How we work
The working agreements, the roles, the loop, and the instruments, for anyone (or any session) picking the project up.
- SCHEMA — the wiki’s working agreements: page types, frontmatter, the decision log’s columns, what a session records
.claude/skills/mo-lead/SKILL.mdin the repo — the lead’s role and loop: brief, fresh worker, verify with probes the brief did not name, record, merge, reportHANDOFF.mdat the repo root — the state and the queue for the next session- roadmap — the board (now, next, waiting on Robert, recently done), the phase table, and the done rows
- log — one entry per session, newest last
- decision-log — the decisions
- The sessions: session-01, session-02, session-03, session-04, session-05, session-06
The roles
Robert’s 19 Sep instructions keep Astra as lead in this Mac session, with fresh
Astra workers at low reasoning in Herdr panes. The oracle requirement and Amp
worker workflow are superseded. The lead writes briefs, verifies, decides and
records; workers own code under toolchain/ and examples/. Robert reviews
the decision log. mo-lead holds the exact launch and acceptance procedure.
Implementation workers have separate worktrees and exact base/write scopes; read-only reviews may share the lead checkout. The lead inspects returned changes, preserves raw evidence and reruns acceptance in its own checkout. A worker’s green report is not acceptance. Preserve historical worktrees and keep sealed suites away from workers. Robert’s overnight authority authorizes lead decisions and bounded setup/implementation while he is AFK; independent work continues around prerequisites requiring his presence.
The auditor (17 Sep 2026)
A separate reader Robert installed after the outside review of 17 Sep: a Perplexity session only he opens, which reads raw evidence and files audit/mo-audit-<date>-<subject>.md in the repo, never Fable’s synthesis first. Three stopping rules are ratified there (the runtime claim and the capabilities claim on program 7, the language’s catch claim at erosion generation ten). Fable’s parallel readings, the evidence bundles under audit/evidence/<date>/, and where the two readings meet: the-audit-workflow.
The rules that were learned the hard way
Nothing is final until measured (d28-nothing-final-until-measured); every step ends in a numbers table, best of five, both runtimes. Zero new syntax where possible; a grammar change is Robert’s call. The laws stay unless a round shows them costing loops without catching bugs. Worktrees are evidence and are never deleted or merged. A control run is pre-registered: predictions on the page before any session starts. The lead commits by path only. Verify with your own probes, on real inputs, with your own client: every finding that mattered came from an input no suite sent.
The instruments
mo-executor-foundation is the accepted bounded fixture executor and protected verdict path for mo-first-coding-harness. mo-agent-terminal-auth, mo-provider-foundation, mo-coding-fixture-v1, mo-workspace-foundation and mo-provider-auth-v1 have independent bounded acceptance, as does mo-provider-bridge-v1 for recorded native-history continuation. mo-application-build-v1 verifies the isolated pinned toolchain; mo-workspace-recovery-v1 adds cleanup ownership; mo-workspace-http-v1 exposes six tools with independent owner lifetime. mo-application-workspace-v1 adds the recorded Mo caller in a separate slice. Fixture acceptance does not replace application checks or historical obligations.
The suites live beside their plans under mo-wiki/plans/*-suite/; see the-rounds for the list. mo-wiki/tools/lint.py checks links, frontmatter, tags, orphans, and sizes. The site is built by Quartz from site/ and published on every push to main.
The research behind the design
- reading-pack-2026-09 and research-summary-2026-09 — the reading pack and the summary
- README — the surveys, author profiles, and standards pages
raw/— every source as it was ingested, articles and research runs