Context

Follow-up to #2020 (dynamic tool schema filtering). Community feedback from @Docat0209 highlighted an important gap: some tools become relevant only after other tasks complete.

Problem

Current filtering exposes tools in two states:

• Iteration 0: filtered by relevance (30-60% token savings ✓)
• Iterations 1+: full tool set available

This doesn't account for sequential dependencies. For example:

• Format results tool should only be visible after search succeeds
• Data validation tool should only be visible after data load completes
• Rollback tool should only be visible after operation attempt

Without explicit dependencies, LLM may request unavailable tools, wasting tokens and confusing the agent.

Solution

Implement a tool dependency graph to make tools conditionally available:

1. ToolDependency struct — define prerequisites

   • must_complete_before: tools that must finish first
   • works_better_after: soft suggestions (for observability)

2. Dependency tracking — in Agent state

   • Track completed tools in current conversation turn
   • Query: "which tools are available?" based on completions

3. Enhanced filtering — combine relevance + dependencies

   • Include tool if: (relevance match) OR (always_on) OR (dependency satisfied)

4. InclusionReason expansion

   • Add: DependencyMet (prerequisite completed)
   • Add: DependencyNotMet (for logging why excluded)

Implementation

• Keep dependencies optional (not all tools need them)
• Config: [tool.<name>.dependencies] section
• Testing: multi-turn scenarios with sequential tool chains
• UI: TUI status showing unavailable tools + unsatisfied dependencies

Acceptance Criteria

• Tool dependencies declared via config
• Unmet dependencies logged with clear reasons
• Tools with unsatisfied deps excluded from LLM schema
• No performance regression (dependency check < 10ms per turn)
• Integration tests verify multi-turn chains
• Docs updated (config schema, examples)

Labels

research, tools, enhancement

Related

• research(tools): dynamic tool schema filtering to reduce context waste and improve selection accuracy #2020 (dynamic tool schema filtering)
• Comment: research(tools): dynamic tool schema filtering to reduce context waste and improve selection accuracy #2020 (comment)