D154 — The generated steering block describes what this client received

Status: implemented (2026-08-28). Amends D65 and D141. Closes #175.

Context. The generated instructions described intended state rather than what provisioning achieved on that client, so any partially-supported client got confidently wrong instructions.

generateKBInstructions emitted, whenever the KB declared any agent: "Subagents installed by this KB: , — their descriptions are in the client's agent registry: delegate to them the tasks they cover." The list came from kbAgentNames(kbRoot) — what the KB declares on disk. It could not come from what the client received: the function has no provider argument and its only caller, BuildManifest, runs exclusively server-side, where no provider is known. Meanwhile whether a provider can receive an agent at all is a static fact the code already states — destinationMatrix["agent"] marks kiro and hermes unsupportedDest. So a KB with two agents synced to Claude Code and Kiro materialized them for the first and not at all for the second, while both were told they were installed. On the reporting deployment the only agents in that client's registry were stale files from an earlier bootstrap, still pointing at a decommissioned source.

Two corrections to the field report, made while implementing.

The omission was not silent, it was not persistent. printApplySummary does print unsupported: agent/<name> … (kind has no destination for this provider) for every artifact routed to Unsupported. What is missing is persistence: the line appears only on a run where the artifact enters the diff, doctor never reported the condition, and the steering sentence was wrong regardless of any output — which is the part that misleads the model rather than the operator.

"Written once per mounted KB" is a reporting defect, not three writes. applyInstructionsGroup accumulates every KB's snippet and calls writeInstructionsBlock once; it then appends one ManagedFile per instructions artifact — i.e. per KB — and the summary printed one line each. Nothing was rewritten N−1 times, so the fix belongs in the printer.

Decision.

• The subagent sentence moves out of the hashed artifact and is emitted client-side, per provider. There is an exact precedent in the same function: the D75 WP4 local-paths table is appended there, client-side only, and deliberately not part of the artifact hash. The sentence has the same nature — a statement about this client, not about the KB.
• It lists only agents whose kind has a destination for the provider, and only those actually authorized: an artifact awaiting approval was not installed, and the whole point is that the sentence describes what happened.
• The rewrite trigger now fires on an agent artifact too. The block's rendered content depends on the agent set while its artifact hash deliberately does not, so without this the sentence would go stale the moment an agent was added or removed. The test that asserted the hash changes when an agent is added now asserts the opposite, and names the trigger as the guarantee that replaced it — the invariant moved, it did not disappear.
• A declared kind the provider cannot receive is warned about on every run, computed from the manifest rather than the diff.
• The wrapper becomes opt-out, not translated. instructions.md may declare <!-- cartographer: preamble: none --> on its first line to suppress the three generated bullets — the redundant, hardcoded-English part. Deliberately not a localisation mechanism: a language: key would make Cartographer carry translations of its own prose forever, so the KB owns the prose instead. The one-line routing sentence stays generated in every case (KB name, server, archives): it is state, not prose, and no KB can write it for itself. The residual is therefore one English line rather than none, and this entry says so rather than claiming the language break is gone. The directive is honoured on the first line only, so a KB can document it without triggering it — the same trap as the placeholder syntax.
• Reporting collapses per path for the instructions kind only. result.Written keeps one entry per artifact, because the lockfile and drift detection depend on it; elsewhere two artifacts never share a path and a blanket dedup would hide a real bug.

Consequences. The steering block changes content on clients that cannot receive agents — the false sentence disappears — so those clients rewrite the block once on the next sync, which is the normal drift path. New KB-side directive; no config migration.