Twelve projects.
One fleet.

In the fleet

• Draft-gate judges reply in JSON. Malformed replies are parsed, shape-checked, range-checked, and retried once before the judge is discarded.
• The video judge runs a repair chain on almost-JSON (bad escapes, trailing commas) before giving up, then asserts exactly five score keys.
• A contract gate on the article handoff raises on any missing or malformed field before the pipeline spends a cent.
• Eval scorers check field, type and range on every scenario run; parse failures are logged with the reason.

In the fleet

• Every wiki note carries source frontmatter. Drafts must cite inline, and a coverage gate hard-blocks publish prep on orphan citations.
• The health coach grounds its advice in a cited knowledge vault and cites notes by path; vault entries carry an evidence-status taxonomy, including "emerging: handle with low confidence".
• Claims with no source are never invented. They get flagged needs-source and routed to Khoji, the research scout, to go and find one.

In the fleet

• The agent loop is capped in config: 50 iterations, spawn depth 1, 600-second child timeouts, 3 concurrent children.
• A failed video-script run retries once after five minutes, then parks: status failed, error posted to Linear, card moved to a failure state with a human assigned. A failure costs one retry, then it stops.
• The draft gate stops auto-redrafting after three failed rounds and escalates to a human.
• Drafting runs two cheap critic loops before anything reaches the judge panel.

In the fleet

• The orchestrator routes by domain through eight spoke contexts, each loading only its own tools and instructions.
• Permissions are scoped per agent. The health coach has no terminal toolset at all: file access inside his own store, plus web for link capture, and nothing else.
• Writes outside an allowlisted root are hard-denied in code. Cron sessions run deny-by-default, so dangerous commands are blocked with no approver present.
• Delegation fans work out to parallel children with the risky tools (delegation, code execution, messaging) stripped from them.

Done

• The health coach keeps an append-only commitments ledger. It remembers what you promised last Tuesday, and asks about it.
• Raw daily metrics are compressed into a trends rollup, so the coaching prompt gets a month of context in a few hundred tokens.
• A wiki knowledge base serves as shared long-term memory across the fleet; session logs with wikilinks carry decisions between sessions.

Left

• No vector store, no embedding-based relevance scoring. Memory is files and git, on purpose: at this corpus size grep beats embeddings and the audit trail comes free. Revisit when the corpus outgrows that.

In the fleet

• A written SOP defines gates as stops, not notifications. Angle confirmation, draft approval and publishing are human decisions, always.
• The budget tool runs suggest, review, apply as separate phases. Nothing writes without per-row human approval, and every write lands in an append-only run log.
• When the draft gate fails three times, the system stops guessing and hands the issue to a human with the full concern history attached.
• Publishing surfaces (Substack, LinkedIn) hold no agent credential by design. A human is the gate.

Done

• Every agent calls through a LiteLLM gateway on its own key, with a hard 30-day budget (breach returns HTTP 429) and a model allowlist (off-list returns 403).
• Models are tiered by task: a cheap tier for chat and routing, a pro tier for anything that writes prose.
• A costs console shows per-key spend for today, month-to-date and 30 days, budget headroom, and cost per produced asset. Telegram fires at 70% of any budget.

Left

• Routing is assigned per task at design time. No runtime classifier picks a model per request, and there is no early exit on confidence.

In the fleet

• Linear webhooks are verified with HMAC signatures (constant-time compare) before anything runs.
• Sweep jobs are idempotent. Already-processed issues are detected by marker comments and skipped, which makes every sweep safe to rerun.
• Failures dead-letter into a dedicated Linear state with a human assignee, instead of retrying forever.
• Provider fallback chains catch upstream stalls, and the public webhook endpoint has its own uptime monitor.

In the fleet

• Article drafts face a blind three-judge panel with decorrelated personas: skeptical scroller, argument editor, brand-voice keeper. The consensus rule is in code: unanimous pass, or it fails.
• The video panel goes further: weighted per-criterion aggregation, a unanimous-veto floor, majority gates, and a contested verdict when judges diverge, instead of averaging the disagreement away.
• Aggregation is deterministic and pytest-covered. The LLM never does the arithmetic.
• An adversarial pair (logic auditor plus pre-mortem critic) red-teams slide decks between versions.

In the fleet

• Failed script verdicts feed only the named concerns back for a surgical revision, then get re-judged fresh. Verdicts are never reused across revisions.
• Slide decks go through versioned audit, revise, re-audit cycles, with the review artifacts kept on disk.
• A fleet eval harness gates every prompt or model change in CI, runs the full suite nightly, and pages Telegram on any red run.

In the fleet

• Every gateway call is traced to Arize Phoenix with token counts. The checklist names Arize; this is Arize Phoenix, self-hosted.
• Eight uptime monitors page Telegram, alongside budget alerts, eval-failure alerts and a nightly registry-drift check.
• Dashboards deploy to a staging environment before production, from separate pipelines.
• Rollback has been rehearsed: pinned image digests per upgrade, backups before every container recreate, restore-drilled data backups, and sha256 stamps verifying deployed configs against the repo.

Closest artifacts

• The agent gateway runs local overlay patches (webhook signature support, payload filters) that would make a reasonable upstream PR. The upstream ask was scoped, then consciously deferred in favour of shipping.
• One tool ships as an installable plugin, but to a self-owned marketplace rather than a public registry.

To close it

• Upstream the request-metadata patch to the Hermes agent framework, or publish the judge-panel pattern as a public plugin with docs and benchmarks.