rdbms-playground/docs/handoff/20260612-handoff-67.md

Session handoff — 2026-06-12 (67)

Sixty-seventh handover. Continues directly from handoff-66 (ADR-0048 seed Phase 2, committed). This was a manual-testing pass: the user exercised the app, found several rough edges, and we triaged each into fix now vs file an issue. Net result: three bug fixes committed and three enhancement issues filed.

§1. State at handoff

Branch: main. Working tree clean; all work committed. Unpushed (push is the user's step).

Tests: 2407 passing / 0 failing / 0 skipped / 1 ignored (the long-standing friendly doctest). Clippy clean (nursery, all targets). +7 over handoff-66's 2400.

Commits since handoff-65:

f7155ce fix(input): thread the `:` one-shot escape into live SQL feedback
4cacb82 fix(completion): don't flag a table alias used before its FROM clause
c3e0103 fix(completion): flag-aware partial so a dash completes flags, not keywords
30b2677 docs: ADR-0048 Phase 2 implemented + handoff 66
a12facc feat(seed): set override clause + column-fill (ADR-0048 Phase 2)

(a12facc/30b2677 are the Phase-2 work documented in handoff-66.)

§2. Bug fixes this session (all committed, all tested)

1. c3e0103 — flag completion ate the dash. Typing a flag at a flag position (add 1:n relationship … -) offered the on keyword and, on accept, produced -on / ---create-fk: the partial-token walk stopped at -, so the dash was outside the replaced range. Fix: flag-aware partial detection (a dash-prefixed token at a word boundary is a flag-in-progress, gated on a flag being expected so where x = -5 stays a number) + a unified flag matcher (trim_start_matches('-')). Affected all flags. 4 tests + 2 partial-flag snapshots updated (they'd captured the latent bug).

2. 4cacb82 — table alias flagged as an unknown column. In a SELECT, the projection (sum(ol.count*…)) can reference an alias whose FROM … OrderLines ol sits after the cursor. The candidate engine recovers that via the §10.6 full-input lookahead (ADR-0032), but invalid_ident_at_cursor only walked text before the cursor — so ol matched no scope and got a red "ERR" overlay on an otherwise valid query. Fix: give the validity check the same full-input lookahead and bail when the partial prefix-matches a binding's alias or table. 1 test.

3. f7155ce — the : one-shot escape broke live SQL feedback. Submission strips the : (ADR-0003), but the live feedback kept it in the buffer handed to the walker, which bailed at the :. Effect: under :, Tab completed nothing and a valid query could flash [ERR] — while the same line in full mode advanced worked. (The hint already stripped it, hence "hint shows the name but Tab does nothing".) Fix: one shared App::feedback_view() (the :-stripped SQL + mapped cursor + stripped offset) routed through completion (with a replaced_range offset shift), the validity verdict, and rendering (new render_input_runs_feedback highlights/overlays the view shifted by the offset; the : renders as plain text); the ambient hint was consolidated onto it (removing the duplicate strip_one_shot_prefix). 3 tests + the 9 existing colon tests still green.

§3. Investigated, no code change (working as designed)

• Comma-FROM implicit join (select … from A, B, C) is deliberately rejected — ADR-0032 §11 / OOS-3: "comma-FROM teaches habits we do not want to encourage; CROSS JOIN covers the same shape explicitly." The explicit equivalent (CROSS JOIN … WHERE …) works.
• sum(…) returning one row with no GROUP BY is correct SQL (the aggregate collapses the result to one row; SQLite/the playground allow the non-aggregated columns where Postgres would error). The user's query needed group by o.id. Verified (1 row).

§4. Open issues filed this session — next session's candidates

All on git.lazyeval.net/oli/rdbms-playground, label enhancement:

• #26 — seed <table> hint omits the optional count. A complete command's optional positional number has no Tab candidate, so it's invisible. IntroProse doesn't fit (it only fires for incomplete required slots; the completing Seq match clears the hint). Needs a way to advertise optional positional non-keyword args. (I attempted + reverted this during Phase 2; see the analysis in the issue.)
• #27 — Bottom status line: keybindings-only, context- and state-aware. Per-nav-focus keybindings (Input vs sidebar), include transient states (Tab-cycle, history) — user preference — and add mode advanced to the empty-input hint. May warrant a small ADR.
• #28 — Reconsider relationship prose in add column (incidental DDL) confirmations. Currently by design (ADR-0044 §1 keeps prose, not diagrams, for incidental DDL). User preference: do NOT show the References: / Referenced by: block in the add-column confirmation at all — focus on the change just made. This revisits a decided area → land as a new ADR superseding the relevant part of ADR-0016 §5 / ADR-0044 §1; confirm scope (just add column, or all incidental DDL).

§5. Other open work (unchanged from handoff-66 §6)

seed is feature-complete (requirements.md SD1 [x], SD2 [x]). Remaining roadmap, user's call:

• H2 hint — the last A1 gap (its own ADR).
• TT5 CI — test infra exists; no CI workflow yet.
• TT4 PTY (Tier-4) — ADR-0008 specifies it; not wired.
• Larger: V4 journal, tutorial/lesson system (each needs an ADR).

A possible quick follow-up: a friendlier "use an explicit JOIN" parse-error for comma-FROM (point 1) — not filed; mention if wanted.

§6. How to take over

1. Read handoffs 65 → 66 → 67, CLAUDE.md, docs/requirements.md.
2. seed Phase 2 is done (ADR-0048 Status block is current). The manual-testing fixes (§2) are committed and green.
3. Pick from §4 (filed issues #26/#27/#28) or §5 (roadmap). #28 is a decision/ADR; #27 is UX (maybe ADR); #26 is a hint-system enhancement.
4. Consider a cargo sweep at this milestone (target/ grows across sessions).