5.1 Autonomous AI Program Manager

chinwag delivery status --repo .

Most engineering organizations do not fail because their developers or AI coding agents write slow code. They fail in the coordination layer between tickets: oversized pull requests stall in review, overlapping branches trigger 3-file merge cascades, and a single flaky CI runner burns 40 minutes of human context-switching.

Chinwag Delivery-2 is Chinwag's autonomous AI Program Manager. It watches your <org>/chinwag-tasks repository, your active Shadow jj agent workspaces, and your CI pipelines in real time · splitting oversized specs before coding begins, routing around dependency collisions, and dispatching instant fix plans when builds fail.

Organizations enable turnkey Chinwag Delivery-2 in one command via chinwag delivery enable (or declaratively by setting delivery.enabled: true in chinwag.workflow.yaml), with every verified coordination action metered transparently from the shared Organization Credit Pool (OrgWallet):

# Enable turnkey Chinwag Delivery-2 autonomous orchestration & empirical seat routing
chinwag delivery enable --repo .

# Inspect active Chinwag Delivery-2 status, queue health, and CP-OMP seat dictionary
chinwag delivery status --repo .

Dual-Brain Architecture: Chinwag Reflex-1 + Chinwag Delivery-2

To operate at sub-50ms edge speed without burning expensive reasoning tokens on routine heartbeats, Chinwag Delivery-2 uses a purpose-built Dual-Brain Architecture:

+-----------------------------------------------------------------------------------+
| BRAIN 1: Chinwag Reflex-1 (<50ms Edge Risk, Cost & Stall Radar)                   |
|   • Monitors task heartbeats, file-scope leases, token burn velocity, and CI logs |
|   • Passive heartbeats & stall checks cost 0 Cr (100% Free!)                      |
|   • 1-way encrypted bus message: 15 Cr                                            |
+-----------------------------------------+-----------------------------------------+
                                          |
                                          v (Escalates only when structural planning is needed)
+-----------------------------------------------------------------------------------+
| BRAIN 2: Chinwag Delivery-2 (System-2 Program Synthesis & Auto-Splitting Engine)  |
|   • Splits oversized tickets (p90 > 150 LOC) into disjoint atomic sub-tasks       |
|   • Computes .codegraph blast radius before allowing cross-crate edits            |
|   • Two-round handoff or broken-CI wakeup & fix cycle (20 physical msgs): 50 Cr   |
+-----------------------------------------------------------------------------------+

Comparison: Chinwag Reflex-1 vs. Chinwag Delivery-2

Five Core Responsibilities of Chinwag Delivery-2

1. Pre-Flight Ticket Sizing & Auto-Splitting: Before an agent claims a ticket, Chinwag Delivery-2 estimates p90 lines-of-code impact against .codegraph. If predicted p90 > 150 LOC, Chinwag Delivery-2 decomposes the ticket into 2..5 disjoint sub-tasks (--file-scope) so every PR merges cleanly in under 15 minutes.
2. Pre-Commit Blast-Radius Prevention: Cross-checks proposed file leases against downstream call graphs (crates/*, packages/*) and blocks silent breaking changes before code is written.
3. Two-Loop Empirical Seat Selection & Orthogonal Quorum Sizing (chinwag omp / chinwag-ci omp-gate):
   • Loop 1 (Continuous Defect-Catch & SURV Dictionary Learning): After every .cwcard review handoff or CI regression, Chinwag Delivery-2 updates the repository's empirical defect-catch dictionary (OmpDictionaryStateV1 in .chinwag/state.crdb) and 7-day post-merge commit survival (SURV) across each subsystem (Rust, TypeScript, SQL, Auth/Crypto, UI/CSS). As frontier models rotate from quarter to quarter, Chinwag Delivery-2 automatically shifts Driver and Reviewer assignments to the strongest model family per domain with zero workflow migration.
   • Loop 2 (<5ms Coherence-Penalized Orthogonal Matching Pursuit CP-OMP & Gram-Schmidt Quorum Sizing): At dispatch and review time, chinwag projects the task's risk vector across uncorrelated model families (Author.models != Reviewer.models), dynamically sizing Pair (100 Cr sandbox / 50 Cr 2-round handoff), Triad (150 Cr), or Quad (200 Cr) quorums in isolated <= 150 LOC Shadow jj workspaces until residual unverified risk satisfies the repository policy floor (||R^K f||_2 <= epsilon_policy).
   • Zero-Smalltalk .cwcard Protocol Filter (ERR_CHWG_CHATTER_REJECTED): Enforces <= 4,000-token schema-locked .cwcard review packets and rejects free-form conversational smalltalk so Reviewer models are never sycophantically primed.
4. Cleanroom CI Wakeup & Direct-to-Agent Failure Triage: Turns every failed CI run into an encrypted root-cause wakeup & verified fix cycle (50 Credits per failed CI run) and dispatches a structured failure triage plan directly to the responsible agent (checking Negative Memory Tombstones to prevent repeated dead ends).
5. Zero-Stall Board Synchronization & Cross-Surface Completeness (chinwag board watch | sync-loop & chinwag-ci surface-gate): Watches GitHub Projects v2 and Linear card transitions in real time (Included at 0 extra credits on every handoff/CI transition), computes a deterministic SurfaceImpactManifestV1 (code_cli, readme, docs_site, marketing_site, telemetry_proxy, ui_qa), fans out parallel sub-issues (#ID.1..#ID.4), and enforces Pillar 12 (chinwag-ci surface-gate) + desktop/mobile chinwag-ci ui-gate (1440×900 & 390×844) so code, README.md, documentation, website, telemetry, and UI geometry never drift out of sync before auto-merging to origin/main.

Next Steps

• Next: <org>/chinwag-tasks & The Autonomous Loop
• Ready to put an autonomous AI Program Manager on your backlog? Reserve Your Organization Credit Pack →