scripts/ab-v9/run_matrix.py runs the ab-v3 corpus (52 prompts) through
the full Rust orchestrator per provider (op-smoke headless,
OPENPENCIL_MANIFEST=1) and scores M3 expected-shape (required roles in
the saved .op tree) + M5 element selection, appending rows per cell for
crash-safe resume. Keys come from env only (MM_KEY/ARK_KEY/DS_KEY).
op-smoke grows OPENPENCIL_SMOKE_KEEP_THINKING=1 to keep MiniMax
reasoning ON: ab-v9 showed M3-nothink emits lazy minimal manifests
(17%, ~10s answers) while M3-with-thinking lands 60% with composite
tied-best at ~110s — the MiniMax production routing target.
14th pre-validation detector + a preventive skill rule.
User-reported 2026-05-10 "Bistro" mobile food app shipped with root
padding [0,16,0,16] AND a "Today's Specials" section padding [0,24].
Effective gutter = 40px on a 375px page → only 295px of usable
content width. Reads as "too much padding" / pinched.
Two pieces:
1. layout.md AESTHETIC HYGIENE block now teaches "page gutter goes
on ONE layer, not both" — pick root horizontal padding OR
per-section horizontal padding, not both. Default convention:
root carries the gutter, sections set vertical-only padding.
Hero / banner / image-bleed sections then sit edge-to-edge by
simply NOT adding horizontal padding (root's gutter shows
through). Preventive teaching at prompt time.
2. detectStackedHorizontalPadding (info-only, detect-only). Walks
every mobile-shaped root (width 320–480 + tall + multi-child),
compares root horizontal padding against each direct child's
horizontal padding; flags the section as the offender when both
are > 0. Page-shape filter mirrors detectEdgeSectionPadding so
the legitimate component-internal padding stacking pattern
(chip → badge → icon, etc.) doesn't trip it. Severity is INFO
because a section may legitimately want a deeper inset for
visual emphasis — let the user/agent decide via audit panel.
Side-quest: scripts/ab-corpus/check-stacked-padding.ts ships with
this commit so the next stacked-padding-style detector calibration
can survey corpus frequency without rebuilding the harness.
Codex stop-hook review caught: the detectTextBgContrast ancestor walk
treated any wrapper with a solid `fill` entry as the bg color, even
when the fill was effectively invisible. The classic miss case:
page { fill: cream }
└─ wrapper { fill: [{ type: 'solid', color: '#FFFFFF', opacity: 0 }] }
└─ text { fill: cream }
Without the guard, the detector picked the wrapper's white fill as bg
and reported a healthy contrast ratio against the cream text — masking
the real cream-on-cream failure that lives one level up.
firstSolidColor() now skips fills with `opacity === 0` and 8-hex colors
whose alpha byte is `00` (e.g. `#FFFFFF00`). Both produce no visible
color, so the ancestor walk continues past them to the real bg.
Semi-transparent fills (opacity 0.5, 8-hex alpha 80, etc.) are out of
scope — the detector still treats them as opaque rather than trying to
math the layered composite. Tests pin both: opacity=0.5 + alpha=80
stay treated as bg.
4 new test cases cover the fix plus the boundary (opacity=0.5, alpha=80
should NOT be skipped). Full corpus replay shows 14 hits unchanged on
the 470-row corpus — no false-positive regression introduced.
Two complement scripts that ride alongside replay-detectors.ts:
- inspect-shape.ts: bucket every applied row's root by (width-bucket /
height-bucket / aspect-ratio / child-count). Used 2026-05-10 to
diagnose why detectEdgeSectionPadding scored 0 hits on a 220-row
mobile subset — turned out 49% of mobile rows produce roots with a
string-typed width ("fill_container" / "fit_content") because the
element-tools path emits component fragments, not pages. The
detector predicate `typeof width === 'number'` correctly skips them
→ 0 hits is the EVAL HARNESS coverage gap, not a detector bug.
- inspect-issue-category.ts: print every issue in a chosen category
with row id + node id + reason. Generic version of inspect-contrast-
hits.ts. 2026-05-10 used it to validate
excessive-frame-effects (4/4 TP — blur 48 cards + OTP slot spread)
and the two borderline mixed-sibling hits (header padding outlier,
spacer cornerRadius outlier — known role-aware limitation, 0.2%
noise rate, accepted).
Together with replay-detectors.ts these three give a fast empirical
loop for tuning a detector against real corpus output without burning
fresh API tokens.
Replayed the 2026-05-08-rank4-gpt55 corpus (104 GPT-5.5 dashboard
outputs, 95 applied) through the new detectTextBgContrast and got
41 hits — 43% of designs flagged. Sampling showed almost all of them
were industry-standard Tailwind palettes used as intentional tertiary
text:
- #94A3B8 (slate-400) caption on #FFFFFF, ratio 2.56 ← Linear/Vercel/Notion
- #2563EB (blue-600) chip on #DBEAFE, ratio 4.24 ← shadcn/ui tag pattern
- #10B981 (emerald-500) delta on #FFFFFF, ratio 2.54 ← stat-positive pattern
- #64748B (slate-500) row text on #F1F5F9, ratio 4.34 ← muted-row pattern
WCAG-AA 4.5:1 is a compliance threshold, not a design-diagnosis
threshold. The user-reported pain point is "white-on-cream" (1.10:1)
and "white-on-white" (1.0:1) — disasters that read as obviously broken
to anyone, not borderline-WCAG cases that production designers ship
on purpose.
Drop default normalThreshold to 2.5 and largeThreshold to 2.0. Open
both as opts so callers needing a stricter audit (e.g. compliance
report) can bring back WCAG-AA without re-implementing the walk.
Replay confirms the new thresholds:
- 41 hits → 6 hits (signal-to-noise from 50% to 0% on the sample)
- All 6 remaining are true positives:
* 3 × slate-400 on slate-100 (caption color used on a non-white
bg — designer mis-paired the palette)
* 3 × white initial on amber-500 avatar (the readability gap the
industry routinely ignores; legitimately worth flagging)
Codex review (a47ef892f72a2d315) confirmed the direction, the
specific numeric pair (2.5 not 3.0 — 3.0 still hits slate-400 at 2.56),
parameterization over a mode-flag, and keeping severity at info-only.
Side-quest: scripts/ab-corpus/replay-detectors.ts +
inspect-contrast-hits.ts ship with this commit so the next detector
calibration doesn't have to rebuild the harness from scratch.
GAP-1 (fontWeight) + GAP-2 (empty-vars early-exit) are fixed in
52f9549b — replace stale knownGap() entries with positive assertions
that lock in the new resolver behavior. Smoke test is now full
regression coverage: 28/28 PASS, 0 known gaps.
Validates full pipeline: createEmptyDocument → applySemanticPalette →
buildHeadingV1/buildSettingRowV1/etc. (theme:'system') → resolveNodeForCanvas
Light/Dark. Surfaces 2 known gaps in the resolver (fontWeight not resolved;
DEFAULT_PALETTE_FALLBACK unreachable for empty-vars docs).
The 4 extra single-value tokens (color-accent-dark, color-info-surface,
color-warning-text-strong, color-danger-text-strong) introduced in P1.1.6
violated spec §3.1 / §7.4 — those hex were INTENDED to merge into existing
tokens with ≤ 5% accepted color drift, not become new tokens.
Replaced with MERGE_MAP in measure-v0-hex-coverage.ts that tracks the 4
near-shade redirections (#1D4ED8→color-accent, #EFF6FF→color-info-bg,
#B45309→color-warning-text, #B91C1C→color-danger-text). Cover rate
calculation now reports direct + merge breakdown.
Final palette token count: 56 (28 color + 18 type + 2 letterSpacing +
5 spacing + 3 radius). Cover rate: 28 direct + 4 merge = 32/32 = 100.0%.
Codex stop-time review caught the previous commit (113bd55a) message
overstating apply.ts's behavior — I claimed "scripts/ab-corpus/apply
both catch per-shape and keep running the remaining tags" but the
loop at line 55 had no inner try/catch. A single throw from any
handleElementToolCall (e.g. the heading invalid-level reject 113bd55a
just added) would bubble up through the outer try at line 40 and
return early, dropping every remaining tag in a composite batch on
the floor — gpt-5.4's 13-tag team-people-page response would lose
tags 12-13 instead of just tag 12.
Wraps each handleElementToolCall in its own try/catch + accumulates
failures into a per-shape list. ELEMENT_TOOL_NAMES miss is also a
push-and-continue (was a return). When `failures.length > 0` we
return ok:false with a message listing every failed tag, AND the
partial PenDocument that DID land — so M3 (role coverage) can still
score the 11 tags that worked. M1 stays strict (any failure → false).
Mirrors apps/web/src/services/ai/element-tools-dispatcher::
dispatchElementToolCalls's "collect-errors-keep-going" semantics —
production already worked this way; ab-corpus now does too.
3772 vitest pass, format clean, tsc silent. Existing dry-run + live
sweeps exercise the path; a focused apply.ts unit test would need
pen-mcp setup that the harness's existing build-prompt test sidesteps,
so leaving that as a followup.
Codex stop-time review caught that the Phase 2 domain filter
corrupted output when two adjacent kept blocks abutted: the regex's
trailing `\s*` consumed every character of whitespace BEFORE the next
match's leading `\s*`, so block N's body ended on its closing ``` and
block N+1's body started on its `### heading` with no separator at
all. Output: ` ```### Audit / activity feed` on a single line — fence
left unclosed, heading swallowed.
Replace the greedy `\s*` on both sides of each marker with a literal
`\n`, so the regex only consumes the single newline immediately
adjoining the comment. Surrounding blank lines stay in the
surrounding text where they belong, separating adjacent kept blocks
naturally.
Adds a regression guard test that searches the output for ` ```###`
(closing fence directly followed by a level-3 heading on the same
line). Codex effectively asked for it.
3767 vitest pass, format clean. Side effect: dropped blocks now
leave their surrounding blank lines intact, so dropped-block savings
shrink by ~3-5 chars per block; total Phase 2 savings stay within
the 500-char floor the existing test asserts.
ab-v3 / ab-v4 showed Phase 1A (cookbook strip on obvious difficulty)
shaved ~4.4k tokens off T-obvious. Phase 2 adds a per-category gate
that strips cookbook recipes whose domain doesn't match the prompt's
category — mobile briefs don't see dashboard recipes, dashboard
briefs don't see mobile / landing recipes, etc.
Mechanism: HTML comment block markers in elements.md
(`<!-- @domain:dashboard --> ... <!-- /@domain -->`) plus a
stripNonMatchingDomains() pass in buildSystemPrompt that drops blocks
whose tag list doesn't include the active category. Untagged content
is "general" and stays in every variant — the safe default.
Tagged 7 single-domain cookbook recipes:
- dashboard: Team / members list, Audit / activity feed, Faceted
search filter sidebar, Dashboard KPI strip
- landing: Pricing section
- mobile: Onboarding "How it works", Support chat thread
Cross-domain recipes (Login, Signup, Settings page, OTP, Empty
inbox) stay untagged so they load for every category. Decision tree
+ PREFER list also untagged today; the per-tool annotations there
would be a much larger judgment pass for marginal additional savings.
Token measurements (chars / 4 estimate):
full mobile dashboard landing
- T + composite 19.0k 17.9k 18.4k 17.7k
(-1.1k) (-0.6k) (-1.3k)
- T + obvious 14.6k 13.5k 14.0k 13.3k
(-1.1k) (-0.6k) (-1.3k)
Modest absolute savings — Phase 2 only filters cookbook RECIPES (in
elements.md), and most cookbook content is in elements-cookbook.md
which Phase 1A already strips on obvious. To hit the 6-8k T target
we still need decision-tree compression or PREFER-list trim, but
both are lossier than this gate. Phase 3 candidates noted in the
ab-v4 results doc.
real-model.ts plumbs call.prompt.category through to buildSystemPrompt.
3766 vitest pass (+6 category filter tests including a 500-char
floor regression guard that the filter actually shaves bytes).
ab-v4 raw output capture on dashboard-search-filters-composite shows
minimax-m2.7 emitting <think>...</think> + 4 op_tool tags that fit
inside the 4096 default — its measured completion-token average for
this run was 697, well under the cap. So thinking-budget truncation
is NOT the actual root cause of minimax's lower multi-tool hit rate
(25% vs gpt+deepseek 50%); the real issues are instruction-following
(mixed Strategy A + B despite the explicit forbidance, invented
"canvas" parent_id placeholder).
Still doubling the cap defensively: composite multi-tool outputs can
chain 12-13 op_tool tags + thinking, and "fit easy" today doesn't
mean "fits headroom-free on a longer brief tomorrow." The bump is
free on the happy path (provider stops generating when done, doesn't
bill unused headroom) and only ever helps when the model would
otherwise hit a real ceiling.
Real follow-up for minimax: instruction compliance — the no-mix rule
needs to land harder than a single trailing sentence. Probably wants
the rule moved to top-of-prompt + a few-shot bad-example contrast.
Out of scope here; tracked under Phase 2 prompt design.
ab-v4's dashboard-search-filters-composite garbaged on gpt-5.4 with
"codex timed out after 300000ms" — Codex's own agent framing
(~20k tokens) plus our 18.9k composite system prompt plus thinking
budget plus a 13-tool composite output is enough to blow the old
5-minute cap. The same model produced 13 chained op_tool tags on
team-people-page-composite within the window, so we know it's
generation latency under heavy briefs, not a hung CLI.
Doubles the default to 600000ms and adds AB_CORPUS_CODEX_TIMEOUT_MS
to override either way (lower it to surface slowness as a hard fail
when iterating, raise it for one-off long-form runs). Pairs with
the existing AB_CORPUS_CALL_TIMEOUT_MS for openai-compat clients.
Doesn't fix the deepseek empty-content failure on the same prompt
(already covered by retries=2 + exp backoff) or the minimax mix
+ invented "canvas" parent_id (model-side instruction skip; out of
scope for this commit).
The earlier T instruction split tool selection into PRIMARY / COMPOSITE
(element tools) and FALLBACK (batch_design). FALLBACK was scoped
"when no element tool fits a given component shape" — implying you
could mix per-component. But output-parser.ts silently drops every
batch_design tag whenever any element call is also present (filter
ELEMENT_TOOL_NAME_RE then return). A mixed response
<op_tool>{"name":"add_section_header_v0",...}</op_tool>
<op_tool>{"name":"batch_design", ...scaffolding...}</op_tool>
would lose the batch_design half and only run the element call —
the brief is half-applied without anyone noticing. Codex stop-time
review caught the mismatch.
Reframes T as a binary choice:
- STRATEGY A: every component fits an add_*_v0 tool — emit one tag
per component
- STRATEGY B: at least one component needs batch_design — emit a
SINGLE batch_design covering the whole brief
Plus an explicit "Do not mix Strategy A and Strategy B" guard line.
Tests assert both strategy markers are present in every T variant
and the dropped per-component-fallback phrase is gone.
The elements.md cookbook still teaches batch_design + element tool
composition for real MCP multi-round usage; the corpus T arm is
single-shot so the trailing T_TOOL_CALL_INSTRUCTIONS overrides via
"last instruction wins".
Codex stop-time review caught a contradiction: elements.md taught
"emit one <op_tool> per component" while T_TOOL_CALL_INSTRUCTIONS
explicitly forbade it ("Respond with one tag, nothing else" + "Do
not combine multiple tags"). Live models were probably reading the
later, more authoritative trailing instructions and dropping back to
batch_design — which matches ab-v3's 0/25 composite-T multi-tool
runs even though the cookbook had recipes.
Rewrites the T output-format block to:
- declare "one or more <op_tool> tags" up front
- add an explicit COMPOSITE clause showing chained tags for
multi-component briefs (settings panel / team list / audit feed /
onboarding) with a 3-tag worked example
- keep the FALLBACK clause for batch_design when no element tool fits
- keep the "no prose between tags" rule
Tests now assert the multi-tool marker is present in every T variant
AND the forbidding phrases ("one tag, nothing else", "Do not combine
multiple tags") are GONE — the regression guard Codex effectively
asked for. Also pulls in an oxfmt auto-format on the measurement
helper that landed unformatted in a3c5bf2d.
One-shot diagnostic that prints char + token-estimate sizes for every
(variant, difficulty) combination of buildSystemPrompt. Used to verify
the elements-cookbook diet is shaving the bytes the test floor
predicts (>10kb) before running an actual sweep. Confirms today's
numbers: B 15kb, T-obvious 57kb, T-composite 75kb — diet saves ~17kb
(~4.4k tokens) on every obvious prompt. Phase 2 (per-domain split)
still needed to hit the 6-8k target.
ab-v3 left T prompts at ~22k tokens (vs ~6-7k for B). The 18kb
elements-cookbook teaches per-tool arg shapes, which is what
composite multi-tool chains genuinely need; single-tool obvious
prompts can route correctly from just the decision tree + PREFER
list alone.
buildSystemPrompt now takes opts.difficulty. T+obvious strips the
cookbook (saves ~18kb on the 47/52 ab-v3 obvious prompts);
T+composite, T+optional, and undefined keep both halves. B variant
unchanged — still strips both, so the A/B comparison stays clean.
Verified by 8 new build-prompt.test.ts cases including a 10kb floor
on the obvious-vs-composite delta. ab-v4 will measure whether the
diet hurts arg compliance on weak models; per-domain split is the
Phase 2 fallback if obvious-T garbage rate creeps up.
ab-v3 left 36 ark empty + 15 timeout + 4 429 + 6 deepseek empty + 4
minimax timeout AFTER the existing retries=1 fired 88 times. Linear
backoff 250ms*(attempt+1) was too tight when stepping up to
retries=2 (500ms then 750ms isn't a typical Ark recovery window).
Switches to exponential 250ms*4^attempt — 250 / 1000 / 4000ms
spacing — and bumps ark + deepseek to retries=2. minimax opts in to
retries=1 (its 4 errors were wall-clock timeouts, not the model's
<think> truncation that retry can't fix anyway).
10 existing retry tests still pass; the retries=2 case now sleeps
1.25s instead of 0.75s, still well under vitest's default timeout.
Codex stop-time review flagged the new ab-v3 composite cookbook
recipes calling add_section_header_v0 with a `subtitle` arg the tool
doesn't accept (silently dropped at runtime today, but teaches live
models to emit invalid shapes). The same bug was in the dry-run stub
fixtures.
Split each header into add_section_header_v0(title) +
add_body_text_v0(content) — semantically what the ab-v3 briefs ask
for, and reinforces the multi-tool chaining the cookbook now teaches.
Also fills in the missing required `number` arg on the onboarding
recipe's final completed step card (schema requires it even when
completed=true renders a check instead of the number).
Codex stop-hook caught: ab-v3 introduced composite-difficulty prompts
that *expect* multi-tool emit (e.g. 5× member_row + 1× invite_row
for a team page), but `ParsedOutput.tool_call` was a single
{name, arguments} so the parser silently dropped every call after
the first. apply.ts only invoked one tool, M3 min_roles couldn't
pass on legitimately-routed multi-tool runs, and byTool stats
under-counted. The composite routing 'multi-tool' bucket was
correctly assigned in classifyRouting, but downstream the pipeline
behaved as if the model emitted a single call.
This commit replaces `kind: 'tool_call'` with
`kind: 'tool_calls'` (NON-EMPTY list) across every consumer:
- types.ts: ParsedOutput tagged union; new ParsedOpToolCall.
ScoreRow.toolName → toolNames: string[].
- output-parser.ts: collects ALL element-tool tags in emit order;
unknown-tool path also surfaces as single-element tool_calls so
routing keeps the same wrong-tool semantics.
- score-run.ts: classifyRouting uses Array.includes for obvious
prompts (right-tool when ANY emitted call matches expected_tool —
over-production isn't a routing miss). Composite stays multi-tool
on any non-empty list.
- aggregate.ts byTool: tallies EVERY name in toolNames, so a
composite row that emits 6× add_activity_log_v0 + 1×
add_section_header_v0 contributes 6+1 = 7 invocations across two
tools (with row-level m1_legal applied to both buckets — apply is
all-or-nothing).
- apply.ts: loops over parsed.calls and invokes
handleElementToolCall in emit order. Any single call failing
aborts the row (M1=false); we don't partial-apply.
- mock-llm.ts mockLlmParsed: collects all `<op_tool>` tags into the
list (composite-prompt mocks can carry multi-call raw strings).
- apps/web design-parser.tryParseElementToolOutput: maps tool_calls
→ its single-shape DesignOutputShape contract using the FIRST
call (the multi-tag path `tryParseAllElementToolOutputs` was
already correct).
Tests: 3746 → 3750 vitest. New cases:
- output-parser: surfaces ALL element-tool tags in emit order with
intermixed batch_design scaffolds dropped (3 element calls from
5 tags).
- score-run: right-tool when expected appears alongside extras;
composite multi-call captures every name in toolNames.
- aggregate: 6× activity_log + 1× section_header → byTool reports
6 and 1 invocations respectively.
dry-run on ab-v3 produces a 208-row report; tsc + format clean.
ab-v3 succeeds ab-v1 (frozen 2026-04-28). Carries forward all 40
v1 obvious yaml files unchanged so the v1↔v3 overlap stays
comparable, then layers in two new dimensions.
**1. Token cost.** All clients (openai-compat, ark, bailian,
deepseek, minimax, codex-cli, stub-model) now return a
`ChatCallResult { content, usage }` instead of bare string.
Provider usage stats (`prompt_tokens` / `completion_tokens`) plumb
through realModelCall → run.ts → scoreRun → ScoreRow.{prompt,completion}Tokens.
aggregate adds avgPromptTokens{Baseline,Treatment} +
avgCompletionTokens{Baseline,Treatment} per ModelSummary.
write-report emits a new "Token cost" table with Δ columns so
narrow-tools-saves-tokens (the ab-v2 hypothesis) is measurable.
avgUsage skips rows with 0/0 usage so codex-cli (CLI doesn't
surface tokens) and harness errors don't deflate the average to
near-zero — they show '—' instead.
**2. Composite difficulty.** New 'composite' value alongside
obvious / optional. Composite prompts express multi-tool intents
where no single expected_tool_if_any applies. classifyRouting
routes composite-treatment runs into multi-tool / fallback /
garbage (3-bucket sum to 1, distinct from obvious's 4-bucket
right/wrong/fallback/garbage). aggregate adds m6_multi_tool +
m6_fallback + m6_garbage; write-report emits a "Composite routing"
table that gracefully degrades to a placeholder when no composite
yaml exists yet.
Harness side: scripts/ab-corpus/run.ts accepts --corpus ab-v3
(enum + parseArgs guard); dry-run on the v1-mirror corpus produces
a 160-row report including populated token table.
Tests: 4 new aggregate cases (composite, token avg with skip-zero,
NaN-when-no-data) + 4 new score-run cases (composite routing
multi-tool/fallback/garbage/baseline-n/a) + 2 new score-run cases
(usage plumbing) + 2 new openai-compat cases (usage parsing,
missing-usage fallback). Existing 5 retry tests updated for new
return shape. 3727 → 3740 vitest tests, all green; tsc + format
clean.
Token-cost docs and composite docs go straight into types.ts /
score-run.ts / aggregate.ts JSDoc — keeps the contract close to
the code that owns it.
ab-v2 (2026-04-28) saw kimi-k2.6 garbage rate hit 17.5% — every
failure was Ark returning empty `choices[0].message.content` or
hitting the 120s wall clock, not a model-quality issue (the model
itself routed to the right element tool 80% of the time when it
did respond). Same pattern at 7.5% on deepseek-v4-pro.
Adds optional `retries` to `callOpenAICompat` with a transient-error
allowlist: empty content, abort/timeout, HTTP 5xx, HTTP 429. Linear
250ms × (attempt+1) backoff. HTTP 4xx other than 429 stays fatal so
auth/bad-request failures don't burn retry budget.
Wires `retries: 1` through clients/ark.ts and clients/deepseek.ts.
MiniMax + Bailian + Codex stay untouched — their ab-v2 failures
were model-quality (DSL escape errors, output truncation), where
retry wastes a call without changing the outcome.
Adds an 8-case fixture in scripts/ab-corpus/clients/__tests__/ that
mocks fetch to verify: first-try success, empty-then-success,
5xx-then-success, 429-then-success, 401 fatal, retries-default-zero,
retries exhausted, and retries=2 (3 attempts total). Extends the
apps/web vitest include glob to pick up scripts/**/__tests__/.
DeepSeek wasn't in the harness; A/B v2 needed it for the 5-model run.
api.deepseek.com is OpenAI-compatible, so the client mirrors the
minimax pattern (callOpenAICompat with DEEPSEEK_API_KEY env, override-
able DEEPSEEK_BASE_URL).
Router: /^deepseek/i routes to the new client. Aliases `deepseek` /
`deepseek-pro` resolve to `deepseek-v4-pro` (current flagship per
docs); `deepseek-v4-flash` and the deprecated `deepseek-chat` /
`deepseek-reasoner` ids pass through verbatim until the 2026-07-24
sunset upstream.
Verified live with mobile-bio-textarea smoke (B emits batch_design,
T emits add_textarea_v0). Full ab-v2 results across all 5 models in
docs/notes/2026-04-28-ab-v2-results.md.
Trimming the Minimal usage block out of elements.md (65c31832) lost
arg-shape templates that the A/B harness depends on. Text-only LLMs
in the treatment arm see only the markdown skill content — no MCP
tools/list, no published inputSchema — so without the per-tool
example payloads they have to guess argument names and break M1.
Restore the full block as a sibling skill `elements-cookbook` (same
hasMcpTools flag, slightly later priority so it loads alongside
elements). Wire it through buildFullPrompt + the
get_design_prompt(section='elements-cookbook') section map. Update
the A/B harness to also strip the cookbook body when building the
baseline prompt — leaving it in B would leak tool names + arg shapes
back into the no-tools variant and re-bias the comparison.
Both files now under the 800-line per-file ceiling.
origin's v0.8.0 had cherry-picks of the v0.7.5 deepseek/image-search
fixes (a727632a, a5952bc8) overlapping local 2073cf5b / 04f4fbc1, plus
new commits (model-selector ark-coding deepseek-v4-pro/flash IDs that
ARK rejects, fetch error.cause unwrap, CI agent-native build, op
export docs cleanup, main merge). Resolved the ark-coding list in
favor of HEAD's deepseek-v3.2 entry (only model ARK Coding Plan
actually supports — see openpencil-docs note).
The Zig NAPI provisioner had a silent failure mode that affected any
matrix entry without a matching ZSeven-W/agent prebuilt: the source-
build fallback dropped `agent_napi.node` at `zig-out/napi/...`, but
electron-builder only ships `packages/agent-native/napi/`. The addon
was therefore absent from the produced .exe / .dmg / .AppImage, and
every chat call died at the dynamic `@zseven-w/agent-native` import.
- Drop the prebuilt-download path; always build from source on the
runner (mlugg/setup-zig is already provisioned for every workflow)
- Always copy the built binary into `napi/agent_napi.node` so
electron-builder packages it
- Honor `ZIG_TARGET` to cross-compile (mac-x64 on arm64 runners now
produces an x86_64 binary instead of a wrong-arch arm64 one)
- Add `OPENPENCIL_REQUIRE_AGENT_NATIVE=1` strict mode plus a
dedicated "Verify agent-native binary" step in build-electron.yml
so missing binaries fail the workflow loudly
- Add `OPENPENCIL_SKIP_AGENT_NATIVE=1` for publish-cli.yml, which
never ships the addon and shouldn't pay for the build
`/models` now returns only deepseek-v4-pro and deepseek-v4-flash;
deepseek-chat / deepseek-reasoner sunset 2026-07-24 and the
deepseek-v3.2 hard-coded in the ark-coding fallback list never
existed. Both v4 models default to thinking enabled and the API
toggles via `{"thinking":{"type":"disabled"}}` — keep
`thinkingMode: 'disabled'` so the app's fast/non-thinking default
stays intact (server reasoning paths honor it; the Zig openai-compat
path doesn't emit the toggle yet, so calls through that path still
get provider-default thinking until it's wired). v4-pro promoted to
full tier; legacy aliases pinned to an exact RegExp so future
deepseek-* variants don't inherit a forced disabled mode.
Bandaid for the unwired toggle: v4-pro gets `timeoutMultiplier: 2`
because its default-on reasoning blows past the orchestrator's
planning timeout on long system prompts (observed in dev: planning
phase falls back, sub-agent then succeeds — UX degraded but
functional). Drop the multiplier once the Zig path actually sends
`thinking:{type:disabled}`.
Don't add a BUILTIN_MODEL_LISTS.deepseek entry — DeepSeek exposes
/v1/models, so let `fetchProviderModels` pull the live catalog
through `/api/ai/provider-models` instead of pinning a snapshot
(the ark-coding `deepseek-v3.2` ghost above shows what those
snapshots drift into).
The first 12-prompt × 2-model × 2-variant sweep (48 API calls)
ran 29 minutes before I killed it. A Kimi call on the 10th
prompt hung indefinitely — no client-side timeout — and the
harness writes scores.jsonl + report.md ONLY at the end, so
partial progress was unrecoverable. Lost 9/12 completed prompts
because the aggregate step never ran.
Two fixes:
1. openai-compat.ts: AbortController with default 120s timeout
(overridable via AB_CORPUS_CALL_TIMEOUT_MS env). When a call
exceeds the budget, the harness catches the abort, records it
as __HARNESS_ERROR__ (routing=garbage, M1=false), and moves on.
Verified by dialing the timeout to 60s — GLM-5.1's first call
took >60s, got aborted cleanly, run continued to completion
instead of hanging.
2. run.ts: append each ScoreRow to scores.jsonl immediately after
scoring. Truncate at start (so re-runs overwrite). Lost-work
window now bounded to "the currently-executing API call," not
"everything since the run started." report.md and report.json
still write once at the end (aggregate needs the full set) but
scores.jsonl alone is enough for any partial-run analysis.
Post-hardening validation (live 方舟 CP runs):
- mobile-upload-dropzone → add_upload_dropzone_v0 ✓ right-tool
- dashboard-dark-modal → add_modal_shell_v1 (theme=dark) ✓
Second one is the first end-to-end proof that the v1 theme-aware
tool family routes correctly with a real LLM — GLM-5.1 inferred
\`theme: "dark"\` from the natural-language prompt.
Two independent changes rolled together since they both serve the
same goal — "can real LLMs actually route to the tools we shipped
today?":
1. scripts/ab-corpus/run.ts gains a `--corpus` flag (ab-v0 |
ab-v1, default ab-v0 for back-compat). The harness was
hardcoded to ab-v0 — adding ab-v1 prompts was worthless
without a way to run them. Validated 17 prompts × 2 models ×
2 variants during a live 方舟-CP run.
2. 5 new ab-v1 prompts cover the 2026-04-24 tool batch:
- mobile-upload-dropzone.yaml → add_upload_dropzone_v0
- mobile-otp-verification.yaml → add_otp_input_v0
- mobile-file-attachment.yaml → add_attachment_row_v0
- mobile-chat-message.yaml → add_chat_bubble_v0
- dashboard-dark-modal.yaml → add_modal_shell_v1
corpus-loader.test.ts bumps its count assertion 12 → 17 and
extends the tool-coverage set. Also loosens the regex to accept
`_v\d+$` (was `_v0$`) so add_modal_shell_v1 passes. No other
test file needed changes — the existing registry-parity and
mock-llm tests already use `_v\d+$` or the registry directly.
.gitignore gains:
- .playwright-mcp/ (MCP Playwright session artifacts)
- editor-*.png (local verification screenshots)
- scripts/ab-corpus/runs/ (live-run outputs / reports)
None of those belong in version control — they're artifacts
from local verification runs.
The Ark router regex was `/^kimi-k2\.6/i` — required the `k`
prefix. But mapKimiArkId's alias list accepted both `kimi-k2.6`
AND `kimi-2.6` (no-prefix form). Result: `kimi-2.6` failed the
Ark regex, fell through to the generic `/^kimi/i` branch, got
routed to Bailian — which doesn't host K2.6. Bailian would
return HTTP 400 "model not supported" with no hint that the id
belonged on Ark.
Fix: Ark router regex now `/^kimi-k?2\.6(-ark)?$/i` — optional
`k` prefix + optional `-ark` suffix, anchored at end to prevent
accidentally matching a hypothetical later version. mapKimiArkId
normalizes all four accepted aliases (kimi-k2.6, kimi-2.6,
kimi-k2.6-ark, kimi-2.6-ark) to the canonical on-Ark id.
Same latent bug fixed on the glm-5.1 route: regex `/^glm-5\.1/i`
would prefix-match a hypothetical `glm-5.10` and wrongly route it
to Ark. Tightened to `/^glm-5\.1(-coding|-ark)?$/i` with the same
anchored-end + suffix-allowlist pattern.
Caught by Codex stop-hook review during 2026-04-22 session.
Volcengine 方舟 (Ark) added GLM-5.1 and Kimi-K2.6 to its coding
plan on 2026-04-22 — single ARK_CODING_KEY covers both. Harness
now prefers this route over the previous paths:
- glm-5.1 was routed to clients/glm.ts (GLM official CP via
open.bigmodel.cn with GLM_OFFICIAL_CODING_KEY). Now routed to
new clients/ark.ts. The old glm.ts file is kept on disk for
historical comparison but not wired into the default router —
callers who want to A/B the old GLM-official path vs. new Ark
path can import callGlm directly.
- kimi-k2.6 is new — added as a dedicated router branch above
the kimi-k2.5 (bailian) branch so the version-specific match
lands on Ark.
Old kimi-k2.5 continues to route through clients/bailian.ts
(DashScope aggregator) for continuity with earlier A/B runs.
Key management (unchanged from the harness convention):
- ARK_CODING_KEY — Volcengine 方舟 CP UUID format key. Export
in shell before running --live; never committed.
- Existing MINIMAX_API_KEY / GLM_OFFICIAL_CODING_KEY /
DASHSCOPE_BAILIAN_CODING_KEY all still honored for their
respective routes.
Throw message updated so missing-key errors surface the correct
env var for each route.
User hit at dev-server startup:
Module "node:fs" has been externalized for browser compatibility.
Cannot access "node:fs.readdirSync" in client code.
Chain: apps/web's design-parser.ts imports `parseModelOutput` from
`@zseven-w/pen-ai-skills`; main barrel re-exports everything from
`./corpus`; `./corpus/index.ts` re-exports `loadCorpus` which imports
`node:fs`. Vite pulls the whole graph into the client bundle → crash
on the first browser-side module evaluation.
Fix: remove `loadCorpus` from `./corpus/index.ts`. The barrel now
only exposes pure-string helpers (parser, scorer, aggregator, types)
— all browser-safe. `loadCorpus` stays in `corpus-loader.ts` but
Node-only consumers (`scripts/ab-corpus/run.ts`) import it directly
via a relative path. Package.json only declares the main entry in
`exports`, so sub-path imports via the package name fail at runtime
(pkg runs under Bun for the harness) — relative file path avoids
that gate.
Verification:
- `bun scripts/ab-corpus/run.ts --only X --dry-run` still runs end
to end
- tsc --noEmit exit 0
- Full test suite 1866/1866
Browser-side verification (user): restart Vite dev server — the
design-parser import no longer pulls node:fs through the barrel.
