Session handoff — 2026-05-31 (53)

Fifty-third handover. A two-issue pair off handoff-52's open list. Closed #15 + #16 in a single commit — they share one piece of infrastructure (a curated SQL-function-name list), exactly as handoff-52 §5 predicted ("do together"). One ADR amendment authored: ADR-0022 Amendment 6, plus an ADR-0031 status note.

The session is worth reading for one process save: the implementation /runda pass (DA hat) found a real duplicate-candidate defect that the green test suite had not — a column named like a function (count, date) was listed twice in the completion line. It also caught a clippy regression I'd introduced and a whitespace/EOF wart. None of these were visible in "tests pass." See §2.

§1. State at handoff

Branch: main. HEAD 6d8c9ee. Ahead of origin/main by 2 — handoff-52's 8311de4 (#9) was still unpushed at the start of this session, and 6d8c9ee (this session) is on top. Tests: 2107 passing / 0 failing / 1 ignored (full cargo test; the lib-only figure is 1538 — see the baseline note below). The 1 ignored is the same long-standing ```ignore doctest in src/friendly/mod.rs as every prior baseline. Clippy: clean (nursery group). Push is the user's step.

Commit since handoff-52's 8311de4:

6d8c9ee feat: curated SQL function list — Tab completion (#15) + typing-time typo hint (#16)

Test-baseline bookkeeping (read once). requirements.md's "Test baseline" had drifted: it still read 1131 (the ADR-0027-era lib figure) because the intervening issue fixes #8/#13/#12/#7/#9 landed tests without bumping it — the doc historically moves the number at ADR milestones, not per issue. With the user's confirmation this session refreshed it to 1538 lib / 2107 full, with an Amendment-6 entry appended to the running list and a note explaining the prior drift. Going forward the two figures are: cargo test --lib = 1538, cargo test (all binaries) = 2107.

§2. The process save that paid off (read this)

Both issues went: read issue bodies + the ADRs they name → trace the exact symbols (completion.rs, sql_expr.rs, the lockdown tests) → escalate the genuine design choices to the user (function set, colour, ADR-or-not) → test-first build → /runda on the implementation → fix findings red-first → commit.

The implementation /runda (DA hat) is where the value was. Three findings, none visible in a green suite:

Duplicate-candidate defect (real). A column named exactly like a function — a column called count or date, both plausible — was listed twice in the completion line (count/Identifier + count/Function), distinguishable only by colour. The existing code deduped identifiers-vs-keywords but I had added no functions-vs-identifiers dedup. Fixed: functions.retain(|f| !identifiers.iter().any(|i| i.eq_ignore_ascii_case(f))) — the column wins (same spirit as the keyword-wins rule), case-insensitive so a Count column also suppresses the count function. Proven red-first: the new test fails without the retain, printing the exact [count/Identifier, count/Function] duplication.
#16 fix not yet proven load-bearing. Reverted the invalid_ident change and confirmed the typo test FAILs without it — the fix is doing the work, not coinciding with another path.
Clippy regression + EOF whitespace I'd introduced (a field_reassign_with_default in the new test, and stray trailing blank lines from a throwaway-probe removal). Both fixed; EOF normalised to be byte-identical to HEAD rather than sneaking in a cleanup of the pre-existing trailing-blank convention.

Lesson reinforced (matches handoff-52 §2.2): /runda on the implementation, not just the design. A green suite is necessary, not sufficient.

§3. Resolved issues

#15 + #16 — curated SQL function list (`6d8c9ee`, ADR-0022 Amendment 6)

Both hinge on the same gap: the advanced-mode SQL expression grammar (ADR-0031 §1) admits a function-call shape at its sql_expr_ident slot but, by design, does not know which names are real functions (the slot is IdentSource::Columns, optimised for the column-reference common case; the walker is a structural matcher, not an evaluator). So both issues wanted the same thing — one source of truth for "what SQL function names does this playground recognise."

New file — src/dsl/sql_functions.rs:

KNOWN_SQL_FUNCTIONS — 23 names, sorted + lowercase (a pinned invariant). The broader curated set (user choice): aggregates (count/sum/avg/min/max), common scalars (length/upper/lower/trim/substr/coalesce/abs/round), and a broader scalar tier (date/datetime/strftime/hex/ifnull/nullif/replace/ instr/typeof/random).
is_known_function_prefix(partial) — case-insensitive prefix test, empty prefix matches all.
cast deliberately excluded — a faithfulness finding during build: SQLite's CAST(expr AS type) is not a plain name(args) call the expression grammar parses (expr AS type is not a valid argument expression), so offering cast would surface a candidate that cannot complete. Documented + test-locked (cast_is_excluded, cast_is_not_offered_as_a_function_candidate).

#15 — Tab completion (completion.rs, "Source 1.8"): at a sql_expr_ident slot the curated functions are offered as candidates (prefix-filtered like every other source), ordered after keywords/types (a learner reads clause keywords first, then discovers callables). New CandidateKind::Function + a ninth Theme colour tok_function (a blue, distinct from keyword purple / identifier teal / type pink-magenta in both dark() and light()) — parallel to how #8 gave types tok_type. ui.rs maps the new kind.

#16 — typing-time typo hint restored (completion.rs, invalid_ident_at_cursor): the issue-#6 fix had dropped the "no such column" flag wholesale at sql_expr_ident slots (to stop the false positive on names like sum). Now it bails only when the partial prefix-matches a known function; otherwise it falls through to the existing schema-column check. So select Agx (no FROM yet — schema diagnostic silent) warns at typing time again, while select sum stays quiet. The issue-#6 lockdown tests (genuine_column_typo_in_complete_select_…, advanced_select_partial_function_name_not_flagged_…) still pass, and the submit-time unknown_column diagnostic path is untouched.

Dedup (the DA finding, §2): a column named like a function appears once; the column wins.

Coverage added (15 tests): sql_functions::{list_is_sorted_and_lowercase, list_has_no_duplicates, cast_is_excluded, prefix_match_is_case_insensitive, empty_prefix_matches_all, unknown_prefix_does_not_match}; completion::{sql_expr_slot_offers_known_function_candidates, projection_slot_offers_known_function_candidates, sql_function_candidates_filter_by_prefix, sql_function_candidates_carry_function_kind, function_candidates_absent_at_non_expression_slots, function_candidate_deduped_against_a_like_named_column, cast_is_not_offered_as_a_function_candidate, invalid_ident_fires_for_genuine_typo_at_sql_expr_slot, invalid_ident_does_not_fire_for_function_prefix_at_sql_expr_slot, invalid_ident_does_not_fire_for_column_prefix_at_sql_expr_slot}; input_render::advanced_select_genuine_column_typo_before_from_warns_at_typing_time; theme::function_colour_is_distinct_from_keyword_identifier_and_type.

Mode-gating confirmed sound (DA probe): DSL where-expressions use role expr_column; only SQL uses sql_expr_ident. Functions therefore never leak into simple mode. Slot coverage probed empirically — functions surface at WHERE, comparison RHS, HAVING, projection list, and post-AND.

§4. ADR / docs work

ADR-0022 Amendment 6 — new (6d8c9ee): the curated function list, the CandidateKind::Function + tok_function colour, the #16 invalid-ident narrowing, the dedup, and the no-validation-allowlist posture (the list drives completion + the typo hint only, never parse-time acceptance — an unknown function still parses and surfaces an engine-neutral execution error, preserving ADR-0031 §6/§7). Coverage list inline.
ADR-0031 status note (6d8c9ee): records that Amendment 6 layers the known-function list on top of the sql_expr_ident slot — grammar unchanged, no-allowlist posture intact; softens §5's "function names are not completed" line to "completed from a curated pedagogical list, not a validation allowlist."
README index (6d8c9ee): ADR-0022 entry gains the Amendment-6 paragraph; ADR-0031 entry gains the status-note reference.
requirements.md (6d8c9ee): I3 (tab-completion) + I4 (highlighting) refinement notes for #15; test baseline refreshed (see §1).

§5. What's open

Bug reports still open (filed handoff-50)

#	Title	Label
10	`[error]` tag colour identical to `[system]` (ADR-0037 gap)	enhancement
11	Copy output panel contents to the system clipboard	enhancement
14	`--resume` should restore the last-used input mode	enhancement

Closed this session: #15, #16 (commit 6d8c9ee — close them on GitHub once pushed).

Categorisation / notes for pickup

#10 ([error]/[system] tag colour) is the most adjacent and was already flagged "most adjacent" in handoff-52: ADR-0040 kept the [error] tag on failure-reason lines, so the colour collision is still live. Needs an ADR-0037 amendment (it owns output tag styling). Smallest of the three; a clean next pick.
#14 ties to ADR-0015 Iter 6 (--resume + persistent input history); small, builds on the existing last_project machinery.
#11 clipboard copy — self-contained but pulls in a new cross-platform clipboard crate (most surface area of the three).

Other tracks (from requirements.md)

Track 2 project storage Iter 5/6 — export/import + --resume + persistent input history + migration framework.
C3a (modify relationship), C4 (m:n convenience).
H1 friendly DB-error layer (partial); tutorial/lesson system (needs its own ADR); V4 session log + Markdown export; I1/I1b multi-line + readline; I3 tab-completion polish (this session advanced it for SQL function names); I4 highlighting beyond input echo.

§6. Process pins (carried forward + this session's reinforcement)

/runda on the implementation, not just the design. This session's save (§2): the impl pass found a real duplicate-candidate defect, a clippy regression, and an EOF whitespace wart — none visible in a green suite. Held from handoff-52 §2.2.
Red-first proof for every fix. Each fix this session was proven by reverting it and watching the relevant test FAIL before restoring. (Global rule: reproduce bugs with a failing test first.)
Escalate genuine design choices — do not decide for the user. The function set, the candidate colour, and whether to write an ADR were all surfaced via AskUserQuestion before building; the stale test-baseline refresh was also confirmed, not assumed.
Faithfulness over convenience. cast looks like an obvious member of "common scalars" but does not fit the call-shape grammar — excluded with a documented reason + a lock test, rather than shipped as a candidate that fails to complete.
Match HEAD, don't reformat unrelated things. The EOF fix restored the file's pre-existing trailing-blank convention byte-for-byte rather than "tidying" it — keeps the diff to the actual change (the user's append-only / no-gratuitous-reformat git philosophy).
Reproduce-first (handoff-51 §2.1) and no easy-way-out on architectural discrepancies (handoff-51 §2.2): both still stand.
Push is the user's step. Commits and pushes are decoupled; unpushed commits are a normal working state. Two are unpushed now (8311de4, 6d8c9ee).

§7. How to take over

Read this note, then CLAUDE.md, then docs/requirements.md (note the refreshed baseline), then docs/adr/README.md.
The codebase is on main at 6d8c9ee, working tree clean, 2 commits ahead of origin (both unpushed — see §1).
If continuing the bug-list run: #10 is the smallest and most adjacent next pick (needs an ADR-0037 amendment for output tag styling); then #14, then #11.
The tooling note from earlier this session: an intermittent empty-output bug in a newer Claude Code build was the cause of the "Bash/Read returning nothing" symptom — the user reverted to a working build mid-session. If output goes blank again, suspect the harness, not the repo.
Honour the process pins in §6 — /runda on implementation + red-first proof are the two that earned their keep this session.

13 KiB Raw Blame History