Project-scoped activity feed exposing real-time subagent events parsed from the Claude Code CLI NDJSON output. Only authenticated users with access to the project can subscribe. Events are persisted in the activities table and pushed via SSE.

• The activity feed is project-scoped, not conversation-scoped — the SSE stream persists across conversation switches in the workspace UI.
• NDJSON lines emitted by the CLI are mapped to activity rows according to four event types: TASK_STARTED, TOOL_USE, AGENT_MESSAGE, TASK_COMPLETED.
• An in-memory map Map<String toolUseId, Long activityId> is maintained per conversation for the lifetime of the CLI process. Events whose parent_tool_use_id is not found in the map are logged as WARNING and skipped.
• Top-level PM text messages (type=assistant, parent_tool_use_id=null) are not written to activities — they belong to the conversation messages stream (UC-04003).
• Historical activities can be retrieved via REST pagination using the GET /activities endpoint with cursor-based afterId parameter.
• The SSE stream pushes only new events as they arrive; it does not replay history.
• System SSE events connected and heartbeat (every 30 s) keep the connection alive.

Agent Role Mapping

Database — activities Table

<!-- changes/0008-create-activities-table.xml -->
<?xml version="1.0" encoding="UTF-8"?>
<databaseChangeLog
        xmlns="http://www.liquibase.org/xml/ns/dbchangelog"
        xmlns:xsi="http://www.w3.org/2001/XMLSchema-instance"
        xsi:schemaLocation="http://www.liquibase.org/xml/ns/dbchangelog
        http://www.liquibase.org/xml/ns/dbchangelog/dbchangelog-latest.xsd">

    <changeSet id="0008-create-activities-table" author="webapp-creator">
        <createTable tableName="activities">
            <column name="id" type="bigint" autoIncrement="true">
                <constraints primaryKey="true" nullable="false"/>
            </column>
            <column name="project_id" type="bigint">
                <constraints nullable="false" foreignKeyName="fk_activities_project" references="projects(id)"/>
            </column>
            <column name="conversation_id" type="bigint">
                <constraints nullable="false" foreignKeyName="fk_activities_conversation" references="conversations(id)"/>
            </column>
            <column name="agent_role" type="varchar(64)">
                <constraints nullable="false"/>
            </column>
            <column name="event_type" type="varchar(32)">
                <constraints nullable="false"/>
            </column>
            <column name="parent_activity_id" type="bigint">
                <constraints nullable="true" foreignKeyName="fk_activities_parent" references="activities(id)"/>
            </column>
            <column name="description" type="varchar(255)">
                <constraints nullable="true"/>
            </column>
            <column name="tool_name" type="varchar(64)">
                <constraints nullable="true"/>
            </column>
            <column name="tool_summary" type="text">
                <constraints nullable="true"/>
            </column>
            <column name="payload_json" type="jsonb">
                <constraints nullable="true"/>
            </column>
            <column name="created_at" type="timestamptz" defaultValueComputed="now()">
                <constraints nullable="false"/>
            </column>
        </createTable>

        <createIndex tableName="activities" indexName="idx_activities_project_created_at">
            <column name="project_id"/>
            <column name="created_at" descending="true"/>
        </createIndex>

        <createIndex tableName="activities" indexName="idx_activities_project_conversation_created_at">
            <column name="project_id"/>
            <column name="conversation_id"/>
            <column name="created_at" descending="true"/>
        </createIndex>
    </changeSet>

</databaseChangeLog>

Columns:

NDJSON → Activity Parsing Rules

Rules applied by ClaudeCliService when processing each NDJSON line:

Note — real Claude CLI format: In stream-json output from Claude CLI, the delegation tool name is "Agent" (not "Task"). The "Task" variant is kept as a fallback for older builds. Additionally, subagent_type carries a "plugin:" prefix (e.g. "plugin:talkide-analyst"). The parser strips this prefix with removePrefix("plugin:") before resolving AgentRole, so the stored agent_role value is always the clean key (e.g. "talkide-analyst").

tool_use IDs are NOT UUIDs. Claude CLI emits Anthropic-format strings such as toolu_015oxKeDDmwYFYuTPoRVL8FG — never UUID-formatted IDs. The parser accepts any non-blank string as a valid tool_use_id and is intentionally prefix-agnostic (if Anthropic changes from toolu_ to another prefix, the parser keeps working without changes).

Async task completion: When an Agent tool is invoked with run_in_background: true, the immediate tool_result carries tool_use_result.isAsync == true. The parser detects this flag and defers TASK_COMPLETED — no activity row is written at that point. The actual completion is read from the subsequent system event with subtype: task_notification and status: completed, matched back to the original delegation by tool_use_id. The payload_json stored for the task_notification event is narrowed to a whitelist (task_id, status, summary, output_file, tool_use_id) to avoid persisting Anthropic-added fields that may appear in future CLI versions.

In-memory map lifecycle: Map<String toolUseId, Long activityId> is held in memory per conversation for the lifetime of the CLI process. It is lost on BE restart. Nested events whose parent_tool_use_id is not found in the map are logged as WARNING and discarded.

REST — Paginated History

sequenceDiagram
    actor User

    User->>+FE: opens workspace activity panel

    FE->>+BE: GET /api/v1/projects/{projectId}/activities?limit=100 <br> Authorization: Bearer {accessToken}

    BE->>BE: validate access token
    alt token missing or invalid
        BE-->>FE: 401 Unauthorized <br> ErrorResponse
    end

    BE->>DB: load project by projectId
    alt project not found
        BE-->>FE: 404 Not Found <br> ErrorResponse
    end

    BE->>BE: check project belongs to user's tenant
    alt project does not belong to tenant
        BE-->>FE: 403 Forbidden <br> ErrorResponse
    end

    BE->>DB: query activities WHERE project_id = ? ORDER BY id DESC LIMIT ?

    BE->>-FE: 200 OK <br> ActivityResponse[]

    FE->>-User: render activity feed history

GET /api/v1/projects/{projectId}/activities?limit=100&afterId=12345

Query parameters:

• limit (optional, default 100, min 1, max 500) — number of rows to return
• afterId (optional) — cursor pagination; returns only activities with id < afterId, ordered by id DESC

200 OK ActivityResponse[]:

[
  {
    "id": 12346,
    "projectId": 7,
    "conversationId": 42,
    "agentRole": "talkide-frontend-dev",
    "eventType": "TASK_COMPLETED",
    "parentActivityId": null,
    "description": "Pozadí aplikace",
    "toolName": null,
    "toolCategory": null,
    "toolSummary": null,
    "createdAt": "2026-04-30T11:24:30Z"
  },
  {
    "id": 12345,
    "projectId": 7,
    "conversationId": 42,
    "agentRole": "talkide-frontend-dev",
    "eventType": "TASK_STARTED",
    "parentActivityId": null,
    "description": "Pozadí aplikace",
    "toolName": null,
    "toolCategory": null,
    "toolSummary": null,
    "createdAt": "2026-04-30T11:24:00Z"
  }
]

payloadJson is intentionally excluded from the REST/SSE response — it is an internal detail. A future UC may expose an activity detail endpoint.

SSE — Live Stream

sequenceDiagram
    actor User

    User->>+FE: workspace view opened (project loaded)

    FE->>+BE: GET /api/v1/projects/{projectId}/activities/stream <br> Authorization: Bearer {accessToken}

    BE->>BE: validate access token
    alt token missing or invalid
        BE-->>FE: 401 Unauthorized <br> ErrorResponse
    end

    BE->>DB: load project by projectId
    alt project not found
        BE-->>FE: 404 Not Found <br> ErrorResponse
    end

    BE->>BE: check project belongs to user's tenant
    alt project does not belong to tenant
        BE-->>FE: 403 Forbidden <br> ErrorResponse
    end

    BE-->>FE: SSE event: connected

    loop CLI outputs NDJSON lines (any conversation in project)
        BE->>BE: parse NDJSON line per parsing rules above
        alt TASK_STARTED detected
            BE->>DB: insert activity (event_type=TASK_STARTED)
            BE-->>FE: SSE event: activity <br> ActivityResponse
        else TOOL_USE detected
            BE->>DB: insert activity (event_type=TOOL_USE)
            BE-->>FE: SSE event: activity <br> ActivityResponse
        else AGENT_MESSAGE detected
            BE->>DB: insert activity (event_type=AGENT_MESSAGE)
            BE-->>FE: SSE event: activity <br> ActivityResponse
        else TASK_COMPLETED detected
            BE->>DB: insert activity (event_type=TASK_COMPLETED)
            BE-->>FE: SSE event: activity <br> ActivityResponse
        end
    end

    loop every 30 seconds
        BE-->>FE: SSE event: heartbeat
    end

    FE->>-User: activity feed updated in real time

GET /api/v1/projects/{projectId}/activities/stream

Authorization: Bearer token. Response is text/event-stream (SSE). The stream is project-scoped and survives conversation switches.

SSE event: connected — stream established:

{}

SSE event: activity — a new activity row was inserted:

{
  "id": 12345,
  "projectId": 7,
  "conversationId": 42,
  "agentRole": "talkide-frontend-dev",
  "eventType": "TASK_STARTED",
  "parentActivityId": null,
  "description": "Pozadí aplikace",
  "toolName": null,
  "toolCategory": null,
  "toolSummary": null,
  "createdAt": "2026-04-30T11:24:00Z"
}

SSE event: heartbeat — keep-alive pulse every 30 s:

{}

For the full SSE event-name convention (connected, activity, heartbeat), auth requirements, and reconnect strategy, see architecture.md — SSE Convention.

401 Unauthorized (missing or invalid access token) ErrorResponse:

{
  "status": 401,
  "code": "AUTHENTICATION_FAILED",
  "message": "Access token is missing or invalid"
}

403 Forbidden (project does not belong to user’s tenant) ErrorResponse:

{
  "status": 403,
  "code": "FORBIDDEN",
  "message": "You do not have access to this project"
}

404 Not Found (project not found) ErrorResponse:

{
  "status": 404,
  "code": "NOT_FOUND_PROJECT",
  "message": "Project not found"
}

UX Guidelines

Context

Activity Panel je pravý sloupec workspace — tab “Activity” v přepínači Preview / Activity / Files (viz WorkspaceScreen.vue). Feed je project-scoped a přežívá přepnutí konverzace. Panel je po refactoru ActivityPanel.vue řízen reálnými daty z REST + SSE; momentálně obsahuje 7 hardcoded záznamů, které budou nahrazeny.

User Flow

1. Uživatel otevře workspace — FE načte historii (GET /activities?limit=100) a zobrazí skeleton.
2. FE současně otevře SSE stream (GET /activities/stream).
3. Panel zobrazí historické záznamy (nejnovější nahoře).
4. Nové SSE eventy se připínají nahoru s animací.
5. Uživatel vidí TASK_STARTED a TASK_COMPLETED — TOOL_USE a AGENT_MESSAGE jsou skryté.
6. Uživatel klikne na ”▶” vedle TASK_STARTED — řádek se rozbalí a zobrazí vnořené eventy.
7. Uživatel klikne znovu (nebo na ”▼”) — sbalí se zpět.
8. (Volitelně) Uživatel klikne na chip agenta v team stripu → filtr feedu na daného agenta.

Layout

┌─────────────────────────────────────────────────────┐
│  LIVE ACTIVITY · PROJECT-SCOPED          [Vše ▾]   │  ← header + (future: agent filter)
├─────────────────────────────────────────────────────┤
│  ● [Eli]  začala · Pozadí aplikace        2 min  ▶ │  ← TASK_STARTED, running dot
│    ┌── [Read]  /src/App.vue               2 min    │  │
│    ├── [Write] /src/main.ts               2 min    │  │ expanded
│    └── [msg]   "Upravuji soubor..."       2 min    │  ↓
│  ✓ [Eli]  dokončila · Pozadí aplikace     5 min    │  ← TASK_COMPLETED, ztlumené
│  ● [Mara] začala · Plánuje aplikaci      10 min  ▶ │  ← TASK_STARTED (Mara = amber dot)
├─────────────────────────────────────────────────────┤
│  (empty state nebo skeleton)                        │
└─────────────────────────────────────────────────────┘

• Screen type: live feed list (pravý panel workspace)
• Responsive: panel má fixní šířku danou layoutem workspace; vnitřně fluid
• Overflow: overflow-y: auto — scrolluje uvnitř panelu; panel nezkresluje vnější layout
• Padding: p-6 (konzistentní s aktuálním ActivityPanel.vue)

Feed Item — Default Layout (TASK_STARTED a TASK_COMPLETED)

Každý viditelný řádek (jeden div s flex items-center gap-3 px-3 py-2.5 rounded-lg):

[ status dot ] [ avatar 20px ] [ agent name ] [ akce ] [ description ]    [ čas ]  [ expand ▶ ]

Vizuální odlišení event types:

Agent colors (mapování agentRole → barva pro dot a avatar):

Expanded State (TOOL_USE a AGENT_MESSAGE)

Po kliknutí na ▶ u TASK_STARTED se zobrazí všechny záznamy se stejným parentActivityId. Vnořené položky jsou vizuálně odsazeny:

  ● [Eli]  začala · Pozadí aplikace         2 min  ▼
    ├── 📄 Read   /src/App.vue               2 min
    ├── ✏️  Edit   /src/main.ts               2 min
    ├── 📝 Write  /src/style.css             2 min
    └── 💬        "Upravuji soubory..."      2 min

Layout vnořeného řádku (ml-8 flex items-center gap-2 px-3 py-1.5):

AGENT_MESSAGE: tool_name je null, zobrazí se MessageSquare ikona + toolSummary truncated na 1 řádek. Klik na řádek rozbalí plný text (tooltip nebo inline expand — detail pro devs).

Animace rozbalení: max-height transition 200 ms ease-out (nebo Vue <Transition name="expand">).

Server-side deduplication (first page only)

The BE pre-deduplicates the activity feed only on the first page (i.e. requests without an afterId cursor). Consecutive activities sharing the same (tool_category, agent_role, parent_activity_id) key are coalesced into a single representative row using a window-function LAG()-based “gaps and islands” query. This prevents the UI from being flooded by, e.g., 20 consecutive Read tool calls from the same agent.

toolCategory values and their meaning:

toolCategory is set by RecordActivityUseCase.recordToolUse at insert time (mapping from tool_name). Raw tool_name is preserved in the database for forensics.

Empty State

Zobrazí se pokud activities pole je prázdné (žádná data z REST ani SSE):

┌─────────────────────────────────────────────────────┐
│                                                     │
│              [Activity ikona 32px]                  │
│         Tým zatím nic neudělal                      │
│   Pošli zprávu Maře a tým se pustí do práce        │
│                                                     │
└─────────────────────────────────────────────────────┘

• Kontejner: flex-1 flex flex-col items-center justify-center gap-3
• Ikona: Activity (lucide) 32px, var(--fg-4)
• Nadpis: text-sm font-medium text-[var(--fg-2)]
• Podnápis: text-sm text-[var(--fg-3)]

Loading State (Skeleton)

Zobrazí se při prvním načítání (REST volání běží, žádná data ještě):

• 5 skeleton řádků: animate-pulse, výška 38px, rounded-lg bg-[var(--bg-2)]
• Šířky střídavě 70 % / 85 % / 60 % / 90 % / 75 % pro přirozený výgled
• Header “LIVE ACTIVITY” je viditelný i během skeleton stavu

Live Update Animation

Nový SSE event activity → prepend na začátek seznamu:

1. Nový řádek se vloží na pozici 0 s opacity-0 translate-y-[-8px]
2. V dalším frame přejde na opacity-100 translate-y-0 — transition 250 ms ease-out
3. Ostatní řádky se posunou plynule dolů (layout shift, ne transform)

Pokud je uživatel odscrollován dolů (historické záznamy) — nový event se přidá nahoru bez přeskrolování. Badge s počtem nových (”↑ 3 nové”) se zobrazí nahoře — klik přeskroluje na top. (Toto je doporučení pro MVP, implementace na uvážení vývojáře.)

Filtrace — Future Enhancement

Filtrace per agent a per conversation je odložena na post-MVP. Důvody: v MVP bude feed relativně krátký, project-scope pokrývá všechny konverzace, přidání filtrů zvyšuje komplexitu bez okamžité hodnoty.

Navrhovaná budoucí implementace:

• Řada tlačítek pod headerem: Vše | Mara | Iris | Theo | Eli | Nia | Kai
• Klik na chip v team stripu (UC-05002) = shortcut pro filtr agenta
• Filtr per conversation: dropdown Aktuální konverzace | Všechny konverzace

Accessibility

• Každý feed item má role="listitem", obalující <ul> má role="list" aria-label="Activity feed"
• Expand button: aria-expanded="true|false" + aria-controls="activity-children-{id}"
• Vnořená sekce: id="activity-children-{id}" (pro aria-controls)
• Pulsující dot: aria-hidden="true" (dekorativní)
• Relativní čas: <time :datetime="createdAt">2 min</time> s ISO atributem
• Fokus na expand button se nemění po sbalení/rozbalení

Error States

i18n Keys

Poznámka k překladu action.started / action.completed: v češtině závisí tvar slovesa na pohlaví. Zatím fixní tvar dle talkide-frontend-dev = Eli (ženský rod). Pokud bude API vracet gender field nebo budou všichni agenti ženského rodu, překlady zůstanou správné. Jinak řešit pluralizací nebo gender-neutral formulací jako “spustila” → “spuštěno” — rozhodnutí pro frontend developera.

Frontend

Validations

Backend

Validations

Test Cases