ideation: close the research arc into a loop — tend, atlas, modeling + agent-facing hardening

README.md

@@ -64,6 +64,8 @@ The harness ships a marketplace of [plugins](plugins/) (Claude Code / Pi skills,
                           └── a Gap becomes ──┘
                               a hypothesis
         trail ── records goal · hypothesis · alternatives · outcome (git trailers)
+        atlas ── indexes every artefact; its frontier seeds the next run (the loop)
+        tend  ── runs the checks on a cadence and self-heals (keeps it running)
 ```
 
 | Stage | Command | What it does |
@@ -73,6 +75,9 @@ The harness ships a marketplace of [plugins](plugins/) (Claude Code / Pi skills,
 | **do** | `/investigate` | *Run* the question, don't read about it. **Frame** it into 2-5 falsifiable hypotheses, **fan out** one contained agent per hypothesis (the swarm), **judge** the outcomes adversarially against their evidence, **record** everything to the trail. |
 | **show** | `/present` | Turn a result into a served, visual presentation: pick the medium per finding (Manim animation · marimo app · static figure · served notebook), render on the right compute, reachable over Tailscale. |
 | **trail** | `trail` | The decision graph: hypotheses, choices, roads not taken, and outcomes recorded as git-commit trailers. Reconstructs from `git log` alone (no LLM), and renders as a B&W graph + tempo timeline with open hypotheses flagged as awaiting a verdict. |
+| **do (quantitative)** | `/modeling` | *Do* the quantitative science: turn data (a file, a simulation, or a digitized figure) into a fitted, model-selected, uncertainty-quantified law: candidate models, parameter covariance, AIC/BIC + cross-validation, symbolic regression. The quantitative twin of `/investigate`; composes the Axiomatic tools. |
+| **compound** | `atlas` | The corpus that makes the arc *compound*: indexes every study, investigation, map and trail into one B&W page and collects their gaps and frontiers into one queue the next run pulls from. The agent arXiv. |
+| **continuity** | `tend` | The layer that turns the arc from a line into a *loop*: register checks that should keep passing, run them on a cadence, record drift, and (armed) spawn a contained agent to self-heal, so work stays healthy and the frontier re-opens without a human in the hot path. |
 
 ```bash
 /study  arxiv.org/abs/2601.06712        # understand: a paper → dossier in the atlas
@@ -82,7 +87,7 @@ trail                                    # see the decision graph reconstruct fr
 /present                                 # show: render the surviving result, serve it
 ```
 
-`/delve` conducts understand→show in one shot; `/map` and `/explore` render a whole repo into a one-page dossier. The full set lives in **[`plugins/`](plugins/)**: `science-writing` (verify-citations, peer-review, rebuttal, arxiv-prep), `tikz`, `manim`, `marimo`, and `writing-styles`.
+`/delve` conducts understand→show in one shot; `/map` and `/explore` render a whole repo into a one-page dossier. `atlas` indexes the whole corpus so the frontier of one run seeds the next, and `tend` keeps the checks healthy over time; together they close the arc into a loop. The full set lives in **[`plugins/`](plugins/)**: `science-writing` (verify-citations, peer-review, rebuttal, arxiv-prep), `tikz`, `manim`, `marimo`, `writing-styles`, and `modeling` (quantitative fits).
 
 ---
 
@@ -345,7 +350,7 @@ When `anu init` finds an existing config it offers three strategies: **merge** (
 <details>
 <summary><b>State & data</b></summary>
 
-anu exposes the repo at `~/.local/share/anu/` and stores runtime state there: `swarms/` (metadata, mailboxes), `reviews/` (cached summaries per SHA), `mesh/` (device cache), `box/` (contained-agent state; `box/claude` holds credentials, gitignored). The **atlas** at `~/.anu/atlas/<repo>` holds dossiers and investigations; the decision **trail** at `~/.anu/trail/<repo>`. The installer link manifest lives at `~/.local/state/anu/manifest` so `anu unlink` restores configs cleanly. All runtime state is gitignored; never commit it.
+anu exposes the repo at `~/.local/share/anu/` and stores runtime state there: `swarms/` (metadata, mailboxes), `reviews/` (cached summaries per SHA), `mesh/` (device cache), `box/` (contained-agent state; `box/claude` holds credentials, gitignored). The **atlas** at `~/.anu/atlas/<repo>` holds dossiers and investigations, with `atlas` rendering one index over all of it at `~/.anu/atlas/index.html`; the decision **trail** at `~/.anu/trail/<repo>`; and `tend` keeps watch state under `~/.local/share/anu/tend/`. The installer link manifest lives at `~/.local/state/anu/manifest` so `anu unlink` restores configs cleanly. All runtime state is gitignored; never commit it.
 
 </details>
 

config/bash/bin/swarm

@@ -1,5 +1,10 @@
 #!/usr/bin/env bash
 # Standalone wrapper so `swarm` is callable as a command (not just a function)
 # This lets AI agents inside Claude Code run `swarm send`, `swarm collect`, etc.
-source "${ANU_PATH:-$HOME/.local/share/anu}/config/bash/fns/swarm"
+ANU="${ANU_PATH:-$HOME/.local/share/anu}"
+# swarm() guards its deps through core (_anu_require); source core first, exactly
+# as bin/tend does, or every non-help subcommand dies here with
+# "_anu_require: command not found" on this non-interactive path.
+source "$ANU/config/bash/fns/core"
+source "$ANU/config/bash/fns/swarm"
 swarm "$@"

config/bash/bin/tend

@@ -0,0 +1,11 @@
+#!/usr/bin/env bash
+# Standalone wrapper so `tend` is callable as a command, not just a shell
+# function, needed by the cron/launchd heartbeat (`tend cron on`), which runs
+# in a shell that has not sourced the anu fns. Mirrors bin/box, bin/delve, bin/swarm.
+ANU="${ANU_PATH:-$HOME/.local/share/anu}"
+# cron/launchd runs with a minimal PATH; make anu's tools resolvable (box/cxc for
+# the heal, jq, homebrew bins) so the headless heartbeat and its healer can run.
+export PATH="$ANU/config/bash/bin:$HOME/.local/bin:/opt/homebrew/bin:/usr/local/bin:$PATH"
+source "$ANU/config/bash/fns/core"
+source "$ANU/config/bash/fns/tend"
+tend "$@"

config/bash/fns/atlas

@@ -0,0 +1,50 @@
+# ==============================================================================
+# atlas: one word, render the corpus index across all anu research artifacts.
+#
+# Every /study, /investigate, /map and trail render lands under ~/.anu/atlas
+# (and ~/.anu/trail). Each is a one-off today. `atlas` indexes them into one
+# fixed B&W page and collects their gaps, open questions and frontiers into a
+# single queue, so finished work exposes what it left open and the next
+# investigation pulls its question from there. The corpus compounds; the agent
+# arXiv. No LLM: the index IS the on-disk JSON (build.py + render.py).
+#
+#   atlas         build + render + open the corpus index
+#   atlas open    re-open the last index without rebuilding
+#   atlas ls      list the corpus records in the terminal
+#
+# The skill `atlas` (plugins/atlas) teaches agents to consult the frontier
+# before starting work and to leave their own gaps behind when they finish.
+# ==============================================================================
+
+_atlas_render() {  # <atlas_root> <trail_root> <skill>
+  python3 "$3/build.py" "$1" "$2" "$1/atlas.json" >/dev/null &&
+    python3 "$3/render.py" "$1/atlas.json" "$1/index.html" >/dev/null
+}
+
+atlas() {
+  local atlas_root="$HOME/.anu/atlas" trail_root="$HOME/.anu/trail" skill
+  skill="${ANU_PATH:-$HOME/.local/share/anu}/plugins/atlas/skills/atlas"
+
+  case "${1:-}" in
+    open)
+      if [[ -f "$atlas_root/index.html" ]]; then
+        open "$atlas_root/index.html" 2>/dev/null || xdg-open "$atlas_root/index.html" 2>/dev/null
+      else
+        echo "no atlas yet. Run: atlas"
+      fi
+      return ;;
+    ls)
+      _anu_require python3 jq || return 1
+      mkdir -p "$atlas_root"
+      [[ -f "$atlas_root/atlas.json" ]] || _atlas_render "$atlas_root" "$trail_root" "$skill" >/dev/null
+      jq -r '.records[] | "  [\(.kind)]\t\(.title)"' "$atlas_root/atlas.json" 2>/dev/null \
+        || echo "no atlas yet. Run: atlas"
+      return ;;
+  esac
+
+  _anu_require python3 || return 1
+  mkdir -p "$atlas_root"
+  _atlas_render "$atlas_root" "$trail_root" "$skill" || { _anu_die "atlas: build failed"; return 1; }
+  echo "atlas → $atlas_root/index.html"
+  open "$atlas_root/index.html" 2>/dev/null || xdg-open "$atlas_root/index.html" 2>/dev/null || true
+}

config/bash/fns/core

@@ -0,0 +1,48 @@
+# ==============================================================================
+# core: shared primitives for anu shell functions.
+# ==============================================================================
+# Sourced into every interactive shell (via the config/bash/fns/* glob, before
+# the commands that call it run) and by the bash test harness. This is the one
+# place anu's diagnostic vocabulary lives, so an unattended agent fails loud
+# with a clear line instead of a cryptic error three calls deep: the single
+# most common way a full-auto cxc run silently dead-ends.
+#
+#   _anu_warn "msg"        # yellow  "anu: msg"  → stderr
+#   _anu_note "msg"        # dim     "anu: msg"  → stderr
+#   _anu_die  "msg"        # red     "anu: msg"  → stderr, returns 1
+#   _anu_require jq tmux   # 0 if all present; else names the missing tool(s)
+
+# Write a prefixed diagnostic to stderr. Colorized only when stderr is a tty,
+# so logs and pipes stay clean.
+_anu_msg() {
+  local level="$1"; shift
+  local color='' reset=''
+  if [[ -t 2 ]]; then
+    reset=$'\033[0m'
+    case "$level" in
+      err)  color=$'\033[31m' ;;
+      warn) color=$'\033[33m' ;;
+      *)    color=$'\033[2m'  ;;
+    esac
+  fi
+  printf '%sanu:%s %s\n' "$color" "$reset" "$*" >&2
+}
+
+_anu_warn() { _anu_msg warn "$@"; }
+_anu_note() { _anu_msg note "$@"; }
+
+# Print a diagnostic and return 1. Callers chain `|| return`.
+_anu_die() { _anu_msg err "$@"; return 1; }
+
+# _anu_require TOOL [TOOL...]: ensure each named command is on PATH. On the
+# first run with any miss, print one line listing every missing tool and return
+# 1. The choke point for "a dependency isn't installed": swarm/ncn route their
+# hard requirements through here so the failure names the tool, not a symptom.
+_anu_require() {
+  local tool missing=()
+  for tool in "$@"; do
+    command -v "$tool" &>/dev/null || missing+=("$tool")
+  done
+  (( ${#missing[@]} )) || return 0
+  _anu_die "missing required tool(s): ${missing[*]}"
+}

config/bash/fns/ncn

@@ -89,7 +89,7 @@ _nc_open() {
   local rcmd
   case "$profile" in
     box)               rcmd="box bash" ;;
-    cluster|apple|ssh) rcmd="ssh ${sshopts:+$sshopts }${target}" ;;
+    cluster|apple|ssh) _anu_require ssh || return 1; rcmd="ssh ${sshopts:+$sshopts }${target}" ;;
     *) echo "nc/ncn: unknown profile '$profile' (cluster|apple|ssh|box)"; return 1 ;;
   esac
 

config/bash/fns/swarm

@@ -93,6 +93,16 @@ _swarm_agent_ids() {
   done
 }
 
+# A one-line "available: a b c" hint for not-found diagnostics (empty if none).
+# Turns a mistyped agent id into a self-correcting message for an unattended
+# conductor instead of a dead end.
+_swarm_agents_hint() {
+  local ids
+  ids=$(_swarm_agent_ids "$@" 2>/dev/null | tr '\n' ' ')
+  ids="${ids% }"
+  [[ -n "$ids" ]] && printf 'available: %s' "$ids"
+}
+
 # Generate a short swarm ID
 _swarm_gen_id() {
   date +%s | shasum | head -c 8
@@ -635,7 +645,7 @@ _swarm_send() {
   swarm_dir=$(_swarm_dir) || { echo "swarm send: no active swarm in this window."; return 1; }
 
   # Validate target exists
-  [[ -f "$swarm_dir/agents/$target.json" ]] || { echo "swarm send: agent '$target' not found."; return 1; }
+  [[ -f "$swarm_dir/agents/$target.json" ]] || { _anu_die "swarm send: no agent '$target' in this swarm. $(_swarm_agents_hint "$swarm_dir")"; return 1; }
 
   local sender
   sender=$(_swarm_whoami)
@@ -655,10 +665,16 @@ TIME: $(date -u +%Y-%m-%dT%H:%M:%SZ)
 $message
 EOF
 
-  # Inject into pane (immediacy)
-  _swarm_send_keys "$target" "$message"
+  # Inject into pane (immediacy). The mailbox write above is the durable
+  # channel; report honestly if the live injection couldn't land.
+  local injected=0
+  _swarm_send_keys "$target" "$message" && injected=1
   command -v _agentlog_append &>/dev/null && _agentlog_append "$(_swarm_current_id 2>/dev/null)" "$target" "send" "[$sender→$target] $message"
-  echo "swarm: sent to $target"
+  if (( injected )); then
+    echo "swarm: sent to $target"
+  else
+    _anu_warn "queued to $target's mailbox; live pane injection failed (agent pane gone? check: swarm status)"
+  fi
 }
 
 # Broadcast message to all agents in the current swarm
@@ -732,7 +748,7 @@ _swarm_capture() {
   local swarm_dir
   swarm_dir=$(_swarm_dir) || { echo "swarm capture: no active swarm."; return 1; }
 
-  [[ -f "$swarm_dir/agents/$agent.json" ]] || { echo "swarm capture: agent '$agent' not found."; return 1; }
+  [[ -f "$swarm_dir/agents/$agent.json" ]] || { _anu_die "swarm capture: no agent '$agent' in this swarm. $(_swarm_agents_hint "$swarm_dir")"; return 1; }
 
   _swarm_capture_pane "$agent" > "$swarm_dir/results/$agent.md"
   echo "swarm: captured $agent → results/$agent.md"
@@ -746,7 +762,7 @@ _swarm_inspect() {
 
   local swarm_dir
   swarm_dir=$(_swarm_dir) || { echo "swarm inspect: no active swarm."; return 1; }
-  [[ -f "$swarm_dir/agents/$agent.json" ]] || { echo "swarm inspect: agent '$agent' not found."; return 1; }
+  [[ -f "$swarm_dir/agents/$agent.json" ]] || { _anu_die "swarm inspect: no agent '$agent' in this swarm. $(_swarm_agents_hint "$swarm_dir")"; return 1; }
 
   _swarm_colors
   local role device cmd
@@ -2882,6 +2898,11 @@ swarm() {
   local subcmd="${1:-help}"
   shift 2>/dev/null
 
+  # Fail loud if the hard runtime deps are missing (help still works without).
+  if [[ "$subcmd" != help && "$subcmd" != --help && "$subcmd" != -h ]]; then
+    _anu_require jq tmux || return 1
+  fi
+
   case "$subcmd" in
     start)              _swarm_start "$@" ;;
     star)               _swarm_star "$@" ;;