Career Agent Workbench

Career Agent Workbench is a local, human-governed portfolio application for researching public job postings, organizing application state, and producing reviewable résumé variants. The reusable Python application stays public-safe; profile material, tracker data, model credentials, and generated documents stay in a separate private workspace selected through one ignored .env file.

The project combines a provider-neutral public-job core, SQLite-backed state, evidence-grounded résumé workflows, deterministic rendering and ATS diagnostics, command-line and Make composition, a small Flask tracker/action adapter, and a proportional FastMCP server. It does not authenticate to LinkedIn, access private member data, submit applications, or contact employers.

Terms

• JOD — Job Opening Description: normalized source and prompt-ready job text.
• MRO — Master Resume Object: structured, reusable evidence derived from the human-maintained résumé source.
• ARO — Application Resume Object: a job-specific résumé mapping derived from the MRO.
• CLO — Cover Letter Object: a separately reviewed, human-owned letter.
• ATS — deterministic local diagnostics for a résumé/JOD pair.
• Variant — an isolated v1, v2, or manual résumé version whose selection is explicit and reversible.

Architecture: public code, private state

The repository contains application code, packaged templates, fictional demo data, synthetic tests, and public documentation. An operator-owned workspace contains the real profile, SQLite database, output, and temporary files. The normal local configuration uses two non-secret selectors in the ignored root .env:

CAREER_AGENT_WORKBENCH_WORKSPACE=../career-agent-workbench-ops
CAREER_AGENT_WORKBENCH_PRIVATE_ENV_FILE=../career-agent-workbench-ops/.env

The bootstrap file is read only to find the workspace and private dotenv. The private file is then loaded with exact mode 0600 where supported. Conventional members—including the explicit configured downloads/ destination—resolve beneath the workspace. Optional overrides are documented in .env.example. Explicit CLI/API values remain highest precedence, and no loader mutates the process environment.

public checkout                        private workspace
career_agent_workbench/   --------->   profile/
scripts/, skills/, Makefile            output/tracking/applications.sqlite3
packaged templates/static              output/, tmp/, and downloads/
fictional examples                     .blacklist and private .env

Evidence and human review

The MRO is built from a bounded YAML file plus its concise source text. Tailored claims must remain grounded in that evidence. Refinement, manual review, and highlighting recheck the configured evidence digests and use conditional database writes so stale work cannot silently replace a newer variant.

The application workflow keeps v1, v2, and manual variants distinct. Automatic preference is useful, but an explicit selection stays pinned until the user changes it. External application submission is deliberately outside the product boundary.

This diagram describes the governed workflow restored in 1.1.0, including workflow-specific GLM routing and local compare/select controls. Model calls remain explicit, evidence-grounded, and outside offline tests.

Fictional offline demo

The tracked demo source uses Avery Demo, Nimbus Quay Example Labs, reserved domains, and fixed synthetic dates. It contains no database or generated binary. Materialize it only into ignored, disposable local state:

make demo

That target creates one fictional tracker row, one selected v1 résumé object, one cover-letter object, and readable YAML, JSON, and HTML examples under tmp/demo-workspace. make demo invokes the Python demo materializer; the materializer itself performs no network, provider, model, browser, or child-process work.

For a normal development/operator checkout:

make install
make console-scripts
cp .env.example .env
make skill-link
make test
make lint

Edit only the ignored .env to point at a private workspace. Do not place real profile or application state in this repository. make install creates the public .venv, installs the editable .[dev,browser] package, and explicitly installs Playwright Chromium; it does not install Ollama or call a provider.

Current interfaces

Installed commands:

career-agent-workbench                  # help and version
career-agent-workbench-seed-jobs        # public-job matching and atomic seed
career-agent-workbench-audit-jods       # bounded local JOD audit
career-agent-workbench-refine-resume    # governed v1-to-v2 refinement
career-agent-workbench-webapp           # local Flask tracker/action adapter
career-agent-workbench-mcp              # FastMCP stdio server

The Makefile composes the reusable job-seeding, draft generation, ARO sync, second-pass refinement, highlighting, manual-pass, local web, demo, test, and lint workflows through public .venv executables. Omitted state variables emit no repository-relative private paths; JOB_IDS=all selects eligible active records where the target supports batch work.

The modular Flask adapter provides tracker filtering/lifecycle, Add/Seed and bounded public-URL ingestion, allowlisted sequential resume actions with progress, explicit ATS refresh, JOD and resume/CLO editing, variant comparison and selection, inline/download artifacts, and explicit configured filesystem copies. GET routes remain read-only. The MCP adapter exposes three workspace-independent public job tools plus one lazy matching tool that uses the same centrally resolved runtime. Public-only tool startup requires no workspace.

Local website, skills, and MCP

make start-website                 # browser stays closed
make start-website OPEN_BROWSER=1  # explicit opt-in
make stop-website
make restart-website

Lifecycle state and the bounded server log live under ignored tmp/website/. Only the exact recorded child PID is stopped; stale records do not trigger port or process scans. make launch-website remains the foreground command.

make skill-link idempotently links exactly the five tracked public skills into the configured CODEX_SKILLS_DIR (normally ~/.codex/skills) and refuses to overwrite a non-symlink. The canonical MCP registration points to the absolute public .venv/bin/career-agent-workbench-mcp executable. MCP stdio arguments contain no private workspace values or tokens.

Synthetic presentation

The annotated PNGs below are synthetic documentation illustrations using the tracked fictional identity. They document the restored tracker, action, progress, JOD, résumé, variant, and cover-letter surfaces without exposing a private operator workspace.

Seed and review tracker state

The Add/Seed surface joins the existing guest provider and generic parser to atomic state seeding and the same app-scoped background action registry.

Inspect stored objects and variants

Operational limits

• LinkedIn integration uses guest-accessible public pages only. There is no authentication, private-member access, or application submission.
• Generic URL ingestion uses a bounded public HTTP seam and normalized parser; tests inject responses and never use the network.
• Public providers and optional model runners can fail, rate-limit, or change shape. Failures remain bounded and content-hidden, and retries are finite.
• No workflow contacts an employer. Human review and an external decision are required before any real application action.
• SQLite state, profile evidence, generated documents, credentials, and model configuration are local and operator-owned.
• This is a single-user portfolio application, not an enterprise deployment or multi-tenant service.

Development and validation

• Representative behavior coverage
• Fictional demo workspace
• 1.0.0 release notes
• 1.1.0 release notes
• Repository guidance
• MIT license

The public-safety checker scans tracked and distribution candidates without printing matched content. CI is read-only and proportional. Version 1.1.0 remains a release candidate until the reviewed PR is merged normally and its merge commit receives the separately authorized v1.1.0 tag.