dr-context-pipeline
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.mdand topic files undermemory/topics/. - Recommended: install
dr-memory-foundationor 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.mdwith 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.
Trigger Gate
Every task loads the small always-on policy. The full context pipeline is required only when one or more of these are true:
- persistent memory or a prior decision is materially relevant;
- the work is complex, risky, production-facing, or otherwise depends on historical context for correctness;
- the user explicitly asks for debug, audit, tracing, receipts, or an evidence dump.
For a routine, self-contained task where persistent memory is not needed:
- record retrieval mode as
not_needed; - answer or act directly;
- do not require a Retrieval Bundle, Context Pack, lint, artifact persistence, or a pipeline-specific footer;
- do not block the task because memory tooling is unavailable.
If the agent is unsure whether history materially affects the result, retrieve the smallest relevant context. If needed history is unavailable, use fresh read-only evidence or ask one targeted question. Stop the affected action only when its correctness genuinely depends on that missing historical state.
Runtime Modes
Apply references/RUNTIME_CHECKLIST.md when the full pipeline is triggered.
Modes:
normal- compact day-to-day operationdebug- full Retrieval Bundle and Context Pack artifactsaudit- 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:
semantichybridfile_fallbacknot_needed
If semantic memory is unavailable, not attempted, or degraded to file/keyword retrieval:
- use file/keyword fallback when it is sufficient;
- in normal mode, mention the limitation only when it materially affects the result;
- in debug/audit mode, include the reason;
- stop only an operational check whose correctness specifically depends on
semantic recall, 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. - Apply the Trigger Gate.
- For
not_neededwork, answer or act directly within the always-on safety and authority rules. - For triggered work, route the message deterministically using
references/router.yml. - Retrieve relevant snippets; prefer semantic/hybrid memory search when persistent memory is relevant.
- Build a Retrieval Bundle that records retrieval mode.
- Compress to a Context Pack and lint it. If lint fails, fall back to raw retrieved snippets.
- Call the main reasoning model with always-on policy + Context Pack + user message.
- For
debugandaudit, persist artifacts described inreferences/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.mdtiny; put detail behind retrieval. - Use snippet IDs only (
S1,S2, etc.) in Context Packsources. - Follow
references/deterministic_ids.mdwhen deterministic snippet IDs matter. - Run
scripts/run_pipeline_tests.pyafter changing routing, retrieval caps, snippet IDs, runtime artifacts, derived queries, daily-log triggers, or transcript evidence. - Run
scripts/memory_watchdog.py --require-semanticbefore claiming semantic memory is healthy.
Related skills
- Downloads
- 1.8kAll time
- Security audit
- Review
- Last updated
- 2w ago
- Current version
- v2.1.2
- License
- MIT-0
