rsi/.pm/board.yaml
2026-09-17 23:12:52 -07:00

68 lines
2.8 KiB
YAML

requirements:
- id: R1
title: Jev research summary
body: |
Summarize what Jev (TypeSafe AI, launched 2026-09-15) does and how it
works at an undergraduate level, with examples. Claims must be tagged
confirmed / self-reported / speculative, with sources.
last_changed: "2026-09-18 06:10"
changed_by: agent
- id: R2
title: Jev use cases inside the pi harness
body: |
Identify concrete use cases for Jev within the pi (Oh My Pi) harness,
each grounded in a real extension surface (hooks, custom tools, skills,
context files, memory pipeline, model routing) with honest limits and
explicitly non-fitting cases.
last_changed: "2026-09-18 06:10"
changed_by: agent
- id: R3
title: Dreaming method proposal
body: |
Propose a method using Jev to "dream": a batch job that replays past
interactions, mines improvement opportunities, and adds useful skills,
AGENTS.md/RULES.md guidance, system-prompt updates, and custom tools to
the pi configuration via human-approved PRs and pm-board proposals.
last_changed: "2026-09-18 06:10"
changed_by: agent
- id: R4
title: Hosted and reviewable
body: |
Repo on Gitea (goodcode/rsi), work lands via pull request (never direct
to main), and the project is enrolled in the deployed pm-board.
last_changed: "2026-09-18 06:10"
changed_by: agent
choices:
- id: C1
title: Board-only project, docs in-repo
body: |
No static site is served; the deliverables are research documents in the
repo. A site would add serving surface without adding information.
last_changed: "2026-09-18 06:10"
changed_by: agent
- id: C2
title: Jev judges, the LLM writes
body: |
Jev is used only for bounded-output decisions (Noul/Score/Choice
triage, routing, gating). All synthesis — drafting skills, diffs,
tool code — stays with the harness LLM, because Jev structurally
cannot generate strings.
last_changed: "2026-09-18 06:10"
changed_by: agent
- id: C3
title: Dream outputs are proposals, never direct edits
body: |
The dreaming pipeline writes PRs and pm-board proposals with evidence
links; the user merges. Accept/reject outcomes are the ground-truth
labels for a calibration loop that tunes Jev thresholds over time.
last_changed: "2026-09-18 06:10"
changed_by: agent
- id: C4
title: Swappable Jev backend behind one interface
body: |
JevClient with three backends: real Jev (needs TYPESAFE_API_KEY early
access), an LLM constrained to the same typed schema (system-one
adapter pattern) for building/auditing now, and a deterministic mock
for tests. Pin jev-1.13.0 when live; log the responding model id.
last_changed: "2026-09-18 06:10"
changed_by: agent