curiositech/hsts-planning-scheduling

SKILL.md: HSTS — Integrated Planning & Scheduling

Wisdom from Muscettola's HSTS framework for building intelligent systems that unify planning (what to do) and scheduling (when/how to do it). Activates when designing constraint-based agents, solving resource allocation problems, building robust execution systems, or reasoning about commitment under uncertainty.

DECISION POINTS

Decision 1: Unified vs. Separated Planning/Scheduling

IF domain has resources with time-varying availability
   AND causal goals interact with resource constraints
   AND execution environment is uncertain
   → Use unified state variable approach
   → Model as evolving constraint network

IF problem is purely causal (no resource contention)
   → Classical planning may suffice

IF problem has fixed task structure with only temporal constraints
   → Classical scheduling may suffice

IF you see "planner feeds into scheduler" architecture
   → QUESTION: Is the seam creating coordination debt?

Decision 2: Commitment Level Selection

FOR each variable binding decision:
  IF decision is reversible AND more information is coming
    → Post ordering constraint only
    → Example: "A before B" instead of "A at 10:00, B at 11:30"
  
  IF decision resolves high-contention bottleneck (measured via simulation)
    → Commit to specific value now
    → Example: Assign spacecraft antenna to high-priority download
  
  IF tempted to fix exact times early in search
    → STOP: Can ordering relation achieve same progress?
    → Preserve temporal flexibility for executor

Decision 3: Search Focus Priority

BEFORE resolving any disjunction:
  1. Run stochastic simulation on current partial solution
  2. Sample 100+ paths without resolving open choices
  3. Count conflicts per resource-time pair
  4. Focus next decision on highest-contention resource

IF search is thrashing (many backtracks, little progress)
  → You're attacking non-bottleneck variables first
  → Re-estimate contention distribution
  → Switch to highest-conflict resource

Decision 4: Domain Modeling Approach

IF domain has objects with internal state evolving over time
  → Model each as state variable with transition graph
  → Example: Satellite antenna {idle, tracking, downloading, slewing}

IF domain has shared resources (consumable/reusable)
  → Model resource levels as state variables
  → Track availability windows, not just capacity numbers

IF expert rules exist ("X and Y can't overlap during Z")
  → Encode as compatibility constraints
  → NOT as search pruning hacks

FAILURE MODES

Anti-Pattern 1: "Nominal Trajectory Planning"

Detection Rule: If schedule output is sequence of exact times/actions with no flexibility Root Cause: Confusing schedule with execution plan; over-committing temporal decisions Symptom: Any runtime deviation requires full replanning Fix: Output behavioral envelopes (constraint networks) not nominal sequences Timeline: Becomes critical when execution uncertainty > 5% of plan duration

Anti-Pattern 2: "Premature Value Commitment"

Detection Rule: If binding variables to specific values before constraints force it Root Cause: Not distinguishing between progress and commitment Symptom: Search space artificially shrunk; solutions exist but aren't found Fix: Post weakest constraint that enables progress; commit only when forced Timeline: Shows up immediately in search statistics (solution quality drops)

Anti-Pattern 3: "Bottleneck Ignorance"

Detection Rule: If resolving disjunctions without measuring which matter most Root Cause: Treating all constraints as equally important Symptom: Search thrashing on easy variables while hard constraints fester Fix: Use stochastic simulation to estimate conflict probability before each decision Timeline: Becomes exponentially worse as problem size grows

Anti-Pattern 4: "Constraint Network Bloat"

Detection Rule: If propagation time grows quadratically with problem size Root Cause: Not respecting module boundaries; monolithic constraint posting Symptom: Solver becomes unusably slow on realistic problems Fix: Decompose into state variable modules; make interaction topology explicit Timeline: Hits performance wall around 50-100 activities

Anti-Pattern 5: "Expert Knowledge as Hardcoded Heuristics"

Detection Rule: If domain knowledge exists as ad-hoc search pruning rules Root Cause: Encoding constraints implicitly instead of declaratively Symptom: Knowledge invisible to debugging; doesn't compose across modules Fix: Express expert rules as formal compatibility constraints with causal semantics Timeline: Maintenance nightmare emerges when domain complexity grows

WORKED EXAMPLES

Example 1: Satellite Mission Planning

Scenario: Mars orbiter with 2 instruments (camera, spectrometer), 1 antenna, 4 GB storage. Plan 8-hour observation session with data downlink.

Novice Approach:

1. Plan observation sequence first
2. Schedule exact start times
3. Add downlink at end Fails when storage constraint discovered late

Expert HSTS Approach:

1. Model state variables: storage_level, antenna_state, instrument_states
2. Run stochastic simulation: storage fills to 95% probability by hour 3
3. Decision: Storage is bottleneck, not antenna time
4. Commit to downlink at hour 3 (value commitment for bottleneck)
5. Leave observation ordering flexible (constraint posting for non-bottleneck)
6. Output: behavioral envelope with storage invariant + flexible observation sequence

Alternative Path Shown: If simulation revealed antenna conflicts instead of storage, would commit to antenna schedule first, leave storage management flexible.

Example 2: Manufacturing Job Shop

Scenario: 5 jobs, 3 machines, delivery deadlines. Machine M2 has 50% uptime, others 90%.

Novice Approach: Schedule assuming all machines available; handle breakdowns reactively.

Expert HSTS Approach:

1. Model machine_availability as time-varying state variable
2. Stochastic simulation: M2 unavailable 40% of time slots
3. Trade-off Analysis:
   • Option A: Route critical jobs around M2 (safer, longer makespan)
   • Option B: Use M2 with backup plans (shorter nominal, replanning risk)
4. Decision: Critical path jobs avoid M2; non-critical use M2 with ordering constraints only
5. Output: schedule robust to M2 failures without pessimistic assumption about availability

What Expert Catches: M2 unreliability creates secondary bottleneck at other machines when jobs reroute. Plans for cascade effects.

QUALITY GATES

Schedule is valid and complete when:

• [ ] All state variables have legal transitions throughout timeline
• [ ] No resource oversubscription in any sampled execution path
• [ ] All goals reachable within temporal flexibility bounds
• [ ] Bottleneck resources identified with >90% confidence via simulation
• [ ] Constraint network admits multiple execution paths (not nominal trajectory)
• [ ] Module interactions explicit and bounded in complexity
• [ ] Search focused on measured conflicts, not assumed priorities
• [ ] Expert domain knowledge encoded as testable compatibility constraints
• [ ] System degrades gracefully under 20% execution deviation
• [ ] Computational cost scales linearly with number of state variables

Automated Validation Signals:

• Stochastic simulation runs complete without contradiction
• Constraint propagation converges in bounded time
• Multiple distinct solutions exist in final envelope
• Module coupling metrics within acceptable bounds

NOT-FOR BOUNDARIES

This skill should NOT be used for:

• Pure optimization problems without temporal/resource constraints → Use mathematical programming instead
• Real-time reactive control → Use control theory instead
• Simple project management with known task durations → Use critical path method instead
• Problems where optimal solution required → Use complete search methods instead
• Domains where expert knowledge unavailable → Use machine learning approaches instead

Delegation Rules:

• For route planning: use path-planning-algorithms instead
• For resource allocation without time: use constraint-satisfaction-methods instead
• For predictive analytics: use time-series-forecasting instead
• For execution monitoring: use plan-execution-monitoring instead

Boundary Signals: If you're spending more effort encoding the domain than solving instances, consider whether classical OR methods might suffice for your use case.