Blame for codebook.md

# Codebook: AI app cloud-deploy threads (GitHub issues)

One row = one public GitHub issue thread (issue + its comments) about deploying an AI app to a remote host.
Published rows contain **no text and no author information**: only the link, metadata and our coded columns.
Coded columns come from keyword rules (`scripts/code.py`). Some rows were read thread by thread by the AI assistant (Claude) that built the dataset: the review samples, and every row the keywords could not settle under the relevance rules (v4). Its judgements replace the keyword values (`relevance_method = agent_reviewed`, `coding_method = keyword+agent_reviewed`). No person has checked these rows yet.

## Identification and metadata (from the GitHub API)

| column | type | definition |
|---|---|---|
| `url` | string | link to the issue on github.com (the only way to read the text, which belongs to its authors) |
| `source` | string | always `github_issues` in this version |
| `repo` | string | `owner/name` of the repository |
| `repo_group` | enum | why the repo was searched: `host` (deploy platform CLI), `framework` (AI app framework or AI app), `framework_generic` (general framework: Streamlit, Gradio; the thread itself must show an AI app), `sample` (cloud vendor AI sample app), `builder` (AI app builder or coding agent) |
| `created_at` | ISO 8601 UTC | issue creation |
| `closed_at` | ISO 8601 UTC or empty | last close time |
| `state` | enum | `open` / `closed` at collection time |
| `state_reason` | enum or empty | GitHub's own value: `completed`, `not_planned`, `duplicate`, `reopened`, empty |
| `n_comments` | int | comment count at collection time |
| `n_reactions` | int | total reactions on the issue body |
| `time_to_first_response_h` | float or empty | hours from creation to the first comment by someone other than the author, bots excluded; empty = no such comment |
| `time_to_close_h` | float or empty | hours from creation to `closed_at` |
| `maintainer_replied` | yes/no | at least one comment by a user whose `author_association` is OWNER, MEMBER or COLLABORATOR (other than the author) |
| `collected_on` | date | snapshot date |

## Coded columns

| column | values | rule (summary) |
|---|---|---|
| `thread_type` | `problem`, `feature`, `docs`, `question`, `pr_ci`, `other` | what the thread is. `problem`: something breaks or blocks; `feature`: asks for a new capability; `docs`: asks for or fixes documentation; `question`: how-to, nothing broken yet; `pr_ci`: a code change, release or CI/test pipeline (issues only in this version, so "PR" means an issue about one); `other`: none of these (measurements, discussions). Rule, first match wins: a title tag (`[CI]`, `test:`, `[Docs]`, `docs:`, `[Feature]`, `feat:`, `[RFC]`, `[Question]`, `[Usage]`, `[Bug]`…); CI/test words in the title or a CI label; docs words in the title (docs, documentation, README, guide, tutorial) with no error word, or a documentation label without a bug label; a feature label (`enhancement`, `feature`, `rfc`, `integration request`) or a feature title pattern (`Add support…`, `Support for…`, `Allow…`, `option to…`) with no error word; an error word (error, fails, broken, cannot, 4xx/5xx, timeout…) or a bug label; a question pattern (`How to…`, `Is it possible…`, `…?`) or a question/support label; an error word in the body; else `other`. Agreement with the reviewer on a fresh sample: 49/60 (82%); most misses are keyword `other` that the reviewer reads as `feature` or `question` |
| `target_host` | `vercel`, `netlify`, `cloudflare`, `fly`, `railway`, `render`, `heroku`, `supabase`, `hf_spaces`, `hf_inference_endpoints`, `streamlit_cloud`, `posit_connect`, `langgraph_platform`, `azure`, `gcp`, `aws`, `digitalocean`, `runpod`, `modal`, `replicate`, `kubernetes` (generic, any cloud), `vps_docker` (self-managed server, Docker, Coolify-like panels), `unspecified` | first host named in the title, else in the body, else the repo's own platform (e.g. `vercel/vercel` → `vercel`, Azure samples → `azure`) |
| `app_type` | `chatbot`, `rag`, `ai_generated_web_app`, `model_serving`, `agent`, `other` | agent cues > RAG cues > builder repo > serving cues > chat cues; repo defaults for single-purpose repos |
| `framework` | one of `streamlit`, `gradio`, `chainlit`, `nextjs`, `remix`, `sveltekit`, `nuxt`, `vite_react`, `astro`, `fastapi`, `flask`, `django`, `express_node`, `hono`, `langgraph`, `langchain`, `llamaindex`, `vercel_ai_sdk`, `flowise`, `langflow`, `open_webui`, `nextchat`, `chatbot_ui`, `cog`, `tgi`, `ollama`, `vllm`, `dify`, `ragflow`, `anythingllm`, `hf_chat_ui`, `litellm`, `mastra`, `google_adk`, `crewai`, `agno`, `skypilot`, `prompty`, `unspecified` | the repo's framework if it is one, else first framework named. In builder repos the framework of the generated app is usually not stated, so `unspecified`; self-hosted bolt.diy is `remix` |
| `stack` | `;`-list of the values above | every framework named (repo framework first) |
| `model_provider` | `;`-list (max 3) of `azure_openai`, `openai`, `anthropic`, `google`, `aws_bedrock`, `local_ollama`, `huggingface`, `mistral`, `deepseek`, `groq`, `openrouter`, `replicate`, `cohere`, or `unspecified` | providers named in title or body; repo default for single-provider samples |
| `topic` | `build`, `runtime_dependencies`, `provisioning_iac` (cloud template / infrastructure provisioning, e.g. `azd up`, Bicep, Terraform), `memory_model_size`, `secrets_api_keys`, `gpu`, `costs_quota`, `cold_start`, `timeouts`, `database`, `auth_permissions`, `networking`, `streaming`, `routing_config`, `agent_behavior` (a coding or deploy agent itself causes the problem: deploys unasked, claims a deploy worked, deploys to the wrong target, or is blocked by its own safety check; `error_type` says which), `observability` (monitoring, tracing and telemetry tooling: Langfuse, OpenTelemetry/OTLP, Sentry, Prometheus, Grafana, Datadog, metrics, ServiceMonitor; plain logs do not count), `performance_scaling` (slowness, latency, throughput, concurrency, autoscaling, replicas, high availability), `other` | what the thread is about, whatever its `thread_type` (a failure, a feature request, a question or a docs request). Called `problem_category` before 2026-10-30. From the `error_type` found in the title; else a topical cue in the title (GPU, cold start, database, auth, cost); else the `error_type` found in the body; `coding_confidence` says which. Only when all of these give `other`: the three topics added on 2026-10-30, title first, then body (`agent_behavior` needs an agent context: the Claude Code repo, or the text names an agent such as Claude, Codex, Copilot, Cursor; `observability` and `performance_scaling` use stricter terms in the body, on its text without code blocks). So they never take a row from the 14 older topics, and `coding_confidence` does not change for them (it stays `none` when no error or older cue was found). Else `other` |
| `error_type` | `secret_exposed` (credentials deployed or published by mistake: a credential word such as secret, API key, password, `.env` must sit next to a leak/expose/publish word; memory leaks are `oom`, exposed ports `port_binding`), `iac_template_error`, `oom`, `size_limit` (incl. disk/volume full), `timeout`, `missing_env_var`, `quota_billing`, `permission_rbac`, `port_binding`, `module_dependency`, `build_failure`, `cors_network`, `streaming`, `not_found_routing`, `server_error_5xx`, `agent_deployed_unasked` (the agent deploys or pushes to production without permission or confirmation, or against instructions), `agent_false_success` (the agent says a deploy worked, was verified or happened when it did not), `agent_wrong_target` (the agent deploys to the wrong project, environment, account, repo or branch), `agent_blocked_refused` (an auto-mode classifier, safety filter or the model blocks or refuses a legitimate, user-authorised deploy: a false positive), `other_unknown` | first matching pattern in the title, else in the first 3,000 characters of the body. For provisioning failures the inner cause (quota, permission) wins when stated. The four `agent_*` types (added 2026-10-30) come with `topic = agent_behavior` only; they replace `other_unknown`, or a `server_error_5xx` found only in the body |
| `resolution_type` | `unanswered`, `self_resolved`, `config_fix`, `version_upgrade`, `version_downgrade`, `workaround`, `platform_limit_no_fix`, `user_error_config`, `duplicate`, `stale_closed`, `open_unresolved`, `closed_unclear` | from replies (non-author, non-bot), labels and `state_reason` |
| `resolved` | yes/no | closed as `completed` (not by a stale bot), or the author says it works / is solved |
| `ai_generated_app` | yes/no/unknown | `yes` if the thread is in an AI app builder or coding-agent repo, or the text says the app was built/generated with an AI tool. `no` only when the thread is about self-hosting the builder itself (e.g. deploying bolt.diy). Otherwise `unknown`: people rarely say they wrote the app by hand |
| `ai_tool` | `lovable`, `bolt`, `v0`, `cursor`, `claude_code`, `replit`, `copilot`, `windsurf`, `dyad`, `codex`, or empty | tool named, or repo default |
| `coding_method` | `keyword` / `keyword+agent_reviewed` | `keyword+agent_reviewed`: the AI reviewer read the thread and corrected the codes (host, app type, framework…) |
| `coding_confidence` | `title` / `body` / `none` | where the error keywords or the older topic cues were found; `title` rows agree best with the review. The three topics added 2026-10-30 do not set it |
| `relevance_method` | `keyword` / `agent_reviewed` | who decided the row is relevant (and its `thread_type`): the keyword rules, or the AI reviewer (Claude), not a person |

## Relevance (rules v4, decided by the dataset owner on 2026-10-28)

A thread is **relevant** if it is about deploying or running an AI app on a host (cloud AND AI). Relevance is `yes` or `no`; the earlier `adjacent` level is gone, and feature and docs requests stay in, told apart by `thread_type`.

1. **Repo context counts.** A Kubernetes or Docker thread in an AI app's repo (Dify, Langflow…) is relevant even if the thread never mentions AI.
2. **A host is a remote machine:** public cloud, VPS, company server, Kubernetes. Docker on a laptop, WSL or a local GPU is not.
3. **Model serving on a host** (vLLM, TGI, Ollama on a remote GPU, a model deployed on Azure) is relevant; `app_type = model_serving`.
4. **An agent that deploys and hits its own limits** (Claude Code blocked by a classifier, permission denied, a false "it's deployed") is relevant; `ai_generated_app = yes`.
5. **Feature and docs requests stay in**, with `thread_type` saying what they are.
6. **An app bug seen on a server but not caused by the deploy, the config or the host runtime is not relevant** (a model looping on a k8s GPU, a UI bug after an update). There must be a link to the host.
7. **Calling a cloud's managed AI API without deploying your own app is not relevant** (Vertex batch, Bedrock, Azure OpenAI calls, "deployment name" settings, LiteLLM router "deployments").
8. **General frameworks** (Streamlit, Gradio, Next.js, host CLIs): the thread must show an AI app (model, LLM, RAG…); the repo alone is not enough. Rule 1 applies to AI-specific repos only.
9. **Tests or CI of a framework's cloud integration** are relevant; `thread_type = pr_ci`.

**How the rules are applied.** Keywords first (`relevance()` in `scripts/code.py`): the title must name a hosting platform or a deploy verb, plus the exclusions above (local cues, managed-API names without a hosting term, router senses of "deployment", CI without a cloud named, Claude Code threads about the desktop app, browser pane, OAuth or connectors). Where the keywords cannot settle a row they set `kw_unsure` (private file only): Claude Code threads whose deploy context is unclear, model-serving threads that name no host, app-behaviour titles with no host cue (rule 6), threads whose only host cue is "in production", managed-API titles that also say "deploy". Every such row was judged by the AI reviewer (`manual/judgement_v4.csv`).

Only rows coded relevant are published.