dr-context-pipeline

Deterministic memory/context pipeline for agents: route a user message, retrieve relevant memory snippets, compress into a cited Context Pack (sources are sn...

DR Context Pipeline (retrieval + compression + lint)

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

Prerequisites

• A file-based memory layout that includes memory/always_on.md (policy header + topic catalog) and topic files under memory/topics/.
• Recommended: install dr-memory-foundation (or implement an equivalent structure).

Quick install commands (copy/paste)

These commands work from anywhere because they set $WORKSPACE (defaults to ~/.openclaw/workspace). Rerun them whenever you need to re-apply the pipeline:

export WORKSPACE=${WORKSPACE:-~/.openclaw/workspace}
cd "$WORKSPACE"
clawhub install dr-context-pipeline --version X.Y.Z --dir skills
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
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 a gap, stop and reply NOT EXECUTED: <reason>.

Apply to this workspace

When the user asks to apply this skill (for example: Apply dr-context-pipeline as default behavior), follow this exact order and paste the evidence for every step: All commands assume you’ve set WORKSPACE=${WORKSPACE:-~/.openclaw/workspace} and are running from that directory.

1. Lay down the files — run python3 ./skills/dr-context-pipeline/scripts/install_pipeline.py --target context_pipeline (adds/updates the workspace copy from assets/context_pipeline/). Keep the script output in the transcript (hash summary + file count).
2. Show the tree — ls -1 context_pipeline so the user sees which files are now present.
3. Patch AGENTS.md — read the file, insert/refresh the “Context Pipeline” instructions, and include a git diff -U20 AGENTS.md (or equivalent) snippet in your reply. Preserve everything else.
4. Validate — run python3 ./skills/dr-context-pipeline/scripts/validate_pipeline.py --context-root context_pipeline and paste the PASS/FAIL summary. If it fails, stop and report NOT EXECUTED with the error.
5. Run the memory watchdog — python3 ./skills/dr-context-pipeline/scripts/memory_watchdog.py --freshness-minutes 240 --min-bytes 200 (tune as needed). Paste the JSON output; if status ≠ OK, stop and reply NOT EXECUTED: memory gap after quoting the issues.
6. Final state — show git status -sb context_pipeline AGENTS.md (or git status -sb if cleaner) so the user can see what changed.
7. Success banner — echo "CONTEXT PIPELINE APPLY COMPLETE" so the transcript clearly shows the run finished cleanly.

This apply flow must be idempotent: if the files already match and AGENTS.md already contains the section, the diff should be empty but you still run the installer + validator and show their outputs.

Memory commit / continue workflow

• When Daniel says “memorize this” (or similar), run the checklist in references/MEMORY_COMMIT.md (daily log, now, open-loops, topic file, MEMORY.md) and confirm which files changed.
• When he says “let’s continue” after a reset, reload memory/now.md, open-loops, and the relevant topic files so you can summarize where things left off before acting.

Runtime evidence contract (non-optional)

Every single task must follow the Runtime Evidence Checklist in references/RUNTIME_CHECKLIST.md:

• Emit the Retrieval Bundle JSON that conforms to references/schemas/retrieval_bundle.schema.json.
• Emit the Context Pack JSON that conforms to references/schemas/context_pack.schema.json (or explicitly state the lint failure and fall back to raw snippets).
• List which snippet IDs are being passed forward.
• Only then provide the user-facing reasoning/result.
• If any step fails, reply NOT EXECUTED: <reason> and stop.

Casual prompts (“go for it”, “review it”, etc.) do not suspend this contract. If the user wants a lighter interaction they can explicitly waive it.

Operating procedure (default)

1. Load the always-on policy + topic catalog (your memory/always_on.md).
2. Route the message deterministically (task type + caps) using references/router.yml.
3. Retrieve top relevant snippets from your memory store; emit a Retrieval Bundle JSON (see schema).
4. Compress Retrieval Bundle → Context Pack JSON using references/compressor_prompt.txt.
   • IMPORTANT: Context Pack sources MUST be snippet IDs only (S1, S2, …).
5. Lint the Context Pack. If lint fails, skip compression and fall back to raw retrieved snippets.
6. Call the main reasoning model with: always-on policy header + Context Pack (+ raw snippets for high-stakes tasks) + user message.

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
• Runtime checklist: references/RUNTIME_CHECKLIST.md
• Golden tests starter suite: references/tests/golden.json
• Installer/validator/watchdog scripts: scripts/install_pipeline.py, scripts/validate_pipeline.py, scripts/memory_watchdog.py

Notes

• Keep “always-on policy header” tiny (invariants only). Put everything else behind retrieval.
• If you need deterministic snippet IDs, follow the stable ordering guidance in references/deterministic_ids.md.