Memory specialist acceptance

Commission and implemented boundary

On 2026-09-05 the operator authorized moving the memory lens to a fresh sub-agent, retaining the useful source-analysis progression of the old memory review. The typed report is the substantive handoff; off-band messages carry progress and access issues, with every analytical finding retained in the report. The operator explicitly rejected keeping legacy paths live or adding compatibility.

The new analyse-agent-memory skill writes one agent-memory-analysis-report inside the parent run. It owns the memory findings and proposed fourteen-axis profile. The main procedure freezes its input, delegates that pass, integrates proposals into canonical IDs and writes one self-contained result. Publication retains the main result and compact review. The local report and input remain required for local completion verification; public consumers use the retained result without needing ignored files.

The legacy skill and its two skill projections are removed. Publication's legacy arguments, output fields and semantic-review branch are removed, with no alias, version switch or fallback. Existing review artifacts remain historical evidence. Historical completion records are not rewritten to pass the current contract. This does not complete the pending historical survey, comparison and citation migrations.

Mechanical acceptance

Runtime tests exercise a complete memory-system publication with the specialist report and no legacy output, and reject missing reports, blocked reports, wrong run/source/boundary, changed input, changed report bytes, invalid report structure and incorrect source quotes. Shared comparison validation also rejects off-vocabulary values, unsupported known classifications, unresolved proposal IDs and contradictory dependencies. Report proposal IDs cannot leak into final-result comparison fields. Existing incumbent, frozen-source and rollback checks remain exercised. The removed CLI options are rejected.

These are deterministic checks of structure, identity and source occurrence. They do not certify analytical truth or establish an independent semantic review of the result.

Source-based handoff trial

Trial run: AAS-2026-09-05-memory-specialist-trial-01. Source: https://github.com/apache/maka at dd7d1d595b7f9284e01fe76cf547c979a6d84a0a. Input SHA-256: 481659e1f9c863b1ee4ce5e949d0dba674ae8168d97a5b888a5b5af76dfc1646.

A fresh worker received only the new skill, frozen source identity/access, SRC-1, scope and question. It received no previous reviews, audit findings, expected classifications or canonical memory conclusions. The scope covers conversation-event retention, both history-checkpoint variants and atomic MemoryItem extraction/storage/read-back. Local MEMORY.md/PENDING.md, goal state, skills, archives and external integrations are excluded. Static source inspection only; no system execution or public regeneration was commissioned.

The worker returned a complete report with SHA-256 a753a4b989ac4d5da4c3481f01225fc9dc6f23cc4f58eccab4ac7d26c99e67fb. Input and method hashes matched before and after execution. Its exact model identifier was unavailable, so the report records unknown. Standing report validation passed cleanly. The parent ran the publication source-anchor checker: 50 checks resolved at the pinned commit, with no failures.

The integration excerpt accepts all 14 proposed records as trial canonical IDs. The profile changes only those IDs; scope, values, statuses, rationales and uncertainties remain unchanged. The shared final-result comparison validator accepted the mapped profile. This excerpt retains its supporting records and findings, so reviewing the bounded integration does not require the ignored report.

The worker distinguished unconnected MemoryItem storage methods from the wired history-tool pull route; kept both text and encrypted checkpoint branches; and left actual task horizon and opaque forms unresolved. Parent spot checks confirmed the encrypted checkpoint type, text-only checkpoint copy admission and absence of production readItem/searchByKeys call sites beyond storage forwarding. Those checks support the bounded integration; they do not audit every analytical claim.

Acceptance: done for the specialist handoff and runtime migration. All 734 tests pass; repository Ruff and diff checks pass. Changed contracts, procedures, reference documents and workshop records validate cleanly.

Limits: this is a static, scoped memory trial plus publication fixtures, not a new complete production Maka review. The starting register contained only SRC-1, so the source trial does not exercise a substantive conflict with existing parent memory findings. A complete main-review rerun should next exercise runtime/epistemic integration and publication from this report contract. Historical public outputs were not regenerated or patched.