11 KiB

Raw Blame History

Feishu Signal — Web Implementation Plan

Scope: Web only. No mobile. No Feishu card integration. Core flow only.

What gets built

AI pipeline — classify message as project-level or signal-level, route accordingly
Bell panel — decision queue for items needing review
Signal drawer — verify AI output, fix errors, pick next action
Project creation drawer — AI pre-fills new project form from Feishu message
SSTC signal source — display + link to charter

Routing logic (core addition)

Every Feishu message goes through a two-step classification:

Step 1 — Level classification (multi-factor, AI综合判断)

AI scores 4 dimensions, each 0–1, then computes weighted average:

Dimension	Weight	Signal of project-level
Build intent	40%	"我们要做", "需要开发", "立项" etc.
Scope	25%	Describes a system / product / feature set, not a single task
Timeline	20%	Explicit deadline or duration mentioned
Owner	15%	Named responsible person or team

project_score = weighted average of 4 dimensions
signal_score = 1 - project_score (not exact inverse, but starting point)

If message contains distinct project intent AND distinct signal content (e.g. "let's build X" + "competitor Y just launched Z"), AI splits into 2 separate items and processes each independently. If ambiguous whether to split, default to signal-level — user can promote later.

Step 2 — Routing

project_score ≥ 0.85, cold start over → auto-create project, silent notification to owner
project_score ≥ 0.85, cold start      → pending queue, card type = "New Project"
project_score 0.60–0.84               → pending queue, card type = "New Project?"
project_score < 0.60                  → treat as signal, route by signal confidence

signal confidence ≥ 0.80, cold start over → auto-enter signal source, no badge
signal confidence < 0.80 or cold start    → pending queue, card type = "Signal"

Cold start period = first 2 weeks after go-live. No auto-creation or auto-entry during this period regardless of confidence.

Phase 1 — AI Pipeline (Week 1)

Backend

On Feishu message received, trigger async AI job
AI job — Step 1: level classification
- level — project or signal
- level_confidence — float 0–1
- level_reasoning — 1–2 sentences why
- If message contains both project intent and signal content: split into 2 items, process each separately
AI job — Step 2 (if level = signal): signal fields
- title — 1 sentence
- type — one of: 竞品动态 / 用户反馈 / 项目风险 / 新需求
- priority — P0–P3
- matched_project_id — existing project this signal belongs to (null if new project)
- summary — 2–3 sentences
- reasoning — 3 bullets
- confidence — float 0–1
- recommended_action — one of: enter_only / create_charter / trigger_alert
AI job — Step 2 (if level = project): project pre-fill fields
- suggested_name — project name extracted from message
- suggested_description — 1–2 sentences
- suggested_owner — extracted from message or inferred
- suggested_timeline — extracted deadline or duration
- confidence — float 0–1
- reasoning — 3 bullets
Routing logic:
- Cold start (first 2 weeks): all items → pending queue, no auto-creation or auto-entry
- project_score ≥ 0.85, cold start over → auto-create project, silent in-app notification to owner, no badge
- project_score 0.60–0.84 → pending queue, bell badge +1, card type = "New Project?"
- project_score < 0.60 → treat as signal
- Signal: confidence ≥ 0.80, cold start over → auto-enter signal source, no badge
- Signal: confidence < 0.80 or cold start → pending queue, bell badge +1
Auto-create project flow: create project with AI pre-filled fields → initialize SSTC board → assign owner → silent in-app notification: "已自动创建项目 [name]，点击查看"
Auto-enter signal flow: create signal node → insert into matched project signal source → generate charter draft (not submitted) → silent in-app notification to project owner
Store: raw Feishu message + AI output + confidence + all edits + reasoning, bound to unique item node ID

DB

feishu_items table: id / feishu_message_id / project_score / signal_confidence / raw_message / split_from_item_id (nullable, for split items) / created_at
signals table: id / feishu_item_id / matched_project_id / title / type / priority / summary / reasoning / confidence / recommended_action / status / created_at
signal_edits table: signal_id / field / old_value / new_value / edited_by / edited_at
project_suggestions table: id / feishu_item_id / suggested_name / suggested_description / suggested_owner / suggested_timeline / confidence / reasoning / status / created_at
Add signal_id foreign key to charters table

Phase 2 — Bell Panel (Week 1)

Frontend

Bell panel opens as 360px dropdown from top-right bell icon.

Two sections:

Pending (default, expanded)

Mixed list of project-level and signal-level items, sorted by priority desc then time
Project card — medium confidence (0.60–0.84) (blue left border, user must confirm):
- Row 1: 🏗 "New Project?" badge + relative time
- Row 2: suggested project name (1 line)
- Row 3: AI reasoning in 1 line + confidence score
- Actions: Create Project (primary) · Open (ghost) · Dismiss (ghost)
- Create Project → opens Project Creation Drawer (pre-filled by AI)
- Open → opens Project Creation Drawer
Project card — auto-created (≥0.85, shown in auto-processed section):
- Read-only. Shows "Auto-created [project name]" + Undo option
Signal card (existing design):
- Row 1: priority badge + type + matched project name + relative time
- Row 2: signal title (1 line, truncated)
- Row 3: AI uncertainty reason
- Actions: Confirm (primary) · Open (ghost) · Dismiss (ghost)
- Confirm → enters signal into matched project's signal source, badge -1
- Open → opens Signal Drawer
Dismiss on either type → inline reason picker, then archives
Clicking project name pill on signal card → popover to switch matched project

Auto-processed (collapsed by default)

Summary line: "X signals auto-entered, Y projects auto-created today"
Expand shows read-only cards with Undo option per card
Undo auto-created project: project moves to pending queue as "New Project?" card

Badge rules:

Red number = pending count only
Clears per item on confirm/dismiss
No badge for auto-processed signals

Phase 3 — Project Creation Drawer (Week 2)

Slides in from right, 480px wide, full page height. Only opened for project-level items.

Content

Top sticky header:

🏗 New Project Detected
From: [Feishu group] · [sender] · [time]

AI pre-filled fields (all editable inline):

Project name (from suggested_name)
Description (from suggested_description)
Owner (from suggested_owner, dropdown to change)
Timeline (from suggested_timeline, date picker)

AI reasoning (collapsed by default)

Toggle: "Why AI thinks this is a new project ▾"
3 bullets, plain language

Original Feishu message (collapsed by default)

Fix section — if AI pre-fill is wrong:

Button	Opens
Not a new project	Converts to signal → opens Signal Drawer
Wrong project name	Inline text edit
Wrong owner	User dropdown
Wrong timeline	Date picker

Bottom bar (full width, fixed):

Create Project (primary) — creates project with pre-filled fields, redirects to new project's SSTC board
Save as Draft — saves without redirecting
Cancel — closes drawer, item stays in pending queue

Slides in from right, 480px wide, full page height. Same component opened from bell, SSTC board, charter page.

Tab 1 — AI Result (default)

Top sticky header:

[type pill]  [project pill]  [priority pill]
Signal title (editable inline)
From: [Feishu group] · [sender] · [time]

All 3 pills tappable → popover to change value, saves immediately.

Body:

AI Summary — 2–3 sentences, editable inline. Click to edit, click away to save.

Reasoning (collapsed by default)

Toggle: "Why AI classified it this way ▾"
3 bullet points, plain language
Exists to catch errors, not to show off

Original Feishu message (collapsed by default)

Toggle: "View original message ▾"
Raw message + 3 messages of context
"Open in Feishu ↗" link

Bottom bar:

Looks good → — goes to Tab 3
Something's wrong — goes to Tab 2

Tab 2 — Fix

6 preset buttons (covers ~95% of correction cases):

Button	Opens
Wrong project	Project dropdown
Wrong signal type	Type selector
Wrong priority	Priority selector
Summary is wrong	Text input with current summary pre-filled
Wrong recommended action	Action selector
Other	Free text input (fallback only)

After any selection → AI updates immediately → shows "Updated. Check Tab 1 to verify →"
No apology message from AI. Just the result.

Tab 3 — Next Action

Three radio options:

A. Enter signal source only (default)
Signal enters SSTC board signal source list. No charter created.

B. Enter + create charter draft
AI pre-fills: goal (from summary) / background (from core facts) / scope (inferred).
Charter preview slides in as a second layer on top of drawer. Confirm to save.

C. Enter + trigger risk alert
Creates risk alert, sends Feishu notification to project owner.

Confirm button (full width, fixed bottom): "Enter into [project name]"
Closes drawer on confirm, shows toast.

Phase 4 — SSTC Integration (Week 2)

Signal source list (existing page)

Each signal card shows: title · type badge · source (Feishu) · status
Status values: 待处理 / 已关联宪章 / 已关闭
Click any card → opens Signal Drawer (Tab 1)

Charter linkage

Charter detail page: top banner "来源信号: [title] ↗" → opens Signal Drawer read-only
Signal card status → "已关联宪章 [charter name]" after charter created
When charter goes active → signal status auto-updates to 已关闭

Phase 5 — Polish & Test (Week 3)

Cold start → mature period threshold transition logic (admin toggle in project settings)
Signal edit history view (in drawer, accessible via small "History" link)
Undo auto-enter from auto-processed section
Empty states for bell panel (no pending, no auto-processed)
Error states (AI job failed, Feishu connection lost)
End-to-end test A (signal flow): Feishu message → AI classifies as signal → bell badge → confirm → signal source → charter
End-to-end test B (project flow): Feishu message → AI classifies as project → bell badge → open drawer → create project → SSTC board
End-to-end test C (split): Feishu message with both intent → AI splits into 2 items → both appear in bell → handled independently

Out of scope (v1)

Mobile
Feishu interactive card integration
Feishu bot conversation
AI model retraining UI
Signal deduplication UI
Analytics / reports
Auto-close rule engine

11 KiB Raw Blame History Unescape Escape