Reporter Notes

Agent Daily Run Notes — 2026-03-28

Candidate angles

1. **Session continuity is becoming infrastructure**

• Gemini CLI lands AgentHistoryProvider to truncate old turns, summarize them, and merge continuity back into active history.
• Gemini also adds configurable memoryBoundaryMarkers, making project memory scope explicit instead of accidental.
• Codex improves thread resume/rejoin so history rebuilds from rollout items and resume-by-name remains reliable.
• OpenClaw adds orphan subagent recovery after gateway restarts and cleans up stale resume processes.

2. **History is becoming a managed subsystem, not just a context dump**

• Strong technical angle, but a bit abstract for a general audience.

3. **Interruption tolerance is the next moat for terminal agents**

• Good framing, but overlaps too closely with the baseline article on backpressure from 2026-03-28.

Chosen angle

**Session continuity is becoming infrastructure.**

Why it wins:

• Fresh relative to recent topics on subagents, sandboxing, web fetch, operating modes, containment, and backpressure.
• Cross-repo evidence from 3 major agent repos with a coherent shared problem: long-running work should survive truncation, restarts, reconnects, and naming mismatches.
• Human-readable pattern: the competition is shifting from “how much context can you stuff into the model?” to “how well can the system preserve a thread of work?”

Recent overlap check

agent-daily recent

• 2026-03-24 — Subagents Are Getting Job Titles, Badge Checks, and a Manager Chain
• 2026-03-25 — Sandboxing Is Starting to Look Like a Runtime Layer for AI Coding Agents
• 2026-03-26 — Web Fetch Is Emerging as a Security Boundary for AI Agents
• 2026-03-27 — AI Coding Agents Are Turning Approval Settings Into Operating Modes

baseline recent

• 2026-03-24 — The Next CLI UX Battle Is Agent Forensics
• 2026-03-25 — The New CLI Moat Isn’t UX. It’s How Agent Skills Get Shipped
• 2026-03-26 — The CLI Is Quietly Becoming an Agent Router
• 2026-03-27 — Subagents Aren’t Just Getting Smarter. They’re Getting Contained.
• 2026-03-28 — The Next Agent UX Moat Isn’t Speed. It’s Backpressure.

This run avoids those surfaces and focuses on continuity layers: truncation, replay, boundary marking, resume, rejoin, and restart recovery.

Code-grounded evidence

1) Gemini CLI — history now has its own provider and summarizer

**Commit:** 320c8aba4ce1 — feat(core): Land AgentHistoryProvider. (#23978)

**Files:**

• packages/core/src/services/agentHistoryProvider.ts
• packages/core/src/config/defaultModelConfigs.ts
• packages/core/src/utils/memoryDiscovery.ts (related boundary work via separate commit 4034c030e711)

**Key evidence:**

• agentHistoryProvider.ts:24-49 evaluates history, decides whether to truncate, and merges a continuity summary back into active messages.
• agentHistoryProvider.ts:153-184 generates an “agent-continuity focused intent summary” that captures original mandate, agent strategy, and evolving context.
• defaultModelConfigs.ts:246-249 assigns a dedicated agent-history-provider-summarizer model config.
• memoryDiscovery.ts:149-170, 216-220 adds configurable boundaryMarkers for finding project roots and memory scope.

**Why it matters:**

Gemini is treating continuity as a managed subsystem with explicit compression and scoped memory boundaries, not as a passive side effect of a long context window.

2) OpenAI Codex — resume/rejoin depends on explicit thread-history reconstruction

**Commit:** b06f91c4fe52 — app-server: improve thread resume rejoin flow (#11776)

**Files:**

• codex-rs/app-server-protocol/src/protocol/thread_history.rs
• codex-rs/tui_app_server/src/lib.rs (related fix in commit 8e24d5aaea1c)

**Key evidence:**

• thread_history.rs:54-64 rebuilds turns from persisted rollout items so resumed history preserves original turn identifiers.
• thread_history.rs:107-160 uses a shared reducer for persisted rollout replay and in-memory current-turn tracking during resume/rejoin.
• tui_app_server/src/lib.rs:1897-1951 regression test shows resume-by-name now finds saved sessions even when the rollout title and stored thread name differ.

**Why it matters:**

Codex is moving continuity from “best effort restoration” into protocol-aware replay logic. Session identity and history reconstruction are becoming product-critical.

3) OpenClaw — restart recovery now actively restitches interrupted agent work

**Commit:** c780b6a6ab2a — fix: ... implement resume context and config idempotency guard

**Files:**

• src/agents/subagent-orphan-recovery.ts
• src/agents/cli-runner.ts (related cleanup commit 8edf2146ae59)

**Key evidence:**

• subagent-orphan-recovery.ts:32-49 builds a synthetic resume message that includes the original task and last user message.
• subagent-orphan-recovery.ts:115-125, 193-229 scans for orphaned sessions after restart, retries resume, and preserves abortedLastRun when recovery fails so the next restart can retry.
• subagent-orphan-recovery.ts:201-203 adds a config-change hint to avoid re-applying changes after interruption.
• cli-runner.ts:39-64 proactively kills stale resume processes that could otherwise leave duplicate resumptions hanging around.

**Why it matters:**

OpenClaw is making interruption recovery an active runtime behavior. The system now works to restore a job’s narrative thread after infrastructure churn.

Web/context signals

Gemini CLI repo page

• Explicitly advertises **conversation checkpointing to save and resume complex sessions**.
• Reinforces that continuity is now user-facing product value, not hidden plumbing.

OpenAI Codex launch page

• Emphasizes independent tasks, real-time progress monitoring, and verifiable evidence through logs and test outputs.
• Supports the framing that long-running agent work needs durable, inspectable state rather than one-shot completions.

Working thesis

The competitive frontier is shifting from raw context-window size to continuity engineering. Agent teams are building systems that compress, replay, scope, recover, and re-identify work so a task can survive interruptions without losing its plot.

Repo list used in article

• google-gemini/gemini-cli
• openai/codex
• openclaw/openclaw

Model review note

• llm is available in this environment.
• Use llm -m gpt-5.4 for review/synthesis. If it fails, fall back to the best available GPT-5.x model and record that in the trail log.

No standalone sources file is available for this article. The article body remains the primary evidence-bearing artifact.