返回资讯中心
外部精选
软件工程
#开发者实时流

JaJa: Job Agency Harness

A job-search agency for Claude Code: a team of agents that finds roles and people, gates and ranks them,

Hacker News 最新khalidmehtabk2026/10/26 · 21 天前30 分钟阅读

以下正文同步自 Hacker News 最新,版权归原站所有,已转换为易读排版。

A job-search agency for Claude Code: a team of agents that finds roles and people, gates and ranks them, tailors pages from your own evidence and drafts every message, and never sends anything without your yes.

jaja is how Spanish writes laughter: because a job search deserves a smile now and then.

Scripts count, gate and check; models judge and draft; you decide and sign. Your profile, contacts, jobs and writing style live in your own data folder and never touch this repository.

The dashboard on the bundled demo data (a fictional candidate). Nothing real appears in this repository.

git clone <repo-url> ~/jaja && cd ~/jaja ./install.sh --demo --yes # synthetic demo data + a health check; nothing real, nothing sent claude plugin marketplace add ./ && claude plugin install jaja@jaja --scope user

Then cd ~/jaja-demo && claude and try /jaja:daily-job-plan, /jaja:score-fit or /jaja:research-company. The installer ends with scripts/doctor.py; on a clean machine it reports 0 failures. For your own search run ./install.sh --data ~/jaja-data (a blank skeleton) and follow getting started. Requirements: macOS or Linux, Python 3.11+, Claude Code; uv is recommended for the MCP server. Optional extras (Gmail and Calendar, resume rendering, PDF checks): pip install -r requirements-optional.txt.

The daily flow. Every box that touches the outside world stops at an approval you sign.

flowchart LR A[Find<br/>boards, posts, people] --> B[Judge<br/>gates with evidence] B --> C[Write<br/>from your evidence] C --> D[Approve<br/>you sign] D --> E[Send<br/>approved files only] E --> F[Track<br/>one event log] F --> G[Learn<br/>outcomes] G -.-> A

Loading

The architecture. The plugin is this repository; everything personal stays in your data folder.

flowchart TB you([you: decide, approve, send]) --> cc[Claude Code session in your data folder] subgraph plugin [jaja plugin: this repo] agents[agents/<br/>30 roles] skills[skills/<br/>52 procedures] hooks[hooks/<br/>9 guards] mcp[mcp/<br/>records server, never sends] end cc --> plugin plugin --> scripts[scripts/<br/>gates, scoring, rendering,<br/>tracking, approvals] scripts <--> data[(your data folder<br/>profile, jobs, prospects,<br/>applications, outbox, hub, config)] scripts <--> rt[(runtime folder<br/>approval queue, run receipts)]

Loading

30 role agents in seven teams, each defined by a card (charter, tools, autonomy tier, model, budget, KPIs). Open a team:

Around them: 52 skills (/jaja:<name>), 8 lanes (apply, focused, hunt, posts, prep, recruiter, research, resume), 16 routines (all disabled), 9 guard hooks and a 24-tool MCP server. Full lists: reference.

How it works

Your profile profile/facts.yaml and profile/evidence.yaml hold what you can prove, one evidence id per claim. Every resume line, answer and letter must cite an id; checks refuse a claim with no source and strings you listed as "holds". One frozen, hash-pinned base resume; a tailored variant ships only if it beats the base in a blind comparison.

Agent profiles A role card (agents/<name>.md) states purpose, allowed tools, autonomy tier (T0 to T4), model, effort, budget, KPIs, stop conditions and denied commands. scripts/kernel/sync.py turns the cards into Claude Code agents (and optional Codex adapters) and --check catches drift.

Memory Your data folder is the single source of truth: records, the hub's rulings, events and registry, append-only logs and run receipts, and the learning loop that turns outcomes into proposals. None of it lives in this repo.

flowchart LR work[agents and scripts do the work] --> rec[(records and receipts<br/>append-only)] rec --> learn[learning loop<br/>outcomes with sample sizes] learn --> prop[proposals for you] prop -->|you rule| hub[(hub: rulings, events, registry)] hub -->|injected into lanes| work prof[(profile: facts and evidence ids)] --> work

Loading

Design nuances

What makes it different, one line each. Each is enforced by a script or hook; docs/NUANCES.md names it.

Tests 5,192 tests pass on the demo data with no real data and no network; 321 test files plus 70 MCP tests

Scope 142 scripts, about 76,000 lines of Python (standard library, PyYAML, requests)

Architecture code, your data and a runtime folder kept apart: architecture, design decisions

Config seam identity, style and policy come from config with neutral defaults; tests run as a synthetic persona

Health python3 scripts/doctor.py checks install, config, hooks, adapters, MCP server and leak guard

Who it is for

For engineers and other technical job seekers who already use Claude Code and want a structured, auditable pipeline for a long search: many roles, many people, many drafts, one true record. Not for you if you want a bot that applies for you without review (the design refuses that), or if you do not want to keep your own evidence and records in files.

Details and honest trade-offs: comparison, FAQ.

Getting started · Your first search · Demo walkthrough · Configuration · Lanes · Customizing agents and skills · Architecture · Concepts · Design nuances · Design decisions · Reference · Charter · Connectors · Roadmap · Changelog

Early, and used by its author for a real search. It works end to end, but expect rough edges: the layout and file formats can still change, only Claude Code is tested (the Codex adapters are generated but not run end to end), and connectors (Gmail, Calendar, X, a logged-in browser) are optional and off until you set them up. Respect each platform's terms; you are responsible for what you approve and send. This project is not affiliated with Anthropic or any job platform.

The product page ends above. This is the developer depth: every agent, gate, hook, deterministic script, skill, lane, routine, MCP tool, configuration key and nuance in the system, one collapsible section each. The tables are generated from the repository (python3 tools/gen-readme.py), so every name matches a file; the guides in docs/ explain how to use them.

Each card in agents/ states purpose, allowed tools, tier_max (T0 read to T4 forbidden), model, effort, budget, KPIs, stop conditions and denied commands.

Nothing reaches you until these have run, in a clean context where it matters.

Shared helpers: lane_common.py, rulings_common.py. The list is hooks/hooks.json; sync.py merges it into .claude/settings.json and .codex/hooks.json.

Scripts count, gate, render, track and check; they never call a model. Cards and skills call them.

Find (discovery, capture and platform terms)

Gate and score (hard gates with evidence, fit and ranking)

Evidence, pages and checks (claims, rendering, lints and gates on every page)

People and outreach (contacts, drafts, approvals' inputs and mail)

Track, report and learn (the one record, the day, the numbers and the learning loop)

/BUNDLE.md: the one page the owner reviews before approving anything on a pursuit

Pacing, safety and policy (caps, block-page stops, run guards, rulings and model policy)

Kernel (scripts/kernel/: adapters, lint, approvals, hub tools, receipts)

A lane is a long-lived session with a charter, a roster and a scope; launch it with claude --agent lane-<type> --settings <repo>/.claude/lane-settings/<type>.json.

Run with uv run --directory mcp job-search-mcp; registered in .mcp.json. Text from postings and mail comes back inside <untrusted_*> tags and is data, never instructions.

code root (this repo) data root (your folder) runtime root agents/ skills/ hooks/ profile/ jobs/ prospects/ work/approvals/ (the approval queue) scripts/ mcp/ dashboard/ applications/ outbox/ state/ logs/runs/ (run receipts) schemas/ templates/ config/ hub/ targets/ config/ views/ vault/ (secret pointers)

Plugin. The repository root is the Claude Code plugin (.claude-plugin/plugin.json): skills from skills/, agents from agents/ (each card carries native description and tools fields that sync.py keeps in step), hooks from hooks/hooks.json, the MCP server from .mcp.json.

Roots. scripts/agency_paths.py resolves the three roots: JAJA_DATA (or JOB_SEARCH_ROOT, then ~/.config/jaja/config.yaml), JAJA_RUNTIME (default <data>/runtime), under tests or JAJA_DEMO=1 the bundled demo. scripts/js_common.py is the shared record library.

Config seam. scripts/agency_config.py reads identity, style and policy from config/defaults.d/.yaml (neutral defaults), overlaid by your <data>/config/.yaml; under pytest a synthetic persona (scripts/tests/persona.d) is used instead.

Generated adapters. scripts/kernel/sync.py renders agents/ to .claude/agents/ and .codex/agents/, lanes to lane-<type> agents with settings in .claude/lane-settings/, skills to .agents/skills/, and merges hooks and MCP config; --check reports drift.

Hub. Lane sessions coordinate through <data>/hub/: programme.json, registry.json, events.jsonl, rulings.yaml, lanes/, briefs/.

Your data folder (tools/init-data.py creates it; it never overwrites a file):

profile/ facts.yaml evidence.yaml PROFILE.md VOICE.md (blank) skills-taxonomy.yaml corpus/ answers/ resumes/ one base resume + a folder per tailored variant jobs/ jobs-pending/ discovered postings with gate results applications/ per-application fit report, pack and events prospects/ graph/ touches/ people, graph, every touch outbox/ drafts waiting for approval pursuits/ company-level campaigns targets/ roles.md platforms.yaml fit-gates.yaml post-queries.yaml x-deny.yaml x-watch.yaml mass.yaml goals.yaml state/ reports, learning loop, presence, browser logs hub/ orchestration files config/ identity.yaml style.yaml policy.yaml views/ optional Obsidian Bases tables sources.yaml discovery channels (at the data root) legal/ your own work-authorization timers (optional)

Configuration keys Every key the code reads has a neutral default in config/defaults.d/.yaml; override any of them in <data>/config/.yaml (files merge in name order). Number in brackets: sub-keys.

The one-line claims in the README, each with the code that enforces it. If a line here stops being true, the README is wrong: fix the code or the claim.

Nuance Where it is enforced

Nothing goes out without your yes. Tiers T0 (read) to T4 (forbidden); every outward action is an approval file. Each role card states tier_max (agents/*.md, schema schemas/role.yaml). scripts/kernel/approvals.py writes and verifies approval files (they expire after 7 days). scripts/outbox.py can-send refuses anything not approved; the sender card is the only one that executes sends.

Denied commands cannot be run by a role that must not run them. Cards list deny_commands; the PreToolUse hook hooks/role_deny.py blocks them (listed in hooks/hooks.json).

Connectors are optional and read-only first; sending uses a separate token and login. scripts/gmail_auth.py keeps separate read, compose and send tokens (login, login --compose, login --send); tokens live in ~/.config/jaja/, never in the repo. Setup: docs/SETUP-CONNECTORS.md.

Pacing caps, gaps and a stop on a block page. scripts/pace.py (daily caps and gaps per platform, from config policy.pace.*), scripts/browser_pace.py (a 429 or block page stops all lanes), hooks/browser_lease.py (one keyboard on a shared browser).

Routines ship disabled, write receipts, and a watchdog flags silence. Every routines/*.yaml has enabled: false; scripts/kernel/lint.py flags an enabled model routine without enabled_by; scripts/kernel/watchdog.py compares the newest receipt in <runtime>/logs/runs/<name>/ with the schedule.

Truth

Nuance Where it is enforced

Every claim cites an evidence id from your records. scripts/check_evidence.py validates the evidence registry and citations; scripts/render_resume.py lints missing-evidence and unknown-evidence; scripts/numbers_ledger.py and scripts/check_holds.py check figures and strings you listed as holds.

Gates fail with the exact line they quoted. scripts/fit_gates.py writes gate, gate_reasons and gate_evidence on each job record.

正文由 探How 从来源站点 RSS 同步,内容未经改写;遇到排版缺失或需要图片、视频时请以原文为准。

打开原文
去社区讨论 →