feat: Switch compaction model from session model - #3

Open

ryangamerdev wants to merge 23 commits into

devfrom

feature/compaction-model

Commits on Mar 15, 2026

feat: add collapse/float compaction modes and knowledge pack support
ryangamerdev
committed
18cc1dc
feat: add minFloat threshold to float compaction mode
ryangamerdev
committed
0fcad9c
fix: use delta token accounting in float sub-collapse minFloat gate
ryangamerdev
committed
12d6635
fix: persist token adjustment to database after each float sub-collapse
ryangamerdev
committed
7c1257b
feat: propagate parent session knowledge packs into subagent sessions
ryangamerdev
committed
bd7c79c
feat: knowledge pack agent prompt overrides
ryangamerdev
committed
0ce7172
fix: restore prompt focus after any sidebar mouse interaction
ryangamerdev
committed
157fec9
fix: scope prompt refocus to knowledge pack interactions only
ryangamerdev
committed
89c2d2c
fix: detect full chain across mid-run user interjections in float/collapse compaction
ryangamerdev
committed
be69668
fix: inject knowledge pack messages into plugin transform hook and toModelMessages
ryangamerdev
committed
abc9f93
fix: resolve sidebar click race condition causing TUI input lockup
ryangamerdev
committed
efbe9cb
fix: defer refetchActive after server responds to update KP enabled state
ryangamerdev
committed
632ff80
fix: mirror global knowledge packs to project config on sidebar add/remove

Ryan Wyler
authored and
ryangamerdev
committed
fbb4acb
fix: write project config to .opencode/opencode.json not config.json

Ryan Wyler
authored and
ryangamerdev
committed
83d3939
fix: guard removeProjectKnowledgePack against unnecessary writes

Ryan Wyler
authored and
ryangamerdev
committed
8fccbb4
fix: run float pre-check on stop finish before exiting prompt loop

Ryan Wyler
authored and
ryangamerdev
committed
fefe0e8
fix: use most recent non-zero-token assistant for float/overflow token count

Ryan Wyler
authored and
ryangamerdev
committed
bfbd009
Revert "fix: use most recent non-zero-token assistant for float/overflow token count"

Ryan Wyler
authored and
ryangamerdev
committed
a38e15a
fix: only absorb user message as mid-run interjection when prev assistant has finish=tool-calls

Ryan Wyler
authored and
ryangamerdev
committed
f5a18a3
Updated

Ryan Wyler
authored and
ryangamerdev
committed
7abce5b
docs: document debug logging added to detectChains in bea7aa2dd

Ryan Wyler
authored and
ryangamerdev
committed
370101d
fix: port 1.3 fixes - filterCompacted error guard and overflow explanation for collapse/float
ryangamerdev
committed
885c467
feat: add compaction model override for session compaction operations
ryangamerdev
committed
4edb177

Navigation Menu

Uh oh!

feat: Switch compaction model from session model - #3

feat: Switch compaction model from session model#3
ryangamerdev wants to merge 23 commits into
devfrom
feature/compaction-model

Commits on Mar 15, 2026

feat: add collapse/float compaction modes and knowledge pack support

feat: add minFloat threshold to float compaction mode

fix: use delta token accounting in float sub-collapse minFloat gate

fix: persist token adjustment to database after each float sub-collapse

feat: propagate parent session knowledge packs into subagent sessions

feat: knowledge pack agent prompt overrides

fix: restore prompt focus after any sidebar mouse interaction

fix: scope prompt refocus to knowledge pack interactions only

fix: detect full chain across mid-run user interjections in float/collapse compaction

fix: inject knowledge pack messages into plugin transform hook and toModelMessages

fix: resolve sidebar click race condition causing TUI input lockup

fix: defer refetchActive after server responds to update KP enabled state

fix: mirror global knowledge packs to project config on sidebar add/remove

fix: write project config to .opencode/opencode.json not config.json

fix: guard removeProjectKnowledgePack against unnecessary writes

fix: run float pre-check on stop finish before exiting prompt loop

fix: use most recent non-zero-token assistant for float/overflow token count

Revert "fix: use most recent non-zero-token assistant for float/overflow token count"

fix: only absorb user message as mid-run interjection when prev assistant has finish=tool-calls

Updated

docs: document debug logging added to detectChains in bea7aa2dd

fix: port 1.3 fixes - filterCompacted error guard and overflow explanation for collapse/float

feat: add compaction model override for session compaction operations