Files
notarius/docs/roadmap/promptkit.md

104 KiB
Raw Blame History

Native PromptKit And CommandKit Migration

Purpose And Status

Make Notarius fully native to PromptKit, ConfigKit, and CommandKit. Use upstream configuration, execution, continuation, admission, inspection, commands, and presentation directly. Retain the application work of producing, assessing, publishing, and reusing typed artifacts. Historical duplication of library behavior is not a requirement to preserve it.

This is the single migration plan for PromptKit, ConfigKit, CommandKit, accepted-artifact caching, and artifact-first output. It owns the selected target, remaining choices, implementation sequence, removal inventory, acceptance criteria, and supporting evidence. The earlier discussion documents are retired; Git history preserves the discussion. All implementation phases are pending. Recording these decisions has not changed application behavior. Unrelated future work remains in the separate backlog.

Notarius predates many modern PromptKit capabilities: the library grew out of mechanics shared by Notarius and related applications. Historical duplication and slight behavioral differences do not express an intentional application requirement. Evaluate remaining machinery against actual domain needs, and consider upstream features when a reusable capability is missing.

Complete one coordinated cutover for configuration, PromptKit resources, and the CLI. Adopt native public conventions fully; preserving old CLI behavior is a non-goal. Supply migration guidance and maintained examples instead of aliases, parallel parsers, or compatibility wrappers for legacy behavior. CommandKit is part of the main migration, not a later follow-up. Release preparation and publication remain separate work.

Adopting the kits together lets standard inspection and execution share the same loaders and native facts. It avoids building temporary command and reporting adapters while those boundaries are already being replaced.

Baseline And Dependency Target

At Notarius revision 12eeb6369e40746b2c18c55397a289a1e1b5908d, the application uses configuration version 5 and Go 1.27. Its dependencies include old-path PromptKit v0.12.0 and ConfigKit v0.5.0, new-path ConfigKit v1.2.0, go-yaml v4.0.0-rc.6, and two external PromptKit provider catalogs.

The selected reviewed baseline is:

Module Version
go.maximumdirect.net/commandkit v0.7.0
go.maximumdirect.net/ckconfig v0.9.0
go.maximumdirect.net/ckprompt v0.10.0
go.maximumdirect.net/configkit v1.2.1
go.maximumdirect.net/promptkit v0.23.3, subject to U1's hash-capable release review

Keep Go's consumer minimum at 1.27, adopt stable go-yaml v4, and resolve compatible validation dependencies. The selected CommandKit integration releases document compatibility with PromptKit v0.23.2. Select v0.23.3 explicitly for its timeout attribution and cancellation/error identity fixes; that patch requires no configuration or request API migration from v0.23.2 and retains its dependencies. The original review used local release tags and tagged dependency manifests; remote release lookup was unavailable, so these are selected versions, not a claim that they are the latest releases. See the integration version guidance and tagged release notes, migration guide, format reference, and Go consumer guide.

The cache requires native MessageContentHash; v0.23.3's RenderedPromptHash is not a substitute. The sibling PromptKit repository has retired the original feature request, and its v0.24.0 release notes document the addition and a ConfigKit v1.3.0 dependency. U1 remains open to review the supplying API/release and verify a compatible assembled dependency set before changing the selected baseline. Independent work can proceed; completed cache integration must use the upstream capability. Do not add a local hash-policy fork. Neither a cross-backend global limit nor a comprehensive execution fingerprint is required. Remove obsolete imports and catalog dependencies when unused; avoid permanent replace directives and update attribution under the OSS policy.

Audita's reviewed v0.23.2 integration demonstrates native source assembly, continuation, observations, and CommandKit wiring. Adopt those mechanisms without copying its RoleClient, structured-output interface, or opaque exchange wrapper. Its per-proposal repair branches and lack of checkpoints do not define Notarius's artifact and recovery policies.

Target Architecture

Application Structure And Ownership

Part Responsibility
Application composition Load configuration and native sources; construct one shared PromptKit engine for producers, validators, and continuation runs.
Artifact-family modules Prepare domain inputs, select prompt IDs and response requirements, interpret responses, apply deterministic domain rules, and provide supported fallbacks.
Pipeline execution Connect generation to configured validation and strict/permissive handling; manage ordered dependencies, typed handoffs, bounded workers, and stable result ordering.
PromptKit Prepare and execute requests; maintain captured exchanges and conversations; enforce native retry, repair, turn, and admission budgets; report execution facts.
Application state and records Publish artifacts, cache final accepted artifacts, provide a small completion summary, and capture optional debug evidence with pipeline context.
ConfigKit and CommandKit ConfigKit owns configuration mechanics; CommandKit and ckconfig/ckprompt own standardized commands and presentation. Notarius supplies loaders, request builders, report content, and outcome mapping.

Use native PromptKit requests, results, exchanges, errors, and observers in model-executing application code. Remove the private provider-neutral completion API. Pure domain models, codecs, source-reference resolution, and deterministic transformations need no PromptKit dependency. Application interfaces continue to express artifact production, typed validation, output encoding, and state access.

Module Execution And Assessment (U3 Resolved)

Keep the existing typed stage boundaries and use native PromptKit values in model-executing code. Place engine construction and shared loaders in focused files in the internal/cli composition root. It constructs one native engine for the run. Retire internal/framework/llm once its adapter, scheduler, asset registry, and reporting responsibilities disappear; do not retain a package solely to hold a small constructor or introduce another LLM service interface.

Place one small shared producer helper alongside assessment in internal/framework/pipeline. Connect modules to it through ordinary functions:

  1. The pipeline supplies an operation-scoped assessment callback capturing the configured validator chain and acceptance criteria. It applies the selected strict/permissive mode when no acceptable result is obtained.
  2. The module prepares source material, references, and any normalization handle mappings once. It builds the native initial request and supplies response interpretation, correction-request, and supported fallback functions.
  3. The helper receives those functions, the shared engine, and relevant cache/debug collaborators. For cacheable work it prepares the initial request and looks up the accepted-artifact cache; on a miss it executes the retained request. Otherwise it uses ordinary native execution.
  4. For each executed response, the helper invokes module interpretation and pipeline assessment. When correction is appropriate, it supplies the module's next request to native coordination. PromptKit owns conversation state and turn iteration; the helper does not implement another retry engine.
  5. Return a typed application outcome. Cache only final accepted artifacts under the initial request's key. A cache hit follows the existing direct-reuse policy without model execution or renewed semantic assessment.

Request builders return promptkit.RunRequest; interpreters consume native results. Candidates, findings, and acceptance outcomes remain application-owned. Keep operation-specific state in local variables or closures, without an execution-plan language, replacement completion API, or copied conversation model.

Use one pipeline assessment implementation for the initial candidate and every correction, including the final permitted response. Do not repeat final validation merely because the helper returns to the outer runner. Deterministic stages keep their ordinary typed production/validation path without synthetic requests or conversation machinery. LLM validators call the shared engine directly and return domain findings; they do not automatically acquire a producer correction loop.

Modules receive operation data and collaborators, not physical storage roots or CommandKit reporters. Pure domain packages retain artifact models, codecs, reference resolution, and deterministic rules. internal/preflight retains its separate known-dependency checks. Exact argument grouping and helper signatures can be refined against representative extraction and normalization during implementation; ownership and placement are settled. U4 selects prompt roles and guidance ownership below; U5 selects retry controls and the correction budget. U6 selects strict-by-default handling and module-owned fallbacks below; configuration precedence and fallback behavior are resolved. Implementation is pending.

Retain the fixed input -> chunk -> extract -> merge -> normalize -> output shape, ordered steps and generated references, typed artifact-family boundaries, and deterministic public ordering. Keep modules independent of physical output, cache, and debug roots. Native adoption does not turn the pipeline into a general workflow engine or redesign durable D&D artifacts.

Command Surface And Native Inputs

Install the application pipeline family alongside ckconfig and ckprompt:

Command Responsibility
pipeline list Discover configured pipelines.
pipeline show <pipeline-id> Report resolved steps, lanes, modules, validators, and reference relationships.
pipeline validate <pipeline-id> Check composition and known prompt/profile dependencies offline.
pipeline run <pipeline-id> --input-file PATH Execute the pipeline and publish artifacts.
run <pipeline-id> --input-file PATH Native alias for pipeline run, with the same handler and inputs.

Use commandkit.Alias("run", "pipeline run"), not a forwarding command. Use notarius run in quick starts and document both routes. Install the full selected standard suite: native help/version; Bash, Zsh, and Fish completion; configuration scopes, validation, show, sources, explain, and diff; PromptKit resource listings and inspection for prompts, profiles, backends, namespaces, and collections; requirements discovery/validation and prepared-request preview; and explicit reachability, authentication, text, structured-output, and aggregate smoke tests. Include upstream convenience aliases and useful completion hints.

Use native typed declarations and supplied-value semantics. Defaults apply to omission, not explicit empty/false/zero values. Preserve native command-path, operand, repeated-flag, and -- parsing conventions. Check every occurrence of a singleton flag and use the last successfully parsed value.

Concern Selected input
Configuration Native --config, --overlay, and supported application --set POINTER=JSON; scope named app.
Pipeline input Required pipeline-ID operand and --input-file PATH.
Lanes Repeatable --lane ID; omission selects the normal lane set. Commas are literal values, not separators.
Output Application --output-dir PATH.
Cache --cache auto|bypass|refresh overrides caching for the entire invocation, including chunk generation. Otherwise resolve module-binding cache, pipeline cache, then auto.
References Repeatable --reference SELECTOR=PATH and --without-reference SELECTOR.
Presentation Native --format, --quiet, --verbose, --debug, and --progress.
Capture destination Application --debug-dir PATH, with the semantics below.

Keep the worker limit and semantic budgets in configuration, adjustable through application --set. U5 selects max_additional_turns; U7 selects concurrency.workers with a default of 64. Use ckprompt's reusable profile, session, and execution flags, including model, temperature, timeout, reasoning effort, and transient retries, for run-wide overrides across producers, validators, and continuations. Omission preserves normal selection; explicitly empty reasoning effort clears inheritance. Use native --llm-profile and --llm-session-id, retiring the separate reasoning-clearing control. Native transient retries remain distinct from semantic additional-turn budgets. Individual prompt selection and named prompt inputs belong on standard PromptKit commands; pipeline modules build their own requests.

Notarius retains reference-selector expansion, specificity, required slots, and external/generated-reference conflict checks. Pipeline reference selectors are not individual-request --llm-input-file bindings. Remove custom dispatch, assembled help, standard flag parsing, comma-separated --only, and bespoke --json switches. Replace config effective with config show and config profile with promptkit show profile. Old flags are migration evidence only.

Application Composition, Loaders, And Preflight

One constructor registers application commands, the native alias, ckconfig's app and promptkit scopes, ckprompt, and native help/version/completion/reporting. It does not load configuration, build an engine, or initialize run storage. Help, version, and completion work without application configuration.

Share two loading responsibilities across application and standard commands:

  1. Select application configuration through explicit --config, then existing environment/discovery policy. Let ConfigKit apply overlays, imports, inheritance, overrides, defaults, normalization, and structural validation. Retain its native result and selected main-source context. Validate an explicitly empty configuration path rather than falling through to discovery.
  2. Build the native engine from that application's source context using the source assembly below. Across commands, --config selects the application document; the PromptKit scope inspects the engine assembled for it. Request overrides do not mutate captured engine configuration. Do not emulate unsupported arbitrary --set overrides for the PromptKit scope.

Use small invocation-owned state through CommandKit's facilities to reuse loaded values. A pipeline shares one engine across all roles. config diff may load two distinct configurations within one invocation: keep their selections separate, including overlays/overrides. No persistent loader cache or configuration hash framework is needed.

Reuse the application resolver for module/options checks, typed compatibility, ordered dependencies, validator chains, and reference bindings. pipeline list only enumerates configuration; show renders resolved composition rather than building an engine to expand every profile. ConfigKit supplies loaded values and native provenance/secret-aware inspection, not domain pipeline semantics.

Centralize known prompt/profile dependency checks in internal/preflight. It accepts the resolved pipeline, module declarations, and native engine. Check every statically known dependency, including LLM validators and correction prompts, plus applicable/default profiles. Partial profiles alone are not errors. Share identifiers and selection logic with execution; no independent inventory, asset registry, or synthetic request execution is needed. Dynamic selections receive their final checks at runtime. Findings identify pipeline, step/lane, stage/validator, and resource where applicable.

pipeline validate and run call the same resolution/preflight functions; handlers do not invoke other commands or parse their reports. Keep config validate focused on configuration and remove its pipeline-specific mode. Offline preflight reads configuration and prompt resources but needs no transcript, reference contents, credential values, or provider calls. Concrete PromptKit preparation/validation owns completed-request checks, including schemas and credentials; provider smoke tests require explicit invocation.

Preflight does not discover configuration, register modules, build engines, initialize storage, or render CLI reports. Modules do not import it. Use native request builders for standard commands and add application callbacks only where needed, preserving supplied operator values. Output/cache/debug allocation belongs to the run path, with debug timing selected below.

Configuration, Sources, And Prompt Assets

Introduce application configuration version 6 as the only supported version. Preserve unrelated discovery, pipeline structure, and resolution behavior. ConfigKit continues to own loading and composition under ADR-0020. Retire promptkit.profile_file, profile_dir, and local_backend in favor of native companion sources; do not introduce a separate promptkit.config_file. Retire concurrency.total_llm and revise retry/worker settings as described below. U5 selects max_additional_turns and removes application structural-repair controls; U7 selects the shared worker pool and retired global-limit migration. U8 replaces the nested legacy cache settings with root-level cache_directory and pipeline/module-binding cache modes.

Use one source assembly for runtime, preflight, and inspection:

  1. PromptKit's bundled OpenRouter underlay.
  2. Notarius's embedded native prompts and profiles under collections.notarius.
  3. An optional, higher-priority operator underlay containing prompts.yml and recursive .yml/.yaml files under prompts/, beside the main application configuration actually selected by ConfigKit.

Anchor companion discovery to that selected main source, whether explicit or discovered. Configurations in the same directory share companions; there is no working-directory fallback. The two companion locations are strict peers, so filenames do not resolve conflicting operator records. Missing companions are fine; loading errors do not silently fall back to embedded defaults. Preserve native imports and source-relative resource paths.

Native declarations become the sole prompt assembly mechanism. Keep the embedded content tree and module ownership of prompt meaning, response requirements, and domain interpretation. Remove prompt/fallback-profile registration in AssetRegistry, per-module PromptAssetManifest plumbing, ModulePromptFS, shared-name maps, synthetic sharedassets mounts, and manual prompt enumeration for checkpoint fingerprints. The resolved schema design below also removes schema registration and filesystem-flattening machinery.

Convert maintained resources, examples, and fixtures to native identified lists, config: source controls, and sibling inheritance (extends: {id: parent}). Same-ID replacement supplies a complete record; distinct inherited variants can customize an existing resource. Explain native profile-null/backend fallback semantics. The optional PromptKit source header is version 2, independently of Notarius's application version. Do not translate legacy resource formats or synthesize compatibility aliases.

Select prompts by ID alone; remove obsolete PromptKit version selectors and related request/debug metadata. Retain current local resource IDs and use exact qualified references, such as notarius/dnd.spells and notarius/dnd-extraction. Keep the extraction profile's existing model/settings and its reference to the bundled unqualified openrouter backend. Inherited application profiles use sibling local IDs in the same collection. The bundled inventory differs from the old external catalogs and selects no model for an application request; document intentional replacements for removed catalog IDs. The bundled backend supplies no concurrency limit. Under U7, Notarius injects no backend concurrency or queue defaults; operators use native backend configuration.

Preserve prompt prose, message ordering, roles, and cache boundaries where their meaning carries forward. Review and deliberately revise initial/repair prompts for accumulated conversation history; preserving every old byte or retry message is not a migration requirement. Shared content should use native filesystem contexts and imports rather than reconstructed virtual trees.

Prompt Assets And Response Schemas (U2 Resolved)

Use the existing root assets package's ordinary embed.FS directly for native prompt sources and response schemas, with fs.Sub only where useful. Shared message files use their actual paths within the native filesystem context. Remove synthetic mounts, file copying, flattened schema namespaces, and module asset manifests. Retire readonlyfs when its prompt-mounting and asset-flattening consumers disappear; no replacement filesystem abstraction is needed.

Declare response format, validation mode, and schema path in native prompt definitions by default. Omit structural-repair attempt settings in bundled prompts to inherit PromptKit's upstream default under U5. Modules select prompt IDs, supply inputs, and interpret responses. U4 selects the same schema file for each initial/correction pair. Execution, inspection, and prepared-request preview then use the same declared contract without repeating schema selection in Go. U4 selects prompt roles and contracts; exact prose is implementation work. U5 selects budget ownership and defaults.

Use a request-level native OutputContract only for a demonstrated runtime- dependent contract. It replaces the whole configured contract, so supply the intended format, validation mode, and schema together while leaving structural repair at its native default. Do not inject a repair count merely to construct the override, or introduce a wrapper or per-call registry solely to reproduce Audita's approach. Operator prompt replacements must still produce responses the selected module can interpret; that module boundary does not require duplicating each prompt's schema selection in Go.

Default to native WithSchemaFS(assets.FS(), "."). Schema paths are relative to that selected root, for example dnd/spells/schemas/dnd_spells_llm.v1.json. They are separate from message-file source contexts: prompt underlays do not implicitly overlay schema files. When an operator configures an external schema directory, select PromptKit's native directory source instead of the embedded root. It replaces the whole schema source and must supply the schemas used by the selected prompts. Do not build an application filesystem overlay to mix embedded and external schemas. Transparent layering would be a separate upstream capability, not a requirement for this migration. Concrete configuration binding is an implementation detail within this selected source policy.

Keep private response schemas as self-contained JSON documents. PromptKit owns loading, compilation, structural validation, provider structured-output settings, and structural repair. Remove Notarius's response-schema byte registry, copied schema wrappers, defensive-copy layer, digest calculation, and duplicate structural-validation machinery. Native prepared snapshots provide diagnostic schema details; no schema fingerprint is added to the message-content cache key.

Keep typed decoding, source/reference checks, reconciliation safety, semantic validators, artifact codecs, and durable artifact schemas in Notarius. Response schemas describe model output; artifact schemas describe published and consumed domain artifacts. Preserve identifiers and versions belonging to actual artifact contracts. Retire private schema keys, registration records, hashes, and copied identity/version/name metadata whose only consumers were adapters, checkpoints, or reporting. No independent schema metadata system replaces them.

Implementation is pending. Verify native preparation of embedded and operator prompt definitions, shared-file paths, schema selection from embedded and external roots, and complete request-contract replacement where actually used. Missing or invalid schemas should follow native preparation behavior. No new schema check requiring synthetic requests is added to offline pipeline preflight.

Native Production, Continuation, And Assessment

Start with one conversation per producer operation: a chunk-plan generation, one lane's extraction of one chunk, or one lane's normalization pass. Use complete replacement model responses. Normalization may continue to return a reconciliation plan that deterministic code applies to produce the artifact. Per-entity branching is not part of this migration.

Prompt Roles And Correction Guidance (U4 Resolved)

Give each correctable producer an initial native prompt and a domain-specific correction prompt. Retain existing initial IDs and identify correction IDs explicitly in the module, for example notarius/dnd.spells.repair. Preflight uses those same declarations to check known prompt dependencies. Shared correction wording can live in ordinary asset files; distinct domain/operation templates supply the task and response expectations without a universal application-aware correction protocol.

The initial request supplies source material, references, and the task. Enduring domain principles belong in its system instructions. The correction template supplies a new user assignment with the current candidate's actionable findings; PromptKit carries the preceding exchange. Do not rerender the original source or manually replay prior responses. Review initial wording so later correction assignments remain valid. Include additional application-derived context only when needed to interpret the findings. Keep source/reference meanings and normalization candidate handles fixed throughout the conversation.

Use the same declared response schema for each initial/correction pair:

Role Response contract
LLM chunk planning Complete replacement chunk plan.
Extraction Complete replacement extracted artifact candidate.
Normalization Complete replacement reconciliation plan, applied afresh to the fixed original candidates.
LLM validation The validator's assessment response; no separate semantic correction conversation by default.

No correction envelope, patch language, or special repair-result type is needed. PromptKit retains structural validation and repair inside each native execution. Notarius assesses the response returned by that public boundary, independently of how many internal attempts produced it.

Use the same native default profile for an initial/correction pair. Preserve explicit application profile selections and run-wide execution overrides across both requests. An operator may select a different default profile in a correction prompt, subject to explicit request overrides. Validators retain their own profile selection. Pass native selections and overrides without copying resolved settings or adding a Notarius-wide repair-model configuration layer.

The component that detects a defect owns its actionable correction guidance:

  • Each rejecting validator supplies guidance describing the defect and the constraints a corrected answer must satisfy. A deterministic validator authors it alongside the check; an LLM validator requests an actionable explanation in its prompt and returns it through the application validation contract.
  • Preserve the explicit CorrectionGuidance boundary separately from operator diagnostics. Guidance need not prescribe an exact replacement, and validators do not construct the complete correction prompt.
  • Module interpretation and normalization safety checks supply their own guidance when defects arise before an artifact candidate can be materialized.
  • The pipeline collects guidance in configured validator order. The module's domain-specific template renders it into the correction assignment. Neither the pipeline helper nor the template duplicates validators' defect-detection logic. Reason codes and operator messages are not automatically instructions.

Assess the same candidate through every applicable configured validator before requesting correction. A rejection does not short-circuit later validators; collect all actionable findings they return. Reassess every revised candidate through the complete applicable chain, including earlier approvals and the final permitted response. The existing validator-owned guidance and ordered-chain semantics remain useful; replace the hard-coded CorrectionRequest message builder with native template rendering and retain useful ordered aggregation.

Applicability requires usable input: if interpretation cannot construct a typed candidate, its module supplies guidance and artifact validators cannot yet run. Cancellation or fatal execution failures can interrupt assessment. An unavailable validator assessment is a separate outcome, not a detected defect or an instruction to change the artifact. Do not convert provider errors into correction guidance; apply U6's strict/permissive handling when no acceptable result is obtained.

Ownership, response contracts, and default profile behavior are resolved. Refine exact template prose, correction IDs, and necessary operation-specific inputs while implementing representative workflows. Implementation is pending; U5 selects semantic budgets and U6 selects strict-by-default terminal handling and fallbacks.

Execution And Assessment Flow

The application flow is:

  1. Prepare source context, references, and candidate-handle mappings once. Build the native initial request. For cacheable work, prepare and check the cache before execution as described below; execute the retained request on a miss. Otherwise use ordinary Engine.Run. Capture an exchange when continuation or debug inspection needs one.
  2. Consume the native usable response or terminal failure, decode, and apply module interpretation and safety checks. A module can supply actionable findings before it has a materialized candidate. Assess each candidate through all applicable configured validators and collect their actionable guidance before requesting correction, as selected in U4.
  3. For actionable correction, call Engine.Coordinate with the captured exchange and the resolved max_additional_turns as native AdditionalTurns. Its callback selects a repair prompt and supplies findings; PromptKit retains the initial exchange and subsequent turns.
  4. Interpret every delivered revised response and rebuild from the fixed original inputs. Run the complete configured validation chain again for changed candidates, including validators that previously approved. Assess the final permitted response before declaring exhaustion.
  5. Stop on acceptance or a terminal condition. If no acceptable result is obtained, fail in strict mode or produce the module's supported fallback in permissive mode. Native stopped and exhausted outcomes do not themselves mean application acceptance. Cancellation stops work without constructing a fallback.

Accumulated history increases context use; U5's allowance bounds additional semantic correction while stopping as soon as the candidate is accepted.

Remove single_response_v1, raw-response ModelCandidate/SemanticCorrection transport, manual assistant/user replay, and correction-protocol capability checks. Replace runProducerAttempts' LLM retry loop with native coordination and application assessment. Adapt NormalizeRetry and producerRetryDirective to findings and explicit fallback information. Retain useful ordered collection and deduplication of actionable findings, rendered through a native repair prompt. Reason codes and operator messages are not automatically model instructions.

Retry Controls And Correction Budget (U5 Resolved)

Delegate transport retries and structural repair to PromptKit. Notarius selects only the allowance for application-directed semantic correction:

Concern Control and selected default
Temporary provider failures Native PromptKit transient retries; inherit the upstream default when omitted and preserve native execution overrides.
Structurally invalid model output Native output-contract policy; bundled prompts and application requests inherit PromptKit's upstream structural-repair default.
Actionable domain defects Application max_additional_turns, passed directly to native AdditionalTurns; application default 2.

max_additional_turns is a nonnegative maximum, excluding the initial producer response. Resolve producer-binding override, then pipeline override, then the application default of 2. Omission inherits; explicit 0 disables semantic correction. It applies to LLM-backed producers. Validators do not receive a separate semantic continuation budget, and deterministic stages do not acquire conversation machinery. Keep the field in configuration, adjustable through the existing application --set mechanism, without another dedicated CLI flag.

For example, a producer binding can override its pipeline allowance:

extract:
  module: dnd/spells
  max_additional_turns: 2

Two additional turns are an allowance, not mandatory executions. An accepted initial candidate needs none; an accepted first correction needs only one. A second turn is available when actionable defects remain and continuation is appropriate. Assess every delivered response, including the final permitted turn, and stop on acceptance or a terminal condition. PromptKit enforces the turn limit; do not add an application counter or retry loop. Internal transport attempts and structural repairs do not consume this semantic-turn allowance. The default is an application policy choice, not a measured claim about how frequently corrections will be needed.

For structural repair, omit output.repair_attempts from all bundled initial, correction, and validator definitions. Remove Notarius's pipeline/binding/ validator structured_output_repair_attempts fields and resolution plumbing. Do not replace them with a hard-coded default, including zero, in request construction or a native output-contract override. Explicit operator customization remains available through native PromptKit prompt configuration.

Notarius deliberately inherits future upstream default changes. An anticipated PromptKit default change is not a migration prerequisite, and its timing or chosen count is not a Notarius contract. Notarius still declares response format, validation mode, and schema; PromptKit owns the mechanics of obtaining a valid response. At the public boundary, handle a usable result or terminal failure, including a returned result reporting unsuccessful structural validation. Do not restart the operation to extend PromptKit's exhausted repair budget.

Remove blanket outer LLM retries, including repeated LLM validator invocations in executeValidationChain. Keep ordered candidate assessment and distinguish a validator execution failure from a domain rejection. Use native OperationTimeout when that operation scope is needed, without adding an arbitrary default or replacing broader application deadlines. Cancellation stops work.

Remove generic retries for non-LLM operations as well. Parsing, chunk materialization, deterministic merging, normalization-plan application, and encoding run without an outer attempt loop. Repeating those operations against unchanged inputs ordinarily cannot correct a defect. Return operation errors through normal pipeline failure handling; apply the selected terminal policy to a rejected deterministic candidate instead of regenerating the identical value. Cancellation stops work.

A demonstrated transient I/O recovery need belongs with that specific operation, where repeatability and failure conditions are understood. Do not retain a generic stage retry setting or framework in anticipation of filesystem or future remote- storage failures. After a fix, an operator can rerun the pipeline using the accepted-artifact cache for successful predecessors while cheap deterministic preparation runs again.

Retire producer and validator retries fields entirely, together with the generic outer attempt loop. Migration guidance maps intent: a producer allowance intended for semantic correction becomes max_additional_turns; native transport controls and prompt-owned structural customization cover their own concerns. Remove non-LLM retry settings without a replacement generic allowance. Do not promise equivalent total provider-attempt counts. U5 is resolved; U6 selects strict-by-default handling and module-owned fallbacks. Implementation is pending.

Terminal Outcomes And Module Fallbacks (U6 Resolved)

Use two modes when an operation does not obtain a successful, validated result:

Mode Terminal behavior
Strict Require a successful, validated result; otherwise fail the pipeline.
Permissive Produce the module's supported fallback, record a warning, and continue with that artifact.

The pipeline selects the mode; the module owns the fallback algorithm. Replace the failure-category policy object and separate fail_run, reject_output, and warn_continue choices with this selection. An empty artifact is a supported fallback where the domain permits it; no separate omit-output action is needed. The application default is strict for every role, including normalization. Modules do not declare their own defaults. Permissive handling requires an explicit configuration selection; normalization has no automatic exception.

Use mode: strict or mode: permissive at pipeline level, with an optional mode override on each configured producer module binding. Resolve the module binding first, then the pipeline setting, then the application default strict. Omission inherits; either explicit value overrides the broader setting. A strict pipeline can therefore permit fallback for one module, and a permissive pipeline can require strict handling for an essential module. This is binding configuration, not a default declared by the module implementation. Validators report findings; the producer's resolved mode governs disposition. Use ordinary configuration and the existing --set mechanism, without a dedicated mode flag.

Attempt application-directed correction when appropriate within the selected budget before terminal handling. Neither mode extends PromptKit's native retry or repair budgets or adds retries to deterministic operations. Preserve the distinction between a detected defect and an unavailable assessment for correction guidance and diagnostics; neither establishes ordinary acceptance. Cancellation remains cancellation and does not initiate fallback work. If a supported fallback cannot be constructed, fail execution.

Fallbacks must satisfy the structural artifact contract and invariants required by downstream stages. They need not satisfy every ordinary semantic acceptance criterion whose failure led to fallback. Do not rerun the complete ordinary validator chain as a condition of using the fallback or claim normal validation passed. Keep the necessary checks with module fallback construction; no separate configurable fallback-validation framework is needed.

Use these domain-specific fallbacks:

Module Fallback direction
D&D spell extraction A valid empty spell artifact, so that the failed chunk contributes no extracted spells.
D&D transcript chunking One chunk containing the entire transcript, preserving source order and references. This is the default fallback; no deterministic slicing machinery or sizing policy is needed for this migration.
Normalization Apply only safe consolidation groups from the final usable response afresh to the fixed originals, retaining everything else. If that response is unusable, use the deterministic merge of the input chunk artifacts. Reuse an already available merged input.

Keep source grounding, reference resolution, candidate isolation, and reconciliation safety checks. Unknown or overlapping handles remain invalid; partial normalization applies only compatible safe groups and leaves other original candidates separate. Use the final response and fixed original inputs, without recovering an earlier candidate or combining proposals across turns. Concrete module algorithms must make their supported fallback and required invariants explicit. Refine module-specific construction without introducing a generic fallback language. Whole-transcript chunking deliberately favors simple machinery for a rarely expected fallback; it does not promise bounded downstream request size.

Fallbacks may be published and consumed downstream, but they and dependent results affected by their incompleteness are ineligible for accepted-artifact cache lookup or storage. Carry the small application outcome facts needed to skip that reuse; do not add dependency fingerprints or a provenance ledger. Fully validated, unaffected predecessors remain reusable, allowing later runs to retry work that previously needed fallback.

Record why fallback was used and expose that qualification in the ordinary completion report, even when diagnostic verbosity hides individual warnings. Use the same application outcome for reporting and optional debug detail. Keep conversation state in process, outside ordinary artifact/cache payloads; explicitly requested detailed debug is separate. U6 is resolved; implementation and module-specific construction details remain in Phases 3–4.

Admission And Pipeline Workers (U7 Resolved)

Remove Scheduler, scheduledClient, and custom global queue/permit plumbing. Retire the separate cross-backend ceiling and the default/maximum extraction worker coupling to concurrency.total_llm. Use one shared pool for ready pipeline jobs, configured by the positive integer concurrency.workers, default 64. Omission uses the default; explicit 1 provides serial job execution. Keep this in application configuration, adjustable through --set, without a dedicated CLI flag or separate extraction/lane worker settings.

The pool serves both kinds of work:

Job Work covered by one worker slot
Extraction One chunk-and-lane extraction, including its validation and semantic correction.
Lane completion A ready lane's merge and normalization, including their validation and correction.

The limit counts total active jobs across both kinds, with no separate reserved capacity. The existing pipeline coordinator tracks dependencies and makes lane completion ready after that lane's extraction finishes. A worker never holds a slot while waiting for prerequisite jobs. Validators and correction turns execute within the current job instead of acquiring another worker slot. Preserve stable public result ordering regardless of completion order. Input preparation, chunking, and final publication retain their sequential placement around lane execution. This remains execution of the fixed pipeline, not a general workflow scheduler.

Count application jobs, not OS threads or all process goroutines. Do not derive the default from CPU count or GOMAXPROCS; model-backed jobs often wait for I/O or native admission. The pool bounds active application state and preparation as well as model-backed work. 64 supplies practical headroom for backend utilization while remaining well below the reviewed native default queue capacity; it is not a benchmark-derived optimum or a guarantee of backend saturation. Operators may tune it for workload and host resources independently of backend settings.

Use one shared native engine across all roles. Native capacity belongs to the engine and backend ID: profiles selecting the same ID share its pool; different backend IDs or separate engines do not automatically share limits, even for the same endpoint. Endpoint-only profiles and backends without a positive limit remain unrestricted by native backend admission.

Inject no Notarius-owned backend concurrency or queue defaults. Preserve native operator configuration and inherit upstream defaults rather than deriving backend settings from the worker limit. In the reviewed PromptKit baseline, omitted/zero backend concurrency is unlimited and has no queue; with positive concurrency, omitted queue capacity defaults to 1024. Explicit queue capacity, including zero, uses native semantics. These are upstream defaults, not values pinned by Notarius. With an unlimited backend, active jobs can generate concurrent calls up to the application's workload; the worker default does not imply a separate LLM ceiling.

For limited backends, the reviewed baseline bounds active model attempts by concurrency_limit and admitted operations by that limit plus queue_capacity. Ordinary execution admission spans preparation, retries, and repairs; split PrepareExecution itself reserves no capacity, with admission occurring during RunPrepared. Active permits are released during retry waits. A completed continuation turn releases admission before assessment, allowing a nested validator to use the same backend. Configure native limits at the intended capacity boundary. Worker counts need not be less than backend concurrency; native admission can queue waiting operations.

Handle ErrCapacityExceeded through U6's existing strict/permissive policy:

  • A producer admission failure fails in strict mode or uses the module's supported fallback in permissive mode.
  • A validator admission failure means assessment is unavailable; apply the producing module's resolved mode. Do not turn it into a semantic rejection or a correction instruction.
  • Cancellation stops work and does not initiate fallback.

Preserve the error and backend identity in diagnostics. Add no application retry, backoff, or second admission queue. A full native queue rejects admission without calling the model; the large default queue makes this unusual with modest worker counts, but explicit operator limits remain authoritative. Use ordinary fallback reporting and cache exclusions when permissive handling continues.

Aggregate concurrency across independent backends may exceed the former global ceiling. That change is intentional. Bounded pipeline workers are application work controls, not a substitute cross-backend LLM ceiling.

Retire concurrency.stage_workers.extract along with concurrency.total_llm. Migration guidance selects concurrency.workers independently and configures each backend natively. Existing explicit extraction-worker values can inform the new shared limit, but its scope now includes lane completion. Do not automatically translate a cross-backend ceiling into per-backend limits or preserve worker inheritance/caps from that field. Provide a native limited-backend example with usable queue capacity, without requiring operators to override the upstream queue default. U7 is resolved; implementation remains in Phase 4.

Accepted-Artifact Cache (U8 Resolved)

Replace both the checkpoint subsystem and the separate source-keyed chunk-plan store with one cache around the complete Notarius producer operation. Its value is an accepted answer to the original task, including an accepted chunk plan for LLM-backed chunk generation. This avoids repeating expensive model work after a downstream failure without caching every cheap deterministic pipeline step.

Prepare initial request -> obtain MessageContentHash -> look up <cache_directory>/<hash>.json
    Hit: decode the accepted artifact and continue the pipeline
    Miss: execute the retained request and complete the Notarius operation
          -> interpret, validate, and perform application-directed corrections
          -> on final acceptance, store the eligible artifact under the INITIAL hash

Use the upstream message-content hash of ordered message roles and exact text. Include message boundaries and order, excluding model/settings, session/execution IDs, timing, provider cache hints, and other non-content metadata. Add no stage, prompt-ID, library-version, whole-configuration, schema-policy, or implementation fingerprint. Actual input or schema text rendered into messages participates as message content; metadata outside those messages does not.

Change Intended reuse behavior
Fix a downstream stage Earlier artifacts remain reusable when their initial messages match.
Change initial message text, roles, order, or boundaries Cache miss.
Change a prompt ID, model, or settings without changing messages Cache hit remains possible.
Change deterministic interpretation or validation Existing accepted artifacts remain reusable; explicit bypass/refresh applies the new policy.

Obtain the key before model execution through native preparation. Preserve the prepared snapshot for execution on a miss using PrepareExecution/RunPrepared where supported, and discard the handle on a hit. OnPrepared execution hooks alone cannot short-circuit execution. v0.23.3's RenderedPromptHash includes session identity and cache hints; it is not the requested hash. The upstream API/release review in U1 must finish before completing this cache integration.

Keep the initial key throughout corrections. Cache the final Notarius artifact through its existing codec, not an intermediate response, RunResult, exchange, or prepared handle. For normalization, save the accepted normalized artifact after applying the reconciliation plan, not the proposed plan itself. Save only after the entire Notarius operation succeeds, including domain validation and any application-directed corrections. Failed, rejected, and validation-incomplete work is not cacheable. U6's fallbacks and dependent results affected by their incompleteness are ineligible for both lookup and storage, even if the downstream operation itself validates successfully. An unrelated accepted predecessor remains reusable. This requires a small eligibility fact, not a dependency fingerprint graph.

A hit loads the accepted artifact directly, with ordinary decoding/usability checks; absent or unusable entries are misses. It makes no model calls, rebuilds no result/exchange, replays no corrections, and performs no new semantic validation merely because the artifact was cached. It makes no claim of acceptance under changed validation policy and contributes no historical token usage to current execution totals.

For an LLM chunker, cache the final accepted plan after materialization, configured validation, and any application-directed corrections succeed. On a hit, decode the plan and deterministically materialize chunks from the current source. Retain ordinary structural and source-reference usability checks; an unusable plan is a miss. Do not rerun semantic validators merely because the plan was cached. The plan remains a domain result, without the old provenance envelope or a cached copy of the full transcript. Deterministic chunkers execute normally; do not manufacture a PromptKit request or a second hash policy to cache them.

Retire the guarantee of one canonical partition per source regardless of the configured chunking request. Matching initial messages reuse the same accepted plan; changed prompt/reference/input text can produce a miss and a different partition. Model/settings-only changes can still hit under the shared hash policy. Source identity alone no longer selects a plan. This deliberately supersedes the policy in ADR-0005 as the migration lands; preserve the accepted ADR and record a successor decision. Chunk-plan semantics and deterministic materialization remain application-owned; their storage and reuse use the same mechanism as other accepted results.

PromptKit owns everything inside each native request, including structural validation, transport retries, and repairs. Notarius consumes its public response or terminal failure and owns domain acceptance. Internal PromptKit attempts do not determine cache eligibility. Native diagnostic evidence does not transfer that ownership. No cached-response injection, replay API, exchange restoration, or PromptKit-owned write-through cache is requested.

Cache Configuration And Reuse Controls

Use root-level cache_directory for the shared cache root. Omission or an empty value selects os.UserCacheDir()/notarius: normally ~/.cache/notarius on Linux, respecting XDG_CACHE_HOME. An explicit directory is the exact root, independent of output/debug roots and run IDs. Document /var/cache/notarius as the convention for an explicitly configured system-wide installation. Directory selection does not enable or disable caching; modes govern cache I/O.

Use cache: auto|bypass|refresh at pipeline level, with optional module-binding overrides. Resolve an explicitly supplied --cache first, then module-binding cache, then pipeline cache, then default auto. Omission inherits. An explicit CLI value governs all cacheable work in the invocation, including chunk generation; without it, module overrides take precedence over the pipeline setting. Ordinary configuration and --set can select more specific bindings. No independent recording switch, separate chunk-cache control, or per-step recomputation selector is needed.

Mode Read existing entries Store newly accepted results
auto (default) Yes Yes
bypass No No
refresh No Yes

Modes never override the accepted-result eligibility rules. refresh replaces an entry only after a newly accepted result is obtained; failure or fallback leaves an existing entry untouched. bypass performs no cache I/O or directory creation. Rerunning with default auto reuses successful predecessors and executes previously unsuccessful work without a separate resume request.

Use a small filesystem lookup/write facility and ordinary temporary-file/rename publication. <cache_directory>/<hash>.json contains the accepted artifact or chunk plan, with no required manifest, acceptance ledger, historical provenance, or versioned checkpoint envelope. Lookup does not require original output or debug directories.

Missing entries are ordinary misses. Unreadable, malformed, or unusable entries are misses with useful diagnostics. Cache write failures warn and preserve the successful application result; do not rerun model work to compensate for a failed cache write. Cache I/O failures do not fail otherwise successful pipeline work. Invalid configuration still follows ordinary configuration-error handling. No expiration policy, garbage collector, or cache migration command is required.

Migration From Both Legacy Caches

Start the unified cache empty and populate it through ordinary execution. Ignore legacy checkpoint and source-keyed chunk-plan records; do not import, reinterpret, rewrite, or automatically delete them, even when their directories are beneath the new root. They do not supply the new message-content keys. Operators may remove obsolete directories when convenient.

Retire cache.checkpoints and cache.chunk_plans, including their separate directories and checkpoint recording switch. Remove --resume, --recompute-step, and the separate chunk-cache flag, including the previously planned --chunk-cache. Release guidance maps those intentions to root-level cache_directory, pipeline/module cache, and invocation-wide --cache. Explain automatic reuse, first-run population, and the move from a canonical source partition to request-content-based plan reuse.

Remove checkpoint identities, dependency fingerprints, invocation-wide barriers, recovery bookkeeping, and the separate canonical-plan lookup/publication policy. Retain domain plan interpretation and materialization without retaining a second cache mechanism. U8 is resolved; implementation remains pending in Phases 3–5, with completion dependent on verification of U1's upstream hash capability.

Inspection And Execution Bookkeeping

Delegate configuration/resource inspection, requirements, concrete preview and validation, and explicit smoke diagnostics to ckconfig/ckprompt through the shared loaders. Remove effective-YAML parsing, the profile-inspector wrapper, and copied effective-parameter types. A configured profile may be partial; use native prepared/result details for the actual execution target, including overrides. Standard commands retain their native reports, support/applicability distinctions, and failure behavior. Smoke status zero alone does not prove a check passed.

Use ordinary Run with OnPrepared when actual request details are needed for diagnostics; split preparation has a separate purpose for cache lookup. Native observations provide terminal facts even without a returned RunResult. OnPrepared is absent on preparation/admission failures and is not a completion record. Keep synchronous callbacks lightweight, compose capture/accounting observers, and synchronize shared state where operations overlap. OnPrepared runs while admission is held; persist retained evidence at normal operation boundaries.

Use native RunID/ParentRunID with application step/lane/chunk/validator scope. Remove LLMExecutionSummary, duplicate usage types, conversions, cloning, independent accumulators, debug completion wrappers, synthetic call IDs, and adapter profile-recording/forwarding. Native snapshots serve actual-target inspection; do not reconstruct a durable profile inventory.

Use one native ExecutionGroup per invocation for diagnostic totals. Explicitly enroll every actual initial, continuation, and nested LLM validator execution once, with application step/lane/chunk/validator context attached to observations. Use native aggregation directly, without application-maintained totals or a hierarchy of overlapping groups. Preserve terminal coverage and unknown-versus-zero usage. Cache hits create no execution and add no historical usage to current totals. Concrete observer composition, enrollment plumbing, and context attachment remain implementation choices within this contract. These facts support diagnostics and capture, not a mandatory consumer execution ledger. Notarius's application outcome supplies publication and fallback qualifications for the completion summary; assessment, failure, and reuse details support optional diagnostics and debug capture without independently maintained reporting state.

Artifact Delivery And Completion Output

Artifacts are the primary deliverable: publish accepted domain artifacts and permitted module fallbacks as files in the output directory. Preserve intrinsic domain content and artifact schemas. CLI --format controls the completion report, not payload encoding. Consumers need no run manifest, metadata bundle, persisted receipt, or execution history to use the artifacts. Retire reporting-only metadata and mandatory manifests instead of rebuilding them with native values.

Preserve the existing run-specific destination and domain artifact paths:

<output-root>/<run-id>/
    lanes/<safe-lane-id>.json
    chunk-map.json             # when explicitly enabled and available
    evidence-context.json      # when explicitly enabled and available

Keep the existing lane filename mapping and intrinsic payload schemas. Remove the mandatory index.json, manifest.json, and reporting-only rejected.json, warnings.json, and diagnostics.json sidecars. Their relevant diagnostic content belongs in ordinary diagnostics and optional debug capture; retain intrinsic domain content in the artifacts themselves. output_directory identifies the absolute run-specific directory, not the configured output root.

Consumers check process status, optionally decode the small completion report, and read documented artifact paths beneath the published directory. For configurable pipelines, the pipeline configuration identifies expected lanes. Migration guidance replaces receipt/index traversal with those paths, without introducing a replacement discovery manifest or a payload-schema change.

Submit one small structured value through Reporter.Report:

Field Meaning
pipeline Selected pipeline identifier.
artifact_count Deliverable artifacts, including reused artifacts; excludes intermediates, cache entries, and debug files.
output_directory Absolute published artifact directory.
debug_directory Optional absolute completed run-specific capture directory; omitted when capture is disabled.
fallback_count Number of producer operations that emitted a fallback; omitted when zero.

Supply a native text template for the ordinary successful case:

Success: dnd-session produced 6 artifacts in /work/results/run-123

When capture is available, append Debug capture: <run-directory>. Provide the path rather than mirroring the bundle. JSON/YAML render the same structured value through native serializers. A plain string would render natively as text but lose named fields in machine output; the template gives both presentations without parallel encoders. No report-version, run-ID, token, profile, warning-count, or validation-history collection is required.

When a completed run used a fallback, qualify the human summary and structured value with that fact, even if native diagnostic verbosity suppresses individual warnings. Count each producer operation that actually emitted a fallback once; do not count rejected attempts, individual validator failures, or downstream artifacts affected by that fallback. JSON/YAML include fallback_count only when positive. The native text template can render:

Success with fallback: dnd-session produced 6 artifacts in /work/results/run-123
2 operations used fallback.

Do not present the run as if all operations obtained normally validated results. Detailed reasons belong in optional diagnostics/debug capture, without per-producer histories in the completion report. Quiet mode continues to suppress stdout.

Wire results to stdout and diagnostics to stderr:

Mode Stdout Stderr
Ordinary Brief completion report. Failure explanation when needed.
--verbose Same completion report. Native information, warnings, and failures.
--quiet Nothing. Failure explanation when needed.
--debug Completion report with capture path when enabled. Native diagnostics and default progress.

Use native mode ordering (last enabled mode wins), explicit progress selection, and quiet suppression. Quiet still permits failure explanations and does not prevent rendering errors from failing the invocation. --format json --verbose keeps stdout valid report JSON. Standard configuration/PromptKit commands retain their own native values, with no universal Notarius envelope. Artifact-quality advisories retain their domain classification.

Optional saved reports use shell redirection, without a report-file flag or automatic duplicate persistence:

notarius run dnd-session --input-file transcript.json --format json > run-report.json

Publish artifacts and finish requested capture before submitting a successful completion report. Use native outcomes: status 0 for command completion, 1 for operational/cancellation/publication/presentation failure, 2 for invalid invocation or identifiable invalid configuration, and 3 for native config diff --check differences. Preserve ckconfig's structured classification and underlying causes; not every preflight failure is invalid input. A permissive run that completes with supported fallbacks succeeds with an explicit fallback qualification. Strict-mode failure, inability to construct a supported fallback, and other terminal pipeline failures fail the command.

On failure, supply a useful native explanation and nonzero status, without a new structured error-envelope protocol. --format does not serialize diagnostics. Retain CommandKit's separation of operation causes from output/diagnostic delivery failures. If publication succeeded before capture/report delivery failed, identify the published path and available debug location. Published artifacts remain valid; do not rerun model work or rewrite execution history. Subprocess callers check status before decoding a successful report; failed writes may leave partial stdout.

Debug Capture And Optional Diagnostics

Capture the most convenient debug representation, rather than optimizing for conciseness, exhaustive completeness, or strict deduplication. Native snapshots and readable views may repeat evidence. Reuse application decisions and native execution facts instead of maintaining a second ledger or transcript model. Debug formats are diagnostic data, not stable consumer contracts.

Controls And Location

Bind an application boolean to native --debug. Capture requires that boolean and a nonempty effective directory; a directory alone does not enable capture. Select --debug-dir PATH over debug.directory over /tmp/notarius-debug. Resolve nonempty configuration-relative paths against their source and CLI-relative paths against the invocation directory. Recognize explicit empty values first: --debug-dir "" or debug.directory: "" disables capture without fallback. A nonempty CLI override can re-enable capture over an empty configured root. An unusable nonempty destination fails requested capture rather than selecting another directory. Remove the configuration-relative ./notarius-debug default.

Capture selection is independent of the final reporting mode. --debug --quiet captures while suppressing stdout; --quiet --debug selects debug verbosity and capture. --debug=false disables capture through native binding semantics. Empty-directory debug retains native verbosity. No separate capture flag or capture-depth switch is needed.

Lifecycle And Failure Handling

Allocate the run bundle after application configuration and run identity are available, before engine construction and pipeline preflight. Argument/config-load failures may have no bundle. Keep allocation out of shared loaders.

Create <debug-root>/<run-id>/ and update <debug-root>/latest to a relative symlink to that run at session start. It points to the newly started session even if that session fails. Do not retarget at completion; concurrent starts use the last successful link update. Runs without capture leave it untouched.

One application recorder accepts initial context, native evidence, and application decisions, then finalizes once with pipeline/publication/capture outcomes. Persist as operations advance, without holding the whole pipeline's content in memory. Keep debug separate from output/cache and do not automatically delete bundles.

Requested capture is required work: allocation/link failures stop before model work; recording failures surface at the next normal application boundary; finalization failure fails the command. Observer callbacks retain errors for boundary checks because they cannot return errors. An existing pipeline failure remains primary, with capture failures reported additionally and causes retained. Already-written data and published artifacts remain available. Capture failures never cause model retries. Finalize before submitting the success report.

Inventory And Depth

Replace separate summary/trace categories with a compact run overview, configuration/composition snapshots, parsed source, and per-operation records. Organize operation directories by meaningful pipeline scope, with native execution IDs within those operations. Provide a readable run overview and clearly named structured snapshots and text views. latest is the selected link name; exact debug filenames, directory nesting, and serialized application fields are explicitly delegated to implementation. Evaluate them by ease of investigating an operation; do not freeze them as a public format or introduce another debug framework.

Capture Selected contents and depth
Run overview Application version, run/pipeline identity, timing, outcome, publication location, effective invocation selections/overrides, and navigation. Terminal facts come from the same application outcome used by the CLI.
Configuration and composition Native application and PromptKit effective snapshots, normally once each, plus resolved steps, lanes, modules, validators, and references.
Parsed source Generic input document with unit identities and content, independent of the original input file. If parsing fails, retain the input location and available diagnostics.
Operation inputs Actual source/chunk and reference content, incoming artifacts, prompt bindings/variables, and fixed original normalization candidates/handle mappings. Include directly even when duplicated elsewhere.
Prepared requests Native PreparedRun, including settings, messages, and exposed output-contract/schema details when available.
Execution events Native per-execution events in occurrence order, including attempts/retries/repairs, timing, usage availability, and terminal metadata.
Results and responses Native RunResult plus exact RawOutput in a separate editor-readable text file.
Conversations Full Exchange.Details() for every application-visible turn, including rejected turns, plus simple readable role-labeled messages. Repeated prefixes are acceptable. Use prepared messages when no completed exchange exists and label the representation.
Interpretation and assessment Candidate via its codec before acceptance; decode/safety failures; each validator's findings and correction guidance; application disposition, any fallback artifact and reason, and links to LLM validator executions.
Normalization and corrections Proposed reconciliation plan and applied candidate; constructed correction inputs even if the next preparation fails.
Deterministic stages and reuse Chunk plans/materialized chunks, merged/normalized artifacts, and the loaded cached artifact with the current hit decision. No reconstructed history or fabricated execution on reuse.

Use indented JSON for native structured snapshots and existing domain codecs. Retain selected content in full, with no arbitrary truncation. Enable exchange capture where debug inspection requires it, including single-turn calls; continuation's in-process exchange needs remain independent of durable capture. Keep operation inputs self-contained enough for direct investigation without requiring original inputs, a cache entry, or reconstruction of deltas. Do not archive every original file or serialize the pipeline's entire internal state.

On failure, retain available evidence, where execution stopped, errors, and application disposition. Preparation/admission failure may have no OnPrepared; operational failure may have no RunResult. Capture a coordination Undelivered response as unassessed. Native internal-attempt metadata is useful but intermediate provider response bodies are not required. No provider wrapper or complete provider-transcript guarantee is part of this design.

Data Handling And Diagnostic Commands

Operators secure debug data as they secure ordinary application inputs and outputs, for example with filesystem permissions. Debug contains internal data; no separate security/retention model or sanitized-bundle promise is required. Do not directly print/serialize ConfigKit-marked secrets, and preserve native credential omissions through secret-aware inspection. Do not heuristically scan, scrub, or rewrite prompts, source material, model output, or arbitrary strings.

Inspection/diagnostic commands use native reports and do not automatically create pipeline debug bundles. Operators may redirect those reports. Neither debug mode nor capture initiates provider probes. Explicit smoke checks retain native support, applicability, and outcome distinctions.

Unresolved Choices

The architecture and public direction above are selected. Resolve each remaining choice before its dependent work; independent work need not wait. Routine API placement and serialization choices are implementation details, not reasons to reopen the architecture. IDs retain continuity with the earlier roadmap.

ID Remaining choice or dependency Resolve for
U1 Review the message-content hash API/release documented upstream in v0.24.0, including its ConfigKit dependency; verify the final compatible dependency set against the selected baseline. Track in Phase 1; required to finish cache integration in Phase 5.

U1 is the only remaining unresolved dependency decision. Observer wiring, debug filenames, and other routine implementation details are delegated under the resolved designs below; all implementation phases remain pending.

U2 resolved: Ordinary embedded assets, prompt-declared output contracts, native schema-source selection, and removal of duplicate schema/filesystem machinery are selected in the schema design. Implementation remains in Phase 2; U4 selects prompt roles and contracts, and U5 selects native repair delegation and the application correction allowance.

U3 resolved: Typed stage boundaries, composition-root engine construction, and the shared pipeline producer helper with module functions and a pipeline assessment callback are selected in the execution design. Implement in Phase 3 and refine signatures during Phase 4 without reopening ownership. U4/U5 select prompt/guidance and retry behavior; U6 selects terminal handling, configuration precedence, and module-owned fallbacks.

U4 resolved: Domain-specific initial/correction prompt pairs, shared response schemas, default profile behavior, validator-owned guidance, and complete applicable-chain assessment are selected in the prompt and guidance design. Implement templates and operation-specific inputs in Phases 3–4; exact prose does not reopen ownership. U5 selects budgets; U6 selects strict/permissive handling.

U5 resolved: max_additional_turns defaults to 2, with binding over pipeline precedence and explicit zero disabling correction. Native transient and structural-repair defaults remain upstream-owned; bundled prompts and application requests do not pin structural repair. The retry-control design also removes generic non-LLM retries, all producer/validator retries fields, and the generic outer attempt loop. Any demonstrated transient I/O recovery stays with its specific operation. Implement in Phases 3–4; U6 selects terminal handling.

U6 resolved: Strict/permissive handling and module-owned fallbacks replace failure-category policies. Fallbacks satisfy artifact/downstream invariants without asserting ordinary semantic acceptance; use only the final response and fixed originals. Fallback and affected downstream work remain outside accepted- artifact caching, and completion reports disclose fallback use. Strict is the application default, including normalization; there are no module-declared defaults. mode resolves module binding, then pipeline, then that default. Chunking falls back to one whole-transcript chunk; spell extraction uses its empty artifact, and normalization uses safe final-response groups or merged input. Implement module-specific construction details in Phases 3–4.

U7 resolved: One pipeline worker pool and native admission replace the separate worker pools and application LLM scheduler. concurrency.workers defaults to 64 and bounds ready extraction/lane completion jobs together, independently of CPU count and backend settings. PromptKit owns backend capacity with no injected application defaults; admission failures use U6's mode without extra retries or queues. Retire total_llm and the old extraction worker field with explicit migration guidance. Implement in Phase 4.

U8 resolved: The unified accepted-result cache includes LLM chunk plans and replaces both legacy caches. Root-level cache_directory selects a per-user default or an explicit root, with /var/cache/notarius documented for system-wide installation. Mode precedence is explicit --cache, module-binding cache, pipeline cache, then auto. auto reads/writes, bypass does neither, and refresh writes newly accepted results without reading. Cache I/O is best effort; legacy records remain untouched and are not imported. Implement in Phases 3–5, subject to U1's hash dependency.

U9 resolved: Execution bookkeeping uses one native execution group per invocation, with all actual initial, continuation, and LLM validator executions enrolled once and application scope attached. Completion output adds only optional positive fallback_count, counting producer operations that emitted fallbacks. Preserve run-specific domain artifact paths and schemas; retire index, manifest, and reporting-only sidecars and migrate consumers to documented paths. Observer wiring and debug filenames/layout are explicitly implementation choices, not unresolved architecture or stable debug contracts. Implement in Phases 3–5.

U10 resolved: CommandKit versions, capabilities, commands, loaders/preflight, output, debug policy, and coordinated cutover sequencing are selected above. The former later CommandKit phase is retired. CLI compatibility is a non-goal; existing artifact schemas remain domain contracts. Any actual artifact-format change discovered during implementation needs an explicit compatibility decision, not an assumption that native CLI adoption authorizes it.

Implementation Sequence

All six phases are pending. Each delivers a reviewable outcome, not a separate release or compatibility layer. Tests and canonical documentation accompany behavior changes; the last phase verifies integration and completeness. Record major boundary changes in successor ADRs without rewriting accepted decisions.

Phase 1: Establish The Migration Baseline

  • Record the selected dependency set and review U1 against the upstream release. Keep the cache dependency explicit without blocking independent work.
  • Map affected ADRs, current callers, contracts, and removal inventory. Carry forward Audita evidence without adopting its application wrappers.
  • Carry resolved U2–U10 designs into their dependent phases without reopening ownership. Leave observer wiring and debug filenames to implementation.

Completion evidence: reviewed dependency baseline, upstream dependency, scope, and outstanding choices are explicit. Upstream availability alone does not establish compatibility of the assembled Notarius dependencies or completion of the migration.

Phase 2: Build Native Composition, Sources, And Preflight

  • Implement the resolved U2 design, update available dependencies, introduce configuration version 6, convert native embedded/companion resources, and use native schema loading with embedded/default or operator-selected roots. Omit bundled structural-repair settings and remove their application override fields.
  • Build the small CommandKit application, shared ConfigKit/engine loaders, standard suites, pipeline declarations/native alias, and internal/preflight.
  • Add maintained embedded-only and operator-override examples. Remove duplicated loading and asset assembly as callers migrate.

Completion evidence: runtime/inspection/preflight share intended definitions; source contexts, shared message paths, prompt-declared output contracts, selected schema roots, strict peer conflicts, and replacement work offline. Help, version, and completion need no configuration. Preflight checks known dependencies without credentials or synthetic requests. config diff handles two selections.

Phase 3: Prove A Representative Native Pipeline

  • Implement the resolved U3–U5 collaboration and retry policy and U6's strict-by-default model, mode precedence, and representative module fallbacks. Connect module request/interpretation functions to the pipeline assessment callback through the shared producer helper; refine concrete signatures here. Implement domain-specific initial/correction prompts using the same schema and default profile, supplied overrides, and validator-owned correction guidance. Resolve max_additional_turns with omission/zero distinctions and pass it to native coordination; inherit native transport and structural-repair defaults.
  • Exercise an LLM validator on the same engine and normalization with fixed handles and fresh plan application. Wire U9's single invocation execution group with application scope for initial, continuation, and validator executions.
  • Run through the new command surface and alias, using the small completion report, native outcomes/streams, and useful debug capture from the outset. Begin the resolved U8 cache integration and U9 reporting/capture implementation without a provisional local message hash.

Completion evidence: fake-provider workflows exercise accumulated history, collection of independent validator findings without rejection short-circuiting, full candidate revalidation through the same assessor, final-turn acceptance without duplicate final validation, strict failure, permissive fallback, and useful failures. Unavailable assessments remain distinct from actionable defects. The executable publishes artifacts and reports them through native presentation without the private completion/correction API.

Phase 4: Migrate All Roles And Remove Superseded Machinery

  • Migrate every LLM-backed chunker, extractor, normalizer, and validator; preserve deterministic paths and ordered validation. Apply U4's prompt/guidance design and U5's retry policy across roles; implement U6's module fallbacks and U7's shared worker pool and native admission handling.
  • Remove the adapter, copied completion contracts, manual replay/capability checks, generic outer attempt loop, all producer/validator retries fields, global scheduler/ceiling, and coupled worker settings.
  • Replace failure-category policy fields with the selected mode configuration and module-owned fallback behavior.
  • Finish native command/inspection migration and remove custom dispatch, flags, encoders, verbosity handling, and obsolete asset registration. Retire the emptied internal/framework/llm package after moving engine assembly to the composition root and producer collaboration to the pipeline.

Completion evidence: every role uses the shared native engine and selected admission policy. One worker limit bounds extraction and lane completion together, including serial operation without dependency or nested-validator deadlock. No production path depends on the private completion API, latest-response-only correction, generic stage retries, global ceiling, or legacy CLI infrastructure. Obsolete tests retire with the machinery; surviving domain requirements remain.

Phase 5: Complete Cache, Output, And Capture Integration

  • Resolve U1's supplying release and implement U8's unified lookup at initial preparation, accepted-artifact/plan storage under that hash, directory and mode precedence, and best-effort cache I/O. Retire both legacy cache mechanisms.
  • Complete U9's artifact-first publication/report contract. Preserve domain paths and schemas, add fallback_count, and retire index/reporting sidecars. Remove checkpoint fingerprint graphs, mandatory manifests/receipts, durable execution ledgers, and duplicate accounting; do not retain them to populate summaries.
  • Complete capture inventory, per-operation evidence, directory/empty-value behavior, latest, and recording/finalization failure handling.

Completion evidence: downstream failure/fix/retry reuses accepted predecessors; message changes miss and settings-only changes can hit; corrected answers use the initial key; hits make no model calls or fresh validation claims. Fallbacks and affected downstream work skip cache lookup/storage; unaffected accepted work remains reusable. Chunk plans use the same cache, rematerialize against current source, and no longer pin one partition across changed messages. Mode precedence, cache I/O failures, and leaving both legacy stores untouched are verified. Artifacts need no metadata bundle and consumers use documented domain paths. Machine output and debug capture obey the selected contracts, including omitted zero/accurate positive fallback counts, failures after publication, and quiet-mode capture. Observer wiring and debug organization satisfy the selected scope and inventory without becoming separate public contracts.

Phase 6: Verify And Document The Coordinated Cutover

  • Complete the acceptance checks and removal inventory, including dependency cleanup, maintained examples, and the assembled command surface.
  • Finish release migration guidance, attribution, successor ADRs, and affected canonical documentation. Update this roadmap's status and reconcile the separate future-work backlog.
  • Confirm no compatibility scaffold, deferred CommandKit adoption, local hash policy, or obsolete metadata requirement remains in the finished migration.

Completion evidence: PromptKit, ConfigKit, and CommandKit adoption, accepted- artifact reuse, artifact delivery, diagnostics, and maintained offline workflows are verified together. Remaining work is explicitly separate future scope. Release preparation/publication is governed separately by release policy.

Removal Inventory

Retire or replace Retained application responsibility
StructuredLLMClient, private request/response types, adapter conversions, emptied internal/framework/llm package Composition-root engine construction, native module request/interpretation functions, and pipeline assessment/helper.
AssetRegistry, PromptAssetManifest, ModulePromptFS, synthetic mounts, flattened schema namespaces, readonlyfs, manual fingerprint enumeration Ordinary embedded content, native source assembly, and module meaning.
Private response-schema registry/wrappers, byte copies, digest and reporting-only identity metadata, duplicate structural validation Native prompt output contracts and schema loading; application interpretation and durable artifact contracts.
single_response_v1, raw ModelCandidate/SemanticCorrection, manual replay, correction capability checks Typed candidates, actionable findings, native exchanges, assessment.
Generic runProducerAttempts outer loop, repeated validator executions in executeValidationChain, all producer/validator retries fields, module rerun directives, hard-coded CorrectionRequest message builder Ordinary deterministic execution, native LLM execution/continuation, ordered validators, validator/module-owned guidance, native domain correction templates, and domain acceptance/rejection/fallback.
Application structured_output_repair_attempts fields/resolution and explicit bundled output.repair_attempts Native upstream structural-repair default, with operator customization through native prompt configuration.
Failure-category validation_policy fields and separate fail_run/reject_output/warn_continue actions Strict/permissive selection, module-owned fallback construction, and explicit fallback outcomes.
Scheduler, scheduledClient, concurrency.total_llm, concurrency.stage_workers.extract, separate extraction/lane worker pools One ready-job pool controlled by concurrency.workers, default 64, and native backend admission.
Checkpoint stage/dependency/configuration fingerprints, invocation-wide barriers, manifests, acceptance/provenance envelopes Small accepted-artifact cache using the upstream message hash, ordinary load checks, explicit controls.
Separate source-keyed chunk-plan store, canonical-plan provenance/publication policy, nested legacy cache configuration, --resume, --recompute-step, separate chunk-cache flag Unified message-content lookup/storage, domain chunk-plan interpretation/materialization, cache_directory, pipeline/module cache, and invocation-wide --cache.
Inspection wrappers, copied effective parameters, usage summaries/conversions/accumulators, debug completion wrapper, synthetic IDs, profile recorder Native facts with application scope for diagnostics.
Custom dispatch/help/version/parsing, command-specific loading, copied inspection, legacy aliases/flags, parallel encoders and verbosity handling Native declarations/loaders, pipeline handlers, domain input checks.
Mandatory index.json/manifest.json, reporting-only rejected.json/warnings.json/diagnostics.json, extensive receipts and durable execution ledgers Preserved domain artifact paths and schemas, one small native completion report with optional fallback_count, ordinary diagnostics, and optional debug evidence.
Summary/trace hierarchy, heuristic debug-content scrubbing, custom transcript/deduplication machinery Convenient run/operation capture and native secret/credential omissions.

Acceptance And Verification

Use real CommandKit, ConfigKit, and PromptKit in representative assembled flows with a fake at the provider boundary. Keep the suite deterministic, offline, and independent of real credentials. Test domain interpretation/assessment at their package boundaries; use subprocess tests for streams, exit status, alias wiring, and representative machine output. Do not duplicate upstream parser/renderer/ retry tests or freeze complete help text and native debug snapshots.

Area Required evidence
Configuration/assets Version 6/new resources work; old formats intentionally rejected; companions/native imports retain main-source context; peer conflicts fail; same-ID replacement, real shared-file paths, prompt-declared contracts, and embedded/external schema roots work; necessary request overrides replace the complete contract.
Commands/loaders Required operands, native alias, supplied-value distinctions, references, and run-wide overrides reach intended handlers/roles; help/version/completion need no config; diff selections remain distinct.
Preflight/inspection Same definitions across execution and inspection; missing statically selected resources fail with context; partial profiles alone pass; no provider calls, transcript, or credential requirement for offline preflight.
Domain workflow Representative extraction and normalization preserve source/reference identity, fixed handles, safe plan application, ordered validators, artifact schemas, and stable output ordering; one assessor checks initial/corrected candidates without an extra final pass; deterministic stages and LLM validators need no producer conversation loop.
Prompt/guidance contracts Domain-specific initial/correction pairs use the same response schema and default profile; explicit profile/execution overrides reach both; operator correction-profile defaults work when not overridden; all applicable validators assess the same candidate despite earlier rejection; their actionable findings reach the correction template; module interpretation defects and unavailable assessments remain distinct.
Retry controls max_additional_turns resolves binding, pipeline, then application default 2; explicit zero disables correction; acceptance stops early and the final allowed turn is assessed. Bundled declarations/application overrides leave structural repair unspecified, native operator customization remains effective, and tests do not freeze upstream default attempt counts.
Non-LLM failure handling Deterministic errors propagate without generic retries; rejected deterministic candidates reach terminal policy without regeneration; cancellation stops work. Producer/validator retries fields and the generic outer attempt loop are retired.
Continuation/failures Accumulated history, full revalidation and final-turn assessment; no blanket retries beyond native exhaustion; validator failure differs from rejection; no earlier-candidate recovery or cross-turn proposal union; cancellation stops without fallback work.
Strict/permissive outcomes mode resolves module binding, pipeline, then strict; omission inherits and either explicit value overrides. Strict is the default for every role, including normalization, without module-specific defaults. Strict failure when no acceptable result is obtained; permissive handling produces module fallback with required artifact/downstream invariants, without rerunning the ordinary full validation chain or claiming normal acceptance. Chunking fallback is one whole-transcript chunk preserving source order/references. Fallback construction failure fails execution; successful fallback use is disclosed in human/machine completion reports.
Admission/workers One shared extraction/lane completion pool obeys concurrency.workers, default 64; explicit 1 runs serially without dependency or nested-validator deadlock. Only ready jobs occupy slots; result ordering remains stable. Backend limits remain independent with no application-injected defaults. Producer/validator admission failures use U6's mode, preserve useful diagnostics, and add no retries; cancellation releases capacity. No separate global LLM ceiling.
Accepted-artifact cache Downstream recovery; initial-message keys; corrected final artifact stored late under the initial key; changed messages miss; settings/ID-only changes can hit; no model/revalidation on hits; unusable entries miss. Failed/rejected/incomplete work is not stored; fallbacks and affected downstream work skip lookup/storage while unaffected accepted predecessors remain reusable; no required history.
Chunk-plan reuse Final accepted plan stored under the initial request hash after materialization/validation/correction; hits rematerialize with structural/source-reference usability checks and no renewed semantic validation. Changed messages can select a new partition; deterministic chunkers run without synthetic cache requests.
Cache controls and migration Root-level directory default/explicit selection; mode precedence CLI, module binding, pipeline, then auto. auto reads/writes; bypass does no I/O or directory creation; refresh only replaces entries after new acceptance. Unusable entries miss; cache I/O failures preserve successful pipeline results. Both legacy stores remain untouched and unread; no conversion or separate canonical-plan mechanism.
Output/outcomes Preserved run-specific domain paths, lane filename mapping, and payload schemas; optional chunk-map/evidence-context publication remains. No index/manifest/reporting sidecars. Native text/JSON/YAML summary includes positive fallback_count and omits zero; count emitting producer operations, not attempts, validator failures, or affected derivatives. Quiet suppression, verbose stderr, debug path, and useful nonzero failures hold. Publication survives later capture/report failure without model reruns.
Execution observations One native invocation group enrolls actual initial/continuation/LLM validator executions once, with application scope; native totals preserve unknown-versus-zero usage without application accumulators or overlapping-group aggregation. Result-less failures remain observable; cache hits add no executions/history; native execution success remains distinct from acceptance.
Debug controls/lifecycle Default/CLI/config/explicit-empty paths and native mode ordering; quiet capture; allocation before engine/preflight; start-time relative latest, no completion retarget; disabled runs leave it untouched; capture failures preserve primary causes and publication facts.
Debug evidence Available input/source/native snapshots, readable text, all visible turns, assessments, normalization and deterministic outputs, cache hits, and failed-run evidence; no invented provider transcript; ConfigKit secrets/native credentials omitted, no heuristic content scrubbing.
End-to-end migration Maintained minimal and complete D&D workflows, session routing, reference behavior, raw artifacts, reports, and migration examples work through the new executable. No postponed CommandKit phase.

Run focused checks during implementation and the repository-wide checks for the completed cutover under the testing policy:

go test ./...
go vet ./...
go build ./cmd/notarius

Check maintained examples, dependencies, attribution, documentation links, and removal coverage. Live provider smoke and human model-quality evaluation remain optional, outside the default suite. These migration checks add no permanent release gates. Documentation completion does not claim that implementation acceptance checks have passed.

Migration Guidance And Documentation

Provide practical release guidance for operators and subprocess callers:

  • New commands/flags, native run alias, --format, stdout/stderr, outcomes, and saving optional reports by redirection. No legacy CLI compatibility promise.
  • Version 6 and native resources, embedded-only versus companion examples in appropriate directories, imports, qualified IDs, replacement/inheritance, native local backends, removed catalogs/versions, and separate source versions.
  • Shared concurrency.workers default 64, ready-job scope, explicit serial execution, native backend concurrency/queue settings, and removal of the global limit and old extraction-worker field without automatic backend translation. Explain strict/permissive handling of producer and validator admission failures.
  • max_additional_turns default/precedence/zero behavior, and intent-based migration from LLM retries. Explain retirement of all producer/validator retries fields and generic non-LLM retries, upstream-owned structural defaults, removal of application repair controls and bundled counts, native operator customization, accumulated-history costs, the universal strict default, pipeline mode with module-binding overrides, and each module's fallback behavior and downstream guarantees, including the whole-transcript chunking fallback.
  • Unified accepted-artifact and chunk-plan caching, message-based misses, settings-only reuse, root-level cache_directory and the per-user/system-wide conventions, pipeline/module cache, and invocation-wide --cache precedence. Explain refresh/bypass, best-effort I/O, fallback/affected-work exclusion, removal of both legacy mechanisms without importing/deleting records, and retirement of canonical-source partition reuse.
  • Artifact delivery through preserved run-specific domain paths and lane filename mapping, without index/manifest/reporting sidecars or receipt traversal. Document fallback_count and its omission/counting rules; consumers use the published directory and pipeline-defined lanes. Preserve intrinsic artifact schemas and document any explicitly selected format change.
  • Debug root/default/empty behavior, latest, native mode ordering, failure behavior, and ordinary operator responsibility for application data.

Update canonical configuration, CLI, operations, internal overview, LLM, pipeline, modules, state, and affected integration/consumer guides as behavior lands. Remove superseded manifest/receipt contracts then; keep current-behavior documentation accurate until implementation. Update architecture policy for native dependencies and changed scheduler, retry, asset, cache, and reporting boundaries.

Record significant changes in successor ADRs, including asset assembly, validation retries, module retries, PromptKit configuration, the canonical chunk-plan policy, admission, caching, and CLI/reporting ownership. Preserve accepted decision text and update supersession status under documentation policy. This roadmap owns implementation status; Git history retains earlier discussions.

Evidence And Reading

These sources support the selected design, not a claim of completed Notarius implementation. Sibling links assume the local repositories are available; the tagged PromptKit links below preserve the reviewed baseline.

Native Library Guidance

Audita Comparison

The comparison was completed on 2026-11-18 against Audita revision da5b8913d1206a8e8f0a784b2322760385548a60, using PromptKit v0.23.2, ConfigKit v1.2.1, CommandKit v0.7.0, ckconfig v0.9.0, and ckprompt v0.10.0. The observations below concern that revision; linked sibling files may evolve.

Evidence Lesson carried into this roadmap
Engine assembly, source tests, and assembly failure tests One captured engine, native underlays and main-source-relative companions, whole-record replacement, preserved file contexts, strict peer conflicts, and no silent fallback on operator-source errors.
Adapter tests Ordinary embedded schemas and complete per-request output contracts work with operator prompt replacement. U2 chooses prompt-declared defaults for Notarius; Audita's schema registry and per-call overrides are not requirements to copy.
Native coordination bridge, repair flow, conversation tests, and repair tests Captured exchanges, accumulated history, complete candidate revalidation, final-turn assessment, and canceled/undelivered response handling.
Prompt guidance Durable system instructions, an initial user task, and subsequent repair assignments suit native continuation.
Observations and CLI composition Native execution totals with application scope and explicit nested-validator enrollment; standard CommandKit suites using shared loaders.

Audita repairs individual transcript corrections in independent branches; Notarius's selected unit is a whole producer operation and typed artifact. Audita's wrappers, response-schema registry, per-proposal policies, and absence of a cache do not define Notarius requirements.

At the reviewed revision, focused LLM, prompt/schema, runner, and proposal-generation tests passed, together with subprocess checks for correction conversations, cancellation during repair, and repair-disabled/output selections using a local fake provider. These historical checks do not establish live-model quality or completion of Notarius's migration acceptance checks.