[frenchie4111]

Bumps the bundled @anthropic-ai/claude-code pin from 2.1.285 → 2.1.301. This binary is used exclusively by Chat tabs (json-mode); Terminal tabs keep spawning the user's PATH claude, so they're unaffected. See the "Dual-claude model" note in CLAUDE.md.

Motivation: the delegation prompt

Older bundled binaries hardcoded an Opus-5-only prompt line that gated delegation on the user alone:

Do not call the AgentTool unless the user requested it
Do not use workflows or deep-research unless the user requested it

Newer versions soften this to honor project config and skills:

Do not use the ${Task} tool, workflows, or deep-research unless the user,
a CLAUDE.md file, or a skill asks for it

Verified in the 2.1.301 binary — 2 occurrences of the softened wording, 0 of the old hard-gated form:

$ strings -n 6 node_modules/@anthropic-ai/claude-code-darwin-arm64/claude \
    | grep -c 'a CLAUDE.md file, or a skill asks for it'
2

So Chat tabs now respect CLAUDE.md and skills when deciding to delegate.

Second commit: Haiku 5.5 pricing

Not cosmetic — the bump changes billing behavior and would otherwise ship a silent bug.

agent-registry.test.ts carries alias-resolution data explicitly marked "re-derive after bumping the pin." Re-deriving against 2.1.301 via the documented procedure (claude --model <alias> -p x --output-format stream-json --verbose, reading the init event):

alias	2.1.285	2.1.301
opus	claude-opus-5-5	unchanged
sonnet	claude-sonnet-5-5	unchanged
haiku	claude-haiku-4-5-20251001	claude-haiku-5-5
fable	claude-fable-5-1	unchanged
best	claude-fable-5-1	unchanged
opusplan	claude-sonnet-5-5	unchanged

claude-haiku-5-5 had no entry in pricing.ts, and rateFor's prefix fallback has no claude-haiku-5 key to catch it — so priceFor returned 0. Every Chat tab session on haiku would have billed $0 silently. The existing "resolves only to models pricing.ts can bill" test is designed to catch precisely this, and did.

The offered alias list itself (BUNDLED_CLI_ALIASES) is byte-identical between the two versions, so no picker changes.

Known limitation

Haiku 5.5 is the only model priced by prompt length:

prompt	input	output	cache hit
≤ 100k tok	$0.10/MTok	$0.50/MTok	$0.01
> 100k tok	$0.50/MTok	$2.50/MTok	$0.05

RATES is keyed by model id alone and rateFor has no access to a request's token count, so only the base tier is carried. Long-context Haiku sessions under-report cost rather than over-report. Flagged in a comment at the callsite; expressing the tier properly would mean threading token counts into rateFor, which is a larger change than this bump warrants. Cache reads are $0.01 against $0.10 input — exactly the existing 0.1 default cacheReadMult, so no override was needed.

Resolver / packaging — no changes needed

Checked the things that could drift with a bundled-binary bump:

• bundledClaudeBinPath() in src/main/json-claude-manager.ts resolves <platform-pkg>/package.json then joins claude. That path exists in 2.1.301 (node_modules/@anthropic-ai/claude-code-darwin-arm64/claude, 250MB, executable) — version-agnostic, no edit.
• asarUnpack globs match on package name, not version.
• .github/workflows/{build-mac,release}.yml and scripts/pack-headless.mjs all read the version out of package.json at build time rather than hardcoding it.

Verification

• node_modules/@anthropic-ai/claude-code/package.json → 2.1.301; platform subpackage → 2.1.301; claude --version → 2.1.301 (Claude Code)
• npm run typecheck — clean
• npx electron-vite build — clean
• npx vitest run — 135 files, 1833 tests, all passing

Lockfile churn is limited to the single dep and its eight platform subpackages.

Note on branch

The bump-bundled-claude-code branch was reused from already-merged PR #213 and sat 154 commits behind main, still pinning 2.1.220 — a PR from it would have reverted main and downgraded the pin. This branch is cut fresh from current main instead.

🤖 Generated with Claude Code