hermes - 💡(How to fix) Fix feat: dedupe loaded skills via active skill context registry

Problem

Hermes currently treats skill_view() output like any other tool result: the full SKILL.md content is appended into the conversation history. This is simple and auditable, but it has poor context behavior in long sessions:

• Re-loading the same skill can duplicate a large block of text in the model context.
• Large skills (for example hermes-agent) become noisy persistent tool outputs even after they have served their purpose.
• The transcript and the prompt context are coupled: preserving a faithful audit trail means also preserving bulky skill bodies in the active model input.
• Mandatory skill-loading policies can repeatedly insert the same static content even when the skill has not changed.
• Client-side deletion/replacement of old tool outputs risks damaging prompt-cache locality if it mutates the message prefix.

This is related to broader context-management work, but skills are a special case: they are named, versionable resources that can be addressed by path/hash and are often semantically closer to "attached context" than to ordinary transient tool output.

Proposed solution

Introduce a first-class skill context registry / active skill block mechanism that decouples skill content from ordinary transcript messages.

High-level model:

Transcript / session DB:
  user/assistant/tool events remain honest and auditable
  skill_view event can be stored as compact metadata:
    loaded skill hermes-agent@sha256:abc123

Prompt builder:
  maintains an active skill context registry:
    hermes-agent@sha256:abc123 -> rendered SKILL.md content
    github-issues@sha256:def456 -> rendered SKILL.md content

Model input:
  system prompt
  memory/profile
  active skill blocks, deduped by skill name + hash
  recent conversation
  compacted/summarized old tool outputs

Behavior:

1. When skill_view(name) is called, compute a stable hash/version of the rendered skill content, including linked file content when requested.
2. If name@hash is already active in the current session, do not append the full skill body again. Return/record a compact event such as:
   • Skill hermes-agent@abc123 already loaded; active context unchanged.
3. If the skill content changed, replace the active block for that skill with the new name@hash content.
4. The prompt builder injects active skill blocks once, in a dedicated section, rather than relying on historical tool outputs to keep them visible.
5. The session transcript can still store full tool outputs for audit/debug if desired, but prompt assembly should be able to collapse old skill-view outputs to metadata/placeholders.

Why this is useful

• Avoids repeated giant skill bodies in the active context.
• Preserves a faithful transcript without forcing every historical skill body into the model input.
• Makes skill freshness explicit through hashes/versions.
• Gives prompt assembly a natural place to budget skill context separately from chat history and generic tool outputs.
• Aligns with the direction of on-demand context, memory, and context editing used in modern coding agents.
• Helps mandatory skill policies stay safe without unnecessary token bloat.

Possible implementation shape

Data model

Add a session-scoped structure, for example:

active_context = {
    "skills": {
        "hermes-agent": {
            "hash": "sha256:abc123",
            "path": "/.../SKILL.md",
            "content": "...rendered skill content...",
            "loaded_at_message_id": 123,
            "source": "skill_view",
        }
    }
}

This could live:

• in memory during a run;
• persisted in session metadata for resume;
• or reconstructed from compact skill-load events in the transcript.

Prompt assembly

Prompt builder renders a bounded block, e.g.:

<active_skills>
## hermes-agent @ sha256:abc123
...

## github-issues @ sha256:def456
...
</active_skills>

Budget policy could be added later:

• max total active skill tokens;
• max per skill;
• evict least recently used skills;
• keep mandatory skills pinned;
• show warnings when a skill is too large.

Tool response behavior

skill_view() can return compact metadata when a skill is already active:

{
  "success": true,
  "name": "hermes-agent",
  "hash": "sha256:abc123",
  "already_active": true,
  "message": "Skill already active in prompt context; full body not reinserted."
}

When content changed:

{
  "success": true,
  "name": "hermes-agent",
  "hash": "sha256:newhash",
  "updated_active_context": true,
  "content": "..."
}

Open questions

• Should full skill content still be stored in the session DB for audit, but excluded from model input at prompt-build time?
• Should active skill context survive /resume? If yes, should it be restored by hash from disk rather than from old tool output?
• How should linked skill files be hashed and represented?
• Should skills be evictable under context pressure, or should used skills remain pinned until /clear/new session?
• How should this interact with prompt caching? A dedicated block near the system prompt may improve dedupe but can also move cache boundaries if it changes frequently.
• Should this be implemented as part of a general attached-context/resource system rather than skill-specific machinery?

Related issues

• #10164 — context-aware skills prompt and system prompt budget management
• #2045 — Lazy skill loading: remove skill listing from system prompt, use on-demand tool instead
• #526 — Anthropic Context Editing API Integration — Server-Side, Cache-Friendly Tool/Thinking Cleanup for Claude Models
• #513 — Two-Phase Context Management — Prune Tool Outputs Before Full Compaction
• #415 — Insertion-Time Tool Result Trimming — Cache-Friendly Context Management
• #15618 — Expose real prompt context usage and compaction metadata in API run events
• #14603 — Context compaction SUMMARY_PREFIX causes cross-session task leakage
• #17344 — Context compression + session resume causes model to re-execute the original first task instead of continuing from compressed state
• #26730 — Kanban workers fail due to context pollution from parent session

Prior art / references

• Claude Code skills docs: https://docs.anthropic.com/en/docs/claude-code/skills
  • Skill bodies are on-demand context rather than always-on prompt text.
• Claude Code extension overview: https://code.claude.com/docs/en/features-overview
  • Separates always-on CLAUDE.md, on-demand skills, subagents for isolation, hooks, and MCP.
• Anthropic context management announcement: https://www.anthropic.com/news/context-management
  • Context editing removes stale tool calls/results while preserving conversation flow.
  • Memory tool stores information outside the active context window.
• Anthropic effective context engineering: https://www.anthropic.com/engineering/effective-context-engineering-for-ai-agents
  • Recommends the smallest high-signal context set and managing system instructions, tools, memory, external data, and history as one context state.

Acceptance criteria

A minimal version would be successful if:

• Loading the same skill twice in one session does not duplicate the full skill body in model input.
• Prompt assembly includes only the latest active version of each loaded skill.
• Skill freshness is tracked with a content hash or equivalent version marker.
• Session transcript remains auditable, but model input can collapse old skill-view outputs to metadata/placeholders.
• Existing skill behavior remains backward-compatible when the feature is disabled.