Changelog

All notable changes to ContextManager.

[Unreleased]

[2.13.0] - 2026-03-22

Removed

  • MessageBus, ContextSync, AgentLauncher, AgentDiscovery — Orchestrator stripped to registry + psmux send-keys.
  • 4 bus MCP tools — orchestrator_post_message, read_messages, peek_messages replaced by orchestrator_send.
  • orchestrator_resume_session — Users run copilot --resume themselves.
  • 4 orchestrator settings — injectBusMessages, maxInjectedMessages, injectFleetStatus, agentStaleTimeout removed. Only orchestrator.enabled remains.
  • ACP Orchestrator — Removed headless ACP agent spawning.
  • 3 prescriptive plugin agents — Removed fleet-monitor, build-coordinator, and session-reviewer from the plugin. Replaced by a single flexible orchestrate agent.
  • Orchestrate skill — Skill content consolidated into the orchestrate agent.
  • 4 _intent MCP write tools — save_card_intent, learn_convention_intent, learn_tool_hint_intent, learn_working_note_intent replaced by direct-write tools.

Changed

  • Orchestrator simplified to registry + psmux send-keys — No bus, no context sync. Agents look each other up in the registry and send messages by typing into psmux/tmux panes.
  • Plugin ships single orchestrate agent — One flexible agent that knows registry + psmux send-keys replaces 3 narrow agents + 1 skill.
  • Agent registry is persistent with status tracking — Agents have status (active/idle/stopped). prune() marks stopped instead of deleting. New purge() cleans entries stopped 7+ days.
  • Direct-write MCP tools — save_card, learn_convention, learn_tool_hint, learn_working_note write directly to projects.json with deduplication. No more queuing intents.
  • orchestrator_list_agents filterable by status — Shows agent status and terminal info.

Added

  • orchestrator_send psmux tool — Send a message to another agent via psmux send-keys (resolves session → pane ID in registry).
  • Auto-bind sessions by cwd — HookWatcher matches session working directory to project rootPaths on SessionStart.
  • Pane ID capture — Capture script reads $env:TMUX_PANE on SessionStart; stored in registry metadata.
  • Claude Code plugin — Full plugin with hooks + MCP server. Install via claude plugin install GittyHarsha/ContextManager:claude-code-plugin. Automatic card queue population works fully.
  • Claude Code hook install command — Quick-start: writes hooks to .claude/settings.json without the full plugin.
  • Copilot CLI agentStop / subagentStop hooks — Automatic card queue population now works from CLI sessions.
  • Copilot CLI preToolUse hook — Tool invocation logging.
  • psmux/tmux terminal detection and tracking — Capture script auto-detects psmux or tmux; stores full terminal info (type, paneId, windowId, sessionName) in agent registry.
  • orchestrator_set_agent_meta terminal parameter — Update psmux/tmux pane info on your agent entry.

Fixed

  • MCP server bundled deps — Esbuild bundle (all deps inlined) fixes MODULE_NOT_FOUND on fresh plugin installs.
  • orchestrator_set_agent_meta registry key — Was overwriting a single entry for all agents sharing the MCP connection. Now uses meta.copilotSessionId so each agent gets its own entry.
  • Removed CLI card capture gap warning — agentStop is now available, resolving the original limitation.
  • CLI agentStop transcript path mismatch — Scripts now check transcriptPath (camelCase) first, then fall back to transcript_path.
  • ConvertFrom-Json -Depth on PS 5.1 — Removed unsupported parameter that silently broke all CLI hooks on Windows PowerShell 5.1.
  • VS Code agent hooks file conflicts with Copilot CLI — Gitignored the generated .github/hooks/contextmanager-hooks.json to prevent format conflicts.

[2.12.0] - 2026-03-15

Changed

  • Per-project workflow execution — Re-entrancy guard scoped per-project; multi-project workflows no longer block each other.
  • Per-project auto-distill rate limit — Each project has its own independent cooldown timer.

Fixed

  • Documented CLI card capture gap — Added warnings that CLI sessions cannot auto-populate the card queue (no Stop hook). Workaround: use MCP contextmanager_save_card_intent.

[2.11.0] - 2026-03-15

Added

  • Session tracking opt-out — Disable session tracking entirely via contextManager.sessionTracking.enabled.
  • Bulk session operations — Select and Dismiss or Delete multiple sessions at once.
  • Copilot CLI plugin hooks — Commands to install CLI plugin hooks for terminal-based Copilot.
  • Installation section in README — Marketplace and CLI install instructions.

Changed

  • PostToolUse capture off by default — hooks.postToolUse defaults to false to prevent observation and card queue noise.
  • Renamed Forget → Delete — Destructive session actions now say “Delete”.

Fixed

  • Settings checkboxes not saving — Webview settings toggles now persist correctly.
  • Bulk session buttons not working — Removed confirm() calls blocked by webview sandbox.
  • Card queue spam — Removed synthetic Stop entry generation from PostToolUse handler.

[2.10.0] - 2026-03-09

Added

  • Folder-aware #saveCard — Chat can now list folders, create folders, and save cards directly into named folders.
  • Custom Workflow upgrades — Workflows now support AI and direct-template create, update, and append actions.
  • Dedicated model selectors — Separate model settings are available for AI workflows and AI Draft / Synthesize Card.

Fixed

  • Workflow markdown preservation — AI workflow output now keeps markdown structure when written to cards.
  • Dashboard draft resilience — Open card drafts survive background updates instead of being replaced mid-edit.

[2.9.0] - 2026-03-07

Added

  • One-shot card injection — Selected cards auto-deselect after being injected into a prompt. Toggle in Dashboard → Context → Inject into Every Prompt.

Fixed

  • Local install reliability — install.ps1 no longer relies on code --install-extension.

[2.8.1] - 2026-03-07

Fixed

  • Data migration — Automatic recovery of projects and data when updating across publisher changes.

[2.8.0] - 2026-03-06

Changed

  • esbuild bundling — VSIX reduced from 34.6 MB to 3.3 MB via single-file bundling.

Fixed

  • Per-prompt context injection — “Inject into Every Prompt” now correctly provides project knowledge on every message, not just at session start.

[2.7.0] - 2026-03-05

Added

  • Skip Pattern — Per-workflow regex; if the LLM output matches, the output action is skipped (⏭️ status). Filters out low-value AI responses before they reach cards.
  • Trigger Filter — Per-workflow regex; auto-triggered workflows only fire when event content matches. Scope event triggers to relevant content without disabling the workflow.
  • Execution history — Last 15 runs tracked per workflow with success/skipped/error counts shown in the dashboard.
  • Card variable resolution for auto-triggers — ,, and `` now auto-resolve from the target card on event-triggered workflows.

Fixed

  • Manual run re-entrancy — Dashboard “Run” button now reuses the singleton WorkflowEngine, preserving the re-entrancy guard.

[2.3.0] - 2026-03-05

Added

  • Custom Workflows — User-defined pipelines with custom prompt templates using `` syntax. Define workflows in the Dashboard Intelligence tab with clickable variable insertion buttons.
  • 7 trigger types — Manual, auto-queue, both, convention-learned, card-created, card-updated, observation-created. Event triggers fire automatically in the background.
  • 3 output actions — Create card, update card, append to collector card (with dated separator).
  • Collection template variables — ,, ,, , — each capped by a per-workflow Max Items setting (default 20).
  • Event template variables — ,, , available when triggered by entity events.
  • Re-entrancy guard — Prevents infinite loops when workflow output actions create/update entities that would re-trigger other workflows.
  • WorkflowEngine singleton — Centralized engine created in extension activation, injected via setters into ProjectManager, AutoCaptureService, and HookWatcher.
  • Purple event badges — Dashboard workflow list shows color-coded trigger badges (blue auto-queue, gray manual, purple event triggers).

[2.2.0] - 2026-03-03

Added

  • Tags in distill-to-approve pipeline — Distilled cards now carry 2–5 keyword tags from the LLM, propagated through the full approve chain (individual and batch)
  • Custom prompt for AI synthesis — ✨ AI Synthesize opens the editor with an optional custom prompt textarea for user-directed generation
  • Generate with AI uses LM API directly — Knowledge tab button now calls the Language Model API directly with cancellable progress and robust JSON parsing
  • Descriptive error messages — All AI failure paths show specific reasons instead of generic messages

Changed

  • Unified card selection for hook injection — Knowledge tab checkboxes now drive the “Inject into Every Prompt” hook system directly. The duplicate card picker in the injection section has been removed.
  • Hook renamed SessionStart → UserPromptSubmit — Capture script now uses UserPromptSubmit with hookSpecificOutput wrapper instead of the old SessionStart event
  • Injection section simplified — Dashboard injection section shows selected card count, custom instruction, and full-content toggle only

Fixed

  • Multi-strategy JSON parsing — 3 fallback strategies for parsing LLM responses with preamble text
  • Config crash on non-string values — Safe getString() helper prevents .trim() on null/undefined

Removed

  • Agents feature — Entire feature removed (dashboard tab, handlers, methods, types, config, hook entries)
  • Overview tab — Removed from dashboard. Intelligence is now the default tab.

[2.1.0] - 2026-02-27

Added

  • Smart-merge on queue approval - When approving a card from the queue, ContextManager checks for similar existing cards (Jaccard ≥ 30%) and offers a QuickPick: “Create new” or “Merge into: [title] (XX% similar)”. Merge appends new content to the target card with a dated separator.
  • Knowledge cards in intelligence injection - Tier 1 now includes pinned and includeInContext cards (full content). Tier 2 uses BM25 search to inject task-relevant cards alongside conventions and working notes.

Changed

  • Auto-capture routes cards to queue - Knowledge cards and architecture notes extracted by the Auto-Capture / Auto-Learn service are no longer silently created. They are staged in the Card Queue for user review and approval, the same as hook-captured cards.
  • Tier token defaults raised to 800 - intelligence.tier1MaxTokens and intelligence.tier2MaxTokens both default to 800 (was 400).
  • Intelligence injection cap raised to 32 000 chars - Previous hard cap of 2 000 characters was far too restrictive; raised to 32 k to support cards with substantial content.

[2.0.0] - 2026-02-27

Added

  • distillQueue() - cross-response card synthesis - New pipeline that reads the entire card queue in one LLM call and synthesizes high-quality knowledge card proposals across all responses
  • Tool call capture - Agent hook system now captures tool.execution_start entries from VS Code Copilot transcripts alongside responses (capped at 10, inputs truncated to 200 chars)
  • Observations distill now produces knowledge cards - distillObservations extended with a cards[] array
  • Approve All - One-click approval for all distilled card proposals
  • cardQueue.minResponseLength setting - Configurable minimum (default 300, range 50–5000)

Changed

  • Card Queue merged into Intelligence tab - Standalone Queue tab replaced with inline section
  • Card Queue redesigned as zero-cost staging buffer - No per-response LLM call; quality assessed at distill time
  • Queue capped at 30 items - FIFO eviction
  • Distill results render inline - In Intelligence tab, not a modal

[1.9.0] - 2026-02-26

Added

  • Intelligence tab - New dashboard tab (🧠 Intelligence) for Auto-Capture, Auto-Learn, observations, conventions, tool hints, working notes, token economics, card queue

Changed

  • TODOs tab removed - TODOs are now user-managed only (no agent TODO tools)
  • Observations and Token Economics moved to Intelligence tab

[1.8.0] - 2026-02-26

Added

  • Disk-backed project storage - JSON files on disk instead of VS Code globalState
  • BM25 full-text search - FTS4 with BM25 ranking from matchinfo('pcnalx')
  • Knowledge card flags - pinned, archived, includeInContext
  • Knowledge index file - ~/.contextmanager/knowledge-index.txt
  • #getCard tool - Read a card by ID (read-only, no confirmation)

Changed

  • FTS4 migration - From FTS5 (not available in sql.js 1.14.0) to FTS4
  • Storage migration - Transparent lossless migration from globalState to disk

[1.7.0] - 2025-02-25

Added

  • Typed observations - bugfix 🔴, feature 🟣, discovery 🔵, decision ⚖️, refactor 🔄, change ✅
  • Content-hash deduplication - SHA-based with 30-second window
  • Privacy tags - <private> content stripped before storage
  • Token economics - discoveryTokens vs readTokens, ROI tracking
  • 3-layer search - search (index), timeline (context), fetch (detail)
  • File path extraction - Auto-indexed from prompts and responses

[1.6.0] - 2025-02-25

Added

  • Auto-Capture Service - Zero-friction observation logging from all chat participants
  • Session Continuity - Injects context from previous sessions into new ones
  • Cross-participant LLM learning - Learns from regular Copilot Chat
  • Observation buffer - Circular buffer (default 50 entries)
  • 8 new settings for auto-capture and session continuity

[1.5.0] - 2026-02-24

Added

  • Progressive disclosure - 3-tier knowledge card injection
  • BM25 OR fallback - Automatic retry with OR when AND returns zero
  • camelCase query expansion
  • Nested folders - Hierarchical folder structure for cards
  • Drag-and-drop cards - Move cards between folders
  • Card staleness detection - 30+ day warning
  • Card Health dashboard - Analytics, duplicates, usage tracking
  • Git-tracked card storage - Export as .md with YAML frontmatter
  • SKILL.md export - Claude Code compatibility
  • #organizeCards tool - listFolders, createFolder, moveCard, autoOrganize
  • Card templates - 6 structured templates
  • Keyboard shortcuts - 1-7 tabs, Ctrl+K search, Ctrl+N new card

[1.4.0] - 2026-02-19

Added

  • Project Intelligence Layer - Conventions, tool hints, working notes
  • Tiered injection - Confirmed always, relevant matched, rest on-demand (800 token cap)
  • #projectIntelligence tool - learnConvention, learnToolHint, learnNote, retrospect
  • Branch session enhancements - Living document model, checkpoint, gitDiff
  • @ctx /done command - End-of-task retrospective
  • Data import/export - Full and per-project export/import
  • LaTeX rendering - KaTeX with MathML output

[1.3.0] - 2026-02-19

Added

  • Inline context menu modals - Replace, Delete, Refine, Create Card
  • /add command - Save last AI response as card
  • Find in Card - Ctrl+F-style search within cards

[1.2.0] - 2026-02-18

Added

  • Global knowledge cards (planned) - Share across all projects
  • Mermaid code block styling - Styled display for ` ```mermaid ` blocks
  • Research & Save as Card inline input
  • Concurrent /save queue - No UI conflicts

[1.0.0] - 2026-02-16

Added

  • BM25 full-text search - SQLite FTS5 via sql.js WebAssembly
  • Branch tracking & git integration
  • Custom prompt system - 6 customizable system prompts
  • Content Security Policy - nonce-based
  • 11 proposed API integrations - auto-activating on VS Code Insiders

[0.1.0] - 2025-07-18

Added

  • Chat participant @ctx with 10 commands
  • #projectContext Language Model Tool
  • Knowledge cards with AI generation and refinement
  • TODO management with AI agents
  • Explanation cache
  • Dashboard with 6 tabs
  • 17+ settings
  • Sidebar tree view and status bar