Skip to content

dev-plan.md

Latest commit

 

History

History
75 lines (57 loc) · 6.84 KB

dev-plan.md

File metadata and controls

75 lines (57 loc) · 6.84 KB
Raw

Plumbline — Development Plan (open work)

Living roadmap of what is not yet shipped to main. Each item states the user value we expect from it. Honest status labels: live on main · on feature branch (reviewed, not merged) · not started · contract-tested only (not yet live-run).

Plumbline discipline applies to this file too: nothing here is claimed done until it is on main and exercised. See metrics/SUMMARY-2026-05-30-dna-investigation.md for the evidence philosophy ("tests green ≠ it works").

Last updated: 2026-06-13

In flight — branch true-line-workflow-completion (reviewed, awaiting merge)

These extend /agileteam into a full customer-value-governed pipeline. Each was TDD'd (contract assertions first) and passed an independent code review.

ID Feature Status User value (1–2 sentences)
G1 Council challenge gate (Phase 0.16, concilium --mode=challenge, token-bounded) branch, reviewed ✓ Before a line of code, three adversarial voices stress-test whether you're even building the right thing — so you don't spend a sprint perfecting an idea that should have changed.
G3 Vision-GO gate → autonomous run branch, reviewed ✓ You confirm the Product Vision once, then the team runs autonomously toward it — you get hands-off delivery without losing the single human checkpoint that matters.
G4 Minimum + dynamic team composition branch, reviewed ✓ Every build always has a reviewer, QA, and product-owner; other specialists are added per architecture — you get right-sized teams instead of a fixed cast that's wrong for the job.
G5 Per-increment Code-reviewer → QA → Watcher chain branch, reviewed ✓ Every increment is checked for customer value as it's built, not just at the end — drift is caught early, when it's cheap to fix.
G6 Graded Watcher escalation (re-align first, escalate only if unreachable) live on main · ⚠ superseded by OD-2 (2026-06-03) maximal autonomy — TIGHTENED: bounded autonomy is restricted to pure code-writing; any deviation from the user-confirmed Product Canvas must hard-block and force an interactive human gate. See Decisions below.
G7 CLI iteration counter N/M + per-iteration Kanban branch, reviewed ✓ You always see "iteration 3/5" and the open tasks for this round — so you can gauge how much is left without reading the agents' minds.

Status correction (2026-06-03): this section is stale. G1/G3/G4 and the per-increment chain (G5) are live on main (merged via the v0.11/v0.12 release flow; verified in concilium.md / agileteam.md and guarded by the PR-1 gate-contract tests). The "branch, reviewed" labels above predate that merge.

Decisions (2026-06-03 — anchored via /agileteam directive)

  • OD-1: Wave A (front-door integrity) started 2026-06-03.
  • OD-2 — Bounded-autonomy hardening (supersedes G6): autonomy is restricted to pure code-writing. Any deviation from the user-confirmed Product Canvas MUST hard-block the run and force an interactive human y/n gate — no autonomous re-alignment across a Canvas boundary. ⚠ This reverses the current shipped behavior (agileteam.md "re-align first; escalate only if the Vision goal is unreachable", ~:219-232) and intersects the G3 contract tests (test_gate_contracts.sh G3-C4 "Watcher escalates to the user" stays satisfied; the re-align-first prose itself changes). Decision recorded here; implementation is a separate governance task (Roadmap Wave B / L3) — NOT yet on main.
  • OD-3 — Council backend: default to OpenRouter free-tier auto-discovery, fail-closed if <2 independent model backends are reachable (no single-model masquerade). Roadmap Wave C.
  • OD-4: website (Roadmap Wave D) scheduled after Wave B.

Not started

ID Feature Status User value (1–2 sentences)
G8 Customer-usage acceptance test + council fallback not started QA tests the product the way a real customer would use it; a grave deviation blocks acceptance and convenes the council for a realistic fix — so "done" means actually usable, not just green.
G9 SMART retrospective with legitimate null-result not started Retros only adopt changes that are Specific/Measurable/Achievable/Realistic/Time-bound — and may correctly decide to change nothing — so the framework improves on evidence, not on the urge to "do something".
T9 Docs: spec-v3 + governance + README for the new gates partial (README updated; spec/governance pending) New users can understand and trust the value gates because they're documented, not just coded — lowering the barrier to adoption.

Validation debt (the honest gap)

Item Status User value of closing it
Live /agileteam dry-run of the True-Line layer partially closed 2026-06-13 — Watcher drift-detection now real-boundary-smoke ✓ (evidence: n=8, catch 3/3 · cry-wolf 0/3, Opus 4.8, verified from subagent logs); full 8-phase orchestration wiring still contract-tested only The Watcher now carries runtime evidence that it catches a blatant planted value contradiction and names the violated Canvas non-goal — that judgment is no longer unit-fake by our own standard. Still open (see the evidence doc's "What this did NOT prove"): the orchestrator wiring that invokes the Watcher end-to-end + records the CONTRA + gates later phases; a multi-model result (the run was Opus-4.8-only — the tier already known to work; no Haiku/Sonnet evidence); and code-level/diff verification. The subtle arm paused 2/2 — disclosed as correct-but-over-sensitive conservatism, not a clean pass.
Council cognitive diversity not wired — OD-3 (2026-06-03): default to OpenRouter free-tier auto-discovery, fail-closed if <2 independent backends reachable (Wave C, own design pass) /concilium's value depends on uncorrelated models. Today it runs as a structured single-model critique. OD-3's OpenRouter route (or a user API key for stronger models) is the path to real triangulation — gated fail-closed so it never silently degrades to single-model (the failure it exists to prevent).

Done & live on main (for context)

  • 87 subagents, /agileteam v3, /concilium four-body council, /honest-status, /bench-oracle, /reflect(-skills).
  • Product Canvas gate, Product Vision gate, Plumbline Watcher (Gate E), Reality Ledger, contradiction ledger — all merged via PRs #6/#7.
  • The mutation-oracle benchmark suite + the honest negative result (metrics/).
  • Live Agent Explorer on GitHub Pages.