feat(runtime): add attention-first semantic compaction by likun666661 · Pull Request #986 · apache/maka

likun666661 · 2026-07-14T13:52:49Z

Summary

preserve prior replay and the exact current-turn user message as an immutable head anchor
replace only the structurally complete middle span before the open provider-protocol tail
use the 8K total-context / 4K completed-span attention trigger and separate generation / accepted-projection budgets
add bounded structured and complete-sentence/line fallback validation, V2 rolling lineage, and shared attention/emergency scheduling
keep the provider-visible continuation strictly exact head anchor + LLM-authored projection + exact protocol tail
keep Runtime coverage, source refs, archive refs, preserved-tail indexes, and legacy state cards in the diagnostic side channel only

Safety properties

the head anchor is signature-checked on every prepareStep and is never semantic coverage
consecutive assistant reasoning/text/tool-call messages form one provider episode; incomplete multi-call episodes remain verbatim in the tail
empty, truncated, source-invalid, private-verifier-invalid, protocol-invalid, or over-budget projections do not replace source messages
token savings, compact-call cost, and cache loss remain diagnostics in attention mode
one prepareStep accepts at most one attention or emergency replacement
emergency compaction reports explicit head_anchor_exceeds_capacity / no_safe_completed_span outcomes instead of rewriting the user instruction
semanticCompact no longer generates, prompts with, records on new blocks, or renders deterministic state cards
rolling successors re-render the predecessor's LLM summary and ignore legacy state cards instead of replaying the old provider message verbatim

MIPS smoke finding

A GLM-5.2 make-mips-interpreter smoke on the earlier version was stopped after it exceeded 135 steps without creating /app/vm.js.

The retained trace showed:

24 accepted semantic compactions and 228 tool calls
19 later summaries whose next action was to write /app/vm.js
provider-visible state cards that instead repeated assistant planning such as “re-read” or “re-verify”
zero creation of /app/vm.js

The deterministic VM matcher treated any assistant text containing vm as operational state. Those inferred cards were rendered after every LLM summary and then copied into rolling successors, creating a positive feedback loop. The fix in 4eebe953 removes that semantic path rather than attempting to repair it with more regexes.

This smoke is mechanism evidence, not an A/B quality result. The old main-branch arm separately failed to activate semantic compaction because of tool_pair_split and eventually timed out near provider capacity, so it is not a usable success baseline.

Validation

After merging current upstream/main:

focused semantic + backend + mid-turn capacity integration: 187 passed, 0 failed
npm --workspace @maka/headless test: 964 passed, 0 failed, 1 skipped
npm run typecheck: passes across all workspaces after rebuilding changed workspace dependencies
Runtime full suite: 1943 passed, 7 skipped, with one unrelated local-only failure in main's macOS sandbox executable-root index assertion because this machine runs Node from /usr/local/bin/node
git diff --check: passes

No attention benefit is claimed yet, and this PR does not expand any surface default. A fresh same-commit A/B is still required after the mechanism correction.

Refs #981

* feat(runtime): default mid-turn capacity compaction on, runtime-owned The mid_turn capacity invariant becomes a runtime-owned default: whenever history compaction is enabled the runtime derives midTurn from the selected model window (contextWindow - reserve), so Desktop, TUI, and headless inherit it without copying config. MAKA_CONTEXT_HISTORY_COMPACT_MID_TURN=off stays as the explicit escape hatch. The backend still gates activation on the checkpoint seams, the persisted head anchor, and a KNOWN context window, so a session without model metadata or a child session with no anchor seam never misfires even though the default is on (covered by new backend safety guards). Refs #882 * feat: default semantic compaction off, runtime-owned across surfaces Semantic compaction is the #981/#986 attention-first experiment, not the standard historyCompact mechanism. Make the runtime the owner of its default: it stays off unless a surface explicitly opts in (MAKA_CONTEXT_SEMANTIC_COMPACT truthy, or a non-off MAKA_CONTEXT_SEMANTIC_COMPACT_MODE). This lets the CLI drop its local semanticCompact strip and eat the runtime default directly, so the surfaces stop duplicating the same off decision. Desktop and headless inherit the off default without any per-surface wiring. Refs #882 * test(runtime): lock the shipped default's proactive long-turn journey Wire buildDefaultContextBudgetPolicy — the exact default every surface now inherits — into the streaming backend and prove a long turn near the window engages mid-turn compaction and continues to a clean end rather than being truncated or surfacing a raw provider error. The fake-backend Playwright suite cannot exercise this path: it is a separate canned-event backend that never runs AiSdkBackend's prepareStep machinery, so a desktop E2E could only fake the very compaction under test. This deterministic runtime-layer test is the honest representative journey for the proactive path. Refs #882 * fix(runtime): bound the derived compaction reserve by the model window The flat 16384 reserve assumed large-window models: on gpt-4 (8192 window) the default policy derived maxHistoryEstimatedTokens = max(1, 8192 - 16384) = 1 — a pre-existing turn-boundary pathology on main — and with mid_turn now on by default the 1-token high water made every multi-step turn run the summarizer for a checkpoint the replay gate could never admit. Make defaultCompactReserveTokens the single owner of the reserve default, bounded by the KNOWN window: min(16384, floor(window / 4)), so both the turn-boundary budget and the mid_turn high water derive from one value (8K window -> 2048 reserve / 6144 budget; >= 64K windows byte-identical to before; unknown window keeps the classic constant). An explicit MAKA_CONTEXT_HISTORY_COMPACT_RESERVE_TOKENS is respected verbatim. Peers bound the same way: opencode caps its compaction buffer by the model's output limit, gemini-cli triggers at a window fraction; pi shares the flat-16384 blind spot. Strengthen the shipped-defaults journey to fully-default derivation with durable-persistence assertions (checkpoint recorded, compact block present, replaced raw span gone), and add the unrescuable path: under the derived default a turn with no safe span ends with the explicit context_budget_exhausted outcome. A small-window guard locks the P2 behavior: on an 8K model the default is completely inert when under budget — never a pointless summarizer call. Refs #882 * docs(runtime): align midTurn policy comments with the shipped runtime default

…t-semantic-compaction # Conflicts: # packages/runtime/src/ai-sdk-backend.ts # packages/runtime/src/context-budget-policy.ts

…c CLI color tests) (#1077) * fix(runtime): let active prune cover the newest step under mid-turn capacity #986 unconditionally excluded the newest completed step from active tool-result pruning, breaking three reviewed capacity-rescue invariants (mid-turn finding C, capacity-replacement re-convergence, overflow round-3 P1) and CI on main. Include the newest step exactly when mid-turn capacity compaction is active, keeping the keep-newest-raw default for every other configuration. * test(cli): pin color level so ANSI assertions are hermetic tui-ansi detects color capability from TERM/COLORTERM at module load, so the pi-transcript and pi-tui-runner ANSI assertions inherited ambient terminal capability: truecolor locally, level 0 on CI runners with no TERM/COLORTERM, where every color function is a no-op and the #1064/#1066 color assertions fail deterministically. Pin level 3 via the existing _setColorLevelForTesting seam, matching tui-ansi.test.ts.

likun666661 · 2026-07-14T13:52:49Z

Summary

preserve prior replay and the exact current-turn user message as an immutable head anchor
replace only the structurally complete middle span before the open provider-protocol tail
use the 8K total-context / 4K completed-span attention trigger and separate generation / accepted-projection budgets
add bounded structured and complete-sentence/line fallback validation, V2 rolling lineage, and shared attention/emergency scheduling
keep the provider-visible continuation strictly exact head anchor + LLM-authored projection + exact protocol tail
keep Runtime coverage, source refs, archive refs, preserved-tail indexes, and legacy state cards in the diagnostic side channel only

Safety properties

the head anchor is signature-checked on every prepareStep and is never semantic coverage
consecutive assistant reasoning/text/tool-call messages form one provider episode; incomplete multi-call episodes remain verbatim in the tail
empty, truncated, source-invalid, private-verifier-invalid, protocol-invalid, or over-budget projections do not replace source messages
token savings, compact-call cost, and cache loss remain diagnostics in attention mode
one prepareStep accepts at most one attention or emergency replacement
emergency compaction reports explicit head_anchor_exceeds_capacity / no_safe_completed_span outcomes instead of rewriting the user instruction
semanticCompact no longer generates, prompts with, records on new blocks, or renders deterministic state cards
rolling successors re-render the predecessor's LLM summary and ignore legacy state cards instead of replaying the old provider message verbatim

MIPS smoke finding

A GLM-5.2 make-mips-interpreter smoke on the earlier version was stopped after it exceeded 135 steps without creating /app/vm.js.

The retained trace showed:

24 accepted semantic compactions and 228 tool calls
19 later summaries whose next action was to write /app/vm.js
provider-visible state cards that instead repeated assistant planning such as “re-read” or “re-verify”
zero creation of /app/vm.js

The deterministic VM matcher treated any assistant text containing vm as operational state. Those inferred cards were rendered after every LLM summary and then copied into rolling successors, creating a positive feedback loop. The fix in 4eebe953 removes that semantic path rather than attempting to repair it with more regexes.

This smoke is mechanism evidence, not an A/B quality result. The old main-branch arm separately failed to activate semantic compaction because of tool_pair_split and eventually timed out near provider capacity, so it is not a usable success baseline.

Validation

After merging current upstream/main:

focused semantic + backend + mid-turn capacity integration: 187 passed, 0 failed
npm --workspace @maka/headless test: 964 passed, 0 failed, 1 skipped
npm run typecheck: passes across all workspaces after rebuilding changed workspace dependencies
Runtime full suite: 1943 passed, 7 skipped, with one unrelated local-only failure in main's macOS sandbox executable-root index assertion because this machine runs Node from /usr/local/bin/node
git diff --check: passes

No attention benefit is claimed yet, and this PR does not expand any surface default. A fresh same-commit A/B is still required after the mechanism correction.

Refs #981

* feat(runtime): default mid-turn capacity compaction on, runtime-owned The mid_turn capacity invariant becomes a runtime-owned default: whenever history compaction is enabled the runtime derives midTurn from the selected model window (contextWindow - reserve), so Desktop, TUI, and headless inherit it without copying config. MAKA_CONTEXT_HISTORY_COMPACT_MID_TURN=off stays as the explicit escape hatch. The backend still gates activation on the checkpoint seams, the persisted head anchor, and a KNOWN context window, so a session without model metadata or a child session with no anchor seam never misfires even though the default is on (covered by new backend safety guards). Refs #882 * feat: default semantic compaction off, runtime-owned across surfaces Semantic compaction is the #981/#986 attention-first experiment, not the standard historyCompact mechanism. Make the runtime the owner of its default: it stays off unless a surface explicitly opts in (MAKA_CONTEXT_SEMANTIC_COMPACT truthy, or a non-off MAKA_CONTEXT_SEMANTIC_COMPACT_MODE). This lets the CLI drop its local semanticCompact strip and eat the runtime default directly, so the surfaces stop duplicating the same off decision. Desktop and headless inherit the off default without any per-surface wiring. Refs #882 * test(runtime): lock the shipped default's proactive long-turn journey Wire buildDefaultContextBudgetPolicy — the exact default every surface now inherits — into the streaming backend and prove a long turn near the window engages mid-turn compaction and continues to a clean end rather than being truncated or surfacing a raw provider error. The fake-backend Playwright suite cannot exercise this path: it is a separate canned-event backend that never runs AiSdkBackend's prepareStep machinery, so a desktop E2E could only fake the very compaction under test. This deterministic runtime-layer test is the honest representative journey for the proactive path. Refs #882 * fix(runtime): bound the derived compaction reserve by the model window The flat 16384 reserve assumed large-window models: on gpt-4 (8192 window) the default policy derived maxHistoryEstimatedTokens = max(1, 8192 - 16384) = 1 — a pre-existing turn-boundary pathology on main — and with mid_turn now on by default the 1-token high water made every multi-step turn run the summarizer for a checkpoint the replay gate could never admit. Make defaultCompactReserveTokens the single owner of the reserve default, bounded by the KNOWN window: min(16384, floor(window / 4)), so both the turn-boundary budget and the mid_turn high water derive from one value (8K window -> 2048 reserve / 6144 budget; >= 64K windows byte-identical to before; unknown window keeps the classic constant). An explicit MAKA_CONTEXT_HISTORY_COMPACT_RESERVE_TOKENS is respected verbatim. Peers bound the same way: opencode caps its compaction buffer by the model's output limit, gemini-cli triggers at a window fraction; pi shares the flat-16384 blind spot. Strengthen the shipped-defaults journey to fully-default derivation with durable-persistence assertions (checkpoint recorded, compact block present, replaced raw span gone), and add the unrescuable path: under the derived default a turn with no safe span ends with the explicit context_budget_exhausted outcome. A small-window guard locks the P2 behavior: on an 8K model the default is completely inert when under budget — never a pointless summarizer call. Refs #882 * docs(runtime): align midTurn policy comments with the shipped runtime default

…t-semantic-compaction # Conflicts: # packages/runtime/src/ai-sdk-backend.ts # packages/runtime/src/context-budget-policy.ts

…c CLI color tests) (#1077) * fix(runtime): let active prune cover the newest step under mid-turn capacity #986 unconditionally excluded the newest completed step from active tool-result pruning, breaking three reviewed capacity-rescue invariants (mid-turn finding C, capacity-replacement re-convergence, overflow round-3 P1) and CI on main. Include the newest step exactly when mid-turn capacity compaction is active, keeping the keep-newest-raw default for every other configuration. * test(cli): pin color level so ANSI assertions are hermetic tui-ansi detects color capability from TERM/COLORTERM at module load, so the pi-transcript and pi-tui-runner ANSI assertions inherited ambient terminal capability: truecolor locally, level 0 on CI runners with no TERM/COLORTERM, where every color function is a no-op and the #1064/#1066 color assertions fail deterministically. Pin level 3 via the existing _setColorLevelForTesting seam, matching tui-ansi.test.ts.

likun666661 · 2026-07-14T13:52:49Z

Summary

preserve prior replay and the exact current-turn user message as an immutable head anchor
replace only the structurally complete middle span before the open provider-protocol tail
use the 8K total-context / 4K completed-span attention trigger and separate generation / accepted-projection budgets
add bounded structured and complete-sentence/line fallback validation, V2 rolling lineage, and shared attention/emergency scheduling
keep the provider-visible continuation strictly exact head anchor + LLM-authored projection + exact protocol tail
keep Runtime coverage, source refs, archive refs, preserved-tail indexes, and legacy state cards in the diagnostic side channel only

Safety properties

the head anchor is signature-checked on every prepareStep and is never semantic coverage
consecutive assistant reasoning/text/tool-call messages form one provider episode; incomplete multi-call episodes remain verbatim in the tail
empty, truncated, source-invalid, private-verifier-invalid, protocol-invalid, or over-budget projections do not replace source messages
token savings, compact-call cost, and cache loss remain diagnostics in attention mode
one prepareStep accepts at most one attention or emergency replacement
emergency compaction reports explicit head_anchor_exceeds_capacity / no_safe_completed_span outcomes instead of rewriting the user instruction
semanticCompact no longer generates, prompts with, records on new blocks, or renders deterministic state cards
rolling successors re-render the predecessor's LLM summary and ignore legacy state cards instead of replaying the old provider message verbatim

MIPS smoke finding

A GLM-5.2 make-mips-interpreter smoke on the earlier version was stopped after it exceeded 135 steps without creating /app/vm.js.

The retained trace showed:

24 accepted semantic compactions and 228 tool calls
19 later summaries whose next action was to write /app/vm.js
provider-visible state cards that instead repeated assistant planning such as “re-read” or “re-verify”
zero creation of /app/vm.js

The deterministic VM matcher treated any assistant text containing vm as operational state. Those inferred cards were rendered after every LLM summary and then copied into rolling successors, creating a positive feedback loop. The fix in 4eebe953 removes that semantic path rather than attempting to repair it with more regexes.

This smoke is mechanism evidence, not an A/B quality result. The old main-branch arm separately failed to activate semantic compaction because of tool_pair_split and eventually timed out near provider capacity, so it is not a usable success baseline.

Validation

After merging current upstream/main:

focused semantic + backend + mid-turn capacity integration: 187 passed, 0 failed
npm --workspace @maka/headless test: 964 passed, 0 failed, 1 skipped
npm run typecheck: passes across all workspaces after rebuilding changed workspace dependencies
Runtime full suite: 1943 passed, 7 skipped, with one unrelated local-only failure in main's macOS sandbox executable-root index assertion because this machine runs Node from /usr/local/bin/node
git diff --check: passes

No attention benefit is claimed yet, and this PR does not expand any surface default. A fresh same-commit A/B is still required after the mechanism correction.

Refs #981

* feat(runtime): default mid-turn capacity compaction on, runtime-owned The mid_turn capacity invariant becomes a runtime-owned default: whenever history compaction is enabled the runtime derives midTurn from the selected model window (contextWindow - reserve), so Desktop, TUI, and headless inherit it without copying config. MAKA_CONTEXT_HISTORY_COMPACT_MID_TURN=off stays as the explicit escape hatch. The backend still gates activation on the checkpoint seams, the persisted head anchor, and a KNOWN context window, so a session without model metadata or a child session with no anchor seam never misfires even though the default is on (covered by new backend safety guards). Refs #882 * feat: default semantic compaction off, runtime-owned across surfaces Semantic compaction is the #981/#986 attention-first experiment, not the standard historyCompact mechanism. Make the runtime the owner of its default: it stays off unless a surface explicitly opts in (MAKA_CONTEXT_SEMANTIC_COMPACT truthy, or a non-off MAKA_CONTEXT_SEMANTIC_COMPACT_MODE). This lets the CLI drop its local semanticCompact strip and eat the runtime default directly, so the surfaces stop duplicating the same off decision. Desktop and headless inherit the off default without any per-surface wiring. Refs #882 * test(runtime): lock the shipped default's proactive long-turn journey Wire buildDefaultContextBudgetPolicy — the exact default every surface now inherits — into the streaming backend and prove a long turn near the window engages mid-turn compaction and continues to a clean end rather than being truncated or surfacing a raw provider error. The fake-backend Playwright suite cannot exercise this path: it is a separate canned-event backend that never runs AiSdkBackend's prepareStep machinery, so a desktop E2E could only fake the very compaction under test. This deterministic runtime-layer test is the honest representative journey for the proactive path. Refs #882 * fix(runtime): bound the derived compaction reserve by the model window The flat 16384 reserve assumed large-window models: on gpt-4 (8192 window) the default policy derived maxHistoryEstimatedTokens = max(1, 8192 - 16384) = 1 — a pre-existing turn-boundary pathology on main — and with mid_turn now on by default the 1-token high water made every multi-step turn run the summarizer for a checkpoint the replay gate could never admit. Make defaultCompactReserveTokens the single owner of the reserve default, bounded by the KNOWN window: min(16384, floor(window / 4)), so both the turn-boundary budget and the mid_turn high water derive from one value (8K window -> 2048 reserve / 6144 budget; >= 64K windows byte-identical to before; unknown window keeps the classic constant). An explicit MAKA_CONTEXT_HISTORY_COMPACT_RESERVE_TOKENS is respected verbatim. Peers bound the same way: opencode caps its compaction buffer by the model's output limit, gemini-cli triggers at a window fraction; pi shares the flat-16384 blind spot. Strengthen the shipped-defaults journey to fully-default derivation with durable-persistence assertions (checkpoint recorded, compact block present, replaced raw span gone), and add the unrescuable path: under the derived default a turn with no safe span ends with the explicit context_budget_exhausted outcome. A small-window guard locks the P2 behavior: on an 8K model the default is completely inert when under budget — never a pointless summarizer call. Refs #882 * docs(runtime): align midTurn policy comments with the shipped runtime default

…t-semantic-compaction # Conflicts: # packages/runtime/src/ai-sdk-backend.ts # packages/runtime/src/context-budget-policy.ts

…c CLI color tests) (#1077) * fix(runtime): let active prune cover the newest step under mid-turn capacity #986 unconditionally excluded the newest completed step from active tool-result pruning, breaking three reviewed capacity-rescue invariants (mid-turn finding C, capacity-replacement re-convergence, overflow round-3 P1) and CI on main. Include the newest step exactly when mid-turn capacity compaction is active, keeping the keep-newest-raw default for every other configuration. * test(cli): pin color level so ANSI assertions are hermetic tui-ansi detects color capability from TERM/COLORTERM at module load, so the pi-transcript and pi-tui-runner ANSI assertions inherited ambient terminal capability: truecolor locally, level 0 on CI runners with no TERM/COLORTERM, where every color function is a no-op and the #1064/#1066 color assertions fail deterministically. Pin level 3 via the existing _setColorLevelForTesting seam, matching tui-ansi.test.ts.

likun666661 · 2026-07-14T13:52:49Z

Summary

preserve prior replay and the exact current-turn user message as an immutable head anchor
replace only the structurally complete middle span before the open provider-protocol tail
use the 8K total-context / 4K completed-span attention trigger and separate generation / accepted-projection budgets
add bounded structured and complete-sentence/line fallback validation, V2 rolling lineage, and shared attention/emergency scheduling
keep the provider-visible continuation strictly exact head anchor + LLM-authored projection + exact protocol tail
keep Runtime coverage, source refs, archive refs, preserved-tail indexes, and legacy state cards in the diagnostic side channel only

Safety properties

the head anchor is signature-checked on every prepareStep and is never semantic coverage
consecutive assistant reasoning/text/tool-call messages form one provider episode; incomplete multi-call episodes remain verbatim in the tail
empty, truncated, source-invalid, private-verifier-invalid, protocol-invalid, or over-budget projections do not replace source messages
token savings, compact-call cost, and cache loss remain diagnostics in attention mode
one prepareStep accepts at most one attention or emergency replacement
emergency compaction reports explicit head_anchor_exceeds_capacity / no_safe_completed_span outcomes instead of rewriting the user instruction
semanticCompact no longer generates, prompts with, records on new blocks, or renders deterministic state cards
rolling successors re-render the predecessor's LLM summary and ignore legacy state cards instead of replaying the old provider message verbatim

MIPS smoke finding

A GLM-5.2 make-mips-interpreter smoke on the earlier version was stopped after it exceeded 135 steps without creating /app/vm.js.

The retained trace showed:

24 accepted semantic compactions and 228 tool calls
19 later summaries whose next action was to write /app/vm.js
provider-visible state cards that instead repeated assistant planning such as “re-read” or “re-verify”
zero creation of /app/vm.js

The deterministic VM matcher treated any assistant text containing vm as operational state. Those inferred cards were rendered after every LLM summary and then copied into rolling successors, creating a positive feedback loop. The fix in 4eebe953 removes that semantic path rather than attempting to repair it with more regexes.

This smoke is mechanism evidence, not an A/B quality result. The old main-branch arm separately failed to activate semantic compaction because of tool_pair_split and eventually timed out near provider capacity, so it is not a usable success baseline.

Validation

After merging current upstream/main:

focused semantic + backend + mid-turn capacity integration: 187 passed, 0 failed
npm --workspace @maka/headless test: 964 passed, 0 failed, 1 skipped
npm run typecheck: passes across all workspaces after rebuilding changed workspace dependencies
Runtime full suite: 1943 passed, 7 skipped, with one unrelated local-only failure in main's macOS sandbox executable-root index assertion because this machine runs Node from /usr/local/bin/node
git diff --check: passes

No attention benefit is claimed yet, and this PR does not expand any surface default. A fresh same-commit A/B is still required after the mechanism correction.

Refs #981

* feat(runtime): default mid-turn capacity compaction on, runtime-owned The mid_turn capacity invariant becomes a runtime-owned default: whenever history compaction is enabled the runtime derives midTurn from the selected model window (contextWindow - reserve), so Desktop, TUI, and headless inherit it without copying config. MAKA_CONTEXT_HISTORY_COMPACT_MID_TURN=off stays as the explicit escape hatch. The backend still gates activation on the checkpoint seams, the persisted head anchor, and a KNOWN context window, so a session without model metadata or a child session with no anchor seam never misfires even though the default is on (covered by new backend safety guards). Refs #882 * feat: default semantic compaction off, runtime-owned across surfaces Semantic compaction is the #981/#986 attention-first experiment, not the standard historyCompact mechanism. Make the runtime the owner of its default: it stays off unless a surface explicitly opts in (MAKA_CONTEXT_SEMANTIC_COMPACT truthy, or a non-off MAKA_CONTEXT_SEMANTIC_COMPACT_MODE). This lets the CLI drop its local semanticCompact strip and eat the runtime default directly, so the surfaces stop duplicating the same off decision. Desktop and headless inherit the off default without any per-surface wiring. Refs #882 * test(runtime): lock the shipped default's proactive long-turn journey Wire buildDefaultContextBudgetPolicy — the exact default every surface now inherits — into the streaming backend and prove a long turn near the window engages mid-turn compaction and continues to a clean end rather than being truncated or surfacing a raw provider error. The fake-backend Playwright suite cannot exercise this path: it is a separate canned-event backend that never runs AiSdkBackend's prepareStep machinery, so a desktop E2E could only fake the very compaction under test. This deterministic runtime-layer test is the honest representative journey for the proactive path. Refs #882 * fix(runtime): bound the derived compaction reserve by the model window The flat 16384 reserve assumed large-window models: on gpt-4 (8192 window) the default policy derived maxHistoryEstimatedTokens = max(1, 8192 - 16384) = 1 — a pre-existing turn-boundary pathology on main — and with mid_turn now on by default the 1-token high water made every multi-step turn run the summarizer for a checkpoint the replay gate could never admit. Make defaultCompactReserveTokens the single owner of the reserve default, bounded by the KNOWN window: min(16384, floor(window / 4)), so both the turn-boundary budget and the mid_turn high water derive from one value (8K window -> 2048 reserve / 6144 budget; >= 64K windows byte-identical to before; unknown window keeps the classic constant). An explicit MAKA_CONTEXT_HISTORY_COMPACT_RESERVE_TOKENS is respected verbatim. Peers bound the same way: opencode caps its compaction buffer by the model's output limit, gemini-cli triggers at a window fraction; pi shares the flat-16384 blind spot. Strengthen the shipped-defaults journey to fully-default derivation with durable-persistence assertions (checkpoint recorded, compact block present, replaced raw span gone), and add the unrescuable path: under the derived default a turn with no safe span ends with the explicit context_budget_exhausted outcome. A small-window guard locks the P2 behavior: on an 8K model the default is completely inert when under budget — never a pointless summarizer call. Refs #882 * docs(runtime): align midTurn policy comments with the shipped runtime default

…t-semantic-compaction # Conflicts: # packages/runtime/src/ai-sdk-backend.ts # packages/runtime/src/context-budget-policy.ts

…c CLI color tests) (#1077) * fix(runtime): let active prune cover the newest step under mid-turn capacity #986 unconditionally excluded the newest completed step from active tool-result pruning, breaking three reviewed capacity-rescue invariants (mid-turn finding C, capacity-replacement re-convergence, overflow round-3 P1) and CI on main. Include the newest step exactly when mid-turn capacity compaction is active, keeping the keep-newest-raw default for every other configuration. * test(cli): pin color level so ANSI assertions are hermetic tui-ansi detects color capability from TERM/COLORTERM at module load, so the pi-transcript and pi-tui-runner ANSI assertions inherited ambient terminal capability: truecolor locally, level 0 on CI runners with no TERM/COLORTERM, where every color function is a no-op and the #1064/#1066 color assertions fail deterministically. Pin level 3 via the existing _setColorLevelForTesting seam, matching tui-ansi.test.ts.

likun666661 · 2026-07-14T13:52:49Z

Summary

preserve prior replay and the exact current-turn user message as an immutable head anchor
replace only the structurally complete middle span before the open provider-protocol tail
use the 8K total-context / 4K completed-span attention trigger and separate generation / accepted-projection budgets
add bounded structured and complete-sentence/line fallback validation, V2 rolling lineage, and shared attention/emergency scheduling
keep the provider-visible continuation strictly exact head anchor + LLM-authored projection + exact protocol tail
keep Runtime coverage, source refs, archive refs, preserved-tail indexes, and legacy state cards in the diagnostic side channel only

Safety properties

the head anchor is signature-checked on every prepareStep and is never semantic coverage
consecutive assistant reasoning/text/tool-call messages form one provider episode; incomplete multi-call episodes remain verbatim in the tail
empty, truncated, source-invalid, private-verifier-invalid, protocol-invalid, or over-budget projections do not replace source messages
token savings, compact-call cost, and cache loss remain diagnostics in attention mode
one prepareStep accepts at most one attention or emergency replacement
emergency compaction reports explicit head_anchor_exceeds_capacity / no_safe_completed_span outcomes instead of rewriting the user instruction
semanticCompact no longer generates, prompts with, records on new blocks, or renders deterministic state cards
rolling successors re-render the predecessor's LLM summary and ignore legacy state cards instead of replaying the old provider message verbatim

MIPS smoke finding

A GLM-5.2 make-mips-interpreter smoke on the earlier version was stopped after it exceeded 135 steps without creating /app/vm.js.

The retained trace showed:

24 accepted semantic compactions and 228 tool calls
19 later summaries whose next action was to write /app/vm.js
provider-visible state cards that instead repeated assistant planning such as “re-read” or “re-verify”
zero creation of /app/vm.js

The deterministic VM matcher treated any assistant text containing vm as operational state. Those inferred cards were rendered after every LLM summary and then copied into rolling successors, creating a positive feedback loop. The fix in 4eebe953 removes that semantic path rather than attempting to repair it with more regexes.

This smoke is mechanism evidence, not an A/B quality result. The old main-branch arm separately failed to activate semantic compaction because of tool_pair_split and eventually timed out near provider capacity, so it is not a usable success baseline.

Validation

After merging current upstream/main:

focused semantic + backend + mid-turn capacity integration: 187 passed, 0 failed
npm --workspace @maka/headless test: 964 passed, 0 failed, 1 skipped
npm run typecheck: passes across all workspaces after rebuilding changed workspace dependencies
Runtime full suite: 1943 passed, 7 skipped, with one unrelated local-only failure in main's macOS sandbox executable-root index assertion because this machine runs Node from /usr/local/bin/node
git diff --check: passes

No attention benefit is claimed yet, and this PR does not expand any surface default. A fresh same-commit A/B is still required after the mechanism correction.

Refs #981

* feat(runtime): default mid-turn capacity compaction on, runtime-owned The mid_turn capacity invariant becomes a runtime-owned default: whenever history compaction is enabled the runtime derives midTurn from the selected model window (contextWindow - reserve), so Desktop, TUI, and headless inherit it without copying config. MAKA_CONTEXT_HISTORY_COMPACT_MID_TURN=off stays as the explicit escape hatch. The backend still gates activation on the checkpoint seams, the persisted head anchor, and a KNOWN context window, so a session without model metadata or a child session with no anchor seam never misfires even though the default is on (covered by new backend safety guards). Refs #882 * feat: default semantic compaction off, runtime-owned across surfaces Semantic compaction is the #981/#986 attention-first experiment, not the standard historyCompact mechanism. Make the runtime the owner of its default: it stays off unless a surface explicitly opts in (MAKA_CONTEXT_SEMANTIC_COMPACT truthy, or a non-off MAKA_CONTEXT_SEMANTIC_COMPACT_MODE). This lets the CLI drop its local semanticCompact strip and eat the runtime default directly, so the surfaces stop duplicating the same off decision. Desktop and headless inherit the off default without any per-surface wiring. Refs #882 * test(runtime): lock the shipped default's proactive long-turn journey Wire buildDefaultContextBudgetPolicy — the exact default every surface now inherits — into the streaming backend and prove a long turn near the window engages mid-turn compaction and continues to a clean end rather than being truncated or surfacing a raw provider error. The fake-backend Playwright suite cannot exercise this path: it is a separate canned-event backend that never runs AiSdkBackend's prepareStep machinery, so a desktop E2E could only fake the very compaction under test. This deterministic runtime-layer test is the honest representative journey for the proactive path. Refs #882 * fix(runtime): bound the derived compaction reserve by the model window The flat 16384 reserve assumed large-window models: on gpt-4 (8192 window) the default policy derived maxHistoryEstimatedTokens = max(1, 8192 - 16384) = 1 — a pre-existing turn-boundary pathology on main — and with mid_turn now on by default the 1-token high water made every multi-step turn run the summarizer for a checkpoint the replay gate could never admit. Make defaultCompactReserveTokens the single owner of the reserve default, bounded by the KNOWN window: min(16384, floor(window / 4)), so both the turn-boundary budget and the mid_turn high water derive from one value (8K window -> 2048 reserve / 6144 budget; >= 64K windows byte-identical to before; unknown window keeps the classic constant). An explicit MAKA_CONTEXT_HISTORY_COMPACT_RESERVE_TOKENS is respected verbatim. Peers bound the same way: opencode caps its compaction buffer by the model's output limit, gemini-cli triggers at a window fraction; pi shares the flat-16384 blind spot. Strengthen the shipped-defaults journey to fully-default derivation with durable-persistence assertions (checkpoint recorded, compact block present, replaced raw span gone), and add the unrescuable path: under the derived default a turn with no safe span ends with the explicit context_budget_exhausted outcome. A small-window guard locks the P2 behavior: on an 8K model the default is completely inert when under budget — never a pointless summarizer call. Refs #882 * docs(runtime): align midTurn policy comments with the shipped runtime default

…t-semantic-compaction # Conflicts: # packages/runtime/src/ai-sdk-backend.ts # packages/runtime/src/context-budget-policy.ts

…c CLI color tests) (#1077) * fix(runtime): let active prune cover the newest step under mid-turn capacity #986 unconditionally excluded the newest completed step from active tool-result pruning, breaking three reviewed capacity-rescue invariants (mid-turn finding C, capacity-replacement re-convergence, overflow round-3 P1) and CI on main. Include the newest step exactly when mid-turn capacity compaction is active, keeping the keep-newest-raw default for every other configuration. * test(cli): pin color level so ANSI assertions are hermetic tui-ansi detects color capability from TERM/COLORTERM at module load, so the pi-transcript and pi-tui-runner ANSI assertions inherited ambient terminal capability: truecolor locally, level 0 on CI runners with no TERM/COLORTERM, where every color function is a no-op and the #1064/#1066 color assertions fail deterministically. Pin level 3 via the existing _setColorLevelForTesting seam, matching tui-ansi.test.ts.

likun666661 · 2026-07-14T13:52:49Z

Summary

preserve prior replay and the exact current-turn user message as an immutable head anchor
replace only the structurally complete middle span before the open provider-protocol tail
use the 8K total-context / 4K completed-span attention trigger and separate generation / accepted-projection budgets
add bounded structured and complete-sentence/line fallback validation, V2 rolling lineage, and shared attention/emergency scheduling
keep the provider-visible continuation strictly exact head anchor + LLM-authored projection + exact protocol tail
keep Runtime coverage, source refs, archive refs, preserved-tail indexes, and legacy state cards in the diagnostic side channel only

Safety properties

the head anchor is signature-checked on every prepareStep and is never semantic coverage
consecutive assistant reasoning/text/tool-call messages form one provider episode; incomplete multi-call episodes remain verbatim in the tail
empty, truncated, source-invalid, private-verifier-invalid, protocol-invalid, or over-budget projections do not replace source messages
token savings, compact-call cost, and cache loss remain diagnostics in attention mode
one prepareStep accepts at most one attention or emergency replacement
emergency compaction reports explicit head_anchor_exceeds_capacity / no_safe_completed_span outcomes instead of rewriting the user instruction
semanticCompact no longer generates, prompts with, records on new blocks, or renders deterministic state cards
rolling successors re-render the predecessor's LLM summary and ignore legacy state cards instead of replaying the old provider message verbatim

MIPS smoke finding

A GLM-5.2 make-mips-interpreter smoke on the earlier version was stopped after it exceeded 135 steps without creating /app/vm.js.

The retained trace showed:

24 accepted semantic compactions and 228 tool calls
19 later summaries whose next action was to write /app/vm.js
provider-visible state cards that instead repeated assistant planning such as “re-read” or “re-verify”
zero creation of /app/vm.js

The deterministic VM matcher treated any assistant text containing vm as operational state. Those inferred cards were rendered after every LLM summary and then copied into rolling successors, creating a positive feedback loop. The fix in 4eebe953 removes that semantic path rather than attempting to repair it with more regexes.

This smoke is mechanism evidence, not an A/B quality result. The old main-branch arm separately failed to activate semantic compaction because of tool_pair_split and eventually timed out near provider capacity, so it is not a usable success baseline.

Validation

After merging current upstream/main:

focused semantic + backend + mid-turn capacity integration: 187 passed, 0 failed
npm --workspace @maka/headless test: 964 passed, 0 failed, 1 skipped
npm run typecheck: passes across all workspaces after rebuilding changed workspace dependencies
Runtime full suite: 1943 passed, 7 skipped, with one unrelated local-only failure in main's macOS sandbox executable-root index assertion because this machine runs Node from /usr/local/bin/node
git diff --check: passes

No attention benefit is claimed yet, and this PR does not expand any surface default. A fresh same-commit A/B is still required after the mechanism correction.

Refs #981

* feat(runtime): default mid-turn capacity compaction on, runtime-owned The mid_turn capacity invariant becomes a runtime-owned default: whenever history compaction is enabled the runtime derives midTurn from the selected model window (contextWindow - reserve), so Desktop, TUI, and headless inherit it without copying config. MAKA_CONTEXT_HISTORY_COMPACT_MID_TURN=off stays as the explicit escape hatch. The backend still gates activation on the checkpoint seams, the persisted head anchor, and a KNOWN context window, so a session without model metadata or a child session with no anchor seam never misfires even though the default is on (covered by new backend safety guards). Refs #882 * feat: default semantic compaction off, runtime-owned across surfaces Semantic compaction is the #981/#986 attention-first experiment, not the standard historyCompact mechanism. Make the runtime the owner of its default: it stays off unless a surface explicitly opts in (MAKA_CONTEXT_SEMANTIC_COMPACT truthy, or a non-off MAKA_CONTEXT_SEMANTIC_COMPACT_MODE). This lets the CLI drop its local semanticCompact strip and eat the runtime default directly, so the surfaces stop duplicating the same off decision. Desktop and headless inherit the off default without any per-surface wiring. Refs #882 * test(runtime): lock the shipped default's proactive long-turn journey Wire buildDefaultContextBudgetPolicy — the exact default every surface now inherits — into the streaming backend and prove a long turn near the window engages mid-turn compaction and continues to a clean end rather than being truncated or surfacing a raw provider error. The fake-backend Playwright suite cannot exercise this path: it is a separate canned-event backend that never runs AiSdkBackend's prepareStep machinery, so a desktop E2E could only fake the very compaction under test. This deterministic runtime-layer test is the honest representative journey for the proactive path. Refs #882 * fix(runtime): bound the derived compaction reserve by the model window The flat 16384 reserve assumed large-window models: on gpt-4 (8192 window) the default policy derived maxHistoryEstimatedTokens = max(1, 8192 - 16384) = 1 — a pre-existing turn-boundary pathology on main — and with mid_turn now on by default the 1-token high water made every multi-step turn run the summarizer for a checkpoint the replay gate could never admit. Make defaultCompactReserveTokens the single owner of the reserve default, bounded by the KNOWN window: min(16384, floor(window / 4)), so both the turn-boundary budget and the mid_turn high water derive from one value (8K window -> 2048 reserve / 6144 budget; >= 64K windows byte-identical to before; unknown window keeps the classic constant). An explicit MAKA_CONTEXT_HISTORY_COMPACT_RESERVE_TOKENS is respected verbatim. Peers bound the same way: opencode caps its compaction buffer by the model's output limit, gemini-cli triggers at a window fraction; pi shares the flat-16384 blind spot. Strengthen the shipped-defaults journey to fully-default derivation with durable-persistence assertions (checkpoint recorded, compact block present, replaced raw span gone), and add the unrescuable path: under the derived default a turn with no safe span ends with the explicit context_budget_exhausted outcome. A small-window guard locks the P2 behavior: on an 8K model the default is completely inert when under budget — never a pointless summarizer call. Refs #882 * docs(runtime): align midTurn policy comments with the shipped runtime default

…t-semantic-compaction # Conflicts: # packages/runtime/src/ai-sdk-backend.ts # packages/runtime/src/context-budget-policy.ts

…c CLI color tests) (#1077) * fix(runtime): let active prune cover the newest step under mid-turn capacity #986 unconditionally excluded the newest completed step from active tool-result pruning, breaking three reviewed capacity-rescue invariants (mid-turn finding C, capacity-replacement re-convergence, overflow round-3 P1) and CI on main. Include the newest step exactly when mid-turn capacity compaction is active, keeping the keep-newest-raw default for every other configuration. * test(cli): pin color level so ANSI assertions are hermetic tui-ansi detects color capability from TERM/COLORTERM at module load, so the pi-transcript and pi-tui-runner ANSI assertions inherited ambient terminal capability: truecolor locally, level 0 on CI runners with no TERM/COLORTERM, where every color function is a no-op and the #1064/#1066 color assertions fail deterministically. Pin level 3 via the existing _setColorLevelForTesting seam, matching tui-ansi.test.ts.

likun666661 · 2026-07-14T13:52:49Z

Summary

preserve prior replay and the exact current-turn user message as an immutable head anchor
replace only the structurally complete middle span before the open provider-protocol tail
use the 8K total-context / 4K completed-span attention trigger and separate generation / accepted-projection budgets
add bounded structured and complete-sentence/line fallback validation, V2 rolling lineage, and shared attention/emergency scheduling
keep the provider-visible continuation strictly exact head anchor + LLM-authored projection + exact protocol tail
keep Runtime coverage, source refs, archive refs, preserved-tail indexes, and legacy state cards in the diagnostic side channel only

Safety properties

the head anchor is signature-checked on every prepareStep and is never semantic coverage
consecutive assistant reasoning/text/tool-call messages form one provider episode; incomplete multi-call episodes remain verbatim in the tail
empty, truncated, source-invalid, private-verifier-invalid, protocol-invalid, or over-budget projections do not replace source messages
token savings, compact-call cost, and cache loss remain diagnostics in attention mode
one prepareStep accepts at most one attention or emergency replacement
emergency compaction reports explicit head_anchor_exceeds_capacity / no_safe_completed_span outcomes instead of rewriting the user instruction
semanticCompact no longer generates, prompts with, records on new blocks, or renders deterministic state cards
rolling successors re-render the predecessor's LLM summary and ignore legacy state cards instead of replaying the old provider message verbatim

MIPS smoke finding

A GLM-5.2 make-mips-interpreter smoke on the earlier version was stopped after it exceeded 135 steps without creating /app/vm.js.

The retained trace showed:

24 accepted semantic compactions and 228 tool calls
19 later summaries whose next action was to write /app/vm.js
provider-visible state cards that instead repeated assistant planning such as “re-read” or “re-verify”
zero creation of /app/vm.js

The deterministic VM matcher treated any assistant text containing vm as operational state. Those inferred cards were rendered after every LLM summary and then copied into rolling successors, creating a positive feedback loop. The fix in 4eebe953 removes that semantic path rather than attempting to repair it with more regexes.

This smoke is mechanism evidence, not an A/B quality result. The old main-branch arm separately failed to activate semantic compaction because of tool_pair_split and eventually timed out near provider capacity, so it is not a usable success baseline.

Validation

After merging current upstream/main:

focused semantic + backend + mid-turn capacity integration: 187 passed, 0 failed
npm --workspace @maka/headless test: 964 passed, 0 failed, 1 skipped
npm run typecheck: passes across all workspaces after rebuilding changed workspace dependencies
Runtime full suite: 1943 passed, 7 skipped, with one unrelated local-only failure in main's macOS sandbox executable-root index assertion because this machine runs Node from /usr/local/bin/node
git diff --check: passes

No attention benefit is claimed yet, and this PR does not expand any surface default. A fresh same-commit A/B is still required after the mechanism correction.

Refs #981

* feat(runtime): default mid-turn capacity compaction on, runtime-owned The mid_turn capacity invariant becomes a runtime-owned default: whenever history compaction is enabled the runtime derives midTurn from the selected model window (contextWindow - reserve), so Desktop, TUI, and headless inherit it without copying config. MAKA_CONTEXT_HISTORY_COMPACT_MID_TURN=off stays as the explicit escape hatch. The backend still gates activation on the checkpoint seams, the persisted head anchor, and a KNOWN context window, so a session without model metadata or a child session with no anchor seam never misfires even though the default is on (covered by new backend safety guards). Refs #882 * feat: default semantic compaction off, runtime-owned across surfaces Semantic compaction is the #981/#986 attention-first experiment, not the standard historyCompact mechanism. Make the runtime the owner of its default: it stays off unless a surface explicitly opts in (MAKA_CONTEXT_SEMANTIC_COMPACT truthy, or a non-off MAKA_CONTEXT_SEMANTIC_COMPACT_MODE). This lets the CLI drop its local semanticCompact strip and eat the runtime default directly, so the surfaces stop duplicating the same off decision. Desktop and headless inherit the off default without any per-surface wiring. Refs #882 * test(runtime): lock the shipped default's proactive long-turn journey Wire buildDefaultContextBudgetPolicy — the exact default every surface now inherits — into the streaming backend and prove a long turn near the window engages mid-turn compaction and continues to a clean end rather than being truncated or surfacing a raw provider error. The fake-backend Playwright suite cannot exercise this path: it is a separate canned-event backend that never runs AiSdkBackend's prepareStep machinery, so a desktop E2E could only fake the very compaction under test. This deterministic runtime-layer test is the honest representative journey for the proactive path. Refs #882 * fix(runtime): bound the derived compaction reserve by the model window The flat 16384 reserve assumed large-window models: on gpt-4 (8192 window) the default policy derived maxHistoryEstimatedTokens = max(1, 8192 - 16384) = 1 — a pre-existing turn-boundary pathology on main — and with mid_turn now on by default the 1-token high water made every multi-step turn run the summarizer for a checkpoint the replay gate could never admit. Make defaultCompactReserveTokens the single owner of the reserve default, bounded by the KNOWN window: min(16384, floor(window / 4)), so both the turn-boundary budget and the mid_turn high water derive from one value (8K window -> 2048 reserve / 6144 budget; >= 64K windows byte-identical to before; unknown window keeps the classic constant). An explicit MAKA_CONTEXT_HISTORY_COMPACT_RESERVE_TOKENS is respected verbatim. Peers bound the same way: opencode caps its compaction buffer by the model's output limit, gemini-cli triggers at a window fraction; pi shares the flat-16384 blind spot. Strengthen the shipped-defaults journey to fully-default derivation with durable-persistence assertions (checkpoint recorded, compact block present, replaced raw span gone), and add the unrescuable path: under the derived default a turn with no safe span ends with the explicit context_budget_exhausted outcome. A small-window guard locks the P2 behavior: on an 8K model the default is completely inert when under budget — never a pointless summarizer call. Refs #882 * docs(runtime): align midTurn policy comments with the shipped runtime default

…t-semantic-compaction # Conflicts: # packages/runtime/src/ai-sdk-backend.ts # packages/runtime/src/context-budget-policy.ts

…c CLI color tests) (#1077) * fix(runtime): let active prune cover the newest step under mid-turn capacity #986 unconditionally excluded the newest completed step from active tool-result pruning, breaking three reviewed capacity-rescue invariants (mid-turn finding C, capacity-replacement re-convergence, overflow round-3 P1) and CI on main. Include the newest step exactly when mid-turn capacity compaction is active, keeping the keep-newest-raw default for every other configuration. * test(cli): pin color level so ANSI assertions are hermetic tui-ansi detects color capability from TERM/COLORTERM at module load, so the pi-transcript and pi-tui-runner ANSI assertions inherited ambient terminal capability: truecolor locally, level 0 on CI runners with no TERM/COLORTERM, where every color function is a no-op and the #1064/#1066 color assertions fail deterministically. Pin level 3 via the existing _setColorLevelForTesting seam, matching tui-ansi.test.ts.

likun666661 · 2026-07-14T13:52:49Z

Summary

preserve prior replay and the exact current-turn user message as an immutable head anchor
replace only the structurally complete middle span before the open provider-protocol tail
use the 8K total-context / 4K completed-span attention trigger and separate generation / accepted-projection budgets
add bounded structured and complete-sentence/line fallback validation, V2 rolling lineage, and shared attention/emergency scheduling
keep the provider-visible continuation strictly exact head anchor + LLM-authored projection + exact protocol tail
keep Runtime coverage, source refs, archive refs, preserved-tail indexes, and legacy state cards in the diagnostic side channel only

Safety properties

the head anchor is signature-checked on every prepareStep and is never semantic coverage
consecutive assistant reasoning/text/tool-call messages form one provider episode; incomplete multi-call episodes remain verbatim in the tail
empty, truncated, source-invalid, private-verifier-invalid, protocol-invalid, or over-budget projections do not replace source messages
token savings, compact-call cost, and cache loss remain diagnostics in attention mode
one prepareStep accepts at most one attention or emergency replacement
emergency compaction reports explicit head_anchor_exceeds_capacity / no_safe_completed_span outcomes instead of rewriting the user instruction
semanticCompact no longer generates, prompts with, records on new blocks, or renders deterministic state cards
rolling successors re-render the predecessor's LLM summary and ignore legacy state cards instead of replaying the old provider message verbatim

MIPS smoke finding

A GLM-5.2 make-mips-interpreter smoke on the earlier version was stopped after it exceeded 135 steps without creating /app/vm.js.

The retained trace showed:

24 accepted semantic compactions and 228 tool calls
19 later summaries whose next action was to write /app/vm.js
provider-visible state cards that instead repeated assistant planning such as “re-read” or “re-verify”
zero creation of /app/vm.js

The deterministic VM matcher treated any assistant text containing vm as operational state. Those inferred cards were rendered after every LLM summary and then copied into rolling successors, creating a positive feedback loop. The fix in 4eebe953 removes that semantic path rather than attempting to repair it with more regexes.

This smoke is mechanism evidence, not an A/B quality result. The old main-branch arm separately failed to activate semantic compaction because of tool_pair_split and eventually timed out near provider capacity, so it is not a usable success baseline.

Validation

After merging current upstream/main:

focused semantic + backend + mid-turn capacity integration: 187 passed, 0 failed
npm --workspace @maka/headless test: 964 passed, 0 failed, 1 skipped
npm run typecheck: passes across all workspaces after rebuilding changed workspace dependencies
Runtime full suite: 1943 passed, 7 skipped, with one unrelated local-only failure in main's macOS sandbox executable-root index assertion because this machine runs Node from /usr/local/bin/node
git diff --check: passes

No attention benefit is claimed yet, and this PR does not expand any surface default. A fresh same-commit A/B is still required after the mechanism correction.

Refs #981

* feat(runtime): default mid-turn capacity compaction on, runtime-owned The mid_turn capacity invariant becomes a runtime-owned default: whenever history compaction is enabled the runtime derives midTurn from the selected model window (contextWindow - reserve), so Desktop, TUI, and headless inherit it without copying config. MAKA_CONTEXT_HISTORY_COMPACT_MID_TURN=off stays as the explicit escape hatch. The backend still gates activation on the checkpoint seams, the persisted head anchor, and a KNOWN context window, so a session without model metadata or a child session with no anchor seam never misfires even though the default is on (covered by new backend safety guards). Refs #882 * feat: default semantic compaction off, runtime-owned across surfaces Semantic compaction is the #981/#986 attention-first experiment, not the standard historyCompact mechanism. Make the runtime the owner of its default: it stays off unless a surface explicitly opts in (MAKA_CONTEXT_SEMANTIC_COMPACT truthy, or a non-off MAKA_CONTEXT_SEMANTIC_COMPACT_MODE). This lets the CLI drop its local semanticCompact strip and eat the runtime default directly, so the surfaces stop duplicating the same off decision. Desktop and headless inherit the off default without any per-surface wiring. Refs #882 * test(runtime): lock the shipped default's proactive long-turn journey Wire buildDefaultContextBudgetPolicy — the exact default every surface now inherits — into the streaming backend and prove a long turn near the window engages mid-turn compaction and continues to a clean end rather than being truncated or surfacing a raw provider error. The fake-backend Playwright suite cannot exercise this path: it is a separate canned-event backend that never runs AiSdkBackend's prepareStep machinery, so a desktop E2E could only fake the very compaction under test. This deterministic runtime-layer test is the honest representative journey for the proactive path. Refs #882 * fix(runtime): bound the derived compaction reserve by the model window The flat 16384 reserve assumed large-window models: on gpt-4 (8192 window) the default policy derived maxHistoryEstimatedTokens = max(1, 8192 - 16384) = 1 — a pre-existing turn-boundary pathology on main — and with mid_turn now on by default the 1-token high water made every multi-step turn run the summarizer for a checkpoint the replay gate could never admit. Make defaultCompactReserveTokens the single owner of the reserve default, bounded by the KNOWN window: min(16384, floor(window / 4)), so both the turn-boundary budget and the mid_turn high water derive from one value (8K window -> 2048 reserve / 6144 budget; >= 64K windows byte-identical to before; unknown window keeps the classic constant). An explicit MAKA_CONTEXT_HISTORY_COMPACT_RESERVE_TOKENS is respected verbatim. Peers bound the same way: opencode caps its compaction buffer by the model's output limit, gemini-cli triggers at a window fraction; pi shares the flat-16384 blind spot. Strengthen the shipped-defaults journey to fully-default derivation with durable-persistence assertions (checkpoint recorded, compact block present, replaced raw span gone), and add the unrescuable path: under the derived default a turn with no safe span ends with the explicit context_budget_exhausted outcome. A small-window guard locks the P2 behavior: on an 8K model the default is completely inert when under budget — never a pointless summarizer call. Refs #882 * docs(runtime): align midTurn policy comments with the shipped runtime default

…t-semantic-compaction # Conflicts: # packages/runtime/src/ai-sdk-backend.ts # packages/runtime/src/context-budget-policy.ts

…c CLI color tests) (#1077) * fix(runtime): let active prune cover the newest step under mid-turn capacity #986 unconditionally excluded the newest completed step from active tool-result pruning, breaking three reviewed capacity-rescue invariants (mid-turn finding C, capacity-replacement re-convergence, overflow round-3 P1) and CI on main. Include the newest step exactly when mid-turn capacity compaction is active, keeping the keep-newest-raw default for every other configuration. * test(cli): pin color level so ANSI assertions are hermetic tui-ansi detects color capability from TERM/COLORTERM at module load, so the pi-transcript and pi-tui-runner ANSI assertions inherited ambient terminal capability: truecolor locally, level 0 on CI runners with no TERM/COLORTERM, where every color function is a no-op and the #1064/#1066 color assertions fail deterministically. Pin level 3 via the existing _setColorLevelForTesting seam, matching tui-ansi.test.ts.

feat(runtime): add attention-first semantic compaction
2298953

This was referenced Jul 14, 2026
feat(runtime): compact active turns before context exhaustion #882
Closed
feat(runtime): mid-turn capacity compaction with a single final-payload verdict owner #996
Merged

fix(runtime): keep semantic compaction LLM-authored
4eebe95

likun666661 mentioned this pull request Jul 15, 2026
RFC: attention shaping over completed history after an immutable user head anchor #981
Closed

likun666661 added 4 commits July 15, 2026 15:02

Merge remote-tracking branch 'upstream/main' into feat/attention-firs…
bb4224b
…t-semantic-compaction # Conflicts: # packages/runtime/src/ai-sdk-backend.ts # packages/runtime/src/context-budget-policy.ts

fix(runtime): preserve compaction continuation momentum
b97eea7

fix(runtime): preserve semantic projection across active pruning
f3b7a76

fix(runtime): preserve execution momentum in semantic compaction
85d54ea

likun666661 merged commit 01fc111 into apache:mainJul 16, 2026
2 of 3 checks passed

likun666661 mentioned this pull request Jul 16, 2026
fix(runtime): lower semantic compaction high water to 128K #1069
Merged

Astro-Han mentioned this pull request Jul 16, 2026
fix: restore green CI on main (runtime prune regression + non-hermetic CLI color tests) #1077
Merged