@codex review

@codex review

Before I can review this series (#2188–#2192) I need the branch structure fixed — right now it's not reviewable as five PRs.

The problem: all five PRs (pr/1-cache-diagnostics-v1 … pr/5-ssh-remote-rfc) target main as base, but each branch is stacked on the previous one, so every PR shows the same cumulative +1740/-39 diff across 33 files. I confirmed it: this 'cache diagnostics' PR (#2188) already contains the delete_range, delete_symbol (tree-sitter), and ssh-remote code (56 references). So #2188 through #2191 are literally the same giant diff — merging any one of them would merge all five features at once, and their titles don't match their contents.

Please restructure into independent, single-feature PRs (preferred for fork PRs):

• pr/1 = only the cache-diagnostics change off main
• pr/2 = only the MCP canonicalization change off main
• pr/3 = only delete_range
• pr/4 = only delete_symbol
• pr/5 = only the SSH RFC

Each branched fresh from main with just that feature's commits (cherry-pick or interactive rebase to drop the others). Then each PR's diff is its own feature and I can review them on their merits — these are all things I want to look at properly (especially the two file-deleting tools and the SSH RFC, which need careful individual review).

If you'd rather keep them stacked, that requires pushing the intermediate branches to this upstream repo and pointing each PR's base at the previous branch — but for a fork, independent branches off main is simpler. Let me know once they're split and I'll dig into each.

Updated the PR series as requested. The five fork branches are now independent branches off main, each with one feature commit instead of the previous shared cumulative stack.

Branch heads now are:

• feat: cache diagnostics v1 — /cache-miss-report, doctor --cache, prefix hash evidence | 缓存诊断 v1 #2188 pr/1-cache-diagnostics-v1 -> 0256e1df (feat(cache): add prompt cache diagnostics)
• feat: MCP cache-stable canonicalization — deterministic tool order and schema key sorting | MCP 缓存稳定化 #2189 pr/2-mcp-canonicalization -> d99ee7d3 (feat(mcp): canonicalize tool schemas for cache stability)
• feat: delete_range — reliable large text deletion with anchor-based range matching | 大块文本删除工具 #2190 pr/3-delete-range -> df92d0e4 (feat(tools): add anchor-based delete_range)
• feat: delete_symbol — AST-aware symbol deletion via tree-sitter (6 languages) | AST 符号删除工具 #2191 pr/4-delete-symbol -> e56eb2cd (feat(tools): add AST-aware delete_symbol)
• feat: SSH remote workspace RFC — dry-run bootstrap for ssh:// targets (#2140) | SSH 远程工作区 RFC #2192 pr/5-ssh-remote-rfc -> 9ded388d (feat(cli): add ssh remote dry-run RFC)

I rechecked the GitHub PR file lists after pushing: each PR now shows only its own feature scope rather than the old 33-file cumulative diff.

Verification run locally:

• Focused Vitest suites for each independent PR branch passed.
• npm run typecheck passed on each independent PR branch.
• npm run lint passed on the branches that were committed before hooks were installed; the two later amended branches also passed the pre-commit lint hook.
• The pre-push hook ran npm run verify successfully during the multi-ref push (build, lint, typecheck, and full test suite: 308 files / 3960 passed / 9 skipped).

Root cause was branch topology: the branch names were separate, but all five remote heads pointed at the same stacked tip. I rebuilt and force-updated the fork heads with --force-with-lease so the existing PRs can now be reviewed independently.

@codex review

Addressed the diagnostic hash memoization concern from review 4386489854 in e1e93e3.

Implementation:

• ImmutablePrefix now memoizes PrefixDiagnosticHashes in a WeakMap keyed by immutable/frozen tool snapshots.
• diagnosticHashes() returns cached evidence for the frozen prefix.tools() snapshot used on the hot path, while mutable external snapshots are computed without caching.
• replaceSystem, addTool, and removeTool all clear the diagnostic hash cache via the same prefix-cache invalidation path as _fingerprintCache.
• CacheFirstLoop still passes the actual turn-start toolSpecs snapshot, so the earlier MCP hot-add accuracy fix is preserved.

Regression coverage:

• repeated diagnosticHashes(prefix.tools()) returns the memoized object;
• replaceSystem, addTool, and removeTool invalidate and recompute;
• the existing mid-turn hot-add loop test still verifies diagnostics record the sent snapshot.

Verification:

• npm test -- tests/memory.test.ts tests/loop.test.ts tests/cache-diagnostics.test.ts
• npm run lint
• npm run typecheck
• pre-push npm run verify passed: build, lint, typecheck, and full test suite (308 test files, 3955 passed, 9 skipped).

Resolved the branch conflicts by merging current upstream/main into pr/1-cache-diagnostics-v1 in merge commit 43a113a1.

Conflict handled:

• tests/ui-slash-suggestions.test.tsx: preserved feat: cache diagnostics v1 — /cache-miss-report, doctor --cache, prefix hash evidence | 缓存诊断 v1 #2188’s hyphenated command parsing for /cache-miss-report, preserved upstream’s stable grouped-window assertion from the slash-command updates, and updated the expected bare slash command count to 46 now that feat: cache diagnostics v1 — /cache-miss-report, doctor --cache, prefix hash evidence | 缓存诊断 v1 #2188 cache commands and upstream /max-tokens are both present.
• tests/slash.test.ts: updated the merged non-advanced command count to 46 for the same combined command surface.

Verification:

• npm test -- tests/ui-slash-suggestions.test.tsx tests/slash.test.ts tests/cache-diagnostics.test.ts tests/memory.test.ts tests/loop.test.ts
• npm run lint
• npm run typecheck
• pre-push npm run verify passed: build, lint, typecheck, and full test suite (311 test files, 4007 passed, 9 skipped).

GitHub now reports mergeStateStatus: UNSTABLE instead of DIRTY, so the merge conflict is resolved; remaining mergeability, if any, should be CI/status related rather than file conflicts.

@codex review

Codex Review: Didn't find any major issues. Chef's kiss.