Initial release: aircontext

templates/README.md

@@ -0,0 +1,26 @@
+# AirContext (per-project)
+
+This directory configures AirContext for **this project**. It was created the
+first time you ran `aircontext` here.
+
+## Files
+
+- `config.yaml` — backend (LLM), trigger threshold, compaction options
+- `rules.md` — compression rules sent to the LLM as system prompt
+- `state.json` — runtime state (do not edit; managed by the plugin)
+- `snapshots/` — backup of each JSONL before compaction; keep or delete freely
+
+## Getting started
+
+1. Open `config.yaml`, set `backend.endpoint` / `backend.model` / `backend.api_key`.
+   The default targets DeepSeek; replace with Ollama or any OpenAI-compatible
+   server as needed. For Ollama set `endpoint: http://localhost:11434/v1` and
+   any non-empty `api_key`.
+2. Tune `trigger.threshold` (default 0.6 = compact at 60% of model context).
+3. Edit `rules.md` to bias summaries toward what your project considers important.
+4. Re-run `aircontext` from this directory.
+
+## Disabling auto compaction temporarily
+
+Run `/aircontext-pause` inside Claude Code, or set `safety.dry_run: true` in
+config.yaml.

templates/config.yaml

@@ -0,0 +1,44 @@
+# AirContext compaction config.
+#
+# YOU MUST FILL these two fields (privacy-sensitive, never auto-populated):
+#   - backend.endpoint   your LLM endpoint URL
+#   - backend.api_key    either a literal value, or the ${env:VAR} placeholder
+#                         pointing at an env var that holds the secret
+#
+# Everything else is auto-configured by `aircontext` on first run
+# (model_context_window inferred from your Claude model setting, etc.) and
+# you generally don't need to touch it. Edit freely if you want to override.
+# `${env:VAR}` placeholders are resolved at runtime against environment
+# variables (including those declared in ~/.claude/settings.json `env` block).
+
+backend:
+  type: anthropic_native                           # or openai_compat (auto-set by aircontext when possible)
+
+  # >>> REQUIRED — fill before re-running aircontext <<<
+  endpoint: ""                                     # e.g. https://api.anthropic.com  |  https://wolfai.top  |  http://localhost:11434/v1
+
+  # >>> REQUIRED — fill or set the env var <<<
+  api_key: ${env:ANTHROPIC_AUTH_TOKEN}             # change to ${env:OPENAI_API_KEY} or paste a literal value
+
+  model: claude-haiku-4-5-20251001                 # cheap+fast for compression; raise to claude-sonnet-4-6 if quality insufficient
+  max_output_tokens: 4000
+  timeout_seconds: 60
+  anthropic_version: "2023-06-01"                  # only used when type = anthropic_native
+
+trigger:
+  strategy: token_ratio
+  threshold: 0.6                                   # compact when active chain reaches 60% of model_context_window
+  cooldown_seconds: 300
+  estimate_method: char_div_3_5
+  model_context_window: 200000                     # auto-overridden on first aircontext run based on your claude model
+
+compaction:
+  preserve_tail_messages: 10
+  drop_tool_results_over_lines: 1000
+  rules_file: rules.md
+  continuation_prompt: "基于上面的压缩摘要继续之前的工作；如果没有进行中的任务则等待我的下一条指令。"
+
+safety:
+  backup: true
+  max_snapshots: 50
+  dry_run: false

templates/rules.md

@@ -0,0 +1,30 @@
+# AirContext Compression Rules
+
+This file is fed to the compression LLM as part of its system prompt. Edit
+freely to bias what the summary keeps versus drops. The defaults below favour
+software-engineering sessions; rewrite them for your domain.
+
+## Always preserve verbatim
+
+- File paths read or modified, with the final intended state of each file
+- Architecture decisions and the reasoning behind them
+- Open TODOs, unresolved bugs, error messages still in scope
+- The user's stated goal for the current session
+- Any user-supplied facts that the model could not derive from the codebase
+  (credentials hints, deployment quirks, deadlines, "we tried X and it failed because Y")
+
+## Aggressively drop
+
+- Exploratory grep/glob results that did not lead anywhere
+- File contents that were superseded by later edits
+- Tool outputs over 1000 lines (keep first 50 lines and last 50 lines, summarise the middle)
+- Repeated similar searches and their near-identical outputs
+- Completed sub-steps whose only output was "looks good, moving on"
+
+## Output format
+
+- Plain prose, no markdown headers or bullet lists unless they materially aid recall
+- Reference files by `path:line` when relevant
+- One paragraph per topic; aim for under 3000 tokens total
+- Do NOT speculate beyond what the conversation contains
+- Do NOT apologise, summarise the act of summarising, or add meta-commentary