redmine/docs/contact_triage_loop_workflow.md

# Contact Triage Loop Workflow (Vision + Operating Pattern)

## Why this exists

This document captures Jason's intended **repeatable contact-review loop** so work can proceed in deterministic batches without reprocessing the same contacts immediately.

Captured from uploaded project note on **2026-07-04** (saved into repo context to avoid losing the design intent).

Raw source capture:
- `/home/jason/projects/redmine/docs/contact_triage_loop_raw_notes_2026-07-04.md`

---

## Core loop pattern

This is a **multi-pass routing system**, not a final one-shot classifier.

1. Select a small batch (10-20 contacts) where `updated_on` is old enough and optional tag constraints match.
2. Sort **newest stale first** (recently stale / last-updated descending), so the loop works through the freshest cooled-down contacts first.
3. For each contact:
   - quick contact-only triage
   - inspect recent helpdesk tickets/messages when the current pass requires it
   - route by adding/removing tags that represent either classifications or workflow state
4. Apply tag updates:
   - if the pass identifies a route, add that route/classification tag
   - otherwise add a status marker like `ai-checked`
5. Because contact metadata was updated, `updated_on` advances.
6. On future runs, recently processed contacts fall out of the immediate stale batch; later passes can pick them up again when stale or when a designated tag routes them to another worker.
7. Later filters/workers may deliberately remove status tags such as `ai-checked` after consuming them, allowing the contact to re-enter another pass of the conveyor belt.

This forms a self-advancing conveyor belt for incremental cleanup/enrichment. Tags are not all semantic classifications: some are routing/status markers. For example, `ai-checked` means this pass touched the contact; it does not mean the contact's final identity is "AI checked." Likewise, `spam`/`promotions` can be step 1 in a longer handling pipeline, not the final disposition.

Operational objective: reducing clutter in the customer-service/sales channel is valid even when a sender is technically a vendor, billing system, automailer, or promotional sender, because Redmine customer-service is not the only channel available to them.

---

## Existing tooling alignment (current)

`redmine_contacts.py` already supports key building blocks:

- batch candidate selection:
  - `batch-update --max-contacts`
  - date/tag filters (`--updated-before`, `--with-tag`, `--without-tag`, `--no-tags`)
  - deterministic ordering (`--order newest-updated-first` etc)
- safe mutation workflow:
  - dry-run by default
  - write only with `--apply`

What is still missing for full loop automation:

- richer FILTER-2 workers that classify contacts and mutate tickets/contacts when needed
- scoring engines that map evidence to tags with thresholds
- downstream queue workers for enrichment/NMI/company/person linking

Implemented tracer pieces:

- `redmine_contacts.py loop-run`
- profile contract in `docs/contact_loop_filter_contract.md`
- sample tracer profile: `profiles/contact-loop/noop-ai-checked.json`
- sample no-op FILTER-2 worker: `profiles/contact-loop/filters/noop_filter2.py`

The intended architecture is now: **runner owns conveyor-belt mechanics; FILTER-2 owns domain-specific ticket/contact mutation logic.**

---

## Suggested scoring framework (v0)

### customer
Signals:
- internal/team response exists (`+1`)
- discussion indicates warehouse packages/items (`+1`)

### promo + spam
Signals:
- sender hints (`no-reply`, `noreply`, `deals`, `promotions`) (`+1`)
- message mentions discount/coupon/deal language (`+1`)

Threshold model (initial):
- tag set applied when score >= `2`
- if no tag set crosses threshold: add `ai-checked`

---

## Diagrams

### Diagram 1 — Conveyor-belt loop

```mermaid
flowchart TD
  A[Candidate pool\nRedmine contacts] --> B[Select stale contacts\nupdated_before + tag filters\nmax 10-20]
  B --> C[Sort newest stale first\nupdated_on descending]
  C --> D{For each contact}
  D --> E[Run current pass\nFILTER-1 / worker gate]
  E -->|pass can route immediately| F[Add route/status/classification tags]
  E -->|needs evidence| G[Fetch recent tickets/messages]
  G --> H[Manual review or FILTER-2 worker]
  H --> I{Route decided?}
  I -->|yes| F
  I -->|no| J[Add status marker\nai-checked]
  F --> K[Update contact/tickets as allowed]
  J --> K
  K --> L[updated_on advances\ncontact leaves immediate batch]
  L --> M[Later passes use staleness + tags\nto route next work]
  M --> N[Later worker may remove status tag\ne.g. ai-checked]
  N --> B
```

### Diagram 2 — Tag/status re-entry lifecycle

```mermaid
stateDiagram-v2
  [*] --> Eligible: updated_on older than cutoff\nand tag filters match
  Eligible --> InCurrentPass: newest-stale-first selection
  InCurrentPass --> Routed: worker adds route tag\nspam/promotions/etc.
  InCurrentPass --> Checked: no route yet\nadd ai-checked
  Routed --> CoolingDown: tag update bumps updated_on
  Checked --> CoolingDown: status tag update bumps updated_on
  CoolingDown --> HeldOut: too fresh for current cutoff
  HeldOut --> Eligible: time passes / becomes stale again
  Checked --> DownstreamQueued: later worker selects ai-checked\nor another status tag
  DownstreamQueued --> ReEntered: worker removes ai-checked\nafter consuming status marker
  ReEntered --> Eligible: tag filters match next pass
```

---

## Future queue/tag orchestration pattern

The long-term architecture should treat tags as explicit work queues, for example:

- `needs-summary` -> summary worker processes recent communications, then removes tag
- `needs-enrichment:*` -> enrichment workers (social/entity/profile data)
- `needs-nmi-mapping` -> account-linkage worker for NMI mapping tasks
- `special-handling:*` -> SLA/VIP/high-risk handling workflows

This lets one loop classify and route, while specialized workers execute downstream tasks.

---

## Proposed implementation phases

### Phase 1 (implemented tracer)
- Add `loop-run` command that performs a **per-contact review loop** in dry-run mode:
  - select candidates
  - run FILTER-1
  - fetch recent tickets/messages
  - call a profile-defined FILTER-2 command
  - print proposed runner-owned tag updates

### Phase 2
- Replace no-op FILTER-2 with a classifier/mutator worker.
- Use FILTER-2 to update contacts/tickets directly through helper CLI calls.
- Keep runner-owned fallback `ai-checked` and RTAG removal semantics.

### Phase 3
- Add queue tags + worker handoff contracts (`needs-summary`, `needs-nmi-mapping`, etc.).

### Phase 4
- Integrate external systems (NMI mapping tables, obsidian profile updates, comm summaries, etc.) with strict scoped guards.

---

## Practical run pattern (current CLI)

Dry-run stale-first loop tracer:

```bash
./redmine_contacts.py \
  --base-url http://192.168.50.170 \
  --project customer-service \
  loop-run \
  --profile profiles/contact-loop/noop-ai-checked.json \
  --source cache \
  --updated-before 2026-06-01 \
  --without-tag ai-checked \
  --max-contacts 10 \
  --order newest-updated-first
```

Apply runner-owned fallback/default tags after reviewing output:

```bash
./redmine_contacts.py \
  --base-url http://192.168.50.170 \
  --project customer-service \
  loop-run \
  --profile profiles/contact-loop/noop-ai-checked.json \
  --source cache \
  --updated-before 2026-06-01 \
  --without-tag ai-checked \
  --max-contacts 10 \
  --order newest-updated-first \
  --apply
```

Older dry-run batch-update skeleton:

```bash
./redmine_contacts.py \
  --base-url http://192.168.50.170 \
  --project customer-service \
  batch-update \
  --source cache \
  --updated-before 2026-06-01 \
  --with-tag customer \
  --max-contacts 20 \
  --order newest-updated-first \
  --set background=ai-checked
```

> Replace payload fields/tags only after reviewing dry-run output.