Operator-facing view of Kora's recent ReasoningEngine calls. Pairs
with CC#3's KR-FEAT-AI-RESPONSE-LOOP ST2 (in flight) — once ST2
extends slack_dm_log.jsonl outbound entries with the model_used /
input_tokens / output_tokens / reasoning_duration_ms /
reasoning_error fields, a small follow-on bucket flips this stub
to read from there. Shape is pinned by the new test suite.

Single-PR scope:
  * GET /api/reasoning/recent stub — 4 representative calls
    deliberately spanning ok @ normal on opus, ok @ warn_75 on
    sonnet (cost-downshift), halted at hard_stop_100 (budget-
    locked refusal), and sdk_timeout failure so the operator's
    first look surfaces the full cost-ladder behaviour + error
    taxonomy. stub:true keeps the FE banner visible.
  * ReasoningPanel.tsx — title + stub banner + 4-column stats
    strip (total calls / token spend / model distribution /
    status distribution) + filter pills (all / ok / failed /
    halted) + newest-first timeline. Per row: status icon,
    model badge (tier-tinted: opus blue / sonnet purple / haiku
    gray / null=halted red), cost-rung pill (color-mapped per
    spec: normal gray / warn_75 yellow / downshift_90 orange /
    hard_stop_100 red), duration with visual bar (warning ≥ 2s),
    input→output token counts, status badge (non-default only),
    error_code chip, response excerpt (truncated to 80 chars
    collapsed; expandable to the 200-char API cap).
  * Dashboard card #12 with Brain icon. Layout keeps lg:grid-
    cols-3 — row 2 grows to 3+3+3+3 (clean 4-row × 3-col closure
    of the layout that's been shepherded since DASHBOARD-V2).
    Headline goes destructive when halted > 0 in 24h (Kora was
    budget-locked; investigate cost rung); warning when failed > 0;
    foreground/success otherwise.
  * Route /reasoning + nav entry between /agent-activity and
    /slack-dm (logical grouping with the agent-facing surfaces).

4-layer security contract (extending the established pattern with
reasoning-specific guards):
  1. response_text_truncated_200 rendered as PLAIN TEXT via
     React's default child escaping. Real responses may contain
     anything Kora generates (HTML / markdown / script fragments).
     FE pins via dangerouslySetInnerHTML grep.
  2. NO Anthropic key shapes (sk-ant- prefix + base64-like body)
     anywhere in payload. Walk-payload regex catches a future
     error-projection bug or log-entry edit that leaks credential
     material into the operator's view.
  3. NO PII (email regex / Slack user-ID regex) leaked from the
     inbound user's message into response_text_truncated_200 —
     Kora's generated text is operator-visible; user message
     content lives in SLACK-DM-PANEL, not here. Belt+braces
     walk-payload sweep covers any field, not just response_text.
  4. TS interface declares all fields with documented contracts;
     no raw_prompt / auth_token / response_html companion fields
     exist on ReasoningCall.

K-DG drift catch (per memory `feedback_k_dg_substrate_field_names`):
  * Bucket §3(a) example payload used uppercase Enum NAMES
    (NORMAL / WARN_75 / DOWNSHIFT_90 / HARD_STOP_100) — same
    convention CC#3 cited in PR #126's K-DG drift note.
  * BUT the canonical wire format per
    agent/cost_state_holder.py:114-117 is the lowercase Enum
    VALUES (NORMAL = "normal" etc), and engine.py:47-49's
    CostLadderRungName literal enforces lowercase.
  * Real CC#3 data will emit lowercase via .value. Stub uses
    LOWERCASE to match — so stub and real data agree at flip
    time. Backend test pins lowercase + flags this as the
    intentional spec divergence.

Tests:
  * tests/kora_cli/test_web_server_reasoning.py — 16 tests:
    shape, 4-call stub pin, status+rung+error diversity, per-
    entry schema with full ReasoningEngine error taxonomy
    validation, cost-rung lowercase pin (K-DG catch), all 4
    security guards (Anthropic key walk-payload + PII walk-
    payload both per-field and whole-blob), FE source-pins
    (no dangerouslySetInnerHTML, response_text as JSX child),
    200-char API-edge cap, by_status_24h reconciliation,
    tokens_total shape, cron-regression sanity.
  * Full admin-panel regression: 287/287 across 24 suites
    (with --extra dev only — slowapi now runtime per PR #128).
  * tsc -b + vite build both clean.

Refs:
  * rafe-walker/kora-docs 17_cc_bucket_prompts/KR-REASONING-PANEL_kora_thinking_lens_stub.md
  * PR #126 — KR-FEAT-AI-RESPONSE-LOOP ST1 (ReasoningEngine +
    error taxonomy + CostRung literal source)
  * PR #128 — slowapi runtime promotion (enables --extra dev only)

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>