Migrate consumers to the main agentic-system review
Commission
The operator commissioned this workshop on 2026-09-05 after reviewing the dependencies on the legacy agent-memory review corpus:
we need to move these dependencies to the new system one by one. Create a new workshop for this. These procedures should directly read the files generated by the main review.
The goal is one system analysis through
analyse-agentic-system,
with downstream procedures reading its generated files directly. Migrate each
consumer separately so its evidence requirements and resulting behavior can be
checked before moving on. The operator authorized execution of this plan on
2026-09-05 after the workshop was created.
Fixed direction
- System characterization comes from the main review's outputs. Do not generate another legacy-format review or introduce an adapter that recreates it for downstream consumers.
- Comparison matrices, tables, surveys, and transfer scans may remain distinct outputs. Their system claims must trace to the main review, without a second independent memory analysis or a fallback to legacy review prose.
- When a consumer needs information the main review does not record, identify that requirement explicitly. Improve the main review's shared method and output contract when warranted, then regenerate from frozen sources. Do not silently infer a classification from an omitted detail or patch generated system-specific findings by hand.
- Existing reviews may remain historical evidence during the migration. Repoint an evidential citation only after checking that its replacement supports the cited claim. A path substitution is not evidence migration.
- Keep Commonplace-specific transfer judgments separate from external-system characterization and public corpus evidence.
Main-review inputs and prerequisites
The exact result contract defines
kb/reports/state/agentic-system-analysis/<run-id>/result.md: the source register,
shared records, runtime account, memory/context and epistemic lenses, synthesis,
and limitations. Use this file for findings requiring the full analysis.
C01 now also retains these exact bytes at
kb/reports/retained/agentic-system-analysis/<run-id>/result.md during
publication. The public review pins that copy's path and hash. Public comparison
readers use the retained result; local transfer scans still check completion
through the run state.
The generated public review contract
defines kb/agentic-systems/reviews/<slug>.md, a compact projection of that result.
It supplies public navigation and can support citations where it actually
retains the needed finding. Do not assume that compact prose contains every
comparison field. The associated run-state.md carries completion and output
hashes; a result's complete disposition alone does not establish publication.
The first consumers resolved these prerequisites as follows:
- Selection and identity. Establish how the consumer selects completed results and joins a result to its public review, source identity, reviewed revision, and hash. Handle repeated runs explicitly; do not treat the newest filename as proof of completion or mix revisions. Reuse existing validation and identity machinery where it supplies the needed checks.
- Field coverage. Compare that consumer's actual evidence requirements with a completed result. Preserve unknown, uninspected, inapplicable, and evidenced absence as distinct cases. Decide comparison normalization only when a concrete field requires it; do not import every legacy column by default.
- Durable evidence — resolved for new runs. C01 publication retains an immutable byte-identical result and the public projection pins its hash. Existing results without this publication contract still need regeneration. Exact working results live in ignored local state. The reports contract forbids making tracked library artifacts depend on those files being present. Consumers needing full findings use that retained copy, preserving the original analysis rather than commissioning a rewritten review.
- Population. Most legacy reviews need source regeneration before they can enter the new corpus. Report the migrated population and missing coverage; never silently combine legacy and main-review evidence to preserve old counts.
Execution and bookkeeping
Consumers is the migration queue and the starting dependency inventory. Its order is not a prescribed implementation sequence. Before the first migration, inspect an available completed run and choose one consumer whose requirements can be tested against it. Matrix readers depend on the new matrix input contract; public syntheses depend on durable evidence; retirement of mandatory legacy publication is now explicitly authorized without waiting for historical publication migrations. The operator authorized a fresh memory specialist and typed report on 2026-09-05, then required removal of legacy paths without compatibility. That report is an internal handoff integrated into the main result, not another independently published memory review.
For each selected consumer, record its owned files, input requirements, chosen
main-review inputs, unresolved prerequisite, and acceptance evidence alongside
its queue entry. Use pending, in progress, blocked, done, or retired;
a blocked entry names the observation or decision that would unblock it.
Reassess the next consumer after each accepted change rather than freezing the
whole implementation now.
A consumer is done when its active procedure reads the main-review files, its output identifies the evidence it used, and a bounded example demonstrates the intended behavior with no legacy fallback. Derived CSV handoffs, where useful for formatting or calculation, must remain traceable to those files; they do not replace reading the main review for qualitative claims. Check a missing or insufficient input explicitly. For code changes, run the required tests; for Markdown changes, validate the changed artifacts and exercise the changed procedure at an appropriate bounded scope. Record checks and material limits, with a commit reference once committed.
Coordination and closure
This workshop owns downstream consumer migration and the eventual removal of
mandatory duplicate legacy publication. The earlier
analyse-agentic-system workshop retains
the procedure's construction and trial history. Its earlier legacy-publication
arrangement is the starting state, not the target architecture here. Changes
to the producer, result contract, publication checks, and corpus regeneration
must be coordinated with the consumer that demonstrates their need.
The workshop closes when every identified active dependency has migrated or been explicitly retired; new analyses no longer need a second legacy memory review; public outputs have reconstructable evidence; and remaining old reviews and references have an explicit historical disposition. Recheck the repository for unaccounted active readers before closure. Historical source citations and old migration scripts need not be erased to satisfy that check.
Extract the durable contracts and procedures into their owning collections, retain any independently needed acceptance evidence under the reports contract, then delete this workshop and remove its Active Workshops entry.
Complete file listing (generated at build time)