v1.3.0

[1.3.0] - 2026-08-25

Added

• Chat portal at /chat: sign in with an API key, threads in a sidebar, plain conversation with web search, per-chat model and effort. Only portal-created chats are listed; admin marks agents "Available in the chat portal".
• Landing page at / linking the chat portal and the admin console.
• Usage broken down by API key: admin Usage table and by_key in GET /v1/usage.
• Admin Sessions page: tenant and source (portal / cli / api) filters; tenant names shown instead of ids across Keys, Sessions, Completions, and Usage.

Changed

• Sessions record their origin (portal, cli, api); session-level model/effort overrides are honoured by the runner.

Full Changelog: v1.2.1...v1.3.0

v1.2.1

[1.2.1] - 2026-08-25

Fixed

• Claude Code conversations are keyed by the client's own session id (metadata.user_id), so a full interactive chat stays in one Session; transcript-prefix matching remains the fallback for other clients.
• Claude Code's next-prompt suggestion side requests no longer hijack or advance a session; they run as stateless completions.
• If a client compacts or rewrites its history, the conversation resyncs into the same Session instead of spawning a new one.

Full Changelog: v1.2.0...v1.2.1

v1.2.0

[1.2.0] - 2026-08-25

Added

• Anthropic Messages API shim (POST /v1/messages, POST /v1/messages/count_tokens) so Claude Code runs against Aegis via ANTHROPIC_BASE_URL; x-api-key auth accepted.
• Agentic CLI conversations (requests carrying tools) are routed into Sessions: follow-up turns are recognized by transcript prefix and resume the same SDK session, sending only new turns. The client system prompt is recorded on the session.
• Delete revoked API keys from the dashboard and DELETE /admin/api/keys/{id}; usage history keeps its tenant attribution.
• Completions: searchable tenant filter, redesigned detail view with collapsible system prompt, tool calls, and markdown rendering.
• Compose host port override via AEGIS_PORT.

Changed

• OpenAI shim returns real tool_calls for requests with tools (previously ignored), reports failures as HTTP 502 / SSE error frames instead of assistant text, returns content: null for tool-only replies, and serves the full model catalog on /v1/models.
• Client system prompts are embedded in the prompt body instead of replacing the CLI system prompt.

Fixed

• Client disconnects mid-stream no longer poison the SQLite connection pool.
• Large client system prompts no longer fail with a spurious "out of extra usage" error.

Full Changelog: v1.1.0...v1.2.0

v1.1.0

Full Changelog: https://github.com/dhpradeep/aegis/commits/v1.1.0

v1.0.0

feat: initial release of Aegis v1.0.0