feat(compaction): 以 history checkpoint 取代压缩覆盖层

.github/workflows/go-ci.yml

@@ -215,9 +215,9 @@ jobs:
           TEST_POSTGRES_BOOTSTRAP_SCHEMA=0 go test -v -race -count=1 -timeout 90s \
             -run '^TestPostgresClaim' ./internal/botworkspace
           TEST_POSTGRES_BOOTSTRAP_SCHEMA=0 go test -v -race -count=1 -timeout 90s \
-            -run '^TestPostgresDiscuss(RecoveryPreservesBatchAcrossCompactorRestart|BacklogOverBudgetIsAnsweredAndNextRoundContinues|BacklogWithoutCompactionIsAnsweredAndRecordsOmission|OversizedInputFailsOnceThenNextMessageIsAnswered|ImageBacklogKeepsImagesOnTheirMessages|OversizedImageInputFailsOnceThenNextMessageIsAnswered|ImageBacklogYieldsAnsweredHistoryFirst|CompactedImageInputIsRecorded|CompactionKeepsBatchTextOverImages)$' ./cmd/agent
+            -run '^TestPostgresDiscuss(RecoveryPreservesBatchAcrossCompactorRestart|BacklogOverBudgetIsAnsweredAndNextRoundContinues|OversizedInputFoldsByItsHeadAndTail|ImageBacklogKeepsImagesOnTheirMessages|OversizedImageInputFailsOnceThenNextMessageIsAnswered|ImageBacklogYieldsAnsweredHistoryFirst|CompactedImageInputIsRecorded|CompactionKeepsBatchTextOverImages)$' ./cmd/agent
           TEST_POSTGRES_BOOTSTRAP_SCHEMA=0 go test -v -race -count=1 -timeout 90s \
-            -run '^TestPostgres(SummaryOnlyRollup|CompleteCompactionRollup)' ./internal/chat/message
+            -run '^TestPostgresCommit' ./internal/chat/checkpoint
           TEST_POSTGRES_BOOTSTRAP_SCHEMA=0 go test -v -race -count=1 -timeout 90s \
             -run '^TestPostgresHistoryToolResultRef$' ./internal/agent/tool
       - name: Verify atomic connector disconnect and tenant isolation

.github/workflows/migrations.yml

@@ -16,6 +16,7 @@ on:
       - "internal/db/team_*_integration_test.go"
       - "internal/db/*postgres_integration_test.go"
       - "internal/chat/message/*postgres_integration_test.go"
+      - "internal/chat/checkpoint/**"
       - "internal/agent/view/*postgres_integration_test.go"
       - "internal/workspacedeps/*integration_test.go"
   pull_request:
@@ -33,6 +34,7 @@ on:
       - "internal/db/team_*_integration_test.go"
       - "internal/db/*postgres_integration_test.go"
       - "internal/chat/message/*postgres_integration_test.go"
+      - "internal/chat/checkpoint/**"
       - "internal/agent/view/*postgres_integration_test.go"
       - "internal/workspacedeps/*integration_test.go"
 
@@ -113,4 +115,4 @@ jobs:
         # web transcript depends on is checked end to end. Nothing else executes
         # them: `go test ./...` runs without a DSN and skips, and the fencing job
         # filters by test name.
-        run: go test -tags=integration -count=1 ./internal/db ./internal/chat/message ./internal/chat/thread ./internal/workspacedeps ./internal/agent/view
+        run: go test -tags=integration -count=1 ./internal/db ./internal/chat/message ./internal/chat/checkpoint ./internal/chat/thread ./internal/workspacedeps ./internal/agent/view

AGENTS.md

@@ -180,7 +180,7 @@ The `Human QA` section contains only one status checkbox, `- [ ] Human QA passed
 - `sdk.ProviderMetadata` is `namespace → name → string`. Memoh-owned annotations on tool-call parts (pending approval, user-input request, execution location) live in the `memoh` namespace through `internal/agent/partmeta`; `internal/messageconv` folds them back into the nested objects `bot_history_messages` has always stored, so the row format did not change with the SDK types.
 - **Tool usage lives with the tool, never in the static prompt.** Per-tool usage goes in `toolexec.Tool.Description`; cross-tool workflow guidance goes in an optional `tools.ToolUsage` `Usage()` method that `assembleTools` injects only when that provider registers tools for the session. `internal/agent/runtime/native/prompt_test.go` guards this.
 - Prompt templates are embedded from `internal/agent/runtime/native/prompts/`. Partials are prefixed with `_`; system prompting combines `system_common.md` with mode-specific prompts such as `mode_chat.md` and `mode_discuss.md`.
-- `internal/chat/` owns internal chat state (`thread`, `message`, `timeline`, `view`) and does not orchestrate Agent execution. Inbound adaptation lives in `internal/channel/inbound/`; discuss-mode driving in `internal/channel/discuss/`.
+- `internal/chat/` owns internal chat state (`thread`, `message`, `checkpoint`, `timeline`, `view`) and does not orchestrate Agent execution. Inbound adaptation lives in `internal/channel/inbound/`; discuss-mode driving in `internal/channel/discuss/`.
 - Browser Use tools (`browser_action` / `browser_observe` / `browser_remote_session`) drive the headed workspace Chrome over CDP; Computer Use (`computer_observe` / `computer_action`) drives the GUI desktop via AT-SPI + RFB. Screenshots are saved to a workspace path and must be explicitly read. Prefer Browser Use for web pages; Computer Use for native dialogs and non-browser GUI states.
 - Runtime internals (Twilight `Stream()`/`Generate()`, compaction, loop detection, tag extraction, full Browser/Computer Use details): `docs/agent-runtime.md`.
 

cmd/agent/abandoned_decision_fold_postgres_test.go

@@ -72,7 +72,7 @@ func TestPostgresInRunFoldTakesAnAbandonedDecisionWithItsCall(t *testing.T) {
 				t.Fatalf("loop: code=%q effects=%v, want every effect run", code, s.effects)
 			}
 			var folded int
-			if err := f.pool.QueryRow(f.ctx, `SELECT count(*) FROM bot_history_messages h JOIN bot_history_message_compacts c ON c.id=h.compact_id AND c.status='ok' WHERE h.session_id=$1 AND h.role='assistant' AND h.content::text LIKE '%call_001%'`, f.sessionID).Scan(&folded); err != nil {
+			if err := f.pool.QueryRow(f.ctx, `SELECT count(*) FROM (`+historyRowsSQL+`) h WHERE h.folded AND h.role='assistant' AND h.content::text LIKE '%call_001%'`, f.sessionID).Scan(&folded); err != nil {
 				t.Fatal(err)
 			}
 			if folded != 1 {