v2.7.0: routing, concurrency and cost

[Param-Harrison]

Summary

• routes in .factory/config.json picks an agent per (type, stage), resolved by resolveRoute
  (routes[type].stages[stage] → stages[stage] → stages.default → "claude"); triage always
  uses stages.triage, since the type isn't known yet. template/.factory/config.example.json
  ships the recommended default (triage/PR on Haiku, plan on Opus, build/verify on Sonnet, docs
  builds on Haiku).
• src/context.ts: buildContextPack assembles AGENTS.md, a route's skills, ARCHITECTURE.md and
  the plan's named files into one pack per stage, capped at 64 KiB, dropped whole (never
  truncated) in priority order. Claude gets it on the same --append-system-prompt as
  STAGE_GUIDANCE; every other agent gets it inlined ahead of the artifact contract.
• Pricing for claude-sonnet-5 and claude-opus-5-5, captured in research/pricing/anthropic.md.
• src/machine.ts: machine-wide concurrency (FACTORY_SLOTS or machine.json), leased in
  machine.db and reclaimed on a dead pid, so two watchers on one machine share one pool. A
  repo's own concurrency becomes a cap on top, and factory doctor advises a slot count.
• Continuous dispatch in watch.ts: the poll loop no longer waits for the whole batch to drain
  before starting new work off a freed slot.
• Spend caps: spend.perIssueUsd and dailyUsd (repo and machine), checked before every stage
  against stage_runs.cost_usd; over a cap parks the issue with a "budget" inbox item.
• proof: "test" | "check" on the plan artifact: check runs named commands instead of a
  failing test, so Markdown-only work (a blog post, a docs page) can ship without one. The type
  list is config-driven everywhere now; TYPE_LABELS is the default, not the only list.

Structural tests (all revert-and-fail proven)

• every (type, stage) pair in routes/TYPE_LABELS resolves to a defined agent
• loadConfig refuses at boot when a routed skill has no SKILL.md on disk, instead of a silent
  mid-run no-op in the context pack (tests/routes.test.ts)
• a proof: check plan (Markdown files, a prose-lint gate, no test file) ships exactly like a
  proof: test plan (tests/scenarios.test.ts, scenario 23)
• the app-agnostic grep now covers every file under template/.claude/skills and
  template/.claude/agents for a named stack tool
• a property test holds running jobs to the machine's slots across two watcher processes and
  confirms a freed slot is picked up within one poll

Test plan

• bunx tsc --noEmit clean
• bun test: 739 pass, 0 fail, 2384 expect() calls, 74 files
• make check: ok (typecheck + test + skills-ref validate)
• Two genuine revert-and-fail proofs shown this release (routed-skill-exists check in
  src/config.ts, and the proof: "check" schema entry in src/schemas.ts)
• Live run on splitbill-demo exercising the new routes and concurrency — pending your reset OK

🤖 Generated with Claude Code