Record decision to retain repair context and limits

This commit is contained in:
2026-11-16 01:08:52 +00:00
parent 073e6cea26
commit a23ce14285
2 changed files with 32 additions and 27 deletions

View File

@@ -10,6 +10,8 @@ await implementation planning. D4 was also resolved on 2026-11-15: improve
input memory use without introducing an application-level input-size limit.
D5 was resolved on the same date: retain best-effort debug capture and operator
ownership of retention and cleanup; efficiency improvements remain pending.
D6 was also resolved on 2026-11-15: retain full repair conversation history and
per-correction limits, without aggregate budgets or context reduction at this time.
This document owns classification, recommended
scope, and pending decisions for the [catalog](catalog.md). The source audits
retain their evidence; their suggested remedies are not approved requirements.
@@ -43,8 +45,8 @@ when substantial restructuring is appropriate.
**Ready** means suitable for implementation planning within the stated scope,
not implemented or validated by this document. **Investigate** denotes technical
work the agent can undertake without a product decision. **D6–D7** identify
pending substantive choices; D1–D5 are recorded below as resolved. **Closed**
work the agent can undertake without a product decision. **D7** identifies
the remaining substantive choice; D1–D6 are recorded below as resolved. **Closed**
denotes no remediation for the specified scope. Mixed entries deliberately
separate ready work from decisions and exclusions.
@@ -107,7 +109,7 @@ behavior from proposed extensions.
| [31. Reduce peak memory for large transcript input and output](catalog.md#31-reduce-peak-memory-for-large-transcript-input-and-output) | E, A | Ready under resolved D4 | Remove generic JSON round trips and release whole-input representations after their last use. The D1 audit additionally identifies source-debug reconstruction: write the acquired bytes directly and retain parsed/normalized views separately, without reopening the input. Share typed output projection; preserve encoding-failure and stream-delivery semantics. Do not introduce an application-level input-size limit; calling integrations own their upload/process budgets. Preserve accepted input formats and stream support. |
| [32. Retain only the run evidence required by enabled outputs](catalog.md#32-retain-only-the-run-evidence-required-by-enabled-outputs) | E, A | Ready | The production consumers establish that much completed evidence is unused without its corresponding output. Retain only required live repair state, compact counts, and selected report/debug detail. Derive not-started counts without a full eager grid where possible; keep complete matrices when enabled and preserve fatal-origin evidence. |
| [33. Limit debug retention costs within a run](catalog.md#33-limit-debug-retention-costs-within-a-run) | E, A | Ready under resolved D5 | Reduce duplicate event representations and serialization; coalesce equivalent warnings while retaining failure counts and affected-artifact information. Preserve requested evidence, self-contained interactions, nonfatal debug failures, and continued best-effort capture. Add no quotas, automatic deletion, or automatic stopping after write failures; operators own retention and cleanup. |
| [34. Account for cumulative correction-repair history and work](catalog.md#34-account-for-cumulative-correction-repair-history-and-work) | E; P | Investigate costs; D6 | Repeated full context follows the deliberately selected repair architecture, so growth alone is not a defect. Measure cumulative requests, bytes, revalidation, and debug costs using existing observations or development instrumentation. Avoid duplicate local representations. New aggregate stopping budgets or reduced model context require D6; do not add ordinary-report fields merely to perform this analysis. |
| [34. Account for cumulative correction-repair history and work](catalog.md#34-account-for-cumulative-correction-repair-history-and-work) | E | Investigate costs under resolved D6 | Retain full conversation history and existing per-correction limits. Measure cumulative requests, tokens/bytes, revalidation, and debug costs using existing observations or development instrumentation; remove unnecessary local duplication. Do not add aggregate stopping budgets or reduce model context at this time. Revisit only if production evidence demonstrates a need; no new ordinary-report fields are required for this analysis. |
| [35. Avoid rebuilding growing normalization prefixes](catalog.md#35-avoid-rebuilding-growing-normalization-prefixes) | E | Ready | Growing merge prefixes are repeatedly scanned and copied. Use incremental text/category accumulation and exact estimator components while retaining joiner behavior, punctuation rounding, limits, and category order. Benchmarks should verify the improvement; large refactoring is not a reason to retain quadratic work. |
| [36. Skip validation context work beyond the usable budget](catalog.md#36-skip-validation-context-work-beyond-the-usable-budget) | E, A, T | Ready | Unusable optional context is serialized needlessly. Stop measurement once required-only or ordered-prefix costs settle the fit. Share trimming/batching policy without losing the independent serialized estimator oracle or changing context priority and oversized-item handling. |
| [37. Reuse checked processing selections and matching views](catalog.md#37-reuse-checked-processing-selections-and-matching-views) | E, A | Ready | Readiness and execution derive the same selections separately. Carry one checked processing plan and reuse immutable effective views and inspections per captured runtime. Preserve provenance and first-failure attribution; cache keys must include ordered selections and exemptions. Do not add cross-invocation caches or reload configuration. |
@@ -289,30 +291,30 @@ repeated failed work, but deliberately omit evidence and add configuration and
reporting obligations. They are not selected. Internal efficiency improvements
should preserve the evidence operators requested.
## Resolved D6: retain repair context and per-correction limits
**Accepted:** 2026-11-15. **Applies to:** #34.
**Status:** decision recorded; cost analysis and efficiency work remain pending.
Retain full conversation history and the existing per-correction limits. Do not
add aggregate section/run repair budgets or reduce conversation context at this
time. Measure cumulative requests, tokens/bytes, revalidation, and debug costs,
and remove unnecessary local copies without changing the model's evidence or
which corrections receive repair attempts. Use existing observations or focused
development instrumentation; no new ordinary-report fields are needed merely
to perform the analysis.
Resource growth partly reflects the deliberately selected repair behavior.
Aggregate budgets could leave later corrections unrepaired because earlier
ones consumed the allowance, while context reduction could reduce repair
quality. Neither change follows from the audit finding alone. Revisit these
options only if production evidence demonstrates a need, through a separate
policy decision.
## Open Decisions
The two decisions D6–D7 below remain pending. Accepting or rejecting an optional
extension must not delay the independent ready portions of its catalog entries.
### D6. Should correction repair gain aggregate budgets or reduced conversation context?
**Applies to:** the optional behavior changes in #34.
**Recommendation:** retain the selected full conversation history and
per-correction limits. Measure cumulative work and remove local duplication
before proposing any new stopping or context policy. The resource growth partly
reflects deliberate useful work rather than an implementation defect.
**Alternative:** add an opt-in section/run budget covering repair and
revalidation, with explicit termination outcomes; or separately evaluate reduced
context. Budgets provide predictable resource envelopes but may leave later
corrections unrepaired. Context reduction can lower costs but can also reduce
editorial quality. Neither follows automatically from the audit's asymptotic
description.
**Why approval:** these changes affect which corrections get another chance or
what evidence the model sees. Measurements using existing observations do not
require approval or new ordinary-report schema fields.
D7 is the only remaining open decision. Its resolution must not delay the
independent ready portions of the catalog.
### D7. Should this remediation add processing progress events?

View File

@@ -347,8 +347,11 @@ failures; operators own retention and cleanup.
### 34. Account for cumulative correction-repair history and work
Each rejected candidate can start its own bounded conversation and repeat the
original prompt history, with additional revalidation. Measure total requests and
bytes across branches before deciding on changes to context or run-level budgets.
original prompt history, with additional revalidation. **Resolved D6:** retain
full conversation history and existing per-correction limits; add no aggregate
budgets or context reduction at this time. Measure cumulative work and remove
unnecessary local duplication. Revisit behavior changes only if production
evidence demonstrates a need.
[Resources §7](05-resources.md#7-correction-repair-replays-growing-history-under-per-correction-budgets)
### 35. Avoid rebuilding growing normalization prefixes