Repository navigation
Model assembly base: per-slot model configuration, shared advice interfaces and contract fields for returned models - #37
Merged
Conversation
Optional RegisteredModelIdentity fields (kind, task_type, criteria, dataset, components, fusion, weights_stage, rubric_version), FeedbackProvenance registered_models and support_model, and RegisteredModelAdvice criterion results, sentence units, aggregation and rule reconciliation; synthetic registrations may show labels. Unset fields keep historical bytes. Move the R2.support input and export projections to models/v3/projection, advice production to models/v3/consumer.AdviceSource, add describe_registration with per-protocol handlers, measure() on every model, return_cli for verify-return, and CAREER_LAB_MODEL_REGISTRATIONS. Behaviour is unchanged: inputs and advice match the pre-move tree byte for byte.
expansion-v3-d4178f30... (previous expansion-v3-14439473...): optional model-assembly fields on RegisteredModelIdentity, FeedbackProvenance and RegisteredModelAdvice, and two registration configuration error codes. Examples are unchanged. engineer-review-v1 follows the new legacy manifest digest. Generated with the official exporters; nothing edited by hand.
CAREER_LAB_<SLOT>_{PROVIDER,KEY_FILE,BASE_URL,MODEL,MAX_TOKENS,TIMEOUT} and
JUDGE/SUPPORT _INPUT_BYTES, read once at startup. SUPPORT falls back to
JUDGE, every slot to the shared variables and --provider, then provider
defaults; with no slot variable set requests match main byte for byte.
openai_compatible needs an explicit base URL and model and may omit the
key only on loopback hosts. Keys without an sk- prefix are accepted and
never logged, stored or echoed in errors. Feedback provenance records the
Judge slot and, when it differs, the support model. The worker prints each
slot's provider, model and non-secret fingerprint at startup; invalid
configuration refuses to start with only an error code. Real calls keep
retries=0.
expansion-v3-a7136a29... (previous expansion-v3-14439473...). The export now fingerprints runtime/model_slots.py and runtime/model_adapter.py with the other assembly inputs, so their startup error codes are published; the reproducibility test's input list follows. The model-assembly history entry also records the slot and registration variables. Examples are unchanged from main; engineer-review-v1 follows the new legacy manifest digest.
A slot with its own _PROVIDER takes key file, base URL and model only from its own variables or that provider's defaults; SUPPORT without a provider takes the whole JUDGE slot; other slots take the whole shared layer. Identity variables without their slot provider refuse to start (model_slot_provider_required), so a shared key is never sent to another provider's endpoint. Feedback provenance records the Judge and support endpoints actually called when slot variables configure them. A criterion registration is now only checked for shape and handed to consumer.criterion_source, which loads nothing yet: feedback, provenance and model loading match a deployment without it. The W14 receipt test pins the local models through configured_slots, which assembly reads.
expansion-v3-bdeee6db... (previous expansion-v3-14439473...): optional FeedbackProvenance.base_url and SupportModelIdentity.base_url, the model_slot_provider_required startup code, and the updated model-assembly history text. Examples are unchanged from main; engineer-review-v1 follows the new legacy manifest digest.
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Base of the model-reception wave (task book 143, line 144): the contract fields, shared advice interfaces and per-slot model configuration that the returned models from the data/model owner will plug into. No new user-visible behaviour; the zh/en feedback page still shows "等待模型接入".
What changes
RegisteredModelIdentitygainskind,task_type,criteria,dataset,components,fusion,weights_stage,rubric_version;RegisteredModelAdvicegainstask_type, per-sentenceunits,aggregation,reconciliationand new error codes;FeedbackProvenancerecords multiple registrations and the support model. Advice from asynthetic_fixtureregistration may now carry labels and evidence (decision 2026-10-12) and is always markedsynthetic_mechanism_only.CAREER_LAB_<SLOT>_{PROVIDER,KEY_FILE,BASE_URL,MODEL,MAX_TOKENS,TIMEOUT}; a slot's identity is taken from one layer only (its own variables with the provider's defaults, or the shared configuration as a whole), never mixed, so a shared key is never sent to another provider.openai_compatibleis a new provider for local services; a non-loopback endpoint without its own key refuses to start. With no slot variables set, requests are byte-identical tomain. Real calls stayretries=0; keys never reach logs, database or provenance.models/v3/consumer.py(AdviceSource),projection.py,registry.describe_registration,measure(),return_cli.py, andCAREER_LAB_MODEL_REGISTRATIONS({"relation": …, "criterion": …}; the criterion entry is parsed only). Moves are byte-identical to the pre-move behaviour (SPLIT-01 against sessions recorded onmain).bdeee6db, engineer-review-v1 →37bedd6f(previous14439473);manifest_historygains one entry;errors.json+9 codes.Review and verification
080-review-144.md: first round sent back (shared key could leak across providers through field-level fallback; criterion entry was loaded instead of parsed), fixed inea82f04, second round 可合入. The reviewer re-recorded themaingoldens, re-generated the contracts byte for byte, and reproduced the three key-routing cases over HTTP.1521d8e: lint 0 new; backend 3298 passed / 31 skipped; frontend 404 / 7; 72 state checks; build; per-ID "Regression gate: 0 problems"; browser 20 steps zh+en; MCP bilingual pass. The branch containsmaind163747, so the merge tree equals the gated tree.Not claimed / registered
No real provider was called; model quality is not verified. Startup fingerprint is printed by the worker only (serve keeps its whitelisted log). P3 follow-ups registered for line 150: invalid port raises an uncaught ValueError at startup;
base_urlhost/port is readable in session provenance. Receipt: 完成回执/claude-coordination/144-契约与装配基座.md.