Skip to content

CLAUDE.md

Latest commit

 

History

History
89 lines (60 loc) · 4.2 KB

CLAUDE.md

File metadata and controls

89 lines (60 loc) · 4.2 KB
Raw

idd-route — CLAUDE.md

Purpose

Data-driven agent routing for IDD methodology. Wraps the PsychQuant/idd-route-swift Swift binary, distributed via plugin auto-download (bin/idd-route-wrapper.sh).

Recommends the best agent (Codex GPT-5.5 xhigh / Claude Opus 4.7 / Sonnet 4.6 / Haiku 4.5) for each new IDD issue based on observed track record per (complexity, scope_class) bucket with exponential decay. Falls back to a static heuristic rubric on cold start (< 5 data points in bucket).

Skills

Skill Purpose
/idd-route:recommend Ad-hoc lookup: "given this issue's features, which agent should implement?"
/idd-route:stats Human-readable markdown summary of routing-stats.jsonl
/idd-route:backfill One-shot seed from past GH verify comments (cold start救星)。Requires binary v0.3.0 (P2 of plan)。

Hooks

(none — idd-route doesn't fire on tool events; the issue-driven-dev plugin's idd-verify / idd-close skills call the binary directly via ~/bin/idd-route)

Binary

Source: PsychQuant/idd-route-swift — universal arm64+x86_64 macOS Swift binary.

Distribution: bin/idd-route-wrapper.sh reads binaries.idd-route.version from plugin.json, compares against ~/bin/.idd-route.version sidecar, atomic-downloads from GH release if stale. Mirrors che-word-mcp-wrapper.sh pattern. Uses gh api if available (auth'd, high rate limit), falls back to plain curl (60/hr unauth limit).

Current pinned version: v0.2.0 (record / recommend / stats / summarize commands; v0.3.0 will add update-outcome + backfill).

Stats data flow

idd-verify (in issue-driven-dev) ─→ ~/bin/idd-route record ... --outcome in_review
                                            └─→ <repo>/.claude/.idd/routing-stats.jsonl
                                            └─→ ~/.cache/idd-route/stats.jsonl (mirror)

idd-close  (in issue-driven-dev) ─→ ~/bin/idd-route update-outcome ... --outcome merged|abandoned
                                            └─→ append a new record (jsonl is append-only)

idd-diagnose (in issue-driven-dev) ─→ ~/bin/idd-route recommend ...
                                            └─→ stdout JSON {recommended, confidence, ...}

The issue-driven-dev skills invoke the binary at ~/bin/idd-route directly (NOT via ${CLAUDE_PLUGIN_ROOT}/../idd-route/bin/...) — decouples plugin path layout from skill code; idd-route binary can also be installed via cli-tools:cli-install for non-IDD users.

Three-tier config injection

1. <repo>/.claude/idd-route.json     ← per-project (highest priority)
2. ~/.cache/idd-route/config.json    ← per-machine
3. Built-in defaults                 ← in PsychQuant/idd-route-swift

Plus kill-switch:

touch ~/.cache/idd-route/disabled    # `idd-verify` skips record call entirely

Schema (all optional):

{
  "enabled": true,
  "global_mirror": true,
  "decay_half_life_days": 30,
  "min_data_points": 5,
  "candidates": ["codex-gpt-5.5-xhigh", "claude-opus-4.7"]
}

Format spec

routing-stats.jsonl schema documented in PsychQuant/idd-route-swift README.md. Append-only; forward-compatible; reader tolerates new optional fields.

References

Why a separate plugin

Could be inlined into issue-driven-dev, but as a separate plugin:

  • Binary lifecycle is independent (idd-route-swift releases on its own cadence)
  • Other tools / CI / standalone scripts can depend on idd-route without bringing in IDD core
  • Skills are discoverable via /idd-route:* namespace
  • User can disable just routing (uninstall idd-route) while keeping IDD core