Production Status

This page is the canonical reference for checked-in current-source facts: implementation version, public surface, capability boundaries, and validated source evidence. It deliberately does not snapshot the moving main commit, latest live CI result, latest published GitHub Release, or live PyPI availability. Consult GitHub Actions, GitHub Releases, and PyPI for those live repository and distribution states. Historical release announcements and proof artifacts remain evidence, but they are not the current-source status record.

Current Source Status

Current source test collection: 1051 tests

A tag is not a GitHub Release, and live CI or package-index state is not host certification, a distribution guarantee, or a productivity result. Installation guidance retains explicit publication and source-checkout gates.

Distribution Status

The 0.32.1 source line introduced the release-side contract for PyPI distribution. Current source 0.32.2 uses the same publication route: a published GitHub Release whose tag matches v<project.version> can build and verify distributions, then publish through PyPI Trusted Publishing with GitHub OIDC. No PyPI API token is stored in the repository workflow. Live PyPI package availability is external state and is not asserted by this checked-in document.

The 0.32.2 release changes product positioning, README/hero presentation, package metadata, and onboarding documentation. It does not alter runtime behavior, durable schema, MCP tools, repository-derived authority, governed durable project-memory authority, retrieval semantics, or automatic-learning boundaries.

Historical Tag Reference

The v0.27.4 tag identifies the historical source snapshot e8210cb204e501650a59876502a2028c7aae9afe. This is a stable tag fact, not a claim about the checked-out main commit, current publication state, or live CI status.

The v0.28.0 tag identifies the historical release merge snapshot c6e3568a59852c5b589d6aba00b89ab580c228e6. This is a stable historical release fact, not a claim about the current source head or current publication state.

Implemented Capability Summary

Governed Project Memory positioning and onboarding

V0.32.2 makes the existing product boundary easier to understand without changing runtime authority. Public positioning now describes AMB as governed project memory that carries useful context across sessions, tools, and time. The README and hero show fragmented project context converging into governed memory and being reused by future coding-agent sessions. Installation is explicitly separate from MCP client registration: each client must be connected, safe setup remains preview-first, and clients that should share memory must use the same persistent AGENT_MEMORY_BRIDGE_HOME.

Project Knowledge Activation

V0.30 adds a persistent, bounded, rebuildable derived repository snapshot. Explicit namespace binding and canonical local Git roots isolate independent clones and worktrees; a clean HEAD is required for commit-bound eligibility, while dirty, stale, moved, or unavailable sources fail closed. The existing MCP recall response keeps selected repository WHAT in repository_knowledge and governed durable WHY in normal items; repository facts are never promoted into durable memory.

Project Learning UX

V0.32.0 introduced the current Project Learning UX. Default Explore presents repository CODE / WHAT separately from governed CONVERSATION / WHY. project init is the preferred first-project path: it detects a local Git checkout, proposes a namespace, requires explicit confirmation, revalidates after confirmation, bootstraps derived repository WHAT, and shows Human-first Explore. Repeat init refreshes repository WHAT and leaves existing project WHY unchanged. This is not automatic learning and does not add an MCP tool.

Project WHY Alignment

V0.31.1 made explicit project decisions and constraints first-class governed task-memory lanes. The same active WHY that recall can see reaches Task Memory, Inspect, the Context Compiler, and Knowledge Explorer. Revision, supersession, validity, lineage, and structured-metadata governance remain in force.

Knowledge Explorer

V0.31 implements a local, read-only, bounded, deterministic, namespace-bound, rebuildable, provenance-bearing projection over existing project knowledge. Repository WHAT remains derived_repository; governed durable decision/constraint WHY remains governed_durable_memory; the namespace connector is derived_projection. Stale repository snapshots fail closed. The primary durable-memory raw scan is bounded at 500 rows, and relation targets use at most 100 unique IDs through direct same-namespace query-only lookup. Missing, existing-but-ineligible, and budget-exhausted targets are distinct diagnostics. Explorer adds no ranking, learning, writeback, graph database, or new MCP tool.

Durable memory and governed recall

AMB stores local engineering memory, revisions, lifecycle state, relations, and coordination signals. Lifecycle-aware recall and governed task-memory assembly suppress ineligible, stale, superseded, unsafe, or otherwise governed-out guidance before it becomes task context. Memory records remain correctable through existing governed mutation paths.

Dynamic State authority

Dynamic State is an internal exact-key release-state authority lane. It provides typed status transition, owner assignment, and restore commands; optimistic version and database-epoch guards; lifecycle idempotency; immutable accepted mutation and terminal request-outcome history; and a deterministically rebuildable state-head projection.

Dynamic State is separate from semantic memory, retrieval, ranking, embeddings, FTS, ordinary memory writes, and the MCP public tool surface.

Context and episode evidence

The Context Compiler is a transient derived-view layer. It accepts explicit bounded Repository Knowledge / WHAT with derived_repository authority, relation-aware governed task memory, exact Dynamic State read snapshots, and explicit session-local items. It does not query storage, rerun recall, rank records, persist a manifest, or add an MCP tool. Repository WHAT remains separate from durable WHY, and rendered context remains in process; manifest serialization retains metadata and digests rather than prompt-facing bodies.

A bounded context attestation can be recorded through the existing episode artifact path. It stores metadata and digests, not raw task text, rendered context, memory bodies, or session bodies. Read-only evaluation linkage can report whether valid context-attestation evidence is bound to the current strong verified outcome. It does not claim that selected context caused an outcome.

Run, event, outcome, artifact, and link rows are durable episode authority; projections and downstream learning/consolidation effects are shadow-only. Strong verification continues to depend on the existing governed episode and verification-receipt authority.

Runtime and Data Boundaries

Validation Status

The checked-in validation contracts verify the source facts on this page. For the current remote workflow state, use GitHub Actions; the repository also checks formatting and linting in its normal validation path.

question_count = 11
memory_expected_top1_accuracy = 1.0
memory_mrr = 1.0
file_scan_expected_top1_accuracy = 0.636
file_scan_mrr = 0.909

sample_count = 16
classifier_exact_match_rate = 0.875
fallback_exact_match_rate = 0.062
classifier_better_count = 13
fallback_better_count = 2
classifier_filtered_low_confidence_count = 2

case_count = 7
flat_case_pass_rate = 0.429
governed_case_pass_rate = 1.0
flat_blocked_procedure_leak_rate = 1.0
governed_blocked_procedure_leak_rate = 0.0
governed_governance_field_completeness = 1.0

signal_contention_case_count = 5
signal_contention_case_pass_rate = 1.0
unique_active_claim_rate = 1.0
duplicate_active_claim_count = 0
active_reclaim_block_rate = 1.0
stale_ack_blocked_rate = 1.0
stale_reclaim_success_rate = 1.0
pending_under_pressure_claim_rate = 1.0
initial_hard_expiry_cap_rate = 1.0

adversarial_case_count = 6
adversarial_task_count = 7
adversarial_governed_task_pass_rate = 1.0
adversarial_governed_blocked_record_leak_rate = 0.0

memory_evolution_case_count = 6
memory_evolution_task_count = 7
memory_evolution_governed_task_pass_rate = 1.0
memory_evolution_governed_blocked_record_leak_rate = 0.0
memory_evolution_governed_disposition_reason_hit_rate = 1.0

review_queue_item_count = 6
review_queue_actionable_count = 6
review_queue_hidden_lane_count = 2
review_queue_writeback_plan_count = 6
review_queue_no_auto_mutation = true
review_queue_public_mcp_surface_change = false
review_queue_item_type_count = 6

review_workflow_source_queue_item_count = 6
review_workflow_item_count = 6
review_workflow_manual_step_count = 27
review_workflow_requires_human_count = 6
review_workflow_auto_write_count = 0
review_workflow_no_auto_writeback = true
review_workflow_public_mcp_surface_change = false
review_workflow_item_type_count = 6

task_brief_used_count = 2
task_brief_ignored_count = 1
task_brief_needs_review_count = 4
task_brief_review_queue_item_count = 2
task_brief_active_signal_count = 1
task_brief_no_auto_writeback = true
task_brief_public_mcp_surface_change = false
task_brief_needs_review_source_type_count = 3

v019_case_count = 12
v019_pass_count = 12
v019_pass_rate = 1.0
v019_retrieval_case_count = 4
v019_retrieval_pass_rate = 1.0
v019_task_brief_case_count = 4
v019_task_brief_pass_rate = 1.0
v019_first_run_adoption_case_count = 4
v019_first_run_adoption_pass_rate = 1.0
v019_public_mcp_tool_count = 10
v019_public_mcp_surface_change = false
v019_client_config_write_count = 0
v019_durable_writeback_count = 0
v019_amh_required = false
v019_native_memory_comparison_required = true

v020_case_count = 6
v020_pass_count = 6
v020_pass_rate = 1.0
v020_import_sanity_pass = true
v020_stdio_round_trip_pass = true
v020_first_run_pass = true
v020_task_brief_pass = true
v020_public_mcp_tool_count = 10
v020_public_mcp_surface_change = false
v020_client_config_write_count = 0
v020_explicit_demo_memory_write_count = 1
v020_explicit_demo_signal_write_count = 0
v020_non_demo_durable_writeback_count = 0
v020_amh_required = false
v020_external_vendor_adoption_claim = false

v021_case_count = 20
v021_category_count = 4
v021_flat_baseline_hazards = 17
v021_governed_case_pass_count = 20
v021_governed_failures = 0
v021_governed_checkpoint_passes = 40
v021_governed_checkpoint_result_count = 40
v021_useful_current_retention_pass = true
v021_suppress_all_can_pass = false
v021_public_mcp_tool_count = 10
v021_public_mcp_surface_change = false
v021_auto_writeback_count = 0
v021_config_write_count = 0
v021_durable_live_writeback_count = 0

Known Boundaries and Non-Claims

AMB does not provide hosted execution, a scheduler, a queue, MCP Tasks, Apps, OAuth/ACL, authenticated actor identity, an HTTP product surface, an ANN/graph database, automatic reranking, automatic prompt or policy mutation, online training, automatic durable writeback, or autonomous skill acquisition.

Caller-declared client, session, model, harness, and evaluator labels are provenance metadata, not authenticated identity. Local benchmark and proof artifacts are implementation evidence, not a vendor certification, external adoption proof, or general productivity result.

Historical Evidence and Detailed References