CSC 4.2 release guide

This guide is for developers who use CSC today or plan to upgrade. It summarizes the changes from 4.2.0 through 4.2.6 and turns the changelog into practical guidance for daily use.

Short version

The CSC 4.2 series focuses on more reliable session recovery, steadier context management, better OpenAI-compatible provider support, stronger Chinese and i18n quality, and timely security fixes.

If you are using 4.2.0, 4.2.1, 4.2.2, or 4.2.3, upgrade directly to 4.2.6. You get more resilient long-session behavior, better Windows and legacy configuration compatibility, and more reliable third-party model integration.

Who should upgrade

• Developers who frequently resume sessions: The 4.2 series improves --resume, --continue, session file discovery, and legacy configuration fallback behavior.
• Developers running long-context tasks: 4.2.6 enables reactive compact by default and restores automatic compression retry when prompts become too long.
• Teams using OpenAI-compatible models: 4.2.4 through 4.2.6 improve retry behavior, error conversion, and structured output for OpenAI-compatible endpoints such as Ollama, DeepSeek, vLLM, and other compatible services.
• Chinese and bilingual teams: 4.2.5 improves i18n coverage, CJK width handling, menu shortcuts, and user-facing copy.
• Security-conscious teams: Multiple 4.2 releases patch high-severity dependency vulnerabilities through upgrades and overrides.

Release map

Version	Focus	User value
4.2.6	Retry, Reactive Compact, security fixes	More resilient APIs, more reliable long-context compression, and several high-severity vulnerability fixes
4.2.5	i18n, build controls, message metadata	Better Chinese UX, consistent streaming message IDs, and build-time API selection
4.2.4-beta	Tool result merging, OpenAI retry, session lookup	Fewer messages, steadier third-party providers, and easier session recovery
4.2.3	/goal command	Set a session goal and preserve it across resumed sessions
4.2.2	Session Warm Pool, Knowledge Hub	Lower first-message latency and better knowledge update flows
4.2.1	Agent panel and UI fixes	Clearer agent state, context usage, and model display
4.2.0	Legacy compatibility, workspace rename	Smoother migration from older legacy paths and package scopes

Key changes and how to use them

1. Session recovery and legacy compatibility are more reliable

Starting in 4.2.0, CSC added broad legacy configuration fallback support across session scanning, file history, paste cache, stats cache, backups, memory, plugin marketplace data, and user rule discovery. Later releases improved session resume, file loading, case-insensitive lookup, and assistant prompt fallback.

This means you do not need to manually migrate every old configuration file before upgrading. CSC prefers the newer CoStrict paths and falls back to legacy directories when needed.

Common commands:

csc --continue
csc --resume <session-id>

This is most useful when you:

• continue a previous task after upgrading;
• recover older sessions on Windows;
• resume a project after path casing changes;
• keep legacy configuration while gradually moving to ~/.costrict/ paths.

2. Long-context work is steadier

4.2.2 introduced Session Warm Pool and context prewarm to reduce first-message latency. 4.2.4 improved message routing and token usage tracking. 4.2.6 enables reactive compact by default and restores automatic compression retry for prompt-too-long failures.

For users, long-running work becomes less fragile. When the context approaches the model limit, CSC can compact more proactively. When a request fails because the prompt is too long, CSC can retry after automatic compression instead of stopping immediately.

Recommended usage:

• use the latest 4.2 release for large refactors, long investigations, and multi-file tasks;
• use /goal to make the task objective explicit;
• let CSC attempt automatic compression before splitting the task manually.

3. OpenAI-compatible providers are more production-ready

4.2.4 wraps OpenAI-compatible chat completions in shared retry logic and converts OpenAI errors into CSC-compatible errors. 4.2.6 adds unified retry behavior for 429 and 529 responses, and moves retry error handling into a shared module.

This is important for teams using third-party model endpoints. Temporary rate limits, overloaded services, and compatibility-layer errors now follow a more consistent retry path.

Example configuration:

export COSTRICT_USE_OPENAI=1
export OPENAI_API_KEY=<your-api-key>
export OPENAI_BASE_URL=<your-openai-compatible-endpoint>
export OPENAI_MODEL=<your-model>

csc

4.2.4 also fixes OpenAI structured output handling by mapping output format requests to response_format with json_schema. If your workflow depends on structured output, tool calls, or automated parsing, use at least 4.2.4-beta; for production usage, prefer 4.2.6.

4. /goal keeps long sessions focused

4.2.3 adds the /goal command to set, clear, and show the current session goal. The goal participates in the query loop and is restored with --resume and --continue.

Example:

/goal Fix the login failure, add regression tests, and make sure bun run typecheck passes

Use it when you:

• want clear acceptance criteria before fixing a complex bug;
• need CSC to stay aligned during a multi-turn refactor;
• resume a session after an interruption;
• collaborate with a team and want the shared objective visible.

5. Chinese and i18n UX are more complete

4.2.5 is the main i18n quality release. It applies a broad audit, standardizes terminology such as Agent, Vault, and Plan, externalizes security messages and pipe/IPC copy, and fixes CJK width handling and empty-string fallback in t().

The locales directory also moves from src/locales/ to the root-level locales/ directory to improve bundled package compatibility. Permission menus, MCP menus, ConfigTool settings, tool validation messages, and shortcut hints also receive better externalization.

Users should notice:

• more complete Chinese menus and prompts;
• more accurate terminal layout with mixed Chinese and English text;
• fewer untranslated permission and MCP messages;
• stronger i18n checks across source scanning, locale symmetry, component contracts, and CLI smoke tests.

6. Security fixes are included throughout the series

The 4.2 series continuously addresses dependency vulnerabilities:

• 4.2.1 upgrades vulnerable dependencies including shell-quote, exporter-prometheus, and tmp;
• 4.2.2 forces @grpc/grpc-js >= 1.14.4;
• 4.2.3 forces esbuild >= 0.28.1;
• 4.2.4-beta overrides vulnerable form-data, protobufjs, and ws versions;
• 4.2.6 fixes multiple high-severity hono and undici vulnerabilities.

If you use CSC in enterprise, CI/CD, or remote development environments, these security patches are a strong reason to upgrade.

What changes after upgrading from 4.2.0 to 4.2.6

The biggest difference is lower day-to-day failure rate, not a dramatic UI redesign:

• historical sessions are easier to recover;
• long-context tasks are less likely to stop because a prompt is too long;
• OpenAI-compatible endpoints can recover from 429 and 529 responses more often;
• Chinese terminal UI is more consistent;
• permission, MCP, and configuration prompts are clearer;
• dependency vulnerability exposure is reduced.

For heavy CSC users, these changes translate into fewer manual recovery steps, fewer repeated prompts, and fewer interruptions from provider instability.

Upgrade recommendation

Recommended version

Use 4.2.6 as the 4.2 baseline. It includes the improvements from 4.2.0 through 4.2.5 and adds important retry, compaction, security, and i18n fixes.

npm install -g @costrict/csc@latest
csc --version

Before upgrading

• If you still use legacy configuration, keep ~/.costrict/ available.
• If you use an OpenAI-compatible provider, confirm OPENAI_BASE_URL, OPENAI_API_KEY, and OPENAI_MODEL are configured.
• If you use the Chinese UI, run a few common flows after upgrading to confirm menu, permission, and MCP copy look correct.
• If you use CSC in CI or enterprise images, reinstall dependencies so security overrides take effect.

Version details

4.2.6: stability and security fixes

4.2.6 focuses on runtime stability. It enables reactive compact by default, restores prompt-too-long automatic compression retry, and fixes API retry message formatting, MCP ENOENT error classification, permission/MCP menu i18n, and shortcut issues.

The OpenAI-compatible provider gains unified retry behavior for 429 and 529 responses, with added test coverage. Dependencies also receive several high-severity hono and undici vulnerability fixes.

4.2.5: i18n and build controls

4.2.5 is the main Chinese and i18n quality release. It improves terminology consistency, CJK width handling, user-facing text externalization, and L1-L4 i18n scan coverage.

It also adds the COSTRICT_ENABLED_API build macro for build-time API selection. Streaming message metadata now reuses one messageUUID, making event IDs, transcript storage, and the final assistant message UUID more consistent.

4.2.4-beta: tool result merging and provider stability

4.2.4-beta adds mergeToolResults, merging tool_result content back into the corresponding tool_use.output field to reduce message count and improve API efficiency. The OpenAI provider gains shared retry behavior and correct structured output mapping.

Session file lookup also improves with legacy fallback and case-insensitive matching, which is especially useful on Windows.

4.2.3: goal management

4.2.3 adds /goal, making the session objective explicit state. Goals can be restored after session resume and are integrated into tracking and completion checks.

This is useful for long tasks, complex bug fixes, and team workflows where staying aligned matters.

4.2.2: cold-start and knowledge updates

4.2.2 introduces Session Warm Pool and context prewarm to reduce first-message latency. Message Routing supports token calculation after compaction, and Knowledge Hub gains update detection, batch update, and UI status indicators.

It also fixes orphaned batch workers, a Ctrl+Z infinite loop, and legacy fallback for session resume and file loading.

4.2.1: agent panel fixes

4.2.1 improves the Agent Detail Panel and related UI. Agent rows show context usage and model information, and the status line removes noisy labels and cost display.

It also fixes model name display, context row layout, slash command suggestion spacing, and session resume diagnostics.

4.2.0: migration foundation

4.2.0 is the migration foundation for the series. It adds broad legacy configuration fallback support and moves workspace scopes from @ant to @costrict.

It also fixes session.error emission for streaming API errors, session deduplication, and /memory path mismatches.

Summary

The CSC 4.2 series delivers three main benefits:

1. Smoother migration: old configuration, sessions, and legacy directories can still be discovered.
2. More stable long-running work: warm pools, reactive compact, prompt-too-long retry, and goal management improve continuity.
3. More reliable provider compatibility: OpenAI retry behavior, error conversion, and structured output support make third-party model integration more practical.

For day-to-day use, treat 4.2.6 as the recommended baseline for the 4.2 series.