Agena

简介

Key Features

AI Pipeline - Autonomous PM → Planner → Developer → Reviewer → Finalizer workflow - CrewAI role-based agents + LangGraph state machine orchestration - Prompt Studio — edit system prompts at runtime without code deploy - Vector memory (Qdrant) — learns from previous tasks for better context

History-Grounded Sprint Refinement - Index closed Azure / Jira work items (with final SPs, assignees, PR titles, branches) into a per-org Qdrant collection - For every new item, the LLM receives the top-5 similar past items and anchors its SP suggestion on your team's real distribution — naming which item it resembles and who did the prior work - See docs/REFINEMENT.md for the full flow and tuning knobs

Team Skill Catalog

Skills are AGENA's compounding-knowledge layer. Every completed task gets distilled into a reusable pattern; every new task pulls the closest patterns into its agent prompt. Past solutions are not re-discovered — they get re-applied.

- Auto extraction. When a task completes, the Skill Librarian asks the LLM (Claude CLI / Codex CLI / OpenAI / Gemini, whichever you're using) to summarise the task's title + description + reviewed code into `{ name, pattern_type, tags, touched_files, approach_summary, prompt_fragment }`. Skipped automatically when the LLM's confidence is < 50%, so one-off oddities don't pollute the catalog.

- Vector retrieval. Each skill is embedded into Qdrant (org-scoped, kind='skill') alongside a MySQL row that drives the catalog UI and usage counters.

- Prompt injection. Before any agent runs, the top-3 skills above the relevance threshold are prepended to the system prompt as a RELEVANT TEAM SKILLS block. Works in Claude CLI, Codex CLI, Gemini, OpenAI, and the classic CrewAI pipeline — same retrieval, four transports.

- Concrete payoff. After ~50 tasks, a Flo-sized team's catalog carries 20+ codebase-specific patterns:

  ┌─ "Stock allocator pattern" ──────────────────────────────────┐
  │  pattern_type: refactor                                       │
  │  tags: [stock, multi-site, allocation]                        │
  │  touched_files:                                               │
  │    app/Library/Stock/SiteBasedAllocator.php                   │
  │    app/Services/V1/StockService.php                           │
  │  prompt_fragment:                                             │
  │    "When site-based stock allocation is needed, use the       │
  │     SiteBasedAllocator class. Don't compute it inline — the   │
  │     allocator already handles overflow, fallback sites, and   │
  │     the per-SKU lock."                                        │
  └───────────────────────────────────────────────────────────────┘
  

Next time someone files a stock-related task, the agent gets that fragment in its prompt — no senior tap on the shoulder needed.

- Manual edits welcome. Hit /dashboard/skills and the "Yeni Skill" button to seed your team's pattern library directly. Useful for encoding "don't do this again" lessons (e.g. "httpx params={} silently strips the URL's query string — always pass None") before AGENA has accumulated enough completed tasks to extract them on its own.

- See docs/SKILLS.md for the extraction + retrieval flow, tuning knobs, and the comparison vs Refinement's similar-past index.

Runtimes Registry - Every compute environment that can execute agent tasks (host CLI bridge, teammate's laptop, cloud daemon) registers as a Runtime - Auto-enrollment: bridge-server.mjs picks up AGENA_JWT + AGENA_TENANT_SLUG from env and calls /runtimes/register on startup, then heartbeats every 30s with its current CLI availability - /dashboard/runtimes shows the live list with status dots, CLI badges, heartbeat age, and daemon version - See docs/RUNTIMES.md for the enrollment + heartbeat + security model

Multi-Repo Orchestration - Assign a single task to multiple repositories simultaneously - Each repo runs its own AI pipeline in parallel — independent branches and PRs - Per-repo locking prevents concurrent conflicts - Unified dashboard shows all PRs and their status in one view

Task Dependencies & Auto-Queue - Define execution order: Task B depends on Task A - Worker checks dependencies before running — blocked tasks wait automatically - When a dependency completes, dependent tasks auto-queue and start - Cycle detection prevents circular dependency chains - Visual dependency flow in the dashboard: A ✓ → B ✓ → C (waiting)

DevOps Automation - Auto-generates branches, commits, and pull requests (GitHub + Azure DevOps) - Sprint import from Jira and Azure DevOps - DORA metrics dashboard (deployment frequency, lead time, change failure rate, MTTR) - Team health symptom analysis (knowledge silos, bus factor, stale PRs, etc.)

Shareable Task Links - One-click "Share" on any task — picks a duration (1–30 days) and a use cap (1–25 imports), gets back a tokenised public URL - Recipients without an account can read the task (description, comments and inline images proxied through the sender's PAT so screenshots still load) and click "Import to my org" to copy the task — including attachments — into their own workspace - Built-in share buttons for WhatsApp, Microsoft Teams, Slack, Telegram, Email and X next to the link, so the URL never has to leave the modal manually - Tokens are revocable, time-limited and use-capped; expired or spent links return a friendly "this link is no longer active" page

Multi-Tenant SaaS - Organization isolation with subdomain routing - JWT auth + RBAC (owner, admin, member, viewer) - Free/Pro/Enterprise plans with usage quotas - Stripe billing integration

AI-powered Sprint Nudges - Every blocked Azure DevOps or Jira work item gets a "Ping" button - Checks the last comment timestamp — if >24h silent, fires a nudge - LLM-generated, single-paragraph, language-picked-per-ping (7 locales) - Pick the generator at send time: Claude CLI (subscription, via host bridge), Codex CLI (subscription), OpenAI, Gemini, or HAL. Claude CLI failures auto-fall-back to Codex CLI on the same bridge. - Posts back with an <at>@Display Name</at> mention so Azure renders it as a real tag, plus a signed "written by Agena via {model}" line for transparency - 48h cooldown + DB-persisted history prevents spam; UI badges already-nudged items with "Agena pinged Xh ago"

Dashboard - Boss Mode — pixel-art office where you manage AI agents visually - Visual flow builder — drag-and-drop automation pipelines - Sprint performance tracking with risk scoring - Real-time task monitoring with live logs and WebSocket updates - Guided tour onboarding for new users - 7 languages (TR, EN, ES, ZH, IT, DE, JA)

Cross-Source Insights (module: insights, default off) - Correlation engine that ties PR merges + deploys + Sentry / NewRelic / Datadog / AppDynamics / Jira / Azure events into one timeline - Confidence-scored clusters (0-100); ≥ 70 surfaces on /dashboard/insights with a one-sentence narrative + full timeline - "Which deploy caused this bug?" answered in seconds — replaces the 20-minute tab-hop during incidents - Confirm / False positive / Undo triage; Confirm can open a one-click rollback PR for the suspected commit - Poller runs every 5 minutes, no extra integration setup — re-uses the data the existing AGENA clients already cache - Marketing page: /cross-source-insights

Stale Ticket Triage (module: triage, source-side scan, AI verdicts) - Source-side scan, not local mirror. Hits Jira's /rest/api/3/search/jql (the new endpoint — the deprecated /search one is gone) with cursor pagination, and Azure DevOps WIQL across every project the org has access to. No internal task table fan-out — the truth lives at the source, and so does the scan. - Background task + progress polling. Scan runs as a strong-ref asyncio task; the dashboard polls /triage/scan/status every 2s for per-row progress so the UI never blocks on a 10-minute crawl. - AI verdict per ticket. The reviewer agent (Claude CLI / Codex CLI / hosted OpenAI — whichever the org configured, never an env fallback) reads title, description, last comment + ticket age and emits one of close / snooze / keep plus a one-sentence reason. - source_updated_at cache. If a ticket's source-side timestamp hasn't moved since the last scan, we skip the LLM call and just backfill any new metadata fields. Fresh decisions only when the source actually changed. - Look-back window. triage_max_age_days caps how far back the scan reaches — useful when you want "only the last 12 months," not history all the way to 2018. - Filters baked in. Source (Jira / Azure), project, state, and status chips. Cancelled / Won't Fix / Rejected excluded by default on both the WIQL query and a defensive Python filter on read. - Confirm-before-apply modal. "Apply all AI suggestions" never fires silently — you see the count, the breakdown, and click to commit. Audit trail per decision in triage_decisions. - Marketing page: /stale-ticket-triage

Review Backlog Killer (module: review_backlog, multi-channel, AI-aware) - Five nudge channels, comma-selectable. Slack DM, WhatsApp, email, Telegram, and a comment posted directly on the PR thread itself (GitHub pulls/{n}/comments or Azure {org}/{project}/_apis/git/repositories/{repo}/pullRequests/{n}/threads — Azure URL needs the project segment, not just the repo guid). - AI-generated comment body, optional. Toggle on and the reviewer agent reads the actual diff (GitHub Accept: application/vnd.github.v3.diff or Azure iterations/{n}/changes) and writes a polite, diff-specific question — not a generic "still around?" template. - Comment language picker — 7 languages. Set the org's preferred tone in nudge_comment_language (en/tr/de/es/it/ja/zh); both static templates and the AI prompt swap accordingly. - Auto-nudge worker (opt-in). Off by default — backlog_auto_nudge boolean toggle. When on, the 30-minute worker poller auto-posts to the configured channels with a hard 24h interval floor (max(24, backlog_nudge_interval_hours)) so nobody ever gets spammed by accident. - Azure deletion detection. Before counting a row as "already nudged," we verify the existing Agena comment still lives on the PR. If a reviewer wiped the comment, last_nudged_at resets and the nudge becomes available again — same logic the production detector uses to dodge stale state. - Per-row cooldown that's actually delivery-aware. Status return is sent / rate_limited / comment_failed; we only stamp last_nudged_at on sent, so a 400 from Azure doesn't lock the row for 6 hours. - Searchbox + chip-tag exempt-repo picker. No more 80-row alt-alta toggle list. Type, autocomplete, click to add as a chip — same UX as Jira/Azure label pickers elsewhere in the app. - Marketing page: /review-backlog-killer

---