Problem
For the Claude Code harness, self-driving-agents install ... --harness claude-code stages agent content under ~/.self-driving-agents/claude-code/<agent> and asks the user to run:
/hindsight-memory:create-agent <agent> from ~/.self-driving-agents/claude-code/<agent>
The installer configures the Hindsight Claude Code plugin with:
dynamicBankGranularity = ["agent", "project"]
That means the create-agent command initializes only the bank derived from the project directory where the user runs it. The seed files are retained into that current bank, and the knowledge pages / mental models are created in that current bank.
However, the generated subagent file is written to ~/.claude/agents/<agent>.md, so the subagent is globally available in Claude Code. If the user switches to another project, the same subagent can still be invoked, but the derived bank is now <agent>::<otherProject>, which has not had the seed content retained or the mental model pages created.
Impact
The agent appears installed and usable globally, but in a new project it starts with an uninitialized Hindsight bank. Its startup instructions call agent_knowledge_list_pages, but the current project bank may have no pages and no retained seed documents.
This is confusing because the subagent definition is global while the seeded memory is project-scoped.
Expected behavior
The user should either be clearly told that each project must initialize the agent's Hindsight bank once, or the subagent should detect that the current project bank has not been initialized and guide the user through initialization.
Possible direction
This probably needs an explicit product decision around whether Claude Code self-driving agents should use project-scoped seed initialization, agent-global seed knowledge, or some hybrid. At minimum, the install flow and generated subagent instructions should make the current scoping behavior clear.
Problem
For the Claude Code harness,
self-driving-agents install ... --harness claude-codestages agent content under~/.self-driving-agents/claude-code/<agent>and asks the user to run:/hindsight-memory:create-agent <agent> from ~/.self-driving-agents/claude-code/<agent>The installer configures the Hindsight Claude Code plugin with:
dynamicBankGranularity = ["agent", "project"]That means the create-agent command initializes only the bank derived from the project directory where the user runs it. The seed files are retained into that current bank, and the knowledge pages / mental models are created in that current bank.
However, the generated subagent file is written to
~/.claude/agents/<agent>.md, so the subagent is globally available in Claude Code. If the user switches to another project, the same subagent can still be invoked, but the derived bank is now<agent>::<otherProject>, which has not had the seed content retained or the mental model pages created.Impact
The agent appears installed and usable globally, but in a new project it starts with an uninitialized Hindsight bank. Its startup instructions call
agent_knowledge_list_pages, but the current project bank may have no pages and no retained seed documents.This is confusing because the subagent definition is global while the seeded memory is project-scoped.
Expected behavior
The user should either be clearly told that each project must initialize the agent's Hindsight bank once, or the subagent should detect that the current project bank has not been initialized and guide the user through initialization.
Possible direction
This probably needs an explicit product decision around whether Claude Code self-driving agents should use project-scoped seed initialization, agent-global seed knowledge, or some hybrid. At minimum, the install flow and generated subagent instructions should make the current scoping behavior clear.