Resume-Builder AGENTS.md

Resume Builder - Codex Context

Overview

Resume Builder creates tailored resumes, cover letters, job-fit reports, DOCX application packages, and tracker updates from job descriptions.

Core targets:

• Authenticity: never invent experience, metrics, titles, dates, education, publications, certifications, or memberships
• Human voice: brevity, burstiness, plain verbs — must pass human_voice_audit.py before DOCX
• ATS score: 75-85% (keywords in Core Competencies first; never stuff bullets)
• HR score: 70%+
• Output: ATS/Workday-compatible DOCX files

Editorial priority (never invert): Authenticity → Human voice → HR impact → ATS match. 75% ATS with human prose beats 90% stuffed AI prose.

Codex Plugin Surface

Codex uses:

• .codex-plugin/plugin.json for plugin metadata
• .agents/plugins/marketplace.json for the local marketplace entry
• skills/resume-team/ for the installed $resume-team workflow
• .codex/agents/ for project-scoped custom role definitions
• .codex.mcp.json for the Codex MCP scorer server

Installed Codex invocation:

• $resume-team [job description]

The Markdown files in commands/ are the shared workflow source and Claude Code slash-command surface; they are not installed Codex /resume-builder:* commands. Resume Team execution requires Python 3.10+, config.json, the configured master resume, and the local deterministic audit helpers.

Native Four-Role Resume Team

• Candidate fit is the mandatory first gate. Before any role/native-team invocation, output/application directory, resume development, DOCX, or tracker work, run candidate_fit_preflight.py against the exact JD and only the configured master_resume_path (never a previous tailored resume). Continue only with a canonical, digest-bound candidate-fit-policy-v3 report scoring at least 70, trustworthy extraction, zero hard knockouts, passed: true, and no codes. A lower score (including 60–69) or hard knockout is REJECTED:CANDIDATE_FIT; unavailable, malformed, stale, or mismatched analysis is FAILED:CANDIDATE_FIT_PREFLIGHT. Both fail closed with no bypass. ATS/HR baseline scores are advisory and cannot override candidate fit.
• Candidate-fit policy v2 calibration: required-tool knockouts are grounded only in requirements/qualifications sections (a tool named in duties prose, e.g. "data mining in Empirica Signal," is never a knockout); doctorates named anywhere in requirement ladders are alternative degree routes (an MD satisfies "PharmD/PhD or equivalent" listings); hard knockouts degrade the score with an informative cap (1 → 55, 2 → 45, 3+ → 35) instead of flattening to 35.
• Stretch-zone review layer: when the deterministic floor passes (trustworthy extraction, zero hard knockouts) but the score is below 70, the coordinator may run python candidate_fit_review.py --resume <master> --job-description <jd> --run-id <id> --case-id <id> --as-of-date <date> --json. It invokes two structurally distinct native reviewers (skeptical-recruiter and hiring-manager lenses), requires every claim to cite exact resume/JD lines, verifies each citation in code, and passes only when both reviewers return PROCEED with fully verified citations (candidate-fit-review/v1, REVIEW_CERTIFIED_PASS). Disagreement, malformed output, unverifiable citations, or reviewer unavailability fail closed. A review-certified pass authorizes coordinator-built manual packages only; the hardened native_resume_team.py runtime still requires the deterministic candidate-fit-policy-v3 pass (passed: true, score ≥ 70) and does not consume review reports.
• Manual user override: the final call belongs to the user. After any gate rejection (deterministic or review), the user may record a deliberate override with python candidate_fit_override.py --resume <master> --job-description <jd> --run-id <id> --case-id <id> --as-of-date <date> --reason "<explicit reason>" [--review-report <path>] --output <path>. The override requires a rejected gate and an explicit reason, binds the exact gate/review reports and digests it overrules (candidate-fit-override/v1, decision PROCEED_MANUAL), and prints its warning. It authorizes only the coordinator-built manual package path — truthful master-sourced content, evidence/human-voice/integrity audits, non-authorized renderers, files labeled MANUAL_DRAFT.txt — and never the native runtime, authorization receipts, authorized DOCX/tracker wrappers, or any claim that candidate fit passed. Overriding a passing gate errors; a missing reason fails closed.
• The root session is the sole coordinator. For an authorized draft, $resume-team must use python native_resume_team.py --host codex; do not replace it with manual prompt-spawn orchestration. The hardened runtime requires macOS or Linux and fails Windows preflight closed with POSIX_RUNTIME_REQUIRED. The project agents resume-researcher, resume-writer, resume-auditor, and resume-editor remain available for interactive inspection.
• Run Researcher → Writer → Auditor in order. Call Editor only after FAIL, then run a fresh Auditor. Maximum two corrections.
• Role agents return proposals only. They never write files, spawn roles, publish, update the tracker, clean up, or authorize their own output.
• Give each role only the scoped payload defined in commands/resume-team.md; validate every handoff digest, parent, attempt, run/case ID, and globally unique per-invocation agent identity, including re-audits and retries.
• Manual Codex sandbox_mode = "read-only" prevents role writes but does not isolate filesystem reads. The production CLI runtime supplies the capability-isolated path by launching roles in empty temporary working directories with interactive tool surfaces disabled. Do not claim equivalent isolation for manual role use.
• Researcher requirements are provenance-bound: hard_requirements followed by soft_requirements must equal the exact uniquely anchored jd_evidence_spans[*].evidence_text values one-for-one, byte-for-byte, and in order.
• Draft-stage publication requires a durable run claim, trusted source attestation, a final Auditor PASS, three fresh independent evidence/human-voice/canonical-integrity votes on the same draft digest, an exact committed publication receipt, and verified readback.
• Runtime PUBLISHED means only an authorized, digest-verified resume.md draft-stage artifact carried by resume-team-result/v2. It is not DOCX, tracker, cleanup, or package completion. /resume or /tailor-resume must finish all command-specific finalization gates before reporting package success.
• A PUBLISHED result is not consumable until its candidate_fit_report and candidate_fit_report_digest exactly match the independently validated passing preflight; the coordinator resolves authorization_receipt_path against the output directory when relative, requires its resolved parent to equal that directory, reads regular non-symlink resume-team-final-receipt/v2 JSON, validates schemas/resume-team-final-receipt.schema.json, recomputes its canonical digest, matches the inline receipt and result run/case IDs, and binds the same passing candidate-fit report/digest plus draft_digest and verified_target_digest to the configured master, exact JD, final_draft_digest, and actual resume.md. Recompute source_digest from the current configured master and job_description_digest from the fixed sibling job_description.txt; require a SHA-256 Researcher artifact and distinct same-host native Researcher/Auditor identities. Require a same-draft PASS auditor_attestation; require the full authorization_report to pass with no codes and exactly three ordered named same-draft PASS votes with distinct IDs; require canonical_digest(authorization_report) == authorization_digest and the receipt vote-ID list to equal those IDs in order. Revalidate immediately before DOCX and preserve the JD and durable receipt during cleanup.
• The hardened Codex runtime ignores user configuration and transient parent-session settings. Do not claim an inherited profile, model, or Ultra mode. Omit pins by default; use --model <exact-model> and/or --reasoning-effort ultra only when explicitly requested. No runtime profile option exists, and Codex-only pins must not be passed to Claude.
• The coordinator alone performs ordered finalization: verified resume DOCX → verified cover-letter DOCX → tracker update → artifact/state verification → cleanup → report. It must retain the captured runtime result, invoke final_receipt_verifier.py with that result's exact sidecar path and digest immediately before every side-effect boundary, and use only create_resume_from_md_authorized, create_cover_letter_from_md_authorized, and add_application_authorized. Advance state only after verified regular non-empty DOCX output or a tracker return value that is literally True; exceptions and false results stop finalization.
• Any malformed, stale, replayed, ambiguous, timed-out, unavailable, side-effecting, or partially published result fails closed.