fix(has_context): skip persisting a generation paused for user input

[TonsOfFun]

Closes #9

Why

The framework is about to gain a way to pause a generation for user input, so a tool call can ask the user a question or wait for approval and then continue. A paused response ends on an unfinished turn: an assistant tool-call turn or a tool result. HasContext would record that turn as a generation and an assistant row. When the generation resumes, its prompt replays the stored conversation, so the after_prompt callback would store the user turn a second time.

This change makes HasContext handle both sides of a pause. Both checks are duck-typed, so framework releases that cannot pause behave exactly as before, and nothing here depends on an unreleased activeagent.

What changes

• persist_generation_to_context returns early, with an info log, when the response answers awaiting_input? with true. The paused response writes no generation record, assistant row or tool rows.
• A generation that pauses still persists its prompt, the user turn. This is a deliberate departure from the earlier proposal to skip both callbacks on pause: the user turn should survive a request that is declined or never answered.
• persist_prompt_to_context returns early while resuming_generation? is true. The paused run already stored the user turn, and the resumed prompt replays it.
• New private resuming_generation?. It is true when the agent defines a public resuming? (the framework flag) that returns truthy, or after self.resuming_generation = true. The private writer is for a host that replays a stored conversation itself and sets the flag from an action or a callback.
  • Both methods are private because every public method on an agent becomes one of its actions, and the action list feeds the framework's release digest.
  • No prompt option is used, because unknown prompt_options keys reach the provider request.
  • solid_agent does not define resuming?, so a framework Base#resuming? is never overridden.
• persist_tool_messages_to_context dedupes by tool_call_id within one response's stack as well as against rows already on the context. A resumed stack repeats the restored tool results, so each is stored once even when the context's messages scope cannot see rows written earlier in the same pass.
• README (HasContext section):
  • what a paused generation persists and what the resumed one persists;
  • that the resumed generation must load the paused generation's context, through a contextual: param naming the same record or load_<name>(context_id:) in the action. Otherwise each agent instance creates its own anonymous context.
• CHANGELOG (Unreleased): "Added" for resuming_generation? and its writer; "Fixed" for the persistence change and the in-stack dedupe. The version is not bumped.

Tests

New file test/solid_agent/has_context_paused_generation_test.rb (13 tests, using fake response and context objects). It covers:

• the resume flag and its private visibility;
• prompt persistence for fresh and resumed generations;
• a paused generation persisting its prompt but not its response;
• paused versus non-paused responses;
• in-stack tool dedupe;
• a full pause-then-resume run that leaves exactly the user turn, both tool results and the final answer.

All runs on Ruby 3.4.8:

• bundle exec ruby -Itest -Ilib test/solid_agent/has_context_paused_generation_test.rb gives 13 runs, 28 assertions, 0 failures, 0 errors, 0 skips. Against the origin/main library the same file gives 13 runs, 4 failures, 5 errors.
• bundle exec rake test gives 251 runs, 595 assertions, 0 failures, 0 errors, 0 skips. Baseline on origin/main: 238 runs, 567 assertions, 0 failures, 0 errors.
• bundle exec rake test:records gives 195 runs, 494 assertions, 0 failures, 0 errors, 0 skips (same as baseline).
• RuboCop with rails-omakase on the changed Ruby files finds no offenses. The repo has no RuboCop config, so I used a temporary inherit_gem: { rubocop-rails-omakase: rubocop.yml } config.

Not run:

• The suite under Ruby 4.0, where require "ostruct" in test_helper fails. That is unrelated to this change, and CI tests 3.2 to 3.4.
• activeagent's cross-repo solid_agent integration suite, after the review fixes. It was unchanged when run against the first commit, and the fixes only change method visibility and docs.

🤖 Generated with Claude Code

[superconductor-for-github]

⏳ Superconductor is working — View implementation

---

I'll get back to you soon!