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
Magic-context itself is model agnostic so we expect all models to run properly. However the quality of compartments, memory extraction, consolidation all depends on model's capabilities and intelligence. If you're already using a high iq main model like opus/gpt5.4 you can still use a lower historian because historian's output is relatively less crucial (I personally still use gpt-5.4 via github-copilot for historian because github-copilot is request based and however long historian needs to run doesn't change cost for me)
You can always experiment with different models and also check their background outputs via our dashboard by analyzing compartment/memory outputs etc.
This may be a bug, but I’ve noticed that whenever I switch between the build and plan agents, if they’re using different models, the usage panel on the right side of the TUI briefly 'shrinks' as though it’s reprocessing something. When both modes use the same model, that flicker does not happen.
My assumption is that this may be configuration-related, possibly because Opencode is re-reading the model settings when switching agents, and the two models may have different context window limits.
To reproduce, it's just to have two different models in build and plan. Then write something in plan agent, switch to build agent say something else and you will see the 'shrink'.
Activity
[-]Model Selection Question[/-][+]Questions[/+]ualtinok commentedon Apr 11, 2026
Magic-context itself is model agnostic so we expect all models to run properly. However the quality of compartments, memory extraction, consolidation all depends on model's capabilities and intelligence. If you're already using a high iq main model like opus/gpt5.4 you can still use a lower historian because historian's output is relatively less crucial (I personally still use gpt-5.4 via github-copilot for historian because github-copilot is request based and however long historian needs to run doesn't change cost for me)
You can always experiment with different models and also check their background outputs via our dashboard by analyzing compartment/memory outputs etc.
null-axiom commentedon Apr 11, 2026
This may be a bug, but I’ve noticed that whenever I switch between the build and plan agents, if they’re using different models, the usage panel on the right side of the TUI briefly 'shrinks' as though it’s reprocessing something. When both modes use the same model, that flicker does not happen.
My assumption is that this may be configuration-related, possibly because Opencode is re-reading the model settings when switching agents, and the two models may have different context window limits.
To reproduce, it's just to have two different models in build and plan. Then write something in plan agent, switch to build agent say something else and you will see the 'shrink'.
ualtinok commentedon Apr 12, 2026
That's the normal behaviour, as you expected, when context size is different it should show the usage differently.
fix: close council audit findings on budget derivation
fix(pi-plugin): close council audit findings (#1, #3, #4, #5, #7) + P…
perf: eliminate redundant tokenization across RPC and system-prompt h…
fix(plugin,scripts): apply council audit post-ship follow-up fixes
fix(pi): round-3 audit — post-commit entry-id ref-map, cutover stamp …
fix(pi): round-4 audit — first-pass reset parity, post-splice ref-mis…
fix(pi): round-7 cluster — model-change-across-restart, marker rehydr…
fix(pi): round-8 cluster — marker isolation, contention fallback, sig…
fix: round-3 council findings (seq=0 m0 gap, dreamer archived underco…
fix: full-codebase audit findings (data-integrity, fail-closed, secur…
14 remaining items
fix(pi): round-4 audit — first-pass reset parity, post-splice ref-mis…
fix(pi): round-7 cluster — model-change-across-restart, marker rehydr…
fix(pi): round-8 cluster — marker isolation, contention fallback, sig…
fix: round-3 council findings (seq=0 m0 gap, dreamer archived underco…
fix: full-codebase audit findings (data-integrity, fail-closed, secur…
fix(audit): per-session map leak, FTS phantom-UNIQUE dedup, column gu…
style(plugin): biome format/import-order on audit-round files + drop …
fix(audit): full-codebase council round — 9 real fixes + false-positi…
fix(cache): resolve transform/cache-stability council audit (5 real b…
fix: resolve remaining v0.22.5 audit findings (cortexkit#5, cortexkit#6…
fix: release-audit batch — node:sqlite CI coverage, embedding-runtime…
Fix round-3 audit blockers: Pi grounding gate (cortexkit#2), config-l…
fix: scope dreamer decay by project, session leak, XML escaping, leas…
perf: add embedding cache with FTS pre-filtering and scoped session r…
fix(parity): hunt #7 — operator-surface truth: rust-authority migrati…