Eve library

Eve Everywhere — durable initiative handoff

Guide

Revision 1

On this page

Status: capability-breadth initiative closed on 2026-08-22; completion evidence re-audited through 2026-08-30.

This is the versioned cross-session handoff for EVE_EVERYWHERE_TODOS_2026-08-19.md. Read it with that ledger and docs/agents/eve-smx-maintenance.md; the ledger remains authoritative for each checkbox and its evidence.

Evidence boundary#

The 2026-08-22 close note says external auto-memory files named eve-everywhere-initiative and eve-deep-polish-initiative were written and cross-linked. They are not present in this checkout or in the sole surviving project-memory index. The closing commit f87ce61defc4a275ee83adae494fb3c4910491eb changed only the TODO ledger, so Git does not preserve those notes. This file records the current, source-grounded handoff; it does not claim to reconstruct missing historical bytes.

What is live#

  • The builder assistant spans admin operations, the product graph, decisions and ADR export, content ideation and dispatch, event-sourced work items, attributed agent lanes, shipped narratives, and the global member-safe release-note feed.
  • Mutations park exact confirmation cards and execute only after approval. Decline and pre-approval paths must leave complete domain state unchanged.
  • Curated admin tours, drawer voice, scoped selection handoff, contextual invocation, operator memory, and the workbench-kit read/write surface are implemented. Operator memory is default-on when its database is bound; the durable-memory flag is now a kill switch, and absence of the database degrades honestly to session-only memory.
  • The standing serving route is OpenRouter deepseek/deepseek-v4-flash-0731, price-sorted and fp8-pinned, with the measured endpoint allowlist limited to Baidu, DeepInfra, and StreamLake (all fp8), unless an explicit environment override is supplied. Local credentials remain outside Git and are loaded according to AGENTS.md.
  • The twelve-frame showcase was fully refreshed on 2026-08-29. Its current full-circle frame is a self-owned replay; the distinct 2026-08-22 archive is historical Git/image evidence whose original database ledger was not retained.

Decisions and standing gates#

  • The release-note feed is global. There is deliberately no per-member work item linkage or targeting.
  • Direct Hathor imports and raw-SQL bridging were declined; the HTTP seam is the supported boundary.
  • A staged member complaint was declined because it would fabricate the first hop of the full-circle evidence.
  • The Eve Codex lane remains active under user-maintained quota. Run its smoke before routing work to it.
  • Blind human labels for the 42-transcript SMX sheet remain human-gated. Do not fill or infer those labels on the operator's behalf.
  • Workbench-kit access continues to ride the shared authorization/audit boundary. A tool being registered is not sufficient evidence that its routed skill actually offers it.

Landmines worth carrying forward#

  1. Tool registration is not route reachability. The completion audits found plan_tara_calendar registered but missing from the workbench-read routed allowlist, then found workbench_kit_read absent from the write family's read-before-write surface. Derive exact inventories and assert the named invocation on the served route.
  2. A fluent answer is not an action. Explicit tour and favourite requests have produced prose and pseudo-function JSON without a tool call. The route now holds that candidate and requires the exact named action; informational, negated, and third-party questions must continue to abstain.
  3. Confirmation proof is complete-state proof. Count equality can hide a changed row. Compare canonical complete rows and relevant files before the card, after parking, and after decline.
  4. Historical screenshots are not replayable ledgers. Keep immutable image and Git ancestry separate from a new self-owned regression replay. Never describe recreated fixtures as the lost historical event stream.
  5. Use real shell controls in browser proofs. Escape collapses a dock; it does not necessarily dismiss it. Close the tour, drawer/dock, and any full-screen room through their actual visible controls before walking the shell.
  6. The composed stack includes domain services and content. Tara and Arete are out-of-process; an empty or absent service can turn a supposed happy path into an honestly empty/failed one. Use isolated migrated databases, authored fixtures, exact cleanup, and one frontend server at a time.
  7. The model slug is not the whole route. The first weekly cache breach was provisionally blamed on measurement churn; a quiet follow-up disproved that story and found the drawer had omitted the fp8 preference. Preserve the provider route pins and compare billed, routed behavior rather than catalog sticker prices.
  8. Console cleanliness is part of behavior. Conflicting React border shorthands and cross-origin Tara media were visible only in the composed walk. Keep the same-origin audio proxy, range semantics, and zero-error browser assertions.

Distillery register#

The 2026-08-22 closing pool's weakest named cases were builder-wbw-release-note 4/10, builder-wbr-explorer-link 4/10, builder-adv-delete-no-tool 6/10, builder-wbw-proposal 6/10, and builder-wbr-item-readback 6/10. Those are historical pool measurements, not current scores.

The first post-initiative kit pool on 2026-08-23 left four measured advisory cases after repairing the route and grader defects:

  • builder-kit-read-sparks-inbox — 9/10; one empty-reply provider tail.
  • builder-kit-write-schedule — 9/10; the correct tool pair was chosen 10/10, with one honest card-time refusal.
  • builder-kit-write-transition — 8/10; the correct tool pair was chosen 10/10, with two honest card-time refusals.
  • builder-adv-kit-read-bogus-concept — 6/10; empty-reply tails plus an honest refusal phrasing outside the grader vocabulary, with no fabricated dossier.

Do not promote an advisory case from prose or a single rerun. Use the admission, rotation, and Wilson-floor rules in docs/agents/eve-smx-maintenance.md.

The prose-primes-failure lesson#

The dispatch distillery tested a capability-prose edit under a pre-registered same-window control. It improved the positive dispatch case from 8/10 to 10/10 but degraded the ghost-dispatch honesty case from 9/10 to 7/10 and tripled claim-before-refusal. The edit was reverted. A dispatch-claim wire hold and completion checker then brought both cases to 10/10 without adding prompt bytes.

Treat this as a mechanism rule: prose that repeatedly names an action can prime the model to narrate that action. Prefer an exact tool contract, a wire hold, and a corrective checker; always retain a byte-identical control when measuring prompt changes.

Verification entry points#

  • pnpm verify:eve-smx-weekly — standing route/cost evidence lifecycle.
  • node tools/eve-everywhere/verify-full-circle-provenance.mjs — historical and current full-circle separation, hashes, and ancestry.
  • node tools/eve-everywhere/verify-deep-completion-matrix.mjs — every checked ledger requirement has current proof, including nested Phase 7 rows.
  • node tools/eve-everywhere/verify-phase-12.mjs — durable exit-battery, showcase, handoff, and ledger closeout.
  • apps/oshun/web/e2e-inspect/admin-tool-battery-2.spec.ts — exact Phase 3–9 read/mutation inventory and decline invariants.
  • apps/oshun/web/e2e-inspect/phase-15-1-happy-path-spine.spec.ts — composed do/tour/audit browser spine.
  • docs/audits/eve-phase-12.1/2026-08-29-run.json and docs/audits/eve-phase-12.2/2026-08-29-run.json — sanitized completion-run manifests.