Skip to content

Command-aware compaction proof

This is the smallest public-safe proof for the command-aware compaction lane.

It demonstrates four things together: - a compacted command view can be stored as an Observe-side receipt - the receipt keeps a deterministic pointer back to raw evidence - pack can prefer the compact text for bounded injection - family metadata can shape advisory policy hints without mutating retrieval ranking

Files

Scenario

Synthetic example: - raw command output is a tiny git diff - compact output is M app.py (+1 -1) - artifact compact-receipt binds the two - artifact rehydrate proves bounded raw recovery - pack selects the receipt and emits both compaction_sideband and compaction_policy_hints

The showcase now also includes a second synthetic proof for test_failures: - raw pytest -q failure output - compact triage summary - receipt + rehydrate + pack + metrics artifacts under docs/showcase/artifacts/test-failures-compaction.*

What the proof shows

1. Observe keeps both sides

The receipt records: - family: git_diff - compact text - raw artifact handle - raw bytes / kind

This means compaction is additive. It does not replace the raw evidence lane.

2. Raw recovery stays deterministic

The rehydrate receipt shows bounded recovery from the raw artifact handle referenced by the compaction receipt.

That keeps the product honest: compact is for usability, raw is still available for exact claims.

3. Pack prefers compact text, not compact truth

The pack output uses the compact text for bundle_text:

- [obs:1] M app.py (+1 -1)

But it also preserves: - compaction_sideband.selected[].rawArtifactHandle - compaction_sideband.raw_rehydrate_hint - trace.extensions.compaction_sideband

4. Family hints stay advisory-only

The same pack output emits: - compaction_policy_hints.mode = advisory_only - preferred_families = ["git_diff"] - one guidance line telling the operator to rehydrate raw diff before exact line-level claims

This is the key product boundary.

Family metadata can improve human/operator behavior now, without becoming ranking truth or write-path logic.

The test_failures proof shows the same boundary with a different operator posture: - compact first for triage speed - raw second for exact stack traces and assertions

Why this matters

This lane upgrades openclaw-mem in the right direction: - smaller bounded prompt bundles - better explainability for compacted command output - safer review posture because raw evidence is still reachable

It also keeps the architecture clean: - no new canonical-memory layer - no new storage tables - no silent automatic rehydrate inside pack

  • compact first for orientation
  • raw second for exactness
  • treat family hints as policy cards, not as ground truth

If you want the broader pack story first, start with the trust-aware context pack proof.