synthesize-agent-memory-landscape
Use when asked to write or refresh a public cross-system synthesis from retained analyse-agentic-system results and their memory-comparison fields. Produces one snapshot-bound analysis; excludes legacy reviews and Commonplace transfer scans.
zby · 已有 AI 审核记录
安装与使用
运行命令前先阅读来源与依赖说明。复制命令不会自动安装或执行。
npx skills add zby/commonplace --skill synthesize-agent-memory-landscape技能说明
name: synthesize-agent-memory-landscape description: Use when asked to write or refresh a public cross-system synthesis from retained analyse-agentic-system results and their memory-comparison fields. Produces one snapshot-bound analysis; excludes legacy reviews and Commonplace transfer scans. type: kb/types/instruction.md user-invocable: true argument-hint: "[public analysis path or response] [selected main reviews] [current or historical]" context: fork
Synthesize the agent-memory landscape
Produce a public comparison whose numbers and qualitative findings come from one frozen population of main-analysis results.
Inputs and authority
Use the requested output, selected systems, and current or historical mode from the user request or invoking packet. A request to refresh a named artifact supplies its output authority. Without a file destination, return the synthesis in the response. Load the output collection's contract before writing there.
The evidence inputs are generated reviews under kb/agentic-systems/reviews/
and the exact retained results named by their analysis-result paths and
analysis-result-sha256 values. Read the full retained results for findings:
source register, shared records, memory/context lens, reconciliation, and limits.
The compact review supplies publication identity and navigation. It cannot
replace a missing full result or comparison assessment.
Use kb/types/agentic-system-analysis-result.md for the memory-comparison
contract. Each row preserves its source revision, run, analysis cutoff, evidence
tier, compared memory boundary, and per-axis assessment, basis, values, and
canonical records. No legacy review, old CSV, transfer scan, or newly acquired
source may supply or repair a finding. Missing required inputs block the
selected population; report the main-analysis regeneration needed. Existing
results must not be hand-patched to make a comparison pass.
Freeze the evidence
-
Select the population. Default a refresh to current inputs. Repeat
--reviewto select the commissioned main reviews; omit it only when the commission covers all generated main reviews. Record the selection rule and exclusions. Select one review per source identity. A small selected set is a bounded comparison, with no implication of historical-corpus coverage. -
Create a new bundle. Run from the repository root:
uv run python scripts/bundle_agentic_landscape.py prepare --output <new-bundle-directory> --review <main-review-path>Repeat
--reviewas needed. Add--ontology <kb/notes/path.md>for each additional ontology artifact actually used. The command reads main results directly, derivesmatrix.csv, and captures the exact review/result bytes, result contracts, reader code, producing and consuming instructions, and dependency declarations under their repository-relative paths. It writessnapshot.jsonand a canonicalMANIFEST.tsvcontaining sortedsha256<TAB>byte-length<TAB>pathrows. Require exit status zero and save the reported manifest hash outside the bundle before interpretation. Existing bundle directories are never replaced. The command does not update public matrix or table files. -
Use only bundled evidence. Treat the bundle as immutable. If a needed finding or ontology input is absent, create a new complete bundle before drafting; do not mix in live files. For an existing historical bundle, verify it with its previously recorded manifest hash and use its bundled instruction and contracts. A method mismatch requires the matching pinned checkout. Legacy-corpus snapshots remain historical evidence, but this procedure does not rebuild or merge them into its population.
A temporary bundle suffices for a response or workshop trial. Before publishing, ensure the exact bundle and every cited original retained result are kept in Git, or identify a repository revision containing every input byte. A commit ID alone is insufficient when an input differs from that revision. Record the manifest hash, matrix hash, input file identities, and the reconstructable revision or retained-snapshot location in the published evidence boundary. A tracked comparison must remain auditable without ignored local run state.
Analyse and write
-
Compute quantitative candidates. Query the bundled CSV mechanically, decoding value cells as JSON arrays. For implementation/operation counts, use code-grounded rows with
knownvalues atwired,observed, orcausally supportedbasis, plusabsentassessments for evidenced negatives. Keep claimed and afforded findings separate. Keep doc-grounded findings in a separate qualitative section. Within each query, report inapplicable, uninspected, and not-determinable rows separately; none is an observed negative. A structurally valid unknown does not block unrelated findings.Retain an executable query and its output in a working query ledger. Each candidate names the fields, value-membership or set-equality test, tier and basis filters, numerator, denominator, included run IDs, and exclusions. Count each system once per query even when its value set contains several stores or routes. An assessed-subset proportion must name that subset; a whole-population prevalence claim requires complete applicable assessment. A change claim requires two verified snapshots, comparable scopes/contracts, and an explicit treatment of population changes.
-
Read and ground the mechanisms. For each selected finding, read the full bundled result and the cited canonical records, including their source evidence and limitations. Preserve the external mechanism and explain why the Commonplace term fits. Trace every qualitative example to a result path, hash, run ID, canonical IDs, and supporting section. Open-ended observations support named examples and contrasts, never prevalenc