jfenal · Oct 10, 2026 · Oct 10, 2026 · Oct 10, 2026
@@ -69,6 +69,7 @@ Things that break silently if you guess wrong.
 - **Manual summarize works**: `POST /session/{id}/summarize` and TUI `/compact` run `Processor.Compact` and return `200` + `{"compacted": bool}` — not `501`.
 - **First-run is doctor + `/connect`**: Use `tinycode doctor` (diagnose) and TUI `/connect` (configure models). There is no `tinycode setup`. Legacy `/tc-doctor` bash skill is obsolete.
 - **Skills paths are plural**: `skills/`. Project agents live at `.tinycode/agent/*.md`.
+- **`go.mod` minimum version is `go 1.25.0` with `toolchain go1.26.8`**: Never run `go mod tidy` or any command that upgrades the `go` directive. `GOTOOLCHAIN=auto` will silently bump it to whatever toolchain is installed (e.g. 1.27.1). Always run tests with `GOTOOLCHAIN=local go test ./...` to prevent this. If `go.mod` shows a version above `1.25.0`, reset it: `go 1.25.0\ntoolchain go1.26.8`.
 
 ## Agent Delegation
 

@@ -1,6 +1,8 @@
 module github.com/bobbyjohnstx/tinycode
 
-go 1.27.1
+go 1.25.0
+
+toolchain go1.26.8
 
 require (
 	github.com/charmbracelet/bubbles v1.0.0

@@ -22,6 +22,7 @@ func RegisterBuiltins(r *Registry) {
 	r.Register(NotepadTool())
 	r.Register(PlanEnterTool())
 	r.Register(PlanExitTool())
+	r.Register(ExitTool())
 }
 
 // BuiltinConfig holds optional configuration for conditional tool registration.

@@ -0,0 +1,46 @@
+package tool
+
+import (
+	"context"
+	"encoding/json"
+)
+
+// ExitTool returns a tool the model can call to exit tinycode cleanly.
+// It publishes an "app.exit" bus event that the TUI handles as tea.Quit.
+func ExitTool() *Def {
+	return &Def{
+		ID:          "exit",
+		Description: "Exit tinycode cleanly. Call this when your task is complete and you want to hand control back to the user or an outer script.",
+		Permission:  "",
+		Parameters: map[string]any{
+			"type": "object",
+			"properties": map[string]any{
+				"reason": map[string]any{
+					"type":        "string",
+					"description": "Optional reason for exiting (e.g. 'task complete', 'PR pushed')",
+				},
+			},
+		},
+		Execute: executeExit,
+	}
+}
+
+func executeExit(ctx context.Context, tc *Context, rawArgs json.RawMessage) (*ExecuteResult, error) {
+	var args struct {
+		Reason string `json:"reason"`
+	}
+	_ = json.Unmarshal(rawArgs, &args)
+
+	reason := args.Reason
+	if reason == "" {
+		reason = "task complete"
+	}
+
+	if tc.Bus != nil {
+		tc.Bus.Publish("app.exit", map[string]any{
+			"reason": reason,
+		})
+	}
+
+	return &ExecuteResult{Output: "Exiting: " + reason}, nil
+}
@@ -0,0 +1,314 @@
+#!/usr/bin/env bash
+# self-improve.sh — run tinycode headlessly to pick up the next open issue,
+# fix it, and open a PR. Loops until the issue backlog is exhausted.
+#
+# Design: https://github.com/bobbyjohnstx/tinycode/issues/35
+#
+# How it works
+#   1. Writes a one-off config (autoApprove: true) into an isolated TINYCODE_CONFIG_DIR
+#   2. Starts `tinycode serve` bound to 127.0.0.1 with token auth
+#   3. For each round: pick the oldest actionable open issue, create a session,
+#      prompt the agent to fix it and open a PR, poll until idle, archive.
+#
+# Safety: this runs an unattended agent with autoApprove. Refuses to start
+# unless SI_ALLOW=1 is explicitly set.
+#
+# Usage
+#   SI_ALLOW=1 scripts/self-improve.sh                  # up to $MAX_ITERATIONS rounds
+#   SI_ALLOW=1 scripts/self-improve.sh --model ollama/qwen3
+#   SI_ALLOW=1 scripts/self-improve.sh --label bug
+#   SI_ALLOW=1 scripts/self-improve.sh 38               # one specific issue
+#   SI_ALLOW=1 scripts/self-improve.sh --rounds 3
+#   scripts/self-improve.sh --list-issues               # no server started
+#
+# Environment
+#   SI_ALLOW=1                   required. Opt-in to the unattended agent loop.
+#   TINYCODE                     binary path (default: ./dist/tinycode or PATH)
+#   MAX_ITERATIONS               total rounds (default 8; --rounds wins)
+#   ROUND_INACTIVITY_TIMEOUT_S   per-round no-progress budget (default 900)
+#   SI_PORT                      server port (default 4096)
+#   SI_LABEL_SKIP                comma-separated labels to skip
+#                                (default: dup,help wanted,wontfix)
+#   SI_SKIP_ISSUES               comma-separated issue numbers to skip
+#                                (default: empty)
+
+set -uo pipefail
+
+PORT="${SI_PORT:-4096}"
+D="$(pwd)"
+MODEL_OVERRIDE=""
+ROUNDS="${MAX_ITERATIONS:-8}"
+LIST_ONLY=0
+SPECIFIC_ISSUE=""
+LABEL_FILTER="${SI_LABEL_FILTER:-}"
+INACTIVITY_TIMEOUT="${ROUND_INACTIVITY_TIMEOUT_S:-900}"
+SKIP_LABELS="${SI_LABEL_SKIP:-dup,help wanted,wontfix}"
+SI_SKIP_ISSUES="${SI_SKIP_ISSUES:-}"
+CFG_DIR="$(mktemp -d "${TMPDIR:-/tmp}/tinycode-si.XXXXXX")"
+LOG_FILE="$CFG_DIR/serve.log"
+SERVER_PID=""
+
+# upstream repo: parent if this is a fork, else bobbyjohnstx/tinycode
+SI_REPO="bobbyjohnstx/tinycode"
+TARGET_REP="$(gh repo view --json parent --jq .parent.nameWithOwner 2>/dev/null)"
+[[ -n "${TARGET_REP:-}" ]] || TARGET_REP="$SI_REPO"
+
+cleanup() {
+  trap - EXIT
+  [[ -n "$SERVER_PID" ]] && kill "$SERVER_PID" 2>/dev/null
+  wait 2>/dev/null
+  rm -rf "$CFG_DIR"
+}
+trap cleanup EXIT
+
+log() { printf '%s %s\n' "$(date +%H:%M:%S)" "$*"; }
+fail() { log "FATAL: $*"; exit 1; }
+api() { curl -sS --max-time 15 "$@"; }  # no auth — server uses TINYCODE_FORCE_NO_AUTH=1
+
+usage() { sed -n '2,34p' "$0" | sed 's/^#\{1,2\} \{0,1\}//'; }
+
+while [[ $# -gt 0 ]]; do
+  case "$1" in
+    --model) MODEL_OVERRIDE="${2:?--model needs a value}"; shift 2 ;;
+    --label) LABEL_FILTER="$2"; shift ;;
+    --rounds) ROUNDS="${2:?--rounds needs a number}"; shift ;;
+    --list-issues) LIST_ONLY=1; shift ;;
+    -h|--help) usage; exit 0 ;;
+    -*) log "unknown flag: $1 (see --help)"; exit 1 ;;
+    *) SPECIFIC_ISSUE="$1"; shift ;;
+  esac
+done
+
+if [[ "$LIST_ONLY" == "1" ]]; then
+  gh issue list -R "$TARGET_REP" --state open --limit 200 \
+    --json number,title,labels \
+    --jq '.[] | [(.number|tostring), (.labels|map(.name)|join(",")), .title] | @tsv' | sort -n
+  exit 0
+fi
+
+# ---------- preflight ----------
+TINYCODE="${TINYCODE:-}"
+if [[ -z "$TINYCODE" && -x "$D/dist/tinycode" ]]; then TINYCODE="$D/dist/tinycode"
+elif [[ -z "$TINYCODE" ]]; then TINYCODE="$(command -v tinycode 2>/dev/null || true)"; fi
+[[ -n "$TINYCODE" && -x "$TINYCODE" ]] || fail "tinycode binary not found. Build it (make build) or set TINYCODE=/path"
+command -v curl >/dev/null || fail "curl not installed"
+command -v gh    >/dev/null || fail "gh not installed"
+command -v jq    >/dev/null || fail "jq not installed"
+gh auth status   >/dev/null 2>&1 || fail "gh not authenticated — run: gh auth login"
+command -v git   >/dev/null || fail "git not installed"
+git config user.email >/dev/null 2>&1 || fail "git user.email not set"
+
+if [[ -z "${SI_ALLOW:-}" ]]; then
+  echo "refusing to start: this script runs an agent unattended with autoApprove"
+  echo "against issues on $TARGET_REP. Read Issue 31.3 in that repo first."
+  echo "If you know what you are doing, re-run with: SI_ALLOW=1 $(basename "$0") $*"
+  exit 2
+fi
+
+BASE="http://127.0.0.1:$PORT"
+
+log "workdir       : $D"
+log "upstream repo : $TARGET_REP"
+log "model         : ${MODEL_OVERRIDE:-<configured default>}"
+log "rounds        : $ROUNDS   (no-progress budget ${INACTIVITY_TIMEOUT}s/round)"
+log "label filter  : ${LABEL_FILTER:-<any>}   skip: $SKIP_LABELS"
+log "log file      : $CFG_DIR"
+
+# ---------- server start ----------
+cat > "$CFG_DIR/tinycode.json" <<'EOF'
+{
+  "autoApprove": true
+}
+EOF
+
+MODEL_ARGS=()
+[[ -n "$MODEL_OVERRIDE" ]] && MODEL_ARGS=(--model "$MODEL_OVERRIDE")
+
+# The server reads its auth token from $DATADIR/web_token (not an env var).
+# Use TINYCODE_FORCE_NO_AUTH=1 — safe here because we bind to 127.0.0.1 only.
+(
+  cd "$D" || exit 1
+  export TINYCODE_CONFIG_DIR="$CFG_DIR"
+  export TINYCODE_FORCE_NO_AUTH=1
+  if [[ ${#MODEL_ARGS[@]} -gt 0 ]]; then
+    exec "$TINYCODE" serve "${MODEL_ARGS[@]}" </dev/null >>"$LOG_FILE" 2>&1
+  else
+    exec "$TINYCODE" serve </dev/null >>"$LOG_FILE" 2>&1
+  fi
+) &
+SERVER_PID=$!
+
+ok=""
+for i in $(seq 1 30); do
+  if curl -sS --max-time 2 "$BASE/global/health" >/dev/null 2>&1; then ok=1; break; fi
+  if ! kill -0 "$SERVER_PID" 2>/dev/null; then
+    log "server exited early — last log lines:"; tail -5 "$LOG_FILE"; exit 1
+  fi
+  sleep 1
+done
+[[ -n "$ok" ]] || { log "server failed to become healthy on $BASE — see $LOG_FILE"; exit 1; }
+log "server up: $BASE  (pid $SERVER_PID)"
+
+# ---------- API helpers ----------
+create_session() {
+  local payload
+  payload=$(jq -nc --arg c "self-improve #$1 round $2" '{title:$c}')
+  api -X POST "$BASE/session" -H 'Content-Type: application/json' -d "$payload"
+}
+
+send_prompt() {
+  local sid="$1"
+  api -X POST "$BASE/session/$sid/prompt_async" -H 'Content-Type: application/json' \
+    -d "$(jq -nc --arg t "$2" '{parts:[{type:"text",text:$t}]}')" >/dev/null
+}
+
+assistant_message_count() {
+  api "$BASE/session/$1/message" \
+    | jq '[.[] | select(.info.role == "assistant")] | length' 2>/dev/null \
+    || echo 0
+}
+
+# The /session/status map lists only in-flight sessions; when the entry
+# disappears the prompt has finished. Success = an assistant message exists.
+wait_idle() {
+  local sid="$1" last=$SECONDS
+  while true; do
+    if api "$BASE/session/status" 2>/dev/null | jq -e --arg s "$sid" '.[$s]' >/dev/null; then
+      last=$SECONDS
+      sleep 3
+      continue
+    fi
+    if (( SECONDS - last > INACTIVITY_TIMEOUT )); then
+      log "no completion for ${INACTIVITY_TIMEOUT}s — aborting round"
+      api -X POST "$BASE/session/$sid/abort" >/dev/null
+      return 1
+    fi
+    sleep 2
+    if (( $(assistant_message_count "$sid") > 0 )); then
+      log "session $sid → done"
+      return 0
+    fi
+    log "session $sid ended with no assistant output (check $LOG_FILE)"
+    return 1
+  done
+}
+
+# ---------- main loop ----------
+skip_jq=$(printf '%s' "$SKIP_LABELS" | jq -Rcs 'split(",") | map(gsub("^\\s+|\\s+$";""))')
+SEEN_JSON='[]'   # issue numbers already attempted this run
+SI_SKIP_JSON=$(printf '%s' "$SI_SKIP_ISSUES" | jq -Rcs 'split(",") | map(gsub("^\\s+|\\s+$";"")) | map(select(length > 0) | tonumber)')
+pick_pickable() {
+  gh issue list -R "$TARGET_REP" --state open --limit 200 \
+    ${1:+--label "$1"} \
+    --json number,title,labels \
+  | jq --argjson skip "$skip_jq" --argjson seen "$SEEN_JSON" --argjson si "$SI_SKIP_JSON" '
+      [ .[]
+        | select(.number as $n | (($seen + $si) | index($n)) | not)
+        | select((.labels | map(.name) | map(ascii_downcase)) as $l
+               | (($l - $skip) | length) == ($l | length))
+      ] | sort_by(.number) | first'
+}
+for round in $(seq 1 "$ROUNDS"); do
+  log "══════════ round $round / $ROUNDS ══════════"
+
+  if [[ -n "$SPECIFIC_ISSUE" ]]; then
+    picked=$(gh issue view "$SPECIFIC_ISSUE" -R "$TARGET_REP" --json number,title --jq '{number, title}')
+  else
+    if [[ -n "$LABEL_FILTER" ]]; then
+      picked=$(pick_pickable "$LABEL_FILTER")
+    else
+      picked=$(pick_pickable)
+    fi
+  fi
+
+  if [[ -z "$picked" || "$picked" == "null" ]]; then
+    log "no actionable issues left — done."
+    break
+  fi
+
+  num=$(echo "$picked" | jq -r .number)
+  title=$(echo "$picked" | jq -r '.title // ""')
+  log "round $round: issue #$num — $title"
+  # Remember this issue so the next round (even after a failed or
+  # "won't fix" outcome) moves to the next one instead of re-picking it.
+  SEEN_JSON=$(jq -cn --argjson s "$SEEN_JSON" --argjson n "$num" '$s + [$n')
+
+  created=$(create_session "$num" "$round")
+  sid=$(echo "$created" | jq -r '.id // empty' 2>/dev/null)
+  if [[ -z "$sid" ]]; then
+    log "session create failed: $created"
+  else
+    read -r -d '' PROMPT <<PROMPT || true
+You are running unattended in a headless tinycode session, fixing issue #$num in $TARGET_REP.
+Working directory: $D
+
+Issue: #$num — $title
+
+Follow this sequence exactly, using your tools (bash, read, write, edit, task, skill).
+
+1. Read the issue fully:
+   gh issue view $num -R $TARGET_REP --json title,body,labels,comments
+
+2. Judge actionability. If the issue is:
+   - already fixed (check recent commits / blame),
+   - an upstream model or provider bug outside this codebase,
+   - a feature request with no clear spec, or needs design discussion,
+   then explain why in a comment and STOP:
+   gh issue comment $num -R $TARGET_REP --body '<a short, evidence-backed reason>'
+   (Do not branch, commit, or open a PR.)
+
+3. Otherwise implement the fix:
+   - Read the relevant source first. This is a Go project (bubbletea TUI,
+     pure-Go SQLite, lipgloss). Match surrounding style.
+   - Keep the change minimal and targeted. No new abstractions for one-off logic.
+   - For multi-file or architectural work, spawn a subagent (architect for
+     design, executor for implementation).
+   - Do not touch packages/ — that is legacy TypeScript, not the Go runtime.
+   - Never disable auth checks or weaken tests just to make them pass.
+
+4. Verify before committing:
+   go build ./...
+   go vet ./...
+   go test ./... -count=1
+   Fix anything you broke before committing.
+
+5. Commit and open a PR (branch from main, not canary):
+   git checkout -b fix-$num-<slug>
+   git add <specific files>
+   git commit -m 'fix(scope): <imperative summary>'
+   git push -u origin fix-$num-<slug>
+   gh pr create -R $TARGET_REP --base main --head origin:fix-$num-<slug> \\
+     --title 'fix(scope): <same as commit>' \\
+     --body "Fixes #$num.\\n\\n<what was wrong, what changed, why minimal. Commands run.>"
+   gh issue comment $num -R $TARGET_REP --body "Fix in #<PR-number>. <one line on what changed>."
+
+6. If gh pr create fails for external reasons (permissions, branch protection),
+   post the exact error as an issue comment and STOP.
+
+If you get stuck, the acceptable outcome is an honest issue comment — a commented
+issue beats a broken PR. Do not silently give up.
+PROMPT
+
+    send_prompt "$sid" "$PROMPT"
+    log "prompt sent"
+
+    if wait_idle "$sid"; then
+      pr=$(gh pr list -R "$TARGET_REP" --state open --limit 10 \
+        --json number,title --jq --arg n "#$num" \
+        '.[] | select(.title | ascii_downcase | contains($n | ascii_downcase))
+         | "\(#): \(.title)"' | head -1)
+      log "round $round result: ${pr:-<no matching PR — check issue comment>}"
+    else
+      log "round $round gave up — details in $LOG_FILE and the session record"
+    fi
+    api -X POST "$BASE/session/$sid/archive" >/dev/null
+  fi
+
+  if [[ -n "$SPECIFIC_ISSUE" ]]; then
+    log "single-issue mode — stopping after round $round"
+    break
+  fi
+done
+
+log "self-improve run finished."
+exit 0