Context
Thread 6 (horizon item) of the code-review-suite forward programme (effectiveness, agreed 2026-06-17).
Observation
The pipeline spends agent tokens on operations that are structurally identical every time — diff parsing, changed-lines computation, file-type routing, significant-deletion detection, trivial-mode evaluation. Each specialist and the orchestrator often re-derives the same facts.
Idea
Extract these into a purpose-built CLI tool (compiled, cross-platform) that runs once, deterministically, at zero token cost. The pipeline calls the tool and consumes structured output.
Candidate operations:
- Diff parsing →
$CHANGED_LINES map, $CHANGED_FILES, line/file counts
- File-type routing flags (
$CSHARP_DETECTED, $JS_DETECTED, etc.)
- Security-sensitive area scanning
- Significant-deletions detection (the
-w contiguous-10-lines check)
- Trivial-mode bar evaluation
- A/B scoring/statistics (currently
ab_stats.py + specialist_score.sh)
Tension
Today's ad-hoc approach (agent writes Python) is maximally flexible. The CLI must capture that flexibility — configurable output, handles the real range of diff shapes. Implementation vehicle (Rust vs Go vs Python) is a later decision; the value is "don't spend tokens on deterministic computation."
Why last
At current review volume the token cost of diff-parsing is small relative to specialist reasoning. Value is real at scale; premature now. Also: building before the pipeline's final shape settles (post threads 1–2) risks the wrong abstractions.
Status
Horizon item. Revisit when review volume or pipeline shape settles. Not started.
Context
Thread 6 (horizon item) of the code-review-suite forward programme (effectiveness, agreed 2026-06-17).
Observation
The pipeline spends agent tokens on operations that are structurally identical every time — diff parsing, changed-lines computation, file-type routing, significant-deletion detection, trivial-mode evaluation. Each specialist and the orchestrator often re-derives the same facts.
Idea
Extract these into a purpose-built CLI tool (compiled, cross-platform) that runs once, deterministically, at zero token cost. The pipeline calls the tool and consumes structured output.
Candidate operations:
$CHANGED_LINESmap,$CHANGED_FILES, line/file counts$CSHARP_DETECTED,$JS_DETECTED, etc.)-wcontiguous-10-lines check)ab_stats.py+specialist_score.sh)Tension
Today's ad-hoc approach (agent writes Python) is maximally flexible. The CLI must capture that flexibility — configurable output, handles the real range of diff shapes. Implementation vehicle (Rust vs Go vs Python) is a later decision; the value is "don't spend tokens on deterministic computation."
Why last
At current review volume the token cost of diff-parsing is small relative to specialist reasoning. Value is real at scale; premature now. Also: building before the pipeline's final shape settles (post threads 1–2) risks the wrong abstractions.
Status
Horizon item. Revisit when review volume or pipeline shape settles. Not started.