Six stages, each with a gate

Automated today

• ✓Intake interview: the agent asks the questions it needs before it writes a spec.
• ✓Research stage with an evidence policy: the codebase first, and two sources for any high-risk claim (not on the Quick profile).
• ✓Cortex code graph of your repository, and a survey of what an existing project already has.
• ✓Session memory (Mnemos) keeps the goal, constraints and errors across turns.

Roadmap

• ○A full as-built baseline of an existing system (inferred ADRs, debt, unknowns) for your review.
• ○Explicit project archetype selection before planning.

Automated today

• ✓Spec drafted from the palgu/spec-v2 template, with acceptance criteria.
• ✓Spec validation by an isolated validator on a different model family: unsupported claims are flagged (not on the Quick profile).
• ✓Product, tech and marketing review boards; critical findings block.
• ✓Human approval checkpoint: the flow pauses until you approve, and never auto-approves.

Roadmap

• ○Threat, privacy and safety modelling as a spec-stage gate.
• ○Accessibility targets and architecture records (C4, ADRs) generated and checked with the spec.

Automated today

• ✓Atomic todos: each has acceptance criteria, test cases and the files it owns; an integration todo is required.
• ✓Todos are grouped into waves so no two parallel todos touch the same file.
• ✓Tech board plan review and a human approval checkpoint.
• ✓A live plan with progress you can trust: an item counts as done only with evidence the agent actually ran.

Roadmap

• ○Per-step checklists and human-in-the-loop approvals set by your org.

Automated today

• ✓TDD per todo: the new tests must fail first (RED), then pass after the change (GREEN); zero or deleted tests are rejected.
• ✓Changed files are checked against the files the todo owns.
• ✓Org skills are injected into matching requests, and signed stack packs (your standards) define the required gates.
• ✓Routing picks the model per role; budgets and policies apply to every request.

Roadmap

Nothing open for this stage.

Automated today

• ✓Functional acceptance tests mapped to the acceptance criteria, a visual check with screenshots, and a demo video.
• ✓Every check is written to the quality ledger with the commit it ran against.
• ✓When a task’s required checks pass, Palgu signs a “done” verdict for that repository and code digest; a task it will not sign is shown as not attested.
• ✓Independent adversarial review by the tech board; critical findings block.
• ✓Security scan on the integration tree (secrets, static analysis, dependency audit). Warn-first by default: critical findings are reported and the flow continues; a stack pack can set them to block.
• ✓Risk reviews at the gateway score changes to sensitive code and nudge the session to fix them.

Roadmap

• ○Red-team and AI-agent evaluation suites.
• ○Control-plane attestation of every stage’s evidence (today the per-stage evidence is recorded locally; only the final “done” verdict is signed).

Automated today

• ✓Draft pull request whose description is built from the recorded evidence.
• ✓Optional preview deploy of the integration head on your own platform account (off by default, opt-in per org).
• ✓Production deploys need an org admin’s approval, bound to the repository, commit, platform, account, environment and command plan; it expires after 30 minutes and a new commit voids it.
• ✓The agent’s shell refuses production deploy commands. Deploy approvals, denials and runs are recorded locally and copied to the org’s audit trail when the gateway is reachable.

Roadmap

• ○Guardrails beyond the agent’s shell (other tools and MCP servers, an org production-deploy policy) and a managed process for production environment keys.
• ○A release evidence bundle with waivers, rollback and cutover checks.
• ○Deploy records in the dashboard (today they are kept in the org’s audit trail, not shown on the Audit log page).