konman (Context Manager) v1.1.1a1 Alpha

Experimental alpha: Konman is being tested by early users. Generated adapters and parser fixtures do not prove that every vendor client/version will invoke its hooks. Verify your own tool with konman doctor; do not treat prompts or Git hooks as a security boundary. Extraction proposes context for human review and can send redacted transcript text to the configured provider.

Konman keeps developer context in version-controlled project data rather than only in chat sessions.

What's New in 1.1.1a1

• Select the context node types to extract with repeatable konman extract --type options. Requirements, decisions, and boundaries can be extracted, and later runs can process additional types from the same captures.
• Extracted boundaries are proposed for review like other extracted nodes. They become active context only after approval with konman review --id B-001.
• Switching providers with konman config --provider now uses the selected provider's defaults rather than silently reusing the previous provider's API format, endpoint, or credential. Partial updates to the currently configured provider continue to reuse its saved settings.

🎯 Core Philosophy

Standard AI-driven development tools lock your project's history inside localized chat interfaces or vendor ecosystems. When your API tokens expire or you switch models, the new assistant lacks the complete picture, often modifying code it shouldn't touch or re-proposing rejected options.

konman solves this by breaking context down into 6 fundamental dimensions:

1. Who: The actor (Human or specific LLM) making the assertion.
2. What: The changes made (Code modifications, symbols, diffs).
3. Where: The scope of the adjustment (File paths, dependency graphs).
4. When: The timeline anchor (Git commit history, session timestamps).
5. How: The mechanism used (Build configs, testing assertions).
6. Why: The human intent, constraints, and evaluated tradeoffs.

While What, Where, When, and How can easily be recovered by indexing raw code, the Why is completely lost once a session vanishes. konman dynamically captures the "Why" at the moment it is generated, standardizing it into an evaluation matrix:

[\text{Object } (O) \xrightarrow{\text{satisfies}} \text{Requirement } (R) \text{ due to its verified Property } (P)]

🚀 Quickstart & Global Installation

1. Install the library via pip

Ensure you have Python 3.10+ installed. For alpha testing from a source checkout, install the current tree:

pip install konman==1.1.1a1

Verify your global environment configuration setup by triggering the structured help display:

konman --help

2. Scaffold a Codebase Repository

Navigate into your target software application codebase folder (e.g., playce-api) and initialize the context filesystem:

cd /path/to/project
konman init

This deploys the underlying .context/ data graph infrastructure:

• requirements/: .yaml files detailing hard structural limitations and developer preferences.
• decisions/: Architectural tradeoffs, chosen technologies, and rejected alternative options.
• boundaries.yaml: Protected codebase paths mapped to explicit verification criteria.
• .raw/: Project-ignored temporary cache for redacted captured/backfilled events before extraction. Redaction is best-effort; inspect policy and data before sharing.

🔌 Multi-Tool Setup & Integration Manual

konman is built to sit between you and your developer assistants. Instead of modifying tool core logic, running konman setup [tool-name] drops target-specific metadata adapters into your project directory. Every adapter explicitly directs its assistant to look up instructions from a centralized, auto-compiled index file named AGENTS.md.

Setup commands generate configuration; they do not prove that a vendor application accepted or fired the hooks. Hook schemas and transcript paths change by tool/version. Confirm the generated files and test a real prompt with the host application before relying on capture. Konman does not sandbox agent file edits.

1. Antigravity Setup

Antigravity utilizes dynamic lifecycle event hooks and dynamic background JSON-RPC configurations.

• Setup Command:

  konman setup antigravity
  

• What it configures: Appends background server hooks directly to your system profile data (~/.gemini/config/mcp_config.json) and drops an event configuration module down at .agents/hooks.json.
• Workflow: When Antigravity invokes a configured hook, Konman stores a redacted event under .context/.raw/. Confirm hook events and payloads with your installed Antigravity version.

2. VS Code Copilot Setup

Agent-mode Copilot (preview) and Copilot CLI do run hooks. Inline completions and classic Copilot Chat do not.

• Setup Command:

  konman setup vsc-copilot
  

• What it configures: .github/copilot-instructions.md (read AGENTS.md) and .github/hooks/konman.json (UserPromptSubmit, Stop).
• Workflow: Agent sessions can capture events when the host invokes its hook. Classic Chat does not use these hooks; attempt transcript backfill where logs are available:

  konman backfill --source copilot
  

3. Cursor Setup

Cursor Agent chat exposes beforeSubmitPrompt (plus sessionEnd / stop).

• Setup Command:

  konman setup cursor
  

• What it configures: .cursor/rules/konman-context.mdc (alwaysApply: true) and merges capture entries into .cursor/hooks.json without wiping existing hooks.
• Workflow: Prompts can be captured when Cursor invokes the configured hook. Tab completions use a different surface and are not covered by this adapter.

4. Claude Code Setup

Claude Code is the cleanest capture path: labeled human prompt on UserPromptSubmit, plus transcript_path on every event.

• Setup Command:

  konman setup claude-code
  

• What it configures: CLAUDE.md (pointer to AGENTS.md) and merges hooks into .claude/settings.json for UserPromptSubmit, Stop, and SessionEnd.
• Workflow: Configured hooks capture events when Claude Code invokes them. Session JSONL locations vary by configuration; use konman backfill --source claude-code to check supported paths.

5. Codex Setup

Codex now has the same nested hook schema as Claude Code, plus JSONL rollouts on disk.

• Setup Command:

  konman setup codex
  

• What it configures: Merges capture hooks into .codex/hooks.json (UserPromptSubmit, Stop, SessionEnd with a 3s timeout).
• Workflow: Trust/enable hooks in the Codex CLI with /hooks where required. Use konman backfill --source codex to scan supported rollout paths.

6. Kiro Setup

Kiro IDE 1.x and CLI V3 load standalone hook files. Steering remains the static instruction channel.

• Setup Command:

  konman setup kiro
  

• What it configures: .kiro/steering/context_steering.md and .kiro/hooks/konman.json (UserPromptSubmit, Stop).
• Workflow: Kiro hook payloads and transcript paths vary by IDE/CLI version. Confirm the prompt event is received, then use konman backfill --source kiro for supported session files.

🛠️ Operational Command Reference

🎯 Core Workspace Management Commands

🔌 Context Extraction & Lifecycle Pipelines

config supports four request formats: openai-compatible, anthropic, gemini, and ollama. Known provider names select a default format and endpoint. For a custom endpoint, provide its API root URL and the format it implements. Konman does not support arbitrary provider-specific REST schemas. Anthropic and Gemini configurations require an explicit model; Ollama works locally without an API key. Keys are stored in the OS credential store when available; existing plaintext keys must be explicitly migrated.

Examples:

# Known hosted providers select their API format and endpoint automatically.
konman config --api-key "$ANTHROPIC_API_KEY" --provider anthropic --model "$ANTHROPIC_MODEL"
konman config --api-key "$GEMINI_API_KEY" --provider gemini --model "$GEMINI_MODEL"
konman config --api-key "$GROQ_API_KEY" --provider groq --model "$GROQ_MODEL"

# Configure any custom endpoint that implements a supported API format.
konman config --api-key "$LLM_API_KEY" --provider my-provider --api-format openai-compatible --base-url https://llm.example/v1 --model my-model

Extract only requirements, decisions, or boundaries by selecting one or more types. Repeat --type to request multiple kinds. Without the option, extraction requests all supported types.

konman extract --type requirement
konman extract --type decision --type boundary

# Local Ollama does not require an API key.
konman config --provider ollama --model llama3.2
konman extract
konman doctor --check-api
konman review

With no arguments, konman config displays saved non-secret settings. Use konman config --migrate for legacy plaintext settings and konman config --clear to remove configuration. konman extract uses saved settings or one-off --provider, --api-format, --base-url, --model, and --api-key values. KONMAN_API_KEY and KONMAN_BASE_URL are supported; OPENAI_API_KEY and OPENAI_BASE_URL remain aliases. When changing to a different provider with konman config --provider, supply that provider's required credentials and any custom endpoint or API format; settings from the previous provider are not inherited. Omitting --provider or updating the currently configured provider preserves its saved settings. Extraction results stay proposed until approved with konman review; use konman review to list proposals and konman review --id B-001 to inspect a boundary proposal.

Raw transcript data is sensitive. Konman applies best-effort credential-pattern redaction before capture and upload, but this is not a guarantee that all secrets or personal data are detected. Inspect captures and use konman clean for explicit retention cleanup. Add project-specific patterns to .context/manifest.yaml:

redaction:
  patterns:
    - '(?i)internal-token-[A-Za-z0-9]+'

Review the full limitations below before enabling capture for a team.

Run konman --help for command forms and common examples, or konman COMMAND --help for a command's options. Setup subcommands also have dedicated help, for example konman setup claude-code --help.

🔒 Enforcement Guardrails & Git Lifecycles

To catch invalid or stale context locally, embed konman in the repository's Git workflow. This is a developer convenience, not a security boundary; hooks can be bypassed or omitted in other worktrees and CI environments.

Activating the Protection Guardrail:

Run this command inside your target codebase application directory:

konman setup-hooks

This adds a marked block to the active pre-commit hook and executes read-only konman check. Stale dependencies are detected at check time without requiring sync. Resolve flagged decisions with konman resolve [id]; use konman review separately for proposed extraction nodes.

Opting Out / Disabling the Guardrail:

If you need to lift the validation gates for temporary testing cycles, bypass the pre-commit checks, or remove konman from your automated lifecycle track, execute the decoupling unlinker:

konman disable-hooks

This removes only Konman's marked block and preserves unrelated hook commands.

Alpha Limitations

• Hook formats, transcript locations, and tool behavior vary by vendor and version. Adapters and parsers need validation with the exact host application; a generated file does not prove a hook fired.
• Transcript parsers cover known layouts, not every release or account configuration. backfill --source reports parsed/failed counts; inspect the result before relying on it.
• Redaction uses patterns and can miss credentials, personal information, or proprietary text. It is best-effort, not a data-loss-prevention system.
• Extraction is model-assisted and may omit, misunderstand, or duplicate context. Results remain proposed until a human reviews them; quotes establish textual provenance, not truth.
• The configured provider receives redacted transcript excerpts during extract. Review the selected provider's retention and data-use policies before use.
• Git hooks and AGENTS instructions are advisory local safeguards, not sandboxing or access control.
• Alpha testing has not validated every adapter or provider against a live vendor environment. Please report tool name/version, OS, and sanitized diagnostics when filing issues; do not attach raw prompts or credentials.