Agent Skills & Protocols
Docs » Agent Skills & Protocols » TPM orchestration plane (Gantt matrix + branch collaboration)
Raw .md

TPM orchestration plane (Gantt matrix + branch collaboration)

Canonical operating geometry for program managers and primary orchestrators. When this file disagrees with live POL-* / CLI validation, trust the kernel.

Matrix geometry

ROADMAP (product / system frame — title over the Gantt)
└── workstream_refs + timeline → chart viewport
    │
    │  columns ≈ priority_plans (feature / functional bundles)
    │  rows    ≈ workstreams (lane / flavor of work — long-lived)
    │
    ├── WS-backend ───── PRI-Sprint-A (BE) ─── PRI-Sprint-B …
    ├── WS-frontend ──── PRI-Sprint-A (FE) ─── …
    └── WS-docs-design ─ PRI-Sprint-A (docs/arch) …
Plane Role Notes
Roadmap Frame around the Gantt — product/system summary Spec: workstream_refs + timeline_*; hosts the collective embodiment of plans, WS, BLIs, goals, reqs, criteria, tests, ATKs
Workstream Horizontal lane — like/complementary activities Reporting lanes; soft lifetime (can stay active indefinitely); e.g. backend+infra, frontend, docs-n-design
Priority plan Column — feature/functional bundle May span multiple workstreams via workstream_refs[]; does not have to
Milestone Time anchors on the Gantt Linked from workstreams / plans / BLIs
BLI → ATK / swarm Cells under a locked plan Execute the frozen set; do not expand mid-flight

Plans reference workstreams whose flavor matches the work (frontend BLIs → frontend WS). Not mandatory for every object, but the intended design for TPM reporting and conflict separation.

3-plan runway (always)

Policy: POL-AGENT-THREE-PLAN-RUNWAY-001. Protocol: WFL-AGENT-THREE-PLAN-RUNWAY-001.

Keep three live priority_plan columns with real BLIs:

  1. One execution-facing lead (active / in_progress) with open children or explicit ship-exit. Hand this column to a primary; hourglass-steer; do not orch it from chat.
  2. ≥2 grooming columns that TPM is shaping for the next lock. Workers do not poach these.

If the count drops below three, TPM work is objectify from undelivered REQ/CRIT onto an unlocked column — not persist align-latest.json, not remint ORCHESTRATE_PLAN, not stuffing exploring BLIs onto the sealed lead (POL-AGENT-TPM-GROOM-AHEAD-001).

Scope lock (definitive start/stop)

Validation refuses draft-plane scope creep onto execution-facing plans (active / in_progress) — e.g. linking exploring BLIs (Scope Creep Protection / plan membership gates).

TPM duty before lock: fill the column (shovel-ready BLIs, milestones, workstream refs).
After lock: primaries execute the sealed set to completion. Missing mid-sprint work → new plan column or strategic replan — never stuff the locked plan.

Process administration only (2026-08-10)

Process administration only (2026-08-10)

Policy: POL-AGENT-TPM-PROCESS-ADMIN-001.

TPM altitude is process admin, not whats-next execution:

  1. zqk system align (+ persist .zqk/state/ambient/align-latest.json) so priorities stay tied to goals/mission/vision/strategic_plan.
  2. Maintain priority_plan.active_order (lower wins on whats-next).
  3. Groom shovel-ready columns; hourglass-steer peers to self-serve whats-next --agent-id <seat>.
  4. Do not agent orchestrate peer-owned PRIs or absorb work via local Ollama sync-loops.

Peers execute the system-proposed plan and scale subagents (WFL-SUBAGENT-DISPATCH).

Plan ownership (not ATK babysitting)

  1. Strategic ambiguity / conflicting priorities → strategic planning meeting → strategic_plan / roadmap rewrite.
  2. TPM decomposes to shovel-ready priority_plans (+ workstream assignment).
  3. Hand each plan to a primary → zqk agent orchestrate <PRI> / subagent swarm to plan completion.
  4. Parallel plans only when workstreams keep lanes non-colliding.
  5. Hourglass (POL-AGENT-ORCH-HOURGLASS-001, WFL-TPM-AGY-MESH-001) = handoff insurance — not a license to micro-mint ATKs as a conveyor.

Primary altitude: .zqk/skills/zqk-orchestration/SKILL.md — assigned plan + workstream lanes (including dedicated tech-debt vs project lanes). Primaries sequence within the sealed plan; they do not rewrite the roadmap Gantt.

Branch collaboration (one plan → one integration branch)

Policy: POL-CODE-BRANCH-001 (Strict TPM Branch Collaboration). Mega-branch pattern: onboarding § branching mandate.

  1. TPM provisions one collaboration / feature branch per plan (pre-PR trunk; e.g. integration/pri-…). Agents must not invent ad-hoc integration branches.
  2. Workers operate in isolated git worktrees / agent branches off that plan branch — never race main or steal peers' tips.
  3. Merge up into the plan branch when pedantic/build/policy gates pass.
  4. TPM opens the single PR to main — not one PR per agent.
  5. Ship-exit (mandatory per column): local pre-commit checks (zqk system check, make verify) → targeted TDD (go test -timeout ≤60s probes) → commit (agent identity) → push → PR → automerge when criteria met (local validation green; optional [ZQK_AGENT_STAMP] / .github/workflows/auto-merge.yml; else gh pr merge --squash after gates). Objectify claims when process objects ship.
  6. No idle: hourglass on directed steers; while waiting, advance the next health/capability action. Peers self-serve whats-next; TPM process-admin only (POL-AGENT-TPM-PROCESS-ADMIN-001).

Value signal: swarm builds significant momentum without random blockers from poor process orchestration (branch fights, open-ended plan scope, TPM ATK babysitting).

Kernel / docs pointers

  • Specs: .zqk/specs/objects/{roadmap,workstream,priority_plan,backlog_item}.yaml
  • Membership / immutability: pkg/validation/priority_plan_membership.go, pkg/kernelcas/compose (Scope Creep Protection)
  • Mesh / hourglass: .cursor/rules/tpm-agy-mesh-wake.mdc, WFL-TPM-AGY-MESH-001
  • Strategic meetings: docs/architecture/STRATEGIC_PLANNING_MEETING_WIRING.md
  • Onboarding mega-branch: docs/onboarding/AI_AGENT_ONBOARDING.md (BRANCHING MANDATE)