feat(plugins): support ephemeral turn context, transforms, state, and response hooks

[binsarjr]

Summary

• extend the existing plugin turn-context result with validated prefill_messages
• inject ephemeral user/assistant priming after the main system prompt and before conversation history
• add a generic read-only plugin summary provider for the existing chat Agent State panel
• restrict prefill roles, message count, and content length
• keep feature-specific behavior outside Evonic core

Prefill messages are request-scoped and are never persisted to SQLite or JSONL session history. Plugin summaries are display-only: they are not persisted into workflow state, do not gate tools, and are not injected into model context.

Verification

• focused turn-context and Agent State summary checks passed
• authenticated chat-state API smoke test passed with the external plugin
• external plugin suite passed: 11 tests
• affected Python modules compile successfully
• live Evonic install and restart verification passed

Existing CI baseline

The earlier Python job reported 2 failures and 2,035 passes. The same two unrelated failures were present on the preceding PR commit (2 failures and 2,034 passes):

• unit_tests/test_ssh_backend_binary_read.py::test_cat_file_bytes_reads_binary_data_through_sftp
• unit_tests/test_token_efficiency.py::test_artifact_instructions_keep_capabilities_with_lower_token_cost

Draft for maintainer review of the generic hook shape and placement.

[binsarjr]

Experimental consumer: Evonic Security Godmode

This draft PR provides generic plugin capabilities used by the external Evonic Security Godmode plugin:

• ephemeral system context and user/assistant prefill at the start of a turn;
• request-scoped preserve, append, and replace system-context modes;
• ordered transformation of only the newest provider-bound user text;
• read-only plugin summaries in Agent State → Plugin States;
• a pre-final response handler that can replace a response or request up to two ephemeral provider retries before chat output is emitted or persisted.

Evonic core contains no Hermes runtime dependency, Godmode prompts, provider catalog, or Godmode-specific policy. The external plugin adapts the Hermes Agent Security Godmode flow to native Evonic lifecycle, authorization, persistence, and UI state.

Branches for testing

Clean draft-PR integration branch:

git clone --branch feature/plugin-prefill-context https://github.com/binsarjr/evonic.git

Server-tested aggregate branch:

git clone --branch dev https://github.com/binsarjr/evonic.git

Plugin branch:

git clone --branch feature/hermes-compatible-flow https://github.com/binsarjr/evonic-security-godmode.git
cd evonic-security-godmode
./scripts/package.sh

Version 0.1.6 makes orchestration programmatic:

• no godmode_* actions are exposed as agent tools;
• discovery runs before the first authorized turn and after exact model changes;
• newest-user transformation runs before the provider request;
• final no-tool responses are scored before chat output;
• hard refusals receive up to two program-selected strategy/encoding retries;
• rejected attempts stay hidden and only the selected final response is persisted;
• an operator-only, default-off fallback can race up to ten other enabled Evonic models;
• preserve, append, and override remain provider-request scoped and never modify SYSTEM.md.

The plugin requires confirmed authorization, exact scope, approver, and future ISO-8601 expiry. Missing or expired authorization blocks discovery, context, transformation, retry, and racing. Agent State reports discovery, retry, score, strategy, and race receipts.

Packaged build: v0.1.6. This integration remains experimental and under active testing.