backend-agentic-marketplace

Explore a backend codebase once with breadth (Sonnet) + depth (Opus) sub-agents, then persist a durable .claude/explorer/ memory so future sessions understand the code without re-exploring.

Spec-driven agentic implementation for Claude Code. Reads the explorer codebase memory as ground truth, plans a requested change with clarity + technical-plan gates, decomposes it into atomic edge-case-hardened tasks (micro-level precision), implements only what the spec says with a per-edit lint/type feedback loop and hybrid retrieval, runs hybrid QA, and keeps the durable memory in sync. Includes a reproduce-first BUG-FIX MODE that turns vague, repro-less bug reports into verified, regression-safe fixes via a failing reproduction + characterization tests + a deterministic regression gate.

Conductor plugin that sequences the explorer -> builder flow and gates it behind a deterministic RELEASE GATE. Reads the per-module STATUS contract and durable memory, runs exploration first when memory is missing or stale, then the spec-driven build (or BUG-FIX MODE), then verifies release-readiness (fresh memory, builder done with full coverage map, bug-fix net green, changelog present, auditor 0-high when present). Advisory by default; enforce via settings.enforce_release or PIPELINE_ENFORCE=1. Pure shell/awk gate — no python dependency.

Productizes this marketplace's own audit methodology (the F1–F13 failure classes) as deterministic, pure-shell/awk REGRESSION detectors plus breadth+depth review sub-agents. Scans every plugin's hooks, guards, manifests, and vendored libs; emits HIGH/MEDIUM/LOW/ADVISORY findings to .claude/auditor/FINDINGS.md; and records a HIGH count in the STATUS contract so the pipeline release gate can enforce 0-high. Silent on a clean tree, fires only on a regression. No python dependency — identical on python-less / Windows hosts.

Reviews a CHANGE (the working-tree diff vs HEAD) against the explorer MEMORY.md invariants and risk map, the approved PLAN scope, and the codebase's surviving callers. Deterministic, pure-shell/awk checks (R1 caller-integrity, R2 convention-regression — BLOCKING; R3 risk-touch, R4 scope-discipline — CONCERN) plus breadth+depth review sub-agents. Emits BLOCKING/CONCERN/NOTE findings to .claude/reviewer/REVIEW.md and records a BLOCKING count in the STATUS contract so the pipeline release gate can enforce 0-blocking. Orthogonal to the auditor (change-vs-invariants, not whole-tree regression). Silent on a clean/in-spec change; no python dependency.

Assesses DEPLOY/RELEASE readiness of the codebase. Deterministic, pure-shell/awk checks kept deliberately LEAN — O1 test-ledger (a recorded RED build/test → BLOCKING; no ledger → CONCERN) and O2 version-consistency (a marketplace entry's version ≠ that plugin's plugin.json version → CONCERN) — plus breadth+depth ops sub-agents that inventory the CI/deploy/observability surface and reason about readiness as ADVISORY CONCERN/NOTE findings. Emits BLOCKING/CONCERN/NOTE to .claude/ops/OPS.md and records a BLOCKING count in the STATUS contract so the pipeline release gate can enforce 0-blocking. Fuzzy deploy judgment lives in the agents, never in a false-firing detector. Silent on a verified, consistent tree; no python dependency.

Always-on "write the least code that fully works" discipline. A SessionStart hook injects the full decision LADDER (skills/minimal-code) and a UserPromptSubmit hook re-injects a compact reminder every turn so the agent ships the SMALLEST code that satisfies the requirement — never thousands of debug-hostile lines — WITHOUT cutting validation, error handling, security, or accessibility. Both hooks are node-guarded and fail-quiet (node is OPTIONAL: absent -> the hooks stay silent and the skill still works on demand; never a hard dependency), and they INJECT CONTEXT ONLY — they make no gate decision, so every gate in this marketplace stays pure shell/awk. Intensity off|lite|full|ultra via /minimize, mirrored into STATUS.json for the dashboard. YAGNI applies only to UNREQUESTED scope: anything the builder PLAN or the user mandates stays in-scope. Ladder adapted from Ponytail by Dietrich Gebert (MIT).