Context as a Resource
Organizations: East China University of Science and Technology, Shanghai, China
Abstract
Managing concurrent tasks across separate agent sessions quickly exhausts human attention, motivating always-on personal agents behind a single continuous interface. Yet when every context shares identical frozen backbone weights, funneling interleaved dialogues and external events into one linear session induces cross-topic constraint bleed and head-of-line stalling. Conversely, router-ticket hierarchies incur lossy prose handoffs, heavy routing overhead, and blind merges - collapsing true-conflict repository merges to a 33.3% three-run mean despite high isolated branch passes. Rather than scheduling anthropomorphic roles, we introduce ClonAct, a routerless architecture that virtualizes context blocks directly through a pluggable single-pass non-generative Context MMU, in-place Context Cells, a footprint-aware Synaptic Bus with an isolated trial merge gate, and a wave-flushed Shared Trunk. Across three programmatic multi-stream benchmarks and three model families, our portable trial merge gate lifts true-conflict merges to 63.3%, while ClonAct's routerless virtualization attains this ceiling with 45.9% fewer prompt tokens and prevents compact-model router collapse. Under continuous surge streams, ClonAct matches top completion (95.0%) with the lowest median wait and zero timed-out status checks.
Figures & tables
| Synaptic-Bench-150 ( Multi-Threaded Episodes) | AppWorld-Micro-30 ( ) | ||||||||
| Orchestration Paradigm | Pass Rate | Safety | C-2 All/True | H-3 | Peek | Tokens | Pass | TGC | Tokens |
| DSH-Mono | 66.0% † ( 3.8) | 93.8% | 70.0 / 70.0% | 33.3% | 28.3s | 2.44M | 93.3% | 96.7% | 439k |
| DSH-Subagent-Fork | 68.7% ( 3.8) | 93.2% | 60.0 / 60.0% | 41.7% | 38.2s | 2.77M | 90.0% | 95.0% | 637k |
| DSH-AgentTeam | 60.0% ‡ ( 4.4) | 90.7% | 50.0 / 60.0% | 41.7% | 42.0s | 3.80M | 90.0% | 95.0% | 851k |
| DSH-SuperAgent | 72.0% ( 3.5) | 98.1% | 45.0 / 10.0% | 66.7% | 11.9s | 6.36M | 76.7% | 85.0% | 1,576k |
| ClonAct (Ours) | 74.7% ( 3.5) | 100.0% | 70.0 / 80.0% | 91.7% | 7.3s | 3.45M | 90.0% | 95.0% | 696k |
| (a) Track C-2 Crossover Control ( : True-Conflict + Disjoint Pairs) | ||||||||
|---|---|---|---|---|---|---|---|---|
| Arm / Control | C-2 All | C-2 True | C-2 Disj. | Safety | Tok/Ep | USD (20 ep) | Wall | Peek |
| DSH-SuperAgent (vanilla) | 45.0% | 10.0% | 80.0% | 100.0% | 10.50M | $2.35 | 488.8s | 15.8s |
| DSH-AgentTeam | 50.0% | 60.0% | 40.0% | 70.0% | 3.97M | $1.17 | 338.3s | 43.5s |
| DSH-SuperAgent + Trial Merge | 70.0% | 80.0% | 60.0% | 96.7% | 12.97M | $2.73 | 549.0s | 13.8s |
| ClonAct (Ours) | 70.0% | 80.0% | 60.0% | 100.0% | 6.33M | $1.49 | 370.7s | 9.4s |
| Ablation / Control Arm | Raw | IPW | C-2 | C-1+O-1 | Wall | M-30 |
|---|---|---|---|---|---|---|
| ClonAct (Full, Jev) | 80.0% | 81.6% | 83.3% | 81.8% | 294.0s | 90.0% |
| w/ Logprob-MMU (same- ) | 81.7% | 83.7% | 75.0% | 81.8% | 272.8s | 80.0% |
| w/o Same-Brain | 63.3% | 64.3% | 66.7% | 27.3% | 325.3s | 93.3% |
| w/o Prefix & Grd. | 70.0% | 71.6% | 75.0% | 63.6% | 341.5s | 96.7% |
| w/ Generative Judge | 66.7% | 73.4% | 41.7% | 90.9% | 320.5s | 100.0% |
| w/o Synapse | 66.7% | 73.6% | 16.7% ‡ | 81.8% | 301.7s | 90.0% |
| Bench-60 Stratified ( ) | Micro-30 ( ) | Combined-90 & System Efficiency | |||||||||
|---|---|---|---|---|---|---|---|---|---|---|---|
| Model Family | Paradigm | Raw | IPW | C-2 | H-3 | Pass | Hyg. | Comb. Ep. | Verif. | Peek | Bench-60 Tok |
| DeepSeek-V4.1-Flash-1017 | DSH-Mono | 70.0% | 72.7% | 75.0% | 60.0% | 93.3% | 100.0% | 77.8% | 89.5% | 38.3s | 2.61M |
| DSH-SuperAgent | 66.7% | 75.4% | 25.0% | 60.0% | 76.7% | 80.0% | 70.0% † | 85.9% | 11.9s | 6.86M | |
| ClonAct | 80.0% | 81.6% | 83.3% | 100.0% | 90.0% | 100.0% | 83.3% | 93.8% | 8.6s | 4.01M | |
| Qwen3.8-Flash | DSH-Mono | 48.3% | 55.6% | 16.7% | 40.0% | 56.7% | 83.3% | 51.1% | 71.7% | 67.0s | 0.91M |
| DSH-SuperAgent | 43.3% | 48.5% | 8.3% | 60.0% | 83.3% | 83.3% | 56.7% | 75.7% | 23.2s | 3.50M | |
Appendix figures & tables14 assets
Supplementary material from the paper’s appendix.
Appendix
| Symbol | Type / Domain | Definition |
|---|---|---|
| Frozen policy | Inner single-session execution kernel (unmodified DSH with backbone weights ). | |
| Orchestrator | Outer context-scheduling policy mapping unfiltered event streams to model calls and tools. | |
| Mutable state | Shared workspace at step (Git repositories, databases, document trees, personal apps). | |
| Event stream | Arrival-ordered sequence ( ) of raw events , timestamps , and channels . | |
| rates | Undisturbed single-task oracle, isolated local pass rate, and joint post-merge pass rate ( Equation 1 ). | |
| gap | Concurrency degradation decomposed into cross-stream bleed and post-merge loss ( Equation 1 ). |
| Suite / Domain | Track | Units | Upstream Substrates | Core Concurrent Interaction Tested |
| Office ( ) | O-1 | 15 | Gaia2, MCPMark | clarification + read-only side query + execute |
| O-2 | 15 | Gaia2, MCPMark (FS/PG) | Multi-channel async bursts with chatter & duplicate alerts | |
| O-3 | 15 | Gaia2, MCPMark | Step-Hook mid-flight constraint amendment / goal overturn | |
| Coding ( ) | C-1 | 15 | Ambig-SWE, SWE-Interact | Root-cause alignment config lookup in-place fix |
| C-2 | 20 | CooperBench | Dual-feature same-repo concurrency ( true conflicts, clean) | |
| C-3 | 20 | SWE-Interact, Verified | Step-Hook mid-fix spec addition or clean rollback on overturn |
| Suite / Instance | Trigger / Arrival | Channel & Sender | Observed Raw Event Excerpt |
| Bench-150 C-3 PA150_C3_01 ( wireservice/csvkit#1274 ) | immediate ( ) | ui://chat/main ( user ) | “Please fix this issue in wireservice/csvkit : Can in2csv add a byte order mark (BOM) so that when opening csv in Excel it correctly formats unicode…” |
| after_worker_tool_calls ( ) | ui://chat/main ( user ) | “Addendum for csvkit-1274 (merge directly into your active implementation): This can be fixed by appending the correct BOM when writing the csv…” | |
| Stream-60 S-1 STREAM_01 (Click/Jinja/DSPy, events) | Mod / Surge | feishu://p2p/user ( user_01 ) | [Code Indep., T01] “Could you work on ‘Add Validation Chain Support to Click Options’ in pallets/click (under projects/click/ )? …” |
| Mod / Surge | feishu://group/team_general ( maya_design ) | [Chatter] “Heads up team: Figma library tokens for the Q3 icon refresh just finished syncing.” | |
| Mod / Surge | feishu://group/ops_alerts ( marco_sre ) | [Repost, T01] “@agent Sharing the earlier request here so everyone has context: Could you work on ‘Add Validation Chain Support to Click Options’…” | |
| Mod / Surge | feishu://p2p/user ( user_01 ) | [Amendment, T01] “One detail for Add Validation Chain Support to Click Options in pallets/click : Also include validators in Option.to_info_dict() …” |
| Component | Parameter / Identifier | Pinned Value / Specification |
| System-2 Backbones | DeepSeek-V4.1-Flash-1017 | Anthropic Messages wire protocol; max_tokens=32768 ; official off-peak USD/1M: 0.15 miss, $0.60 out. |
| Qwen3.8-Flash | OpenAI-compatible completions; enable_thinking=True , reasoning_effort="xhigh" ; CNY/1M: ¥0.05 / ¥0.50 / ¥2.00. | |
| GPT-6-Luna | OpenAI Responses API; reasoning.effort="high" ; USD/1M: 0.10 miss, $0.50 out. | |
| ClonAct Context MMU | Single-Pass System-1 Backend | TypeSafe jev-1.14.0 ( AsyncTypeSafeClient , Choice / Noul ); official USD/1M: 0.00 output (structured probabilities unbilled); same-weight max_tokens=1, logprobs=True fallback ( top_logprobs=5 for binary Noul ; uppercase letter mapping A,B,C,… with top_logprobs=min(20, max(5, |options|+2)) for Choice ). |
| Decision Thresholds | Actionability ; coreference ; L3 recall . | |
| State Serialization Caps | Normalized topic snippet chars; workspace AGENTS.md excerpt chars. |
| System-2 Backbone (DeepSeek-V4.1-Flash) | System-1 | Runtime | Dual-Tier Cost (USD) | ||||||||
| Suite / Regime | Paradigm | S2 Hit | S2 Miss | S2 Out | S2 Hit% | Jev In | Sess. | WT | S2 | S1 | Total |
| Bench-150 ( totals) | DSH-Mono | 354.8M | 6.54M | 4.40M | 98.2% | — | 150 | 0 | $4.69 | — | $4.69 |
| DSH-Subagent-Fork | 401.7M | 8.45M | 5.92M | 97.9% | — | 245 | 22 | $6.02 | — | $6.02 | |
| DSH-AgentTeam | 549.4M | 12.15M | 7.79M | 97.8% | — | 377 | 127 | $8.15 | — | $8.15 | |
| DSH-SuperAgent | 928.1M | 15.57M | 10.34M | 98.4% | — | — | 295 | $11.33 | — | $11.33 | |
| ClonAct (Ours) | 497.1M | 10.62M | 6.20M | 97.9% | 2.86M | 290 | 148 | $6.80 | $0.12 | $6.92 | |
| Artifact / Arm | Verbatim Prompt / Schema Specification |
|---|---|
| ClonAct Static Trunk Prompt ( in ) | “You are Synaptic-PA [codebase alias for ClonAct ] , the user’s unified Personal Agent operating over a shared epistemic trunk . Core principles: (1) Speak in first-person concisely; (2) Never invent unverified workspace, database, API, or document facts; inspect the workspace and run checks to verify; (3) Execute requested tasks end-to-end inside your workspace while respecting explicit user safety/read-only instructions; (4) Respect any [SYNAPTIC ...] or [USER FOLLOW-UP MESSAGE] blocks appended to your context.” Followed by workspace <workspace_agents_md> conventions and . |
| ClonAct Read-Only Tool ( peek_peer_cell ) | peek_peer_cell(cell_id: str, rel_path?: str) : “Instant zero-side-effect read-only inspection ( 0ms ) of a sibling Context Cell’s live progress, verified state deltas, recent tool outputs, file diff, and final completion summary. Pass cell_id (e.g. ’Cell_E_1’ or ’ALL’ ) and optional rel_path to inspect the sibling cell’s detailed git diff on a specific shared file.” |
| ClonAct Context MMU (4 Parallel Non-Generative Choice / Noul Queries) | (Q1) External Disposition: Choice over { new_workspace_task , non_actionable_chatter , duplicate_already_handled } + Noul actionability score . (Q2) Turn Granularity: Choice over { single_intent , coupled_causal_workflow , orthogonal_parallel_tasks }. (Q3) Cell Coreference & Settled Follow-Up: Choice over active/parked/completed cell IDs using compact cell timeline summaries ( chars/topic), plus settled_followup_kind distinguishing pure_duplicate_already_handled from genuine follow-up edits. (Q4) Read/Write Mode & Turn-Local Scoping: Choice over align_or_query vs. execute_now , direct_deliverable vs. derived_observer_brief , and turn_local_readonly_guard vs. unrelated , scoping read-only requests to zero-worktree inspection/observer cells without suppressing concurrent write cells. Same-Weight Logprob Fallback: Evaluates binary Noul queries via max_tokens=1, logprobs=True, top_logprobs=5 on , and maps categorical Choice options to single uppercase letters ( A, B, C, … ) with top_logprobs=min(20, max(5, |options|+2)) , normalizing exponentiated candidate log-probabilities. |
| Shared Baseline Rules (DSH-Mono, DSH-Subagent-Neutral, DSH-Subagent-Fork, DSH-AgentTeam) | “(1) Speak concisely in first-person; (2) Never fabricate workspace, code, DB, or app facts; inspect files/APIs and run verification checks; (3) When a user request is underspecified/ambiguous or explicitly asks for read-only analysis before modifying code or data, inspect in READ-ONLY mode (zero file edits, zero DB/app mutations, zero new git worktrees) and ask concise clarifying questions; (4) Ignore non-actionable group-chat banter, facility/HR broadcasts, or courtesy acks without modifying workspace state or spawning workers, and do not re-execute duplicate cross-channel alerts; (5) Honour the latest user overturn or constraint update cleanly.” All baselines also receive the identical <workspace_agents_md> workspace conventions block. |
| Baseline Mode Prompts | DSH-Mono: Single continuous ReAct session ( subagent disabled); include next tool call when answering live status queries while tasks remain open. DSH-Subagent-Neutral: Exposes subagent , subagent_fork , list_agents , send_message , interrupt_agent ; agent decides autonomously whether to run directly or delegate (optional worktree note). DSH-Subagent-Fork: Our Delegation-First prompt, adapted from Qiao (2026) with harness-specific tool names: “Prefer handing each distinct piece of implementation or execution work to a background subagent ( background: true or subagent_fork ), and do only small, fast things yourself” , supplemented in our harness with explicit Git worktree creation/merge commands under ../worktrees/<task_slug> . DSH-AgentTeam: Write-gated Team Lead ( write / edit denied) delegating via team_spawn_teammate(use_worktree) and integrating via team_board , team_mailbox , and team_merge_worktree . |
| Arm | Off. | Cod. | Hyb. | O-1 | O-2 | O-3 | C-1 | C-2 | C-3 | H-1 | H-2 | H-3 | H-4 |
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| DSH-Mono | 68.9 | 74.5 | 54.0 | 46.7 | 73.3 | 86.7 | 73.3 | 70.0 | 80.0 | 46.7 | 73.3 | 33.3 | 62.5 |
| DSH-Subagent-Fork | 77.8 | 69.1 | 60.0 | 66.7 | 80.0 | 86.7 | 73.3 | 60.0 | 75.0 | 80.0 | 53.3 | 41.7 | 62.5 |
| DSH-AgentTeam | 68.9 | 60.0 | 52.0 | 60.0 | 66.7 | 80.0 | 66.7 | 50.0 | 65.0 | 60.0 | 53.3 | 41.7 | 50.0 |
| DSH-SuperAgent | 73.3 | 67.3 | 76.0 | 60.0 | 80.0 | 80.0 | 73.3 | 45.0 | 85.0 | 80.0 | 80.0 | 66.7 | 75.0 |
| ClonAct (Lite Loop) | 53.3 | 63.6 | 54.0 | 53.3 | 40.0 | 66.7 | 66.7 | 50.0 | 75.0 | 53.3 | 46.7 | 66.7 | 50.0 |
| ClonAct (DSH) | 71.1 | 74.5 | 78.0 | 60.0 | 80.0 | 73.3 | 73.3 | 70.0 | 80.0 | 73.3 | 73.3 | 91.7 | 75.0 |
| Oracle (DSH-Mono) | ClonAct ( Independent Runs) | DSH-SuperAgent ( Independent Runs) | |||||||||
| Episode ID | Repo / Task | Feat A | Feat B | Run 1 | Run 2 | Run 3 | Loc (/3) | Run 1 | Run 2 | Run 3 | Loc (/3) |
| PA150_C2_11 | pallets/click#2068 | – | – | – | 3/3 | – | 3/3 | ||||
| PA150_C2_12 | pallets/jinja#1465 | 3/3 | – | – | – | 3/3 | |||||
| PA150_C2_13 | dirty-equals#43 | – | 3/3 | – | – | – | 3/3 | ||||
| PA150_C2_14 | outlines#1371 | 3/3 | 3/3 | ||||||||
| PA150_C2_15 | datasets#3997 | – | – | 3/3 | – | 3/3 | |||||
| Ablation / Control Arm | Off. | Cod. | Hyb. | O-1 | O-2 | O-3 | C-1 | C-2 | C-3 | H-1 | H-2 | H-3 | H-4 |
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| ClonAct (Full, Jev) | 83.3 | 86.4 | 70.0 | 83.3 | 83.3 | 83.3 | 80.0 | 83.3 | 100.0 | 66.7 | 50.0 | 100.0 | 66.7 |
| w/ Logprob-MMU (same- ) | 88.9 | 81.8 | 75.0 | 66.7 | 100.0 | 100.0 | 100.0 | 75.0 | 80.0 | 83.3 | 50.0 | 80.0 | 100.0 |
| w/o Same-Brain | 61.1 | 59.1 | 70.0 | 16.7 | 66.7 | 100.0 | 40.0 | 66.7 | 60.0 | 83.3 | 50.0 | 80.0 | 66.7 |
| w/o Prefix & Grd. | 66.7 | 77.3 | 65.0 | 50.0 | 66.7 | 83.3 | 80.0 | 75.0 | 80.0 | 83.3 | 16.7 | 100.0 | 66.7 |
| w/ Generative Judge | 72.2 | 68.2 | 60.0 | 83.3 | 66.7 | 66.7 | 100.0 | 41.7 | 100.0 | 83.3 | 50.0 | 80.0 | 0.0 |
| w/o Synapse | 83.3 | 50.0 | 70.0 | 83.3 | 66.7 | 100.0 | 80.0 | 16.7 | 100.0 | 83.3 | 50.0 | 80.0 | 66.7 |
| Verified Survival | Waiting Area ( ) | Latency (s) | Observability | Verified by Category (%) | Stability | Compute | ||||||||
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| Regime | Paradigm | RMW | vs. ClonAct ( CI) | All | Cond | Peek (Ans/48) | QA | AppWr | Code | Regr. | Tok/Str | |||
| Surge ( rate, /req) | ClonAct (Ours) | 50.0% | 76.7% | 95.0% | 119.4s | — | 53.6s | 51.9s | 9.2s (48/48) | 94.4 | 97.0 | 88.9 | 0 / 0 | 27.4M |
| DSH-AgentTeam | 33.3% | 83.3% | 95.0% | 128.4s | [-17.0, 32.8] | 87.9s | 87.2s | 46.2s (15/48) | 94.4 | 97.0 | 88.9 | 0 / 0 | 22.1M | |
| DSH-SuperAgent | 33.3% | 80.0% | 93.3% | 142.7s | [-12.2, 56.3] | 79.6s | 78.0s | 5.9s (45/48) | 94.4 | 97.0 | 77.8 | 0 / 0 | 50.0M | |
| DSH-Subagent-Fork | 33.3% | 76.7% | 88.3% | 172.8s | [9.5, 100.6] | 91.7s | 77.9s | 54.7s (9/48) | 88.9 | 87.9 | 88.9 | 0 / 0 | 21.4M | |
| DSH-Mono | 11.7% | 45.0% | 90.0% | 222.5s | [72.9, 134.1] | 187.6s | 175.1s | 43.7s (8/48) | 94.4 | 87.9 | 88.9 | 1 / 0 | 11.4M | |
| Surge Arrival Regime ( /req) | Moderate Arrival Regime ( /req) | ||||||||||
|---|---|---|---|---|---|---|---|---|---|---|---|
| Stream (3-Project Monorepo) | Paradigm | All | Peek (Ans/16) | Tokens | All | Peek (Ans/16) | Tokens | ||||
| S-1 (Click / Jinja2 / DSPy) | ClonAct | 20/20 (100%) | 69.3s | 38.1s | 6.1s (16/16) | 17.1M | 20/20 (100%) | 59.4s | 30.1s | 8.1s (16/16) | 16.5M |
| DSH-SuperAgent | 18/20 (90%) | 138.1s | 54.6s | 4.8s (16/16) | 47.9M | 17/20 (85%) | 160.8s | 64.2s | 4.3s (15/16) | 80.4M | |
| DSH-AgentTeam | 20/20 (100%) | 100.4s | 78.0s | 47.0s (7/16) | 20.2M | 18/20 (90%) | 148.8s | 63.9s | 28.6s (10/16) | 29.1M | |
| DSH-Subagent-Fork | 19/20 (95%) | 113.1s | 53.3s | 53.5s (6/16) | 19.7M | 19/20 (95%) | 104.3s | 39.7s | 36.5s (10/16) | 21.9M | |
| DSH-Subagent-Neutral | 19/20 (95%) | 114.5s | 74.6s | 56.2s (1/16) | 14.0M | 20/20 (100%) | 44.4s | 22.4s | 31.5s (11/16) | 15.5M | |
| Bench-60 Domains (%) | Bench-60 Safety & Slices (%) | Micro-30 (%) | Bench-60 Compute & Cache | ||||||||||
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| Model Family | Paradigm | Off. | Cod. | Hyb. | Safety | O-1 | C-2T | M-1 | M-3 | Wall | Pr. Tok | Out. Tok | Cache |
| DeepSeek-V4.1 | DSH-Mono | 66.7 | 81.8 | 60.0 | 92.7 | 50.0 | 70.0 | 90.0 | 100.0 | 212.1s | 154.6M | 1.80M | 98.3% |
| DSH-SuperAgent | 83.3 | 54.5 | 65.0 | 97.1 | 83.3 | 10.0 | 50.0 | 90.0 | 431.8s | 407.0M | 4.47M | 98.4% | |
| ClonAct | 83.3 | 86.4 | 70.0 | 100.0 | 83.3 | 80.0 | 80.0 | 100.0 | 294.0s | 237.8M | 2.77M | 97.6% | |
| Qwen3.8-Flash | DSH-Mono | 72.2 | 36.4 | 40.0 | 88.2 | 66.7 | 0.0 | 70.0 | 30.0 | 318.5s | 53.6M | 1.25M | 94.0% |
| DSH-SuperAgent | 66.7 | 27.3 | 40.0 | 86.8 | 16.7 | 0.0 | 50.0 | 100.0 | 660.5s | 207.0M | 3.10M | 95.5% | |
| Benchmark / Regime | Error Stratum | Count | Exact Track / Category Breakdown & Root Mechanism |
|---|---|---|---|
| Synaptic-Bench-150 ( failed; Tier-1 safety) | Tier-1 Safety / Hygiene | ( ) Tier-1 checks passed; zero bleed or rollback violations. | |
| Shared Backbone Ceiling (All Arms Fail) | Coding (2 C-1, 4 C-2, 3 C-3), Hybrid (2 H-1, 3 H-2, 1 H-3), Office (2 O-1, 2 O-2, 1 O-3); bounded by . | ||
| ClonAct -Specific Residual (Tier-2 Only) | Coding (2 C-1, 2 C-2, 1 C-3), Hybrid (2 H-1, 1 H-2, 2 H-4), Office (4 O-1, 1 O-2, 3 O-3); incomplete multi-turn Office follow-through or edge-case patch misses. | ||
| Stream-60 Surge ( failed; regressions) | Independent Code Edge Case | STREAM_02_T01 (Click feature3 ): tests failed due to missing batch_size keyword forwarding in Parameter.init . | |
| Multi-Step AppWorld Write | STREAM_02_T14 ( acct_08 ): partial assertion pass within . | ||
| Read-Only QA Lookup | STREAM_03_T09 ( acct_14 ): answer snapshot unverified within ( QA passed overall). |
| Episode & Setup | DSH-SuperAgent (Router + Blind Tickets) | ClonAct (Virtualized Context Cells) |
|---|---|---|
| PA150-C2-12 (Jinja2 filters.py : case_sensitive + reverse on groupby() ) | 1. Router dispatches two isolated mandates into blind worktrees. 2. Worker 1 and Worker 2 each edit sync_do_groupby / do_groupby and pass local unit tests. 3. Worker 1 merges cleanly; Worker 2’s sequential merge overwrites the signature without case_sensitive , failing main-branch pytest with TypeError . | 1. MMU routes Task 1 Cell 1 and Task 2 Cell 2. 2. Tier-0 AST hook detects shared symbol sync_do_groupby ; Cell 2 invokes four read-only peek_peer_cell diff reads. 3. Cell 1 settles first; Cell 2 receives trunk-advance notice, unifies (..., case_sensitive=False, reverse=False) in trial merge, and passes full pytest on main. |
| PA150-H3-10 (Outlines types.guid + read-only release brief) | 1. External issue opens coding ticket for types.guid . 2. User asks in chat for a release cross-check brief: “read-only tracking—do not modify business code.” 3. Top-level router conflates foreground caveat with active mandate, suppressing all file edits ( files changed AttributeError ). | 1. Cell 1 executes types.guid implementation inside isolated worktree. 2. MMU classifies foreground brief as derived read-only request and spawns zero-worktree Observer Cell ( ) under . 3. Cell 1 lands passing tests undisturbed; Observer Cell reads settled milestone via Synapse and delivers verified brief. |
| Micro-30 M-1 (GPT-6-Luna: dual-app errands + chatter) | 1. Router decodes full event log on every chatter message ( k tokens/episode). 2. Compact-model MasterAssessor drops exact recipient/date constraints while paraphrasing prose tickets or mis-rejects personal app requests ( M-1 pass, hygiene). | 1. Single-pass MMU drops bystander chatter and duplicate alerts at entry ( hygiene). 2. Actionable errands route verbatim into independent Context Cells ( ) with zero ticket paraphrasing, sustaining M-1 pass with fewer tokens ( k). |
LLM Context Management
Large language model context management studies how to select, compress, retrieve, and organize information within bounded model inputs.
See all 267 papers