v1.8.8

Harden the agent output contracts.

• Implementer can report it is blocked. sdd-implementer gains an IMPLEMENTATION BLOCKED exit state, returned when verification cannot be made to pass or the step's approach is fundamentally invalid. implement-plan routes a blocked step to you and asks how to proceed, instead of hardening a falsely-complete step.
• Step breakdown can flag an unbreakdownable plan. sdd-step-breakdown gains a PLAN TOO AMBIGUOUS TO BREAK DOWN state; write-plan surfaces what is unclear and stops before declaring the plan ready.
• Verification is shown, not asserted. The implementer and the three committing agents (step hardener, standards enforcer, final reviewer) now quote the tail of the actual verification output alongside each command.
• Committing agents never bypass hooks. The step hardener, standards enforcer, and final reviewer are instructed never to use git commit --no-verify, so a blocking pre-commit hook is not defeated in-prompt.
• Skip path discovers commit conventions. When you skip a hardener's findings, implement-plan now discovers the project's commit conventions and stages files by name instead of committing as-is.
• Hardener remarks surface in the summary. Trade-offs a hardener flags at any step now appear in the final implement-plan summary.
• Removed the stale "never chain commands with `&&`" rule from the execution agents.

v1.8.7

Fixed: /implement-plan argument handling and plan hand-off

• The command resolves its argument to a concrete plan path once, so it works whether you pass a path, a natural-language reference ("the plan we just wrote"), or nothing (it locates the plan in plans/, or asks). Every pass uses the resolved path.
• The implementer receives the plan by path and reads it in fresh context, like the hardener and final reviewer, instead of having the full plan pasted into its prompt each step. The orchestrator stays lightweight across the run.

v1.8.6

New: due-diligence pass in /write-plan

A new isolated pass runs after the standards review and before step breakdown, via a new sdd-plan-diligence agent. It:

• Web-verifies the external facts a plan cites (library and CI action versions, third-party API contracts, platform behavior) against live sources, fixing outright errors and flagging valid-but-not-latest ones for you to decide.
• Surfaces the risks and decisions you would implicitly approve by executing the plan, split into a FIXED bucket (factual corrections it applied) and a REQUIRES YOUR JUDGMENT bucket (nothing it will touch on its own).
• Collects <!-- REVIEW: --> markers left by earlier passes so they reach you before implementation.

/write-plan now also proposes committing the plan at the end, so it exists in git before /implement-plan.

Improved: consistent verification-command discovery (1.8.5)

Every agent that runs tests, lint, or typecheck now discovers project commands from the same set of config files, with coverage extended to more stacks (JVM, PHP, Ruby, .NET, Deno).

v1.8.4

Current stable release