Summary

Emits a counter at every agent-loop termination so operators can finally alert on non-success exits.
Before this, the agent loop recorded no metric when it aborted on repeated tool failures, max iterations, a loop-guard circuit break, or a provider content-filter — the only signal was reading transcripts.
A cron / trigger fire that aborted on repeated tool failures still recorded librefang_cron_fires_total{outcome="ok"} because the loop returned (did not throw).

New metric:

librefang_agent_loop_exits_total{agent, reason}

Acceptance alert (per #6227): rate(librefang_agent_loop_exits_total{reason!="completed"}) > 0 per agent.

Reason set

Matched to the real termination branches in run_streaming.rs, mod.rs, and loop_guard.rs:

• completed — any Ok(_) return: a finalized EndTurn reply, a silent completion (NO_REPLY / cascade-leak / progress-leak / mid-stream cascade-leak abort), a MaxTokens partial return, an interrupt-cancel early return, or the provider-not-configured early return.
• max_iterations — Err(MaxIterationsExceeded): the for loop ran out.
• repeated_tool_failures — Err(RepeatedToolFailures): consecutive_all_failed >= MAX_CONSECUTIVE_ALL_FAILED.
• circuit_break — loop-guard global circuit breaker tripped.
• content_filtered — Err(ContentFiltered): provider safety / content filter blocked the reply.
• error — any other propagated Err (LLM call failure, memory error, etc.).

Deviations from the issue's suggested set (documented per the task)

• No empty_response reason. The issue listed it, but there is no distinct empty-response termination branch: an empty LLM response is a one-shot in-loop retry (EndTurnRetry::EmptyResponse → continue); the turn then either retries to a real reply or finalizes via finalize_end_turn_text, both of which land on completed.
• Added content_filtered. ContentFiltered is a real, distinct terminal Err branch operators will want to separate from generic error.

How exactly-one increment is guaranteed

Structural, not by discipline at each site:

• run_agent_loop and run_agent_loop_streaming become thin instrumented wrappers (the #[instrument] span name and fields are unchanged, so the instrument_span_fields / CLI span-name tests stay green).
• Each wrapper calls its *_inner body once and feeds the single returned Result to record_agent_loop_exit(agent, &result).
• classify_exit_reason maps that one Result to one reason; the counter is incremented once.
• No terminal site inside the loop touches the metric, so there is no path where one branch falls through into another and double-counts. ?-propagated exits (circuit-break, LLM/memory errors) are caught at the wrapper too — they would have been missed by per-return-site increments.

Circuit-break surfaces as Err(Internal(msg)) propagated up through the ? on execute_single_tool_call. It is classified by matching the shared loop_guard::CIRCUIT_BREAKER_MSG_PREFIX constant — a single source of truth now referenced by both the producer (LoopGuard::check) and the classifier, so the match is compiler-anchored rather than a loose duplicated string literal.

Label cardinality

The agent label is sanitized (control chars stripped, length capped at 64) by sanitize_agent_label, mirroring sanitize_tool_label in tool_call.rs and the bounded-cardinality contract on record_cron_fire_metric in the kernel's cron.rs.

Changes

• crates/librefang-runtime/src/agent_loop/mod.rs — sanitize_agent_label, classify_exit_reason, record_agent_loop_exit; run_agent_loop split into instrumented wrapper + run_agent_loop_inner.
• crates/librefang-runtime/src/agent_loop/run_streaming.rs — run_agent_loop_streaming split into instrumented wrapper + run_agent_loop_streaming_inner, calling super::record_agent_loop_exit.
• crates/librefang-runtime/src/loop_guard.rs — extract CIRCUIT_BREAKER_MSG_PREFIX const; use it in the circuit-break message.
• crates/librefang-runtime/src/agent_loop/tests/utilities.rs — three injection-site tests (below).
• CHANGELOG.md — [Unreleased] / Added entry.

Tests

Added at the injection site (agent_loop::tests::utilities):

• test_classify_exit_reason_covers_every_branch — pins each reason mapping (completed, max_iterations, repeated_tool_failures, content_filtered, circuit_break, error) so a future error-variant or branch change can't silently re-bucket a termination.
• test_record_agent_loop_exit_increments_once_with_labels — DebuggingRecorder (same pattern as the existing record_tool_call_metric tests) asserts the counter increments exactly once with the right agent and reason labels.
• test_sanitize_agent_label_strips_control_and_caps_length — cardinality guard.

Verification

All run in the sanctioned librefang-rust-dev Docker image (no native cargo on this host), scoped to librefang-runtime:

• cargo check -p librefang-runtime --lib — Finished, clean.
• cargo clippy -p librefang-runtime --all-targets -- -D warnings — Finished, zero warnings.
• mold -run cargo test -p librefang-runtime --lib -- test_classify_exit_reason test_record_agent_loop_exit test_sanitize_agent_label — 3 passed; 0 failed.
• mold -run cargo test -p librefang-runtime --lib agent_loop::tests::utilities — 70 passed; 0 failed.
• cargo fmt -p librefang-runtime applied.

Live LLM verification is not required: this change does not alter any LLM call path or prompt/metering wiring; it only adds an observability counter classified from the loop's existing return value, fully covered by the unit tests above.

Closes #6227