Fleet next milestone: unbuilt tools, docs hygiene, safety gaps

[lorenzoliuzzo]

MyThingsLab fleet — next milestone (org-wide tracking issue)

Snapshot as of 2026-07-08: 12 repos live — the above plus my-planner and
my-projector, both shipped 2026-07-07 (main CI green; my-projector needed the
new mythings.projects core module, landed via mythings-core#29). 9 My[X]
tools are now shipped, including my-researcher (live arXiv+Tavily topic
research → cited study brief + study path; design PR mythings-core#28). The
remaining designed tools exist only as design docs under
mythings-core/docs/tools/. Note: my-planner/my-projector fail-closed on public
edits when unattended — they need an ASK channel (TelegramPolicy) to post in CI. This issue tracks the fleet's next milestone
across repos — sub-tasks get filed as issues in their own repo and linked
back here as they're picked up.

1. Docs hygiene — shipped tools missing the required historical banner

Per docs/CONVENTIONS.md, a docs/tools/<name>.md must get a one-line
historical banner once its tool ships. Only my-reporter.md has one; five
shipped tools don't:

• docs/tools/my-orchestrator.md — banner pointing at my-orchestrator's README
• docs/tools/my-tester.md — banner pointing at my-tester's README
• docs/tools/my-telegram-bot.md — banner pointing at my-telegram-bot's README
• docs/tools/my-changelogger.md — banner pointing at my-changelogger's README
• docs/tools/mythings-template.md — reconcile naming: the design names a
  dedicated mythings-template template repo, but CONVENTIONS.md's
  "Starting a new tool" already points at the existing my-template repo.
  Confirm whether my-template is the shipped resolution (then banner +
  redirect) or whether a second, GitHub-native template repo is still
  wanted for the "Use this template" button use case.

2. Designed tools not yet built (recommended order, per docs/tools/README.md)

• MySearcher — ranks files relevant to an issue; unblocks MyReviewer and MyGroomer
• MyReviewer — flags correctness bugs on an open PR; needs github.GitHub.diff() (see §3)
• MyGroomer — splits/labels raw issues; needs create_issue/add_labels/list_labels (see §3)
• MyWiki — Q&A over this project's own ledger history
• MyGrapher — keeps a repo's knowledge graph fresh (graphify incremental --update only)
• MyKnowledger — Q&A over external literature via a separate graphify corpus
• MyScaffolder — bootstraps a new tool repo from a proposal
• MyDriftWatcher — flags cross-repo convention drift
• MyAdvisor — recommends a course of action with trade-offs
• MyDescriber — writes/improves a PR's title+description after it opens; needs diff() too
• MyCoder (issue → diff → PR) — was deferred pending a real Engine backend.
  ClaudeCLIEngine shipped 2026-07-07 in mythings-core, so the blocker is
  gone — worth re-scoping this one now, it's the highest-leverage
  remaining tool (the only one that writes the actual feature/fix).

3. Core contract additions — confirm before implementing (per workspace CLAUDE.md)

• github.GitHub.diff() — now has two callers-in-waiting (MyReviewer, MyDescriber)
• github.GitHub.create_issue() / add_labels() / list_labels() — for MyGroomer
• Decide the shared "shortlist-from-a-corpus-then-cite" reuse mechanism —
  5 tools independently want it (MyWiki, MySearcher, MyAdvisor, MyDescriber,
  MyKnowledger). Options on the table: tool-to-tool package deps (like
  my-guard → mythings-core), or promote into mythings-core once
  duplicated. Leaning toward the latter per the design doc; not decided.

4. Safety/process gaps not covered by any existing design doc

Things that matter more now that ClaudeCLIEngine makes unattended, real
(non-Noop) runs possible, and that fleet_dispatch.py already runs multiple
concurrent workers against real accounts:

• No dependency-vulnerability scanning anywhere in the fleet. No
  Dependabot config, no pip-audit/osv-scanner/CodeQL in any repo's CI.
  Every repo pulls mythings-core + a small dep set; one compromised
  transitive dependency currently has no automated tripwire.
• No GitHub App yet — every tool acts under the personal PAT via the
  gh CLI (mythings/github.py has no App/token-scoping logic).
  ARCHITECTURE.md calls out a permission-scoped GitHub App as the
  intended identity model; today a bug in any tool has the same blast
  radius as the human account. Worth scoping once MyCoder makes
  unattended code-writing real.
• Spend/usage guardrails are ad hoc. fleet_usage.py parses
  transcripts for denials after the fact; there's no hard per-run/per-day
  token or dollar cap enforced before a headless session starts. Worth
  a --max-cost/--max-turns guard in fleet_dispatch.py or a small
  policy rule in MyGuard, given --execute already runs real paid
  sessions across two accounts concurrently.
• No secret-leak check on ledger/PR content. dev-ledger/*.jsonl and
  PR bodies are free-text (detail, data); nothing scans them before
  they're committed/pushed. Low risk today (no tool touches real
  credentials yet) but worth a cheap regex/gitleaks-style pre-commit gate
  before MyCoder starts writing arbitrary diffs.
• Kill switch is implicit, not explicit. fleet_dispatch.py defaults
  to --dry-run, which is good, but there's no single documented command
  to halt an in-flight --execute run across both accounts if something
  misbehaves. Worth a one-line runbook note at minimum, a real
  abort-and-cleanup command at best.

Not in scope here

Anything already flagged inside a design doc as "not decided here" stays
undecided in this issue too — it's tracked, not resolved, by listing it.
Actual implementation of any checkbox above happens as its own issue/PR in
the owning repo, following the normal harness loop.

[lorenzoliuzzo]

Light-integration follow-ups drafted for the token/tooling items in §4 and the MyCoder note in §2:

• Make the Engine model knob real (cheapest-capable-first, coordinating) my-things-core#12 — make the Engine model knob real (cheapest-capable-first); children my-reporter#7, my-orchestrator#6
• Optional rtk output-compression hook for headless dispatch workers #2 — optional rtk output-compression hook for headless workers

Deliberately scoped narrow: no semantic-index/graph tooling (repos too small yet), no external router, no rtk on human sessions.