Project memory with a plan for the next step

At task intake, propose the steps, advisory roles, dependencies, checks, repair paths and approval boundaries. A small task can stay with one agent; additional roles do not start automatically.

The opt-in local coordinator binds results to the checkout and task contract, preserves failures and enforces configured repair and time limits. Execution is sequential in this release. It does not replace host permissions, an operating-system sandbox or a deployment adapter.

Classic project notes are useful until old and new decisions conflict or a summary loses its source. ITHZ MCP keeps an immutable episode trail and builds typed abstractions above it: a decision, risk, gate or claim can point to evidence, declare its scope and be superseded without erasing history.

The practical result is smaller, role-specific context. An implementer can receive the decisions and constraints needed for a task, while a reviewer or cross-lab opponent can inspect a more independent evidence path. The model is not made magically smarter; it becomes more consistent and easier to audit.

A long-lived memory can preserve a mistake as efficiently as a correct decision. The canary therefore tests suspicious or conflicting memory in a deliberately constrained lane: it starts disabled, reads evidence only, has time and case limits, and can be stopped with a kill switch.

Its verdict cannot execute a change or write itself into the archive as truth. Promotion still depends on traceable receipts, independent review and the human or workflow authority that already governs the project.

Existing project.ithz archives remain readable. The safe upgrade is to back up the archive, install the current package in a fresh Python 3.10+ environment, run the memory-integrity and canary checks, and keep the canary disabled until the project explicitly opts in.

Older records can remain as legacy evidence while new work starts producing stronger provenance and v2 review receipts. Very old or customized installations still need a project-specific trial and a tested rollback; compatibility is not a reason to skip verification.