Plan the prompt quality sprint

This commit is contained in:
2026-11-15 01:42:04 +00:00
parent fc607cfdf0
commit d6413b3996
2 changed files with 131 additions and 74 deletions

View File

@@ -0,0 +1,104 @@
# Prompt Quality Implementation Plan
Follow the stages in order within one sprint. Intermediate stages may leave
dependent edits in progress. Verify the integrated result at the end of the
sprint. Read the [feature roadmap](prompt-quality.md) for the selected behavior
and [developer guide](../development.md) for the applicable policies.
## Stage 1 — Finish the PromptKit Adoption Follow-ups
1. In the debug case of `TestSubprocessCorrectionRepairConversation`, read the
run's `llm-execution.json` and compare its top-level
`summary.provider_attempts` with the number of requests captured by the
provider fixture. Use the existing fixture with an initial proposal, repair
branches, and nested validation. This checks that the shared proposal is
counted once and that repair and validation calls join the invocation total.
[Subprocess fixture](../../cmd/audita/correction_repair_integration_test.go)
2. In `repairCandidates`, remove the extra JSON unmarshal and use the typed
response from one decode for the domain decision and the interaction
artifact. Reuse `DecodeRepairResponse` and its existing proposal-reference
and action checks. Where a decoded response fails a correspondence check,
retain it as diagnostic evidence alongside the raw response and failure.
Assess only delivered, structurally valid content. Keep the existing
invalid-output disposition, failure attribution, and revalidation behavior.
[Repair runner](../../internal/framework/runner/repair.go)
[Repair decoder](../../internal/framework/proposal_generation/repair.go)
3. In `BuildLLMExecution`, call `llm.Summarize` once per call and use the result
for both terminal coverage and the serialized summary. Keep the invocation
summary sourced from the native group. Preserve field names, ordering, and
the distinction between unknown usage and zero usage.
[Execution report](../../internal/framework/processreport/llm.go)
## Stage 2 — Review Proposal and Repair Conversations
1. Use the existing subprocess fixture to exercise the production adapter and
coordinator with scripted responses. Inspect its captured request messages
and response schemas for the initial proposal and the two-turn repair.
Inspect the existing withdrawal case in the runner fixture as well. Use
prepared request artifacts for the rendered instructions and wire requests
for the effective schema. Inspect the shipped definition and an operator
replacement from the existing fixture. Use temporary test logging if
captured requests or artifacts are otherwise discarded after a passing run.
[Subprocess fixture](../../cmd/audita/correction_repair_integration_test.go)
[Runner fixtures](../../internal/framework/runner/repair_test.go)
2. Check whether the original system instructions, the initial assignment,
and the subsequent repair assignment give a coherent task. Check that the
continuation retains the relevant glossary and protected spellings when
supplied, selects the repair definition, role profile, and one-result
schema, and includes each branch's own revision history. Compare the latest
target and before/after preview with the candidate under review, including
a retargeted edit and the existing omission of an inapplicable after-preview.
Use `TestRoleClientContinuesIndependentRepairConversations` for distinct
proposal and repair profiles.
[Adapter conversation fixture](../../internal/framework/llm/conversation_test.go)
3. Check whether the findings give the model a specific defect to address,
whether the current preview identifies the edit being assessed, and whether
the instruction makes revision and withdrawal clear. Fix concrete
ambiguities in the affected bundled prompt or feedback presentation. Keep
module focus, editorial acceptance, and required evidence as specified by
the feature roadmap. If the requests are coherent, record the reviewed
cases and the outcome in the sprint summary. [Prompt registry](../internal/prompts.md)
## Stage 3 — Review Evidence at the Validator Boundary
1. Follow a rejected candidate and, where available, its revision through the
captured proposal, validator request, decision, and next repair request.
Compare the target speaker, literal edit, before/after text, supplied
neighbors, matching glossary entries, and validator's stated reason.
Account for optional context omitted by the existing payload fitting.
[Validator payload and fitting](../internal/validators.md#supplemental-validation-context)
2. If evidence is already present but the instruction obscures how to use it,
revise the relevant prompt guidance or example. Preserve each validator's
remit, including the full-assertion reversal criterion. If the case needs
unavailable evidence, describe the missing evidence and the smallest
candidate payload change in the sprint summary. Broader context expansion
and combined semantic validation remain deferred. If no concrete gap
appears, the captured comparison completes this stage.
3. Use the native invocation summary to describe any additional requests
attributable to repair and revalidation. Use per-call evidence to identify
their roles. Treat group totals as encompassing potentially overlapping
work and account for terminal coverage separately from known or complete
usage. Report editorial findings from the requests themselves.
[Diagnostics and reporting](../internal/diagnostics-reporting.md)
## Stage 4 — Verify the Integrated Result
1. Review the existing runner, adapter, and subprocess evidence together for
independent branches, full revalidation, final-turn acceptance, withdrawal,
exhaustion, recoverable unavailability, and cancellation before delivery.
Confirm that accepted siblings survive recoverable failures and that
undelivered content cannot become an approved correction. Extend the nearest
existing fixture only if a changed behavior needs coverage. Use the real
adapter and coordinator with controlled model responses for execution tests.
2. For any prompt edits, inspect representative rendered requests and verify
supplied-data rendering and affected examples against the applicable
response schemas. Use durable assertions for changed behavior and inspect
the prose as part of review. Follow the [testing policy](../policy/testing.md).
Update the [prompt registry](../internal/prompts.md),
[module guide](../internal/modules.md), or [validator guide](../internal/validators.md)
where their owned behavior changes.
3. Use focused tests for iteration in the touched packages. At sprint
completion, run `go test ./...` with the repository's Go toolchain and
offline fixtures. Verify affected links and run `git diff --check`.
Summarize the cases inspected, any evidence-backed edits, and the results
of verification. Live provider comparisons remain deferred.

View File

@@ -3,7 +3,8 @@
**Status:** The earlier six prompt improvements are implemented. This follow-up
uses Audita's PromptKit v0.23.0 integration to review and develop proposal,
validation, and repair prompts through the same execution path used in production.
Live-model comparisons and broader context expansion remain deferred.
Live-model comparisons and broader context expansion remain deferred. See the
[implementation plan](implementation.md) for the sprint sequence.
## Baseline and Boundaries
@@ -32,83 +33,35 @@ run accounting. Audita owns candidate identity, findings, previews, acceptance,
withdrawal, and transcript mutation. Follow-up prompt work must use these APIs;
it must not recreate message replay, turn loops, or token aggregation locally.
## 1. Review Complete Native Conversations
## Target State
Complete the PromptKit adoption review follow-ups before revising prompts.
Proposal and repair instructions should remain coherent across a complete
conversation. The original context and supplied glossary and protected
spellings remain available, each repair branch keeps its own history, and
continuations select the repair definition, profile, and output schema.
Both shipped and operator-replaced definitions use native prompt selection.
Repair feedback identifies the current edit and the validator's findings so
that the model can revise the proposal or withdraw it.
- Extend the debug case of `TestSubprocessCorrectionRepairConversation` to
compare the top-level `llm-execution.json` provider-attempt total with the
provider fixture's request count. The total should include the initial proposal
once and count each repair and validator request.
[Subprocess fixture](../../cmd/audita/correction_repair_integration_test.go)
- Decode each repair response once in the runner. Reuse the decoded value for
proposal-reference and action checks and for debug output. Preserve the existing
invalid-output disposition and raw response evidence.
[Repair runner](../../internal/framework/runner/repair.go)
- Compute each per-call execution summary once and reuse it for terminal coverage
and usage in `llm-execution.json`.
[Execution report](../../internal/framework/processreport/llm.go)
Audita assesses every delivered revision, including the last allowed response.
Coordination stopping and correction acceptance are distinct outcomes.
Undelivered content is diagnostic evidence. Independent approvals survive
recoverable failures. Transient retries, structural repair, and correction
revision retain their separate budgets.
Use captured proposal exchanges and independent coordinator branches to inspect
representative proposal-to-repair conversations, including a second revision and
withdrawal. Review actual `OnPrepared` snapshots for instruction authority,
response-format transitions, glossary/protected-term context, and fresh previews.
A follow-up selects its repair definition, role profile, and schema directly.
Prompt and feedback changes should address concrete issues in captured
requests while preserving validator acceptance and the allowance for minor
editorial cleanup. Broader context and combined semantic validation remain
deferred. Any proposal to expand validator context should identify a concrete
case and the evidence missing from its request. This follow-up retains current
context fitting and apply-time conflict handling.
Exercise operator replacements through native prompt discovery and precedence.
Compare the request that actually executes rather than separately rendering an
approximation. Reusing captured history must preserve the proposal context while
keeping each correction's repair history private.
Record concrete editorial concerns and revise affected prompts when warranted.
Do not change validator acceptance or narrow allowable cleanup merely to make
proposal and repair assignments identical.
## 2. Assess Repair Feedback and Outcomes
Use the real coordinator with controlled model responses for deterministic
integration checks. Cover revision, withdrawal, final-turn acceptance, exhaustion,
unavailable responses, and independent accepted siblings. Revalidate each revision
from the beginning. A normal coordination stop is not evidence of acceptance;
Audita's domain outcome determines that. Canceled undelivered content is evidence
only and must never enter the approved set.
Use current target previews and validator-owned findings to assess whether a
repair request explains the defect sufficiently. Keep presentation changes in
prompt assets and domain feedback construction; keep history and budget mechanics
upstream. Retain separate transient retry, structural repair, and correction
revision budgets.
## 3. Diagnose Evidence Gaps Using Actual Requests
For a concrete rejected correction, compare the captured proposal evidence with
what the validator actually received. The target speaker is already included.
Broader validation context remains evidence-dependent: propose a focused change
only when a captured case demonstrates relevant evidence missing from validation.
Use native execution-group totals to distinguish additional model work from
shared initial history. Select validator membership explicitly, count the initial
proposal once, and do not add overlapping group totals. Preserve unknown usage
and terminal coverage separately. These measurements describe execution and cost;
they do not establish editorial quality.
Existing local examples have not established a need for broader validation windows
or combined semantic validation. Keep context fitting and apply-time conflict
handling unchanged until concrete cases justify further work.
## 4. Verification and Documentation
Follow the [testing policy](../policy/testing.md). Use the real PromptKit adapter
and coordinator with deterministic model-boundary fixtures. Protect supplied-data
rendering, schemas, profile selection, current previews, branching, revalidation,
withdrawal, and operator replacement. Inspect representative rendered prose as
part of review; avoid phrase checks, prose snapshots, and tests that duplicate
upstream history or accounting algorithms.
Update canonical prompt and validator documentation when behavior changes.
Keep examples valid against their response schemas and verify affected links.
Run focused tests and the full offline suite for runtime changes. Pure wording
changes need proportionate review, not new change-detector tests.
Native execution diagnostics count the shared proposal once and include each
new repair and validator run. Terminal coverage and usage availability
remain distinct. Execution measurements describe work and cost. Review
prompt clarity using the supplied evidence and rendered instructions.
The adoption follow-up also removes duplicate repair decoding and per-call
summarization while preserving the existing diagnostic evidence.
## Deferred Evaluation