Files
correx/docs/epics/epic-13-resolution.md
T

251 lines
11 KiB
Markdown

# Epic 13 — Interfaces (TUI + CLI + API)
## completed deliverables
### task 0 — orchestrator resume mechanism
fixed the approval gate in `SessionOrchestrator` to truly suspend and resume rather than immediately failing.
changes:
- `handleApproval` is now `suspend` — holds a `CompletableDeferred<ApprovalDecision>` per pending request, keyed by `ApprovalRequestId`
- emits `ApprovalRequestedEvent` before suspending so the server layer can surface the request id to clients
- `DefaultSessionOrchestrator.submitApprovalDecision(requestId, decision)` completes the deferred and resumes the workflow coroutine
- `ReplayOrchestrator.mapValidationOutcome` updated to `suspend` to match the parent signature
- `approvalEngine` removed from `SessionOrchestrator` (was evaluated synchronously; now decisions arrive externally)
files changed:
- `core/kernel/.../orchestration/SessionOrchestrator.kt`
- `core/kernel/.../orchestration/DefaultSessionOrchestrator.kt`
- `core/kernel/.../orchestration/ReplayOrchestrator.kt`
---
### task 1 — websocket protocol
defined the wire protocol as sealed class hierarchies with kotlinx.serialization in `apps/server/protocol/`.
final structures:
* `ServerMessage` — 16 subtypes across session lifecycle, stage execution, inference, tool execution, approval, provider status, and error
* `ClientMessage` — 5 subtypes: StartSession, ResumeSession, CancelSession, ApprovalResponse, Ping
* `ApprovalDecision` (protocol enum) — `APPROVE`, `REJECT`, `STEER` — separate from the domain `ApprovalDecision` data class
* `RiskSummaryDto`, `ProviderHealthDto`, `SessionConfigDto`, `PauseReason` — protocol-only DTOs
* `ProtocolSerializer` — encodes `ServerMessage` to JSON, decodes `ClientMessage` from JSON; unknown/malformed input throws `ProtocolException` (never crashes)
key properties:
- `Json { classDiscriminator = "type"; ignoreUnknownKeys = true }`
- no domain types leak into the protocol layer except pure value aliases (`SessionId`, `StageId`, `ApprovalRequestId`, `Tier`)
files created:
- `apps/server/src/main/kotlin/com/correx/apps/server/protocol/ServerMessage.kt`
- `apps/server/src/main/kotlin/com/correx/apps/server/protocol/ClientMessage.kt`
- `apps/server/src/main/kotlin/com/correx/apps/server/protocol/Dtos.kt`
- `apps/server/src/main/kotlin/com/correx/apps/server/protocol/ProtocolSerializer.kt`
---
### task 2 — ktor server + websocket
implemented the full Ktor 3.x server with REST and WebSocket endpoints.
rest endpoints:
- `GET /health` — provider status, 200 in degraded mode
- `GET/POST /sessions` — list and start sessions
- `GET /sessions/{id}` — single session state
- `POST /sessions/{id}/cancel`
- `GET /sessions/{id}/events?from=<eventId>&limit=50` — paginated event replay
- `GET /workflows` — available workflows
- `GET /providers` — provider health
websocket endpoints:
- `WS /sessions/{id}/stream` — session-scoped stream
- `WS /stream` — global stream across all sessions
websocket behavior:
- on connect: snapshot of current state, then live events
- on reconnect with `?lastEventId=<id>`: replays missed events from that point
- 30s server heartbeat; client must respond within 10s or connection is dropped
- unknown client message → `ProtocolError` sent, connection kept open
`ServerModule` holds all dependencies as interfaces; no concrete instantiation inside it.
files created:
- `apps/server/src/main/kotlin/com/correx/apps/server/Application.kt`
- `apps/server/src/main/kotlin/com/correx/apps/server/ServerModule.kt`
- `apps/server/src/main/kotlin/com/correx/apps/server/routes/SessionRoutes.kt`
- `apps/server/src/main/kotlin/com/correx/apps/server/routes/WorkflowRoutes.kt`
- `apps/server/src/main/kotlin/com/correx/apps/server/routes/ProviderRoutes.kt`
- `apps/server/src/main/kotlin/com/correx/apps/server/ws/SessionStreamHandler.kt`
- `apps/server/src/main/kotlin/com/correx/apps/server/ws/GlobalStreamHandler.kt`
- `apps/server/src/main/kotlin/com/correx/apps/server/registry/WorkflowRegistry.kt`
- `apps/server/src/main/kotlin/com/correx/apps/server/registry/ProviderRegistry.kt`
---
### task 3 — approval interaction
implemented the full approval lifecycle bridging the orchestrator suspension and the WebSocket clients.
`ApprovalCoordinator` responsibilities:
- tracks connected WS clients per session
- on `ApprovalRequestedEvent`: broadcasts `ServerMessage.ApprovalRequired` to all clients for that session
- schedules a timeout job (`coroutineScope.launch { delay(timeoutMs) }`) that auto-denies if no response within `ApprovalConfig.timeoutMs` (default 5 minutes)
- on `ClientMessage.ApprovalResponse`: cancels timeout job, translates protocol enum to domain `ApprovalDecision`, calls `orchestrator.submitApprovalDecision()`
- double-response guard via `ConcurrentHashMap<ApprovalRequestId, Boolean>` — second response returns `ProtocolError`, no double-resume
protocol → domain mapping:
- `APPROVE``ApprovalOutcome.APPROVED`
- `REJECT``ApprovalOutcome.REJECTED`
- `STEER``ApprovalOutcome.APPROVED` + `UserSteering(text = steeringNote, ...)`
rest fallback: `POST /sessions/{id}/approve` for non-WS clients.
files created:
- `apps/server/src/main/kotlin/com/correx/apps/server/approval/ApprovalCoordinator.kt`
- `apps/server/src/main/kotlin/com/correx/apps/server/approval/ApprovalConfig.kt`
- `apps/server/src/main/kotlin/com/correx/apps/server/routes/ApprovalRoutes.kt`
---
### task 4 — mosaic tui
implemented a new `apps/tui` Gradle module with a Mosaic-based interactive terminal UI.
architecture:
- single WS listener coroutine updates `mutableStateOf(TuiState)` — no polling
- separate stdin reader coroutine sends `KeyEvent` to a `Channel`
- key handler coroutine dispatches protocol messages to the server
- exponential-backoff reconnect (1s → 30s max) on server disconnect
- `applyServerMessage` pure reducer maps `ServerMessage` to state updates
layout: status bar / session list (last 5) / active session panel / approval panel (collapsed when no pending approval) / input bar
keybinds: `n` new, `c` cancel, `a/r/s` approve/reject/steer (only active when approval pending), `↑↓` navigate, `q` quit — inactive binds are no-ops
files created:
- `apps/tui/build.gradle`
- `apps/tui/src/main/kotlin/com/correx/apps/tui/TuiApp.kt`
- `apps/tui/src/main/kotlin/com/correx/apps/tui/StateReducer.kt`
- `apps/tui/src/main/kotlin/com/correx/apps/tui/KeyEvent.kt`
- `apps/tui/src/main/kotlin/com/correx/apps/tui/state/TuiState.kt`
- `apps/tui/src/main/kotlin/com/correx/apps/tui/ws/TuiWsClient.kt`
- `apps/tui/src/main/kotlin/com/correx/apps/tui/components/StatusBar.kt`
- `apps/tui/src/main/kotlin/com/correx/apps/tui/components/SessionList.kt`
- `apps/tui/src/main/kotlin/com/correx/apps/tui/components/ActiveSession.kt`
- `apps/tui/src/main/kotlin/com/correx/apps/tui/components/ApprovalPanel.kt`
- `apps/tui/src/main/kotlin/com/correx/apps/tui/components/InputBar.kt`
---
### task 5 — clikt cli
implemented the full `apps/cli` command suite using Clikt.
commands:
- `correx run --workflow <path> [--session <id>] [--auto-approve]`
- `correx session list/resume/cancel/events`
- `correx approve <session-id> --decision <approve|reject|steer> [--note]`
- `correx status`
- `correx provider list`
output modes: `--json` (valid parseable JSON), `--quiet` (errors only), default (human-readable with ANSI color).
`correx run` behavior:
- starts session via `POST /sessions`, subscribes to WS stream
- prints stage transitions and tool calls to stdout
- on `ApprovalRequired`: interactive TTY prompt if stdin is a terminal; auto-deny with warning if no TTY and `--auto-approve` absent
- exit codes: `0` complete, `1` failed, `2` approval denied
`CliWsClient` wraps the Ktor WebSocket client for stream subscriptions.
files created:
- `apps/cli/src/main/kotlin/com/correx/apps/cli/CorrexCli.kt`
- `apps/cli/src/main/kotlin/com/correx/apps/cli/CliConstants.kt`
- `apps/cli/src/main/kotlin/com/correx/apps/cli/ws/CliWsClient.kt`
- `apps/cli/src/main/kotlin/com/correx/apps/cli/commands/RunCommand.kt`
- `apps/cli/src/main/kotlin/com/correx/apps/cli/commands/SessionCommand.kt`
- `apps/cli/src/main/kotlin/com/correx/apps/cli/commands/ApproveCommand.kt`
- `apps/cli/src/main/kotlin/com/correx/apps/cli/commands/StatusCommand.kt`
- `apps/cli/src/main/kotlin/com/correx/apps/cli/commands/ProviderCommand.kt`
---
### task 6 — shared config
implemented configuration loading in `core/config`.
schema (`~/.config/correx/config.toml`, override via `$CORREX_CONFIG`):
```toml
[server]
host = "localhost"
port = 8080
[tui]
theme = "dark"
session_list_limit = 5
[cli]
default_output = "human"
[approval]
timeout_ms = 300000
```
`ConfigLoader.load()` — reads from file, returns `CorrexConfig()` defaults on missing file, logs warning and returns defaults on malformed input. no new library dependencies — manual TOML parser for the required subset (sections + key=value pairs).
files created:
- `core/config/src/main/kotlin/com/correx/core/config/CorrexConfig.kt`
- `core/config/src/main/kotlin/com/correx/core/config/ConfigLoader.kt`
---
# final architecture after Epic 13
```text
apps/
├── server ← Ktor 3.x, REST + WebSocket, approval coordinator
├── tui ← Mosaic reactive terminal (new module)
├── cli ← Clikt commands, WS stream client
└── shared ← (config lives in core/config)
core/
└── config ← CorrexConfig, ConfigLoader (new module)
```
---
# major architectural outcomes
Epic 13 established:
* real approval pause/resume — orchestrator now truly suspends and resumes externally
* a wire protocol separating transport from domain (no domain types in protocol layer)
* a Ktor server as the single point of access — TUI and CLI are pure clients
* WebSocket event streaming with snapshot-on-connect and missed-event replay
* approval timeout and double-response safety
* a Mosaic TUI with reactive state and context-aware keybinds
* a full Clikt CLI with machine-readable JSON output and correct exit codes
* shared config via TOML with graceful fallback to defaults
---
# what Epic 13 intentionally does NOT include
not implemented:
* router context isolation (explicitly deferred — see CLAUDE.md)
* streaming inference responses (explicitly deferred)
* parallel agent execution (explicitly deferred)
* GPU residency scheduling (explicitly deferred)
* TUI config editor (`,` keybind present, editor not implemented)
* `correx config --edit` command (not in server surface)
---
# final state
Correx now has:
> a complete interface layer — server, TUI, and CLI — all backed by a single Ktor WebSocket bus, with approval decisions flowing from any client through to the orchestrator's suspended coroutine and resuming execution cleanly.