ci: read the Go limits page by what it carries, and keep the catalog refresh going when it breaks by viztor · Pull Request #167 · viztor/dsh-opencode-patch

viztor · 2026-11-12T08:17:23Z

Why

Every nightly catalog freshness run since 2026-11-07 has failed in the Go limits generator:

Error: go-limits: no <TabItem label="Go"> section in the doc
    at parseGoLimitsDoc (src/go-limits.ts:143:13)
    at scripts/generate-go-limits.ts:133:30

Cause, and what the live page carries now

The reader handed any 2xx body to the markdown parser. go.md answered 404 on 2026-10-12, but since 11-07 it answers 200. Measured by this PR's CI on 2026-11-12:

Both https://opencode.ai/docs/go.md and the old https://opencode.ai/v2/docs/console/go/ now redirect to https://opencode.ai/docs/console/go/. It is an HTML page of about 98,000 characters.
The page carries three monthly-limit tables, one per Starlight tab:

Tab Model rows First row

Go 40 GLM-5.3-Flash, $60

Go Plus 40 GLM-5.3-Flash, $180

Go Teams 40 GLM-5.3-Flash, $65
The third tab means a reader that takes the first two tables by position can read the wrong pair. An earlier commit on this branch did exactly that, before the review caught it.

What changed

The reader is in scripts/go-limits-html.ts, which is pure and tested.

docFrom(answer)

Reads a 2xx answer by what its body carries. The status alone is never trusted, and an answer outside 2xx is not read.
It tries the tabbed markdown first, when it carries both tabs, then the HTML page.
Either is taken only when parseGoLimitsDoc reads it without an error. Anything else falls through to the next source.

Table choice by plan label (monthlyLimitTables, chooseTables)

Each table's label comes from its Starlight tab: the role="tabpanel" ancestor's aria-labelledby, matched to the role="tab" element's text. A panel's own label or the nearest heading is the fallback.
Go is the one table labelled Go, and Go Plus the one labelled Go Plus. Other tabs, such as Go Teams, are ignored.
A missing or duplicated plan label is refused. Only two tables with no plan or tab label keep the order the page had when it was first measured.

tableRows

Reads a cell as a browser shows it: <br> becomes a space, character references are decoded, and whitespace collapses.

firstDoc(GO_LIMITS_SOURCES, read)

The walk asks the canonical page, then go.md, and takes the first answer the parser reads.

noDocMessage

When no source works, it names every table: label, model rows, first row. It also gives the parser's or the chooser's reason.
On success, the generator's source: line names which tab it read as Go and as Go Plus, and which it ignored.

Elsewhere

scripts/generate-go-limits.ts is glue only. --from takes a saved page in either shape.
AGENTS.md, docs/testing.md and the GO_LIMITS_DOC_URL comment in src/go-limits.ts match the code.
main (ci: propose the catalog refresh even when the limits generator fails #168, test: a harness that is the real host #110, fix: disabling or reloading the plugin restores the host methods it wrapped #112) is merged in. The nightly-step change this PR used to carry is on main through ci: propose the catalog refresh even when the limits generator fails #168, so this diff is the reader only.

Review

Four adversarial rounds. Each finding went to two skeptics, and only upheld findings were fixed.

Upheld and fixed:

acceptance weaker than the parser;
a third table read by position;
the source walk untested inside the script;
"never by its status" contradicting the code;
unasserted noDocMessage branches;
a stale generator header;
multi-line cells;
undecoded references;
firstDoc's JSDoc;
a heading inside a hidden tab panel labelling a later table;
"no Go tab" read as the markdown tab;
a test asserting a constant against its own literal.

Declined by both skeptics:

catching fetch rejections in the walk;
two tables under contradicting non-plan headings, which are already refused unless both are unlabelled.

Verification

Red first: each behaviour change opened with a test that failed before it. Seven mutants of the label logic were each caught.
Gate on Node 26 at 89e7049:
- vp check: 0 errors, 18 warnings.
- Unit suite: 102 files, 1764 passed, 1 skipped.
- Mock suite: 52 passed.
- pnpm run build and scripts/check.ts: passed.
Offline round trip: an HTML page rebuilt from the shipped table reads back through --from as up to date (28 models). A scratch three-tab page reads Go and Go Plus from their tabs and ignores the third.

Live, on this PR's CI:

source: https://opencode.ai/docs/console/go/ (98163 characters), read by label: Go from table 1 (tab "Go", 40 model row(s), first "GLM-5.3-Flash" → "$60"); Go Plus from table 2 (tab "Go Plus", 40 model row(s), first "GLM-5.3-Flash" → "$180"); ignored table 3 (tab "Go Teams", 40 model row(s), first "GLM-5.3-Flash" → "$65")
go-limits-data.ts is STALE — run: pnpm run limits:shim -- --write

Why `catalog freshness` is red here

The reader is right. The table is out of date: the vendor now lists 40 models per plan, where the shipped src/go-limits-data.ts has 28, and the amounts moved.

This PR changes the limits generator, so under D23 it gates on that table. The table is regenerated, never hand-edited:

pnpm run limits:shim -- --write && pnpm exec vp fmt --write src/go-limits-data.ts

That needs opencode.ai, which the session that wrote this PR cannot reach. Run those two commands on this branch and commit the result, and the gate turns green.

Once this is on main, the nightly refresh regenerates the table on its own and proposes it.

🤖 Generated with Claude Code

https://claude.ai/code/session_0141fKp5k6hiaJcP5ovhGtSD

…refresh going when it breaks Every nightly refresh since 2026-11-07 failed in the limits generator with `no <TabItem label="Go"> section in the doc`, and because the two generators ran in one fail-fast step, no catalog refresh was proposed either, while the vendor added six models. The cause is in the reader, not the parser. The HTML path always emits the tab sections, so the error can only come from the first branch: `go.md` answered 200 again, with a body that has no tab sections, and the reader handed any 2xx body to the markdown parser. - `scripts/go-limits-html.ts` gains `docFrom`, which reads one answer by its body: the tabbed markdown as it is, an HTML page carrying both monthly-limit tables through `htmlToDoc`, anything else (or a status outside 2xx) not at all. `noDocMessage` names what every source answered (status, redirect, content type, size, which shape it carried), so the next layout change names itself in the log. - The generator asks the page the docs live on, then the old URL, and takes the first answer that carries the tables. `--from` accepts a saved page in either shape. - The nightly step keeps `src/go-limits-data.ts` as it was when the limits generator fails, still proposes the catalog refresh, and fails at the end, naming the cause. Red first: the five new cases in test/go-limits-html.test.ts failed on main. Offline round trip: an HTML page rebuilt from the shipped table reads back through `--from` as "up to date (28 models)". The live page is measured by this PR's own `limits:shim` gate, since the vendor host is not reachable from the session that wrote it. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_0141fKp5k6hiaJcP5ovhGtSD

viztor · 2026-11-12T08:20:06Z

catalog freshness is red on this PR, and only on limits:shim. It is the live measurement the description asked for, and the reader passes it:

source: https://opencode.ai/v2/docs/console/go/ (98282 characters)
go-limits-data.ts is STALE — run: pnpm run limits:shim -- --write

The fixed reader took the moved page and parsed it with no unreadable rows. The vendor has changed the allowances since the last refresh on 11-06, so the shipped src/go-limits-data.ts no longer matches. Under D23 this PR gates on that, because it changes the generator.

Why the fix is not pushed from here: the table must be regenerated, never hand-edited, and opencode.ai is outside this session's network policy. It needs pnpm run limits:shim -- --write run on this branch by anyone who can reach the page, followed by vp fmt --write src/go-limits-data.ts and a commit. I will do it as soon as the host is reachable here.

The nightly step's change is also on its own PR, #168. It touches no generator input, so it can merge without the table, and from then on the catalog refresh is proposed even while the limits generator fails.

Generated by Claude Code

…test the source walk reader-1, tests-2: docFrom took bodies the parser then threw on. The markdown arm now needs both the Go and the Go Plus tab; the HTML arm needs the allowance tables to each carry a model row of six cells (modelRows, shared with htmlToDoc). Tests for a Go-only body, empty tbodies, short rows, and one readable table beside an unreadable one. reader-2: htmlToDoc took the first two monthly-limit tables by position. It now throws unless there are exactly two, naming the count, and docFrom requires exactly two, so a third table falls through to the next source. The "fewer than two" test now reads "exactly 2"; a three-table case ($5/$60/$120) is added for both functions. reader-5, tests-1, rules-docs-2: the walk moved out of the import-time generator into scripts/go-limits-html.ts as firstDoc (injected reader, answering { doc, source, answer }; the script logs), with the ordered GO_LIMITS_SOURCES beside it. Tested with a fake reader: first accepted -> second never asked; first refused -> second read; none -> rejects with the whole noDocMessage, both sources in order, the 2026-11-07 go.md shape as the second; no sources -> the bare message. freshness-scope still answers "report" for catalog:shim. tests-3, rules-docs-4: "by what its body carries, never its status" now reads "a 2xx answer is read by what its body carries; the status alone is never trusted" in docs/testing.md, AGENTS.md, the docFrom JSDoc, the module doc and the generator header. tests-5: the full 404 line is asserted, with no "(redirected"; an empty content type reads "no content type"; HTML bodies report their table count and, when fewer have a model row, how many; a Go tab without Go Plus is named as such. rules-docs-5: the generator header says what the code does now (the moved page first, go.md second, read by content), with the history in the past tense; the future-work paragraph is gone. GO_LIMITS_DOC_URL's comment names it the old URL, asked second (src/go-limits.ts is an input of limits:shim only). Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_0141fKp5k6hiaJcP5ovhGtSD

Make this branch carry exactly #168's versions of the files the two pull requests share, so the nightly "propose the refresh" step and its tests are one change wherever they land. - .github/workflows/ci.yml: the nightly step now guards both generators the same way (each failure annotated where it happens, its file put back as committed, the other refresh still proposed, the step failing at the end), with the catalog restored before the limits generator starts. Byte-identical to #168. - CONTRIBUTING.md: the "one broken source" bullet takes #168's text. This branch had no other CONTRIBUTING.md edit, so nothing was merged by hand. Byte-identical to #168. - test/freshness-scope.test.ts: #168's tests that run the step itself in a scratch repository. Byte-identical to #168. No findings carried in this commit. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_0141fKp5k6hiaJcP5ovhGtSD

…multi-line cells reader1-1: docFrom accepted a body on its shape (both tab strings, or two allowance tables each with a six-cell row), which is weaker than what parseGoLimitsDoc reads, so the walk could stop at a source the parser then threw on and never ask the next one. Acceptance is now the parser's own verdict: a body is tried as the tabbed markdown (both tabs) and then as the converted HTML page (exactly two allowance tables), and a shape is taken only when parseGoLimitsDoc reads it without throwing. noDocMessage appends the parser's reason for each shape it was handed. The module, docFrom and generator headers and docs/testing.md say so. Tests: both tabs with no rows or a five-cell table; two tables whose rows end in Unlimited, $15 or a seventh empty cell; a page that quotes both tab strings and carries valid tables; the walk asking the second source when the first only looks readable; the parser's reasons in the error. reader1-2: tableRows now reads a <br> as a space and collapses whitespace inside a cell to one space, so a cell spanning lines no longer splits the emitted markdown row and drops the model without an unmapped entry. Tests: the cell reading, and a round trip through docFrom and the parser with a newline in an input cell and <br> plus newline in a label. ci1-1: carried byte-identical from the refresh-decouple branch: the nightly step warns, after a successful push, which files a partial refresh withdrew from an open refresh pull request; CONTRIBUTING.md says so; and test/freshness-scope.test.ts pins it with a pre-existing refresh branch. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_0141fKp5k6hiaJcP5ovhGtSD

…ghtly step's fixes reader2-2: tableRows stripped tags and never decoded character references, so `$15` was no dollar amount (a valid page refused), `$15 ` read as NaN, `Grok 4.7` keyed as `grok-nbsp-4.7` and `Q&A 2` as `q-amp-a-2`, and names shipped with a literal `>`. tableRows now decodes after stripping tags and before collapsing whitespace, in one pass (`&lt;` stays the text `<`): `& < > " '   ≤ ≥` and numeric `&#N;`/`&#xN;`, leaving an unknown name (looked up in a Map, so `&constructor;` is unknown) or a number that names no character as written. The `$15` case no longer stands for "no dollar amount"; new cases read entity-encoded amounts, key a label with ` `, `>`, `&` and `≤`, decode in one pass and leave unknown references alone. docs/testing.md's row says so. reader2-3: firstDoc's JSDoc said a later source is asked only when every earlier one carried neither shape; it is asked whenever the earlier ones answered nothing the parser reads (a status outside 2xx, neither shape, or a shape the parser rejected). Reworded. ci2-1, ci2-2, ci2-3: .github/workflows/ci.yml, CONTRIBUTING.md and test/freshness-scope.test.ts copied byte-identical from #168 (no inherited GIT_* reaches the scratch-repo tests; the "both current" line only when both regenerated; an open refresh pull request is asked for, not assumed, and a failed creation fails the step). Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_0141fKp5k6hiaJcP5ovhGtSD

…ble the page carries The 2026-11-12 CI run measured the page carrying three monthly-limit tables, behind both URLs, which now redirect to the canonical https://opencode.ai/docs/console/go/. A choice by position or by count cannot say which two are Go and Go Plus, so each table is now chosen by the label the page attaches to it. - scripts/go-limits-html.ts: monthlyLimitTables labels each allowance table from (a) its innermost role="tabpanel" ancestor's aria-labelledby, resolved to the element whose id attribute it names, (b) the panel's own data-label/aria-label (an unprocessed <starlight-tab-item data-label> counts as a panel), else (c) the nearest preceding h2-h4. Tags are scanned with attributes in any order and either quote, comments and script/style contents blanked first. chooseTables reads Go and Go Plus from the one table labelled each (case-insensitive), ignores other labelled tables, and refuses a missing or repeated plan label; with no plan or tab label, exactly two tables keep the measured order and any other count is refused. The Starlight markup is assumed, not measured, and the module header says so. - Logs: noDocMessage lists every allowance table with its label source and text, its model-row count and its first row, and why the tables were not chosen; FoundDoc.read names which table was read as Go and as Go Plus and which were ignored, and the generator's source: line prints it, with the redirect. - GO_LIMITS_DOC_MOVED_TO is the canonical https://opencode.ai/docs/console/go/; GO_LIMITS_SOURCES stays [that page, go.md]. - test/go-limits-html.test.ts: a three-tab Starlight page in a non-plan order, reordered single-quoted attributes with decoy ids, innermost and unprocessed panels, a tag inside a script or comment, heading labels, two unlabelled tables in order, three refused with each table named, a duplicated Go label, two tabs naming no plan, an empty labelled Go table refused by the parser, and the success description. - AGENTS.md, docs/testing.md, the generator header and src/go-limits.ts's URL comment describe label-based selection and the unmeasured markup. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_0141fKp5k6hiaJcP5ovhGtSD

… it, and name the markdown tabs as markdown - scripts/go-limits-html.ts: labelAt's heading fallback skips an h2-h4 that lies inside a tab panel not holding the table, so a table after a tab group is no longer labelled by the last panel's heading (which refused an unambiguous page, or with a "Plus" tab read the wrong table as Go Plus). - scripts/go-limits-html.ts: tabsCarried names the markdown shape ("no markdown Go <TabItem>" and its two siblings), so the error line no longer says "no Go tab" beside a table described as tab "Go". - test/go-limits-html.test.ts: three cases for the panel-scoped heading (refused-before page now read by label, wrong-table page now refused, a heading inside the table's own panel still labels it), the reworded phrases in every expected error line, and the Go/Go duplicate case asserting the markdown phrase beside the tab labels. - AGENTS.md and the module header: describe the choice as the code makes it (panel label step, panel-scoped headings, two tables with no plan label and no tab label kept in order, "rather than" in place of "never" by position). Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_0141fKp5k6hiaJcP5ovhGtSD

… literal The "asks the page the docs live on first" case compared a test-local constant with its own literal, so it passed whatever the reader exported. - test/go-limits-html.test.ts: import GO_LIMITS_DOC_MOVED_TO and assert it names the canonical page, in place of the self-comparing assertion. Under a mutation that points the constant back at the /v2/ URL, the old line passed and the new one fails. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_0141fKp5k6hiaJcP5ovhGtSD

…9n-go-limits Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_0141fKp5k6hiaJcP5ovhGtSD

…bctz9n-go-limits Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_0141fKp5k6hiaJcP5ovhGtSD

viztor · 2026-11-15T11:38:30Z

Merging with catalog freshness red, on purpose.

On c81a80c, with current main merged in, this PR's reader reads the vendor's live page: https://opencode.ai/docs/console/go/ (99,695 characters). It takes Go from tab "Go" (40 models, first GLM-5.3-Flash → $60) and Go Plus from tab "Go Plus" (40 models, $180), and ignores "Go Teams".

The only remaining error is that the committed src/go-limits-data.ts predates the vendor's change. Regenerating it needs network access to opencode.ai. The nightly refresh has that access, but it runs main's reader, which cannot read this page until this PR is on main.

So the order is: merge this, let tonight's nightly regenerate the table and propose it as a refresh PR, then port and merge that refresh. Every other check is green.

Generated by Claude Code

viztor mentioned this pull request Nov 12, 2026

ci: propose the catalog refresh even when the limits generator fails #168

Merged

claude added 8 commits November 12, 2026 10:51

Merge origin/main (#168, #110, #112) into claude/vigilant-mendel-bctz…

89e7049

…9n-go-limits Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_0141fKp5k6hiaJcP5ovhGtSD

This was referenced Nov 13, 2026

chore: refresh the generated catalog #159

Merged

ci: one setup, one unit run, and stricter scripts #173

Open

Merge origin/main (#113-#116, #171-#174) into claude/vigilant-mendel-…

c81a80c

…bctz9n-go-limits Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_0141fKp5k6hiaJcP5ovhGtSD

viztor merged commit 4b854ed into main Nov 15, 2026
8 of 9 checks passed

Navigation Menu

Repository navigation

ci: read the Go limits page by what it carries, and keep the catalog refresh going when it breaks - #167

ci: read the Go limits page by what it carries, and keep the catalog refresh going when it breaks#167
viztor merged 10 commits into
mainfrom
claude/vigilant-mendel-bctz9n-go-limits

viztor commented Nov 12, 2026 •

edited

Loading

Uh oh!

viztor commented Nov 12, 2026

Uh oh!

viztor commented Nov 15, 2026

Uh oh!

Uh oh!

Reviewers

Uh oh!

Assignees

Labels

Projects

Milestone

Development

Uh oh!

2 participants

Tab	Model rows	First row
`Go`	40	GLM-5.3-Flash, $60
`Go Plus`	40	GLM-5.3-Flash, $180
`Go Teams`	40	GLM-5.3-Flash, $65

Conversation

viztor commented Nov 12, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Why

Cause, and what the live page carries now

What changed

Review

Verification

Why catalog freshness is red here

Uh oh!

viztor commented Nov 12, 2026

Uh oh!

viztor commented Nov 15, 2026

Uh oh!

Uh oh!

Reviewers

Uh oh!

Assignees

Labels

Projects

Milestone

Development

Uh oh!

2 participants

viztor commented Nov 12, 2026 •

edited

Loading

Why `catalog freshness` is red here