v0.15.0 — a new engine underneath, and plans that survive real work

MandoCode's AI engine now runs on Microsoft Agent Framework. The chat layer, its tools, and plan execution have all moved off Semantic Kernel onto Microsoft's agent and workflow runtime — the largest change to the engine's internals since it was written.

Your models, providers, tools, skills, MCP servers, and approval prompts are unchanged, and nothing about how you use MandoCode changes. If none of the below sounds different to you, MandoCode will simply behave as it always has.

What the new foundation buys is plans that survive real work.

Install / upgrade

dotnet tool update -g MandoCode

New install: dotnet tool install -g MandoCode. Run mandocode --doctor to confirm the version and which runtime you're on.

Why this matters

• A plan no longer disappears with the process. If MandoCode closes after step two, the first two steps stay settled and the remaining work resumes instead of starting over. An unfinished plan is offered at startup with Resume or Discard.
• You stay in charge when the plan changes. /plan <goal> forces a plan on demand. Steps can be selected and edited in place before execution, and later steps refresh so a renamed file doesn't leave the rest of the plan stale. If execution disproves the plan, MandoCode proposes a revised remainder and waits for approval.
• Plans are faster and cheaper. A step used to be gated by a second model call judging whether it was really done. That's gone — a step now completes on the executor's own evidence against acceptance criteria it carries, with an explicit test-and-repair phase at the end. It halves the model calls in a plan run, and a plan can no longer finish green on work that was never exercised.
• Failures are reported honestly. Retry, skip, revise, and cancel are distinct decisions. A run that finishes after skipping failed work says so rather than reporting an unqualified success.
• .NET 10 and .NET 8 ship side by side. NuGet picks the match for your machine at install time. Nothing to choose or configure.
• .NET 8 stops getting security fixes in November 2026. MandoCode carries both builds through that window and drops the .NET 8 one afterwards, so the removal is announced well ahead of time.

New commands

Command	What it does
/plan <goal>	Force a plan for a goal, or show an unfinished one
/plan-resume	Continue an unfinished plan where it left off
/plan-discard	Forget an unfinished plan
/compact	Compress the conversation into a recap and keep going — unlike /clear, the thread of the work survives

/config: the planner and plannerEngine settings are retired (the workflow planner is always on), and strictPlanVerification is gone with the verifier it controlled. Existing config files carrying them still load fine — the keys are simply no longer read.

Also in this release

Fixed

• Long plans no longer stall when the host stops reading progress
• A failed verification response no longer discards completed implementation work — retries reuse evidence already gathered, confirmed valid by file hash
• Session token totals use provider-reported usage only, so reads and attachments no longer inflate the visible count
• A broad directory reference like @MandoCode/ can no longer flood a model with a recursive listing
• Browser tab and frame listings are always read live rather than answered from cache

Added

• Images can be delivered to models that can actually see them, with vision support detected rather than assumed — and images now count toward the context estimate
• Hosts can supply their own agent tools, and can watch shell commands run as they happen
• Automatic planning follows task shape instead of message length

Changed

• Dependencies updated, including Model Context Protocol 2.2.0 and YamlDotNet 18.1.0
• The startup banner keeps prerelease tags, so a test build no longer looks identical to the release it was cut from

Verification

733/733 tests passing on both .NET 8 and .NET 10. Coverage includes workflow topology, checkpoint storage, resume context, retry behaviour, plan revision, proposal handoff, scoped directory references, and bounded tool results. The planner was also exercised through real CLI and Desktop sessions, including forced process termination between steps.

Full detail in the changelog.