Pre-implementation planning release. No application code exists yet —
this tags the point where the design is considered complete enough to
start building against.
What's in this release
- Full plan pack (
docs/project-plan/): product scope, security/auth/
secrets, vault runtime & git strategy, agent runtime & chat, files/reports/
derived indexes, implementation roadmap, deployment/ops/threat model,
feature backlog, design system, and devops/proxy/SSE notes. - AGENTS-PLAYBOOK.md: the entry point for implementation sessions,
including weaker models. Defines a session-start protocol, a
log-intent → implement → verify → commit work loop scoped to one
checklist item at a time, a recovery protocol for trees left dirty by a
crashed session, and the plan's hard rules distilled to a checklist. docs/progress/: crash-proof progress tracking — a rewritable
STATUS.md(single source of truth for where things stand), an
append-onlyjournal.md, and per-milestone checklists with one
verification command per item, starting with Milestone 0.
Key decisions locked in this pass
- Implementation begins with a Milestone 0 feasibility spike proving
Cline SDK behaviour (multi-turn sessions, resume across restarts,
headless approvals, rules/workflow/skills loading) before any web
scaffolding — the SDK is an agent framework, not a headless build of the
Cline extension, so these can't be assumed. - The app-side tool-policy guard blocking non-catalog writes under
library/is mandatory, since the vault's own PreToolUse hook won't
fire under the SDK. - Runtime is a plain long-lived Node service (not Next.js route
handlers) so agent runs, SSE streams, the writer lock, and git operations
live outside HTTP request lifecycles. - Provider API keys are encrypted only with a dedicated
SECOND_BRAIN_WEB_SECRETS_KEY; storage is refused if it's unset. - MVP is slimmed: automatic context compaction and the source-coverage view
move to near-term backlog, manual compaction ships first. - All writing sessions are expected to serialise behind the single-writer
vault lock — acceptable for a single-principal app, but the UI must show
waiting/lock-holder state. - git-lfs for
library/originals noted as a likely future need, not
adopted yet.
Next
Point a fresh implementation session at AGENTS-PLAYBOOK.md to begin the
Milestone 0 SDK spike.