chore: add agentic development toolkit (#33)

.agents/README.md

@@ -0,0 +1,49 @@
+# `.agents/` — harness-agnostic agentic dev toolkit
+
+Portable skills, prompts, and scripts that drive the box-box development lifecycle.
+Any harness (Claude, Codex, opencode, …) can read these — the canonical workflows
+live here, not in a tool-specific folder. The shared project context every harness
+reads is `AGENTS.md` (→ `CLAUDE.md`).
+
+## Layout
+
+```
+skills/          Codex/open agent skills: groom, write-spec, implement, review, lenses
+personas/        grill.md (base) + lens overlays (architect, …) — the interrogation voices
+prompts/         ready-spec.md (groomed spec), implement.md/review.md dispatch prompts
+lib/gh.sh        GitHub issue + Project (#2) state helpers: issue_*, set_stage/effort/priority
+lib/dispatch.sh  dispatch(): Ready issue → worktree → harness → gate → PR
+harnesses.sh     headless adapters (one fn per harness) + run_gate — the ONLY tool-specific code
+bin/dev          CLI: `dev implement <issue#> --harness <name> [--dry-run]`
+```
+
+## The lifecycle
+
+`Icebox → Research → Ready → In Progress → In Review → Done` (the Project `Stage` field).
+
+- **Groom** (interactive, Claude): `/groom <issue#>` runs a seeded grill-me → writes a
+  Ready spec into the issue body → sets Effort/Priority → leaves Stage at `Research`.
+  You review and flip to `Ready`.
+- **Implement** (any harness): `.agents/bin/dev implement <issue#> --harness <name>`
+  (or `/implement …` in Claude to supervise) → isolated worktree → runs the harness
+  headless on the spec → build gate → opens a PR → sets Stage `In Review`.
+- **Review + merge**: use the `review` skill from a harness different from the
+  implementer to create a local review packet and PR comment, then you merge.
+
+## Skills and harnesses
+
+`.agents/skills` is the canonical home for reusable workflows. Codex discovers
+repo skills from that path directly, and Claude can use the same files through
+`.claude/skills -> ../.agents/skills`. Other harnesses can read the same
+`SKILL.md` files explicitly or enter the workflow through `.agents/bin/dev`.
+Do not put canonical workflow instructions under `.claude/`; that directory is
+local adapter state.
+
+## Adding / fixing a harness
+
+Edit one function in `harnesses.sh`: `harness_<name> <workdir> <promptfile>`, running the
+tool non-interactively in `<workdir>` on the prompt. `claude`/`codex`/`opencode` are
+wired; `pi`/`cursor` are stubs — confirm their headless flags before trusting.
+
+Always `--dry-run` a new harness first: it renders the exact prompt and plan, touching
+nothing (no worktree, PR, or state change).

.agents/bin/dev

@@ -0,0 +1,45 @@
+#!/usr/bin/env bash
+# .agents/bin/dev — box-box agentic dev CLI. Works from any harness's shell.
+#
+#   dev implement <issue#> --harness <claude|codex|opencode|pi|cursor> [--dry-run] [--base <branch>]
+#
+# Grooming is driven interactively via the Claude Code /groom skill; this CLI covers
+# the implement lane (dispatch a Ready issue to a harness → worktree → gate → PR).
+
+set -o pipefail
+ROOT="$(cd "$(dirname "${BASH_SOURCE[0]:-$0}")/../.." && pwd)"
+
+usage() {
+  cat >&2 <<EOF
+box-box dev CLI
+
+  dev implement <issue#> --harness <name> [--dry-run] [--base <branch>]
+
+harnesses: claude, codex, opencode (supported) · pi, cursor (verify flags in .agents/harnesses.sh)
+  --dry-run   render the prompt + plan, touch nothing (no worktree/PR/state change)
+  --base      base branch for the worktree/PR (default: main)
+EOF
+  exit 2
+}
+
+cmd="${1:-}"; shift 2>/dev/null || true
+case "$cmd" in
+  implement)
+    issue="${1:-}"; shift 2>/dev/null || true
+    harness=""; passthru=()
+    while [ $# -gt 0 ]; do
+      case "$1" in
+        --harness) harness="${2:-}"; shift 2 ;;
+        --dry-run) passthru+=(--dry-run); shift ;;
+        --base)    passthru+=(--base "${2:-}"); shift 2 ;;
+        *) echo "unknown arg: $1" >&2; usage ;;
+      esac
+    done
+    [ -n "$issue" ] && [ -n "$harness" ] || usage
+    # shellcheck source=/dev/null
+    source "$ROOT/.agents/lib/dispatch.sh"
+    dispatch "$issue" "$harness" ${passthru[@]+"${passthru[@]}"}
+    ;;
+  ""|-h|--help|help) usage ;;
+  *) echo "unknown command: $cmd" >&2; usage ;;
+esac

.agents/harnesses.sh

@@ -0,0 +1,52 @@
+#!/usr/bin/env bash
+# .agents/harnesses.sh — headless harness adapters.
+#
+# ONE function per harness:  harness_<name> <workdir> <promptfile>
+# Each runs the harness NON-INTERACTIVELY inside <workdir>, feeding it the rendered
+# prompt, expected to edit files (and ideally commit). This is the ONLY harness-specific
+# code in the toolkit — tune the exact flags here per tool/version.
+#
+# Autonomy note: these run unattended, so each uses the tool's "just do it" mode
+# (accept edits / full-auto). Only run harnesses you trust on code you can review via PR.
+
+# ---- MUST-HAVE ----
+
+harness_claude() {                      # Claude Code — print mode, auto-accept edits
+  local dir="$1" prompt="$2"
+  ( cd "$dir" && claude -p "$(cat "$prompt")" --permission-mode acceptEdits )
+}
+
+harness_codex() {                       # OpenAI Codex CLI — non-interactive exec, full auto
+  local dir="$1" prompt="$2"
+  ( cd "$dir" && codex exec --full-auto "$(cat "$prompt")" )
+}
+
+harness_opencode() {                    # opencode — non-interactive run
+  local dir="$1" prompt="$2"
+  ( cd "$dir" && opencode run "$(cat "$prompt")" )
+}
+
+# ---- NICE-TO-HAVE (verify the exact invocation for your version before trusting) ----
+
+harness_pi() {                          # pi — CONFIRM headless CLI + flags
+  local dir="$1" prompt="$2"
+  ( cd "$dir" && pi run "$(cat "$prompt")" )        # placeholder — verify
+}
+
+harness_cursor() {                      # Cursor CLI agent — CONFIRM flags
+  local dir="$1" prompt="$2"
+  ( cd "$dir" && cursor-agent -p "$(cat "$prompt")" --force )   # placeholder — verify
+}
+
+# ---- build/typecheck gate (fast, local) ----
+# Returns non-zero on failure. This is a smoke gate — CI runs the full suite. Tune freely.
+run_gate() {
+  local dir="$1"
+  ( cd "$dir" && go build ./... ) || return 1
+  if [ -d "$dir/frontend/node_modules" ]; then
+    ( cd "$dir/frontend" && npx tsc --noEmit ) || return 1
+  else
+    echo "  (gate: frontend deps absent in worktree — tsc/vitest deferred to CI)" >&2
+  fi
+  return 0
+}

.agents/lib/dispatch.sh

@@ -0,0 +1,121 @@
+#!/usr/bin/env bash
+# .agents/lib/dispatch.sh — implement a Ready issue with a chosen harness in an
+# isolated git worktree, run the build gate, and open a PR.
+#
+# Source it, then:  dispatch <issue#> <harness> [--dry-run] [--base <branch>]
+# (or use the CLI:  .agents/bin/dev implement <issue#> --harness <name> [--dry-run])
+
+_AGENTS_DIR="$(cd "$(dirname "${BASH_SOURCE[0]}")/.." && pwd)"
+# shellcheck source=/dev/null
+source "$_AGENTS_DIR/lib/gh.sh"
+# shellcheck source=/dev/null
+source "$_AGENTS_DIR/harnesses.sh"
+
+_slug() { echo "$1" | tr '[:upper:]' '[:lower:]' | sed -E 's/[^a-z0-9]+/-/g; s/^-+//; s/-+$//' | cut -c1-40; }
+
+_render_prompt() { # <issue#> <title> <body>
+  echo "# Implement: $2"
+  echo
+  cat "$_AGENTS_DIR/prompts/implement.md"
+  echo; echo "---"; echo
+  echo "## Spec — issue #$1"
+  echo
+  echo "$3"
+}
+
+_pr_body() { # <issue#> <harness> <gate>
+  cat <<EOF
+Implements #$1.
+
+- **Harness:** $2 (dispatched via \`.agents/bin/dev\`)
+- **Local gate** (\`go build\` + \`tsc --noEmit\`): **$3**
+- Full test suite + independent review run in CI / by a reviewer harness.
+
+See #$1 for the groomed spec, Test Plan, and Definition of Done.
+
+Closes #$1
+EOF
+}
+
+dispatch() { # <issue#> <harness> [--dry-run] [--base <branch>]
+  local issue="$1" harness="$2"; shift 2 || { echo "usage: dispatch <issue#> <harness> [--dry-run] [--base <branch>]"; return 2; }
+  local dry=0 base="main"
+  while [ $# -gt 0 ]; do
+    case "$1" in
+      --dry-run) dry=1 ;;
+      --base)    base="$2"; shift ;;
+      *) echo "dispatch: unknown flag '$1'" >&2; return 2 ;;
+    esac; shift
+  done
+
+  # adapter must exist
+  if ! declare -f "harness_$harness" >/dev/null 2>&1; then
+    echo "no adapter for harness '$harness' — add harness_$harness() to .agents/harnesses.sh" >&2; return 2
+  fi
+
+  local repo_root title body slug branch wt prompt
+  repo_root="$(git rev-parse --show-toplevel)" || return 1
+  title="$(issue_title "$issue")" || { echo "issue #$issue not found on $REPO" >&2; return 1; }
+  body="$(issue_body "$issue")"
+  slug="$(_slug "$title")"
+  branch="feat/issue-${issue}-${slug}"
+  wt="$repo_root/.worktrees/issue-${issue}"
+  prompt="$(mktemp "${TMPDIR:-/tmp}/boxbox-prompt-${issue}.XXXX")"
+  _render_prompt "$issue" "$title" "$body" > "$prompt"
+
+  echo "── dispatch #$issue → $harness ──"
+  echo "  title    : $title"
+  echo "  branch   : $branch"
+  echo "  worktree : $wt"
+  echo "  base     : $base"
+  echo "  prompt   : $prompt"
+
+  if [ "$dry" = 1 ]; then
+    echo "  [dry-run] no worktree / harness / PR / state change. Prompt preview:"
+    sed 's/^/    | /' "$prompt"
+    return 0
+  fi
+
+  # preflight: warn (don't block) if not Ready
+  local stage; stage="$(get_field "$issue" Stage)"
+  [ "$stage" = "Ready" ] || echo "  ⚠ Stage is '$stage' (not Ready) — dispatching anyway"
+
+  # isolated worktree
+  if [ -d "$wt" ]; then
+    echo "  worktree exists — reusing"
+  else
+    git -C "$repo_root" worktree add -b "$branch" "$wt" "$base" || return 1
+  fi
+
+  set_stage "$issue" "In Progress"
+
+  echo "  running $harness (headless)…"
+  ( harness_"$harness" "$wt" "$prompt" ); local hrc=$?
+  echo "  $harness exited ($hrc)"
+
+  # fallback commit: guarantee a PR-able branch even if the harness didn't commit
+  if [ -n "$(git -C "$wt" status --porcelain)" ]; then
+    git -C "$wt" add -A
+    git -C "$wt" commit -q -m "feat(#$issue): $title
+
+Implemented by $harness via .agents/dev dispatch." && echo "  committed leftover changes"
+  fi
+
+  # gate
+  local gate="passed"
+  run_gate "$wt" || gate="FAILED"
+  echo "  gate: $gate"
+
+  # PR (only if there are commits ahead of base)
+  if [ -n "$(git -C "$wt" log "$base..$branch" --oneline 2>/dev/null)" ]; then
+    git -C "$wt" push -u origin "$branch" || { echo "  push failed — inspect $wt" >&2; return 1; }
+    local draft=""; [ "$gate" = "FAILED" ] && draft="--draft"
+    local pr
+    pr="$(gh pr create -R "$REPO" --head "$branch" --base "$base" $draft \
+      --title "$title (#$issue)" --body "$(_pr_body "$issue" "$harness" "$gate")")" || { echo "  gh pr create failed" >&2; return 1; }
+    echo "  PR: $pr${draft:+  (draft — gate failed)}"
+    set_stage "$issue" "In Review"
+  else
+    echo "  no commits on $branch — leaving Stage 'In Progress'. Inspect the worktree: $wt" >&2
+  fi
+}

.agents/lib/gh.sh

@@ -0,0 +1,68 @@
+#!/usr/bin/env bash
+# .agents/lib/gh.sh — harness-agnostic helpers for the box-box agentic dev lifecycle.
+#
+# Source it, then call the functions:
+#   source .agents/lib/gh.sh
+#   issue_json 9 ; set_stage 9 Research ; set_effort 9 M
+#
+# Requires: gh (authed, with `project` scope), jq.
+# Config is overridable via env vars.
+
+REPO="${BOXBOX_REPO:-AmanTahiliani/box-box}"
+PROJECT_OWNER="${BOXBOX_PROJECT_OWNER:-AmanTahiliani}"
+PROJECT_NUMBER="${BOXBOX_PROJECT_NUMBER:-2}"
+
+_BOXBOX_CACHE="${TMPDIR:-/tmp}/boxbox-agent"
+mkdir -p "$_BOXBOX_CACHE" 2>/dev/null
+
+# ---------- issues ----------
+issue_json()     { gh issue view "$1" -R "$REPO" --json number,title,body,labels,url,state; }
+issue_body()     { gh issue view "$1" -R "$REPO" --json body  -q .body; }
+issue_title()    { gh issue view "$1" -R "$REPO" --json title -q .title; }
+issue_url()      { gh issue view "$1" -R "$REPO" --json url   -q .url; }
+set_issue_body() { gh issue edit "$1" -R "$REPO" --body-file "$2"; }        # <issue#> <file>
+add_comment()    { gh issue comment "$1" -R "$REPO" --body-file "$2"; }     # <issue#> <file>
+
+# Native sub-issue children of an epic (issue numbers, one per line).
+sub_issues() {
+  gh api graphql -H "GraphQL-Features: sub_issues" -f query='
+    query($owner:String!,$repo:String!,$num:Int!){
+      repository(owner:$owner,name:$repo){
+        issue(number:$num){ subIssues(first:50){ nodes{ number } } } } }' \
+    -F owner="${REPO%/*}" -F repo="${REPO#*/}" -F num="$1" \
+    -q '.data.repository.issue.subIssues.nodes[].number'
+}
+
+# ---------- project fields (cached per shell invocation) ----------
+_fields_file="$_BOXBOX_CACHE/fields.json"
+_items_file="$_BOXBOX_CACHE/items.json"
+_pid_file="$_BOXBOX_CACHE/project_id"
+
+_refresh_fields() { gh project field-list "$PROJECT_NUMBER" --owner "$PROJECT_OWNER" --format json > "$_fields_file"; }
+_refresh_items()  { gh project item-list  "$PROJECT_NUMBER" --owner "$PROJECT_OWNER" --format json --limit 200 > "$_items_file"; }
+project_refresh() { _refresh_fields; _refresh_items; }   # call once at the start of a session to get fresh state
+
+_project_id() { [ -s "$_pid_file" ] || gh project view "$PROJECT_NUMBER" --owner "$PROJECT_OWNER" --format json | jq -r .id > "$_pid_file"; cat "$_pid_file"; }
+_field_id()   { [ -s "$_fields_file" ] || _refresh_fields; jq -r --arg n "$1" '.fields[]|select(.name==$n)|.id' "$_fields_file"; }
+_option_id()  { [ -s "$_fields_file" ] || _refresh_fields; jq -r --arg f "$1" --arg o "$2" '.fields[]|select(.name==$f)|.options[]?|select(.name==$o)|.id' "$_fields_file"; }
+_item_id()    { [ -s "$_items_file" ]  || _refresh_items;  jq -r --arg n "$1" '.items[]|select(.content.number==($n|tonumber))|.id' "$_items_file"; }
+
+# set_field <issue#> <FieldName> <OptionName>   (single-select fields: Stage/Priority/Effort/Phase)
+set_field() {
+  local item opt fld pid
+  item="$(_item_id "$1")"; fld="$(_field_id "$2")"; opt="$(_option_id "$2" "$3")"; pid="$(_project_id)"
+  if [ -z "$item" ] || [ -z "$fld" ] || [ -z "$opt" ]; then
+    echo "set_field: could not resolve issue=$1 field=$2 option=$3 (item=$item field=$fld opt=$opt)" >&2; return 1
+  fi
+  gh project item-edit --id "$item" --project-id "$pid" --field-id "$fld" --single-select-option-id "$opt" >/dev/null \
+    && echo "set #$1 $2=$3"
+}
+set_stage()    { set_field "$1" Stage    "$2"; }
+set_priority() { set_field "$1" Priority "$2"; }
+set_effort()   { set_field "$1" Effort   "$2"; }
+
+# get_field <issue#> <FieldName>  -> current value (single-word field names only)
+get_field() {
+  [ -s "$_items_file" ] || _refresh_items
+  jq -r --arg n "$1" --arg f "$2" '.items[]|select(.content.number==($n|tonumber))|.[($f|ascii_downcase)] // "-"' "$_items_file"
+}

.agents/personas/architect.md

@@ -0,0 +1,35 @@
+# Lens overlay: Architect
+
+Apply this **on top of** the base grill (`grill.md`). Bias every question toward
+technical soundness and fit with the existing box-box architecture.
+
+## box-box context to hold
+
+- **Backend (Go):** `internal/api` (OpenF1 client, cache, 37 endpoints), `internal/web`
+  (REST handlers + SSE hub, route table in `server.go`), `internal/store` (domain
+  SQLite), `internal/query` (read models over the domain DB), `internal/ingest`,
+  `internal/live` (SignalR).
+- **Frontend (React+Vite+TS):** TanStack Router/Query, `src/api.ts` (typed fetchers),
+  `src/types.ts` (payload mirrors), `src/lib` (client helpers), `src/pages`,
+  `src/components` (incl. `components/live`).
+- **Patterns to respect** (see CLAUDE.md "How To Extend"): ServeMux longest-prefix
+  route ordering in `server.go`, cache TTL tiers, `?source=openf1|local|auto`
+  resolution, two-phase standings load, lazy tab loads, stale-data fallback banner.
+
+## Grill especially on
+
+- **Reuse vs new:** does existing code already do this (a TUI equivalent in
+  `internal/ui/*.go`, a query model, an `api.ts` fetcher)? Port vs rebuild.
+- **Data flow & source:** OpenF1 live vs domain DB vs cache; payload size; rate
+  limits; how `?source` is handled.
+- **Seams:** which files/modules change; new endpoint (mind registration order!) vs
+  extend an existing one; new component vs extend; where shared logic lives
+  (`frontend/src/lib/*`).
+- **Testability:** how does this land in `go test` / `vitest` / hermetic Playwright?
+  What seam makes it testable without live OpenF1?
+- **Risk:** domain-DB migrations, perf on large sessions, backward compat, and
+  failure / stale-data behavior.
+
+Keep questions concrete and decision-shaped — e.g. *"port the GPS normalization from
+`internal/ui/trackmap.go`, or recompute in a shared `frontend/src/lib/trackmap.ts` so
+it's unit-testable?"* — each with your recommendation.

.agents/personas/grill.md

@@ -0,0 +1,51 @@
+# Persona: The Grill
+
+You run a **grill-me** session. Instead of the human prompting you, **you interrogate
+the human** until you share a design concept for one specific piece of work (a GitHub
+issue or epic). The shared understanding — not the document — is the real output.
+
+## Before you ask anything
+
+Load the full context of the target:
+
+- Read the issue title + body (and any notes already on it).
+- Read `CLAUDE.md`.
+- Explore the code paths the work implicates.
+
+## Rules
+
+1. **One question at a time.** Walk each branch of the design tree and resolve
+   dependencies in order — a later question often depends on an earlier answer.
+2. **Recommend, consequence-first.** Every question carries your recommended answer
+   and a short "why". **Calibrate to a technically fluent reader who does not know
+   *this project's* internals.** Assume general engineering literacy (APIs, streaming,
+   latency, front/back-end, caching, etc.) and don't explain those. **Do** unpack
+   anything project-specific: internal file/type/endpoint names, bespoke architecture
+   choices, and why they matter *here* — a few words is enough, no lectures. Above all,
+   lead each option with the **practical consequence** a decision can be made on
+   (effort, risk, what ships sooner, how it feels to use), so the reader can choose
+   without needing the implementation detail. Recommendation first, with why it's the
+   better call **for them**.
+   - *Example — keep the mechanism, but lead with the tradeoff:* "**A (recommended):**
+     reuse the existing SSE snapshot — cheapest to build, but cars jump a little
+     between updates. **B:** a dedicated ~4Hz position stream — more work now, but
+     motion is smooth and it sets up interpolation later." (Names the real mechanism;
+     the choice is still obvious from the consequences.)
+3. **Hybrid asking.**
+   - Decision with clear discrete options → present a **structured choice**, the
+     recommendation first. *(In Claude Code: use the AskUserQuestion tool; put the
+     recommended option first and end its label with "(Recommended)".)* Write each
+     option's description in the plain-language, consequence-first style from rule 2 —
+     the label can be terse, but the description must be understandable on its own.
+   - Genuinely open-ended → ask in **prose**.
+4. **Explore before you ask.** If the codebase or the issue already answers a
+   question, do **not** ask — state what you found and the assumption you're
+   proceeding with, then move on. Only ask about real forks the human must decide.
+5. **Stay in scope.** Grill the design of *this* work, not the whole app. Note
+   out-of-scope temptations instead of chasing them.
+
+## Termination
+
+Stop when no unresolved branches remain and you could write the spec yourself with no
+open questions. Summarize the shared design concept in 3–6 bullets, confirm it with
+the human, then hand off to `write-spec`.

.agents/prompts/implement.md

@@ -0,0 +1,21 @@
+You are an autonomous coding agent working in an **isolated git worktree** on the
+**box-box** repo. Implement the groomed spec below as a single, focused, story-sized
+change — then commit it.
+
+## Ground rules
+- Read `AGENTS.md` / `CLAUDE.md` first and follow the project's conventions exactly
+  (architecture, route-registration order in `server.go`, cache TTL tiers,
+  `api.ts`/`types.ts` mirrors, test layout).
+- Implement ONLY this story's scope. Honor the spec's **Out of Scope** — do not build
+  deferred items, even if tempting.
+- If the spec has an **early spike / risk** step, do that FIRST and note the result in
+  your commit message (and adjust the approach if the spike says to).
+- Add or extend tests per the **Test Plan**. Make the relevant suites pass:
+  `go build ./...`, `go test ./...`, and in `frontend/`: `npm run test`, `tsc --noEmit`.
+- Keep the change reviewable and story-sized. **Commit your work** with a clear,
+  conventional message when done (the dispatcher opens the PR).
+- Satisfy every item in the spec's **Definition of Done**.
+
+If something in the spec is ambiguous or turns out to be wrong once you're in the code,
+make the smallest reasonable decision, implement it, and call it out clearly in the
+commit message / PR so the reviewer can catch it — do not silently expand scope.

.agents/prompts/ready-spec.md

@@ -0,0 +1,27 @@
+## Context
+{one paragraph: the problem and why it matters, grounded in the grooming}
+
+## Acceptance Criteria
+- [ ] {observable behaviour 1}
+- [ ] {observable behaviour 2}
+
+## Technical Approach
+- **Files to touch:** {real paths}
+- **Data source:** {OpenF1 live | domain DB | cache | ...}
+- **Endpoints/components:** {new or extended — note route-registration order if new}
+- **Key decisions:** {the forks resolved during the grill}
+
+## Test Plan
+- {cases mapped to the real suites: go test · vitest · hermetic Playwright}
+
+## Out of Scope
+- {explicitly deferred}
+
+## Definition of Done
+- [ ] Tests added and green (`go test` · `vitest` · `tsc --noEmit` · hermetic Playwright as applicable)
+- [ ] Matches CLAUDE.md conventions (route order, cache TTLs, `api.ts`/`types.ts` mirrors)
+- [ ] No console / preview errors (UI verified in preview)
+- [ ] Story-sized PR, linked to this issue
+
+---
+_Groomed {date} · Effort {S|M|L} · Priority {P0|P1|P2} · via /groom_

.agents/prompts/review.md

@@ -0,0 +1,20 @@
+You are an independent reviewer for a box-box pull request. Review the PR against
+the linked GitHub issue spec, not against your own preferred scope.
+
+## Ground rules
+- Use a harness different from the implementer when possible.
+- Read `AGENTS.md` / `CLAUDE.md`, the PR body/diff, and the linked issue body.
+- Run the relevant local gates and capture logs under `.review/issue-<n>-pr-<pr>/logs`.
+- For UI changes, create a local visual packet under `.review/issue-<n>-pr-<pr>/`
+  with desktop and mobile screenshots for the affected routes.
+- If screenshots should appear inline on GitHub, publish only review artifacts to
+  a separate artifact branch, never to `main` or the product PR branch.
+- Post a PR comment with pass/fail status, acceptance-criteria alignment, local
+  artifact paths, screenshot links when available, and caveats.
+- Do not merge. The human owns the merge gate.
+
+## Output
+- A local packet with `summary.md`, optional `index.html`, screenshots, logs, and
+  any visual diffs.
+- A GitHub PR comment that makes the review visually scannable.
+- A clear recommendation: pass, pass with caveats, or needs changes.

.agents/skills/groom/SKILL.md

@@ -0,0 +1,42 @@
+---
+name: groom
+description: Groom a box-box GitHub issue into a Ready spec via a seeded grill-me interrogation. Use when the user asks to groom an issue, for example "/groom <issue-number>" or "$groom <issue-number>" with optional "--lens architect". Single-issue path: asks targeted questions, writes a structured spec into the issue body, sets Effort/Priority, and leaves Stage at Research for approval.
+argument-hint: <issue-number> [--lens architect]
+---
+
+# /groom — refine one issue into a Ready spec
+
+Groom issue **$ARGUMENTS** on the box-box roadmap (Project #2): drive it from the
+backlog into a fully-specified Ready ticket through a grill-me session.
+
+## 0. Setup
+- Parse the first token of the arguments as the **issue number**. An optional
+  `--lens <name>` pulls in a lens overlay (currently: `architect`).
+- Run: `source .agents/lib/gh.sh && project_refresh` (fresh field/item state).
+- Load context **before asking anything**: `issue_json <n>`, read `CLAUDE.md`, and
+  explore the code paths the issue implicates.
+- Move it into grooming if it isn't already there: `set_stage <n> Research`.
+
+## 1. Grill
+- Read `.agents/personas/grill.md`. If `--lens <name>` was given, also read
+  `.agents/personas/<name>.md` and apply it on top.
+- Run the grill exactly per those rules: **one question at a time**, recommendation
+  first, **hybrid** asking (AskUserQuestion for discrete decisions with the
+  recommended option first and labelled "(Recommended)"; prose for open-ended), and
+  **explore the code to self-answer** wherever possible — only ask about genuine forks.
+- Track the resolved decisions as you go.
+
+## 2. Synthesize
+- When no open branches remain, summarize the shared design concept in 3–6 bullets and
+  confirm it with the user.
+- Then invoke the **write-spec** skill for issue `<n>`, handing it the resolved
+  decisions, so it renders `.agents/prompts/ready-spec.md` into the issue body and
+  sets Effort + Priority.
+
+## 3. Hand back (human gate)
+- Do **not** auto-advance to Ready — that's the user's call. Report that the spec is
+  written, Stage is `Research`, and they should review the issue and flip Stage →
+  `Ready` when satisfied (`set_stage <n> Ready`).
+- Print the issue URL (`issue_url <n>`).
+
+Stay focused on THIS issue's design throughout. Note but don't chase out-of-scope ideas.

.agents/skills/implement/SKILL.md

@@ -0,0 +1,32 @@
+---
+name: implement
+description: Dispatch a Ready box-box issue to a coding harness (claude/codex/opencode/pi/cursor) in an isolated git worktree, run the build gate, and open a PR. Use when supervising implementation via "/implement <issue-number> --harness <name> [--dry-run]" or from any harness terminal with .agents/bin/dev.
+argument-hint: <issue-number> --harness <claude|codex|opencode|pi|cursor> [--dry-run]
+---
+
+# /implement — dispatch an issue to a harness (supervised)
+
+Supervise the implementation of issue **$ARGUMENTS**. You are SUPERVISING, not coding —
+a fresh harness does the work in its own isolated worktree with clean context. Do not
+edit project files yourself.
+
+## Steps
+1. Parse: `<issue-number> --harness <name> [--dry-run] [--base <branch>]`.
+2. **Preflight (report, don't hard-block):** `source .agents/lib/gh.sh` and check
+   `get_field <n> Stage` is `Ready` and the issue body has an "## Acceptance Criteria"
+   section (a groomed spec). If it's not Ready or has no spec, say so and recommend
+   `/groom <n>` first — proceed only if the user confirms.
+3. **Dispatch:** run `.agents/bin/dev implement <n> --harness <name> [flags]`. For a
+   first run against an unfamiliar harness, suggest `--dry-run` first so the user can
+   eyeball the prompt.
+4. **Report the outcome:** branch, worktree path, gate result (pass/FAILED → draft PR),
+   and the PR URL. On success, Stage will be `In Review`.
+5. If no PR was created (no changes, or push failed), surface exactly why and point at
+   the worktree (`.worktrees/issue-<n>`) so the user can inspect. Diagnose from the
+   dispatcher output; recommend a fix or re-run — don't silently take over the coding.
+
+## Notes
+- The harness adapters and the gate live in `.agents/harnesses.sh` — the single place
+  to tune per-tool flags.
+- From any harness shell, run the same thing directly:
+  `.agents/bin/dev implement <n> --harness <name>`.

.agents/skills/lens-architect/SKILL.md

@@ -0,0 +1,22 @@
+---
+name: lens-architect
+description: Grill/analyze a box-box issue or epic from a software-architecture perspective and post the findings as a comment. Use standalone as "/lens-architect <issue-number>" for an on-the-fly architecture review, or let /groom compose it via "--lens architect". Reads the base grill + architect persona and focuses on reuse, data flow, seams, testability, and risk.
+argument-hint: <issue-number>
+---
+
+# /lens-architect — architecture lens
+
+Target: issue **$ARGUMENTS** (box-box, Project #2).
+
+1. `source .agents/lib/gh.sh`; load context (`issue_json <n>`, `CLAUDE.md`, and the
+   relevant code paths).
+2. Read `.agents/personas/grill.md` + `.agents/personas/architect.md` and run a focused
+   grill from the architecture lens: hybrid asking (AskUserQuestion for discrete
+   decisions, recommendation first; prose otherwise), recommend every answer, and
+   explore the code to self-answer before asking.
+3. When aligned, write an **"## Architecture review"** summary (decisions taken,
+   files/seams affected, risks, the test seam) to a temp file and `add_comment <n> <file>`.
+4. Print the issue URL.
+
+If invoked from **within /groom**, skip the comment — instead return the architecture
+decisions inline so groom can fold them into the spec.

.agents/skills/review/SKILL.md

@@ -0,0 +1,53 @@
+---
+name: review
+description: Locally review a box-box PR against its linked GitHub issue spec, run tests, capture visual screenshots when applicable, create a .review packet, and post a GitHub PR comment. Use when a ticket implementation is ready for independent local review before the human merge gate.
+argument-hint: <pr-number> [--harness <name>] [--publish-screenshots]
+---
+
+# /review — local independent PR review
+
+Review PR **$ARGUMENTS** for box-box using the local-only lifecycle. You are
+reviewing, not implementing. Do not merge the PR.
+
+## 0. Setup
+- Parse the PR number. Optional `--harness <name>` records which reviewer harness is
+  acting; optional `--publish-screenshots` allows pushing visual artifacts to a
+  dedicated artifact branch.
+- Read `.agents/prompts/review.md`, `AGENTS.md`, the PR metadata/diff, and the linked
+  issue body.
+- Identify the implementer harness from the PR body when present. If it matches the
+  reviewer harness, call that out as a reduced independence caveat.
+- Create `.review/issue-<issue>-pr-<pr>/logs`, `screenshots`, and `artifacts`.
+
+## 1. Verify
+- Run the smallest meaningful gates first, then broaden based on risk:
+  `go test` for touched Go packages, `npm run test` and `npm run build` for frontend
+  changes, and hermetic Playwright when user-facing routes changed.
+- Save all command output to `.review/issue-<issue>-pr-<pr>/logs`.
+- Inspect the diff for spec conformance, scope leaks, missing tests, and known project
+  conventions from `AGENTS.md`.
+
+## 2. Visual Packet
+- If the PR changes UI, start a local seeded or mocked preview and capture desktop and
+  mobile screenshots for affected routes.
+- Prefer hermetic mocks/seeded data over live external state.
+- Save screenshots under `.review/issue-<issue>-pr-<pr>/screenshots`.
+- Create a concise `summary.md`; create `index.html` when screenshots exist.
+
+## 3. Publish
+- If `--publish-screenshots` is present, publish only review artifacts to a dedicated
+  branch such as `review-artifacts/pr-<pr>/` and use raw GitHub URLs in the comment.
+- Post a PR comment with:
+  - reviewer harness and implementer harness
+  - result: pass, pass with caveats, or needs changes
+  - local packet path
+  - gates run and results
+  - acceptance-criteria checklist
+  - screenshots or artifact links when available
+  - caveats that the human must inspect
+- If the result is pass/pass-with-caveats, set the linked issue's custom Project
+  `Stage` to `In Review` using `.agents/lib/gh.sh`. Do not set `Done`.
+
+## 4. Hand Back
+- Tell the human exactly what to open locally and what decision remains theirs.
+- Do not merge or delete worktrees.

.agents/skills/write-spec/SKILL.md

@@ -0,0 +1,30 @@
+---
+name: write-spec
+description: Render a groomed design into the box-box Ready-spec template and write it into a GitHub issue body, then set Effort and Priority. Called by /groom after a grill session, or run standalone as "/write-spec <issue-number>" to (re)write an issue's spec from agreed decisions. Does not change Stage.
+argument-hint: <issue-number>
+---
+
+# /write-spec — write the Ready spec into an issue
+
+Target issue: **$ARGUMENTS** (box-box, Project #2).
+
+## Steps
+1. `source .agents/lib/gh.sh`
+2. Gather the agreed design decisions: from the current grooming conversation if one
+   is in progress; otherwise ask the user for the key points, or read the issue and
+   explore the code to draft them and confirm.
+3. Read `.agents/prompts/ready-spec.md` and fill every placeholder:
+   - Concrete, **behavioural** acceptance criteria (checkboxes).
+   - Technical approach grounded in **real files/paths** and the chosen data source.
+   - Test plan mapped to the actual suites (`go test` · `vitest` · hermetic Playwright).
+   - Explicit out-of-scope.
+   - Keep the Definition of Done checklist verbatim.
+   - Stamp the footer: date (from the current-date context), Effort (S/M/L),
+     Priority (P0–P2).
+4. Write it into the issue body: save the filled template to a temp file under the
+   scratchpad and `set_issue_body <n> <file>`. The spec is the single source of truth —
+   only preserve prior body text that captures decisions the spec doesn't.
+5. Set fields: `set_effort <n> <S|M|L>` and `set_priority <n> <P0|P1|P2>`.
+6. Print the issue URL and a one-line summary of what was written.
+
+Do **not** change Stage — `/groom` owns state transitions.

.claude/skills

@@ -0,0 +1 @@
+../.agents/skills

.gitignore

@@ -22,7 +22,8 @@ frontend/node_modules/
 frontend/dist/
 
 # Local Claude workspace settings
-.claude/
+.claude/*
+!.claude/skills
 
 # Playwright
 node_modules/
@@ -31,3 +32,9 @@ node_modules/
 /blob-report/
 /playwright/.cache/
 /playwright/.auth/
+
+# agentic dev: isolated implementation worktrees
+.worktrees/
+
+# agentic dev: local review packets
+.review/

docs/PRODUCT_ROADMAP.md

@@ -0,0 +1,198 @@
+# box-box — Product & UX Roadmap
+
+> An F1-geek + product/UX analysis of box-box as an F1 companion, and a
+> prioritized plan to make it richer than "data in tables" — for both live race
+> weekends and the quiet stretch between rounds.
+
+## Where the app is today (honest read)
+
+box-box has a strong data layer with two front doors (TUI + web) and six web
+surfaces:
+
+| Surface | What it does today | Where it's thin |
+|---|---|---|
+| **Command Center** (`/`) | Hero weekend + countdown, season calendar, championship snapshot, weekend schedule, news | Strong launcher, but tells you *when*, rarely *what to care about*. It's a menu, not a companion |
+| **Race Hub** (`/race-hub`) | Per-session tabs: Overview, Race Story (position-evolution scrubber), Strategy (stints), Laps, Conditions, Race Control | `RaceStoryCanvas` is the best "beyond tables" work in the app. Everything else is still tables |
+| **Live Timing** (`/live`) | SSE timing tower, battle chips, gap sparklines, pinned drivers, RC feed | Solid tower, but it's "a spreadsheet that updates" — no track map, no telemetry, no "what just happened" |
+| **Championship** (`/championship`) | Standings + hub stats (wins/poles/form) + points simulator | Good; the simulator is a genuine differentiator |
+| **Briefing** (`/briefing`) | 7 RSS feeds (FIA, BBC, Autosport, RaceFans, Guardian, RACER, F1 YouTube) + readability extraction | Passive reading list, not tied to the season narrative |
+| **Data Library** (`/admin`) | Ingestion / coverage admin | Fine as-is |
+
+**Biggest finding:** the app already *pulls* the richest data in F1 — GPS
+`Location`, full car telemetry (throttle / brake / DRS / gear / rpm / speed),
+team-radio audio, mini-sector `Segments`, speed traps, overtakes — and the web
+app visualizes almost none of it. The **track map exists only in the TUI**
+(`internal/ui/trackmap.go`); telemetry is fetched and shown to nobody. This is a
+large latent asset.
+
+## Core product thesis
+
+An F1 companion serves two different jobs; the app currently treats them the same
+(data, in tables, per session):
+
+1. **On a race weekend — the "second screen."** The user is watching the
+   broadcast (or can't, and wants to *feel* it). They want: what's happening now,
+   why it matters, what to watch next. The broadcast supplies emotion; box-box
+   should supply **the data the broadcast doesn't show** — the delta the director
+   cut away from, the undercut developing, the tyre cliff approaching.
+
+2. **Between races — "understand the season."** No live action. The user wants to
+   make sense of what happened and anticipate what's next: rewatch the story of
+   the last race, argue strategy, track the title fight, get smart before the next
+   round.
+
+Everything below makes each job *feel* like a companion instead of a database.
+
+---
+
+## Race-weekend experience ("second screen")
+
+### 1. Live Track Map — highest-leverage missing feature
+Port the TUI's GPS outline (`trackmap.go`, `GetLocation`) to a web SVG dot-map:
+cars as team-colored dots on the circuit outline, DRS zones highlighted, sectors
+tinted by status. Turn mini-sectors purple/green live using the `Segments` data
+already fetched. Tap a car → mini-telemetry readout. This is what turns "updating
+spreadsheet" into "I'm watching the race."
+
+### 2. "What just happened" synthesized event rail — the companion voice
+Combine Race Control + overtakes + pit stops + position deltas into plain-English
+beats: *"LAP 34 — VER pits (2.4s), rejoins P4 behind NOR — undercut on RUS is
+live."* The inputs all exist; this is a synthesis layer, not new data. It's the
+difference between *data* and *commentary*.
+
+### 3. Telemetry compare overlay
+On the tower / a driver panel, pick two drivers → overlaid speed/throttle/brake
+traces + delta-time graph for their last comparable lap. *The* tifosi feature —
+"where did Leclerc lose the lap." `/api/v1/laps/comparison` + car data already
+exist; this needs a chart, not a table.
+
+### 4. Live strategy / tyre-degradation view
+Stints + tyre age + pit-lane times → live pit window and tyre-cliff panel, with
+**undercut/overcut threat** indicators from gap-to-car-behind vs. pit-loss time.
+The pit-window calculator exists in the TUI (`pitwindow.go`); bring it to the web
+and make it live.
+
+### 5. Team radio, surfaced
+`GetTeamRadio` returns audio clips. Add a "Radio" ticker on the live page — play
+button + driver + timestamp. Peak emotional content, and no rival dashboard has it
+inline.
+
+### 6. Session-aware Command Center
+When a session is live, the home hero should pull the top battles, leader gap, and
+last RC flag onto the front page instead of only saying "LIVE." Make the front
+page reactive to the moment.
+
+---
+
+## Between-races experience ("understand the season")
+
+### 7. Race replay as a first-class story
+Grow `RaceStoryCanvas` from a scrubber into a narrative replay: auto-generated
+"chapters" (start, first pit phase, VSC, decisive overtake, finish) with a
+headline each (derived from RC + position swings), and scrub the position graph
+and track map together. The "watch the race in 90 seconds" mode that makes people
+open the app on a Tuesday.
+
+### 8. Driver pages + rivalry view (a real gap)
+No driver profile exists in the web app. Add a **driver page** (season form,
+teammate H2H — already computed in the champ hub — quali vs. race pace, tyre
+management, track-by-track) and a **rivalry view** (two drivers → cumulative
+points, H2H, gap-over-season). Feeds the argument every fan has.
+
+### 9. Next-race preview / "get smart" page
+Between races the app goes quiet. Fill it: circuit characteristics, last year's
+result (2023/24 is cached), typical strategy (1 vs 2 stop), DRS zones, weather
+outlook, and the storylines (title-fight math to watch). Turn dead air into
+anticipation.
+
+### 10. Championship scenario narratives
+Extend the simulator from "drag points around" to narrative permutations:
+*"VER clinches if he outscores NOR by 9 this weekend"* / *"first race McLaren can
+seal constructors'."* The stuff fans actually search for.
+
+### 11. Briefing → calendar-aware season digest
+Reframe the news reader into a paddock digest tied to the calendar: group by GP,
+tag by team/driver, surface a "since last race" summary. Optionally an
+LLM-generated weekly briefing (a `claude-haiku` summarization pass over ingested
+RSS — cheap and on-brand).
+
+---
+
+## New / better data sources
+
+- **Jolpica (Ergast successor, `api.jolpi.ca`)** — free historical results back to
+  1950: qualifying, pit stops, lap times, circuit metadata. Unlocks all-time
+  records, "best-ever at this track," and career stats OpenF1 (2023+) can't give.
+  High value for driver pages and the "get smart" preview.
+- **OpenF1 weather timeseries** — already fetched; plot it as a session-long strip
+  (track temp / rain / wind) instead of a table. Weather narrates strategy.
+- **Circuit metadata / DRS zones / corner names** — enriches the track map and
+  previews. Some is in OpenF1 circuit info; curate the rest once as static data.
+- **Static per-race context** — a tiny curated JSON per round (tyre allocation,
+  notable stats) goes a long way for previews.
+
+---
+
+## Cross-cutting UX principles
+
+1. **Replace tables with a shape wherever a shape carries the meaning.** Lap times
+   → a trace with fastest lap marked. Stints → a horizontal tyre timeline. Gaps →
+   the sparkline already shipped. Keep tables only where data is genuinely tabular
+   (standings) — but annotate them.
+2. **Always answer "so what?"** Every number sits next to its meaning (a gap next
+   to "undercut live," a tyre age next to "5 laps from the cliff").
+3. **One primary "moment" per screen.** The Command Center should always have a
+   single obvious "here's what to watch/do now."
+4. **Make between-races feel alive.** The app currently rewards you only on
+   Sundays. Previews, digests, replays, and rivalries give a reason to open it
+   midweek.
+
+---
+
+## Suggested sequencing (impact × effort)
+
+**Phase 1 — turn live into a companion (highest impact; data already in hand)**
+1. Web track map (SVG + GPS + mini-sectors)
+2. Telemetry compare overlay (speed/throttle/brake + delta)
+3. "What just happened" synthesized event rail
+4. Team radio ticker
+
+**Phase 2 — own the between-races window**
+5. Narrative race replay (grow `RaceStoryCanvas`)
+6. Driver pages + rivalry view
+7. Next-race preview page
+
+**Phase 3 — depth & reach**
+8. Jolpica/Ergast historical integration + all-time records
+9. Championship scenario narratives
+10. Calendar-aware briefing digest (optional LLM summaries)
+
+---
+
+## Execution tracking (GitHub)
+
+This roadmap is tracked on GitHub:
+
+- **Project board:** https://github.com/users/AmanTahiliani/projects/2 ("box-box Roadmap")
+- **Epics:** issues #2–#8 (label `epic`), one per epic above, on Phase milestones
+- **Stories:** issues #9–#31, wired as native **sub-issues** under their epic (progress rolls up automatically)
+- **Labels:** `epic`, `enabler`, `research`, `area:{live,viz,between-races,championship,data,ux}`
+- **Milestones:** `Phase 1 — Live Companion`, `Phase 2 — Between-Races`, `Phase 3 — Depth & Reach`
+
+**Board fields:** `Stage` (Icebox → Research → Ready → In Progress → In Review → Done),
+`Priority` (P0–P2), `Effort` (S/M/L), `Phase` (1–3).
+
+**Working model — active vs. bank:**
+- Only **1–2 epics active** at a time (currently **E1 Live Race Companion** + **E2 Viz
+  Primitives**); the other five epics are the theme-level idea bank.
+- `Stage = Icebox` is the story-level bank. The 10 E1/E2 stories are seeded to
+  `Ready`; everything else is `Icebox`. Promote a handful to `Ready`/`In Progress`
+  per cycle; use `Research` to scope a vague idea before it's `Ready`.
+**Board views** (built):
+- **Backlog** — Table, all epics with expandable sub-issues: the full bank.
+- **Board** — grouped by `Stage`, filtered `-stage:Icebox`: the active-WIP wall.
+- **Roadmap** — Table grouped by `Phase` (Phase 1/2/3). Note: this is a
+  phase-grouped table, not a timeline. A true timeline roadmap needs a date or
+  iteration field (GitHub won't draw/persist a roadmap layout without one); add a
+  `Target date` field and populate it once phases have real target dates, then
+  switch this view to the Roadmap layout.