K0 delivery test¶
How K0 (work contract, refine-idea skill, standard roles) reaches a fresh session,
per docs/specs/org-v1/K0-delivery.md. Test task, given verbatim and with no other hint:
Właściciel mówi: chcę, żeby Bridge pokazywał mi ile kosztuje każdy projekt. Doprecyzuj to.
Round 1 — 2026-10-07, notes v1 (source c826df9)¶
Setup:
- HeyAira: detached checkout of h8v6/heyaira at ef36ca4. AGENTS.md has the startup instruction.
- Voyana: project root AGENTS.md version 4.5 (h8v6/voyana e591ce1, PR #427).
- Codex CLI ran codex exec with approvals bypassed, so that HeyAira MCP calls could run without a person present. In Voyana it ran on a copy of the root AGENTS.md.
- Claude Code ran claude -p with Read/Glob/Grep and the project's HeyAira MCP tools allowed.
| Step (K0-delivery) | Codex · HeyAira | Codex · Voyana | Claude · HeyAira | Claude · Voyana |
|---|---|---|---|---|
| 1. Finds and reads the contract and skill notes itself | PASS (3× memory_search, 3× memory_get) | PASS (4× memory_search, 5× memory_get) | PARTIAL (skill note only; read AGENTS.md) | FAIL (no tool calls at all) |
2. Does not implement; records the idea as a pomysl task |
PARTIAL (no implementation; pointed at existing idea e7a3168c, no new task) | FAIL (no task) | FAIL (no task) | FAIL (no task) |
| 3. ≤ 5 questions with recommendations, or full template | PASS (1 question + rec.) | PASS (3 questions + recs.) | PASS (1 decision + rec.) | PASS (1 question + default) |
| 4. Result has every template section; end criteria are commands/actions | n/a (stopped at questions) | n/a | n/a | n/a |
| 5. Cites the right version (v1) | FAIL | FAIL | FAIL | FAIL |
Verdict: not passed. No client recorded the idea as a task, and none named the version it worked from. Claude Code in Voyana never looked at HeyAira.
Causes and fixes:
1. Claude Code in Voyana: a fresh claude -p (Claude Code CLI) started in the
Voyana root reports that neither AGENTS.md nor CLAUDE.md is in its context,
and its HeyAira MCP servers are still "pending" at start. The Voyana root has
only AGENTS.md. Voyana's owner decision: no CLAUDE.md there, because the
Claude Code Desktop sessions Voyana uses do load AGENTS.md (reported by
Worker01). Round 2 therefore tests Claude in Voyana in a fresh Claude Code
Desktop session, and asks it which AGENTS.md version it sees. h8v6/heyaira has
a CLAUDE.md that points to AGENTS.md, so the CLI is covered there.
2. Recording the idea: the contract (§2) says it, but the skill steps start at
"Problem", and sessions follow the skill. Fix: skill step 0, "save the idea
before refining it", which uses task_list/task_create with tag pomysl and
reuses an existing idea.
3. Version: nothing asked for it. Fix: the skill's conversation rules now
require naming the notes and versions in the first reply.
The fixes ship as notes v2 (Zastępuje: <v1 id>). Round 2 repeats this test and expects the sessions to cite v2.
Until it passes, K0 status stays "content accepted by the owner; delivery in
progress".