Imported from marqHQ/marq-marketplace-sales (AGENTS.md). Install upstream with npx skills add marqHQ/marq-marketplace-sales. Copyright stays with the author.

Agent guide

Dual-compatible plugin marketplace (Claude Code + Codex) for Marq sales skills. Human-facing overview and install steps: README.md.

Commands

No build step. Run the relevant test suites:

python plugins/marq-sales-suite/skills/audit-hubspot-pipeline/scripts/run_tests.py
python3 plugins/marq-sales-suite/skills/marq-deal-acceleration/scripts/run_tests.py
python3 plugins/marq-sales-suite/skills/marq-post-call-execution/scripts/check_acceptance_fixtures.py
python3 .claude/skills/review-submission/scripts/test_unpack_submission.py

Run the audit suite after any change under audit-hubspot-pipeline/scripts/ or to references/scoring-fixtures.json. Run the deal-acceleration suite after any change to its momentum scorer, contract, or fixtures. Run the post-call fixture checker after changes to its workflow, evidence model, CRM mapping, or acceptance cases.

Layout

• plugins/marq-sales-suite/skills/<name>/ — each skill: SKILL.md, references/, optional scripts/ and agents/openai.yaml (Codex interface metadata).
• Marketplace manifests: .claude-plugin/marketplace.json (Claude Code) and .agents/plugins/marketplace.json (Codex).
• Plugin manifests: plugins/marq-sales-suite/.claude-plugin/plugin.json and plugins/marq-sales-suite/.codex-plugin/plugin.json. Keep name, description, and version in sync across both when editing either.
• under-construction/ — draft docs, not skill-formatted. Don't load or promote them without adding proper SKILL.md frontmatter.
• .claude/skills/review-submission/ — owner-side review skill run by .github/workflows/review-skill-proposal.yml, plus the deterministic unpack_submission.py and verify_repo.py scripts. Not part of the shipped plugin. See CONTRIBUTING.md for the full intake → review → approval pipeline.

Invariants — do not weaken when editing skills

• Every CRM write is gated behind an exact approval table; approval of a different proposal is insufficient.
• Scheduled/unattended runs are read-only regardless of prompt wording.
• No negative inference from silence (absence of discussion never becomes No/None).
• Portal fields, stages, and enums are discovered live, never hardcoded.
• Writes are verified by an independent connector re-read, never by a UI success banner.
• hs_next_step remediation is append-only.

Scoring

Green/yellow/red scoring is deterministic Python in scripts/audit_utils.py — never reproduce it in prose or let a model score heuristically. references/scoring.md is the contract; keep it and the code in sync. Fixtures live in references/scoring-fixtures.json; add a fixture when adding a scoring rule.

The deal-acceleration skill has its own deterministic customer-momentum classifier in scripts/momentum_utils.py. Its contract is references/momentum-and-evidence.md; fixtures live in references/momentum-fixtures.json. Keep all three synchronized.

Definition of done

• Test suite passes.
• Both plugin manifests (.claude-plugin/plugin.json and .codex-plugin/plugin.json) still agree on name, description, and version.
• Any scoring behavior change is reflected in references/scoring.md and covered by a fixture.
• Skill edits preserve every invariant above.