docs: align README with current surfaces - #1180
Merged
localai-bot merged 2 commits intoAug 21, 2026
Merged
Conversation
Update registry, parser, and ABI claims from the current source. Add the gated MiniMax-Music3 server path to News and retire a superseded benchmark item from that section. FOLLOWING_AGENTS_PROTOCOL Assisted-by: Codex:gpt-5.4 [Codex]
localai-bot
added a commit
that referenced
this pull request
Aug 18, 2026
…h whole-file budget (#498) (#1245) `README.md` measured 29,993 characters against `MAX_README_CHARS = 30000` (`scripts/check-readme-structure.py:47`). Seven free. The next README edit reds CI, and two concurrent edits that each fit on their own merge into a file that fails. AGENTS.md Records forbids that shape in its own words, "Limit an entry, not a shared file", and gives the reason: a shared-file budget forces each addition to remove another entry, and merging two such edits cleanly is worse than conflicting because it applies both removals. ## This is the fourth budget of this shape, not a new argument | Budget | Retired by | |---|---| | per-class line limits | 2026-08-10, recorded in AGENTS.md | | `MAX_CHARS` in `check-now-current.py` + `STATUS_RATCHET` `chars` | `87308dea` under #364 | | `PageRules.max_chars` in `check-public-doc-tables.py` | #460 (PR #494) | | `MAX_README_CHARS` | this PR, under #498 | `87308dea` is the governing precedent and it is explicit about what survives: it kept "every local per-cell and per-paragraph cap" while removing the whole-file budget. This change mirrors it exactly. ## What changed `MAX_README_CHARS` and the branch that reads it are removed. The obligation relocates rather than disappearing: - `MAX_CELL_CHARS = 220` unchanged, one table cell is one author's entry - `MAX_PARAGRAPH_CHARS = 900` unchanged, one prose paragraph is the largest unit one author writes alone - required sections, ordering, em-dash, `STATUS_LINK` and `CONTRIBUTOR_LINK` unchanged, because they count defects rather than length and so cannot collide between concurrent pull requests **`README.md` is byte-identical at 29,993 characters.** This PR does not add or remove a single character of the landing page, so it cannot be confused with buying headroom. ### Rejected: a per-section budget Issue #498 offers "a per-section or per-paragraph budget". A per-`##`-section cap is rejected, and the spec records why: a section such as News is written by many pull requests, so capping it recreates the same lock at smaller granularity. The per-paragraph cap is the real entry cap and it already existed. ## Evidence Red-before, on the unmodified checker: ``` test_oversized_readme_passes_when_every_entry_is_small AssertionError: ['README is 69427 chars, over the 30000-char landing-page budget...'] != [] test_checker_declares_no_whole_file_budget Ran 23 tests -- FAILED (failures=2) ``` The suite moves 21 to 23 tests, so the new cases demonstrably run. The retired `test_oversized_readme_fails` is deleted because it pins the behavior being removed; the two new cases plus the entry-cap guard replace its coverage. Green after: `Ran 23 tests ... OK`, and `check-readme-structure.py` still passes on the shipped, unchanged README. Mutation, each applied to the committed tree and restored byte-for-byte (`sha256` asserted after every restore): | Mutation | Result | |---|---| | M1 reintroduce `MAX_README_CHARS` and its branch | 2 tests RED | | M2 `MAX_PARAGRAPH_CHARS` to 10,000,000 | 2 tests RED | | M3 `MAX_CELL_CHARS` to 10,000,000 | 2 tests RED | M2 and M3 are the ones that matter for this change: they prove removing the file budget did not quietly disable the entry budgets, which is the failure mode a "the tests still pass" report would otherwise hide. `test_checker_declares_no_whole_file_budget` asserts the constant is absent from the module rather than checking its value, so the gate fails if it returns under any threshold. Full `scripts/agent-preflight.sh`: all gates green. ## Also unblocks half of #1098 #1098 records that `README.md` cannot be corrected, and names two independent blockers. Blocker 1 was this cap: the LTX-2.5 matrix row could only land by deleting another architecture's row. That blocker is now gone, which also clears the way for PR #1180's 37-to-40 architecture-count correction. Blocker 2, `check-doc-checkpoint.py` refusing any README change that does not also touch a landing source, is untouched and still owns #1098. No issue-index row was appended here, because #1098 already has one and AGENTS.md forbids editing an existing row. Fixes #498. FOLLOWING_AGENTS_PROTOCOL Following-Agents-Protocol: true AI-Assisted: true Assisted-by: claude:claude-opus-5 [claude-code]
…ABI version main had already refreshed the architecture and tool-parser counts this branch was written to fix, so both README conflicts are resolved in main's favour: its News text is newer and carries an LTX-2.5 entry the branch predates, and its capability sentence names image, video, and audio input separately. What the branch still contributes is the C ABI line, and its value was stale too. include/vllm.h:329 defines VLLM_ABI_VERSION 23, so the branch's v21 is as wrong as main's v19; it is set to v23 here. The "36 functions" count is dropped rather than corrected, because a live count in a landing document goes stale on the next commit that adds a function. FOLLOWING_AGENTS_PROTOCOL Following-Agents-Protocol: true AI-Assisted: true Assisted-by: Claude:claude-opus-5 [Claude Code]
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
3 participants
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
What changed
include/vllm.h:329actually defines (#define VLLM_ABI_VERSION 23).it. A live count in a landing document is a drift lock: it is wrong again on
the next commit that adds an exported function, and nothing gates it.
its caveat, in the hardware table: "Qwen3.6-27B decode matches llama.cpp
Vulkan (4.36 vs 4.35, denominator SUPERSEDED, Every recorded llama.cpp floor was measured against a local-only fork carrying 65 of our own performance commits, and must be re-taken against the stock pin #1003)". A headline item
whose own denominator is superseded does not belong in News.
What this pull request no longer changes
It was opened on 18 August to raise the architecture count from 37 to 40, the
tool-parser counts from 36/40 to 38/42, and the ABI from v19 to v21.
mainhassince refreshed every count except the ABI, so those edits are gone from the
diff. The ABI value it proposed was stale as well — v21 against a header that
now reads 23 — which is the same failure the count edits illustrate.
Both README conflicts with
mainwere resolved inmain's favour: its Newstext is newer and carries an LTX-2.5 entry this branch predates, and its
capability sentence names image, video, and audio input separately rather than
collapsing them to "multimodal input".
Verification
check-readme-structure.py,check-supported-models.py,check-public-doc-tables.pyandcheck-model-checklist.pyall exit 0 on themerged tree. Net diff against
mainis +1/-4 inREADME.mdand nothing else.FOLLOWING_AGENTS_PROTOCOL
Following-Agents-Protocol: true
AI-Assisted: true
Assisted-by: Claude:claude-opus-5 [Claude Code]