[Bug] /goal keeps hitting every 2-3 seconds

[alvinycheung]

Affected area

Coding agent and CLI

What happened?

Output:

 Luna is actively verifying the dotfun-owned RichText, route, and deployment seams that must stay outside IWK. Workflow state is otherwise unchanged.

 Goal continuation · Execute the reviewed IRIS-2717 ledger at /Users/alvinycheung/tmp/wo... (Ctrl+O to expand)

 ✓ python · workflow_progress=await rlm_workflow.run(str(run_dir),max_paral… · ↑ 1 ↓ 22 lines · 29ms

 Contract analysis is still active; Sol implementation remains blocked.

 Goal continuation · Execute the reviewed IRIS-2717 ledger at /Users/alvinycheung/tmp/wo... (Ctrl+O to expand)

 ✓ python · workflow_progress=await rlm_workflow.run(str(run_dir),max_paral… · ↑ 1 ↓ 22 lines · 27ms

 No state change; Luna continues the Steps contract analysis.

 Goal continuation · Execute the reviewed IRIS-2717 ledger at /Users/alvinycheung/tmp/wo... (Ctrl+O to expand)

 ✓ python · workflow_progress=await rlm_workflow.run(str(run_dir),max_paral… · ↑ 1 ↓ 22 lines · 23ms

 Contract analysis remains active; workflow state is unchanged.

 Goal continuation · Execute the reviewed IRIS-2717 ledger at /Users/alvinycheung/tmp/wo... (Ctrl+O to expand)

 ✓ python · workflow_progress=await rlm_workflow.run(str(run_dir),max_paral… · ↑ 1 ↓ 22 lines · 26ms

 No transition yet; Luna continues the contract work.

 Goal continuation · Execute the reviewed IRIS-2717 ledger at /Users/alvinycheung/tmp/wo... (Ctrl+O to expand)

 ✓ python · workflow_progress=await rlm_workflow.run(str(run_dir),max_paral… · ↑ 1 ↓ 22 lines · 28ms

 Contract analysis remains active; Sol implementation is still blocked.

 Goal continuation · Execute the reviewed IRIS-2717 ledger at /Users/alvinycheung/tmp/wo... (Ctrl+O to expand)

Steps to reproduce

1. run prime-agent
2. set a goal
3. see it keep hitting repeatedly

Expected behavior

It shouldn't hit every 3 seconds, it's really expensive like that and the agent just got the message already anyway.
It looks like it's not really a timing thing, it's more like it's spamming /goal between every python command.

Prime Agent version

0.7.4

Environment

macOS Tahoe 26.6

Additional context

No response

[snimu]

Investigating, thanks for the report.

[alvinycheung]

Oh good, I saw a very claude-coded response in here earlier lol. Came back to try to redirect it but @snimu you're probably a person yeah?

Example 2

gpt-5.6-sol xhigh

It's set a goal for itself, and is waiting for another agent to let it know it's time to start working because the block of work it has planned is dependent on the other agent landing it's work first.

Here's the loop I noticed it looping in:

 The prerequisite release is still in progress. Production remains untouched.

 Goal continuation · Complete IRIS-2795 end to end after session 01a0252d-b977-70e9-ad79... (Ctrl+O to expand)

 Still waiting for READY — IRIS-2758 RELEASED. No production changes made.

 Goal continuation · Complete IRIS-2795 end to end after session 01a0252d-b977-70e9-ad79... (Ctrl+O to expand)

 The dependency gate remains closed. IRIS-2795 production work has not started.

 Goal continuation · Complete IRIS-2795 end to end after session 01a0252d-b977-70e9-ad79... (Ctrl+O to expand)

 ✓ bash · gh pr view 1977 --repo Dot-Fun/iris --json state,mergeStateStat… · ↑ 2 ↓ 7 lines · 580ms · (Ctrl+O to expand)

 No readiness signal yet. IRIS-2795 remains paused safely.

 Goal continuation · Complete IRIS-2795 end to end after session 01a0252d-b977-70e9-ad79... (Ctrl+O to expand)

 Still awaiting READY — IRIS-2758 RELEASED. No production changes made.

 Goal continuation · Complete IRIS-2795 end to end after session 01a0252d-b977-70e9-ad79... (Ctrl+O to expand)

 The dependency gate remains closed. Production stays unchanged.

 Goal continuation · Complete IRIS-2795 end to end after session 01a0252d-b977-70e9-ad79... (Ctrl+O to expand)

 No readiness handoff yet. IRIS-2795 remains safely paused.

 Goal continuation · Complete IRIS-2795 end to end after session 01a0252d-b977-70e9-ad79... (Ctrl+O to expand)

Maybe it'd be good to be able to give them the ability to pause their goal? I know codex has this... the problem with that, then, is I consistently see GPT-5.6-sol in Codex pause the goal so it can go do other things and it never actually gets back to the goal... Maybe a nontrivial solution. I have an agent that will respond to this thread as a comment once it figures out how /goal works in the other harnesses because I don't really have a good solution in mind for this :/ just seeing it as a problem. Hopefully the extra input helps ya'll to make a good decision about how you think prime-agent should work.

I'm really digging it as a harness. It's a massive upgrade from codex. 10/10

[alvinycheung]

Confirmed: Prime Agent 0.7.4 has a goal continuation bug. The reported 2–3 seconds are model latency, not a polling interval.

Root cause in Prime Agent 0.7.4

The agent loop calls getContinuationMessages() after each natural turn end. An active goal immediately returns another visible goal_context message. There is no delay, debounce, or continuation limit.

The 0.7.4 goal check does not treat live RLM descendants as active work. A parent can correctly start a child and end its turn to wait. Prime then starts another full-context model turn at once. That turn often reports that it is still waiting, ends, and starts the cycle again.

This also explains why the loop appeared after each Python call. The hook does not run after the cell itself. It runs after the model receives the cell result and ends its next response.

#1610 fixes this exact case and is now merged. It defers goal continuation while any RLM descendant remains unsettled. It admits one continuation after the last descendant settles. The latest release is still 0.7.4, so the fix is not in the installed release yet.

Comparison with other installed clients

I traced the installed builds and matching source for each client.

Client	When it continues	Durable state	How it stops	Budget
Prime Agent 0.7.4	At each natural turn end. The bug ignores live RLM descendants.	A thread_goal_state entry in the session JSONL.	The model calls await goal.complete(). The user can pause, resume, clear, or stop it.	Optional token budget. It counts input and output tokens.
Claude Code 2.1.235	A session Stop hook runs when Claude tries to stop. Background tasks defer the check.	goal_status entries in the same session transcript.	A separate evaluator accepts success or impossibility. The user can clear, stop, or cancel it.	No goal token budget. The UI count is output tokens used since goal start.
AGY 1.1.17	A built-in stop hook continues unless the response has a goal sentinel.	A protobuf GoalState in the conversation SQLite trajectory.	The model emits <!-- GOAL_COMPLETE --> or <!-- GOAL_CANCELLED -->.	No current iteration limit or goal token budget.
Codex 0.148.0	The goal extension starts another turn when the thread becomes idle.	thread_goals in ~/.codex/goals_1.sqlite.	The model calls update_goal with complete or blocked. The user can pause, resume, edit, or clear it.	Optional token budget. It counts uncached input plus output tokens.

Claude uses a small evaluator instead of trusting a completion marker. It also knows when background tasks still run. Its default deferred check-in is 30 minutes.

AGY uses a deterministic sentinel contract and a five-point completion audit. Its state has active, completed, and cancelled values. Version 1.0.14 removed the old maximum, so a goal can run until completion or cancellation.

Codex has the most detailed durable lifecycle. Its states include active, paused, blocked, usage_limited, budget_limited, and complete. Its prompt says that one blocker must persist for three goal turns before the model marks the goal blocked.

One AGY detail explains a help mismatch. agy -p /help does not list /goal, but the compiled interactive slash registry does. The command is interactive-only and uses /goal <free-form task>.

The main design difference for this bug is background-work gating. Claude already defers while background tasks run. Prime 0.7.4 continued as soon as the parent became idle, even when its children were still working. PR #1610 adds that missing gate.

[alvinycheung]

@snimu looks like you already handled it, this can be closed.