feat: auto-continue after max-iteration exhaustion

[stefanpieter]

Summary

• add config-gated auto-continue when a turn exhausts agent.max_turns
• preserve truthful cumulative api_calls accounting across auto-continue cycles
• mark synthetic continuation prompts clearly and skip them in CLI, gateway, and TUI retry/undo flows
• document the new config in defaults/example config and add focused regressions

Motivation

Long autonomous workflows were stopping at the iteration/tool-turn ceiling only to require the user to say "continue". This patch lets Hermes continue boundedly when configured, while preserving the existing fallback summary path once the auto-continue cap is reached.

Validation

• pytest tests/run_agent/test_run_agent.py tests/cli/test_cli_retry.py tests/gateway/test_retry_replacement.py tests/tui_gateway/test_protocol.py -k "auto_continue or retry" -q
• pytest tests/run_agent/test_run_agent.py::TestHandleMaxIterations::test_auto_continue_helper_injects_prompt_and_resets_budget tests/run_agent/test_run_agent.py::TestRunConversation::test_auto_continue_on_iteration_exhaustion_resumes_without_summary -q
• python -m py_compile agent/chat_completion_helpers.py agent/conversation_loop.py cli.py gateway/run.py tui_gateway/server.py hermes_cli/config.py run_agent.py

Independent review

• bounded fresh-context review: PASS
• verified the prior retry/undo regression and api_calls accounting regression are fixed

[alt-glitch]

Implements feature request #16004. Related: #18607 (emergency compression before max_iterations), #30709 (iteration_budget reset on cached agent reuse).

[teknium1]

Thanks for the bounded, config-gated continuation design and the retry/undo coverage. The feature remains relevant: current main still moves exhausted turns to the tool-less summary fallback in agent/turn_finalizer.py:82-97.

Problems

• agent/chat_completion_helpers.py:127 appends the continuation as a user message. After a tool-calling iteration exhausts the budget, the tail can be tool → user. Current main explicitly repairs that malformed shape before requests (agent/conversation_loop.py:773-790), and documents strict-provider failures for it in agent/message_sanitization.py:282-311. Existing synthetic continuations preserve alternation by first appending an assistant turn (agent/conversation_loop.py:4927-4944).
• The loop was refactored after this branch: budget initialization now lives in agent/turn_context.py:222-259, while exhaustion fallback and completion semantics live in agent/turn_finalizer.py:54-157. The counter reset needs to be reworked through those seams.

Suggested changes

• Rebase the implementation concept onto the current continuation/finalization seams, preserving a valid assistant/tool/user sequence before the next API request, and test an exhausted tool-result tail.
• Keep total versus per-cycle API-call accounting explicit through finalize_turn.

Automated hermes-sweeper review.

[stefanpieter]

Current-main remediation pushed at a2a9e60b79cc7eee867f51bf82bb36cef60d5cc6.

What changed:

• Reimplemented bounded max-iteration continuation on the current turn-finalizer/conversation-loop architecture.
• Opens a fresh provider-safe continuation boundary rather than appending user after a trailing tool result.
• Preserves the original user target/session semantics across CLI, TUI, gateway, and resumed sessions; replacement prompts do not become the durable target.
• Clears stale continuation state on completion/error/cleanup and enforces the configured continuation cap.
• Exposes the current config through CLI/schema and Desktop controls.

Verification on the final tree merged onto current main (36f2a966c7f9f69987494b867c3dcf96b69a5766):

• 951 Python focused/integration tests passed.
• Desktop auto-continue tests: 2 passed; Desktop TypeScript typecheck passed.
• Ruff and Python compilation passed.
• Exact feature diff: 24 paths; SHA-256 f98e1109ce16daa84fe2af4da1e48f3d60c5e478c52268d852f350e9993775da.
• Independent exact-diff review: PASS with no security or logic blockers.
• PR head is live and mergeable.

[stefanpieter]

Final current-main sync: head 1c6e894165ee62d3dceef21b008ee7a7e6b95a85 is merged onto upstream main at 0385e155444d36203a6ef431f635223d78490d34. The exact 24-path feature diff and SHA-256 remain unchanged (f98e1109ce16daa84fe2af4da1e48f3d60c5e478c52268d852f350e9993775da). Reverification on this final tree: 951 Python tests passed, 2 Desktop tests passed, Desktop typecheck passed, and Ruff/Python compilation passed.

[stefanpieter]

@teknium1 current-main merge blocker is resolved on exact head d51a630553210effad4786a9cebf0235e45f1f98.

• Merged upstream/main f4df260f26c93f15694698869f3ea8e965eea301 into local/hermes-auto-continue-max-turns with a normal fast-forward branch update (no history rewrite / no force-push).
• Resolved the two conflict sites by preserving both behaviors:
  • conversation_loop -> finalize_turn(...) now passes _pending_verification_response_previewed, _defer_iteration_limit_fallback, and total_api_call_count together.
  • finalize_turn() keeps max-iteration continuation tool-tail closure and verification synthetic-nudge stripping before persistence checks continue.
• Verification on the merged tree:
  • scripts/run_tests.sh tests/agent/test_turn_finalizer_iteration_limit_exit.py tests/agent/test_turn_finalizer_cleanup_guard.py tests/run_agent/test_verification_continuation_budget.py → 35 tests passed, 0 failed
  • python3 -m py_compile agent/conversation_loop.py agent/turn_finalizer.py → pass
  • git diff --check → pass
  • ruff check agent/conversation_loop.py agent/turn_finalizer.py → All checks passed!
  • added-line security scan on the resolved hunks → no findings
  • fresh exact-scope independent review on the resolved conflict regions → PASS
• Current live PR state: headRefOid=d51a630553210effad4786a9cebf0235e45f1f98, mergeable=MERGEABLE, checks: none reported on this branch.

Please re-review the current head if the branch still looks useful on top of current main.