# Show HN Draft ## Title (choose one) **Option D** (research-backed — RECOMMENDED for HN): ``` Show HN: ForgeDock — Codebase LLM Wiki for AI agents, built on GitHub Issues ``` *Rationale: "LLM Wiki" has 17M views of cultural priming from Karpathy's April 2026 gist. Developers searching for the pattern recognize it immediately. Direct vocabulary alignment with the activated term.* **Option E** (pain-first — RECOMMENDED for product pages and README hero): ``` Show HN: ForgeDock — AI agents forget everything. This gives them memory that lives in GitHub. ``` *Rationale: "Memory" is the #1 vocabulary by adoption (Mem0 48K stars). Names the felt pain ("forget") directly. No explanation required.* **Option A** (technical): ``` Show HN: ForgeDock – Open-source pipeline that turns Claude Code into a deterministic dev team ``` **Option B** (problem-first): ``` Show HN: ForgeDock – GitHub as institutional memory for AI coding agents ``` **Option C** (concise): ``` Show HN: ForgeDock – Autonomous investigate-build-review-merge pipeline for Claude Code ``` --- ## URL https://github.com/RapierCraftStudios/ForgeDock --- ## Maker Comment (post immediately after submission) Hi HN. AI coding agents suffer from context rot: the gradual degradation of output quality as context fills with noise, stale file contents, and conventions re-explained from scratch every session. Every context reset, the agent forgets what it investigated, what decisions it made, what it already tried. You end up re-explaining your codebase every time. Most teams try to fight context rot with config files — CLAUDE.md, AGENTS.md, .cursorrules, Copilot instructions, Kiro steering. Five files, five formats, five things to keep in sync, and the agent still can't reconstruct what happened in the last PR. ForgeDock takes a different approach: it turns GitHub itself into the agent's persistent memory. When the pipeline runs on an issue, each stage (investigate, build, review, merge) posts a machine-readable FORGE: annotation to the issue/PR. Downstream stages read what upstream stages wrote. When context compacts, the agent runs `gh issue view` and reconstructs full state from GitHub alone — no external memory layer, no proprietary database, no vendor lock-in. Unlike DeepWiki (which generates a documentation snapshot of your codebase once), ForgeDock writes structured annotations during each pipeline execution — so context reflects the current state of every open issue and in-flight decision, not a snapshot that goes stale. It's the LLM Wiki pattern (Karpathy's April 2026 gist), but applied to your entire development pipeline: issues are the wiki entries, FORGE: annotations are the structured knowledge, and GitHub's API is the query interface. The pipeline has 9 phases, a 9-agent specialist review system (security, database, auth, concurrency, etc.), and a 14-domain quality gate — all running as Claude Code slash commands. It's not a wrapper or thin CLI; the command specs are 50KB+ prompt engineering documents that constrain agent behavior at every step. Some technical details: - Zero runtime code — just markdown command specs symlinked into Claude Code - Install: `npx forgedock` - Uses `gh` CLI and `yq` as only dependencies - AGPL licensed, free forever for the core pipeline - Currently Claude Code only, investigating multi-model support Known limitations: - Claude Code only (no Cursor/Codex/Aider support yet) - Token-heavy on first run before adaptive scripts cache project patterns - The pipeline is opinionated — it enforces a specific workflow that may not fit every team I'm a solo developer building this to solve my own problems. Happy to answer questions about the architecture, the FORGE annotation protocol, or the prompt engineering behind the command specs. --- ## Posting Guidelines - **When**: Tuesday–Thursday, 9AM–12PM ET - **First 60 minutes**: Respond to every comment - **Tone**: Technical, honest, no marketing language - **If asked about monetization**: "Open-core model — the pipeline is AGPL and always free. Planning a commercial dashboard for pipeline observability." - **If compared to Devin**: "Devin is a cloud service; ForgeDock runs locally with your own API key. Different trade-offs — Devin is turnkey, ForgeDock gives you full control and traceability." - **If compared to DeepWiki**: "DeepWiki generates a documentation snapshot of your repo. ForgeDock writes structured annotations during each pipeline run — context stays current with your actual work, not a snapshot." - **If asked about scale**: Be honest about current adoption. Don't inflate numbers. --- ## Pre-Post Benchmark Gate (#1264) **Do not submit to HN until all items are checked.** - [ ] Cost accounting landed (#1255) - [ ] Topology fix landed (#1254) - [ ] Benchmark suite run (#1259) and scorecard published - [ ] Scorecard includes: cost per issue by difficulty class, one-shot success rate, before/after optimization wave - [ ] Every quantitative claim in the maker comment above traces to a row in the scorecard - [ ] `` comments in the draft have been replaced with measured figures - [ ] Reproducible method documented: demo repo + runner script so readers can re-run the benchmark