v0.4.0

Cleaner Agent Timeline: Tool calls, tool results, reasoning, warnings, and restored session history now render with clearer spacing and compact summaries so short model responses no longer disappear in terminal noise.
Compact Tool Output by Default: Large tool outputs are summarized in one line with size and saved-output metadata, while detailed head/tail previews remain available in detailed mode.
Session-Restore Display Metadata: SuperCoder now stores UI-only summaries, previews, policies, and metadata alongside messages so restored sessions can replay the same clean visual timeline without changing the API payload sent to the model.
Safer Interrupted Session Recovery: Incomplete tool-call exchanges left by an interrupted session are repaired before API replay, preventing invalid assistant/tool history from breaking the next turn.

v0.3.10

Host-Side Loop Detection: SuperCoder now detects repeated assistant responses, repeated identical tool calls, repeated tool errors, and repeated no-progress edits before a local model can spin indefinitely.
Corrective-Then-Stop Recovery: The first loop gets a short corrective instruction so the model can change strategy; repeated loops stop the turn cleanly instead of wasting context.
Configurable Stability Guardrails: Loop detection is enabled by default and configurable through loop_detection thresholds for tool calls, tool errors, no-progress edits, assistant repetition, and corrective attempts.
Checkpoint-Safe Aborts: If a loop is stopped during an edit-capable turn, active checkpoints are rolled back to avoid partial file changes.

v0.3.9

Native Tool-Call Recovery: If a local/OpenAI-compatible backend writes a tool call as text in native mode, SuperCoder can parse the supported fallback formats and keep the turn moving.
Malformed Tool-Call Retries: Responses that look like broken tool calls get a short retry correction before the agent gives up.
API-Based Context Usage: Context usage and auto-compact thresholds now use the latest API usage.total_tokens when available, matching proxy/provider accounting more closely.
Interrupt-Safe Compaction: Double-ESC abort state is reset before manual /compact, so an interrupted turn no longer poisons the next compaction request.
Cleaner Terminal UI: Large tool outputs show user-facing previews instead of internal compact markers, long shell commands get bounded previews with Show full command, and Unicode tool arguments render as readable text.

v0.3.8

Persistent Command Approvals: Save one-time, session, always-allow, or always-deny command decisions in project-local permission rules.
Permission Management: Use /permissions to inspect, remove, or clear saved command rules for the current project.
CODE Edit Approval: code mode asks before file edits, while accept-edits keeps the low-friction editing path.
Approval Menus + Diff Preview: Command and edit confirmations use arrow-key menus; file edits show a highlighted diff before approval.
Apply and Auto-Accept: An edit approval can apply the current change and switch the rest of the active loop to accept-edits.

v0.3.7

Host-Enforced Modes: SuperCoder now has ask, plan, code, and accept-edits modes enforced by the host before tools run.
Cache-Friendly Mode Switching: Mode changes no longer rebuild the system prompt or tool schema list; SuperCoder announces mode policy in-band only when needed.
Plan Files: plan mode can save dated plans under .supercoder/plans/ while blocking project file edits and shell commands.
Mode UX: Use /plan, /accept-edits, /accept, /edit, or Shift+Tab to switch modes with a live toolbar indicator.

v0.3.6

Read-Before-Edit Enforcement: code-edit now requires fresh file context before modifying existing files.
Stale Edit Protection: If a file changes after file-read or @file attachment, edits are blocked until the file is read again.
Attachment-Aware Editing: Successful @file attachments count as fresh context, while directory listings, glob, and search results do not.
Freshness Debug Logs: Debug JSONL logs now include freshness_check events without exposing file contents.

v0.3.5

Host-Side Permission Policy: Shell commands now pass through deterministic allow / ask / deny rules before confirmation or execution.
Sensitive Path Protection: .env, private keys, credentials, and similar files are blocked across read, edit, search, glob, project tree, and @path attachments.
Safe Config Defaults: The config template now includes a permissions section for command and path policies.
Audit-Friendly Debug Logs: Denied and cancelled command-exec requests now log permission_decision, tool_call, and tool_result events.

v0.3.4

Explicit Context References: Mention files or directories with path to attach bounded context before the model call.
path Autocomplete: Typing "ma" now suggests matching files and folders, while ignoring runtime/cache directories.
Compact-Safe Attachments: Attached context is stored as context_attachment and kept with the related user prompt during compaction.
Debug Visibility: Debug JSONL logs include context_attachment metadata without dumping full attached file contents into the event.

v0.3.3

Context-Efficient Tool Results: Large tool outputs are compacted before they enter model context, with full output offloaded to .supercoder/tool-outputs/ for inspection.
Improved file-read: File reads now include byte metadata, binary-file protection, maxBytes caps, and nearby path suggestions for typos.
Faster Search + glob: code-search now prefers ripgrep (rg) with a Python fallback, and the new glob tool lists matching paths without reading file contents.
Structured Debug Diagnostics: Debug JSONL logs now preserve native tool-call metadata and include tool_output_masked events with size/offload details.
Cleaner RepoMap Context: RepoMap ignores runtime and environment directories such as .supercoder, .venv, cache folders, and dependency trees.

v0.3.2

Cache-Aware Compact for Local Models: /compact now runs as an in-band chat turn instead of a separate summarization prompt, allowing llama.cpp/Ollama/LM Studio-style backends to reuse prompt cache.
Automatic Context Compaction: SuperCoder can auto-compact around 75% of usable context before the emergency trimming fallback is needed.
Protected Recent Steps: Compact keeps the summary plus the last 6 exact conversation steps, preserving the current working state without relying only on abstraction.