fix(agent): self-correct when the model calls an unknown tool

[walcz-de]

What

When the LLM selects a tool name that matches no available action, the agent now feeds the valid tool list back to the model so it can re-select an existing tool, instead of silently proceeding with a nil action.

Why

In consumeJob's WithToolCallBack, the chosen tool is resolved with allActions.Find(tc.Name). If the model emits a name that was never offered — common with local models, which sometimes hallucinate tool names — Find returns nil and the code falls through:

• execution proceeds with a nil action → the model gets an empty/unhelpful result, no signal to recover, and typically repeats the bad call or gives up;
• if an observer is attached, the decision path dereferences chosenAction.Definition() on a nil action → nil-pointer panic.

There was no path that told the model "that tool doesn't exist; here are the ones that do."

How

cogito already has the right primitive: ToolCallDecision.Adjustment re-runs tool selection with feedback (see tools.go). This change uses it:

• On an unknown tool name, return {Approved: true, Adjustment: "<name> doesn't exist; available tools are […]; re-issue using one of these"} so the model self-corrects.
• The correction is bounded (max 3 attempts per tool name); past that the call is Skipped, so a model that keeps hallucinating can't drive an unbounded re-selection loop (cogito's goto reprocessCallbacks).
• Built-in control verbs (stop / send_message / update_state) are dispatched by name and are explicitly excluded.
• The observer creation path is guarded against a nil action (fixes the latent nil-deref).

The decision logic is factored into a small pure helper, correctUnknownToolCall, with a hermetic unit test (no live LLM required).

Notes

• Approved: false is deliberately not used for this case — it raises ErrToolCallCallbackInterrupted and aborts the whole run; Adjustment is the mechanism designed for "let the model try again."
• Complements feat(agent): per-agent tool allow/deny-list to constrain the effective tool set #480 (per-agent allow/deny-list) and Per-agent option to exclude built-in knowledge-base search action (MCP-only agents); enable_kb=false does not remove it when KB is configured globally #479: an allow-list constrains which tools are offered, but cannot prevent the model from naming a tool that doesn't exist — this handles that case.

Test

go test ./core/agent/ -run TestCorrectUnknownToolCall -v   # passes, no LLM needed
go build ./core/agent/... && go vet ./core/agent/...        # clean

Assisted-by: Claude:claude-opus-4-8 [Claude Code]