@jolli.ai/space-cli 0.99.5 → 0.99.7

package/README.md

@@ -14,11 +14,43 @@ Once both are installed, the following commands become available:
 - `jolli space` — Inspect Jolli auth state and select a space for sync (`status`, `switch`)
 - `jolli source` — Manage source repositories for impact analysis (`add`, `remove`, `list`)
 - `jolli sync` — Sync markdown files with the server (`pending`, `changeset`, `up`, `down`, `full`, `strip`)
-- `jolli impact` — Documentation impact-analysis tools (`extract`, `search`, `agent`)
+- `jolli impact` — Documentation impact-analysis tools: server-backed (`extract`, `search`, `agent`) and purely-local (`init`, `settings`, `local`, `attention` — see below)
 - `jolli agent` — Interactive LLM agent with local tool execution (`list`, `resume`)
 
 Without `@jolli.ai/space-cli` installed, the host CLI displays these commands as stubs that prompt the user to install this package.
 
+## Purely-local impact (no Jolli login)
+
+`jolli impact` also has a **fully local mode**: point it at a docs directory and one or more local source repos, bring your own LLM key, and it updates the affected docs in place over your new commits — **no Jolli login, no space, no server round-trip, no GitHub App.**
+
+```bash
+# Set up .jolli/settings.local.yaml
+jolli impact init                                    # auto-detect: current git repo = source, find a docs/ dir (within 2 levels)
+jolli impact init --docs ./docs --source ../backend  # or specify explicitly (repeat --source for multiple repos)
+
+# Configure an LLM — default is Anthropic + claude-sonnet-4-6, so just set a key
+export ANTHROPIC_API_KEY=sk-...                      # or OPENAI_API_KEY / GEMINI_API_KEY / JOLLI_LLM_API_KEY
+jolli impact settings --provider openai --model gpt-4o   # only to use a NON-default provider/model
+jolli impact settings --set-key                      # or store the key via a hidden prompt (never in shell history)
+jolli impact settings --max-tokens-per-turn 16384    # raise the per-turn output budget if edits get truncated (default 8192, range 4096..200000)
+
+# Run impact over your new commits
+jolli impact local                                   # edit docs affected by new commits, in place
+jolli impact local --uncommitted                     # preview against the working tree
+jolli impact local --dry-run                         # list impacted docs only (no writes, no LLM, no key needed)
+jolli impact local --thinking                        # stream the model's reasoning above the live panel (default: off)
+
+# Attention = the document -> source-file mapping that impact matches against
+jolli impact attention                               # generate / update the cache
+jolli impact attention --refresh                     # rebuild from scratch
+jolli impact attention --show                        # view the current mapping
+```
+
+- A **source must be a git repository** (impact diffs it); the **docs directory does not** — it can be a standalone folder, its own repo, or a subfolder of the source.
+- **Doc file types**: defaults to `.md` / `.mdx` / `.json`. Pass `--doc-extensions .md,.yaml` to `init` to stamp a different list at init time, or hand-edit the `docExtensions` field in `.jolli/settings.local.yaml` later. Replace semantics — the configured list replaces the defaults.
+- Providers: `anthropic` | `openai` | `gemini` | `openai-compatible`. LLM config + key live in `.jolli/llm.local.env` (git-ignored) or the environment.
+- Generated mappings are cached in `.jolli/attention.local.yaml` and are **never written back into your docs**. `init` adds `.jolli/` to your global git-ignore.
+
 ## Architecture
 
 This package is a **commander-shaped shell** that the host CLI loads in-process: when both packages are installed, `@jolli.ai/cli`'s plugin loader calls this package's `register(ctx)` and it extends the host's commander `program` with the commands above. It does **not** value-import `commander` or `@jolli.ai/cli/api` — both are provided by the host process at load time and are therefore externalized from the bundle.