D0017 — The Navigator: how the operating officer works (N1–N8)

[obotclaw[bot]]

One page for the whole Navigator role, because you asked for one. It folds in the two pages from last night that each answered part of it — the worker closeout contract and the supervision question — so there is one place to decide instead of three.

Read it here: https://jwildfire.github.io/obot.roadmap/reports/decisions/2026-08-16-navigator-design/

The role, in your words: improving operations — templates, dashboards, the audit framework, process mechanics. Turning your asks into requirements and checking that workers delivered are instruments of that, not the job itself. It is a builder, not an inspector.

Why stand it up today: six pieces of work shipped overnight with nothing in the roadmap recording them — and almost every one was operations improvement. That backlog is the Navigator's own work, dispatched ad hoc because the role did not exist to hold it. The five untracked issues were not a discipline lapse; they were a job without an owner.

Settled and built in, not offered as a choice: the audit now covers all seven project repos, not just the hub — "tasks live across the project". That is also what makes the headline failure detectable by a machine at all: the six untracked issues were invisible for exactly one reason, that they sat in a spoke repo while the audit only read the hub. Two things come with it so day one is not a mess — the rule gates on work done rather than on filing (the blunt version fires on 26 of 41 open spoke issues today), and the resulting backlog is the Navigator's to triage, not yours. Cost is stated on the page rather than waved past: several times the API calls per run.

Six things you settled this morning are built in rather than re-asked: the Navigator is an agent; its job is improving operations; it owns requirements, tasks, lifecycle and milestones while you own goals and strategy; it decides within its domain and escalates only critical items; and it reports to prime, who reports to you and owes you a recommendation on anything it proposes.

The eight calls, one line each:

• N1 — Day one, and the boundary. Ship the four audit checks that missed last night — now across all seven repos — take every ask as a requirement first, judge every closeout. And accept that it now authors — requirements and machinery changes — which the version you approved days ago was forbidden to do. Recommend: yes to both; the boundary becomes "authors and repairs the plan, never touches the work."
• N2 — Where structure ends. Structure is goals plus the shape of the plan; everything inside it is the Navigator's. Splitting a requirement is inside; adding a lifecycle stage is the shape. Retiring a noisy audit rule is operations; redefining what a milestone means is not. Recommend: adopt that line.
• N3 — What is exempt. The floor is consequence, not size: a typo or dead link is exempt; a one-line change that alters a rule is not. Every exemption written in one line so the rate is countable. Recommend: adopt.
• N4 — The escalation line. Inherit the "critical" bar that already shipped in your dashboard — capped at three, earned by evidence, never by opinion — plus one clause: an escalation must name the thing only you can do. Recommend: inherit unchanged, add the clause.
• N5 — Recording its own calls. A delegated decision that leaves no record is indistinguishable from no decision. Every judgment gets a line; plan-changing calls get a permanent id you can point at and reverse. Primary reader is prime, daily — so it is written dense, not trimmed for a phone. Recommend: adopt.
• N6 — Standing or waking. It stays up so you can reach it, but thinks only on a trigger. Polling every five minutes buys nothing a free script does not already do, and costs $10–50 a day. Recommend: standing to talk to, never polling.
• N7 — Does the separate supervisor survive? Under this model it has no separate job left, and its watching half is a script either way. Recommend: fold it in — four roles, not five.
• N8 — What "delivering" means, and last night's six. A pull request advances a requirement; a question or config request annotates it. Five of the six orphans are missing a link, not a requirement. Recommend: link them, cite the decision on the sixth, forward-only.

Two things on the page are corrections to what earlier drafts told you — both caught before you read them, both left visible rather than quietly fixed:

• All six spoke issues are unattached, not the four then five that earlier counts carried. One of the six was filed by an agent, not asked for by you.
• "The hub side was clean" is withdrawn. Requirements filed overnight were not all correctly parented and boarded — one was never linked to its goal and nine were boarded after the fact by a later check. They are correct now; they were corrected, not filed right. So the gap is not confined to the spoke repos.

And one thing left unresolved on purpose: two independent checks disagreed about the same GitHub state within hours — the nightly audit named four off-board requirements, a separate verifier found nine more. Which one was wrong is not established, and neither is assumed right. Since this design leans on the audit, resolving that is proposed as the Navigator's first task.

Requirement: #195

---

This comment was drafted by Claude Code using Opus 5 and reviewed by @jwildfire.

[obotclaw[bot]]

Decided, 2026-08-16, in chat. @jwildfire:

"I'm good with D0017 recommendations. Implement. Let me know when the agent is active. One thing for the queue: I want the opsdb/sessions page refactored to show a list of all agents along with thier ID, status, cost and the impact they had on the roadmap."

All eight calls (N1–N8 / D0017.1–.8) are adopted as recommended. The record now lives on the artifact itself, which is the source of truth: https://jwildfire.github.io/obot.roadmap/reports/decisions/2026-08-16-navigator-design/

One sequencing change came with the approval and is binding: the disagreement between the nightly audit and the wrapup verifier is resolved before the four new audit checks ship. Two independent checks reported different board state within hours of each other this morning — the audit named four requirements off the board, the verifier found nine more — and neither is assumed correct. Rules added on top of a check whose reliability is unverified are built on sand, so that is the Navigator's first job.

What happens next, in order:

• Resolve the audit-versus-verifier discrepancy and fix whichever check is defective.
• Ship the four audit checks that missed last night, scoped to all seven project repos, with the ancestor rule gated on work done rather than on filing.
• Translate the queued sessions-page ask into a requirement of its own — it is not part of N1–N8.

The queued ask itself: refactor the Operations Dashboard sessions page to list every agent with its identifier, status, cost, and the effect it had on the roadmap.

---

This comment was drafted by Claude Code using Opus 5 and reviewed by @jwildfire