mirror of
https://github.com/dtzp555-max/ocp.git
synced 2026-07-27 07:55:07 +00:00
fdd6dadf96
* feat(models): add Claude Opus 5 and repoint the `opus` alias to it Adds `claude-opus-5` to models.json (the SPOT) and moves the `opus` alias from `claude-opus-4-8` to it. `/v1/models` goes 6 -> 7; OpenClaw picks the new entry up on the next `ocp update` via scripts/sync-openclaw.mjs. server.mjs is NOT modified by this PR — models.json is the single source of truth (ADR 0003) and every consumer (server.mjs MODEL_MAP, setup.mjs, sync-openclaw.mjs, ocp-connect) derives from it. ocp-connect needs no change either: its metadata table matches on the `claude-opus` FAMILY prefix, which `claude-opus-5` hits (the guard added in PR #152 review). Alignment evidence. The claim "the CLI itself defaults opus -> claude-opus-5" is verified from the compiled CLI binary 2.1.220 via `strings`, the same protocol used for #152/#168: latest_per_family:{fable:"claude-fable-5",opus:"claude-opus-5", sonnet:"claude-sonnet-5",haiku:"claude-haiku-4-5"} The same registry entry gives context:{window:1e6,native_1m:true}, max_output_tokens:{default:64000,upper:128000} and pricing:"tier_5_25" -- i.e. identical $5/$25 per MTok to Opus 4.8, so the alias repoint carries no cost change. Availability confirmed with a live subscription-pool completion (`claude -p --model claude-opus-5` -> "OK"). This mirrors #168 (sonnet -> sonnet-5), which established the precedent that following the CLI's own latest_per_family is the correct default behavior. contextWindow is deliberately 200000, not the native 1M: 1. MAX_PROMPT_CHARS is a SINGLE GLOBAL budget, not per-model. derivePromptCharBudget (lib/prompt.mjs) returns `max(floor, max(...contextWindows) * 3)` across ALL entries, so a 1M entry would lift the truncation ceiling to 3,000,000 chars for claude-haiku-4-5-20251001 as well -- genuinely a 200k-native model -- turning clean OCP-side truncation into an upstream API rejection. 2. OpenClaw scales its history budget linearly off this value: contextWindow * maxHistoryShare * SAFETY_MARGIN (= x0.6), plus an oversized-message threshold at x0.5 (compaction-planning, OpenClaw 2026.7.1). Its own bundled registry hardcodes 200000 for Claude models, and the upstream request to raise it to 1M (openclaw#22979) was closed "not planned" -- 1M needs a beta header its compaction path omits. Raising the window for real requires per-model budgets and is ADR-level; a new regression test pins the invariant so that change has to be deliberate. Tests: 452 passed, 0 failed. Three mutations verified to bite: - revert aliases.opus -> 4-8 => 1 failure (opus-alias SPOT) - delete the claude-opus-5 entry => 2 failures (presence + dangling alias) - set its contextWindow to 1000000 => 2 failures (invariant + budget SPOT) E2E on an isolated port (:3999, prod on :3456 untouched and verified so): /v1/models returns 7 ids with claude-opus-5 first; a request with model:"opus" logs event=claude_spawned model=claude-opus-5 tier=opus and returns a real completion. Derived MAX_PROMPT_CHARS stays 600000. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_017gbqUZ8HfBZpjjbzQ85oH8 * fix(docs,test): restore the dropped Opus 4.8 line; assert the window CEILING Two review findings from the independent Iron-Rule-10 reviewer, both confirmed before applying: 1. docs/lan-mode.md: the sample `ocp update` output lost `ocp/claude-opus-4-8` entirely — the edit substituted the id instead of inserting the new one, so the block listed 6 bullets while line 219 of the same document claims "7 models available". Self-contradictory. Restored; the block is 7 again. 2. test-features.mjs: "every contextWindow is 200000" was too rigid and contradicted ADR 0009, which states the budget "scales automatically — no code change". Asserting every entry turned that documented no-code-change path into a must-edit-tests path, and would have failed on a future entry with a legitimately SMALLER window. Now asserts the MAX instead: raising the ceiling (the actual hazard, since the budget is global) still fails, while a smaller-window model stays legal. Verified the reformulation is strictly better, not just different: - raise a window to 1000000 -> 2 failures (ceiling + derivePromptCharBudget SPOT) - add a legitimate 128k model -> 452 passed, 0 failed (old test would have failed here) Tests: 452 passed, 0 failed. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_017gbqUZ8HfBZpjjbzQ85oH8 --------- Co-authored-by: dtzp555 <dtzp555@gmail.com> Co-authored-by: Claude Opus 5 (1M context) <noreply@anthropic.com>