Skip to content

v0.2 foundation: benchmark lock and context inventory - #3

Merged
HayronHgh merged 1 commit into
mainfrom
agent/adaptive-context-inventory
Aug 9, 2026
Merged

v0.2 foundation: benchmark lock and context inventory#3
HayronHgh merged 1 commit into
mainfrom
agent/adaptive-context-inventory

Conversation

@HayronHgh

Copy link
Copy Markdown
Owner

Scope

This is the first v0.2 development change and is intentionally limited to M0/M1. It does not connect a semantic Planner, execute promotion, or change the frozen v0.1.2 deterministic context policy.

What changed

  • merge and pin the resolved v0.1.2 control at tag v0.1.2 / ff2944581c7dd6200a170bc0bd94128e58ffd533
  • record the exact llama.cpp build, GGUF filename/bytes/SHA-256, chat-template hash, prompt/config/tool blobs, 64K/8K/4K budgets, thresholds, and host fingerprint
  • add an oracle-backed benchmark fixture with protected facts, checkpoints, and probes
  • add bilingual RFC-001 for Adaptive Semantic Context Planning
  • add validated Context Units with:
    • stable session-prefix + monotonic-sequence IDs
    • authority separate from future Planner importance
    • recoverability and artifact references
    • Runtime-owned protected reasons
    • typed dependencies, token cost, and lifecycle
  • add a bounded observational Context Inventory and /inventory
  • keep all inventory metadata behind the existing model serialization boundary
  • classify durable externalized exchanges as artifact-backed memory references
  • report inventory statistics without using them to authorize context reduction
  • mark the development package/runtime as 0.2.0-dev.1

Safety boundary

The Planner is not implemented in this PR. src/context-manager.js, deterministic thresholds, durability gates, path containment, State Transfer validation, and I1-I8 remain unchanged from v0.1.2.

Runtime protection cannot be cleared by a custom Context Unit descriptor. Invalid IDs, enums, dependencies, artifact recovery claims, duplicate active IDs, and malformed baseline fixtures fail closed.

Validation

  • npm.cmd test: 49 tests, 48 passed, 1 Windows file-symlink test skipped because the host denies file symlink creation
  • all source files passed node --check
  • all JSON files parsed
  • all PowerShell files parsed
  • 22 Markdown files passed relative-link validation
  • sensitive-information scan clear
  • git diff --check clean
  • real llama.cpp + Qwen3.6 E2E passed: ContextOS 0.2.0-dev.1 -> V020_M1_E2E_OK
  • merge-main CI for the frozen v0.1.2 baseline passed before the tag was created

Next milestones

  • M2: strict CompactionPlan protocol with fake Planner fixtures
  • M3: Runtime Validator where proposal never equals permission
  • M4: bounded Qwen semantic Planner with deterministic fallback
  • M5: deterministic/semantic/hybrid A/B/C benchmark

@HayronHgh
HayronHgh marked this pull request as ready for review August 9, 2026 11:52
@HayronHgh
HayronHgh merged commit 267a003 into main Aug 9, 2026
4 checks passed
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant