Releases·Tool Release
Dev builds deterministic state plugin to fix long-horizon task drift in local models after context compaction
A developer shared an open-source plugin that counters goal drift in local agent setups after context-window compaction. It hooks into the agent loop and maintains a structured state file outside the chat buffer, re-anchoring the model to deterministic milestones after each compaction. The tool targets a common failure mode — hallucinated task completion and loops on long tasks — and is available on GitHub for adaptation.
11 engagement1 sourceSat, Oct 3, 2026, 11:17 AM
AI Overview
- Plugin hooks into the agent loop and maintains structured state outside the chat buffer, tracking active milestone, current micro-step, next 3 queued steps, and a verification log
- After compaction, the plugin re-anchors the model to an exact state file instead of the lossy compaction summary, preventing goal drift, hallucinated completion, and loops
- Code for the plugin is open-sourced at github.com/janpauldahlke/dsh-local-long-horizon, with feedback and PRs invited
Related
Releases·Tool Release
Solving Agentic Amnesia: Dev builds Craft, an open-source Claude Code plugin using the file system as a state machine for context
A developer built Craft, an open-source Claude Code Plugin addressing "agentic amnesia" — the context-window overflow that makes agents hallucinate or forget earlier architectural decisions. Craft shifts context management from prompts to the file system, treating prompt rules as guidance the model can drop under context pressure.
A developer built Craft, an open-source Claude Code Plugin addressing "agentic amnesia" — the context-window overflow that makes agents hallucinate or forget earlier architectural decisions. Craft shifts context management from prompts to the file system, treating prompt rules as guidance the model can drop under context pressure.
11 engagement1 sourceredditMon, Sep 7, 2026, 05:28 PM
Social·Community
Context Drift Auditor skill audits silent context drift in long agent runs
A Reddit user shared a custom Claude skill, context-drift-auditor, that audits whether an agent's working context has silently degraded across turns, summarizations, tool results, and prompt version changes. The skill builds a cumulative 'active contract' baseline, extracts a context ledger with provenance, classifies drift per item, and gates the next action with a cited decision. It targets the failure mode where summaries drop negative constraints or stale assumptions harden into decisions during long agent tasks.
A Reddit user shared a custom Claude skill, context-drift-auditor, that audits whether an agent's working context has silently degraded across turns, summarizations, tool results, and prompt version changes. The skill builds a cumulative 'active contract' baseline, extracts a context ledger with provenance, classifies drift per item, and gates the next action with a cited decision. It targets the failure mode where summaries drop negative constraints or stale assumptions harden into decisions during long agent tasks.
3 engagement1 sourceredditTue, Sep 22, 2026, 09:27 AM
Social·Community
Builder details goal-drift failure mode in multi-step AI agents and mitigation approach in Richoo runtime
A developer building Richoo, described as an AI Goal Runtime, reports that multi-step agents drift after 10-15 sequential tool calls because raw execution logs flood the model's context window and dilute attention. The post frames the problem as context pollution rather than model incapability and previews an architectural mitigation based on a bounded execution envelope. It offers practitioners a concrete failure-mode diagnosis for long-horizon agent reliability.
A developer building Richoo, described as an AI Goal Runtime, reports that multi-step agents drift after 10-15 sequential tool calls because raw execution logs flood the model's context window and dilute attention. The post frames the problem as context pollution rather than model incapability and previews an architectural mitigation based on a bounded execution envelope. It offers practitioners a concrete failure-mode diagnosis for long-horizon agent reliability.
2 engagement1 sourceredditTue, Sep 1, 2026, 02:21 PM
Social·Community
Dev's system-prompt proxy keeps long-horizon agents sane through 62 compactions on a 128k budget
A Reddit developer describes a proxy technique for long-horizon agents: agents edit memory files that the proxy appends to their system prompt each turn, giving a cache-stable source of truth. One agent has compacted 62 times on a 128k budget (~8M tokens processed) and stayed coherent, evolving from coder to project manager. The author reports this nearly eliminates post-compaction hallucination, though the agent still forgets things when it fails to take notes.
A Reddit developer describes a proxy technique for long-horizon agents: agents edit memory files that the proxy appends to their system prompt each turn, giving a cache-stable source of truth. One agent has compacted 62 times on a 128k budget (~8M tokens processed) and stayed coherent, evolving from coder to project manager. The author reports this nearly eliminates post-compaction hallucination, though the agent still forgets things when it fails to take notes.
43 engagement1 sourceredditTue, Oct 13, 2026, 01:19 PM
Social·Community
Engineer tests typed reasoning graph to stop agent context amnesia in 5-hour session
An engineer experimented with forcing a frontier model to operate through a typed reasoning graph instead of prompt compression to prevent context amnesia in autonomous agents. The agent ran a continuous 5-hour discovery session on a complex repository, exhausting raw context window limits. The approach aims to maintain context and reduce hallucinated progress on large codebases.
An engineer experimented with forcing a frontier model to operate through a typed reasoning graph instead of prompt compression to prevent context amnesia in autonomous agents. The agent ran a continuous 5-hour discovery session on a complex repository, exhausting raw context window limits. The approach aims to maintain context and reduce hallucinated progress on large codebases.
10 engagement1 sourceredditFri, Aug 21, 2026, 01:42 PM