Context is the product.
Every project carries two artifacts: the code, and the understanding that produced it. The code is visible, versioned, backed up. The understanding — why Yjs and not Automerge, why reads never touch the network, what the team decided in March and why — lives in Slack scrollback, stale READMEs, and someone's head.
When an AI agent opens your repo, it gets the first artifact and none of the second. So it guesses. Confidently.
The real deliverable isn't the scaffold. It's the shared memory the project, its agents, and its team operate from.
RepoCrew treats that memory as a first-class system: typed, queryable, versioned, served live. Code you can regenerate from good context. Context, once lost, you cannot regenerate from code.
Brain, not hands.
RepoCrew never runs coding agents on a server. It remembers, plans, and hands the right context to the right tool at the right moment — over MCP, to whichever hands you've chosen: Claude Code, Cursor, Codex, Gemini CLI.
This is not a limitation. It's the design. Your tools have fast, streaming, local loops that a server farm can't match — and you already trust them. We don't compete with your hands. We make them start every session already knowing the project.
Surface one-way doors early.
Some decisions are cheap to make and brutal to unmake: a CRDT library, local-first reads, the tenancy model. Most teams discover these doors after walking through them — when the retrofit costs a quarter.
Discovery is built to find the one-way doors before code constrains them, name them out loud, and record them as decisions with rationale. An agent that knows a decision is a one-way door doesn't relitigate it at 2am.
Deterministic floor, intelligent ceiling.
LLMs are powerful and unreliable. The answer isn't to avoid them — it's to never depend on them for the floor. RepoCrew's bundle generation runs in three stages: Copy is deterministic templates. Compose is rule-based selection from tested blocks. Adapt is one structured, validated LLM pass on top.
If the model fails, you still ship. If it succeeds, you ship something personal.
Nondeterminism is tamed by structured output and deterministic rendering — never by hoping the model behaves.
Humans hold authority. Always.
Agents propose; humans approve. Anything strategic — a milestone, a schema migration, code leaving a branch, a deploy — is gated behind explicit approval. No flag turns this off.
And authority cuts deeper than approvals: you can move any task to any state, and the system will never revert your change. It records the override, feeds it to the next agent run, and reduces repeated mistakes. A system that fights its operator has its priorities inverted.
Version-stamp everything.
Every generated file opens with a stamp: which template, which prompt version, which model, which context. It reads like bureaucracy. It's actually the moat.
Stamps make outcomes attributable: which prompts produce files people keep, which get rewritten within a day. That telemetry is how the planner is built to improve as the platform sees more project outcomes — a system that gets better at shaping new projects because it has seen the outcomes of old ones.
Never lie in the interface.
If an integration is deferred, the button says so. If a merge would violate branch protection, the tool refuses and explains — it doesn't pretend. If a sync is broken, a banner says it's broken before you notice missing data.
Trust in an AI-native tool is spent in one fake success and earned back over months.
The same honesty applies inward: guardrails before refactors, contracts at every boundary, red CI blocks merge. The unglamorous work is the work.
Adopt conventions; don't invent them.
Closes #7 has worked on GitHub since 2013. So that's what RepoCrew uses — not a proprietary prefix that breaks every historical link and retrains every engineer. Where a convention exists and works, we adopt it. Where a provider has a rules engine, we surface it rather than replicate it.
Invention is reserved for the places nothing exists: the live memory, the context compiler, the lifecycle handoff. Everything else should feel like it was already there.