dsh-spec-forge

The most frustrating part of a coding conversation is rarely the code itself. It is the repeated inefficiency: starting work before the requirement is clear, re-explaining project no-go zones every new session, describing the same pagination-interface pattern as if it were the first time. dsh-spec-forge was built for exactly that: before you start, it sharpens a vague requirement into a concrete one; after you finish, it stores the experience; and the next time a similar task shows up, it brings the right template back automatically.

The core is a five-tool closed loop. spec_recall builds a weighted keyword fingerprint (path, tech term, identifier, Chinese 2-gram scored at different weights) and matches it against a plaintext template library, injecting the best one or two templates plus project no-go rules into context. spec_triage grades the task as L1 atomic (emit an execution checklist, ask nothing), L2 module change (at most 3 questions, each with an inferred default), or L3 architecture refactor (full Grill-me). spec_distill turns the clarified requirement into a single implementation prompt. spec_retro captures the session lessons into a Markdown template, idempotent by name and overwriting rather than piling up. spec_library lets you browse the full library, hit stats, and expired-template report on demand. Everything is stored as zero-dependency plaintext Markdown under ~/.dsh/spec-forge, readable, editable, Git-friendly, and not a black box. Project-level do-not-touch rules accumulate in a per-repo profile.md and are auto-injected into every future matching task, so you never have to repeat them.

It fits developers working daily in DeepSeek Harness who repeatedly handle CRUD, field add/remove, pagination, and similar patterns, and who want their project conventions to persist across sessions. It does not make network calls, execute shell commands, or read credentials; the source is public and fully auditable before installation.