job-forge

name	job-forge
description	Run the JobForge AI job-search pipeline for job URLs and descriptions, offer evaluation, tailored CV/PDF generation, portal scanning, application submission, tracking, follow-up, outreach, rejection analysis, and negotiation. Use when the user invokes $job-forge or /job-forge, pastes a job posting, or asks to manage any stage of a JobForge search.

JobForge

Run JobForge from the consumer project root. Treat the current user request as the invocation input; do not expect a template variable to contain arguments.

Start safely

1. Read AGENTS.harness.md when it exists. It is the shared operational contract. Continue to obey the consumer project's root AGENTS.md and any closer overrides.
2. Before responding, check for cv.md, config/profile.yml, and portals.yml. If any are missing, enter onboarding and follow modes/reference-setup.md instead of evaluating or applying.
3. When the request selects a mode, announce it before taking mode actions. For a bare discovery invocation, show the menu without inventing a mode.
4. Read only the mode-specific context listed below. Do not preload the whole tracker, portal catalog, scan history, or every mode.

Select a mode

User intent	Mode
Pastes a job description or URL	auto-pipeline
Evaluates one offer	offer
Compares offers	compare
Drafts LinkedIn outreach	contact
Researches a company	deep
Generates a tailored CV/PDF	pdf
Evaluates training or certification	training
Evaluates a portfolio project	project
Reviews application status	tracker
Fills or submits an application	apply
Searches configured portals	scan
Processes pending URLs	pipeline
Batch-processes offers	batch
Finds follow-ups	followup
Records or analyzes a rejection	rejection
Handles a received offer	negotiation

If the invocation contains no mode, URL, job description, or recognizable intent, show the discovery menu below. If an intent remains ambiguous, ask which mode fits instead of guessing.

Load context

Read modes/_shared.md plus modes/{mode}.md for:

• auto-pipeline, offer, compare, pdf, contact, apply, pipeline, scan, and batch.

Read only modes/{mode}.md for:

• tracker, deep, training, project, followup, rejection, and negotiation.

Load data only when the selected mode needs it:

Data	Modes
data/applications.md or data/applications/*.md	tracker, followup, rejection, compare, and dedup/numbering in auto-pipeline or batch
data/pipeline.md	pipeline, scan
data/scan-history.tsv and portals.yml	scan
batch/batch-prompt.md, batch/batch-state.tsv	batch
config/profile.yml	modes that load _shared.md
cv.md	pdf, auto-pipeline, and CV tailoring in apply

Execute

Follow modes/{mode}.md exactly, including its gates and output format. Apply the hard limits in AGENTS.harness.md to any multi-job workflow.

Map shared orchestration terms to the active client instead of inventing tool names:

• In OpenCode, use its task subagent primitive.
• In Codex, use the available collaboration/subagent tools.
• If the client has no subagent capability, run safe inline work and report any mode requirement that cannot be satisfied.

Use the MCP tools actually exposed in the session. Follow the documented fallback when Geometra or Gmail is unavailable; never claim that a live page, form, submission, or mailbox action succeeded without tool evidence.

Discovery menu

Show both native invocation forms so the user can continue in the same client:

JobForge command center

Codex:    $job-forge <mode or job URL>
OpenCode: /job-forge <mode or job URL>

Modes: auto-pipeline, offer, compare, contact, deep, pdf, training, project,
       tracker, pipeline, apply, scan, batch, followup, rejection, negotiation

Paste a job URL or description to run auto-pipeline.