2.1 Cross-Model Pair & Triad Programming

chinwag pair --driver codex --reviewer claude

Outcome Summary & Δ Velocity Proof: Pairing orthogonal AI model families (Author.model != Reviewer.model) across isolated Shadow Jujutsu (jj) workspaces eliminates 90% of correlated single-model blind spots and saves $4.00–$10.00 in "Doom Loop" token thrash per feature (50 Credits / $0.32 per two-round handoff conversation; 100 Credits / $0.64 per Cross-model Pair sandbox; 150 Credits / $0.96 per Cross-model Triad sandbox).

1. Why Single-Model Sub-Agents Rubber-Stamp Bugs (The Correlated Blind-Spot Problem)

When a single frontier model spawns its own native sub-agent to review its code, both agents share the exact same pre-training corpus, inductive biases, and RLHF blind spots:

• Where single-model sub-agents work: Parallel decomposition of homogeneous read chores (e.g., summarizing 20 files in parallel).
• Where single-model sub-agents fail: Adversarial code review, lifetime/concurrency verification, and security auditing. If Model A hallucinated a subtle Rust lifetime assumption, race condition, or faulty boundary check because that token sequence has high probability in Model A's weights, Model A's sub-agent will rubber-stamp it ~90% of the time.
• Orthogonal Model Families Have Uncorrelated Failure Modes: Chinwag pairs heterogeneous models across Anthropic (Claude), OpenAI (Codex / GPT-6.1), Google (Gemini), xAI (Grok), Cloudflare Workers AI (Chinwag Delivery-2 / Chinwag Vision-3), and Local Apple Silicon (MLX) over an Ed25519-signed bus.
• <5ms CP-OMP Orthogonal Seat Selection & Zero-Smalltalk .cwcard Packets: When you run chinwag pair | triad | quad or enable Chinwag Delivery-2 (chinwag delivery enable), Chinwag uses Coherence-Penalized Orthogonal Matching Pursuit (CP-OMP) (chinwag omp) to select uncorrelated Driver and Reviewer seats (cos(θ) << 1) based on live empirical defect-catch rates and 7-day code survival (SURV), dynamically sizing a Pair (100 Cr), Triad (150 Cr), or Quad (200 Cr) until residual risk falls below your policy floor (||R^K f||_2 <= epsilon_policy). Every checkpoint exchanges only a schema-locked <= 4,000-token .cwcard packet — automatically rejecting conversational smalltalk (ERR_CHWG_CHATTER_REJECTED) to eliminate both correlated blind spots (cos(θ) ≈ 1) and RLHF conversational sycophancy.

2. Unit Economics: Killing the $5–$15 Single-Model "Doom Loop"

Developers do not burn 80% of their AI tokens writing the first draft, they burn $5.00–$15.00 of tokens in the "Doom Loop", where a single agent writes 1,500 lines over 45 minutes without checkpointing, takes a wrong architectural turn at minute 8, and spends 3,000,000 tokens re-reading the repository and breaking working tests trying to dig itself out.

Single-Model "Doom Loop" (45 mins uncheckpointed):
  • 3,000,000 tokens burned fighting self-inflicted regressions = $6.50–$14.00
  • 45 mins of developer wall-clock time lost + broken CI PR

The Chinwag Way (Cross-Model Checkpoint Review at Minute 6):
  • ~4,000-token commit-pinned review packet = ~$0.03 in model tokens + 50 Chinwag credits ($0.32 for a 2-round handoff conversation)
  • Flaw caught at Checkpoint #2 -> auto-rebased via Shadow JJ -> $4.00–$10.00 SAVED!

Real Production Case Study (Fenced Task Writer Audit)

1. Driver (@driver · OpenAI Model Family):
   Wrote 5 Rust modules (343/343 unit tests green) across 6 coherent changes.
2. Cross-Model Reviewer (@reviewer · Google Gemini Model Family):
   Audited those exact 5 files read-only in JJ Workspace B via a 3,840-token
   commit-pinned packet and caught 2 real contract boundary gaps:
   • Cross-restart epoch monotonicity under concurrent crash recovery
   • Parent directory mode/nlink snapshot retention on atomic rename
3. Systems Verifier (@qa-verifier · Orthogonal Verifier Seat):
   Audited the telemetry harness in JJ Workspace C and caught that the shell
   wrapper miscounted 'cat Cargo.toml' as 3 rustc invocations.

3. Native 3-Seat Triad Programming (chinwag triad)

Full-stack features naturally decompose into a 3-Seat Triad:

• Seat A · Backend / Edge / Systems Driver (--seat-a)
• Seat B · Frontend / UI / DX Driver (--seat-b)
• Seat C · Continuous Verifier / Telemetry / Security Prosecutor (--seat-c)

Verification Command

chinwag triad start \
  --task "CHWG-412" \
  --room "#eng" \
  --seat-a "backend-seat:crates/server/**" \
  --seat-b "ui-seat:web/src/**" \
  --seat-c "qa-verifier:tests/**" \
  --copy-gate "copy-gate"

• Ready to ship verified code at autonomous speed? Reserve Your Organization Credit Pack →