Claude Code's Hook Enforcement Gap and Haiku 5.5 Pricing

Claude Code v2.1.294: Instruction-Based Hooks Were Not Blocking What You Told Them To Block

If your Claude Code deployment uses natural-language prompt hooks — instructions written as "Block commands that modify production config" rather than explicit allow/deny lists — you had a regression. Claude Code v2.1.294 patches a bug where hooks written in that instruction style were passing the very actions they were meant to block. The release also tightened how Stop and SubagentStop hooks are evaluated, so a hook like "Carry on if the build is broken" no longer causes Claude to halt early when it shouldn't.

This matters most if you've deployed Claude Code to engineering teams under a managed policy and relied on instruction-style hooks as your guardrail layer rather than explicit command lists. Go check your hook definitions now. If any of them read like a natural-language rule rather than a structured predicate, test them against the v2.1.294 runtime before assuming the behavior you intended is what was enforced last week.

The fix to Stop/SubagentStop hook judgment is a separate correctness improvement, but it's consequential for long-running workflow agents — a subagent that should keep iterating after a build failure will now actually do so.

Claude Haiku 5.5: Sub-Agent Economics Just Changed

Claude Haiku 5.5 landed on October 7, and the pricing shift is significant for anyone running multi-agent architectures where Haiku carries the high-frequency work. The new model prices at $0.10/$0.50 per million tokens for prompts under 100K tokens — a substantial drop from Haiku 4.5's $1/$5, and now matched to OpenAI's GPT-6 Luna at the same tier.

For a client running a document-processing workflow where Haiku handles extraction and classification at scale, this directly changes the unit economics of the pipeline. The 100K token threshold matters: prompts that spill past it jump to $0.50/$2.50, so context hygiene on high-volume sub-agents is no longer just a latency concern — it's a cost control lever.

Claude Code v2.1.293 made Haiku 5.5 the default Haiku on the Anthropic API and added the claude-haiku-5-5 model identifier. If you're pinning model versions in your agent definitions, update them explicitly. If you're not pinning, verify the new default behaves as expected in your eval suite — capability changes between Haiku generations have surprised teams before.

The 1M context window on Haiku 5.5 is also worth noting. Prior Haiku models weren't candidates for tasks requiring deep context. That constraint is gone.

Claude Code v2.1.295: onFailure: "block" and Per-Hook Failure Semantics

v2.1.295 added onFailure: "block" to command and HTTP hooks. Previously, a hook that couldn't start, timed out, or exited with an unexpected code would silently let the action through. Now you can declare that a hook failure itself should block the action.

This is the right default for security-sensitive hooks and the wrong default for availability-sensitive ones. The distinction matters operationally: a hook that calls an internal approval service needs onFailure: "block" so a service outage doesn't become a permission bypass. A hook that logs to a telemetry endpoint should probably fail open. You now have to make that call explicitly per hook, which is the correct design — but it means your existing hook inventory needs a review pass to assign the right failure semantics before you upgrade to v2.1.295 in production.

The same release added Program Status Protocol (OSC 7501) support. Terminals that implement it can surface whether Claude Code is working, waiting, or done — a small quality-of-life change for engineers who run Claude Code in a multiplexer alongside other processes.

Claude Code v2.1.296 and v2.1.292: Subagent Control Surfaces Expanding Fast

Two releases this week added controls that matter for teams building multi-agent workflows. v2.1.296 introduced CLAUDE_CODE_WORKFLOW_SUBAGENT_MODEL, which lets you run every workflow agent on a single specified model while other subagents keep their own model assignments. It also added autoCompactWindow to subagent frontmatter and --agents definitions, so a subagent can trigger context compaction earlier than the main conversation's window — useful for long-running subagents that would otherwise hit the context ceiling mid-task.

v2.1.292 added an effort parameter to the Agent tool, so you can now tell Claude to run a sub-agent at a specified effort level. It also added prompt caching to $.model.complete for mods, and introduced prompt.autocomplete as an event mods can hook to extend the prompt box's autocomplete list.

Taken together, these three releases add up to a meaningful expansion of the control surface for subagent orchestration. Teams that built workflow architectures six months ago against a much simpler Claude Code API are now running on a substantially different runtime. A configuration audit — model assignments, effort levels, compaction thresholds, hook failure semantics — is warranted before the end of the quarter.

Anthropic TypeScript SDK v0.133.0: Managed Agents Gets Workflow and Thread Filtering

SDK v0.133.0 added types for Chat and Cowork unified analytics metrics, and — more relevantly for enterprise deployments — added workflows, multiagent configuration, and thread status filtering to Managed Agents. The prior SDK surfaced Managed Agents as a relatively flat API. Thread status filtering in particular is necessary for any operational dashboard that needs to distinguish running, waiting, and completed agent threads at scale.

The same release cycle (v0.132.0) added display_name to RBAC roles and deprecated name, added lifecycle stage fields and filtering to /v1/models, and surfaced url_sources to the Managed Agents web_fetch tool config. If you're building tooling on top of the Anthropic API for enterprise governance — model lifecycle tracking, role management, agent observability — this is a meaningful week for SDK surface area.

Copilot Studio: Shared Plugin Registry Enters Preview

Microsoft 365 Roadmap item 570669 puts Copilot Studio's connection to the shared plugin registry into preview this month, with GA in November. For organizations where multiple teams are building Copilot Studio agents independently, this is the governance surface that matters: a centralized registry means IT can control what plugins are discoverable before makers start wiring them into production agents.

The preview window is the right time to inventory what your teams are building and decide which plugins belong in the shared registry versus which should stay scoped to individual environments. November GA will come faster than most IT governance cycles move.

Watch next week for whether Anthropic publishes any Haiku 5.5 eval comparisons against GPT-6 Luna — the pricing match invites direct benchmarking, and the results will determine whether Haiku 5.5 holds at the default for high-volume sub-agent roles or gets displaced.