docs: the routing rulings, the credential store, the report inbox and the two-session run - #658
Merged
Merged
Conversation
…gements get a decision adapter measured in a lab adr-2609221009491186: a provider adapter serves only the models it lists, under a vendor denylist no listing overrides; everything else runs on the host. The OpenAI-compatible API adapter (OpenRouter as configuration) is written and planned with the allowlist, the denylist, the named key, the run record and a one-time setup at ahoy that offers three homes for the key and never touches the harness. A decision adapter for the closed-option judgements is planned: one interface with the host's own judgement, shadow mode as an abcd lab, a research note as the evidence a ruling turns a judgement type on with, Jev the first candidate. The model tier takes the escalation rule (one tier up on a failed fix round) and the allowlist; itd-17's learned router is superseded. The research pass is filed as a note. Assisted-by: Claude:claude-opus-5
…ne for the routing rulings Assisted-by: Claude:claude-opus-5
…configured fallback recorded as intel, and itd-22 widened to no harness Assisted-by: Claude:claude-opus-5
…tomorrow Both interviewed with the product thinker on 2026-09-22 and planned READY: the inbox in the user account with identity in the inbox and a fingerprint in the record, and the second session with all three division modes measured a window at a time under bounds that keep it from stopping or overloading the run. Assisted-by: Claude:claude-opus-5
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
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Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be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.
What
The product thinker's routing interview of 2026-09-22, after two independent research passes and an ideate gauntlet:
adr-2609221009491186: a provider adapter serves only the models it lists, under a vendor denylist no listing overrides; everything else runs on the host. adr-2609221017021499: every external credential abcd holds is named in configuration and kept in one store the person chose.
Planned and READY: the OpenAI-compatible API adapter with OpenRouter as configuration, the allowlist, the named key, the run record and a one-time
ahoywalkthrough (itd-2609081951381895); the credential store with three homes and one reader (itd-2609221017023290); a decision adapter for abcd's typed judgements that runs in shadow as anabcd laband whose research note is the evidence a ruling turns a judgement type on with (itd-2609221009495079); the command-line runner, with a configured fallback host and every fall back recorded as intel (itd-2609201916056194).Amended: the model tier takes the escalation rule (one tier up on a failed fix round) and the allowlist; itd-22 is widened so "any harness" includes none; itd-17's learned router is superseded.
Recorded, not adopted: the control-programme idea went through the ideate gauntlet and came out reframed — an operator console served by the binary over local web, with a product-thinker pane in their own register, sequenced after the runner. Two kill attempts were fatal to the idea as put.
Two more, interviewed the same evening for tomorrow's run: a managed repository reports back to abcd through an inbox in the user account, with identity in the inbox and a fingerprint in anything committed (itd-2609221656361680); and two orchestrator sessions share one run, all three division modes measured a window at a time under bounds that keep the second from stopping or overloading the first (itd-2609221656373558).
Why
Jev (TypeSafe AI, on OpenRouter) is a typed-decision model, not a router: it fits abcd's closed-option judgements and nothing generative. Reaching it means reaching an aggregator that also serves the frontier models a person already pays for through their host, so the allowlist is the condition of the integration, not a refinement of it. The same walkthrough is the pattern for every credential abcd will hold, which is why the store is its own record rather than the adapter's private habit.
Review
Three independent passes fed this: a state-of-the-art pass on Jev and model routing (filed as a research note with its sources), a primary-source pass on headless harness invocation, and an adversarial review of the console idea by an evaluator that saw the artefact without its provenance. Every record went through the readiness gate; the record, reviews-charter, decisions-append and docs gates and the full preflight pass.
A note on this branch
It replaces an identical branch whose secret scan failed on a false positive: a spec's footprint line read
internal/core/oracle, internal/adapter/openaiapi, internal/surface/cli, and the generic-api-key rule read the package name plus the next token as a key and its value. No credential was involved. Because the scan reads history and force-pushing is not done here, the work was rebuilt commit by commit offmainwith the two footprint lines reworded; this branch scans clean.Records
Assisted-by: Claude:claude-opus-5