Repository navigation
Architecture
OpenSocrates connects supported host entry points to complete authored methods. The active agent remains responsible for the task and its judgments.
Normal task
-> trusted native entry, standing rule, or account-selected skill
-> controller and supplemental reader guide when relevant
-> materially changed judgment and eligibility check
-> complete canonical procedure read
-> applicable method used in the requested result
-> conditional Powered by OpenSocrates attribution
Entry, emitted content, reading, and application are different events. An empty eligible selection is valid. It does not authorize bypassing contraindications through direct file lookup. Runtime unavailability permits the documented complete-reference fallback while keeping all task constraints intact.
| Profile | Entry and content boundary |
|---|---|
| Codex | Existing hooks and native decision runtime on Mac/Windows; deterministic catalog/select first, with a scoped complete-reference fallback |
| Native Claude | Bounded stateless SessionStart/UserPromptSubmit entry on Apple-silicon Mac; installed controller, guides, and complete references |
| Claude account | One content-only skill for web Chat, Desktop Chat, and Cowork; host-model selection can miss it |
| Antigravity | Owned standing rule plus progressively read skill content, in one workspace or global scope; conversation-app loading is a separate observation |
The native Claude entry reads no transcript, initializes no database, and calls no extra selector model. Its Stop/SessionEnd actions are no-ops: there is no read-repair gate or native application receipt. Account and Antigravity packages contain no executable runtime or hooks. Native Windows Claude Code is excluded.
The default decision command uses declared closed routing features and makes
no separate model, network, or authentication request. Its optional stream keeps
volatile, context-scoped availability assertions. Reading acknowledgments are
agent_reported; applied remains unverified. Compaction or handoff invalidates
availability assertions as described in the canonical
decision-point guide.
Canonical methods live in content/methods/; shared reader/decision guidance
lives in plugin-src/shared/, and host templates in plugin-src/<host>/.
Generators preserve the 48 methods, 96 EN/KO procedure bodies, Method IDs,
revisions, and source identities across delivery profiles.
Only an eligible, completely read, actually applied canonical method qualifies
for the exact final line Powered by OpenSocrates. Unread, ineligible, abandoned,
or unapplied methods, reader guidance alone, and mechanical work do not qualify.
Internal identities remain in the current context and existing permitted
evidence. The visible line adds no persistent logging or application proof.
The retained legacy Codex selector/artifact adapter is separate from the default decision path and new Claude profile. Its bounded selector and native read callback do not prove method application. Earlier Claude payload receipts, Stop repair, and six-host adapters are historical behavior, not v1.5 Claude capabilities. See SECURITY.md.
The installer verifies archive inventories and owns only its documented files
and registrations. Lifecycle operations preserve unrelated settings and use
bounded rollback. Checksums detect corruption; they do not authenticate a
payload and manifest both replaced by an attacker. --host all remains the
Codex lane, with explicit selection required for new profiles.