Conflict Detector ability

[hassan1731996]

What is Conflict Detector?

A passive ability that silently extracts your plans and commitments from natural conversation and cross-checks every new one against existing ones for scheduling conflicts. When a clash is detected, it interrupts immediately — not a vague reminder, a specific alert naming both commitments and explaining exactly why they conflict.

Just talk naturally. Say "I'll call Marcus on Friday" today and "flying out Thursday night" tomorrow — it will catch the conflict and tell you.

Key Features

Passive Capture

• Listens every 15 seconds, zero friction
• Word count gate + LLM commitment extractor
• Captures type (meeting/call/travel/task/social/deadline), people involved, date, time, and duration per commitment
• Natural date resolution: day names, relative phrases ("after the weekend", "in 3 days"), specific dates ("May 10th")
• 70% word overlap dedup prevents capturing the same plan twice
• Only captures definite commitments — skips maybes, hypotheticals, and vague plans with no time reference

Conflict Detection

• Two-phase check: free same-date scan first, LLM only when a candidate is found
• Adjacent-day travel detection — overnight travel on Thursday flagged against Friday morning commitments
• Conflict pair dedup so the same clash never fires twice
• Severity classification: hard (impossible to do both) vs soft (tight but possible)
• Fires a targeted interrupt: "Heads up — you said you'd [A] but you also mentioned [B]. [reason]."

Interactive Queries

• LIST: Upcoming commitments sorted by date, with conflict count
• CONFLICTS: Open conflicts sorted by severity, dismiss individually or all at once
• ADD: Manually log a commitment with date and optional time
• DISMISS: Remove a specific conflict (with confirmation)
• CLEAR: Wipe all commitments and conflicts (with confirmation)

Smart Maintenance

• Daily stale commitment expiry — past commitments auto-archived so they don't cause false positives
• Startup notification if unalerted conflicts exist from a previous session
• Periodic pointer persistence for reconnect recovery

How It Works Under the Hood

background.py (daemon)

• Polls message history every 15 seconds
• Skip phrase filter blocks agent-directed and hypothetical utterances before LLM
• Phase 2 LLM extracts structured commitments: text, type, people, date_hint, time_hint, duration_hint
• _resolve_date: pure Python natural language → ISO date
• _run_conflict_check: same-date scan (free) → LLM conflict check (targeted)
• _llm_check_conflict: returns conflicts/reason/severity JSON
• _expire_stale_commitments: runs once per day, marks past commitments expired
• Conflict alert fires with send_interrupt_signal() + speak(), flag set before speak to prevent double-fire

main.py (interactive skill)

• Triggered by 20+ hotword variants
• Five intents: LIST, CONFLICTS, ADD, DISMISS, CLEAR
• run_confirmation_loop on all destructive operations (DISMISS single, DISMISS all, CLEAR)
• resume_normal_flow() in finally on all exit paths

Platform Compliance

• No print/asyncio.sleep/asyncio.create_task
• resume_normal_flow() called immediately in watch_loop and in finally in _run
• All daemon speak calls preceded by send_interrupt_signal
• Pydantic constraints respected — all mutable state in local s dict
• __init__.py empty

[github-actions]

✅ Community PR Path Check — Passed

All changed files are inside the community/ folder. Looks good!

[github-actions]

🔀 Branch Merge Check

PR direction: feature/conflict-detector → dev

✅ Passed — feature/conflict-detector → dev is a valid merge direction

[github-actions]

✅ Ability Validation Passed

📋 Validating: community/conflict-detector
  ✅ All checks passed!

[github-actions]

🔍 Lint Results

✅ __init__.py — Empty as expected

Files linted: community/conflict-detector/background.py community/conflict-detector/main.py

✅ Flake8 — Passed

✅ All checks passed!

[uzair401]

Great work on this @hassan1731996 — tested it with several conflicting flows and it captures conflicts clearly and explains them well.

One thing I'd flag for the next background daemon: please keep the polling interval between 60–90 seconds rather than 15s. A tighter loop sounds responsive but in practice it doesn't add much — most real commitments come in over the course of a conversation, not in 15-second bursts — and the LLM calls stack up fast. With MAX_LLM_CALLS_PER_POLL = 3 plus the per-commitment conflict checks, this can fire 5–10+ LLM calls per minute on a normal conversation. A 60–90s cadence gives the daemon room to batch newly-arrived messages, run extraction once over the batch, and only invoke the conflict check when there's a real same-date or adjacent-day candidate — same end result, fraction of the API spend.

Bigger picture: we have several abilities now built around the same passive-capture idea — Conflict Detector, Decision Journal, Social Memory — each running its own daemon, its own poll loop, its own LLM extraction over the same message history. That's three independent LLM extractions on every message. There's real value in consolidating them into a single passive-capture daemon that polls the message history once per cycle, runs one LLM call that classifies and extracts whatever it finds (commitment, decision, person/relationship note, etc.), and then routes the structured output to the right storage bucket. One graceful poll, one inference, multiple abilities served. Happy to sketch that architecture if you want to take it on next.

[hassan1731996]

Thanks @uzair401, really appreciate the detailed feedback!

The 15s polling was something I went back and forth on - I wanted it to feel reactive but you're right, the LLM cost doesn't justify it. Will bump it to 60–90s in the next revision.

The shared passive-capture daemon idea is genuinely interesting. The redundancy is already obvious looking at how Conflict Detector, Decision Journal and Social Memory all do the same extraction pass independently - it's wasteful. Once I wrap up Portfolio Monitor I'm going to take this on - consolidating all three into a single daemon feels like the right next move.

[uzair401]

Thanks @uzair401, really appreciate the detailed feedback!

The 15s polling was something I went back and forth on - I wanted it to feel reactive but you're right, the LLM cost doesn't justify it. Will bump it to 60–90s in the next revision.

The shared passive-capture daemon idea is genuinely interesting. The redundancy is already obvious looking at how Conflict Detector, Decision Journal and Social Memory all do the same extraction pass independently - it's wasteful. A single daemon that does one extraction per cycle and routes structured output to each ability's storage would clean that up a lot. I'd be up for taking a stab at that architecture, would be great to have a rough sketch from you if you're open to sharing it.

Sure @hassan1731996, I’ll share a rough sketch with you soon. Your ideas are quite creative as well, so I’d definitely like your input too so we can build something genuinely interesting and publish it on the marketplace together.