Decision lifecycle and evidence

Goal

Map the consequences of concrete choices about what work enters the project's durable record and how later readers use it. Use that map to support justified, provisional decisions for particular kinds of work—not a generally optimal retention architecture.

The operator opened this workshop on 2026-09-25 around a prematurely accepted ADR. The current commission broadens the question from ADR placement and access to the economics of retaining products and accounts of automated work. Git is the main record in this project. Requiring it in every Commonplace KB is not part of the present task.

A useful result may be a small change, different arrangements for different workflows, or a reasoned decision to keep existing practices. We need enough comparison to understand consequential commitments, not an exhaustive taxonomy, a complete cost model or proof of optimality.

Fixed boundaries

These are operator-set premises for this investigation, not hypotheses it must first test:

  • Use the available harnesses. Their session traces exist, contain the session information and remain available for some time. Trust that record; do not audit its completeness or redesign its capture, retention or deletion rules. A future harness change is a separate decision.
  • Preserve intermediate artifacts by default in the second repository. The operator set this provisional direction on 2026-11-11 for debugging and inspection, without per-file justification. Reconsider workflows with substantial churn or repository growth. This does not by itself implement capture, synchronization or permission to publish sensitive material.
  • Do not create duplicate primary process records. Preserve workflow files already produced; this does not require copying the harness's entire session archive or generating new accounts. Extra production must still earn its cost.
  • Preservation is not main-record admission. Operational files may be kept durably without curated extraction, endorsement or inclusion in the main repository. Lasting value may warrant separate extraction; debugging value already supplies the accepted class-level reason for operational preservation.
  • Artifacts admitted to the main record must serve their intended uses without session traces. Preserve the necessary context, evidence, reasons and qualifications, directly or through other durably retained material. Trace links can be optional provenance, not indispensable dependencies. This does not require preserving everything needed for every conceivable investigation.
  • Retention, adoption, implementation and observed benefit are distinct. A rejected option can deserve retention. An accepted choice is not thereby implemented or corroborated. Existing publication authority remains binding; no universal manual-approval gate is proposed here.

These distinctions would still matter under a future arrangement without the current harnesses. That possibility does not expand this commission.

Current state — 2026-11-11

Two concrete arrangements now have dispositions. The ingest working directory is kept provisionally for one decision about the repair round, with a recorded end condition. The cache/state investigation led to ADR 128, implemented in 9a854f3d0: ignored report outputs share state/, and cleanup is decided per workflow. This resolves the directory split, not admission to the tracked record or connect freshness.

The ADR-versus-commit comparison remains a design comparison without a selected carrier policy. The operator has now commissioned a workflow intermediate files ledger to show current handling against the preservation default. The first pass identifies deletion, replacement and location gaps; it does not implement producer changes. Complete its stated coverage gaps before claiming all workflows are covered. The second repository's publication and setup remain with the public-state-repository workshop.

On 2026-11-12 four families from the ledger moved under ADR 129: multistage runs, AutoReason bundles, article staging packages and the memory-landscape synthesis scratch now live under kb/reports/state/ and are kept after their workflow ends. The agentic-system analysis process record is the recorded exception until integration copies it into the state tree. Version capture, commit cadence and pruning remain open.

Working map

File Role
Admission to the main record Main design-space map: choices, benefits, costs, dependencies, reversibility and open questions.
Git and retained artifacts Bounded carrier comparison for decision accounts: ADR files, commit messages and access routes, with equivalent content.
ADR 089 premature placement Historical evidence that committing a decision account can communicate more than the implementation supports; not a current defect or a measured error rate.
Review storage carrier jump Historical evidence that leaving Git and choosing a database were decided as one choice, without comparing an untracked file layout; read from the retained ADRs, not from the original session.
Cache and state hold the same kind of record Dated inventory and reader comparison, followed by ADR 128's implemented removal of the directory split; connect freshness remains deferred.
Ingest working directory Current handoff retention and the repair-round observation trigger; later debugging-retention direction requires reconsidering its original deletion end condition.
Workflow intermediate files ledger Current paths, overwrites, deletions and retention gaps across workflow families; explicit remaining coverage, with no producer changes.

The general map remains preliminary. The two bounded dispositions above do not select a universal retention threshold, mining schedule, cache lifetime, marker, index or record type. The workflow examples are starting cases, not an obligation to test every combination. Agentic-system analysis remains outside the original carrier comparison, but its intermediate files are included in the ledger because worktree integration does not preserve all local debugging evidence.

How to advance the work

Intent: resolve a consequential choice at proportionate cost while preserving the reasons and uncertainty needed to reconsider it.

  1. Start with one real use and a choice it makes necessary. Name the future operation, its consumer and what information it needs. Existing practice is an alternative, not a deficient baseline by assumption.
  2. Compare only plausible arrangements. Trace what each requires, saves, loses or makes harder. State assumptions rather than inventing numerical precision. For carrier comparisons, hold decision content constant; do not compare a rich ADR with a terse commit and attribute the difference to the carrier.
  3. Identify interactions and reversibility. For example, independent retained use requires more extraction than a trace pointer, and trace deletion can close an option that is inexpensive to preserve temporarily.
  4. Use mechanism inspection and ordinary-work evidence first. A new execution test is warranted only by an unresolved question that could change the decision, with realistic conditions and proportionate setup cost. Do not build a benchmark, new index or alternative KB merely to make a test runnable.
  5. Retain the bounded conclusion: chosen or recommended arrangement, intended use, deciding reasons, important alternatives, acceptable losses, uncertainty and conditions for revisiting it. Keep original expectations and contrary outcomes distinguishable from later explanations. No fixed record template is required.

A partial comparison is useful if it identifies a real constraint or eliminates an unsuitable arrangement. If the remaining uncertainty does not justify more work, state it and stop rather than expanding the investigation indefinitely.

Current choices worth examining

  • Which workflows delete or overwrite intermediates before the second repository could preserve them? Complete the ledger's coverage, then select separately authorized implementation changes.
  • Which operational material deserves curated extraction into the main record, beyond preservation in the second repository?
  • Which workflows, if any, produce enough churn to justify an exception? No measured exception or numerical threshold is adopted.
  • For a decision worth retaining, what must survive trace deletion, and which existing file or Git access route makes it usable at acceptable cost?
  • Do the per-workflow cleanup rules adopted by ADR 128 identify acceptable losses in actual use? The cache/state split no longer supplies that policy; the investigation records why.
  • What does the ingest case show about the repair round, and which extra record-producing machinery still earns its cost? Historical preservation now has an accepted debugging use beyond that decision; it no longer follows that all records should be deleted afterwards. At this update, this machine has no handoff.yaml records under kb/reports/state/ingest/; that local count establishes neither activity on other machines nor failure of the arrangement. Trace recovery remains the stated fallback.
  • When would on-demand extraction be insufficient compared with selective proactive mining? No recurring mining programme is preselected.

Choose among these from actual work. No fixed order, required case count or claim of broad representativeness is imposed.

Scope and coordination

This commission permits design work in this workshop and its active-list entry. It does not authorize paid workers, implementation, evidence deletion, harness changes, history rewriting or framework-wide policy adoption.

Start from the shipped storage contract, Git history convention, commit trailers and content routing. The ADR type still governs ADR placement; mapping alternatives does not change it.

  • Lineage mechanisms owns general derived-material provenance and retention design. Coordinate any broader policy rather than duplicating it here.
  • Use-and-outcome records owns growing observational histories. Design-space studies explores richer design accounts.
  • ADR testing examines existing record usefulness; ADR routing owns consultation routes.
  • Conjecture-store remains independent. Its schema and thresholds do not bind this investigation; possible unification is not a prerequisite for completing it.

Completion

Close when the selected concrete choices have sufficiently understood consequences and justified provisional dispositions, including keeping current practice where appropriate. State where each conclusion applies, its evidence limits and what would warrant reconsideration. Unresolved broader questions can be explicitly deferred; a universal solution is not a completion condition.

Promote any warranted reusable finding or authorized proposal, preserve the selected evidence needed without sessions, and then close the workshop under its collection contract. A proposal is not an implemented change, and a design comparison is not measured proof of better agent performance.

Complete file listing (generated at build time)