Client adapter contract

Project Cognition and Goal Governor are portable cores. A client adapter is the authority boundary that turns native client events into project-cognition/host-event/v1 and enforces the resulting decision. Sharing prompts alone is not conformance.

Stable invocation surface

Host integrations preserve two equivalent user entry styles:

await researcher.ask({ question: 'What is this project trying to prove?' })
await researcher.mode.set({ state: 'on', question: 'Watch for architecture drift' })
researcher.mode.get()
await researcher.mode.set({ state: 'off' })

• ask is one-shot and ends at the next host-observed turn_end.
• mode on persists across turns and resume until a host-authenticated user switches it off.
• model-authored or model-asserted control events cannot switch the mode.
• adapters may expose native command syntax in addition to this control API. DSH maps /researcher <question> and /researcher on|off to the same reducer.

HostEvent v1

Every normalized event carries seq, session_id, optional runtime_goal_id, a native evidence reference, actor, source, identity assurance, kind, and data. Frozen kinds are:

user_action, tool_call, tool_result, goal_transition, usage, turn_end, session_resume, guard_violation.

The host must retain the native event stream. HostEvent is a deterministic projection used by portable reducers; it is not permission to discard native evidence or trust assistant prose.

The package-root adapterCore export exposes this base envelope and the Researcher control reducer as an experimental additive API. It freezes the ask and mode.set/get user semantics, not a governed event profile or a compatibility claim. Typed verifier receipts, usage completeness, atomic sequencing, raw-first storage, and enforcement receipts remain requirements for the later E2-gated adapter API.

Conformance boundary

An adapter manifest v2 reports two different facts:

• capability declaration: the adapter claims to provide human approval, hard stop, event storage, trusted verification, and project-root confinement;
• conformance status: PENDING, PASS, or FAIL, backed by independently scoreable evidence.

Only conformance=PASS plus all five capabilities produces governed=true. The DSH adapter is currently PENDING until Live E1 completes. Claude Code is the frozen second-adapter target, but implementation is gated on E2 PASS; Codex, Zcode, and OpenClaw remain design candidates rather than compatibility claims.

Discovery is not conformance

Version-locked discovery records live under evaluation/adapter-discovery/ and may contain official-source mappings, redacted no-model traces and replay designs. Their only allowed results are DISCOVERY_QUALIFIED, HOLD and NO_GO. They never update an adapter manifest or create an installation path.

The current Claude Code Agent SDK 0.3.251 and Codex App Server stdio 0.150.0-alpha.12.2 records are both HOLD: Codex exposes a promising thread/turn/item and approval surface, but raw durable command replay, principal assurance and write enforcement are unproven; its complete schema bundle and method inventory are reproducible under a credential-stripped capture. Claude's locked SDK module and bundled CLI load under a credential-stripped no-session capture, but no authentic query/tool/resume trace exists. Run npm run adapter:discovery:check to verify these boundaries offline.

The derived evaluation/adapter-discovery/host-event-convergence-v1.json currently confirms only that both locked discovery mappings name the same seven candidate HostEvent kinds and target the same 28 normalized binding fields. The current locked documents mark 25 fields for each client as DOCUMENTED and three for each as explicit GAP; none is OBSERVED. Every documented field must reference a reproducible proof locator bound to the exact Claude type hash or Codex generated-schema tree and bounded source-file hashes. These counts describe the reviewed interface documents, not runtime reliability or a client ranking. The checker binds the source mapping bytes and keeps guard_violation, authenticated human receipts, usage completeness, raw-first durability, restart checkpoints and terminal write enforcement as unresolved boundaries. This comparison cannot promote either client or substitute for E2.

Field presence is not event integrity. The locked event-cohesion audit finds that Claude can assemble five mappings from one native frame, but its approval path is unjoined: PermissionRequestHookInput has session/prompt context and no toolUseID or requestId, while canUseTool has the call/request IDs and decision without session_id or prompt_id. Its interrupt-return-to-Stop join also depends on host context rather than a shared native receipt ID. Codex has five single-frame mappings plus two native-key joins through JSON-RPC and thread/turn identity. This makes Codex's discovery surface structurally more cohesive for a future adapter, but still does not establish principal identity, usage completeness, durable replay, hard-stop enforcement, conformance, or outcome value.

The discovery checker now executes one locked synthetic semantic fixture per client. The fixtures are host-authored native-shape data with zero model and network calls, not captured client output. They verify deterministic projection and fail-closed behavior for duplicate IDs, missing bindings, and incorrect terminal joins. Claude produces six candidate projections plus one unresolved approval path; Codex produces eight candidate projections with no unresolved fixture event. Those counts describe only the chosen synthetic examples and do not upgrade either discovery from HOLD or create a product adapter.

Each fixture also has one hash-locked synthetic restart boundary. Full replay must preserve every completed prefix projection and distinguish a genuinely resolved pending join from a retained or merely changed unresolved condition. This catches checkpoint drift, event reordering and missing terminal suffixes, but it is not raw-first persistence or a native process-exit/resume trace. Accordingly resume_prefix_checkpoint remains a convergence GAP.