[Daanish2003]

Summary

Spec #946: the SDK follow-ups, in two groups.

CI reliability

• bot-hand-off 'shows on the to-do list as waiting' fails in CI #929: bot-hand-off "shows on the to-do list as waiting" read calls.at(-1), which under CI timing was the Memory pass's Model call, not the Turn's. Tests now pick the Turn's call with one shared isTurnCall (@ai-chat/test-support/bot-turns). This was a test bug, not a product bug.
• bot-demonstration-capture 'ends as Done at 200 stored steps' fails under load and in CI #945: demonstration-capture "ends as Done at 200 stored steps" raced the drafting pass (drafting → failed about 65 ms later) and stored 200 steps one by one. It now seeds 197 steps, presses to the cap, scripts the drafting round, and asserts saved.
• account e2e: the rename test changes the shared seeded user #931: no change needed. seedUser already makes one user per call.

Boundaries

• Routines and the Helper card read the job tables through the runner #922: Routines and the Helper card read jobs through the runner's internal store (runner/internal), not the job tables. Reconcile is one anti-join again (idsWithoutActiveJob), and fireKeyPrefix is the single place the key prefix is built.
• Feature jobs reach only the services in their needs #923: coreContextOf and the [coreContext] symbol are gone. Memory's pass reaches Models and the thread hint through needs (models, threads core services).
• The Turn core still computes Skill and Knowledge changes #919: features compute their own Turn changes through a synchronous turnChanges hook, using the shared writtenFiles(transcript) helper. The core only stores the results, in the same Turn columns, with no migration. With neither feature enabled, no scan runs.
• Features declare their enums, so --drop removes them #932: a feature's schema can declare enums and sequences. --drop drops what the feature declares plus what its tables' columns and defaults use. It never drops a core enum, or anything another table uses or an enabled feature declares.

Closes #946

How it was tested

• bot-hand-off 'shows on the to-do list as waiting' fails in CI #929: forcing CI's ordering turned the test red with CI's exact message. With the fix it went green, then passed 5/5 alone and under load. CI was green on its own PR (test(bot): CI proof for #929 (do not merge) #947, based on main).
• bot-demonstration-capture 'ends as Done at 200 stored steps' fails under load and in CI #945: an instrumented run found the race. The new test went red with the cap broken, and red with failed before the drafting round was scripted. It passed 5/5 alone, and passed in CI on its own PR (test(demonstrations): CI proof for #945 (do not merge) #948). That run's one failure was bot-hand-off, which bot-hand-off 'shows on the to-do list as waiting' fails in CI #929 fixes and test(demonstrations): CI proof for #945 (do not merge) #948 did not carry.
• account e2e: the rename test changes the shared seeded user #931: the full e2e suite passed 3 runs in a row under --workers=2 (58/58 each).
• The Turn core still computes Skill and Knowledge changes #919: a new test went red on main (it stored both kinds of changes with neither feature enabled). A @ts-expect-error test rejects an async turnChanges.
• Features declare their enums, so --drop removes them #932: these tests were each seen red first: an unused declared enum, an undeclared enum or sequence that a column uses, the guard for another feature's enum and the core-enum guard, and an enabled feature that declares the same enum.
• Routines and the Helper card read the job tables through the runner #922: the reconcile query-count test was red on the two-read version.
• Full verify (lint, format, typecheck, all vitest projects, including Inngest) is green on the integration tip, plus one more full e2e run after the review fixes.
• Branch review (code-review, high) found 10 findings, R1–R10, all fixed in 86200dc..96c4952.

Decisions I made

• Public API (AGENTS.md §6, please approve):
  • New CoreService names models and threads (ThreadsService) that a feature can list in needs (Feature jobs reach only the services in their needs #923, which allowed "or a core service").
  • New exports on @botsdk/bot/feature: the turnChanges hook with FeatureEndedTurn / FeatureTurnChanges (The Turn core still computes Skill and Knowledge changes #919, per the triage), the writtenFiles helper, and schema.enums / schema.sequences (Features declare their enums, so --drop removes them #932, approved 2026-10-27).
• Not public: the runner's new job reads are internal only (runner/internal), and sdk.runner is unchanged from main.
• Build order: --drop takes the union of what a feature declares and what its tables' columns use, rather than "exactly what the feature declares", so a feature that doesn't declare its enums keeps main's behaviour. buildBotSdk runs the Models part's init before enabling features, so a feature's context holds models. This relies on the existing "core parts start with Models" assertion.
• Test trade-off: the bot-demonstration-capture 'ends as Done at 200 stored steps' fails under load and in CI #945 test no longer drives 200 real steps through the in-page queue, so the queue's own 200-entry cap is untested.

Risk

• No migration and no schema change.
• --drop drops enums and sequences the feature declares, in addition to the ones its columns use. This removes more than before, but only things the feature owns, and the guards above still apply.
• The core no longer computes Skill and Knowledge changes. A Host without skills() or knowledge() stores none, as it should.

Out of scope, noted by #919: core modules that still import feature internals are computer-tools (isSkillFile), demonstration-draft (skill files), review (memory) and bot/knowledge-files.ts (knowledgeDirectory, walk scripts).

🤖 Generated with Claude Code