You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
WS1 — The pack interface: one manifest per domain, score as a field of a pack
Important
Problem — Domain knowledge today loads through the score machinery alone: a score's stages compile into the runtime instructions, and there is no declared unit bundling skills + templates + corrector + catalog + eval corpus. The autoresearch reframe (parent #368) needs a single harness-agnostic seam that every domain pack implements.
Approach — Introduce the pack manifest as the unit of generality, owned by the schema package with one validator consumed by the extension, the CLI, and CI. The loader/compiler treats the onboarding score as one field of a pack. No physics content moves in this slice — the pulse repertoire is re-expressed as a pack manifest, byte-identical in output.
Scope — in: manifest schema, validator, loader refactor, golden compiled-output parity · out: the instruction rewrite (WS2), new packs (WS3), runtime pack switching (follow-up).
Acceptance Criteria
A pack manifest schema exists (skills, template registry, corrector with integrity manifest, catalog schema, onboarding score, eval corpus pointer) with one shared validator
The extension loads the existing pulse-designer repertoire through a pack manifest with byte-identical compiled output against today's golden
A second pack can be registered by manifest alone — no loader code change
Boot-time selection behaves exactly as today for the default pack
Testing Decisions
Extend the existing score-loader unit tests with manifest fixtures; add a schema round-trip in the schema package's suite; add a golden compiled-output parity test (new loader vs current output). This is the TDD spec — the parity test goes red first.
Key Decisions
A score is one field of a pack, not a peer of it.
The manifest names the corrector and its integrity manifest explicitly — the threshold condition is a load-time property, not a convention.
Boot-time selection stays in v1; router-time recompilation is explicitly deferred.
Constraints & Invariants
No file outside the loader/compiler seam changes behavior; the corrector paths a manifest names must live outside agent-editable trees.
Prior Art
The score loader/compiler and entitlement filter (extension session-prep); the schema package's single-validator pattern.
Source
Part of #368 · design-of-record: vault spec note spec-20260813-162300-autoresearch-studio
WS1 — The pack interface: one manifest per domain, score as a field of a pack
Important
Problem — Domain knowledge today loads through the score machinery alone: a score's stages compile into the runtime instructions, and there is no declared unit bundling skills + templates + corrector + catalog + eval corpus. The autoresearch reframe (parent #368) needs a single harness-agnostic seam that every domain pack implements.
Approach — Introduce the pack manifest as the unit of generality, owned by the schema package with one validator consumed by the extension, the CLI, and CI. The loader/compiler treats the onboarding score as one field of a pack. No physics content moves in this slice — the pulse repertoire is re-expressed as a pack manifest, byte-identical in output.
Scope — in: manifest schema, validator, loader refactor, golden compiled-output parity · out: the instruction rewrite (WS2), new packs (WS3), runtime pack switching (follow-up).
Acceptance Criteria
Testing Decisions
Extend the existing score-loader unit tests with manifest fixtures; add a schema round-trip in the schema package's suite; add a golden compiled-output parity test (new loader vs current output). This is the TDD spec — the parity test goes red first.
Key Decisions
Constraints & Invariants
Prior Art
Source
Part of #368 · design-of-record: vault spec note
spec-20260813-162300-autoresearch-studio