…uation

Adds opt-in `claude --resume` continuation across consecutive webhooks
of the same logical thread (ClickUp task, GitHub PR, Linear issue).
Without this, every webhook spawned a fresh Claude subprocess that had
to re-read all prior comments, re-derive context, and re-fetch data via
API calls — losing 90%+ of the model's reasoning trace between turns.

## Why

Real-world incident driving this: 2026-05-02 ClickUp task 86c9kyquv.
User asked the Oracle for a marketing report (Phase 1, ~$3.21, 22min,
43k tokens), then 30 minutes later asked to "implement recommendation 3".
The fresh follow-up Oracle had to:
  - Re-read the entire Google Doc via API call
  - Re-derive what "rec 3" meant from raw text
  - Re-discover the cuid Vitalmadente
  - Re-fetch GA4/GBP/Ads data baseline
  - LOSE the entire reasoning trace (rejected hypotheses, alternatives
    considered) that the prior Oracle had in context

With session resume, the second Oracle inherits the full conversation
history — knows what was tried, what was rejected, and why — at zero
re-derivation cost.

## How — backend

- `Trigger.resume_sessions BOOLEAN` column (default false). Existing
  triggers keep current "fresh subprocess" behaviour. Operator opts in
  per-trigger via dashboard checkbox or YAML.
- `trigger_session_threads` table maps `(trigger_id, dedup_key)` →
  `claude_session_id`. The dedup_key is extracted per-source by
  `_extract_dedup_key()` in `routes/triggers.py`:
    - ClickUp (source=custom + slug contains 'clickup'): task.id
    - GitHub: pull_request.number / issue.number
    - Linear: issue.id
  Extensible for new sources.
- `run_claude(..., resume_session_id=...)` in `ADWs/runner.py`
  prepends `--resume <id>` to the CLI invocation. Captures `session_id`
  from output JSON and returns it for upsert. Silently no-ops on
  non-Anthropic providers (OpenClaude doesn't expose --resume yet).
- `_execute_trigger` flow:
  1. If `trigger.resume_sessions`: extract dedup_key, lookup prior
     session_id from `trigger_session_threads`.
  2. Pass to `run_claude` (None = fresh).
  3. After run: if claude_session_id captured, upsert into
     `trigger_session_threads`.
  4. If resume failed (stale session): retry once without resume.

## How — UI

- `/settings → Sessions` tab (new):
  * Default-on toggle for new triggers
  * Auto-cleanup window (1-365 days) + daily cleanup hour (0-23)
  * Force compaction turn count (1-500)
  * Storage stats (count + disk usage of `~/.claude/projects/`)
  * Live list of active threads with manual reset + bulk cleanup-stale
- Trigger edit form: "Enable session resume" checkbox with explanatory
  inline tooltip.

## Endpoints

- `GET /api/settings/sessions` — global config + storage stats
- `PUT /api/settings/sessions` — update defaults
- `GET /api/sessions` — list active threads with staleness flag
- `DELETE /api/sessions/<id>` — manual reset
- `POST /api/sessions/cleanup-stale` — bulk delete rows older than
  `auto_cleanup_days`

## Schema

Alembic migration `0012_clickup_session_resume.py`:
- ALTER `triggers` ADD `resume_sessions BOOLEAN NOT NULL DEFAULT FALSE`
- CREATE `trigger_session_threads` (id, trigger_id FK, dedup_key,
  claude_session_id, last_used_at, created_at) with UNIQUE
  (trigger_id, dedup_key) and index on (trigger_id, last_used_at).
- Reversible via downgrade().

## Backward compatibility

- All existing triggers keep `resume_sessions=false` (no behaviour change).
- `run_claude` callers unchanged unless they explicitly pass
  `resume_session_id`.
- Schema migration is idempotent (checks `_has_column` / `_has_table`
  before alter/create).

## Cost / quality impact (measured on 86c9kyquv)

Without resume (status quo):
- Oracle 1: 22min, 43k tok, $3.21
- Oracle 2 (re-read everything): 11min, 18k tok, $2.19
- Total: $5.40 + degraded continuity

With resume (projected, prompt cache + reused context):
- Oracle 1: 22min, 43k tok, $3.21
- Oracle 2 (resume): 4min, 8k tok, $0.80
- Total: ~$4.00, 25% saved + full reasoning continuity

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