Milestones
List view
Charter: one correct, isolated major. Consolidates the 2026-08-08 convergence audit (4 workflows, 42 agents, every claim through an adversarial refuter). Contents: 29 verified findings (4 critical), 9 redundancy retirements (re-home BEFORE delete), 14 HIGH-value unused CC capabilities adopted, 2 unregistered CC hook events wired. Governing rule: CC mechanism is the SUBSTRATE, our need is a LAYER on it via CC extension points. EXTEND is default. RETIRE = re-home the opinion first, then delete the pipe. OVERRIDE and DIVERGE are banned. KEEP only where CC has no mechanism. Definition of done: every fix carries a test that FAILED before it and PASSES after, with the failing output recorded in the PR. "Tests pass" is not evidence — this audit found 103 unit tests passing on a helper that was dead in production because the fixtures mirrored the bug. Train: 10.0.0-alpha.N -> 10.0.0-beta.N -> 10.0.0, cut from release/v10.
No due date•95/103 issues closedFix the auto-triage pipeline signal quality exposed by the 2.1.209-214 wave: filer truncation (#2950), an unfiled-high-signal lane, and grep-gated flagger precision. This wave: 46/49 filed issues were keyword false positives while the one real adoption (sessionstart fork source) was never filed.
No due date•3/4 issues closed**Source:** Assessment of Forward Future AI's [Loop Library](https://signals.forwardfuture.ai/loop-library/) (31 autonomous loop recipes) vs OrchestKit. Interactive playground: `docs/loop-library-assessment/index.html`. ## Thesis — orthogonal, not competing The Loop Library is the **loop envelope** (run-until-condition: budgets, streaks, holdouts, heartbeats). OrchestKit is the **work each pass does** (parallel sub-agents, rubric gates, 211 safety hooks, memory, 111 skills). Plan: **consume their recipes, keep our machinery** — don't rebuild either. ## Verdict Composite **7.3 / B**. We lead on loop **engine** (primitives), **safety** for unattended runs, and stopping-condition **rigor**. Two genuine gaps: **holdout-promotion** and a **consecutive-pass streak gate**. Theme coverage across the Loop Library: **7 Direct · 8 Partial · 2 Gap** (of 17 collapsed themes). ## Scope (4 issues) - **Adopt now:** Loop Recipe Book; Streak gate - **Later:** Holdout-promotion gate; Cross-model adversarial reviewer ## Deferred (not scheduled) - **Heartbeat maintainer** (Loop-Library "Five-Minute Maintainer" — wake every 5 min, one bounded task): conflicts with the no-paid-background-LLM principle (bg LLM calls bill per-session-per-dev). Opt-in only if ever built; intentionally not an issue.
No due date•14/16 issues closedFull adoption of the Fable 5 loop-design patterns (Lance Martin, Anthropic, 2026-06-09): self-correction loops with rubric-as-environment-feedback, verifier sub-agents with stop-gating authority, and closing the memory progression fail-investigate-VERIFY-distill-consult. Gap analysis 2026-06-10: composite 5.9/10 (loops 6.2, verifiers 7.0, memory 4.4). Sequencing: Approach A (rubric contract + stop-gating graders) then B (memory verify loop); C (loop primitive) deferred pending A. Cross-refs: milestone 97 (Memory Hardening), 156 (Eval v2), PR #2339 (model vocabulary).
No due date•6/9 issues closedSupersedes #86 (foundation shipped). Forward gaps: walk-all auto-eval to close the 21-skill + 7-agent coverage gap, a CC-idiom conformance grader ("up to latest-CC standard"), an aggregate quality index wiring check-eval-regression.sh as a gate, and connecting the DeepEval/RAGAS metrics hook. Goal: auto-evaluate every skill and subagent against the latest CC.
No due date•12/14 issues closedReplace monitors.json with 4-layer SQLite-backed coordination architecture. BREAKING; no backwards compat. v7.94.0 target. Tracks #1908.
No due date•21/22 issues closed- No due date•6/8 issues closed
Unified per-event hook contract package (npm + PyPI) with codegen-generated discriminated union types, JSON Schema publish, OpenAPI 3.1 for the HTTP sink, HMAC signing protocol RFC, and cross-repo parity gate. Closes the 'platform Python schema drifts from OrchestKit TS schema' class of bugs surfaced during the #1794 → #1798 work. Driving constraints: - Single source of truth: one hook event registry per event, derived types in both languages - Drift detection: parity gate test in CI catches field-by-field mismatch between npm + PyPI packages - Generic-client compatible: any HTTP sink implementation can consume the contract (not just yonatan-hq platform) Non-goals: M141 ships the contract + parity gate. Migrating all OrchestKit hooks to the typed events is M142.
No due date•18/19 issues closedAdopt new capabilities from emulate 0.5.0, portless 0.12.0, agent-browser 0.26.0, @json-render/core 0.18.0, plus the not-yet-tracked vercel-labs/wterm. 8 ideas ranked by impact/effort. Top 3 quick wins (priority >= 2.5) ship in ~1 day; cross-package idea (#6) is bigger lift. See PR #1558 companion playground + NotebookLM infographic for the full ranking.
No due date•19/22 issues closedBreaking architecture release: lib/ reorganization (54 flat → 7 modules), shared sync dispatcher, .claude/ path consolidation (18 → 5 namespaces), hook-registry.json generation. Zero functional changes — same hooks, same events, same payloads. yonatan-hq HTTP contract untouched.
No due date•7/10 issues closedCherry-picked improvements from CC internals analysis. Packaging security, cache optimization, token budget, MCP output transform.
No due date•38/43 issues closedFix 17 bugs found in memory audit, auto-persist decisions, add health dashboard, transcript extraction. Driven by competitive analysis vs supermemory (March 2026 audit).
No due date•8/13 issues closedDocumentation site overhaul. Skill search UX, interactive examples, narrative content refresh.
No due date•21/34 issues closedDeferred — content production. Gallery UX, 6 demo videos, hero montage, CI automation, TerminalFlow fixes and variants.
Overdue by 1 month(s)•Due by June 30, 2026•2/9 issues closed