Compaction obscures session context and silently drops prior work

[roadrunner21]

The current compaction flow rewrites history so aggressively that it obscures the overall framing. The new inline auto-compaction just makes it even harder to notice. I’d like to gather experiences + ideas before proposing changes.

What’s happening

1. Compaction drops assistant context. Every run replaces the full history with a single bridge containing only user prompts plus a summary, so confirmations like “Bug X is fixed” vanish. That’s true for manual /compact and for the auto version.
2. The rewrite is easy to miss Codex does emit a quick “Compact task completed” message, but when a single turn goes past the limit, Codex compacts right away and resumes sampling, so the user rarely sees that anything happened. So sometimes it looks like Codex suddenly changed topics or started editing unrelated files
3. Very little feedback There’s no bridge preview or list of trimmed items, so the notification gives no real insight into what is actually delivered to the next session (which is why I opened issue #5366)
4. No first-class opt-out for the mid-turn compactor. The only workaround is to set model_auto_compact_token_limit to a massive number in config.toml, which isn’t documented and disables the safety entirely rather than just the mid-turn behavior.

Why it matters

• Multi-step work becomes fragile: once assistant confirmations vanish, Codex can redo finished tasks or edit files that should be left alone.
• Silent token drain: Codex can compact mid-turn, lose track of prior completions, then re-run expensive steps again, so the session chews through context even faster without actually moving forward.
• Lack of a mid-turn targeted opt-out means you either live with the surprises or disable auto compaction entirely.

Possible directions

I don’t have the perfect resolution, but here are a few ideas that might be worth exploring:

• Make compaction more transparent, e.g., show the bridge preview (see Emit a concise post‑Compact context message after manual Op::Compact #5366), highlight what got trimmed, maybe even require an explicit “continue” when the rewrite happens mid-turn. This is something that would have helped me in the past to get up to date with what the next session even knew.
• Add a targeted mid-turn opt-out flag or confirmation prompt so auto-compaction can be paused without disabling manual /compact.
• Carry over the last few assistant turns or let users mark steps as “done” to keep completed work in the bridge.
• Include a lightweight summary of the assistant’s outputs (not just user prompts) so both sides of the conversation survive compaction instead of reflecting only one side of the exchange

Curious to hear whether others have run into similar issues or have ideas for balancing safety vs. autonomy. Even incremental transparency might help, but it feels like we need a broader rethink of how compaction behaves, especially now that it can trigger mid-turn without any user input.

[imajes]

these are good thoughts. I have noticed similar behaviors also!

[Bud-Open]

I've encountered a similar situation where, after the compression is triggered, the recent conversation history is discarded, and it returns to the user's original prompt to start processing the already completed task.

[roadrunner21]

I wanted to update the discussion since codex has addressed some of the compaction behavior in the latest updates.

A rundown of what changed, what it means in practice, and what that means within the scope of this discussion:

What changed since v0.50.0

• You can customize the compaction prompt (feat: compaction prompt configurable #5959)

  • Both manual /compact and the automatic, inline compaction now respect a user‑defined prompt.

• The model sees a cleaner input during compaction (Filter out reasoning items from previous turns #5857)

  • When Codex creates the summary, it now filters out internal noise before sending content to the model.

• Conversation history keeps internal “snapshots” as breadcrumbs (aka ghost snapshots) (chore: undo nits #5631)

  • What this means: Codex now keeps lightweight receipts of important state (e.g., what files looked like after a step) inside the saved history when a compact happens.

• These snapshots are not shown as chat messages and are not sent back to the model. They don’t expand the model’s memory by themselves - they exist so the system can resume/fork more reliably after compaction.

How to use the new settings

Add these to your Codex config (in the profile you use). Exact layout may vary in your setup, but the keys are:

Example: set a custom compact prompt inline

[profile.my-workflow]
compact_prompt = """
Summarize for the next agent:
- Completed steps (with file paths)
- Outstanding TODOs (file:line) and blockers
- Decisions taken (why) and constraints (env/latency/tokens)
- What not to redo (explicitly list finished items)
One screen of bullets; be concise.
"""

Example: load the prompt from a file

[profile.my-workflow]
experimental_compact_prompt_file = "compact_prompt.txt"

How the changes stack up against the original concerns

• “Compaction drops assistant context”

  • Still true in spirit: the history is rewritten to a single bridge. The improvement is that you can now steer the summary to capture the key assistant confirmations you care about.

• “It’s hard to notice when a rewrite happened”

  • Slightly better: you’ll see a short notice when older items are trimmed, and you still get the “compact completed” message. There isn’t a preview or a “click to continue” step yet.

• “Very little feedback about what changed”

  • Still limited: there’s no bridge preview or explicit list of trimmed items, just the count of items trimmed when it happens.

• “No first‑class way to opt out of mid‑turn compaction”

  • Still the same: there’s now a clearer threshold you can tune, but there isn’t a targeted mid‑turn opt‑out or confirmation flow.

What’s still on the wish list (to be discussed)

• Preview the bridge (or at least show a concise post‑compact context message) so it’s obvious what the next turn will “know”.
• A targeted mid‑turn opt‑out or confirmation, so compaction can be paused without disabling the safety entirely.
• Carry a small number of recent assistant turns as first‑class items alongside the summary, not just inside it to avoid codex turning over the same issues again.