Project overview
A local AI-work control plane that compiles bounded, task-specific context and coordinates evidence, provider handoffs and human-approved capabilities without duplicating project data.
Problem
AI coding sessions can waste context, confuse plans with current state and lose decisions between providers. Smaller prompts alone do not establish correctness.
Engineering scope
Lazy project discovery, structural indexing, line-range retrieval, bounded context compilation, shared coordination and scoped capability routing.
Role: System design and AI-assisted implementation
Context and constraints
- Source files remain authoritative; the system is not an autonomous agent or a second project database.
- Credentials stay in provider or operating-system stores, outside retrieved context and metrics.
Technology stack
- Python
- MCP
- Context engineering
- Evidence gates
- Capability routing
Contributions
- Designed task-specific context routing with explicit evidence gaps and a hard context budget.
- Integrated project-scoped coordination, evidence gates and capability routing across AI providers.
- Separated context-size estimates from answer quality, provider billing and unsupported adoption claims.
Validation and outcome
- The reviewed implementation records bounded context receipts and rejects credential-shaped material from retrieval.
- An initial calibration pack was rejected after rendered-browser review found missing task-critical evidence. The revised pack was assessed against source rather than accepted for size alone.
A 28 August 2026 local calibration selected 3,600 estimated tokens from an 8,502-token project estimate in 2.199 seconds, reporting a 57.66% context-size reduction.
Evidence scope
- One task-specific local calibration, documented in the project metrics record. Counts are deterministic estimates, not provider billing.
- Implementation and calibration documentation reviewed on 8 September 2026. Private source is not presented as a public repository.
Public scope
Private implementation and confidentiality boundaries limit the detail shared here. This overview focuses on the system's purpose, Bongo's role and the technologies used.
Constraints and limits
- The calibration does not establish improved answer quality, complete evidence coverage or lower API cost.
- Machine-specific timing and one task cannot establish general performance. Independent cross-provider review must be recorded separately from self-review.