You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
While these issues mention different triggers, they all involve the TUI becoming unresponsive. The specific behavior after compaction might be a distinct angle worth investigating separately, or could be related to the broader responsiveness issues being tracked in those tickets.
Experiencing the same issue. It shows up after compaction. If it helps I am using Opus 4.6 and had a similar issue with Opus 4.5. And like OP, keep going fixes it and the model keeps on iterating. But would be nice to not have to resort to manually asking the model to continue.
I am experiencing the same issue. I am using GitHub Copilot, and I have less context than with other providers. Most of my tasks that are delegated to the agents end in the middle of the exploration. The agent explores the codebase, the context becomes full before it completes the exploration, it writes a compaction, and ends the task. It causes multiple round trips of agent task delegation.
One more voice I'm on OpenCode 1.2.15 but 1.2.4 was the same. Like others here Im using Opus 4.6 via Github Copilot. Exactly the same issues with compaction and havint to prod, the amount and regularity of the compacting is a bit random too. Sometimes it compacts whilst planning which is extremely frustrating.
@2fns nailed the shape back in Feb: "it writes a compaction, and ends the task." With @gergo-hortobagyi now reproducing on GPT-5.4 (1.14.18), plus Opus 4.5/4.6, Copilot-Sonnet, local qwen3.5 and minimax m2.5 — that breadth across models/providers usually points at a harness-level invariant, not a model quirk.
One diagnostic framing: after auto-compaction the summary becomes the trailing item. Models are trained to read summary-as-last-item as the terminal state of a turn — so without a trailing user-message-like anchor or an explicit continuation signal, stopping is the expected behavior. "keep going" works because it restores the anchor.
One documented pipeline splits the policy by trigger. Mid-turn compaction inserts initial context before the last real user message, with precedence last_real_user_index ?? last_user_or_summary_index ?? last_compaction_index. Pre-turn/manual uses DoNotInject and defers to the next regular turn. The invariant: mid-turn compaction must preserve a trailing user-anchor or inject a continuation hint — never leave the summary as the final item.
Cheap verification: log the exact prompt sent on the first post-compaction API call. If the last message is the summary with nothing after it, that's the mechanism.
Activity
github-actions commentedon Feb 11, 2026
This issue might be related to existing TUI responsiveness problems. Please check if your issue is similar to:
While these issues mention different triggers, they all involve the TUI becoming unresponsive. The specific behavior after compaction might be a distinct angle worth investigating separately, or could be related to the broader responsiveness issues being tracked in those tickets.
yasoob commentedon Feb 26, 2026
Experiencing the same issue. It shows up after compaction. If it helps I am using Opus 4.6 and had a similar issue with Opus 4.5. And like OP,
keep goingfixes it and the model keeps on iterating. But would be nice to not have to resort to manually asking the model to continue.2fns commentedon Feb 26, 2026
I am experiencing the same issue. I am using GitHub Copilot, and I have less context than with other providers. Most of my tasks that are delegated to the agents end in the middle of the exploration. The agent explores the codebase, the context becomes full before it completes the exploration, it writes a compaction, and ends the task. It causes multiple round trips of agent task delegation.
pschork commentedon Feb 28, 2026
Experiencing the same issue when using local llama.cpp serving qwen3.5_35b with 16384 context size.
viralbitcraft commentedon Mar 1, 2026
Experiencing the same issue when using local ik llama cpp serving minimax m2.5 with 196k context size.
TUI is responsive, just the agent is not proceeding after that.
opencode --version
1.2.10
NicolaiOksen commentedon Mar 1, 2026
Seeing the same - Version 1.2.15.
using Github Copilot - Opus 4.6
MattOates commentedon Mar 1, 2026
One more voice I'm on OpenCode 1.2.15 but 1.2.4 was the same. Like others here Im using Opus 4.6 via Github Copilot. Exactly the same issues with compaction and havint to prod, the amount and regularity of the compacting is a bit random too. Sometimes it compacts whilst planning which is extremely frustrating.
Mte90 commentedon Mar 4, 2026
same issue
l0seq commentedon Mar 12, 2026
Same issue, i use Sonnet 4.6 with github copilot
fix: improve compaction continuation to prevent agent from stopping
zjjott commentedon Mar 17, 2026
+1
sunhs commentedon Apr 16, 2026
+1
It's been happening regularly. I start to be afraid of using opencode with copilot.
Any info I can provide to help address the issue?
gergo-hortobagyi commentedon Apr 20, 2026
+1
GPT-5.4, using 1.14.18 on MacOS 26.3.1
dhruvkej9 commentedon Apr 20, 2026
@rekram1-node are we getting it solved?
AlexKenbo commentedon Apr 20, 2026
@2fns nailed the shape back in Feb: "it writes a compaction, and ends the task." With @gergo-hortobagyi now reproducing on GPT-5.4 (1.14.18), plus Opus 4.5/4.6, Copilot-Sonnet, local qwen3.5 and minimax m2.5 — that breadth across models/providers usually points at a harness-level invariant, not a model quirk.
One diagnostic framing: after auto-compaction the summary becomes the trailing item. Models are trained to read summary-as-last-item as the terminal state of a turn — so without a trailing user-message-like anchor or an explicit continuation signal, stopping is the expected behavior. "keep going" works because it restores the anchor.
One documented pipeline splits the policy by trigger. Mid-turn compaction inserts initial context before the last real user message, with precedence
last_real_user_index ?? last_user_or_summary_index ?? last_compaction_index. Pre-turn/manual usesDoNotInjectand defers to the next regular turn. The invariant: mid-turn compaction must preserve a trailing user-anchor or inject a continuation hint — never leave the summary as the final item.Cheap verification: log the exact prompt sent on the first post-compaction API call. If the last message is the summary with nothing after it, that's the mechanism.
Rules: https://github.com/AlexKenbo/codex-harness-internals/blob/main/SRC-0011/rules.json (
initial-context-injection-policy,initial-context-insertion-precedence).— @AlexKenbo
github-actions commentedon Jul 3, 2026
To stay organized issues are automatically closed after 60 days of no activity. If the issue is still relevant please open a new one.
ilarramendi commentedon Jul 7, 2026
+1