C. Dynamic derived resources + track adoption #119

New issue

Closed

opened 2026-09-24 15:23:10 +00:00 by kilo · 1 comment

kilo commented

2026-09-24 15:23:10 +00:00

Member

Child C of #116 (Dynamic Execution Sources program).

Server derives/materializes executable resources from the ModelInventory so administrators never manually create one resource per new model. Resource stays EXACT; the source owns the lifecycle.

Behaviors

New model appears (e.g. gpt-6.1-sol): next discovery observes it → classified → if adoption policy allows, becomes routable WITHOUT manual resource creation.
Model disappears: never immediately erase historical identity; existing audit records/pins remain interpretable; resource becomes unavailable/stale/retired by an explicit deterministic rule; no existing pinned request silently rewritten.
Unknown model appears (e.g. some-new-unclassified-model): appears in inventory as discovered/unclassified/not routable; capabilities never guessed; never routed merely because the provider lists it.

Conservative adoption (NOT "latest model always wins")

States DISCOVERED → CLASSIFIED → ROUTABLE. A newly discovered model becomes ROUTABLE automatically only when: its track is known; naming/version structure safely understood; the track has an approved conservative capability floor; the runtime reports the required effort; source/auth/health gates pass; compatibility evidence permits the requested feature. Inheritance only as a deliberately approved TRACK FLOOR (never a copy of an older model version's full ratings); stronger ratings come later through explicit evidence.

Reasoning efforts

Runtime-reported; not duplicated as static provider facts when avoidable. Policy may require low/medium/high/xhigh/max/ultra, but routing may only choose an effort THIS discovered physical model currently advertises. A model/effort pair absent from source inventory fails closed.

GPT-6 migration

Reconcile: model-policy.json already has GPT-6 Astra references (D-033 gate: reference-only, not selector-eligible); model-catalog.json is still GPT-5.6-only; live runtime exposes GPT-6 Luna/Sol/Astra. Review and reconcile gpt-6-luna/sol/astra using current official evidence, the real runtime inventory, existing owner-approved workflow roles and conservative capability assumptions. No mechanical rename of GPT-5.6 entries; historical entries may remain while still available/usable. Current/default policy must no longer require obsolete manual GPT-5.6 configuration merely because it was the previous generation.

Catalog/rating changes require provenance, date/version, confidence, rationale and a human-reviewable diff.

Child C of #116 (Dynamic Execution Sources program). Server derives/materializes executable resources from the ModelInventory so administrators never manually create one resource per new model. Resource stays EXACT; the source owns the lifecycle. ## Behaviors - **New model appears** (e.g. `gpt-6.1-sol`): next discovery observes it → classified → if adoption policy allows, becomes routable WITHOUT manual resource creation. - **Model disappears**: never immediately erase historical identity; existing audit records/pins remain interpretable; resource becomes unavailable/stale/retired by an explicit deterministic rule; no existing pinned request silently rewritten. - **Unknown model appears** (e.g. `some-new-unclassified-model`): appears in inventory as `discovered`/`unclassified`/`not routable`; capabilities never guessed; never routed merely because the provider lists it. ## Conservative adoption (NOT "latest model always wins") States DISCOVERED → CLASSIFIED → ROUTABLE. A newly discovered model becomes ROUTABLE automatically only when: its track is known; naming/version structure safely understood; the track has an approved conservative capability floor; the runtime reports the required effort; source/auth/health gates pass; compatibility evidence permits the requested feature. Inheritance only as a deliberately approved TRACK FLOOR (never a copy of an older model version's full ratings); stronger ratings come later through explicit evidence. ## Reasoning efforts Runtime-reported; not duplicated as static provider facts when avoidable. Policy may require low/medium/high/xhigh/max/ultra, but routing may only choose an effort THIS discovered physical model currently advertises. A model/effort pair absent from source inventory fails closed. ## GPT-6 migration Reconcile: `model-policy.json` already has GPT-6 Astra references (D-033 gate: reference-only, not selector-eligible); `model-catalog.json` is still GPT-5.6-only; live runtime exposes GPT-6 Luna/Sol/Astra. Review and reconcile `gpt-6-luna/sol/astra` using current official evidence, the real runtime inventory, existing owner-approved workflow roles and conservative capability assumptions. No mechanical rename of GPT-5.6 entries; historical entries may remain while still available/usable. Current/default policy must no longer require obsolete manual GPT-5.6 configuration merely because it was the previous generation. Catalog/rating changes require provenance, date/version, confidence, rationale and a human-reviewable diff.

kilo referenced this issue

2026-09-24 19:14:20 +00:00

Dynamic Execution Sources, Model Discovery, Multi-Account & Security #116

adrian.tkacz referenced this issue from a commit

2026-09-24 19:53:19 +00:00

feat(server): execution sources, conservative track adoption and derived exact resources (issue #119)

kilo referenced this issue

2026-09-24 19:53:46 +00:00

Dynamic Execution Sources, Model Discovery, Multi-Account & Security (D-053, #116) #124

kilo commented

2026-09-26 17:53:23 +00:00

Author

Member

Closeout — implemented in merged PR #124 (merge commit a4370f0 into develop, head 9789c67; program #116 / D-053).

Child C delivered: conservative track adoption (DISCOVERED → CLASSIFIED → ROUTABLE) against the reviewed model-tracks.json floors; per-advertised-effort derived exact resources (<source_id>:<slug>:<effort>) so a resource can never bind an unadvertised variant; deterministic lifecycle (absence → unavailable, retirement after 3 consecutive authenticated misses; audit/pins never rewritten); Daybreak and unclassified models stay visible but never routable; capabilities inherit only the approved track floor, never copied ratings. GPT-6 Luna/Sol/Astra reconciled into model-catalog.json (catalog v4, dated live-runtime evidence; D-033 Astra gate exercised), ultra added to the effort vocabulary. Verified live: a real authenticated source auto-discovered and materialized GPT-6 targets with zero configuration changes (T0→T1 proven in e2e).

**Closeout — implemented in merged PR #124** (merge commit `a4370f0` into `develop`, head `9789c67`; program #116 / D-053). Child C delivered: conservative track adoption (DISCOVERED → CLASSIFIED → ROUTABLE) against the reviewed `model-tracks.json` floors; per-advertised-effort derived exact resources (`<source_id>:<slug>:<effort>`) so a resource can never bind an unadvertised variant; deterministic lifecycle (absence → unavailable, retirement after 3 consecutive authenticated misses; audit/pins never rewritten); Daybreak and unclassified models stay visible but never routable; capabilities inherit only the approved track floor, never copied ratings. GPT-6 Luna/Sol/Astra reconciled into `model-catalog.json` (catalog v4, dated live-runtime evidence; D-033 Astra gate exercised), `ultra` added to the effort vocabulary. Verified live: a real authenticated source auto-discovered and materialized GPT-6 targets with zero configuration changes (T0→T1 proven in e2e).

kilo closed this issue

2026-09-26 17:53:49 +00:00