Changelog

Release history — all CodePulse updates in one place - Page 10

All Releases Features Fixes Improvements

fixMarch 30, 2026

CodePulse v2.3.47 — Hang Timer Fix

Fixed CLI hang timer killing Claude during legitimate long operations like approval waits and MCP calls.

Bug Fixes

Fixed: CLI hang timer kills Claude during legitimate operations — HANG_TIMEOUT_MS was set to 180 seconds but APPROVAL_BRIDGE_TIMEOUT is 360 seconds. The hang timer killed the CLI process during long approval waits or MCP calls that exceeded three minutes. Increased the timeout to 420 seconds with smart approval-aware detection — when the timeout fires, it checks if an approval request is pending and extends the deadline instead of killing the process.

fixMarch 29, 2026

CodePulse v2.3.46 — Delegate Mode Intent Fix

Fixed delegate mode misclassifying task messages as simple questions, blocking natural language task delegation.

Bug Fixes (TAB-533)

Fixed: Delegate mode misclassifies task messages as 'simple_question' — Users in delegate mode sending natural language tasks received "I'm in question-only mode" because the intent classifier used word-pattern detection that misclassified messages not starting with action verbs. Now overrides simple_question to multi_step when in delegate mode, ensuring task messages are routed to Claude CLI for execution. All other intents (git_action, project_switch, continuation) are preserved.

fixMarch 28, 2026

CodePulse v2.3.45 — Content Push & Claude Behavior Fixes

Fixed TCVF falsely blocking content-only git pushes and prevented Claude from suggesting PR links after delegate tasks.

Bug Fixes (TAB-531)

Fixed: T3 falsely blocks content-only git push — The TCVF (Test Coverage Verification Framework) T3 gate blocked git push for changelog and blog .mdx file changes with "no tests observed during session." Now checks if all edited files are content-only formats (.md, .mdx, .txt, .yaml, images) before requiring tests. Code changes still require tests as before.
Fixed: Claude suggests PR links after delegate tasks — Added a behavioral rule to all intent types instructing Claude not to suggest creating pull requests or GitHub compare links. The user manages the git workflow via the Task Complete card instead.

improvementMarch 28, 2026

CodePulse v2.3.44 — Diagnostic Logging Improvements

Improved diagnostic logging with user message previews, raw Haiku responses, CLI prompt prefixes, and increased log capture size.

Improvements (TAB-530)

Improved diagnostic logging — Added user message preview (200 characters) during plan generation for easier debugging. Added raw Haiku JSON response logging (2000 characters) to trace plan extraction issues. Added CLI prompt prefix logging (500 characters) in both approval and delegate bridges. Increased /support log capture from 200KB to 500KB to ensure complete diagnostic data in support requests.

fixMarch 28, 2026

CodePulse v2.3.43 — Unicode Escape Revert

Reverted the v2.3.41 Unicode escape to literal UTF-8 conversion that broke all emoji and special characters in Telegram cards.

Reverts

Reverted: Unicode escape to literal UTF-8 conversion — Bun's compiled binary handles \uXXXX escape sequences correctly but mangles literal multi-byte UTF-8 characters. The v2.3.41 change broke all emoji and special characters across every Telegram card, including buttons. Restored the original escape sequences to fix all rendering issues.

fixMarch 28, 2026

CodePulse v2.3.42 — Plan Card Cost & JSON Repair

Fixed plan card showing estimated cost instead of $0 for Max Plan users, and added JSON repair for truncated Haiku responses.

Bug Fixes (TAB-529)

Fixed: Plan card shows estimated cost instead of $0 for Max Plan — Delegate mode uses the Max Plan subscription at zero actual cost. The plan card now correctly shows "$0.00 (Max Plan)" instead of the misleading estimated cost that implied API charges.
Added: JSON repair safeguard for truncated Haiku responses — When Haiku exceeds maxTokens, the JSON response is truncated mid-stream. Instead of falling back to the generic three-step plan, the parser now attempts to repair truncated JSON by closing unclosed strings, arrays, and objects, preserving as much of the structured plan as possible.

fixMarch 28, 2026

CodePulse v2.3.41 — Unicode Escape Change (Reverted)

Replaced Unicode escape sequences with literal UTF-8. This change was reverted in v2.3.43 due to compiled binary incompatibility.

Bug Fixes (Reverted in v2.3.43)

~~Replace Unicode escapes with literal UTF-8~~ — This change replaced \uXXXX escape sequences with literal UTF-8 characters across source files. However, Bun's compiled binary handles escape sequences correctly but mangles literal multi-byte UTF-8 characters. This caused diamond replacement characters across all Telegram cards. Reverted in v2.3.43.

featureMarch 28, 2026

CodePulse v2.3.40 — Extractive Plan Generation

New lossless extractive plan generation that preserves full user intent, richer plan cards with constraints, and increased maxTokens.

Features (TAB-529)

Extractive plan generation preserving full user intent — Replaced the lossy creative plan summarization with a lossless extraction approach. Haiku now extracts constraints, cross-references, and negative constraints directly from user messages. The original user message travels alongside the structured plan to Claude CLI, ensuring no detail is lost.
Richer plan cards — The plan card now displays Constraints, Cross-check, and Do NOT sections when extracted from the user's message, giving full visibility into how the plan was interpreted before execution begins.
Increased maxTokens for plan generation — Bumped from 512 to 1024 tokens to prevent truncated JSON responses from the more detailed extractive prompt.

fixMarch 28, 2026

CodePulse v2.3.39 — MCP Wildcard Syntax Fix

Fixed MCP allowedTools wildcard syntax that silently failed, preventing MCP tools from being available in delegate mode.

Bug Fixes (TAB-528)

Fixed: MCP allowedTools wildcard syntax — The mcp__server__* wildcard pattern silently fails in Claude Code's --allowedTools flag (Claude Code GitHub issue #13077). Changed to mcp__server without the trailing __* suffix to correctly match all tools from each configured MCP server.

fixMarch 28, 2026

CodePulse v2.3.38 — MCP Tools in Delegate Mode

Fixed MCP tools being blocked in delegate mode by dynamically discovering configured MCP servers from settings.

Bug Fixes (TAB-528)

Fixed: MCP tools blocked in delegate mode — The --allowedTools whitelist contained only 10 standard tools. MCP tools from Linear, GitHub, Supabase, Stripe, and other configured servers were excluded, causing delegate mode tasks to fail when they required external integrations. Now dynamically discovers configured MCP servers from ~/.claude/settings.json permissions and adds them to the allow list.

Bug Fixes

Fixed: CLI hang timer kills Claude during legitimate operations — HANG_TIMEOUT_MS was set to 180 seconds but APPROVAL_BRIDGE_TIMEOUT is 360 seconds. The hang timer killed the CLI process during long approval waits or MCP calls that exceeded three minutes. Increased the timeout to 420 seconds with smart approval-aware detection — when the timeout fires, it checks if an approval request is pending and extends the deadline instead of killing the process.

Bug Fixes (TAB-533)

Fixed: Delegate mode misclassifies task messages as 'simple_question' — Users in delegate mode sending natural language tasks received "I'm in question-only mode" because the intent classifier used word-pattern detection that misclassified messages not starting with action verbs. Now overrides simple_question to multi_step when in delegate mode, ensuring task messages are routed to Claude CLI for execution. All other intents (git_action, project_switch, continuation) are preserved.

Bug Fixes (TAB-531)

Fixed: T3 falsely blocks content-only git push — The TCVF (Test Coverage Verification Framework) T3 gate blocked git push for changelog and blog .mdx file changes with "no tests observed during session." Now checks if all edited files are content-only formats (.md, .mdx, .txt, .yaml, images) before requiring tests. Code changes still require tests as before.

Fixed: Claude suggests PR links after delegate tasks — Added a behavioral rule to all intent types instructing Claude not to suggest creating pull requests or GitHub compare links. The user manages the git workflow via the Task Complete card instead.

Improvements (TAB-530)

Improved diagnostic logging — Added user message preview (200 characters) during plan generation for easier debugging. Added raw Haiku JSON response logging (2000 characters) to trace plan extraction issues. Added CLI prompt prefix logging (500 characters) in both approval and delegate bridges. Increased /support log capture from 200KB to 500KB to ensure complete diagnostic data in support requests.

Reverts

Reverted: Unicode escape to literal UTF-8 conversion — Bun's compiled binary handles \uXXXX escape sequences correctly but mangles literal multi-byte UTF-8 characters. The v2.3.41 change broke all emoji and special characters across every Telegram card, including buttons. Restored the original escape sequences to fix all rendering issues.

Bug Fixes (TAB-529)

Fixed: Plan card shows estimated cost instead of $0 for Max Plan — Delegate mode uses the Max Plan subscription at zero actual cost. The plan card now correctly shows "$0.00 (Max Plan)" instead of the misleading estimated cost that implied API charges.

Added: JSON repair safeguard for truncated Haiku responses — When Haiku exceeds maxTokens, the JSON response is truncated mid-stream. Instead of falling back to the generic three-step plan, the parser now attempts to repair truncated JSON by closing unclosed strings, arrays, and objects, preserving as much of the structured plan as possible.

Bug Fixes (Reverted in v2.3.43)

~~Replace Unicode escapes with literal UTF-8~~ — This change replaced \uXXXX escape sequences with literal UTF-8 characters across source files. However, Bun's compiled binary handles escape sequences correctly but mangles literal multi-byte UTF-8 characters. This caused diamond replacement characters across all Telegram cards. Reverted in v2.3.43.

Features (TAB-529)

Extractive plan generation preserving full user intent — Replaced the lossy creative plan summarization with a lossless extraction approach. Haiku now extracts constraints, cross-references, and negative constraints directly from user messages. The original user message travels alongside the structured plan to Claude CLI, ensuring no detail is lost.

Richer plan cards — The plan card now displays Constraints, Cross-check, and Do NOT sections when extracted from the user's message, giving full visibility into how the plan was interpreted before execution begins.

Increased maxTokens for plan generation — Bumped from 512 to 1024 tokens to prevent truncated JSON responses from the more detailed extractive prompt.

Bug Fixes (TAB-528)

Fixed: MCP tools blocked in delegate mode — The --allowedTools whitelist contained only 10 standard tools. MCP tools from Linear, GitHub, Supabase, Stripe, and other configured servers were excluded, causing delegate mode tasks to fail when they required external integrations. Now dynamically discovers configured MCP servers from ~/.claude/settings.json permissions and adds them to the allow list.