@holt-os/holt 0.8.4 → 0.9.0

package/README.md

@@ -90,6 +90,7 @@ holt memory                    # stats for this folder
 holt memory facts              # show the distilled facts (facts.md)
 holt memory search <query>     # find remembered moments
 holt memory embed              # embed older moments for semantic recall
+holt memory global             # share high-value facts across your folders
 holt memory clear              # wipe this folder's memory
 ```
 
@@ -97,6 +98,30 @@ Turns saved before semantic memory was enabled are upgraded in one pass with `ho
 
 Long conversations stay cheap: only recent turns are replayed verbatim, older context comes back through recall.
 
+### Optional global memory (`holt memory global`)
+
+By default every folder's memory is **isolated**: recall never crosses folders. If you want a folder to draw on what you learned elsewhere, opt it into a shared store:
+
+```bash
+holt memory global on          # this folder now contributes + reads shared facts
+holt memory global status      # who is sharing, and global store stats
+holt memory global off         # stop sharing + reading (recall goes local-only)
+holt memory global off --purge # also delete this folder's rows from the store
+```
+
+What opting in does:
+
+- **Facts only.** Only distilled facts (the `fact` rows, which includes wiki pages) are promoted, never raw turns. Each promoted fact is **tagged with the absolute path of the folder it came from**, so you always know its origin.
+- **One switch = contribute + read.** An opted-in folder both pushes its facts to the shared store and, during recall, also scores the shared store and merges the hits, so `holt memory search` and chat recall surface knowledge from your *other* opted-in folders. Your own folder's rows are excluded from the global read (they are already in local memory, so there is no double counting). Global hits are shown tagged with their source folder.
+- **`on` backfills** the folder's existing facts, then new facts mirror automatically. Saving the same fact twice does not duplicate it (dedup by normalized content + folder).
+
+Everything lives outside your projects, so a folder that never opts in is completely unaffected:
+
+- **Store:** `~/.holt/global/turns.jsonl`, one row per promoted fact plus a `workspace` field.
+- **Registry:** `~/.holt/memory-scopes.json`, shape `{ "enabled": ["<abs folder path>", ...] }`. This is the whole list of folders that participate. Per-folder `config.json` is not touched.
+
+A missing or corrupt global store degrades silently to local-only recall. Per-folder isolation stays the default.
+
 ## Brains
 
 A brain is an agent CLI installed and logged in on your machine. No API keys to paste.
@@ -122,7 +147,39 @@ A brain can also be a direct provider connection: no CLI install, your own key.
 - Turns that mean similar things (when local embeddings exist) get a strong link.
 - Drag to pan, scroll to zoom, click a turn to read it in full, click a concept to light up its turns, and search to highlight matching memory live.
 
-Options: `--out <path>` to write elsewhere, `--no-open` to skip the browser.
+Options: `--out <path>` to write elsewhere, `--no-open` to skip the browser. When a wiki exists (see below), the graph auto-includes wiki pages as green nodes wired by their `[[links]]`; force it on or off with `--wiki` / `--no-wiki`. This is the graph as a view over your wiki.
+
+## Knowledge wiki: `holt wiki`
+
+Memory has three layers, each derived from the one below it:
+
+1. **Raw turns** (`turns.jsonl`): every exchange, append-only, authoritative.
+2. **Facts** (`facts.md`): 1 to 5 durable facts distilled per session.
+3. **Wiki** (`.holt/wiki/`): a cross-linked set of Markdown pages, synthesized and maintained by an LLM. This is the "kept and connected" layer.
+
+The core rule: **wiki pages are derived and regenerable, never a sole source of truth.** Turns and facts stay authoritative. Every page records which fact ids it drew from (frontmatter `sources:`), so `holt wiki rebuild` can regenerate the whole wiki from scratch. A bad synthesis is therefore never lossy: rebuild recovers it. Pages are plain Obsidian-compatible Markdown with `[[wikilinks]]` and an `index.md`, so you can point Obsidian at `.holt/wiki` as a vault, or edit pages by hand (a rebuild will overwrite hand-edits, so keep the folder under git if you edit).
+
+```bash
+holt wiki                     # status: maintainer, model, page count, last sync, RAM hint
+holt wiki sync                # fold new facts into pages (route + merge)
+holt wiki rebuild             # wipe and regenerate every page from facts (asks first)
+holt wiki lint                # audit pages for contradictions, duplicates, gaps
+holt wiki list                # list pages (title, updated, size)
+holt wiki show <page>         # print one page
+holt wiki open                # open the wiki in your default app (Obsidian reads it natively)
+holt wiki setup               # recommend a local model for this machine's RAM
+```
+
+**How sync works.** `holt wiki sync` gathers facts added since the last sync (a marker in `.holt/wiki/.state.json`), embeds each one locally, and routes it to the nearest existing page by cosine similarity. Related facts land on the same page; a fact that matches nothing new starts a fresh page. Facts are grouped by target page and the maintainer is called **once per changed page** (not once per fact), so it only rewrites the few pages that actually changed. Routing is always local and free, which is what keeps the maintainer's cost minimal. `holt wiki rebuild` folds every fact from scratch; `holt wiki lint` reports issues but does not rewrite (propose-only).
+
+**Who maintains it (the key knob).** Synthesis is a reasoning task, and you choose who does it with `wiki.maintainer`:
+
+- `brain` (default): the folder's configured brain, exactly like fact distillation. When the brain is the Claude Code CLI this rides your existing Claude plan, so there are no marginal dollars per sync.
+- `local`: a local generative model via Ollama (set `wiki.localModel`, default `qwen2.5:7b`). Free and offline, lower quality, and it trades dollars for RAM. If the model is not pulled, Holt prints the exact `ollama pull` line and falls back to the brain rather than failing.
+
+`holt wiki setup` (or the hint in `holt wiki status`) reads your total RAM and recommends a model: under 16 GB it discourages local (use the brain); 16 GB suggests `qwen2.5:7b` (tight alongside the embed model, an always-on machine is a better host); 24 to 32 GB suggests `qwen2.5:14b`; 48 GB and up suggests `qwen2.5:32b`.
+
+Wiki pages also participate in recall: each page is embedded and indexed like a high-value fact, so `holt memory search` and chat recall can surface synthesized knowledge, not just raw turns.
 
 ## Skills
 
@@ -173,6 +230,28 @@ claude mcp add holt -- holt mcp
 
 Point the client at the folder whose memory you want it to use; Holt serves whatever folder the server process starts in.
 
+## Ambient memory for Claude Code (`holt hook`)
+
+The MCP server above lets Claude Code recall and remember **when it decides to**. `holt hook` goes further: it makes Holt's per-folder memory work **ambiently**, with no `holt chat` and no manual tool call. Two directions, both wired as [Claude Code hooks](https://docs.claude.com/en/docs/claude-code/hooks):
+
+- **Inject** (`UserPromptSubmit`): before each prompt, Holt quietly recalls the most relevant remembered notes for the current folder and adds them to the model's context.
+- **Capture** (`Stop`): when a session ends, Holt distills durable facts from the transcript (via this folder's configured brain) and saves them to the folder's memory.
+
+```bash
+holt hook install                 # install both hooks into ~/.claude/settings.json
+holt hook install --inject-only   # only the before-prompt recall hook
+holt hook install --capture-only  # only the end-of-session fact hook
+holt hook install --project       # write ./.claude/settings.json instead of the global one
+holt hook status                  # show what is installed and which directions are active
+holt hook remove                  # remove ONLY Holt's hooks, leaving everything else intact
+```
+
+Install **merges** into any existing `hooks` config, never clobbers unrelated settings or other hooks, is **idempotent** (re-running does not duplicate), and backs the file up to `settings.json.holt-bak` before writing. `remove` deletes only the entries whose command is `holt hook inject` / `holt hook capture`.
+
+**Trusted-folder guard (important).** The two runtime hooks (`holt hook inject` and `holt hook capture`, invoked by Claude Code, not you) no-op **silently** unless the folder is a trusted Holt workspace with an existing `.holt/memory`. So Holt never injects your private notes into an unrelated project, and never creates memory in a folder you never set up. To make a folder ambient, run `holt init` (or `holt chat`) there once so it becomes trusted and gets its `.holt/memory`. The inject hook keeps stdout clean (only the context block is printed, since Claude Code injects it verbatim); all diagnostics go to stderr, and both hooks always exit 0 so a memory step never blocks or slows your prompt.
+
+To uninstall, run `holt hook remove` (add `--project` if you installed with `--project`).
+
 ## Output format
 
 Replies print as markdown. `/output html` (or `markdown`) switches the save format and persists it. `/save [name]` writes the last reply to the current folder: `.md`, or a small self-contained dark-theme `.html` page.
@@ -216,6 +295,99 @@ holt run "daily brief" | holt notify   # pipe any output to your phone
 
 The token is stored in `~/.holt/telegram.json` (mode 600) and never printed in full.
 
+### holt routine
+
+A routine is a named, reusable job: it bundles the pieces above into one object you can run by name or on a schedule. A routine is a **task source** (an installed skill or an inline prompt) plus an optional **daily schedule** plus **output routing** (stdout, a file, or Telegram). It is the generic version of a recurring "agent" such as a daily brief.
+
+```bash
+# an inline task you can rerun by name
+holt routine add brief --task "summarize what changed and what is open here" --notify
+
+# a routine backed by an installed skill
+holt routine add triage --skill inbox-triage --out triage.md
+
+# a scheduled routine: installs an OS timer that fires "holt routine run digest --quiet"
+holt routine add digest --task "digest today's changes" --at 07:00 --notify
+
+# a built-in template (daily-brief or standup)
+holt routine add myday --template daily-brief
+
+holt routine run brief          # run it now and stream the reply
+holt routine run brief --out b.md   # also write the reply to a file this run
+holt routine list               # name, source, schedule, outputs, workspace
+holt routine show brief         # full detail
+holt routine remove brief       # deletes the routine and its OS timer, if any
+```
+
+Sources: give exactly one of `--skill <name>`, `--task "<prompt>"`, or `--template <t>`. A skill routine splices that skill's `SKILL.md` body into the prompt at run time; a task routine uses the prompt as-is.
+
+Scheduling: `--at HH:MM` (24h, daily) installs an OS timer through the same scheduler `holt schedule` uses (launchd on macOS, cron on Linux) whose command is `holt routine run <name> --quiet`. Omit `--at` for a manual, run-on-demand routine. Removing the routine removes its timer.
+
+Output routing: by default the result prints to stdout. `--out <file>` also writes it to a file (relative to the routine's workspace); `--notify` pushes it to Telegram (guarded cleanly when Telegram is not set up). Both `--out` and `--notify` can also be passed to `holt routine run` for a single run. In `--quiet` mode (what the scheduler uses) stdout is suppressed but `--out` and `--notify` still fire.
+
+Routines are stored in `~/.holt/routines.json` (one entry per routine, each carrying its absolute workspace). A routine with a schedule also produces an entry in `~/.holt/schedules.json`, keyed by the routine name; the two are kept consistent.
+
+## Machine check (`holt doctor`)
+
+Not sure how best to run Holt on your hardware? `holt doctor` looks at this machine and its installed tools and tells you, in plain language, what to do.
+
+```bash
+holt doctor
+```
+
+It reports, section by section:
+
+- **Machine**: platform, CPU, RAM (total and free), and Node version.
+- **Brains**: which agent CLIs (`claude`, `codex`, `gemini`) are installed, plus a recommendation (prefer an installed CLI brain; otherwise install one or add an API brain).
+- **Semantic memory**: whether a local Ollama with the embed model is reachable, and how to enable it (`ollama pull nomic-embed-text`) so recall is semantic and fully private.
+- **Knowledge wiki maintainer**: `brain` (the default, best quality, rides your existing plan) versus a local model sized to your RAM. On a 16 GB machine it names `qwen2.5:7b`; more RAM gets a larger model; under 16 GB it steers you to `brain`.
+- **Always-on / Telegram**: whether a Telegram bot is set up, and a note on hosting the bot on a low-power always-on machine while keeping heavy local models on a bigger one.
+- **Recommended next steps**: a short checklist built from whatever gaps it found.
+
+It is read-only advice: it changes nothing, needs no trust, and always exits cleanly even when a probe fails.
+
+The RAM-to-model table lives in `src/specs.ts` (`LOCAL_MODEL_RECS` / `recommendLocalModel`), the single source of truth the wiki maintainer reads too.
+
+## Your writing voice (`holt voice` + `holt write`)
+
+Holt can learn how *you* write, then draft in that voice while steering clear of the patterns that make text read as machine written.
+
+### Teach it your voice: `holt voice`
+
+At the end of `holt init` you are offered a short optional interview. You can also run or redo it anytime:
+
+```bash
+holt voice            # run the interview, or add samples, or both
+holt voice add <file> # learn from a writing sample you already have
+holt voice show       # print your profile
+holt voice edit       # show where the profile lives (it is a plain JSON file)
+holt voice clear      # remove it
+```
+
+The interview asks you to pick a **depth** up front: `quick` (a few questions) or `detailed` (more). Questions are about **writing and communication style only**: tone, sentence length, first vs third person, emoji habits, words you love or ban, who you write for, and what you want to sound like and not sound like.
+
+Privacy is a hard rule. The interview never asks for personal details (no name, job, location, or life story). If you happen to mention something personal in an answer, Holt stores only what you typed and never asks follow ups.
+
+You can also feed it real writing. `holt voice add <file>` reads a file; inside the interview you can paste a sample instead. Samples are stored as a hash and length by default; an excerpt is kept only if you say yes, so Holt can match your rhythm directly.
+
+From the answers and samples, Holt asks your configured brain to build a **style profile**: tone, formality (1 to 5), average sentence length, person, emoji and formatting habits, signature moves, banned words, target audiences, and a short "sounds like / does not sound like" summary. If no brain is set yet, your raw answers are saved and the profile synthesizes the next time you run `holt voice` with a brain configured.
+
+The profile lives at `~/.holt/voice.json` (mode 600, since it can hold writing excerpts). It is global to you, not per folder, and you can hand edit it.
+
+### Draft in your voice: `holt write`
+
+```bash
+holt write "a linkedin post about shipping Holt" --type linkedin
+holt write "a reply to a customer asking for a refund" --type email --out reply.txt
+holt write "a short post about testing" --fast
+```
+
+`holt write` composes three things into one prompt: your voice profile, a built in **anti-AI rubric**, and your request. Then it runs your default brain. `--type` shapes length and format (`linkedin`, `email`, `tweet`, `blog`, `generic`), `--out` also writes the draft to a file, and `--brain <id>` picks a brain.
+
+By default a second pass checks the draft against the rubric and fixes any tells; `--fast` skips it for a single call. The anti-AI rubric bans em-dashes, "rule of three" cliches, "in today's fast-paced world" openers, "it is not X, it is Y" reframes, tidy bow-tie conclusions, uniform sentence rhythm, hedging filler, corporate buzzwords, and emoji unless your profile allows them. The output is always em-dash free.
+
+With no profile yet, `holt write` still works using a plain, natural voice.
+
 ## Commands
 
 ```
@@ -223,12 +395,18 @@ holt init            set up (trust, brains, sign-in, defaults) for this folder
 holt chat            start a session that remembers past ones
 holt run <task>      run one task non-interactively (recall, brain, remember)
 holt schedule        run a task on a timer: add | list | remove
+holt routine         named, reusable, scheduled jobs: add | run | list | show | remove
 holt telegram        chat with Holt from your phone: telegram [setup]
 holt notify [msg]    push a message to your phone over Telegram (stdin-friendly)
+holt doctor          check this machine and recommend how best to run Holt here
+holt voice           teach Holt your writing voice: add <file> | show | edit | clear
+holt write <what>    draft content in your voice with anti-AI checks (--type, --out, --fast)
 holt graph           see your memory as an interactive knowledge graph
 holt mcp             run an MCP server so other tools use this folder's memory (holt mcp setup)
+holt hook            ambient memory for Claude Code: install | remove | status
 holt skill           manage skills: list | show | create | add | remove
 holt memory          inspect memory: holt memory [search <query> | facts | embed | clear]
+holt wiki            derived knowledge wiki: holt wiki [sync | rebuild | lint | list | show | status]
 holt setting         configure brains, API brains, and launch command
 holt login <brain>   sign in to claude, codex, or gemini
 holt version         print version
@@ -253,7 +431,8 @@ Built in always-shippable phases toward a full-vision v1:
 3. **Skills**: portable SKILL.md skills, create/add/run *(shipped)*
 4. **Knowledge graph**: see and navigate your own memory with `holt graph` *(shipped)*
 5. **Everywhere**: MCP server so other tools read Holt's memory, plus `holt run`, scheduling, and Telegram *(shipped)*
-6. **Next**: docs site, skill registry publishing, more channels
+6. **Knowledge wiki**: LLM-maintained, cross-linked, regenerable pages over your memory with `holt wiki` *(shipped)*
+7. **Next**: docs site, skill registry publishing, more channels
 
 ## Contributing