Feature request
Consider making manualMode.enabled: true the default (or otherwise making autonomous compression opt-in), because the current default lets the agent compress context on its own, and in practice that sometimes destroys content the user still needed.
The problem
In automatic mode the agent decides both when and what to compress. Under context pressure (DCP's own nudges push it harder past maxContextLimit), the model tends to make increasingly aggressive and imprecise choices — compressing ranges that are not actually "closed and stale", losing details the user was still actively working with. Once a bad compression lands, recovering attention to the pruned content is hard even with /dcp decompress, because the user has to notice the loss first.
Real-world data point: my config already pushes maxContextLimit to 180k / minContextLimit to 120k specifically to fight over-eager autonomous compression on a 200k-window model, and I still hit cases where the agent compressed things it shouldn't have. Manual mode is currently the only way to fully stop autonomous compression, but it's buried three levels deep in docs (compress.permission: "deny" and manualMode.enabled are non-obvious alternatives with different trade-offs).
Why Manual mode is a good default candidate
With manualMode.automaticStrategies: true (also default), Manual mode keeps all the harmless automatic cleanup running:
- deduplication still runs
- error purging still runs
- nudges/stats/panel still work
The only thing lost is autonomous range/message compression — which is exactly the destructive part. The user keeps an explicit trigger (/dcp-compress [focus]) that also produces better results, since human-directed focus text gives the model a much clearer target than self-initiated compression under nudge pressure.
Alternatives considered
- Default
compress.permission: "ask" — safer, but adds a permission prompt on every compression; noisy for users who are happy with auto mode. - First-run notification — mention Manual mode prominently in DCP's first-run chat notification so users know the off-switch exists.
- Softer nudges by default — doesn't fully prevent bad autonomous decisions, just delays them.
Option 2 (or changing the default outright) would be cheap and would prevent the worst failure mode. Happy to send a PR for whichever direction you prefer.
Feature request
Consider making
manualMode.enabled: truethe default (or otherwise making autonomous compression opt-in), because the current default lets the agent compress context on its own, and in practice that sometimes destroys content the user still needed.The problem
In automatic mode the agent decides both when and what to compress. Under context pressure (DCP's own nudges push it harder past
maxContextLimit), the model tends to make increasingly aggressive and imprecise choices — compressing ranges that are not actually "closed and stale", losing details the user was still actively working with. Once a bad compression lands, recovering attention to the pruned content is hard even with/dcp decompress, because the user has to notice the loss first.Real-world data point: my config already pushes
maxContextLimitto 180k /minContextLimitto 120k specifically to fight over-eager autonomous compression on a 200k-window model, and I still hit cases where the agent compressed things it shouldn't have. Manual mode is currently the only way to fully stop autonomous compression, but it's buried three levels deep in docs (compress.permission: "deny"andmanualMode.enabledare non-obvious alternatives with different trade-offs).Why Manual mode is a good default candidate
With
manualMode.automaticStrategies: true(also default), Manual mode keeps all the harmless automatic cleanup running:The only thing lost is autonomous range/message compression — which is exactly the destructive part. The user keeps an explicit trigger (
/dcp-compress [focus]) that also produces better results, since human-directed focus text gives the model a much clearer target than self-initiated compression under nudge pressure.Alternatives considered
compress.permission: "ask"— safer, but adds a permission prompt on every compression; noisy for users who are happy with auto mode.Option 2 (or changing the default outright) would be cheap and would prevent the worst failure mode. Happy to send a PR for whichever direction you prefer.