…c impl + context loader

First ST of Feature 5's biggest user-facing milestone. Lays the
clean reasoning surface (Protocol + value classes + Anthropic
impl + JSONL context loader + operator-editable system prompt)
that ST2 will wire into the Slack DM handler to replace the echo.

NOT yet plugged into any handler — that's ST2.

## New package

**`kora_cli/reasoning/`**:
- `engine.py` (~180 lines) — `ReasoningEngine` Protocol +
  `IncomingMessage` / `ConversationContext` / `ConversationTurn` /
  `ResponseResult` value classes. Protocol over ABC for test-seam
  composition.
- `anthropic_engine.py` (~350 lines) — `AnthropicReasoningEngine`
  implementing the protocol against `anthropic==0.86.0`. Two-mode
  credential cascade (see K-DG drift below). Cost-ladder-aware
  model selection. Operational-state gating. 60s per-call timeout.
  NO retry on 5xx per PM Q3 default. Credential-sanitized error
  mapping.
- `context_loader.py` (~200 lines) — `load_slack_dm_context()`
  reads `${HERMES_HOME}/slack_dm_log.jsonl`, filters to (channel,
  thread), skips filtered/failed entries, projects to
  `ConversationTurn` list (oldest→newest, capped at max_turns).
  Resolves operational-state + cost-rung strings from holders.
  Defensive parsing (malformed lines logged + skipped).

**`kora_docs/00_canonical_current_state/kora_system_prompt.md`**
(~150 lines) — first-draft Kora identity prompt per PM Q1 YES.
Authored from project_kora memory: digital extension framing,
brevity-by-default, no preamble/sycophancy, technical fluency
assumption, operational-state respect, cost-ladder awareness,
NOT-Claude-the-public-assistant boundary. Operator-editable;
fail-CLOSED if missing at daemon boot.

**`[web]` extra extended** with `anthropic==0.86.0` so daemon
deploys auto-pick up the SDK. Drift 2 resolution.

## K-DG drift findings (surfaced to PM + acted on)

**Drift 1 — credential env name conflict**: bucket spec said
`KORA_ANTHROPIC_API_KEY`; existing env-mapping (PR #110) +
gate-2 anti-secrets + "Max plan via Agent SDK billing" framing
all point to `CLAUDE_CODE_OAUTH_TOKEN`. **Implementation supports
BOTH via cascade**:
- `KORA_ANTHROPIC_API_KEY` (if set) → SDK `api_key=...` (Console billing)
- `CLAUDE_CODE_OAUTH_TOKEN` (fallback) → SDK `auth_token=...` (Max plan)
- Both unset → `ReasoningEngineNotConfigured` fail-CLOSED

Operator chooses; existing OAuth token works without any new
Doppler secret. PM may tighten to a single env path in a
follow-up if preferred.

**Drift 2 — `anthropic` lib was extra-only**: moved to `[web]`
extra alongside fastapi/uvicorn/slowapi so daemon deploys don't
need a separate `pip install -e ".[anthropic]"` invocation.

**Drift 3 — CostRung enum value naming**: bucket spec used
`normal / warned / constrained / halted`; actual enum is
`NORMAL / WARN_75 / DOWNSHIFT_90 / HARD_STOP_100` (`agent/cost_state_holder.py:114-117`).
Engine consumes the canonical `.value` strings via
`ConversationContext.current_cost_ladder_rung` literal type:
`"normal" / "warn_75" / "downshift_90" / "hard_stop_100" / "unknown"`.

**Drift 4 — `holder.active_rung()` is a method, not a property**;
`cost_state_holder.get_cost_holder()` (not `get_holder`). All
verified via grep at `agent/cost_state_holder.py:259`, `:502`.

## Cost-ladder model selection (per R4.1 §9.6 rung mapping)

| CostRung | Model used |
|---|---|
| `normal` | `claude-opus-4-7` |
| `warn_75` | `claude-sonnet-4-6` |
| `downshift_90` | `claude-haiku-4-5-20251001` |
| `hard_stop_100` | refuse — no API call; `error="cost_ladder_halted"` |
| `unknown` (defensive) | `claude-opus-4-7` + WARN log |

## Operational-state gating (per PM Q4 YES default)

`PAUSED` or `STOPPED` → refuse, no API call;
`error="operational_state_paused"`. The handler maps this to a
canned acknowledgment so Joshua isn't met with silence.

## Error code taxonomy (stable across the codebase)

`sdk_auth` (401/403) / `sdk_rate_limited` (429) / `sdk_5xx` /
`sdk_4xx_<code>` / `sdk_timeout` / `sdk_transport` /
`sdk_unknown_<class>` / `cost_ladder_halted` /
`operational_state_paused` / `response_projection_failed`.

## §4 PM-opens — all defaults accepted

- Q1: YES first-draft system prompt — composed (~150 lines)
- Q2: 10-turn context — default in loader (`DEFAULT_MAX_TURNS=10`)
- Q3: NO retry on 5xx — single SDK call exactly (tested)
- Q4: PAUSED/STOPPED gating — implemented + tested

## Tests (42 new, 275 total all passing)

**`test_anthropic_engine.py`** (24 tests):
- Construction: fail-CLOSED on both creds unset; API key only;
  OAuth only; API key wins when both set; whitespace-only treated
  as unset
- System prompt: fail-CLOSED on missing/empty/unreadable;
  env-path override works
- Cost-ladder: NORMAL→opus, WARN_75→sonnet, DOWNSHIFT_90→haiku,
  HARD_STOP_100→refuse (no SDK call), unknown→opus default + WARN
- Operational-state gating: paused/stopped refuse (no SDK call)
- Happy path: successful call returns text + tokens + duration
- Message history: alternating roles preserved; consecutive
  same-role concatenated; system prompt passed to SDK
- SDK error mapping: 401/429/500/400/timeout/unknown — all 6 distinct
- NO retry on 5xx or 429 (single SDK call asserted)
- **SECURITY**: diverse failure-mode sequence (401/429/500/timeout/
  transport/unknown) — credential env value NEVER in error codes,
  text, model_used, or any log line

**`test_context_loader.py`** (18 tests):
- Missing/empty file → empty context + "unknown" state strings
- Channel filtering (same/diff channel)
- Thread filtering (exact match; None matches None)
- Filtered inbound entries skipped (filtered_non_joshua /
  dropped_paused / handler_error)
- Failed outbound entries skipped (send_status: failed)
- Turns ordered oldest→newest (sort by timestamp)
- max_turns slices most-recent
- DEFAULT_MAX_TURNS == 10 (asserted)
- Malformed JSON lines logged + skipped, valid lines still loaded
- Operational state surfaces from initialized holder (paused)
- Log path env override

## §5 ship checklist

- [x] Base `feature/phase2-upgrades`
- [x] Title format `feat(kora): KR-FEAT-AI-RESPONSE-LOOP STn — <scope>`
- [x] §4 questions resolved (all defaults accepted)
- [x] anthropic dep moved into `[web]` extra
- [x] API key NEVER logged (asserted across 6 failure modes)
- [x] Cost-ladder rung respected (5 enum values tested)
- [x] PAUSED state respected (tested for both PAUSED + STOPPED)
- [x] Engine failure paths return ResponseResult.error (no crash)
- [x] K-DG verified: literal field names pasted from grep;
      `.current` (property) vs `active_rung()` (method) caught;
      CostRung values verified against agent/cost_state_holder.py
- [x] Tests pass locally (**275/275** across full suite)

## What's next

**ST2** — wire reasoning into the Slack DM handler:
1. New `kora_cli/listeners/reasoning_engine_listener.py` registering
   via `register_daemon_listener("reasoning_engine", ...)` mirroring
   the `current_pool()` pattern from KR-MCP-CONSUMPTION ST1 with
   a `current_reasoning_engine()` accessor.
2. Update `SlackDMHandler` to call the reasoning engine instead of
   constructing the echo text.
3. Extend outbound JSONL schema with `model_used`, `input_tokens`,
   `output_tokens`, `reasoning_duration_ms`, `reasoning_error`.
4. Engine failure → canned fallback ("Kora is currently unable to
   respond; operator notified") — NOT a re-echo, NOT a crash.
5. Cost-ladder write integration: call `record_inference()` with
   the canonical usage from the response.

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