Adopt PromptKit v0.23.0 conversations and execution accounting

This commit is contained in:
2026-11-15 01:25:55 +00:00
parent d28b560265
commit d98d419b36
25 changed files with 689 additions and 924 deletions

View File

@@ -27,7 +27,7 @@ PromptKit loads its default catalog and Audita's embedded prompts, followed by
optional `prompts.yml` and recursive `prompts/` YAML files beside the application
main. Native imports, resource collections, inheritance, and replacement apply.
Profiles, backends, models, credentials, admission, and provider parameters belong
to PromptKit configuration; see its [format reference](https://forge.maximumdirect.net/base-layer/promptkit/src/tag/v0.21.0/docs/formats.md).
to PromptKit configuration; see its [format reference](https://forge.maximumdirect.net/base-layer/promptkit/src/tag/v0.23.0/docs/formats.md).
The engine captures its sources for an invocation. Edit files and run a new
invocation to pick up changes. Inspection does not require model credentials.

View File

@@ -19,7 +19,7 @@ failures; explicit zero makes one attempt. `llm.repair_attempts` counts
additional generations for schema-invalid model output. The operation timeout
covers preparation, capacity waiting, retries, generation, and repair. Caller
cancellation or an earlier deadline also applies. See [configuration](../config.md#settings)
for defaults and the [PromptKit consumer contract](https://forge.maximumdirect.net/base-layer/promptkit/src/tag/v0.21.0/docs/consumers/pkg-promptkit.md)
for defaults and the [PromptKit consumer contract](https://forge.maximumdirect.net/base-layer/promptkit/src/tag/v0.23.0/docs/consumers/pkg-promptkit.md)
for protocol and retry details.
A model operation that exhausts structural repair, a live-parent local operation

View File

@@ -48,6 +48,14 @@ artifacts can still contain sensitive transcript and prompt content. Execution
position and optional operator ID are carried separately from generated safe
artifact path coordinates.
Execution diagnostics include a native-group invocation summary and per-call
summaries. Terminal coverage is separate from usage completeness; unknown usage
is distinct from known zero. Overlapping summaries must not be added together.
The collector snapshots after work joins and retains domain labels and stable
presentation order. Prepared evidence comes from `OnPrepared`, so preparation or
admission failures may have no request snapshot. Coordinator execution failures
still finalize the attempted logical call and its available evidence.
## Key Tests
Useful tests: `internal/core/reporting/*_test.go`,

View File

@@ -9,27 +9,44 @@ it never replaces the requested engine.
Each resolved pipeline entry receives role-bound clients with its selected
profiles and effective budgets. Clients share engine-local backend admission.
The adapter uses `PrepareExecution` and `RunPrepared` on the same handle, passes
The adapter uses ordinary `Run` with `OperationTimeout` and `OnPrepared`, passes
Audita's per-call schema and repair budget, and decodes only validated results.
The operation budget includes preparation, admission, retries, and structural
repair. `ErrOperationTimeout` remains recoverable even when wrapped by a
preparation error. Parent cancellation remains distinct. The prepared notification
provides the actual request snapshot; failures before notification have no snapshot.
PromptKit owns backend and credential selection. The adapter uses captured
credential source names for artifact redaction without inspecting unrelated
environment values. See the [integration contract](../integrations/openai-compatible-llm.md)
for failure behavior and the [configuration reference](../config.md) for sources.
When correction repair is enabled, the proposal call retains its prompt,
effective variables, prepared messages, and full final response for section-local
replay. The adapter renders the repair definition with the selected profile,
appends only that branch’s history, and checks the original prepared prefix
before execution. The complete conversation is sent to the selected provider;
provider rejection ends that branch as unavailable. The requested repair
definition is retained in call attribution alongside native base-prompt metadata.
When correction repair is enabled, the proposal call captures a native immutable
PromptKit exchange. Each actionable correction starts an independent
`Engine.Coordinate` branch from that exchange. PromptKit assembles continuation
history, owns the additional-turn budget, and delivers the last permitted response
for assessment. Follow-ups select the repair prompt, profile, and output schema
directly; Audita does not rerender the original proposal or assemble replay messages.
An opaque exchange in the structured-output contract keeps provider state outside
modules and validators. The adapter shares request construction, notifications,
observers, and result/failure finalization between ordinary and coordinated calls.
Audita owns actionable rejection, revision decoding, withdrawal, current previews,
and full revalidation. Accepted siblings remain independent. Requesting more work
at the budget boundary produces exhaustion without allocating another logical call.
A canceled, undelivered response supplies diagnostic evidence only; it cannot enter
the approved set. Repair disabled in configuration bypasses coordination entirely.
The runner supplements each repair instruction with its current proposal preview
as described in the [module guide](modules.md#failure-and-outcome-behavior).
The bundled prompts separate durable system principles from the initial user
assignment so later repair instructions can allow acceptable minor cleanup;
see [prompt instruction ownership](prompts.md).
The observation collector tracks logical calls and execution events. Ordinary
The observation collector tracks logical calls and execution events, and enrolls
selected runs in a native `ExecutionGroup`. Proposal, repair, and nested validator
clients explicitly share the invocation observer; history reuse adds no extra run.
Native summaries deduplicate terminal totals by run ID and retain terminal coverage
separately from usage presence and completeness. Logical calls and run membership
can differ, especially on failures before execution. Ordinary
reports use compact safe outcomes; request and response evidence is optional
debug material. The runner preserves ordered pipeline traversal and deterministic
section collection while canceling and joining peers on fatal errors.

View File

@@ -20,7 +20,9 @@ formats; [configuration](../config.md) owns settings; the
Recoverable failed batches mark their candidates unavailable; successful
survivors continue through all required validators.
6. Actionable rejections can enter independent section-local repair branches.
Each revision is mapped to the original section and revalidated from the
PromptKit coordinates each branch from the captured proposal exchange, owning
its history and additional-turn limit. Each revision, including the final
permitted response, is mapped to the original section and revalidated from the
start. Withdrawal, exhaustion, and recoverable failures finish the branch
without affecting independent approvals.
7. Only fully approved proposals reach application. Corrections apply in

View File

@@ -210,7 +210,7 @@ License and notices: [Apache 2.0](licenses/Apache-2.0.txt).
## PromptKit
Selected module: `go.maximumdirect.net/promptkit` at `v0.21.0`.
Selected module: `go.maximumdirect.net/promptkit` at `v0.23.0`.
Copyright (C) 2026 Eric Rakestraw
@@ -218,9 +218,9 @@ SPDX-License-Identifier: LGPL-3.0-or-later
License and notices: [LGPL v3](licenses/LGPL-3.0.txt) |
[GPL v3](licenses/GPL-3.0.txt) |
[COPYRIGHT](https://forge.maximumdirect.net/base-layer/promptkit/src/commit/171fb11f4e967a1da3af57843c6467954b14ba09/COPYRIGHT).
[COPYRIGHT](https://forge.maximumdirect.net/base-layer/promptkit/src/commit/fcbff80ab227397a03935ab3bf8114481c0736a7/COPYRIGHT).
[Source](https://forge.maximumdirect.net/base-layer/promptkit/src/commit/171fb11f4e967a1da3af57843c6467954b14ba09).
[Source](https://forge.maximumdirect.net/base-layer/promptkit/src/commit/fcbff80ab227397a03935ab3bf8114481c0736a7).
## yaml.v3

View File

@@ -1,229 +1,106 @@
# Prompt Quality Roadmap
**Updated:** 2026-11-11
**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.
**Status:** All six stages implemented and verified. Broader context changes and
live-model editorial evaluation remain deferred as described below. The plan
followed the review of revision `d155bf4`, which covered all eleven prompt
definitions, shared instructions, response schemas, and request assembly.
That review identified design concerns, not measured live-model regressions.
## Baseline and Boundaries
## Target State and Boundaries
The [prompt registry](../internal/prompts.md),
[validator guide](../internal/validators.md), and
[LLM runtime](../internal/llm-runtime.md) own implemented behavior. The existing
prompts separate durable system principles from initial proposal assignments,
guide independent and overlapping edits, supply current before/after repair
previews, include the target speaker in validation, and illustrate editorial
judgments, revision, and withdrawal.
Audita's proposal, validation, and repair prompts will give each model call a
coherent assignment and the evidence needed to carry it out. Proposals will
better account for interacting edits, repair will receive an explicit view of
its current proposed result, and validation will identify the target speaker.
Focused examples will clarify difficult editorial judgments.
Preserve these selected decisions:
Preserve the established design:
- Module focus guides generation; otherwise acceptable minor editorial cleanup
remains permissible across modules under [ADR 0003](../adr/0003-module-focus-and-editorial-acceptance.md).
- Keep the carefully tuned meaning-reversal criterion and distinct validator
responsibilities. Approvals require no explanation; confidence scores remain absent.
- Preserve literal spans, deletion, replacement policies, glossary context,
protected spellings, full revalidation, and deterministic application.
- Retain English defaults and native operator prompt replacement. Module sequence
reevaluation follows the [module guidance](../internal/modules.md).
- Module focus guides proposal generation while otherwise acceptable minor
editorial cleanup remains permissible across modules, as selected in
[ADR 0003](../adr/0003-module-focus-and-editorial-acceptance.md).
- Validators retain distinct responsibilities. Preserve the carefully tuned
meaning-reversal criterion and its focus on reversal of the full assertion.
- Keep literal-span matching, deletion semantics, separate glossary context and
protected spellings, and full before/after validation.
- Keep independent repair conversations with full history, successful withdrawal,
and full revalidation of revisions. Validators explain defects; repair owns
revision or withdrawal. Approvals continue to need no explanation.
- Retain confidence-free responses, English defaults, existing module identities,
and native operator prompt replacement. The current module sequence is outside
this pass; ongoing reevaluation follows the [module guidance](../internal/modules.md).
PromptKit now owns immutable captured exchanges, continuation assembly, bounded
coordination, per-operation deadlines, prepared-request notification, and selected
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.
The [prompt registry](../internal/prompts.md), [validator guide](../internal/validators.md),
and [LLM runtime](../internal/llm-runtime.md) describe the implemented baseline.
Update these canonical owners when the corresponding changes land.
## 1. Review Complete Native Conversations
## Implementation Sequence
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.
### 1. Align Proposal and Repair Instruction Authority
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.
Proposal system messages currently include categorical task restrictions, such
as prohibiting punctuation or grammar edits in the homophone module. Repair
inherits those messages and then permits acceptable minor editorial cleanup in a
later user message. These instructions can compete across authority levels.
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.
- Put durable correction principles in the system messages: evidence-supported
work, established protection and meaning constraints, and the distinction
between requested focus and acceptable editorial corrections.
- Put the initial proposal's focused assignment and task-specific exclusions in
its user message. Preserve the focused tasks rather than asking every module
to perform every kind of cleanup.
- Let each repair instruction give the subsequent reassessment assignment,
including the selected editorial allowance, without conflicting with an
inherited system prohibition.
- Review each complete proposal-to-repair conversation, including later repair
attempts, for coherent instructions and response-format transitions.
## 2. Assess Repair Feedback and Outcomes
This implements the existing acceptance decision; it does not tighten validators
or change the repair lifecycle. Keep the validator prompts' separate remits,
especially meaning-reversal review, intact.
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.
### 2. Guide Independently Useful and Overlapping Corrections
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.
The shared proposal contract already makes each edit refer to the supplied
segment rather than the imagined result of another proposal. Add guidance for
edits that overlap or depend on one another:
## 3. Diagnose Evidence Gaps Using Actual Requests
- Prefer independently useful corrections with literal spans that identify the
intended changes.
- Combine changes into one proposal when their spans overlap or when one edit
requires the other. Preserve unchanged context needed for unique matching or
for selecting an occurrence under the glossary replacement policy.
- Continue to allow multiple independent corrections to the same segment.
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.
For example, separate proposals to capitalize `hello` and replace
`hello world` with `hello world` may both pass validation against the original,
but applying the capitalization first prevents the second span from matching.
A combined replacement can express both edits against the original text.
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.
These are prompt instructions. Add no new runtime overlap rejection, proposal
limit, or application-order rule. Assess combined semantic effects separately
if practical examples establish a need for additional processing changes.
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.
### 3. Supply Repair with the Current Before/After Segment
## 4. Verification and Documentation
The repair request currently identifies the proposal, its current correction,
and validation findings. Although the original section remains available in
conversation history, repair must locate it and reconstruct the proposed result.
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.
- Add the original target segment and its proposed resulting text to repair
context when the current proposal is applicable.
- Reuse the same domain preview semantics used by validation, including literal
deletion, whitespace, and replacement of every matching occurrence where
applicable. Avoid a second implementation of replacement behavior.
- Refresh this view for each repair attempt, using the current proposal version
and target rather than a stale view of the initial proposal.
- When a preview cannot be produced, retain the relevant findings and available
source context without fabricating a resulting segment or implying application
succeeded.
- Keep the full conversation history, proposal references, and revision or
withdrawal response contract. Explain the added view as evidence for assessing
the current correction, not an already-applied transcript change.
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.
This should make spacing, deletion, and replacement consequences easier to
judge without requiring the model to reproduce Audita's application mechanics.
## Deferred Evaluation
### 4. Improve Validation Evidence
Formal corpus preparation and live-model comparisons remain optional follow-up
work. Keep live calls opt-in and outside the offline suite. When undertaken,
compare needed and unnecessary edits, missed corrections, validator disagreement,
and repair outcomes using independent editorial judgment. Native groups can
supply execution measurements without introducing a second accounting system.
Add the target segment's speaker to each validation item. Neighboring segments
already carry their speakers; the candidate's identity completes that context
and helps distinguish continued speech, interruptions, and another speaker's
response. Carry the supplied value through the existing payload construction
and token accounting without inferring a missing speaker.
Investigate the broader evidence imbalance using available practical cases:
proposal sees a section, while validation sees immediate neighbors and selected
glossary entries. A correction supported several turns earlier may appear
unsupported to a validator. Compare the evidence actually supplied to both
calls before attributing a rejection to prompt wording.
Adding the target speaker is agreed implementation work. Substantial expansion
of validation context remains evidence-dependent: record a concrete case and a
focused follow-up recommendation if needed. Retain the current context selection
and fitting behavior otherwise. Stronger instructions cannot supply missing
evidence, and this roadmap does not require a broader context redesign.
### 5. Add Focused Editorial and Repair Examples
Add a small number of examples where they clarify judgments that current
instructions leave implicit:
- Apparent filler that contributes meaning, such as `like` expressing
approximation or `I mean` introducing a substantive self-correction. Use
surrounding context to contrast meaningful usage with removable dysfluency.
- One compact repair revision and one withdrawal, using the existing response
schema and making withdrawal a successful preservation of the original text.
- A combined overlapping edit illustrating step 2.
Place examples in the relevant proposal, validation, or shared repair guidance.
Keep existing useful contextual contrasts. Examples should illustrate reasoning
without becoming unconditional token-based editing rules or an extensive
rulebook. Maintain the narrow meaning-reversal instructions.
### 6. Verify Integration, Documentation, and Test-Policy Compliance
Follow the [testing policy](../policy/testing.md) throughout implementation and
finish with a focused review of affected tests:
- Verify repair context reflects the current proposal and the shared preview
behavior, including representative deletion, replacement-policy, subsequent
revision, and unavailable-preview cases where coverage is warranted.
- Verify target-speaker data reaches validation and participates in existing
payload accounting. Preserve coverage of response handling, revision,
withdrawal, revalidation, and operator prompt replacement.
- Review rendered messages and representative complete repair conversations
directly for instruction authority, editorial intent, and format consistency.
Check response examples against their schemas.
- Remove any affected tests that freeze bundled prompt wording through phrase
checks, exact prose comparisons, or snapshots. Do not replace them with checks
for the new wording. Retain tests of supplied data and meaningful machine
contracts where their protective value justifies their maintenance cost.
- Update the relevant canonical prompt, module, validator, and runtime documents
as behavior lands. Update maintained examples where affected; check links and
whitespace.
Run the appropriate package checks, build, and full offline suite for the
implementation under the [development guidance](../development.md#baseline-validation).
Report the test-policy review's meaningful removals, revisions, and retained
coverage. Offline tests establish data and execution contracts; they do not
establish live-model editorial quality.
## Deferred Evidence and Follow-Up
The evidence review examined the maintained local, campaign, and prompt-edit
transcript examples, adjacent-context validation fixtures, and proposal
application fixtures. The glossary example is a single-segment correction;
the validation fixture supplies the troupe definition and adjacent evidence
needed to assess its proposed name change. These sources contain no recorded
live-model rejection demonstrating a need for more distant context. The target
speaker is now included; broader selection and context fitting remain unchanged.
Revisit the window when a captured case shows material evidence in the proposal
request that was absent from validation.
The existing stale-span application fixture demonstrates the mechanical overlap
addressed by the new guidance. It does not establish a need for new combined
semantic validation, so application ordering and rejection rules remain unchanged.
Formal evaluation corpus preparation and live-model comparisons remain deferred
and are not prerequisites for the agreed changes. Broader validation context and
additional handling of combined semantic effects should follow concrete evidence
rather than hypothetical restrictions. Any later evaluation should examine
needed and unnecessary edits, missed corrections, validator disagreement, and
repair outcomes using independent editorial judgment. Keep live calls opt-in
and separate from the default offline suite.
No blocking product decisions remain for the implementation scope above.
## Completion and Verification
- The four proposal definitions now separate durable system principles from
their initial user assignments. The shared proposal guidance covers overlapping
and dependent edits; the grammar prompt illustrates a combined edit.
- Every repair attempt receives the current target's original text when
available and its previewed result when applicable, using the same domain
replacement function as validation. Revision and withdrawal examples accompany
the shared repair guidance.
- Validation includes the target speaker in rendered input and token accounting.
Spoken-word proposal and editorial review examples distinguish meaningful
apparent filler from removable hesitation. The meaning-reversal definition and
all response schemas retain their existing contracts.
- Behavioral tests cover repair deletion and replace-all previews, ambiguous or
missing spans, blank results, missing targets, and refresh after retargeting.
Existing validation tests now cover target-speaker propagation, absence, and
budget fitting, including parity with serialized payload estimates.
- The affected prompt, module, request-assembly, and conversation tests were
reviewed for wording checks. No further bundled-prose change detectors were
found to remove; the earlier cleanup remains effective. Retained assertions
protect supplied-data rendering, schema selection, operator replacement,
conversation history, revalidation, withdrawal, and outcomes. New assertions
check request data and behavior rather than prompt phrases.
- All eleven definitions prepared through the CLI. Four illustrative complete
conversations, including subsequent repair requests, were reviewed, and the
sixteen embedded JSON response examples passed schema checks. The application
build and full offline Go suite passed. Canonical documentation and maintained
preview examples were updated; links and whitespace were checked.
These checks verify implementation and editorial intent, not live-model quality.
No blocking product decisions remain for this follow-up scope.

View File

@@ -1,168 +0,0 @@
# PromptKit v0.23.0 Implementation Plan
**Status:** Ready for implementation; all tasks below are pending.
The [adoption roadmap](promptkit.md) owns the target behavior and boundaries.
This plan sequences the direct v0.21.0-to-v0.23.0 migration and the evidence needed
to establish full adoption. It does not authorize application code changes during
the current documentation pass.
## 1. Upgrade and unify execution
- [ ] Update `go.mod` and `go.sum` directly to
`go.maximumdirect.net/promptkit v0.23.0`. Use released module resolution; remove
temporary replacements/workspace files. Check the selected module graph and
compatibility with the existing CommandKit/ckprompt integration.
- [ ] Refactor [the LLM adapter](../../internal/framework/llm/adapter.go) around
shared request construction, response interpretation, and error mapping that
can serve both ordinary calls and coordinator-owned turns. Keep those helpers
narrow; do not wrap `Coordinate` in another turn loop.
- [ ] Move whole-operation timing to `RunRequest.OperationTimeout`. Use ordinary
`Run` for calls whose prepare/run split exists only to obtain diagnostics.
Avoid retaining a redundant child timer around the same ordinary operation.
- [ ] Supply `OnPrepared` to collect effective metadata and selected credential
source names, and retain prepared content for debug when enabled. Use the
delivered snapshot without a separate `Prepare` call. Clear per-call state
so a failure before notification cannot reuse a prior turn's evidence.
- [ ] Preserve per-role profile/schema/variable selection, transient and
structural-repair budgets, ordinary observers, output decoding, and error
disposition. Map `ErrOperationTimeout` explicitly, including timeouts during
preparation. Distinguish parent cancellation and provider-local deadlines.
Retain the public processing reason codes where their meanings still apply.
- [ ] Ensure call finalization runs on both results and errors; coordinator
execution errors do not arrive as ordinary response callbacks. Preserve the
last attempted operation's diagnostic and failure context for final reporting.
**Verification:** focused adapter tests for exact prepared evidence after a
provider failure, absent evidence before notification, independent operation
budgets, timeout classification, metadata redaction, and ordinary observer
composition. Use controlled model clients and cancellation. Retain explicit
prepared execution only where an identified consumer actually requires it.
## 2. Replace history and repair-loop mechanics
- [ ] Replace the local `Exchange`/`Replay` model in
[structured completion contracts](../../internal/framework/structuredoutput/call.go)
with a minimal opaque bridge to native captured exchanges. The adapter owns
PromptKit execution; modules and validators must not reconstruct messages or
depend on provider-specific state.
- [ ] Capture the original proposal run through `CaptureExchange`. Select the
repair prompt directly for a continuation; remove original-prompt rerendering,
manual assistant/suffix assembly, prefix comparison, and `RepairInstruction`
history plumbing. Preserve prepared snapshots used only for diagnostic output.
- [ ] Replace the attempt loop in
[repair coordination](../../internal/framework/runner/repair.go) with one
adapter-backed `Engine.Coordinate` invocation per actionable rejected proposal.
Keep scheduling across proposals in Audita. Pass the resolved correction repair
limit as `AdditionalTurns`, and bypass coordination when repair is disabled.
- [ ] Keep domain branch state in the callback: original proposal identity,
current revision, local segment mapping, current before/after preview, safe
rejection findings, and domain outcome. The initial callback uses existing
findings without rerunning the initial batch validation.
- [ ] Build each follow-up request with the repair definition, role profile,
repair schema, variables, operation timeout, prepared notification, and selected
observers. Let the coordinator attach `ContinueFrom` and manage history.
- [ ] Process `ConversationTurn.Result` only when present. Check structural
validation before schema decoding, interpret withdrawal, and revalidate every
revision from the beginning using the parent callback context. Validator calls
select their own profiles, timeouts, and observers through the shared adapter.
- [ ] Return a nil next request for accepted, withdrawn, or unavailable domain
outcomes. If rejection remains actionable, request a further turn even at the
budget boundary so PromptKit reports `CoordinationExhausted`; do not recreate
an independent exhaustion loop. The final permitted revision is still assessed.
- [ ] Keep diagnostic attempt coordinates and operation records, but allocate
execution evidence only for runs actually started. The callback may construct
a request that is never executed because the turn budget is exhausted.
- [ ] Map execution errors, callback errors, stopping, exhaustion, and
`Undelivered` separately. Unassessed or structurally invalid content never
reaches the approved set. Preserve accepted sibling proposals and apply-time
checks; fatal/canceled work still cancels and joins peers.
**Verification:** a controlled workflow with one initial approval and two rejected
proposals demonstrates independent histories, withdrawal, revised acceptance,
full revalidation, and deterministic application. Cover final-turn acceptance,
exhaustion, zero repair, structural invalidity, recoverable failures, parent
cancellation before delivery, and capacity-one validator calls. Assertions should
exercise Audita outcomes through the real upstream coordinator with a fake model
boundary, rather than test a fake coordinator's loop behavior.
## 3. Adopt native execution accounting
- [ ] Introduce an invocation-scoped `ExecutionGroup` and compose its `Observe`
callback with existing diagnostic observers using `ComposeObservers`. Ensure
observation remains enabled wherever required accounting is consumed,
independently of whether debug artifacts are enabled.
- [ ] Explicitly enroll proposal, repair-generation, and validator runs. Use
`CoordinateRequest.Observer` for continuation membership and select membership
separately for nested validation requests. Keep conversation parent identity
distinct from Audita's stage/section/proposal/validator labels.
- [ ] Replace the terminal-map deduplication and usage summation in
[observation summaries](../../internal/framework/llm/observations.go) with
`ExecutionGroup.Snapshot()`. Convert upstream summary fields only at presentation
boundaries. Retain logical call metadata, safe event evidence, statuses, and
deterministic ordering that the upstream group does not supply.
- [ ] Carry terminal coverage separately from usage known/completeness into the
internal summary and its diagnostic representation. Do not present missing
terminals as complete zero accounting. Keep the established ordinary report
scope; adding token payloads to the ordinary errors/recoveries report is not
part of this upgrade.
- [ ] Count shared initial work once in the invocation group; count each new
continuation and validator run separately. Do not sum overlapping subgroup
totals or attempt events on top of their terminal totals. Sort any presented
member lists deterministically rather than relying on native snapshot ordering.
- [ ] Snapshot final totals after all selected work has joined. Groups add no
separate close/wait lifecycle. Audit and delete local aggregation that became
redundant; document the purpose of any remaining application-specific counters.
**Verification:** integration tests demonstrate composition with existing
observers, explicit validator membership, shared initial work, failure terminal
facts, unknown versus zero usage, missing terminal coverage, and concurrent calls.
Reuse upstream group semantics; test Audita's enrollment and presentation instead
of duplicating the library's unit suite.
## 4. Cleanup, documentation, and acceptance
- [ ] Audit references to the old exchange/replay types, prefix checks,
`RepairInstruction`, local repair attempt loop, redundant timeout wrapper, and
terminal aggregation. Delete dead production helpers and replace tests tied to
those implementations with focused behavior coverage. A thin local adapter is
justified; a parallel conversation/coordinator/accounting implementation is not.
- [ ] Update the canonical [LLM runtime](../internal/llm-runtime.md),
[pipeline](../internal/pipeline.md),
[diagnostics/reporting internals](../internal/diagnostics-reporting.md), and
[LLM integration contract](../integrations/openai-compatible-llm.md) as behavior
lands. Update pinned upstream references and dependency attribution as needed.
Document the selected notification boundary and timeout mapping without claiming
that preparation or admission failures always have a prepared snapshot.
- [ ] Update configuration/operations references only for actual externally
visible changes. No prompt rewrite, validator tuning, new configuration format,
compatibility shim, new scheduler, or unavailable-decision regeneration is
required by this migration.
- [ ] Follow the [testing policy](../policy/testing.md): keep tests that protect
domain and integration behavior, remove obsolete implementation-coupled tests,
and avoid prompt word/phrase change detectors. Review final tests for redundancy.
- [ ] Run the focused migration tests, the complete Audita suite, and a race-enabled
check of changed concurrent adapter/runner/accounting paths. Build the CLI and
exercise existing offline processing and inspection coverage. No paid provider
calls are required for acceptance.
- [ ] Review the final diff, verify local Markdown links and `git diff --check`,
and confirm no local dependency replacements or workspace artifacts remain.
Verify the selected PromptKit version is v0.23.0.
- [ ] Confirm all three upstream capability families are used in production
processing and the agreed local machinery is retired. Record completion only
after this end-state audit; move useful current contracts into their canonical
owners and retire these temporary planning documents under repository policy.
Suggested final checks, using the locally selected Go 1.27.0-or-newer toolchain:
```sh
go list -m go.maximumdirect.net/promptkit
go test ./...
go test -race ./internal/framework/llm ./internal/framework/runner ./internal/framework/processreport
go build -o ./bin/audita ./cmd/audita
git diff --check
```
Adjust focused package checks if code ownership moves. A green dependency-only
suite is not evidence that the new workflow has been adopted; acceptance also
requires the coordinator, conversation, and group integration scenarios above.

View File

@@ -1,136 +0,0 @@
# PromptKit v0.23.0 Adoption Roadmap
**Status:** Accepted target; implementation pending. This pass changes planning
documents only. The [implementation plan](promptkit-implementation.md) owns the
ordered tasks and verification.
## Objective and Baseline
Upgrade Audita directly from PromptKit v0.21.0 to v0.23.0 and fully adopt its
conversation, coordinator, and execution-group APIs. Removing redundant local
machinery is an explicit completion requirement. A dependency-only upgrade or
continuation adoption with the old repair loop retained is an intermediate state.
The current [dependency manifest](../../go.mod) pins v0.21.0. Audita captures its
own exchange and replay values, reconstructs and compares conversation prefixes,
appends repair history, manages the revision loop, and aggregates terminal
observations locally. Each role client wraps explicit preparation and execution
in a local deadline. See the [LLM runtime](../internal/llm-runtime.md),
[pipeline](../internal/pipeline.md), and
[diagnostic implementation](../internal/diagnostics-reporting.md).
[PromptKit v0.23.0](https://forge.maximumdirect.net/base-layer/promptkit/src/tag/v0.23.0/docs/releases/v0.23.0.md)
includes the v0.22.0 conversation and accounting capabilities plus the execution
controls needed for full coordinator adoption. The former upstream timeout and
prepared-request gaps are closed; no intermediate v0.22.0 rollout or further
upstream feature gate is required. The Go minimum and dependency versions are
unchanged by these upstream releases.
The release review verified Audita's existing suite against the tagged v0.23.0
source through a temporary module override and selected upstream packages with
race detection. That is compatibility evidence, not verification of the migration
described here. Implementation must use the released dependency, without a local
replacement or committed workspace file.
## Target Ownership
| Responsibility | Target owner |
| --- | --- |
| Immutable captured history, continuation assembly, and parent run identity | PromptKit `Exchange`, `CaptureExchange`, and `ContinueFrom` |
| Sequential turns, additional-turn budget, callback delivery, and coordination termination | PromptKit `Engine.Coordinate` |
| Whole-operation deadlines and timeout attribution | PromptKit `OperationTimeout` and `ErrOperationTimeout` |
| Exact prepared-request notification before generation | PromptKit `OnPrepared` |
| Selected run membership, deduplication, and terminal count/usage aggregation | PromptKit `ExecutionGroup` and `ComposeObservers` |
| Candidate identity, actionable rejection, feedback, revision decoding, withdrawal, and full revalidation | Audita |
| Branch scheduling, accepted sibling proposals, application order, and transcript mutation | Audita |
| Domain operation labels, failure disposition, reports, redaction, and debug artifact storage | Audita |
Use the supported public API through Audita's framework adapter boundaries.
Keep module and validator contracts domain-oriented. Any bridge needed by the
runner must delegate to the real coordinator and retain opaque conversation
state, rather than implementing another conversation engine or general workflow
framework.
## Selected Behavior
### Execution and diagnostics
Use ordinary `Engine.Run` with per-request `OperationTimeout` and `OnPrepared`
where explicit preparation currently serves only to inspect and execute the same
request. Map the resolved Audita operation timeout onto each proposal, validation,
and repair operation. Keep provider, transport-retry, structural-repair, and
correction-revision budgets distinct.
Use `ErrOperationTimeout` to identify the selected local budget, while retaining
Audita's parent-cancellation and recoverable-provider-failure handling. Do not
classify every wrapped preparation error as fatal before considering its timeout
identity. Ordinary configuration failures remain fatal. If split preparation is
still needed for a concrete consumer, one caller context must cover both calls
when a total preparation-plus-execution deadline is required; the upstream
prepared-handle duration starts only at execution.
`OnPrepared` receives the actual detached snapshot. Retain the metadata needed
for attribution and credential-source redaction, and capture content for optional
debug evidence. Keep the callback short; application validation belongs in the
next-turn callback after admission is released. Notification may be absent on
preparation/admission failure or cancellation before delivery. Reports must not
invent missing snapshots. Optional debug write failures remain nonfatal.
### Conversations and correction repair
Capture the proposal exchange when repair needs it. Every rejected candidate's
branch starts from that same immutable exchange; subsequent turns remain private
to that branch. Follow-ups select the module's repair prompt, role profile,
variables, and repair output contract through ordinary PromptKit requests.
Retain native operator prompt overrides and the existing domain prompt content.
Use one coordinator per actionable correction. Its initial callback uses existing
rejection findings. Each newly delivered response must pass structural validation
before decoding; captured content alone does not imply validity. A decoded
revision is fully revalidated before acceptance. Withdrawal, unavailable results,
and accepted revisions remain Audita outcomes; a normal coordinator stop is not
itself an approval.
The last permitted response follows the same assessment path. Further requested
work beyond the budget maps to exhaustion. Skip coordination when Audita's
correction-repair setting disables it: upstream's zero budget still invokes an
initial callback. Cancellation's `Undelivered` response is diagnostic evidence,
not an approved correction, and must not trigger model work after cancellation.
Recoverable branch failures preserve independent approvals; fatal failures and
caller cancellation retain Audita's cancel-and-join behavior.
### Execution groups
Use native groups as the source of aggregate provider-attempt, retry, repair,
and available token totals. Explicitly select membership for proposal calls,
coordinated continuations, and validator calls. Coordinator observers cover its
own turns; nested validators require their own observer selection.
Maintain an invocation group for combined execution totals. Add narrower groups
only for accounting views Audita consumes. A shared initial proposal run belongs
once to the invocation total; replaying its history does not enroll it again.
Each new call's reported history tokens still count. Overlapping group totals
must not be summed as disjoint work.
Preserve upstream usage presence/completeness and terminal coverage separately.
Keep domain logical-call records and deterministic presentation where needed;
their count need not equal observed run membership. Group summaries do not replace
Audita's errors/recoveries, validation decisions, or per-operation failure facts.
## Delivery and Completion
Implement in ordered, reviewable stages: dependency and execution adapter;
captured conversations and coordinator; execution groups; cleanup, integration
verification, and canonical documentation. All stages belong to this adoption.
Completion requires removal of local exchange/replay history, prefix checks,
repair turn-budget loops, and duplicate terminal aggregation where replaced by
PromptKit. Retain local behavior only where it has a concrete Audita domain or
presentation responsibility. Preserve transcript correctness, repair prompts,
validation semantics, deterministic reports, standard inspection commands, and
operator configuration unless a deliberate change is documented and tested.
No product-level open decisions remain. Exact internal bridge signatures are
implementation choices governed by this ownership split. Future validator
decision recovery and new scheduling behavior remain in the
[deferred work catalog](future.md).

4
go.mod
View File

@@ -3,11 +3,12 @@ module gitea.maximumdirect.net/eric/audita/v3
go 1.27.0
require (
github.com/santhosh-tekuri/jsonschema/v6 v6.0.2
go.maximumdirect.net/ckconfig v0.7.0
go.maximumdirect.net/ckprompt v0.7.0
go.maximumdirect.net/commandkit v0.6.0
go.maximumdirect.net/configkit v1.2.0
go.maximumdirect.net/promptkit v0.21.0
go.maximumdirect.net/promptkit v0.23.0
go.maximumdirect.net/releasekit v0.9.0
gopkg.in/yaml.v3 v3.0.1
)
@@ -16,7 +17,6 @@ require (
github.com/inconshreveable/mousetrap v1.1.0 // indirect
github.com/kr/pretty v0.3.1 // indirect
github.com/rogpeppe/go-internal v1.14.1 // indirect
github.com/santhosh-tekuri/jsonschema/v6 v6.0.2 // indirect
github.com/spf13/cobra v1.10.1 // indirect
github.com/spf13/pflag v1.0.9 // indirect
github.com/yuin/goldmark/v2 v2.1.6 // indirect

4
go.sum
View File

@@ -32,8 +32,8 @@ go.maximumdirect.net/commandkit v0.6.0 h1:4PHhMIlZuuvreIxZhUazeweg6zoYixEvhM9LJn
go.maximumdirect.net/commandkit v0.6.0/go.mod h1:ZS98DewKNVBygKVDoh5MFD09CnPlP37dS9pVsID5Eoc=
go.maximumdirect.net/configkit v1.2.0 h1:c4tDJ6Xeh/Jg90UZplLxsnf500lvrZBVj0E3i1ysir8=
go.maximumdirect.net/configkit v1.2.0/go.mod h1:s1zLJ9GGB/1Q4a+aOHMLISonJxWFgNWpA6SpPlOafeA=
go.maximumdirect.net/promptkit v0.21.0 h1:Ws4y28lnW+lIXAq4tGhXVFnQV/6AqkDChhg8knIZFrU=
go.maximumdirect.net/promptkit v0.21.0/go.mod h1:pi9acdCi80HWyzY+XYwRvMEoCiTcDWKyeBh3mzHdIFo=
go.maximumdirect.net/promptkit v0.23.0 h1:Q17wmRHWSznc1yW1MuB7dxuT3KVYRCzHI9pKBzpXhq0=
go.maximumdirect.net/promptkit v0.23.0/go.mod h1:pi9acdCi80HWyzY+XYwRvMEoCiTcDWKyeBh3mzHdIFo=
go.maximumdirect.net/releasekit v0.9.0 h1:sK+Aoj6gS6ATU/1reocjwBekp4YETYENwkcX5ti6mrE=
go.maximumdirect.net/releasekit v0.9.0/go.mod h1:QE4BLrAhVJTzi+/YTcVl21K/qZp9ojRLg5kuoy3Zb44=
go.yaml.in/yaml/v4 v4.0.0-rc.6 h1:1h7H1ohdUh93/FyE4YaDa1Zh64K6VVbjF4K6WUxMtH4=

View File

@@ -5,7 +5,6 @@ import (
"encoding/json"
"errors"
"fmt"
"reflect"
"sync"
"time"
@@ -47,13 +46,20 @@ func (r *Runtime) NewRoleClient(role, profileID string, settings config.LLM, hoo
return &RoleClient{runtime: r, role: role, profileID: profileID, settings: settings, hooks: hooks}, nil
}
func (c *RoleClient) CompleteStructured(parent context.Context, req structuredoutput.Request, out any) (response structuredoutput.Response, err error) {
ctx, cancel := context.WithTimeout(parent, c.settings.OperationTimeout)
defer cancel()
type pendingCall struct {
request pk.RunRequest
finish func(*pk.RunResult, error) (structuredoutput.Response, error)
}
func (c *RoleClient) prepareCall(parent context.Context, req structuredoutput.Request, out any, observer pk.ExecutionObserver) (*pendingCall, error) {
var response structuredoutput.Response
var err error
var finishHooks []func()
ctx := parent
var terminalMu sync.Mutex
var terminal pk.ExecutionEvent
terminalAvailable := false
defer func() {
finishHooks = append(finishHooks, func() {
terminalMu.Lock()
if terminalAvailable {
response.Execution = outcomes.ExecutionCounts{
@@ -69,7 +75,7 @@ func (c *RoleClient) CompleteStructured(parent context.Context, req structuredou
}
}
terminalMu.Unlock()
}()
})
var record structuredoutput.CallRecord
if c.hooks.OnCall != nil || c.hooks.OnEventWithCall != nil {
record.CallID = c.runtime.nextCallID.Add(1)
@@ -81,7 +87,7 @@ func (c *RoleClient) CompleteStructured(parent context.Context, req structuredou
record.Started = time.Now().UTC()
record.Prompt = structuredoutput.PromptMetadata{PromptID: req.PromptID, ProfileID: c.profileID}
onCall(record)
defer func() {
finishHooks = append(finishHooks, func() {
record.Finished = time.Now().UTC()
if err != nil {
var failure *structuredoutput.Failure
@@ -92,123 +98,120 @@ func (c *RoleClient) CompleteStructured(parent context.Context, req structuredou
}
}
onCall(record)
}()
})
}
finish := func() {
for i := len(finishHooks) - 1; i >= 0; i-- {
finishHooks[i]()
}
}
fail := func(failure error) (*pendingCall, error) {
err = failure
finish()
return nil, failure
}
if parent.Err() != nil {
return response, classifiedFailure(parent, ctx, parent.Err(), nil)
return fail(classifiedFailure(parent, ctx, parent.Err(), nil))
}
if err = validateOutputTarget(out); err != nil {
return response, &structuredoutput.Failure{Class: structuredoutput.FailureFatal, Reason: outcomes.ReasonSetupFailure, Message: "invalid structured output target", Cause: err}
return fail(&structuredoutput.Failure{Class: structuredoutput.FailureFatal, Reason: outcomes.ReasonSetupFailure, Message: "invalid structured output target", Cause: err})
}
contract, contractErr := OutputContract(req.Schema, c.settings.RepairAttempts)
if contractErr != nil {
return response, &structuredoutput.Failure{Class: structuredoutput.FailureFatal, Reason: outcomes.ReasonSetupFailure, Message: "unknown structured response schema", Cause: contractErr}
return fail(&structuredoutput.Failure{Class: structuredoutput.FailureFatal, Reason: outcomes.ReasonSetupFailure, Message: "unknown structured response schema", Cause: contractErr})
}
vars := prompts.WithSharedVariables(req.Variables)
for name, value := range prompts.SharedVariables() {
vars[name] = value
}
baseID, baseVars := req.PromptID, vars
var suffix []pk.RenderedMessage
if req.Replay != nil {
if req.Replay.Exchange == nil {
return response, &structuredoutput.Failure{Class: structuredoutput.FailureFatal, Reason: outcomes.ReasonSetupFailure, Message: "repair exchange is unavailable"}
var exchange *pk.Exchange
if req.ContinueFrom != nil {
var ok bool
exchange, ok = req.ContinueFrom.State.(*pk.Exchange)
if !ok || exchange == nil {
return fail(&structuredoutput.Failure{Class: structuredoutput.FailureFatal, Reason: outcomes.ReasonSetupFailure, Message: "repair exchange is unavailable"})
}
instruction, renderErr := c.runtime.Engine.Prepare(ctx, pk.RunRequest{PromptID: req.PromptID, ProfileID: c.profileID, Vars: vars})
if renderErr != nil {
return response, classifiedFailure(parent, ctx, renderErr, nil)
}
response.RepairInstruction = convertMessages(instruction.Messages)
exchange := req.Replay.Exchange
baseID, baseVars = exchange.PromptID, exchange.Variables
suffix = append(suffix, pk.RenderedMessage{Role: pk.RoleAssistant, Content: exchange.FinalContent})
for _, message := range req.Replay.Turns {
native := pk.RenderedMessage{Role: message.Role, Content: message.Content}
if cache := message.CacheControl; cache != nil {
native.CacheControl = &pk.CacheControl{Type: pk.CacheControlType(cache.Type), TTL: cache.TTL}
}
suffix = append(suffix, native)
}
suffix = append(suffix, instruction.Messages...)
}
handle, prepErr := c.runtime.Engine.PrepareExecution(ctx, pk.RunRequest{
PromptID: baseID, ProfileID: c.profileID, Vars: baseVars, AppendedMessages: suffix,
runRequest := pk.RunRequest{
PromptID: req.PromptID, ProfileID: c.profileID, Vars: vars,
CaptureExchange: req.CaptureExchange, ContinueFrom: exchange,
TransientRetries: &c.settings.TransientRetries,
Validation: contract,
Observer: func(event pk.ExecutionEvent) {
Observer: pk.ComposeObservers(observer, func(event pk.ExecutionEvent) {
if event.Kind == pk.ExecutionFinished {
terminalMu.Lock()
terminal = event
terminalAvailable = true
terminalMu.Unlock()
}
if c.hooks.OnEvent != nil {
c.hooks.OnEvent(event)
}
if c.hooks.OnEventWithCall != nil {
c.hooks.OnEventWithCall(record.CallID, record.Context, event)
}
},
})
if prepErr != nil {
return response, classifiedFailure(parent, ctx, prepErr, nil)
}),
}
defer handle.Discard()
details := handle.Details()
if req.Replay != nil {
prefix := req.Replay.Exchange.Messages
if len(details.Messages) < len(prefix) || !reflect.DeepEqual(convertMessages(details.Messages[:len(prefix)]), prefix) {
return response, &structuredoutput.Failure{Class: structuredoutput.FailureFatal, Reason: outcomes.ReasonInternalFailure, Message: "repair conversation prefix changed"}
var details pk.PreparedRun
capturePrepared := func(_ context.Context, prepared pk.PreparedRun) {
details = prepared
if name := details.EffectiveModelParams.APIKeyEnv; name != "" {
c.credentialMu.Lock()
if c.credentialNames == nil {
c.credentialNames = make(map[string]struct{})
}
c.credentialNames[name] = struct{}{}
c.credentialMu.Unlock()
}
response.Prompt = promptMetadata(details)
record.Prompt = response.Prompt
response.Model = details.EffectiveModelParams.Model
response.Prepared = preparedSnapshot(details)
record.Prepared = response.Prepared
}
if name := details.EffectiveModelParams.APIKeyEnv; name != "" {
c.credentialMu.Lock()
if c.credentialNames == nil {
c.credentialNames = make(map[string]struct{})
runRequest.OperationTimeout = c.settings.OperationTimeout
runRequest.OnPrepared = capturePrepared
return &pendingCall{request: runRequest, finish: func(result *pk.RunResult, runErr error) (completed structuredoutput.Response, finalErr error) {
defer func() { err = finalErr; finish(); completed = response }()
if result != nil {
response.FinalContent = result.RawOutput
record.FinalContent = result.RawOutput
response.Model = result.ModelName
response.PromptTokens = result.Usage.PromptTokens
response.CompletionTokens = result.Usage.CompletionTokens
response.TotalTokens = result.Usage.TotalTokens
}
c.credentialNames[name] = struct{}{}
c.credentialMu.Unlock()
}
response.Prompt = promptMetadata(details)
record.Prompt = response.Prompt
response.Model = details.EffectiveModelParams.Model
response.Prepared = preparedSnapshot(details)
record.Prepared = response.Prepared
result, runErr := c.runtime.Engine.RunPrepared(ctx, handle)
if result != nil {
response.FinalContent = result.RawOutput
record.FinalContent = result.RawOutput
response.Model = result.ModelName
response.PromptTokens = result.Usage.PromptTokens
response.CompletionTokens = result.Usage.CompletionTokens
response.TotalTokens = result.Usage.TotalTokens
}
if runErr != nil {
return response, classifiedFailure(parent, ctx, runErr, result)
}
if result == nil {
return response, &structuredoutput.Failure{Class: structuredoutput.FailureFatal, Reason: outcomes.ReasonInternalFailure, Message: "structured model call returned no result"}
}
if !result.Validation.IsValid {
return response, classifiedFailure(parent, ctx, nil, result)
}
if decodeErr := json.Unmarshal([]byte(result.RawOutput), out); decodeErr != nil {
var typeError *json.UnmarshalTypeError
if errors.As(decodeErr, &typeError) {
return response, &structuredoutput.Failure{Class: structuredoutput.FailureRecoverable, Reason: outcomes.ReasonInvalidOutput, Message: "structured output contains an unusable value", Cause: decodeErr}
if runErr != nil {
return response, classifiedFailure(parent, ctx, runErr, result)
}
return response, &structuredoutput.Failure{Class: structuredoutput.FailureFatal, Reason: outcomes.ReasonInternalFailure, Message: "validated structured output could not be decoded", Cause: decodeErr}
}
if req.CaptureExchange && req.Replay == nil {
frozen := make(map[string]string, len(vars))
for key, value := range vars {
frozen[key] = value
if result == nil {
return response, &structuredoutput.Failure{Class: structuredoutput.FailureFatal, Reason: outcomes.ReasonInternalFailure, Message: "structured model call returned no result"}
}
response.Exchange = &structuredoutput.Exchange{PromptID: req.PromptID, Variables: frozen, Messages: convertMessages(details.Messages), FinalContent: result.RawOutput}
if !result.Validation.IsValid {
return response, classifiedFailure(parent, ctx, nil, result)
}
if decodeErr := json.Unmarshal([]byte(result.RawOutput), out); decodeErr != nil {
var typeError *json.UnmarshalTypeError
if errors.As(decodeErr, &typeError) {
return response, &structuredoutput.Failure{Class: structuredoutput.FailureRecoverable, Reason: outcomes.ReasonInvalidOutput, Message: "structured output contains an unusable value", Cause: decodeErr}
}
return response, &structuredoutput.Failure{Class: structuredoutput.FailureFatal, Reason: outcomes.ReasonInternalFailure, Message: "validated structured output could not be decoded", Cause: decodeErr}
}
if result.Exchange != nil {
response.Exchange = &structuredoutput.Exchange{State: result.Exchange}
}
response.Content = json.RawMessage(result.RawOutput)
response.Provider = "openai-compatible"
return response, nil
}}, nil
}
func (c *RoleClient) CompleteStructured(ctx context.Context, req structuredoutput.Request, out any) (structuredoutput.Response, error) {
call, err := c.prepareCall(ctx, req, out, c.hooks.OnEvent)
if err != nil {
return structuredoutput.Response{}, err
}
response.Content = json.RawMessage(result.RawOutput)
response.Provider = "openai-compatible"
return response, nil
result, err := c.runtime.Engine.Run(ctx, call.request)
return call.finish(result, err)
}
// OutputContract maps an Audita domain schema to the native request contract.
@@ -231,6 +234,12 @@ func classifiedFailure(parent, local context.Context, cause error, result *pk.Ru
failure.Cause = parent.Err()
return failure
}
if errors.Is(cause, pk.ErrOperationTimeout) {
failure.Class = structuredoutput.FailureRecoverable
failure.Reason = outcomes.ReasonOperationTimeout
failure.Message = "model operation timed out"
return failure
}
if isSetupFailure(cause) {
failure.Reason = outcomes.ReasonSetupFailure
failure.Message = "structured model setup failed"
@@ -298,10 +307,7 @@ func promptMetadata(details pk.PreparedRun) structuredoutput.PromptMetadata {
}
func preparedSnapshot(details pk.PreparedRun) *structuredoutput.PreparedSnapshot {
messages := make([]structuredoutput.Message, len(details.Messages))
for i, message := range details.Messages {
messages[i] = structuredoutput.Message{Role: string(message.Role), Content: message.Content}
}
messages := convertMessages(details.Messages)
return &structuredoutput.PreparedSnapshot{Messages: messages, Prompt: promptMetadata(details)}
}

View File

@@ -482,3 +482,14 @@ func TestPreparedAdapterUnknownValidationFailureRemainsFatal(t *testing.T) {
t.Fatalf("unknown validation failure: %#v", failure)
}
}
// Preparation errors can carry the native operation deadline as their cause.
// That deadline is recoverable even when the outer category is a setup phase.
func TestOperationTimeoutDuringPreparationIsRecoverable(t *testing.T) {
ctx := context.Background()
cause := errors.Join(pk.ErrPromptLoad, pk.ErrOperationTimeout, context.DeadlineExceeded)
failure := classifiedFailure(ctx, ctx, cause, nil)
if failure.Class != structuredoutput.FailureRecoverable || failure.Reason != outcomes.ReasonOperationTimeout || !errors.Is(failure, pk.ErrOperationTimeout) {
t.Fatalf("failure = %#v", failure)
}
}

View File

@@ -19,141 +19,9 @@ import (
"gitea.maximumdirect.net/eric/audita/v3/internal/framework/structuredoutput"
"gitea.maximumdirect.net/eric/audita/v3/internal/prompts"
ck "go.maximumdirect.net/configkit"
pk "go.maximumdirect.net/promptkit"
)
func TestNativeConversationReplayWithRepairContract(t *testing.T) {
dir := t.TempDir()
files := map[string]string{
"config.yml": "version: 3\n",
"prompts.yml": `profiles:
- id: proposal-profile
endpoint: https://example.invalid/v1
model: proposal-model
- id: repair-profile
endpoint: https://example.invalid/v1
model: repair-model
prompts:
- id: proposal-fixture
messages:
- role: system
content: 'System {{.Subject}}'
- role: user
content: 'Proposal {{.Subject}}'
output:
format: json
validation_mode: json_schema
schema_path: correction_set.v1.json
- id: repair-fixture
messages:
- role: user
content: 'Repair {{.Finding}}'
output:
format: text
validation_mode: none
`,
}
for name, content := range files {
if err := os.WriteFile(filepath.Join(dir, name), []byte(content), 0644); err != nil {
t.Fatal(err)
}
}
loaded, err := config.Load(ck.Input{Source: ck.File(filepath.Join(dir, "config.yml"))}, "")
if err != nil {
t.Fatal(err)
}
var requests []struct {
Model string `json:"model"`
Messages []struct{ Role, Content string } `json:"messages"`
ResponseFormat map[string]any `json:"response_format"`
}
outputs := []string{`{"corrections":[]}`, `{"proposal_ref":1,"action":"revise","correction":{"id":1,"original_text":" ","corrected_text":""}}`, `{"proposal_ref":1,"action":"withdraw","correction":null}`}
transport := transportFunc(func(r *http.Request) (*http.Response, error) {
var request struct {
Model string `json:"model"`
Messages []struct{ Role, Content string } `json:"messages"`
ResponseFormat map[string]any `json:"response_format"`
}
if err := json.NewDecoder(r.Body).Decode(&request); err != nil {
t.Fatal(err)
}
requests = append(requests, request)
content, _ := json.Marshal(outputs[len(requests)-1])
body := fmt.Sprintf(`{"model":%q,"choices":[{"message":{"content":%s}}]}`, request.Model, content)
return &http.Response{StatusCode: 200, Header: make(http.Header), Body: io.NopCloser(strings.NewReader(body)), Request: r}, nil
})
runtime, err := NewRuntime(context.Background(), loaded, &http.Client{Transport: transport})
if err != nil {
t.Fatal(err)
}
ctx := context.Background()
base := pk.RunRequest{PromptID: "proposal-fixture", ProfileID: "proposal-profile", Vars: map[string]string{"Subject": "source"}}
original, err := runtime.Engine.PrepareExecution(ctx, base)
if err != nil {
t.Fatal(err)
}
prefix := original.Details().Messages
initial, err := runtime.Engine.RunPrepared(ctx, original)
if err != nil {
t.Fatal(err)
}
suffix := []pk.RenderedMessage{{Role: pk.RoleAssistant, Content: initial.RawOutput}}
contract, err := OutputContract(responseschema.CorrectionRepairKey, 0)
if err != nil {
t.Fatal(err)
}
for _, finding := range []string{"first", "second"} {
instruction, err := runtime.Engine.Prepare(ctx, pk.RunRequest{PromptID: "repair-fixture", ProfileID: "repair-profile", Vars: map[string]string{"Finding": finding}})
if err != nil {
t.Fatal(err)
}
suffix = append(suffix, instruction.Messages...)
req := base
req.ProfileID = "repair-profile"
req.AppendedMessages = suffix
req.Validation = contract
handle, err := runtime.Engine.PrepareExecution(ctx, req)
if err != nil {
t.Fatal(err)
}
details := handle.Details()
if !reflect.DeepEqual(details.Messages[:len(prefix)], prefix) || !reflect.DeepEqual(details.Messages[len(prefix):], suffix) {
t.Fatalf("prepared conversation = %#v", details.Messages)
}
result, err := runtime.Engine.RunPrepared(ctx, handle)
if err != nil {
t.Fatal(err)
}
if !result.Validation.IsValid {
t.Fatalf("invalid result: %+v", result.Validation)
}
suffix = append(suffix, pk.RenderedMessage{Role: pk.RoleAssistant, Content: result.RawOutput})
}
if len(requests) != 3 || requests[0].Model != "proposal-model" || requests[1].Model != "repair-model" || requests[2].Model != "repair-model" {
t.Fatalf("models: %+v", requests)
}
for i, wantLen := range []int{2, 4, 6} {
req := requests[i]
if len(req.Messages) != wantLen {
t.Fatalf("request %d messages: %+v", i, req.Messages)
}
for j, message := range req.Messages {
if message.Role != string(append(prefix, suffix...)[j].Role) || message.Content != append(prefix, suffix...)[j].Content {
t.Fatalf("request %d message %d: %+v", i, j, message)
}
}
encoded, _ := json.Marshal(req.ResponseFormat)
expected := "corrections"
if i > 0 {
expected = "proposal_ref"
}
if !strings.Contains(string(encoded), expected) {
t.Fatalf("request %d contract: %s", i, encoded)
}
}
}
func TestRoleClientCapturesAndReplaysRepairConversation(t *testing.T) {
func TestRoleClientContinuesIndependentRepairConversations(t *testing.T) {
dir := t.TempDir()
for name, content := range map[string]string{
"config.yml": "version: 3\n",
@@ -207,17 +75,13 @@ func TestRoleClientCapturesAndReplaysRepairConversation(t *testing.T) {
t.Fatalf("capture: %+v %v", initial, err)
}
vars["SectionJSON"] = "changed"
replay := &structuredoutput.Replay{Exchange: initial.Exchange}
exchange := initial.Exchange
for _, finding := range []string{"first", "second"} {
result, err := repair.CompleteStructured(context.Background(), structuredoutput.Request{PromptID: prompts.PromptIDModuleGrammarRepair, Schema: responseschema.CorrectionRepairKey, Variables: map[string]string{"RepairRequestJSON": finding}, Replay: replay}, &map[string]any{})
result, err := repair.CompleteStructured(context.Background(), structuredoutput.Request{PromptID: prompts.PromptIDModuleGrammarRepair, Schema: responseschema.CorrectionRepairKey, Variables: map[string]string{"RepairRequestJSON": finding}, ContinueFrom: exchange}, &map[string]any{})
if err != nil {
t.Fatal(err)
}
if len(result.RepairInstruction) != 1 {
t.Fatalf("instruction: %+v", result)
}
replay.Turns = append(replay.Turns, result.RepairInstruction...)
replay.Turns = append(replay.Turns, structuredoutput.Message{Role: "assistant", Content: result.FinalContent})
exchange = result.Exchange
}
if !reflect.DeepEqual(models, []string{"proposal", "repair", "repair"}) {
t.Fatalf("models = %v", models)
@@ -235,21 +99,18 @@ func TestRoleClientCapturesAndReplaysRepairConversation(t *testing.T) {
!strings.Contains(string(first.Content), `"text":"first"`) || !reflect.DeepEqual(first, messages[2][len(prefix)+1]) {
t.Fatalf("cached instruction changed during replay: first=%s replay=%s", first.Content, messages[2][len(prefix)+1].Content)
}
_, err = repair.CompleteStructured(context.Background(), structuredoutput.Request{PromptID: prompts.PromptIDModuleGrammarRepair, Schema: responseschema.CorrectionRepairKey, Variables: map[string]string{"RepairRequestJSON": "independent"}, Replay: &structuredoutput.Replay{Exchange: initial.Exchange}}, &map[string]any{})
_, err = repair.CompleteStructured(context.Background(), structuredoutput.Request{PromptID: prompts.PromptIDModuleGrammarRepair, Schema: responseschema.CorrectionRepairKey, Variables: map[string]string{"RepairRequestJSON": "independent"}, ContinueFrom: initial.Exchange}, &map[string]any{})
if err != nil || len(messages) != 4 || len(messages[3]) != len(prefix)+2 || !strings.Contains(string(messages[3][len(prefix)+1].Content), "independent") {
t.Fatalf("independent branch: %v %+v", err, messages)
}
bad := *initial.Exchange
bad.Messages = append([]structuredoutput.Message(nil), bad.Messages...)
bad.Messages[0].Content = "changed"
_, err = repair.CompleteStructured(context.Background(), structuredoutput.Request{PromptID: prompts.PromptIDModuleGrammarRepair, Schema: responseschema.CorrectionRepairKey, Variables: map[string]string{"RepairRequestJSON": "failure"}, Replay: &structuredoutput.Replay{Exchange: &bad}}, &map[string]any{})
_, err = repair.CompleteStructured(context.Background(), structuredoutput.Request{PromptID: prompts.PromptIDModuleGrammarRepair, Schema: responseschema.CorrectionRepairKey, Variables: map[string]string{"RepairRequestJSON": "failure"}, ContinueFrom: &structuredoutput.Exchange{}}, &map[string]any{})
var failure *structuredoutput.Failure
if !errors.As(err, &failure) || failure.Class != structuredoutput.FailureFatal || len(messages) != 4 {
t.Fatalf("prefix failure: %v, calls %d", err, len(messages))
}
}
func TestRepairReplayUsesReplacementAndCountsStructuralRepair(t *testing.T) {
func TestRepairContinuationUsesReplacementAndCountsStructuralRepair(t *testing.T) {
dir := t.TempDir()
files := map[string]string{
"config.yml": "version: 3\n",
@@ -291,11 +152,11 @@ func TestRepairReplayUsesReplacementAndCountsStructuralRepair(t *testing.T) {
if err != nil {
t.Fatal(err)
}
result, err := client.CompleteStructured(context.Background(), structuredoutput.Request{PromptID: prompts.PromptIDModuleGrammarRepair, Schema: responseschema.CorrectionRepairKey, Variables: map[string]string{"RepairRequestJSON": "fixture"}, Replay: &structuredoutput.Replay{Exchange: base.Exchange}}, &map[string]any{})
result, err := client.CompleteStructured(context.Background(), structuredoutput.Request{PromptID: prompts.PromptIDModuleGrammarRepair, Schema: responseschema.CorrectionRepairKey, Variables: map[string]string{"RepairRequestJSON": "fixture"}, ContinueFrom: base.Exchange}, &map[string]any{})
if err != nil {
t.Fatal(err)
}
if len(requests) != 3 || result.Execution.ProviderAttempts != 2 || result.Execution.StructuralRepairs != 1 || result.FinalContent != outputs[2] || len(result.RepairInstruction) != 1 || result.RepairInstruction[0].Content != "Replace fixture" {
if len(requests) != 3 || result.Execution.ProviderAttempts != 2 || result.Execution.StructuralRepairs != 1 || result.FinalContent != outputs[2] {
t.Fatalf("result=%+v calls=%d", result, len(requests))
}
}

View File

@@ -0,0 +1,63 @@
package llm
import (
"context"
"encoding/json"
"gitea.maximumdirect.net/eric/audita/v3/internal/framework/structuredoutput"
pk "go.maximumdirect.net/promptkit"
)
var _ structuredoutput.Coordinator = (*RoleClient)(nil)
// Coordinate shares request construction and finalization with ordinary calls.
// PromptKit alone advances history and consumes the additional-turn budget.
func (c *RoleClient) Coordinate(ctx context.Context, start *structuredoutput.Exchange, budget int, next func(context.Context, structuredoutput.Turn) (*structuredoutput.Request, error)) (bool, error) {
var exchange *pk.Exchange
if start != nil {
exchange, _ = start.State.(*pk.Exchange)
}
var pending *pendingCall
result, err := c.runtime.Engine.Coordinate(ctx, pk.CoordinateRequest{
Start: exchange, AdditionalTurns: budget, Observer: c.hooks.OnEvent,
Next: func(ctx context.Context, turn pk.ConversationTurn) (*pk.RunRequest, error) {
var delivered structuredoutput.Turn
if turn.Result != nil {
response, failure := pending.finish(turn.Result, nil)
pending = nil
delivered = structuredoutput.Turn{Response: &response, Err: failure}
}
selected, err := next(ctx, delivered)
if err != nil || selected == nil {
return nil, err
}
// Exhaustion needs a requested turn, but must not allocate a logical call.
if turn.RemainingTurns == 0 {
return &pk.RunRequest{}, nil
}
var output json.RawMessage
pending, err = c.prepareCall(ctx, *selected, &output, nil)
if err != nil {
_, domainErr := next(ctx, structuredoutput.Turn{Response: &structuredoutput.Response{}, Err: err})
if domainErr != nil {
return nil, domainErr
}
return nil, err
}
return &pending.request, nil
},
})
if pending != nil {
// Execution errors and withheld responses do not enter Next. Finalize their
// diagnostic evidence and let the domain record the terminal failure once.
response, failure := pending.finish(result.Undelivered, err)
_, domainErr := next(ctx, structuredoutput.Turn{Response: &response, Err: failure})
if domainErr != nil {
return false, domainErr
}
if structuredoutput.IsRecoverable(failure) {
return false, nil
}
}
return result.Outcome == pk.CoordinationExhausted, err
}

View File

@@ -0,0 +1,141 @@
package llm
import (
"context"
"errors"
"testing"
"gitea.maximumdirect.net/eric/audita/v3/internal/framework/outcomes"
"gitea.maximumdirect.net/eric/audita/v3/internal/framework/responseschema"
"gitea.maximumdirect.net/eric/audita/v3/internal/framework/structuredoutput"
"gitea.maximumdirect.net/eric/audita/v3/internal/prompts"
pk "go.maximumdirect.net/promptkit"
)
func TestCoordinatorFinalTurnFailuresAndAccounting(t *testing.T) {
for _, tc := range []struct {
name string
reply reply
again bool
reason outcomes.ReasonCode
}{
{name: "last turn assessed", reply: reply{200, `{"proposal_ref":1,"action":"withdraw","correction":null}`}},
{name: "exhaustion", reply: reply{200, `{"proposal_ref":1,"action":"withdraw","correction":null}`}, again: true},
{name: "invalid structure", reply: reply{200, `{"wrong":true}`}, reason: outcomes.ReasonInvalidOutputExhausted},
{name: "provider error", reply: reply{400, ""}, reason: outcomes.ReasonProviderFailed},
} {
t.Run(tc.name, func(t *testing.T) {
runtime, count, _ := adapterFixture(t, []reply{{200, `{"corrections":[]}`}, tc.reply}, 0, 0, nil)
collector := NewObservationCollector()
client, err := runtime.NewRoleClient("repair", "base", runtime.Config.LLM, collector.Hooks())
if err != nil {
t.Fatal(err)
}
req := adapterRequest(responseschema.CorrectionSetKey)
req.CaptureExchange = true
base, err := client.CompleteStructured(context.Background(), req, &map[string]any{})
if err != nil {
t.Fatal(err)
}
deliveries := 0
exhausted, err := client.Coordinate(context.Background(), base.Exchange, 1, func(ctx context.Context, turn structuredoutput.Turn) (*structuredoutput.Request, error) {
if turn.Response != nil {
deliveries++
if tc.reason != "" {
var failure *structuredoutput.Failure
if !errors.As(turn.Err, &failure) || failure.Reason != tc.reason {
t.Fatalf("failure=%v", turn.Err)
}
return nil, nil
}
if turn.Err != nil || len(turn.Response.Content) == 0 {
t.Fatalf("turn=%+v", turn)
}
if !tc.again {
return nil, nil
}
}
return &structuredoutput.Request{PromptID: prompts.PromptIDModuleGrammarRepair, Schema: responseschema.CorrectionRepairKey, Variables: map[string]string{"RepairRequestJSON": "{}"}}, nil
})
if err != nil || exhausted != tc.again || deliveries != 1 || count.Load() != 2 {
t.Fatalf("exhausted=%v delivered=%d calls=%d err=%v", exhausted, deliveries, count.Load(), err)
}
snapshot := collector.Snapshot()
summary := Summarize(snapshot)
if len(snapshot.Calls) != 2 || summary.ProviderAttempts != 2 || !summary.TerminalCoverageComplete {
t.Fatalf("summary=%+v calls=%+v", summary, snapshot.Calls)
}
for _, call := range snapshot.Calls {
if call.Finished.IsZero() {
t.Fatal("unfinished logical call")
}
}
})
}
}
func TestCoordinatorCanceledDeliveryRetainsEvidence(t *testing.T) {
runtime, _, _ := adapterFixture(t, []reply{{200, `{"corrections":[]}`}, {200, `{"proposal_ref":1,"action":"withdraw","correction":null}`}}, 0, 0, nil)
collector := NewObservationCollector()
proposal, _ := runtime.NewRoleClient("proposal", "base", runtime.Config.LLM, collector.Hooks())
req := adapterRequest(responseschema.CorrectionSetKey)
req.CaptureExchange = true
base, err := proposal.CompleteStructured(context.Background(), req, &map[string]any{})
if err != nil {
t.Fatal(err)
}
ctx, cancel := context.WithCancel(context.Background())
defer cancel()
hooks := collector.Hooks()
hooks.OnEvent = func(e pk.ExecutionEvent) {
if e.Kind == pk.ExecutionFinished {
cancel()
}
}
client, _ := runtime.NewRoleClient("repair", "base", runtime.Config.LLM, hooks)
deliveries := 0
_, err = client.Coordinate(ctx, base.Exchange, 1, func(_ context.Context, turn structuredoutput.Turn) (*structuredoutput.Request, error) {
if turn.Response != nil {
deliveries++
if !errors.Is(turn.Err, context.Canceled) || turn.Response.Prepared == nil || turn.Response.FinalContent == "" || len(turn.Response.Content) != 0 {
t.Fatalf("canceled turn=%+v", turn)
}
return nil, turn.Err
}
return &structuredoutput.Request{PromptID: prompts.PromptIDModuleGrammarRepair, Schema: responseschema.CorrectionRepairKey, Variables: map[string]string{"RepairRequestJSON": "{}"}}, nil
})
if !errors.Is(err, context.Canceled) || deliveries != 1 {
t.Fatalf("deliveries=%d err=%v", deliveries, err)
}
if got := Summarize(collector.Snapshot()); got.ProviderAttempts != 2 || !got.TerminalCoverageComplete {
t.Fatalf("summary=%+v", got)
}
}
func TestCoordinatorReleasesAdmissionBeforeNestedValidation(t *testing.T) {
runtime, count, _ := adapterFixture(t, []reply{{200, `{"corrections":[]}`}, {200, `{"proposal_ref":1,"action":"withdraw","correction":null}`}, {200, `{"validations":[]}`}}, 0, 0, nil)
collector := NewObservationCollector()
proposal, _ := runtime.NewRoleClient("proposal", "base", runtime.Config.LLM, collector.Hooks())
repair, _ := runtime.NewRoleClient("repair", "base", runtime.Config.LLM, collector.Hooks())
validation, _ := runtime.NewRoleClient("validation", "base", runtime.Config.LLM, collector.Hooks())
req := adapterRequest(responseschema.CorrectionSetKey)
req.CaptureExchange = true
base, err := proposal.CompleteStructured(context.Background(), req, &map[string]any{})
if err != nil {
t.Fatal(err)
}
_, err = repair.Coordinate(context.Background(), base.Exchange, 1, func(ctx context.Context, turn structuredoutput.Turn) (*structuredoutput.Request, error) {
if turn.Response != nil {
if turn.Err != nil {
return nil, turn.Err
}
_, err := validation.CompleteStructured(ctx, adapterRequest(responseschema.ValidatorDecisionSetKey), &map[string]any{})
return nil, err
}
return &structuredoutput.Request{PromptID: prompts.PromptIDModuleGrammarRepair, Schema: responseschema.CorrectionRepairKey, Variables: map[string]string{"RepairRequestJSON": "{}"}}, nil
})
summary := Summarize(collector.Snapshot())
if err != nil || count.Load() != 3 || summary.ProviderAttempts != 3 || summary.LogicalCalls != 3 || !summary.TerminalCoverageComplete {
t.Fatalf("calls=%d summary=%+v err=%v", count.Load(), summary, err)
}
}

View File

@@ -115,7 +115,7 @@ func TestRoleClientExecutionBudgetsAndFailureClasses(t *testing.T) {
capturing := request
capturing.CaptureExchange = true
response, err := client.CompleteStructured(context.Background(), capturing, &result)
if tc.success && (response.Exchange == nil || response.Exchange.FinalContent != tc.outputs[len(tc.outputs)-1]) {
if tc.success && (response.Exchange == nil || response.FinalContent != tc.outputs[len(tc.outputs)-1]) {
t.Fatalf("captured final response: %+v", response.Exchange)
}
if (err == nil) != tc.success || calls != tc.wantCalls {

View File

@@ -64,21 +64,24 @@ type LogicalCall struct {
}
type ObservationSnapshot struct {
Calls []LogicalCall `json:"calls"`
Events []ExecutionObservation `json:"events"`
Group *pk.ExecutionGroupSummary `json:"-"`
Calls []LogicalCall `json:"calls"`
Events []ExecutionObservation `json:"events"`
}
type ExecutionSummary struct {
LogicalCalls uint64 `json:"logical_calls"`
ProviderAttempts uint64 `json:"provider_attempts"`
TransientRetries uint64 `json:"transient_retries"`
StructuralRepairs uint64 `json:"structural_repairs"`
Usage Usage `json:"usage"`
TerminalCoverageComplete bool `json:"terminal_coverage_complete"`
LogicalCalls uint64 `json:"logical_calls"`
ProviderAttempts uint64 `json:"provider_attempts"`
TransientRetries uint64 `json:"transient_retries"`
StructuralRepairs uint64 `json:"structural_repairs"`
Usage Usage `json:"usage"`
}
// ObservationCollector accepts short synchronous callbacks from parallel calls.
// Snapshot should be taken after the runner has joined active section work.
type ObservationCollector struct {
group pk.ExecutionGroup
mu sync.Mutex
calls map[uint64]LogicalCall
events []ExecutionObservation
@@ -137,6 +140,7 @@ func (c *ObservationCollector) recordEvent(callID uint64, context structuredoutp
if c == nil {
return
}
c.group.Observe(event)
count := func(value pk.ObservedTokenCount) TokenCount {
return TokenCount{Value: value.Value, Known: value.Known, Complete: value.Complete}
}
@@ -176,52 +180,28 @@ func (c *ObservationCollector) Snapshot() ObservationSnapshot {
}
return snapshot.Events[i].Time.Before(snapshot.Events[j].Time)
})
group := c.group.Snapshot()
snapshot.Group = &group
return snapshot
}
// Summarize counts terminal usage once per logical call. PromptKit's terminal
// usage already aggregates its attempts, retries, and repairs.
// Summarize delegates deduplication and terminal accounting to PromptKit. A
// filtered diagnostic view selects its own group; overlapping views are not additive.
func Summarize(snapshot ObservationSnapshot) ExecutionSummary {
result := ExecutionSummary{LogicalCalls: uint64(len(snapshot.Calls))}
terminal := make(map[uint64]ExecutionObservation)
for _, event := range snapshot.Events {
if event.Kind == string(pk.ExecutionFinished) {
terminal[event.CallID] = event
summary := snapshot.Group
if summary == nil {
var group pk.ExecutionGroup
count := func(c TokenCount) pk.ObservedTokenCount {
return pk.ObservedTokenCount{Value: c.Value, Known: c.Known, Complete: c.Complete}
}
}
dispatched := false
aggregate := func(dst *TokenCount, source TokenCount) {
if source.Known {
dst.Value += source.Value
dst.Known = true
for _, event := range snapshot.Events {
group.Observe(pk.ExecutionEvent{Kind: pk.ExecutionEventKind(event.Kind), RunID: event.RunID, Attempts: event.Attempts, TransientRetries: event.TransientRetries, RepairCalls: event.StructuralRepairs, Usage: pk.ObservedTokenUsage{PromptTokens: count(event.Usage.Prompt), CompletionTokens: count(event.Usage.Completion), TotalTokens: count(event.Usage.Total), CachedTokens: count(event.Usage.Cached), CacheWriteTokens: count(event.Usage.CacheWrite)}})
}
dst.Complete = dst.Complete && source.Complete
value := group.Snapshot()
summary = &value
}
result.Usage = Usage{Prompt: TokenCount{Complete: true}, Completion: TokenCount{Complete: true}, Total: TokenCount{Complete: true}, Cached: TokenCount{Complete: true}, CacheWrite: TokenCount{Complete: true}}
for _, call := range snapshot.Calls {
event, ok := terminal[call.CallID]
if !ok {
continue
}
result.ProviderAttempts += event.Attempts
result.TransientRetries += event.TransientRetries
result.StructuralRepairs += event.StructuralRepairs
if event.Attempts == 0 {
continue
}
dispatched = true
aggregate(&result.Usage.Prompt, event.Usage.Prompt)
aggregate(&result.Usage.Completion, event.Usage.Completion)
aggregate(&result.Usage.Total, event.Usage.Total)
aggregate(&result.Usage.Cached, event.Usage.Cached)
aggregate(&result.Usage.CacheWrite, event.Usage.CacheWrite)
count := func(c pk.ObservedTokenCount) TokenCount {
return TokenCount{Value: c.Value, Known: c.Known, Complete: c.Complete}
}
if !dispatched {
result.Usage.Prompt.Known = true
result.Usage.Completion.Known = true
result.Usage.Total.Known = true
result.Usage.Cached.Known = true
result.Usage.CacheWrite.Known = true
}
return result
return ExecutionSummary{LogicalCalls: uint64(len(snapshot.Calls)), TerminalCoverageComplete: summary.TerminalCoverageComplete, ProviderAttempts: summary.Attempts, TransientRetries: summary.TransientRetries, StructuralRepairs: summary.RepairCalls, Usage: Usage{Prompt: count(summary.Usage.PromptTokens), Completion: count(summary.Usage.CompletionTokens), Total: count(summary.Usage.TotalTokens), Cached: count(summary.Usage.CachedTokens), CacheWrite: count(summary.Usage.CacheWriteTokens)}}
}

View File

@@ -3,6 +3,7 @@ package llm
import (
"context"
"encoding/json"
"fmt"
"strings"
"sync"
"testing"
@@ -23,8 +24,8 @@ func TestObservationSummarySeparatesLogicalCallsAttemptsAndUsage(t *testing.T) {
}
call(1) // preparation failed before PromptKit emitted events
call(2) // capacity rejected without a dispatched attempt
hooks.OnEventWithCall(2, structuredoutput.DiagnosticContext{Role: "proposal"}, pk.ExecutionEvent{Kind: pk.ExecutionAttempt, Dispatched: false})
hooks.OnEventWithCall(2, structuredoutput.DiagnosticContext{Role: "proposal"}, pk.ExecutionEvent{Kind: pk.ExecutionFinished, Attempts: 0})
hooks.OnEventWithCall(2, structuredoutput.DiagnosticContext{Role: "proposal"}, pk.ExecutionEvent{RunID: "run-2", Kind: pk.ExecutionAttempt, Dispatched: false})
hooks.OnEventWithCall(2, structuredoutput.DiagnosticContext{Role: "proposal"}, pk.ExecutionEvent{RunID: "run-2", Kind: pk.ExecutionFinished, Attempts: 0, Usage: pk.ObservedTokenUsage{PromptTokens: pk.ObservedTokenCount{Known: true, Complete: true}, CompletionTokens: pk.ObservedTokenCount{Known: true, Complete: true}, TotalTokens: pk.ObservedTokenCount{Known: true, Complete: true}, CachedTokens: pk.ObservedTokenCount{Known: true, Complete: true}, CacheWriteTokens: pk.ObservedTokenCount{Known: true, Complete: true}}})
summary := Summarize(collector.Snapshot())
if summary.LogicalCalls != 2 || summary.ProviderAttempts != 0 || summary.Usage.Total != (TokenCount{Known: true, Complete: true}) {
t.Fatalf("zero-attempt accounting: %+v", summary)
@@ -32,16 +33,16 @@ func TestObservationSummarySeparatesLogicalCallsAttemptsAndUsage(t *testing.T) {
call(3)
context := structuredoutput.DiagnosticContext{Role: "validation", ModuleInstance: "grammar", ValidatorName: "editorial_review"}
hooks.OnEventWithCall(3, context, pk.ExecutionEvent{Kind: pk.ExecutionAttempt, Dispatched: true, Usage: pk.ObservedTokenUsage{TotalTokens: pk.ObservedTokenCount{Value: 100, Known: true, Complete: true}}})
hooks.OnEventWithCall(3, context, pk.ExecutionEvent{Kind: pk.ExecutionFinished, Attempts: 3, TransientRetries: 1, RepairCalls: 1, Usage: pk.ObservedTokenUsage{
hooks.OnEventWithCall(3, context, pk.ExecutionEvent{RunID: "run-3", Kind: pk.ExecutionAttempt, Dispatched: true, Usage: pk.ObservedTokenUsage{TotalTokens: pk.ObservedTokenCount{Value: 100, Known: true, Complete: true}}})
hooks.OnEventWithCall(3, context, pk.ExecutionEvent{RunID: "run-3", Kind: pk.ExecutionFinished, Attempts: 3, TransientRetries: 1, RepairCalls: 1, Usage: pk.ObservedTokenUsage{
PromptTokens: pk.ObservedTokenCount{Value: 5, Known: true, Complete: false},
TotalTokens: pk.ObservedTokenCount{Value: 8, Known: true, Complete: false},
}})
call(4)
hooks.OnEventWithCall(4, context, pk.ExecutionEvent{Kind: pk.ExecutionFinished, Attempts: 1, Usage: pk.ObservedTokenUsage{PromptTokens: pk.ObservedTokenCount{Known: false, Complete: false}, TotalTokens: pk.ObservedTokenCount{Known: false, Complete: false}}})
hooks.OnEventWithCall(4, context, pk.ExecutionEvent{RunID: "run-4", Kind: pk.ExecutionFinished, Attempts: 1, Usage: pk.ObservedTokenUsage{PromptTokens: pk.ObservedTokenCount{Known: false, Complete: false}, TotalTokens: pk.ObservedTokenCount{Known: false, Complete: false}}})
snapshot := collector.Snapshot()
summary = Summarize(snapshot)
if summary.LogicalCalls != 4 || summary.ProviderAttempts != 4 || summary.TransientRetries != 1 || summary.StructuralRepairs != 1 || summary.Usage.Total.Value != 8 || !summary.Usage.Total.Known || summary.Usage.Total.Complete || summary.Usage.Prompt.Value != 5 || !summary.Usage.Prompt.Known || summary.Usage.Prompt.Complete || summary.Usage.Cached.Known || summary.Usage.Cached.Complete {
if summary.LogicalCalls != 4 || summary.ProviderAttempts != 4 || summary.TransientRetries != 1 || summary.StructuralRepairs != 1 || summary.Usage.Total.Value != 8 || !summary.Usage.Total.Known || summary.Usage.Total.Complete || summary.Usage.Prompt.Value != 5 || !summary.Usage.Prompt.Known || summary.Usage.Prompt.Complete || !summary.Usage.Cached.Known || summary.Usage.Cached.Complete {
t.Fatalf("partial usage or double count: %+v", summary)
}
raw, err := json.Marshal(snapshot)
@@ -68,7 +69,7 @@ func TestObservationCollectorConcurrentAttributionAndSnapshotCopy(t *testing.T)
context := structuredoutput.DiagnosticContext{Role: "proposal", ModuleInstance: "grammar", SectionIndex: &index}
started := time.Now().UTC()
hooks.OnCall(structuredoutput.CallRecord{CallID: uint64(id), Context: context, Started: started})
hooks.OnEventWithCall(uint64(id), context, pk.ExecutionEvent{Kind: pk.ExecutionFinished, RunID: "run", Attempts: 1, Usage: pk.ObservedTokenUsage{TotalTokens: pk.ObservedTokenCount{Value: 1, Known: true, Complete: true}}})
hooks.OnEventWithCall(uint64(id), context, pk.ExecutionEvent{Kind: pk.ExecutionFinished, RunID: fmt.Sprint(id), Attempts: 1, Usage: pk.ObservedTokenUsage{TotalTokens: pk.ObservedTokenCount{Value: 1, Known: true, Complete: true}}})
hooks.OnCall(structuredoutput.CallRecord{CallID: uint64(id), Context: context, Started: started, Finished: time.Now().UTC()})
index = -1
}(i)
@@ -124,3 +125,18 @@ func TestRoleClientFeedsCollectorAcrossPreparationFailure(t *testing.T) {
t.Fatalf("logical failures: %+v", snapshot.Calls)
}
}
func TestObservationCoverageIsIndependentOfUsage(t *testing.T) {
collector := NewObservationCollector()
hooks := collector.Hooks()
hooks.OnEventWithCall(1, structuredoutput.DiagnosticContext{}, pk.ExecutionEvent{RunID: "pending", Kind: pk.ExecutionAttempt})
got := Summarize(collector.Snapshot())
if got.TerminalCoverageComplete || got.Usage.Total.Known || got.Usage.Total.Complete {
t.Fatalf("incomplete group=%+v", got)
}
hooks.OnEventWithCall(2, structuredoutput.DiagnosticContext{}, pk.ExecutionEvent{RunID: "finished", Kind: pk.ExecutionFinished, Attempts: 1, Usage: pk.ObservedTokenUsage{TotalTokens: pk.ObservedTokenCount{Value: 7, Known: true, Complete: true}}})
got = Summarize(collector.Snapshot())
if got.TerminalCoverageComplete || !got.Usage.Total.Known || !got.Usage.Total.Complete || got.Usage.Total.Value != 7 {
t.Fatalf("partial coverage=%+v", got)
}
}

View File

@@ -20,15 +20,16 @@ func mapLLMSummary(source llm.ExecutionSummary) reporting.LLMSummary {
}
type executionCall struct {
RequestedPromptID string `json:"requested_prompt_id,omitempty"`
CallID uint64 `json:"call_id"`
Context structuredoutput.DiagnosticContext `json:"context"`
Prompt structuredoutput.PromptMetadata `json:"prompt"`
Started time.Time `json:"started"`
Finished *time.Time `json:"finished,omitempty"`
ElapsedMS int64 `json:"elapsed_ms"`
Failure structuredoutput.FailureKind `json:"failure,omitempty"`
Summary reporting.LLMSummary `json:"summary"`
TerminalCoverageComplete bool `json:"terminal_coverage_complete"`
RequestedPromptID string `json:"requested_prompt_id,omitempty"`
CallID uint64 `json:"call_id"`
Context structuredoutput.DiagnosticContext `json:"context"`
Prompt structuredoutput.PromptMetadata `json:"prompt"`
Started time.Time `json:"started"`
Finished *time.Time `json:"finished,omitempty"`
ElapsedMS int64 `json:"elapsed_ms"`
Failure structuredoutput.FailureKind `json:"failure,omitempty"`
Summary reporting.LLMSummary `json:"summary"`
}
type executionEvent struct {
@@ -72,7 +73,7 @@ func BuildLLMExecution(snapshot llm.ObservationSnapshot) any {
value := call.Finished
finished = &value
}
calls = append(calls, executionCall{RequestedPromptID: call.RequestedPromptID, CallID: call.CallID, Context: call.Context, Prompt: call.Prompt, Started: call.Started, Finished: finished, ElapsedMS: call.WallDuration.Milliseconds(), Failure: call.Failure, Summary: mapLLMSummary(llm.Summarize(group))})
calls = append(calls, executionCall{TerminalCoverageComplete: llm.Summarize(group).TerminalCoverageComplete, RequestedPromptID: call.RequestedPromptID, CallID: call.CallID, Context: call.Context, Prompt: call.Prompt, Started: call.Started, Finished: finished, ElapsedMS: call.WallDuration.Milliseconds(), Failure: call.Failure, Summary: mapLLMSummary(llm.Summarize(group))})
}
for _, event := range snapshot.Events {
var generation, attempt *uint64
@@ -87,7 +88,8 @@ func BuildLLMExecution(snapshot llm.ObservationSnapshot) any {
events = append(events, executionEvent{CallID: event.CallID, Context: event.Context, Kind: event.Kind, RunID: event.RunID, Time: event.Time, BackendID: event.BackendID, ProfileID: event.ProfileID, Model: event.Model, Generation: generation, Attempt: attempt, Dispatched: event.Dispatched, CapacityWaitMS: event.CapacityWait.Milliseconds(), GenerationMS: event.GenerationDuration.Milliseconds(), RetryDelayMS: event.RetryDelay.Milliseconds(), DurationMS: event.Duration.Milliseconds(), Outcome: event.Outcome, Failure: event.Failure, StatusCode: event.StatusCode, RetriesExhausted: event.RetriesExhausted, Attempts: event.Attempts, TransientRetries: event.TransientRetries, StructuralRepairs: event.StructuralRepairs, Usage: mapLLMSummary(llm.ExecutionSummary{Usage: event.Usage}).Usage})
}
return struct {
Calls []executionCall `json:"calls"`
Events []executionEvent `json:"events"`
}{calls, events}
Summary llm.ExecutionSummary `json:"summary"`
Calls []executionCall `json:"calls"`
Events []executionEvent `json:"events"`
}{llm.Summarize(snapshot), calls, events}
}

View File

@@ -19,7 +19,7 @@ func TestExecutionReportingKeepsSafeMilliseconds(t *testing.T) {
{CallID: 1, Context: structuredoutput.DiagnosticContext{Role: "proposal", ModuleInstance: "grammar-1"}, Prompt: structuredoutput.PromptMetadata{BackendID: "primary", ProfileID: "draft", Model: "m"}, Started: start, Finished: start.Add(1500 * time.Millisecond), WallDuration: 1500 * time.Millisecond},
{CallID: 2, Context: structuredoutput.DiagnosticContext{Role: "validation"}, Prompt: structuredoutput.PromptMetadata{BackendID: "secondary"}, Started: start, Finished: start.Add(5 * time.Millisecond), WallDuration: 5 * time.Millisecond, Failure: structuredoutput.FailureRuntime},
},
Events: []llm.ExecutionObservation{{CallID: 1, Kind: string(pk.ExecutionFinished), BackendID: "primary", Attempts: 2, TransientRetries: 1, CapacityWait: 20 * time.Millisecond, GenerationDuration: 500 * time.Millisecond, RetryDelay: 100 * time.Millisecond}},
Events: []llm.ExecutionObservation{{CallID: 1, RunID: "run-1", Kind: string(pk.ExecutionFinished), BackendID: "primary", Attempts: 2, TransientRetries: 1, CapacityWait: 20 * time.Millisecond, GenerationDuration: 500 * time.Millisecond, RetryDelay: 100 * time.Millisecond}},
}
raw, err := json.Marshal(BuildLLMExecution(snapshot))
if err != nil {
@@ -46,7 +46,7 @@ func BenchmarkExecutionReports(b *testing.B) {
snapshot := llm.ObservationSnapshot{}
for i := 0; i < n; i++ {
snapshot.Calls = append(snapshot.Calls, llm.LogicalCall{CallID: uint64(i + 1), Prompt: structuredoutput.PromptMetadata{BackendID: "local"}, Context: structuredoutput.DiagnosticContext{ModuleInstance: "grammar", Role: "proposal"}})
snapshot.Events = append(snapshot.Events, llm.ExecutionObservation{CallID: uint64(i + 1), Kind: string(pk.ExecutionFinished), Attempts: 1})
snapshot.Events = append(snapshot.Events, llm.ExecutionObservation{RunID: fmt.Sprint(i), CallID: uint64(i + 1), Kind: string(pk.ExecutionFinished), Attempts: 1})
}
b.Run(fmt.Sprintf("execution/%d", n), func(b *testing.B) {
for i := 0; i < b.N; i++ {
@@ -64,9 +64,9 @@ func TestIndexedReportsPreserveOrderAndPartialUsage(t *testing.T) {
{CallID: 2, Context: structuredoutput.DiagnosticContext{Role: "validation", ModuleInstance: "grammar_1"}, Prompt: structuredoutput.PromptMetadata{BackendID: "a"}},
},
Events: []llm.ExecutionObservation{
{CallID: 1, Kind: string(pk.ExecutionFinished), Attempts: 1, Usage: llm.Usage{Total: llm.TokenCount{Value: 7, Known: true, Complete: true}}},
{CallID: 3, Kind: string(pk.ExecutionFinished), Attempts: 1},
{CallID: 2, Kind: string(pk.ExecutionFinished), Attempts: 1},
{CallID: 1, RunID: "run-1", Kind: string(pk.ExecutionFinished), Attempts: 1, Usage: llm.Usage{Total: llm.TokenCount{Value: 7, Known: true, Complete: true}}},
{CallID: 3, RunID: "run-3", Kind: string(pk.ExecutionFinished), Attempts: 1},
{CallID: 2, RunID: "run-2", Kind: string(pk.ExecutionFinished), Attempts: 1},
},
}
raw, err := json.Marshal(BuildLLMExecution(snapshot))

View File

@@ -18,7 +18,7 @@ import (
)
func TestExecuteModuleProposalBuildsMessagesFromSectionAndDescription(t *testing.T) {
exchange := &structuredoutput.Exchange{PromptID: "captured", FinalContent: "response"}
exchange := &structuredoutput.Exchange{}
client := &fakeStructuredClient{
completion: structuredoutput.Response{Exchange: exchange},
responses: []StructuredCorrectionSet{

View File

@@ -2,6 +2,7 @@ package runner
import (
"context"
"encoding/json"
"errors"
"fmt"
"path/filepath"
@@ -117,8 +118,11 @@ func repairCandidates(ctx context.Context, input validateSectionCandidatesInput,
return findings
}
findings := safeFindings(rejected)
replay := &structuredoutput.Replay{Exchange: input.Exchange}
for attempt := 0; attempt < input.Spec.Settings.LLM.CorrectionRepairAttempts; attempt++ {
coordinator, ok := input.Spec.RepairClient.(structuredoutput.Coordinator)
if !ok {
return finish(fmt.Errorf("correction repair coordinator is unavailable"))
}
buildRequest := func() (*structuredoutput.Request, error) {
request := proposal_generation.RepairRequest{
ProposalRef: position + 1,
Correction: proposal_generation.StructuredCorrectionProposal{SegmentReference: localRef, OriginalText: current.OriginalText, CorrectedText: current.CorrectedText},
@@ -134,12 +138,23 @@ func repairCandidates(ctx context.Context, input validateSectionCandidatesInput,
}
payload, err := request.JSON()
if err != nil {
branches = append(branches, branch)
return finish(err)
return nil, err
}
proposalIndex, attemptNumber := original.ProposalIndex, branch.AttemptCount+1
return &structuredoutput.Request{PromptID: promptID, Variables: map[string]string{"RepairRequestJSON": payload}, Schema: responseschema.CorrectionRepairKey, Context: structuredoutput.DiagnosticContext{StageIndex: &input.Spec.Index, StageID: input.Spec.ID, Role: "correction_repair", ModuleKey: input.Spec.ModuleKey, ModuleInstance: input.Spec.InstanceName, SectionIndex: input.SectionIndex, ProposalIndex: &proposalIndex, RepairAttempt: &attemptNumber}}, nil
}
exhausted, coordinationErr := coordinator.Coordinate(ctx, input.Exchange, input.Spec.Settings.LLM.CorrectionRepairAttempts, func(ctx context.Context, turn structuredoutput.Turn) (*structuredoutput.Request, error) {
if turn.Response == nil {
return buildRequest()
}
completion, callErr := *turn.Response, turn.Err
proposalIndex, attemptNumber := original.ProposalIndex, branch.AttemptCount+1
var response proposal_generation.RepairResponse
proposalIndex, attemptNumber := original.ProposalIndex, attempt+1
completion, callErr := input.Spec.RepairClient.CompleteStructured(ctx, structuredoutput.Request{PromptID: promptID, Variables: map[string]string{"RepairRequestJSON": payload}, Schema: responseschema.CorrectionRepairKey, Replay: replay, Context: structuredoutput.DiagnosticContext{StageIndex: &input.Spec.Index, StageID: input.Spec.ID, Role: "correction_repair", ModuleKey: input.Spec.ModuleKey, ModuleInstance: input.Spec.InstanceName, SectionIndex: input.SectionIndex, ProposalIndex: &proposalIndex, RepairAttempt: &attemptNumber}}, &response)
if callErr == nil {
_ = json.Unmarshal(completion.Content, &response)
}
operation := outcomes.Operation{
Reference: outcomes.Reference{
StageIndex: &input.Spec.Index, StageID: input.Spec.ID, ModuleKey: input.Spec.ModuleKey,
@@ -183,26 +198,23 @@ func repairCandidates(ctx context.Context, input validateSectionCandidatesInput,
finishAttempt(callErr, nil, validateSectionCandidatesResult{})
if structuredoutput.IsRecoverable(callErr) {
branch.Outcome = "unavailable"
break
return nil, nil
}
branches = append(branches, branch)
return finish(callErr)
return nil, callErr
}
decoded, err := proposal_generation.DecodeRepairResponse(completion.Content, position+1)
if err != nil {
operation.FailureReason = outcomes.ReasonInvalidOutput
finishAttempt(err, nil, validateSectionCandidatesResult{})
branch.Outcome = "unavailable"
break
return nil, nil
}
operation.Succeeded = true
operation.RecoveryReasons = outcomes.ExecutionRecoveries(operation.Counts)
replay.Turns = append(replay.Turns, completion.RepairInstruction...)
replay.Turns = append(replay.Turns, structuredoutput.Message{Role: "assistant", Content: completion.FinalContent})
if decoded.Action == "withdraw" {
finishAttempt(nil, nil, validateSectionCandidatesResult{})
branch.Outcome = "withdrawn"
break
return nil, nil
}
revision := decoded.Correction
localRef = revision.SegmentReference
@@ -225,7 +237,7 @@ func repairCandidates(ctx context.Context, input validateSectionCandidatesInput,
operation.RecoveryReasons = append(operation.RecoveryReasons, outcomes.RecoveryProposalsFiltered)
}
for i := range validated.operations {
proposalIndex, attemptNumber := original.ProposalIndex, attempt+1
proposalIndex, attemptNumber := original.ProposalIndex, attemptNumber
validated.operations[i].Reference.ProposalIndex = &proposalIndex
validated.operations[i].Reference.RepairAttempt = &attemptNumber
}
@@ -236,32 +248,35 @@ func repairCandidates(ctx context.Context, input validateSectionCandidatesInput,
finishAttempt(validationErr, &current, validated)
initial.operations = append(initial.operations, validated.operations...)
if validationErr != nil {
branches = append(branches, branch)
return finish(validationErr)
return nil, validationErr
}
if approved {
branch.Outcome = "approved"
initial.approved = append(initial.approved, current)
superseded[original.ProposalIndex] = true
break
return nil, nil
}
if actionable(original.ProposalIndex, validated.rejected) {
if attempt+1 == input.Spec.Settings.LLM.CorrectionRepairAttempts {
branch.Outcome = "exhausted"
break
}
findings = safeFindings(validated.rejected)
continue
return buildRequest()
}
branch.Outcome = "unavailable"
for _, d := range validated.rejected {
if d.Disposition != validators.DispositionUnavailable {
branch.Outcome = "rejected"
break
return nil, nil
}
}
break
return nil, nil
})
if exhausted {
branch.Outcome = "exhausted"
}
if coordinationErr != nil {
branches = append(branches, branch)
return finish(coordinationErr)
}
branches = append(branches, branch)
}
return finish(nil)

View File

@@ -4,10 +4,17 @@ import (
"context"
"encoding/json"
"errors"
"gitea.maximumdirect.net/eric/audita/v3/internal/framework/llm"
"gitea.maximumdirect.net/eric/audita/v3/internal/framework/responseschema"
"gitea.maximumdirect.net/eric/audita/v3/internal/prompts"
ck "go.maximumdirect.net/configkit"
"io"
"net/http"
"os"
"path/filepath"
"strings"
"testing"
"time"
"gitea.maximumdirect.net/eric/audita/v3/internal/core/config"
"gitea.maximumdirect.net/eric/audita/v3/internal/core/schema"
@@ -20,6 +27,7 @@ import (
)
type repairClient struct {
t *testing.T
outputs []string
calls []structuredoutput.Request
err error
@@ -27,11 +35,7 @@ type repairClient struct {
}
func (c *repairClient) CompleteStructured(_ context.Context, req structuredoutput.Request, out any) (structuredoutput.Response, error) {
if req.Replay != nil {
copyReplay := *req.Replay
copyReplay.Turns = append([]structuredoutput.Message(nil), req.Replay.Turns...)
req.Replay = &copyReplay
}
c.calls = append(c.calls, req)
if c.err != nil {
return structuredoutput.Response{}, c.err
@@ -40,18 +44,18 @@ func (c *repairClient) CompleteStructured(_ context.Context, req structuredoutpu
if err := json.Unmarshal([]byte(value), out); err != nil {
return structuredoutput.Response{}, err
}
return structuredoutput.Response{Prepared: &structuredoutput.PreparedSnapshot{Messages: []structuredoutput.Message{{Role: "user", Content: "synthetic-secret"}}}, Content: json.RawMessage(value), FinalContent: value, RepairInstruction: []structuredoutput.Message{{Role: "user", Content: "instruction"}}}, nil
return structuredoutput.Response{Prepared: &structuredoutput.PreparedSnapshot{Messages: []structuredoutput.Message{{Role: "user", Content: "synthetic-secret"}}}, Content: json.RawMessage(value), FinalContent: value, Exchange: &structuredoutput.Exchange{}}, nil
}
func (c *repairClient) ConfiguredSecrets(context.Context) []string { return c.secrets }
func TestRepairBranchesRetargetAndPreserveOtherApprovals(t *testing.T) {
client := &repairClient{outputs: []string{`{"proposal_ref":2,"action":"revise","correction":{"id":2,"original_text":"uh ","corrected_text":""}}`}}
client := &repairClient{t: t, outputs: []string{`{"proposal_ref":2,"action":"revise","correction":{"id":2,"original_text":"uh ","corrected_text":""}}`}}
candidates := []proposals.EnrichedCorrectionProposal{
{CorrectionProposal: proposals.CorrectionProposal{TargetSegmentID: 1, OriginalText: "cat", CorrectedText: "dog"}, ProposalMetadata: proposals.ProposalMetadata{ProposalIndex: 4}},
{CorrectionProposal: proposals.CorrectionProposal{TargetSegmentID: 0, OriginalText: "x", CorrectedText: "y"}, ProposalMetadata: proposals.ProposalMetadata{ProposalIndex: 5}},
}
input := validateSectionCandidatesInput{Spec: contracts.ModuleRunSpec{ModuleKey: "grammar", RepairClient: client, Settings: config.Stage{LLM: config.LLM{CorrectionRepairAttempts: 1}}}, Policy: proposals.ReplacementPolicyRequireUnique, WorkingTranscript: &schema.Transcript{Segments: []schema.Segment{{ID: 1, Text: "cat"}, {ID: 2, Text: "uh hello"}}}, Validators: []contracts.Validator{validators.ProposalShapeValidator{}, validators.NoEffectValidator{}, validators.OriginalTextPresenceValidator{}, validators.NonEmptyResultingSegmentValidator{}}, SectionEnriched: candidates, Exchange: &structuredoutput.Exchange{PromptID: "proposal", FinalContent: "original response"}}
input := validateSectionCandidatesInput{Spec: contracts.ModuleRunSpec{ModuleKey: "grammar", RepairClient: client, Settings: config.Stage{LLM: config.LLM{CorrectionRepairAttempts: 1}}}, Policy: proposals.ReplacementPolicyRequireUnique, WorkingTranscript: &schema.Transcript{Segments: []schema.Segment{{ID: 1, Text: "cat"}, {ID: 2, Text: "uh hello"}}}, Validators: []contracts.Validator{validators.ProposalShapeValidator{}, validators.NoEffectValidator{}, validators.OriginalTextPresenceValidator{}, validators.NonEmptyResultingSegmentValidator{}}, SectionEnriched: candidates, Exchange: &structuredoutput.Exchange{}}
initial, err := validateSectionCandidates(context.Background(), input)
if err != nil {
t.Fatal(err)
@@ -85,7 +89,7 @@ func TestRepairRequestCarriesLiteralPreview(t *testing.T) {
{"blank result", "uh", "uh", "", "", proposals.ReplacementPolicyRequireUnique, false},
} {
t.Run(tc.name, func(t *testing.T) {
client := &repairClient{outputs: []string{`{"proposal_ref":1,"action":"withdraw","correction":null}`}}
client := &repairClient{t: t, outputs: []string{`{"proposal_ref":1,"action":"withdraw","correction":null}`}}
candidate := proposals.EnrichedCorrectionProposal{CorrectionProposal: proposals.CorrectionProposal{TargetSegmentID: 81, OriginalText: tc.original, CorrectedText: tc.replacement}}
input := validateSectionCandidatesInput{
Spec: contracts.ModuleRunSpec{RepairClient: client, Settings: config.Stage{LLM: config.LLM{CorrectionRepairAttempts: 1}}},
@@ -93,7 +97,7 @@ func TestRepairRequestCarriesLiteralPreview(t *testing.T) {
WorkingTranscript: &schema.Transcript{Segments: []schema.Segment{{ID: 81, Text: tc.segment}}},
Validators: []contracts.Validator{validators.ProposalShapeValidator{}, validators.NoEffectValidator{}, validators.OriginalTextPresenceValidator{}, validators.NonEmptyResultingSegmentValidator{}, rejectUntilRepair()},
SectionEnriched: []proposals.EnrichedCorrectionProposal{candidate},
Exchange: &structuredoutput.Exchange{PromptID: "proposal"},
Exchange: &structuredoutput.Exchange{},
}
initial, err := validateSectionCandidates(context.Background(), input)
if err != nil {
@@ -125,7 +129,7 @@ func TestRepairRequestCarriesLiteralPreview(t *testing.T) {
}
func TestRepairRequestRefreshesPreviewAfterRetarget(t *testing.T) {
client := &repairClient{outputs: []string{
client := &repairClient{t: t, outputs: []string{
`{"proposal_ref":1,"action":"revise","correction":{"id":2,"original_text":"uh ","corrected_text":""}}`,
`{"proposal_ref":1,"action":"withdraw","correction":null}`,
}}
@@ -138,7 +142,7 @@ func TestRepairRequestRefreshesPreviewAfterRetarget(t *testing.T) {
WorkingTranscript: &schema.Transcript{Segments: []schema.Segment{{ID: 81, Text: "cat"}, {ID: 104, Text: "uh hello"}}},
Validators: []contracts.Validator{validators.ProposalShapeValidator{}, validators.NoEffectValidator{}, validators.OriginalTextPresenceValidator{}, reject},
SectionEnriched: []proposals.EnrichedCorrectionProposal{{CorrectionProposal: proposals.CorrectionProposal{TargetSegmentID: 81, OriginalText: "cat", CorrectedText: "dog"}}},
Exchange: &structuredoutput.Exchange{PromptID: "proposal"},
Exchange: &structuredoutput.Exchange{},
}
initial, err := validateSectionCandidates(context.Background(), input)
if err != nil {
@@ -175,9 +179,9 @@ func TestRepairBranchTerminalOutcomes(t *testing.T) {
{"provider failure", nil, &structuredoutput.Failure{Class: structuredoutput.FailureRecoverable, Reason: outcomes.ReasonProviderFailed, Message: "provider failed"}, 1, "unavailable", 1},
} {
t.Run(tc.name, func(t *testing.T) {
client := &repairClient{outputs: tc.outputs, err: tc.err}
client := &repairClient{t: t, outputs: tc.outputs, err: tc.err}
candidate := proposals.EnrichedCorrectionProposal{CorrectionProposal: proposals.CorrectionProposal{TargetSegmentID: 1, OriginalText: "cat", CorrectedText: "dog"}, ProposalMetadata: proposals.ProposalMetadata{ProposalIndex: 0}}
input := validateSectionCandidatesInput{Spec: contracts.ModuleRunSpec{RepairClient: client, Settings: config.Stage{LLM: config.LLM{CorrectionRepairAttempts: tc.budget}}}, Policy: proposals.ReplacementPolicyRequireUnique, WorkingTranscript: &schema.Transcript{Segments: []schema.Segment{{ID: 1, Text: "cat"}}}, Validators: []contracts.Validator{validators.ProposalShapeValidator{}, validators.NoEffectValidator{}, validators.OriginalTextPresenceValidator{}, rejectUntilRepair()}, SectionEnriched: []proposals.EnrichedCorrectionProposal{candidate}, Exchange: &structuredoutput.Exchange{PromptID: "proposal"}}
input := validateSectionCandidatesInput{Spec: contracts.ModuleRunSpec{RepairClient: client, Settings: config.Stage{LLM: config.LLM{CorrectionRepairAttempts: tc.budget}}}, Policy: proposals.ReplacementPolicyRequireUnique, WorkingTranscript: &schema.Transcript{Segments: []schema.Segment{{ID: 1, Text: "cat"}}}, Validators: []contracts.Validator{validators.ProposalShapeValidator{}, validators.NoEffectValidator{}, validators.OriginalTextPresenceValidator{}, rejectUntilRepair()}, SectionEnriched: []proposals.EnrichedCorrectionProposal{candidate}, Exchange: &structuredoutput.Exchange{}}
initial, err := validateSectionCandidates(context.Background(), input)
if err != nil {
t.Fatal(err)
@@ -189,8 +193,8 @@ func TestRepairBranchTerminalOutcomes(t *testing.T) {
if len(branches) != 1 || branches[0].Outcome != tc.outcome || len(client.calls) != tc.calls || branches[0].AttemptCount != tc.calls || len(branches[0].Attempts) != 0 {
t.Fatalf("branches=%+v calls=%d", branches, len(client.calls))
}
if tc.calls == 2 && len(client.calls[1].Replay.Turns) != 2 {
t.Fatalf("history: %+v", client.calls[1].Replay.Turns)
if tc.calls == 2 && client.calls[1].Context.RepairAttempt == nil {
t.Fatalf("history: %+v", client.calls[1].ContinueFrom)
}
})
}
@@ -202,11 +206,11 @@ func (m repairModule) RepairPromptID() string { return "repair" }
func TestRunnerAppliesRevisionsInOriginalOrder(t *testing.T) {
candidates := []proposals.CorrectionProposal{{TargetSegmentID: 1, OriginalText: "cat", CorrectedText: "dog"}, {TargetSegmentID: 1, OriginalText: "cat", CorrectedText: "fox"}, {TargetSegmentID: 1, OriginalText: "bird", CorrectedText: "owl"}}
client := &repairClient{outputs: []string{`{"proposal_ref":2,"action":"revise","correction":{"id":1,"original_text":"cat","corrected_text":"fox"}}`,
client := &repairClient{t: t, outputs: []string{`{"proposal_ref":2,"action":"revise","correction":{"id":1,"original_text":"cat","corrected_text":"fox"}}`,
`{"proposal_ref":3,"action":"revise","correction":{"id":1,"original_text":"bird","corrected_text":"owl"}}`,
}}
module := repairModule{outcomeModule{fakeModule: fakeModule{key: "grammar", policy: proposals.ReplacementPolicyRequireUnique, validators: []contracts.Validator{validators.ProposalShapeValidator{}, validators.NoEffectValidator{}, validators.OriginalTextPresenceValidator{}, rejectUntilRepair(0)}}, proposeResult: func(req contracts.ProposalRequest) (contracts.ProposalResult, error) {
return contracts.ProposalResult{Proposals: candidates, SegmentReferences: []int{1, 1, 1}, Exchange: &structuredoutput.Exchange{PromptID: "proposal"}}, nil
return contracts.ProposalResult{Proposals: candidates, SegmentReferences: []int{1, 1, 1}, Exchange: &structuredoutput.Exchange{}}, nil
}}}
runner := New(fakeFactory{modules: map[string]contracts.TranscriptModule{"grammar": module}})
out, err := runner.Run(context.Background(), RunInput{Transcript: &schema.Transcript{Segments: []schema.Segment{{ID: 1, Text: "cat and bird"}}}, ModuleSpecs: []contracts.ModuleRunSpec{{ModuleKey: "grammar", InstanceName: "grammar", RepairClient: client, Settings: config.Stage{LLM: config.LLM{CorrectionRepairAttempts: 1}}}}})
@@ -246,11 +250,11 @@ func TestRepairRevalidatesEarlierApprovalAndRetainsUnrelatedApprovalOnFailure(t
}
return validators.Result{Decisions: decisions}, nil
}}
client := &repairClient{outputs: []string{`{"proposal_ref":2,"action":"revise","correction":{"id":1,"original_text":"bird","corrected_text":"owl"}}`}}
client := &repairClient{t: t, outputs: []string{`{"proposal_ref":2,"action":"revise","correction":{"id":1,"original_text":"bird","corrected_text":"owl"}}`}}
if tc.failure {
client.err = &structuredoutput.Failure{Class: structuredoutput.FailureRecoverable, Reason: outcomes.ReasonProviderFailed, Message: "failed"}
}
input := validateSectionCandidatesInput{Spec: contracts.ModuleRunSpec{RepairClient: client, Settings: config.Stage{LLM: config.LLM{CorrectionRepairAttempts: 1}}}, Policy: proposals.ReplacementPolicyRequireUnique, WorkingTranscript: &schema.Transcript{Segments: []schema.Segment{{ID: 1, Text: "cat bird"}}}, Validators: []contracts.Validator{check, rejectUntilRepair(0)}, SectionEnriched: candidates, Exchange: &structuredoutput.Exchange{PromptID: "proposal"}}
input := validateSectionCandidatesInput{Spec: contracts.ModuleRunSpec{RepairClient: client, Settings: config.Stage{LLM: config.LLM{CorrectionRepairAttempts: 1}}}, Policy: proposals.ReplacementPolicyRequireUnique, WorkingTranscript: &schema.Transcript{Segments: []schema.Segment{{ID: 1, Text: "cat bird"}}}, Validators: []contracts.Validator{check, rejectUntilRepair(0)}, SectionEnriched: candidates, Exchange: &structuredoutput.Exchange{}}
initial, err := validateSectionCandidates(context.Background(), input)
if err != nil {
t.Fatal(err)
@@ -271,9 +275,9 @@ func TestRepairRevalidatesEarlierApprovalAndRetainsUnrelatedApprovalOnFailure(t
}
func TestRepairInterruptionRetainsAttempt(t *testing.T) {
client := &repairClient{err: &structuredoutput.Failure{Class: structuredoutput.FailureCanceled, Reason: outcomes.ReasonCanceled, Message: "canceled", Cause: context.Canceled}}
client := &repairClient{t: t, err: &structuredoutput.Failure{Class: structuredoutput.FailureCanceled, Reason: outcomes.ReasonCanceled, Message: "canceled", Cause: context.Canceled}}
candidate := proposals.EnrichedCorrectionProposal{CorrectionProposal: proposals.CorrectionProposal{TargetSegmentID: 1, OriginalText: "cat", CorrectedText: "dog"}, ProposalMetadata: proposals.ProposalMetadata{ProposalIndex: 0}}
input := validateSectionCandidatesInput{Spec: contracts.ModuleRunSpec{RepairClient: client, Settings: config.Stage{LLM: config.LLM{CorrectionRepairAttempts: 1}}}, Policy: proposals.ReplacementPolicyRequireUnique, WorkingTranscript: &schema.Transcript{Segments: []schema.Segment{{ID: 1, Text: "cat"}}}, Validators: []contracts.Validator{rejectUntilRepair()}, SectionEnriched: []proposals.EnrichedCorrectionProposal{candidate}, Exchange: &structuredoutput.Exchange{PromptID: "proposal"}}
input := validateSectionCandidatesInput{Spec: contracts.ModuleRunSpec{RepairClient: client, Settings: config.Stage{LLM: config.LLM{CorrectionRepairAttempts: 1}}}, Policy: proposals.ReplacementPolicyRequireUnique, WorkingTranscript: &schema.Transcript{Segments: []schema.Segment{{ID: 1, Text: "cat"}}}, Validators: []contracts.Validator{rejectUntilRepair()}, SectionEnriched: []proposals.EnrichedCorrectionProposal{candidate}, Exchange: &structuredoutput.Exchange{}}
initial, err := validateSectionCandidates(context.Background(), input)
if err != nil {
t.Fatal(err)
@@ -285,12 +289,12 @@ func TestRepairInterruptionRetainsAttempt(t *testing.T) {
}
func TestRepairArtifactsAreUniqueAndRedacted(t *testing.T) {
client := &repairClient{secrets: []string{"synthetic-secret"}, outputs: []string{`{"proposal_ref":1,"action":"withdraw","correction":null}`, `{"proposal_ref":2,"action":"withdraw","correction":null}`}}
client := &repairClient{t: t, secrets: []string{"synthetic-secret"}, outputs: []string{`{"proposal_ref":1,"action":"withdraw","correction":null}`, `{"proposal_ref":2,"action":"withdraw","correction":null}`}}
candidates := []proposals.EnrichedCorrectionProposal{
{CorrectionProposal: proposals.CorrectionProposal{TargetSegmentID: 1, OriginalText: "cat", CorrectedText: "dog"}, ProposalMetadata: proposals.ProposalMetadata{ProposalIndex: 0}},
{CorrectionProposal: proposals.CorrectionProposal{TargetSegmentID: 1, OriginalText: "cat", CorrectedText: "fox"}, ProposalMetadata: proposals.ProposalMetadata{ProposalIndex: 1}},
}
input := validateSectionCandidatesInput{Spec: contracts.ModuleRunSpec{InstanceName: "grammar_1", RepairClient: client, Settings: config.Stage{LLM: config.LLM{CorrectionRepairAttempts: 1}}}, Policy: proposals.ReplacementPolicyRequireUnique, WorkingTranscript: &schema.Transcript{Segments: []schema.Segment{{ID: 1, Text: "cat"}}}, Validators: []contracts.Validator{rejectUntilRepair()}, SectionEnriched: candidates, Exchange: &structuredoutput.Exchange{PromptID: "proposal"}, DiagnosticsDir: t.TempDir()}
input := validateSectionCandidatesInput{Spec: contracts.ModuleRunSpec{InstanceName: "grammar_1", RepairClient: client, Settings: config.Stage{LLM: config.LLM{CorrectionRepairAttempts: 1}}}, Policy: proposals.ReplacementPolicyRequireUnique, WorkingTranscript: &schema.Transcript{Segments: []schema.Segment{{ID: 1, Text: "cat"}}}, Validators: []contracts.Validator{rejectUntilRepair()}, SectionEnriched: candidates, Exchange: &structuredoutput.Exchange{}, DiagnosticsDir: t.TempDir()}
initial, err := validateSectionCandidates(context.Background(), input)
if err != nil {
t.Fatal(err)
@@ -318,14 +322,14 @@ func TestRepairArtifactsAreUniqueAndRedacted(t *testing.T) {
}
func TestRepairArtifactFailureDoesNotChangeOutcome(t *testing.T) {
client := &repairClient{outputs: []string{`{"proposal_ref":1,"action":"withdraw","correction":null}`}}
client := &repairClient{t: t, outputs: []string{`{"proposal_ref":1,"action":"withdraw","correction":null}`}}
path := filepath.Join(t.TempDir(), "file")
if err := os.WriteFile(path, []byte("x"), 0600); err != nil {
t.Fatal(err)
}
warnings := 0
candidate := proposals.EnrichedCorrectionProposal{CorrectionProposal: proposals.CorrectionProposal{TargetSegmentID: 1, OriginalText: "cat", CorrectedText: "dog"}, ProposalMetadata: proposals.ProposalMetadata{ProposalIndex: 0}}
input := validateSectionCandidatesInput{Spec: contracts.ModuleRunSpec{InstanceName: "grammar", RepairClient: client, Settings: config.Stage{LLM: config.LLM{CorrectionRepairAttempts: 1}}}, Policy: proposals.ReplacementPolicyRequireUnique, WorkingTranscript: &schema.Transcript{Segments: []schema.Segment{{ID: 1, Text: "cat"}}}, Validators: []contracts.Validator{rejectUntilRepair()}, SectionEnriched: []proposals.EnrichedCorrectionProposal{candidate}, Exchange: &structuredoutput.Exchange{PromptID: "proposal"}, DiagnosticsDir: path, ReportDiagnosticError: func(error) { warnings++ }}
input := validateSectionCandidatesInput{Spec: contracts.ModuleRunSpec{InstanceName: "grammar", RepairClient: client, Settings: config.Stage{LLM: config.LLM{CorrectionRepairAttempts: 1}}}, Policy: proposals.ReplacementPolicyRequireUnique, WorkingTranscript: &schema.Transcript{Segments: []schema.Segment{{ID: 1, Text: "cat"}}}, Validators: []contracts.Validator{rejectUntilRepair()}, SectionEnriched: []proposals.EnrichedCorrectionProposal{candidate}, Exchange: &structuredoutput.Exchange{}, DiagnosticsDir: path, ReportDiagnosticError: func(error) { warnings++ }}
initial, err := validateSectionCandidates(context.Background(), input)
if err != nil {
t.Fatal(err)
@@ -337,7 +341,7 @@ func TestRepairArtifactFailureDoesNotChangeOutcome(t *testing.T) {
}
func TestNoEffectAndUnavailableOnlyDoNotStartRepair(t *testing.T) {
client := &repairClient{}
client := &repairClient{t: t}
for _, rejected := range [][]ValidatorRejectedChange{
{{ProposalIndex: 0, Disposition: validators.DispositionRejected, ReasonCode: validators.ReasonNoEffect}},
{{ProposalIndex: 0, Disposition: validators.DispositionUnavailable, ReasonCode: "provider_failed"}},
@@ -352,10 +356,10 @@ func TestNoEffectAndUnavailableOnlyDoNotStartRepair(t *testing.T) {
}
func TestRepairFeedbackRedactsValidationCredentials(t *testing.T) {
client := &repairClient{outputs: []string{`{"proposal_ref":1,"action":"withdraw","correction":null}`}}
validation := &repairClient{secrets: []string{"validation-secret"}}
client := &repairClient{t: t, outputs: []string{`{"proposal_ref":1,"action":"withdraw","correction":null}`}}
validation := &repairClient{t: t, secrets: []string{"validation-secret"}}
candidate := proposals.EnrichedCorrectionProposal{CorrectionProposal: proposals.CorrectionProposal{TargetSegmentID: 1, OriginalText: "cat", CorrectedText: "dog"}, ProposalMetadata: proposals.ProposalMetadata{ProposalIndex: 0}}
input := validateSectionCandidatesInput{Spec: contracts.ModuleRunSpec{InstanceName: "grammar", RepairClient: client, Settings: config.Stage{LLM: config.LLM{CorrectionRepairAttempts: 1}}}, ValidationLLMClient: validation, SectionEnriched: []proposals.EnrichedCorrectionProposal{candidate}, Exchange: &structuredoutput.Exchange{PromptID: "proposal"}, DiagnosticsDir: t.TempDir()}
input := validateSectionCandidatesInput{Spec: contracts.ModuleRunSpec{InstanceName: "grammar", RepairClient: client, Settings: config.Stage{LLM: config.LLM{CorrectionRepairAttempts: 1}}}, ValidationLLMClient: validation, SectionEnriched: []proposals.EnrichedCorrectionProposal{candidate}, Exchange: &structuredoutput.Exchange{}, DiagnosticsDir: t.TempDir()}
initial := validateSectionCandidatesResult{rejected: []ValidatorRejectedChange{{ProposalIndex: 0, ValidatorName: "editorial_review", Disposition: validators.DispositionRejected, ReasonCode: "rejected", Message: "model echoed validation-secret"}}}
_, branches, err := repairCandidates(context.Background(), input, []int{1}, initial, "repair")
if err != nil {
@@ -393,3 +397,66 @@ func rejectUntilRepair(preserved ...int) contracts.Validator {
return validators.Result{Decisions: decisions}, nil
}}
}
type repairTransport func(*http.Request) (*http.Response, error)
func (f repairTransport) RoundTrip(r *http.Request) (*http.Response, error) { return f(r) }
// Exercise the real adapter and upstream coordinator; only model responses are scripted.
func (c *repairClient) Coordinate(ctx context.Context, _ *structuredoutput.Exchange, budget int, next func(context.Context, structuredoutput.Turn) (*structuredoutput.Request, error)) (bool, error) {
t := c.t
dir := t.TempDir()
for name, content := range map[string]string{
"config.yml": "version: 3\n",
"prompts.yml": "profiles:\n - id: fixture\n endpoint: https://example.invalid/v1\n model: fixture\nprompts:\n - id: repair\n messages:\n - role: user\n content: 'synthetic-secret {{.RepairRequestJSON}}'\n output:\n format: text\n validation_mode: none\n",
} {
if err := os.WriteFile(filepath.Join(dir, name), []byte(content), 0600); err != nil {
t.Fatal(err)
}
}
loaded, err := config.Load(ck.Input{Source: ck.File(filepath.Join(dir, "config.yml"))}, "")
if err != nil {
t.Fatal(err)
}
var selected *structuredoutput.Request
initial := true
runtime, err := llm.NewRuntime(ctx, loaded, &http.Client{Transport: repairTransport(func(*http.Request) (*http.Response, error) {
content := `{"corrections":[]}`
if initial {
initial = false
} else {
c.calls = append(c.calls, *selected)
content = c.outputs[len(c.calls)-1]
}
body, _ := json.Marshal(map[string]any{"choices": []any{map[string]any{"message": map[string]any{"role": "assistant", "content": content}, "finish_reason": "stop"}}})
return &http.Response{StatusCode: 200, Header: make(http.Header), Body: io.NopCloser(strings.NewReader(string(body)))}, nil
})})
if err != nil {
t.Fatal(err)
}
client, err := runtime.NewRoleClient("repair", "fixture", config.LLM{OperationTimeout: time.Second}, llm.CallHooks{})
if err != nil {
t.Fatal(err)
}
var output map[string]any
base, err := client.CompleteStructured(ctx, structuredoutput.Request{PromptID: prompts.PromptIDModuleGrammarProposal, Schema: responseschema.CorrectionSetKey, CaptureExchange: true, Variables: map[string]string{"TranscriptDescriptionBlock": "", "GlossaryJSON": "[]", "SectionJSON": "[]"}}, &output)
if err != nil {
return false, err
}
return client.Coordinate(ctx, base.Exchange, budget, func(ctx context.Context, turn structuredoutput.Turn) (*structuredoutput.Request, error) {
req, err := next(ctx, turn)
if req != nil && err == nil {
selected = req
if c.err != nil {
c.calls = append(c.calls, *req)
_, err = next(ctx, structuredoutput.Turn{Response: &structuredoutput.Response{}, Err: c.err})
return nil, err
}
// Test prompt IDs vary; all use the same supplied-data fixture definition.
copy := *req
copy.PromptID = "repair"
return &copy, nil
}
return req, err
})
}

View File

@@ -31,7 +31,7 @@ type Request struct {
Schema responseschema.Key `json:"schema"`
Context DiagnosticContext `json:"context"`
CaptureExchange bool `json:"-"`
Replay *Replay `json:"-"`
ContinueFrom *Exchange `json:"-"`
}
// PromptMetadata is the execution snapshot returned by the runtime when available.
@@ -44,33 +44,23 @@ type PromptMetadata struct {
Model string `json:"model,omitempty"`
}
// Exchange holds a section's original prepared conversation for read-only replay.
// Exchange is an opaque immutable conversation owned by the runtime adapter.
type Exchange struct {
PromptID string
Variables map[string]string
Messages []Message
FinalContent string
}
// Replay carries a section exchange and the prior turns of one repair branch.
type Replay struct {
Exchange *Exchange
Turns []Message
State any
}
type Response struct {
Exchange *Exchange `json:"-"`
RepairInstruction []Message `json:"-"`
Execution outcomes.ExecutionCounts `json:"execution"`
Content json.RawMessage `json:"content,omitempty"`
Provider string `json:"provider,omitempty"`
Model string `json:"model,omitempty"`
PromptTokens int `json:"prompt_tokens,omitempty"`
CompletionTokens int `json:"completion_tokens,omitempty"`
TotalTokens int `json:"total_tokens,omitempty"`
Prompt PromptMetadata `json:"prompt"`
Prepared *PreparedSnapshot `json:"prepared,omitempty"`
FinalContent string `json:"final_content,omitempty"`
Exchange *Exchange `json:"-"`
Execution outcomes.ExecutionCounts `json:"execution"`
Content json.RawMessage `json:"content,omitempty"`
Provider string `json:"provider,omitempty"`
Model string `json:"model,omitempty"`
PromptTokens int `json:"prompt_tokens,omitempty"`
CompletionTokens int `json:"completion_tokens,omitempty"`
TotalTokens int `json:"total_tokens,omitempty"`
Prompt PromptMetadata `json:"prompt"`
Prepared *PreparedSnapshot `json:"prepared,omitempty"`
FinalContent string `json:"final_content,omitempty"`
}
// PreparedSnapshot is the exact rendered request snapshot. Its messages may
@@ -108,3 +98,16 @@ type CallRecord struct {
type Client interface {
CompleteStructured(ctx context.Context, req Request, out any) (Response, error)
}
// Turn delivers a completed correction request, or its failure. The initial
// callback has no response. A failure is terminal for this coordination.
type Turn struct {
Response *Response
Err error
}
// Coordinator delegates bounded conversation progression to the runtime.
// Returning no request stops the branch; requesting beyond the budget exhausts it.
type Coordinator interface {
Coordinate(context.Context, *Exchange, int, func(context.Context, Turn) (*Request, error)) (bool, error)
}