Skip to main content
Introduction · v9.0

Implementation Roadmap

The order is deliberate: everything depends on fresh, structured data. Build that first. Everything else is additive.

Build Order​

01
Atomic commit-index (post-commit hook + SQLite)
Everything else depends on fresh, structured data
02
tree-sitter parsing + entity normalizer
Populates modules, dependencies, edges tables
03
Multi-label router + index queries
Enables basic dev plan queries against the index
04
LSP integration (tsserver for TypeScript)
Adds symbol precision for the primary language
05
Graph CTE with hub handling
Enables blast radius analysis
06
ECO construction with 5 dimensions
Produces the contract that drives the pipeline
07
Lane classification + strategy selection
Pipeline can now consume the ECO
08
Task decomposition + TEE generation
Implementer agents have scoped boundaries
09
Decision trace + replay engine
Debugging and calibration infrastructure
10
Ground truth sampling + calibration file schema
Measurement infrastructure for continuous improvement

20-Query Benchmark​

Before team rollout, run 20 representative queries against the target codebase to validate:

MetricWhat it proves
ECO dimension accuracyKnowledge layer correctly characterizes codebase state
Lane classification correctnessForced constraints trigger when they should
Strategy selection appropriatenessIntent-biased defaults make sense
Evidence checkpoint reliabilityBinary pass/fail correlates with real evidence

Key threshold: If Lane C rate exceeds 30% of non-trivial tasks, loosen escalation thresholds.

Measure: % blocked · % escalated to Lane C · % reclassified · % eventually judged correct

Calibration Timeline​

Week 1–4:   System runs with default weights. Traces accumulate.
Week 4:     First calibration review.
            ~50–100 ground truth samples.
            Adjust penalty weights, routing keywords, severity thresholds.
Monthly:    dev replay --all --since 30d
            Verify improvements, detect regressions.
Ongoing:    Ground truth sampling at 5%.
            Calibration file grows.

Pending Items (v9 Delta)​

The v9 specification is complete. Implementation is pending for all v7–v9 additions:

  • Build the v9 implementation (all specs defined, no code for v7–v9 additions)
  • Run the 20-query benchmark against the target codebase
  • Implement multi-label router with bitmask classification
  • Implement severity-weighted cluster accumulation
  • Implement replay engine (dev replay command)
  • Set up ground truth sampling (5% capture + calibration file schema)
  • Create ast-grep rules scoped to XState for state_transition edge type
  • Define entity normalization rules between tree-sitter / LSP / ctags output formats
  • First monthly calibration cycle (after 4 weeks of trace data)