Case Cockpit: surface dropped state, per-stage actions, live data

[toasterman234]

Plan — Mira Case Cockpit: surface dropped state, per-stage actions, live data

Written 2026-07-15. Self-contained: assumes zero prior context. Intended to be filed as a
GitHub issue on byteseek/Mira and picked up by an implementing agent.

Context (why)

The dashboard at dashboard/ (see dashboard/HANDOFF.md) is a read-only mirror of Mira's
research cases. It shows where a case sits (board stage + cockpit: route, readiness ladder,
expectation map, evidence bars, artifacts) but gives the operator nothing to do from there.

Mira itself is not a state store — it is a set of loops/skills (actions) in loops/ and
skills/. So the dashboard today drops the parts that make Mira Mira, and offers no way to act
on a card. Ben reviewed it and named three gaps to close:

1. Surface dropped state — the cockpit throws away thesis-ledger, decision-log, refresh
   triggers, expectation-map surprise columns, and evidence conflict detail.
2. Per-stage action menus — each stage has specific legitimate next moves (real Mira loops).
   The dashboard should offer them, turning the mirror into a launcher.
3. Live data connector — every case is connector: none. A real tools/mira_data Futu/IBKR
   connector already exists; wire its output in.

Architecture decision: stay static for v1 (Python build step → data/cases.json → vanilla JS,
no running backend), matching Mira's stdlib-only convention (tools/mira_pipeline_sync.py,
tools/mira_data). Per-stage actions copy the exact Mira prompt/command to paste into an
agent; live prices are folded in at build time from existing private/futu|ibkr/ snapshots.
Leave a clean, documented seam so a small local backend can later execute actions and fetch live
prices on click. Do not build the backend in v1.

Hard invariant: the dashboard is read-only over cases/. build_data.py must never write into
case files. (This is why parsing helpers were copied out of mira_pipeline_sync.py rather than
imported — that module writes frontmatter at import time.)

Files touched

• dashboard/build_data.py — new parsing (ledger, decision-log, refresh radar, conflict detail,
  live snapshot fold). Stdlib only.
• dashboard/app.js — new cockpit panels, expectation-map columns, per-stage action menu +
  copy-to-clipboard, board overdue badge, live-data rendering.
• dashboard/style.css — styles for new panels, action popover/modal, badges.
• dashboard/index.html — minor (popover/modal container if needed).
• dashboard/HANDOFF.md — document the live-snapshot fetch step and the backend seam.

---

Feature 1 — Surface dropped state

build_data.py (new pure helpers, wired into build_case())

• parse_ledger(case_dir) → read thesis-ledger.md if present. Extract:
  • the - key: value metadata block (state, last_updated, selected_framework,
    selected_overlays, stale_after) with the same line-anchored regex style as fm_get;
  • the first paragraph under ## Current Thesis;
  • ## Key Assumptions items, and any kill_criteria / reversal_condition lines.
    Return a dict or None. (Sample shape confirmed in cases/aapl-2026-04/thesis-ledger.md.)
• parse_decision_log(case_dir) → load_csv_rows("decision-log.csv"); keep
  date, research_object, decision_type, basis, risk, required_followup. Return list or None.
  (Header confirmed in cases/aapl-2026-04/decision-log.csv.)
• Refresh radar — add to each case:
  • refresh_status ∈ overdue | due_soon | ok | unknown and days_to_stale (int|null),
    computed by extracting the first DATE_RE match from the stale_after frontmatter field and
    comparing to datetime.date.today() (allowed — this is the build step, not the sandboxed
    research workflow). due_soon = within 14 days.
  • must_refresh_if and kill_criteria from README frontmatter when present.
• Evidence conflict detail — in build_evidence(), alongside conflict_count, add
  conflicts: a short list of {claim, conflict_status, source_id} for rows whose
  conflict_status ∉ {none, ""}. Cap at ~10 to keep JSON small.
• Expectation map already passes full CSV rows — no builder change; the extra columns
  (consensus_proxy, evidence_status, next_check) are surfaced in the frontend.

app.js (new/updated cockpit panels)

• Thesis Ledger panel — state badge (stale/active), current-thesis text, kill/reversal
  criteria, key assumptions, framework + overlays. Empty-state when no ledger.
• Decision Log panel — compact timeline: date · decision_type · basis (truncated) · followup.
• Expectation Map — extend renderExpectationMap() to show consensus_proxy and next_check
  (and evidence_status as a chip). This is the "where's the surprise vs consensus" view.
• Refresh chip in the case header — red "overdue Nd" / amber "due in Nd" from refresh_status.
• Evidence panel — make conflict_count expandable to list evidence.conflicts.
• Board cards — add a small overdue dot/badge when refresh_status === "overdue".

---

Feature 2 — Per-stage action menus (launcher)

Add a STAGE_ACTIONS config in app.js: pipeline_stage → [{ label, kind, template }].
kind ∈ prompt | shell | open (this typing is the backend seam — a future server executes by
kind; v1 copies template to clipboard via navigator.clipboard.writeText). template is a
function of the case (interpolates slug / ticker / label / loop path).

Stage → actions (sourced from AGENT_QUICKSTART.md routing table + loops/):

Stage	Actions (each yields a real Mira prompt/command)
idea	Route it → loops/analysis-routing.md · Promote to Queued
queued	Run first-pass loops/research-loop.md (paste task card for this case)
intake	Confirm route · pick framework/overlay · run the routed skill
researching	Continue loop · complete expectation-map.csv · run gates
package_complete	Review → refresh / extend / archive · loops/monitoring-loop.md · loops/thesis-update-loop.md · check actionability-bridge.md
needs_refresh	Refresh now → scripts/check_updates.sh then loops/monitoring-loop.md / loops/event-delta-loop.md (show the stale-trigger date)
archived	Reference only · reopen as new case

Card-level actions (any stage): Open memo / evidence-log / expectation-map (copy path or
file:// link) · Copy next-step prompt (from next_action) · Copy case folder path
cases/<slug>/.

UI: clicking an action opens a small popover/modal showing the exact text with a Copy button;
no execution in v1.

---

Feature 3 — Live data connector (static snapshot-fold + seam)

build_data.py

• load_live_snapshot(ticker) → look under private/futu/<TICKER>* and private/ibkr/<TICKER>*
  for the newest emitted snapshot JSON. Implementation step: read tools/mira_data/emit.py and
  tools/mira_data/adapters/ to match the exact emitted JSON shape (last price, as-of timestamp,
  source id). Return {price, as_of, source, connector} or None.
• build_live_data() — when a snapshot exists, set connector to futu_opend_local /
  ibkr_gateway_local, fill real price / as_of / freshness (compare as-of to today), else keep
  the honest connector: "none" default. Never fabricate a price.

app.js

• renderLiveData() — when connector !== "none", show live price / as_of / source / freshness
  with a live-vs-stale badge; keep the current honest empty state otherwise.

Operator step (document in HANDOFF.md)

PYTHONPATH=tools python -m mira_data fetch futu_market_price US.<TICKER> --out private/futu/<TICKER>
python3 dashboard/build_data.py   # re-fold the snapshot into cases.json

private/ is gitignored — snapshots stay local. No account ids / positions enter tracked files
(per docs/futu-private-data.md, docs/ibkr-private-data.md).

Backend seam (v2, documented not built)

A thin dashboard/server.py (stdlib http.server) exposing /api/live/<slug> (shells
mira_data fetch) and /api/run (executes a stage action by kind). Frontend prefers these
endpoints when reachable, falls back to static JSON / clipboard otherwise. The kind typing on
stage actions and the connector field already make this a drop-in.

---

Verification

1. python3 dashboard/build_data.py → runs clean, 36+ cases, no writes into cases/
   (git status cases/ clean after run). Spot-check aapl-2026-04 in data/cases.json: ledger
   fields, decision-log rows, refresh_status, expectation-map extra columns, conflicts list.
2. Serve: python3 -m http.server 8420 --directory dashboard, open in browser. Confirm:
   • Cockpit shows Thesis Ledger + Decision Log panels for aapl-2026-04.
   • Expectation map shows consensus_proxy + next_check.
   • needs_refresh cases past their stale date show an overdue badge on board + cockpit.
   • Each card/stage shows an action menu; clicking an action copies the correct Mira prompt.
   • Live panel: drop a sanitized test snapshot into private/futu/AAPL, re-run build, confirm
     the live price renders; then remove the test snapshot.
3. Re-run build_data.py; confirm data/cases.json still gitignored and cases/ untouched.

Out of scope (v1)

• The local backend server (seam only).
• Drag-to-change-stage / any write-back into case frontmatter (stage is owned by
  mira_pipeline_sync.py, not the browser).
• The mira_pipeline_sync.py regex fix (tracked separately in HANDOFF.md).

Filing

After approval, file this as a GitHub issue on byteseek/Mira via gh issue create (title:
"Case Cockpit: surface dropped state, per-stage actions, live data"), body = this plan.
byteseek/Mira is a public repo — confirm that's the right target before filing.