
file_fallback
not_needed

If semantic memory is unavailable, not attempted, or degraded to file/keyword retrieval:

normal mode must say so compactly in the footer
debug/audit mode must include the reason
operational checks that depend on semantic recall must stop with NOT EXECUTED: memory/embedding gap

Normal footer shape:

Context: task type <type>; snippets <ids>; retrieval <semantic|hybrid|file_fallback|not_needed>; compression <succeeded|fallback>.

Notification Safety

The watchdog can emit a dry-run notification payload on GAP, but this skill must not enable live delivery by itself.

Live memory/embedding alerts require a separate approval gate covering:

schedule/frequency
Discord DM target
cooldown/suppression behavior
exact message format
retry behavior

Until approved, the watchdog prints JSON only.

Operating Procedure

Load memory/always_on.md.
Route the message deterministically using references/router.yml.
Retrieve relevant snippets from memory; prefer semantic/hybrid memory search when persistent memory is relevant.
Build a Retrieval Bundle object that matches the schema and records retrieval mode.
Compress to a Context Pack using references/compressor_prompt.txt.
Lint the Context Pack. If lint fails, fall back to raw retrieved snippets.
Call the main reasoning model with always-on policy + Context Pack + user message.
For debug and audit, persist artifacts described in references/RUNTIME_ARTIFACTS.md.
Emit evidence according to the active runtime mode.

What To Read / Use

Router + caps: references/router.yml
Compressor prompt: references/compressor_prompt.txt
Retrieval Bundle schema: references/schemas/retrieval_bundle.schema.json
Context Pack schema: references/schemas/context_pack.schema.json
Receipt Ledger schema: references/schemas/receipt_ledger.schema.json
Runtime artifact layout: references/RUNTIME_ARTIFACTS.md
Runtime checklist: references/RUNTIME_CHECKLIST.md
Semantic memory checklist: references/SEMANTIC_MEMORY_ASSURANCE.md
Starter example file: references/tests/golden.json
Positive-path harness fixtures: references/tests/harness_cases.json
Anti-fabrication fixtures: references/tests/anti_fabrication_cases.json
Installer/validator/watchdog scripts: scripts/install_pipeline.py, scripts/validate_pipeline.py, scripts/memory_watchdog.py
Behavioral test runner: scripts/run_pipeline_tests.py

Notes

Keep memory/always_on.md tiny; put detail behind retrieval.
Use snippet IDs only (S1, S2, etc.) in Context Pack sources.
Follow references/deterministic_ids.md when deterministic snippet IDs matter.
Run scripts/run_pipeline_tests.py after changing routing, retrieval caps, snippet IDs, runtime artifacts, derived queries, daily-log triggers, or transcript evidence.
Run scripts/memory_watchdog.py --require-semantic before claiming semantic memory is healthy.