[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"skill:01M6DZAN9JRPQTHP3YFRQCWEF1":3,"skill-same:01M6DZAN9JRPQTHP3YFRQCWEF1":75},{"skillNo":4,"name":5,"nameZh":6,"description":7,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":15,"licenseFileUrl":6,"sourceHtmlUrl":16,"categories":17,"keywords":23,"publishedAt":39,"linkOnly":40,"markdownContent":41,"markdownBody":42,"frontMatter":43,"refName":55,"commitSha":56,"skillDirectory":57,"forksCount":14,"repositoryHtmlUrl":58,"repositoryHomepageUrl":6,"ownerHtmlUrl":59,"ownerBlogUrl":6,"ownerDescription":6,"updatedAt":60,"install":61,"resources":74},"01M6DZAN9JRPQTHP3YFRQCWEF1","hermes-insights",null,"Generate insights about your Claude Code usage — what topics you work on most, common patterns, productivity trends.","pshpjr\u002Fcoding-agent-test","pshpjr","https:\u002F\u002Favatars.githubusercontent.com\u002Fu\u002F30226755?v=4","01M6DZ9Q9KB110G8J47FD3ZH3B","SeongHawnPark","01M6DZ9QDVFESFAMG3FGJGJQAZ",0,"MIT","https:\u002F\u002Fgithub.com\u002Fpshpjr\u002Fcoding-agent-test\u002Fblob\u002Fmaster\u002F.claude\u002Fskills\u002Fhermes-insights\u002FSKILL.md",[18],{"categoryNo":19,"categoryName":20,"taxonomyType":21,"taxonomyTypeDesc":22},"01JBSEEDCAT000000000000023","Skill与智能体开发","SCENARIO","使用场景",[24,25,26,27,28,29,30,31,32,33,34,35,36,37,38],"hermes","hermes insights","analytics","generate","code","search","productivity","work on","hermes agent","common","work","honcho","claude code usage","patterns","ccc","2026-11-10 10:12:30",false,"---\nname: hermes-insights\ndescription: Generate insights about your Claude Code usage — what topics you work on most, common patterns, productivity trends.\nversion: 1.0.0\nauthor: hermes-CCC (ported from Hermes Agent by NousResearch)\nlicense: MIT\nmetadata:\n  hermes:\n    tags: [Analytics, Insights, Usage, Productivity]\n    related_skills: [hermes-search, honcho]\n---\n\n# hermes-insights\n\nAnalyze your Claude Code usage patterns across sessions and produce a structured insight report: what topics you work on most, which skills you use, what types of projects dominate your workflow, and where your productivity is strongest.\n\n## Invocation\n\n```\n\u002Fhermes-insights\n\u002Fhermes-insights --days \u003CN>\n```\n\n---\n\n## Subcommands\n\n### `\u002Fhermes-insights` — Full Analysis\n\nAnalyzes all available memory files and session history with no time constraint.\n\n### `\u002Fhermes-insights --days \u003CN>` — Time-Bounded Analysis\n\nRestricts analysis to session files created in the last N days.\n\n**How Claude filters by date:**\n\n```bash\nfind ~\u002F.claude\u002Fprojects\u002F*\u002Fmemory\u002Fsession_*.md \\\n     -mtime -\u003CN> -type f\n```\n\n---\n\n## Data Sources and What Claude Analyzes\n\n### Source Priority\n\nThis project collects precise telemetry via Claude Code hooks. When JSONL log files exist in `_workspace\u002F`, they are the **primary data source** — they contain exact API-reported numbers, not estimates. Memory files are the fallback for projects without instrumentation.\n\n```bash\n# Check which sources are available\nls _workspace\u002Ftoken-log.jsonl _workspace\u002Fskill-usage-log.jsonl _workspace\u002Fcommand-log.jsonl 2>\u002Fdev\u002Fnull\n```\n\n### 1. Token Usage — `_workspace\u002Ftoken-log.jsonl` (primary)\n\nEach line is a JSON object with exact API usage per session checkpoint:\n\n| Field | Meaning |\n|-------|---------|\n| `input_tokens` | API-reported input tokens |\n| `output_tokens` | API-reported output tokens |\n| `cache_creation_input_tokens` | KV cache creation cost |\n| `cache_read_input_tokens` | KV cache hits (discounted rate) |\n| `total_billed_tokens` | `input_tokens + output_tokens` |\n| `turns` | Number of assistant turns in session |\n| `models` | Models used in session |\n| `session_id` | Session identifier (deduplicate by taking max-turns entry per session) |\n\n**Analysis to produce:**\n- Daily\u002Fweekly billed token totals and trends\n- Average tokens per session and per turn\n- Cache hit ratio: `cache_read \u002F (cache_read + cache_creation + input)` — higher is better\n- Model distribution across sessions\n- Cost efficiency: `total_billed_tokens \u002F changed_lines` (cross-reference with `git log --numstat`)\n- Use `python tools\u002Fshow_token_usage.py --days N` for quick pre-formatted reports\n\n**Fallback:** If this file does not exist, skip token analysis and note \"no token instrumentation found\".\n\n### 2. Skill & Agent Utilization — `_workspace\u002Fskill-usage-log.jsonl` (primary)\n\nEach line records an exact Skill or Agent tool invocation:\n\n| Field | Meaning |\n|-------|---------|\n| `tool` | `\"Skill\"` or `\"Agent\"` |\n| `name` | Skill name or subagent_type |\n| `args` \u002F `description` | Invocation context (truncated to 120 chars) |\n\n**Analysis to produce:**\n- Exact invocation counts per skill\u002Fagent, ranked by frequency\n- Last-used date per skill\n- Skills with zero invocations (trigger problem candidates)\n- Agent type distribution (which specialist agents are most used)\n- Skill usage trend over time (daily\u002Fweekly)\n\n**Fallback:** If this file does not exist, infer skill usage from memory files (mark as `inferred`).\n\n### 3. Command Patterns — `_workspace\u002Fcommand-log.jsonl` (primary)\n\nEach line records a Bash tool call:\n\n| Field | Meaning |\n|-------|---------|\n| `command` | Executed command (max 300 chars) |\n| `purpose` | Agent-written description of intent |\n| `script` | Extracted script path (`.py`, `.sh`, etc.) |\n\n**Analysis to produce:**\n- Most-used scripts\u002Ftools ranked by frequency\n- Purpose categories (build, test, deploy, query, etc.)\n- Commands without purpose descriptions (quality signal for agent prompting)\n- **Repeated command sequences** — same command appearing 3+ times in a session indicates a retry loop or missing automation\n- **Failed commands** — if `exit_code` is non-zero, count failure rate per script\u002Fcommand\n\n**Fallback:** If this file does not exist, skip command analysis.\n\n### 4. Waste & Repetition Detection (cross-source)\n\nThis is the core analysis for the recursive improvement loop. Cross-reference all three JSONL logs to find token waste candidates:\n\n| Signal | How to detect | Improvement action |\n|--------|--------------|-------------------|\n| **Retry loops** | Same command ≥3 times in one session in command-log | Diagnose root cause; add error handling to skill or create a script |\n| **High-turn sessions** | Sessions with turns ≫ median in token-log | Check if scope was too broad or agent got stuck; tighten exec-plan scoping |\n| **Repeated skill+fail** | Same skill invoked multiple times with similar args in skill-usage-log | Fix skill instructions or add precondition checks |\n| **Low cache ratio sessions** | `cache_read \u002F (cache_read + cache_creation + input) \u003C 0.8` | Investigate context instability — frequent file switches or large tool outputs |\n| **Outlier sessions** | Billed tokens > 2× median | Identify what drove the cost — large file reads, excessive tool calls, or retry loops |\n| **Manual work patterns** | Same purpose description appearing across multiple sessions | Candidate for new skill or script automation |\n\n**Output:** A \"Waste Candidates\" table listing the top 3–5 patterns with estimated token savings if automated\u002Ffixed, linked to harness-improvement queue actions.\n\n### 5. Memory Files — Topic and Theme Extraction (fallback)\n\n```bash\nls ~\u002F.claude\u002Fprojects\u002F*\u002Fmemory\u002Fsession_*.md\n```\n\nFor each session file, Claude reads the `decisions`, `facts_learned`, `artifacts_created`, and `open_issues` sections and extracts noun phrases as topic signals.\n\n**Output:** Top 10 topics, each with session count and a representative example decision or fact.\n\n### 5. Data Hygiene and Confidence Rules\n\n- **Deduplication:** For token-log, keep only the entry with the highest `turns` per `session_id` (later checkpoints are cumulative).\n- Normalize topic strings by lowercasing, trimming punctuation, and folding obvious singular\u002Fplural variants before counting.\n- Prefer JSONL data over memory-file inference whenever both are available.\n- Downgrade trend claims to `low confidence` when fewer than 5 sessions match the selected date range.\n- Report `no data matched the filter` instead of fabricating empty charts when `--days \u003CN>` returns zero entries.\n- Separate unresolved carry-over work from newly opened issues so the resolution rate is not overstated.\n\n---\n\n## Full Output Structure\n\n```\n## Claude Code Usage Insights\nGenerated: 2026-04-08 | Sessions analyzed: 12 | Date range: 2026-04-05 – 2026-04-08\n\n### Token Usage (from _workspace\u002Ftoken-log.jsonl)\n| Date       | Sessions | Turns | Billed     | Cache Read | Cache Ratio |\n|------------|----------|-------|------------|------------|-------------|\n| 2026-04-08 |        3 |    45 |     52.1k  |     8.2M   |       94%   |\n| 2026-04-07 |        5 |   120 |    118.3k  |    22.1M   |       91%   |\n| 2026-04-05 |        4 |   219 |    109.6k  |    13.1M   |       89%   |\n\nTotals: 280.0k billed | 43.4M cache read | avg 23.3k\u002Fsession | avg 1.2k\u002Fturn\n\n### Skill & Agent Utilization (from _workspace\u002Fskill-usage-log.jsonl)\n| Name              | Type   | Invocations | Last Used  |\n|-------------------|--------|-------------|------------|\n| opencode-review   | Skill  |           2 | 2026-04-08 |\n| codex-delegate    | Skill  |           1 | 2026-04-08 |\n| opencode-reviewer | Agent  |           2 | 2026-04-08 |\n\nZero-invocation skills (potential trigger issues): hermes-persona, hermes-traj\n\n### Command Patterns (from _workspace\u002Fcommand-log.jsonl)\n| Script\u002FCommand         | Count | Top Purpose                    |\n|------------------------|-------|--------------------------------|\n| tools\u002Fhi_done.py       |     2 | Mark queue items as done       |\n| git status             |     1 | Show working tree status       |\n| git log                |     1 | Show recent commits            |\n\n### Efficiency Metrics\n- Billed tokens per changed line: 142 tokens\u002Fline (lower is better)\n- Cache hit ratio trend: 89% → 91% → 94% (improving)\n- Median session tokens: 18.2k | Outlier threshold (2×): 36.4k\n\n### Waste Candidates (recursive improvement targets)\n| Pattern | Evidence | Est. Savings | Action |\n|---------|----------|-------------|--------|\n| `git status` retry loop | 4× in session abc123 | ~2k tokens\u002Fsession | Add pre-check to commit skill |\n| codex process timeout | `taskkill \u002FF \u002FIM codex.exe` in 3 sessions | ~5k tokens\u002Fsession | Add timeout guard to codex-delegate |\n| Outlier session 7e58b64a | 47k billed (2.5× median) | — | Scope was too broad; split exec-plan |\n\n→ Register top candidates in `docs\u002Fworkflow\u002Fharness-improvement-queue.md`\n\n### Recommendations\n- **Cache efficiency improving** — recent sessions show 94% cache hit ratio, keep using stable context patterns.\n- **Zero-invocation skills** — hermes-persona, hermes-traj have never triggered. Review trigger descriptions or remove.\n- **Retry loops are the #1 waste source** — automate or add error handling to eliminate re-runs.\n\n### Key Takeaway\n12 sessions consumed 280k billed tokens with 92% average cache hit ratio. Top waste candidate: retry loops (~2–5k tokens\u002Fsession recoverable).\n```\n\n---\n\n## How Claude Gathers the Data\n\n1. Check for instrumented JSONL logs in `_workspace\u002F`:\n   - `token-log.jsonl` → token usage analysis (deduplicate by max turns per session_id)\n   - `skill-usage-log.jsonl` → exact skill\u002Fagent invocation counts\n   - `command-log.jsonl` → command pattern analysis\n2. If `--days N` is specified, filter JSONL entries by `date` field.\n3. For token efficiency, cross-reference with `git log --numstat --since=\u003Cdate>` for changed-line counts.\n4. If JSONL logs are absent, fall back to memory files: `ls -t ~\u002F.claude\u002Fprojects\u002F*\u002Fmemory\u002Fsession_*.md`\n5. For quick token reports, run `python tools\u002Fshow_token_usage.py --days N` directly.\n6. Aggregate and format the insight report.\n\nClaude does not write any files during `\u002Fhermes-insights`. The command is read-only.\n\n---\n\n## Relationship to Hermes `\u002Finsights` Command\n\nThis skill mirrors the `\u002Finsights` command in Hermes Agent (NousResearch), which produces usage analytics from the SQLite session store. In hermes-CCC, the equivalent data source is the Markdown session memory files written by `hermes-compress`. The output categories (topics, skill utilization, productivity trends, recommendations) match the Hermes insights contract.\n\n---\n\n## Notes\n\n- When `_workspace\u002Ftoken-log.jsonl` exists, token analysis is precise (API-reported). No hermes-compress runs needed for quantitative insights.\n- Qualitative analysis (topics, decisions) still benefits from hermes-compress session files.\n- `\u002Fhermes-insights` works best after at least 5 sessions of data accumulation (from either JSONL or memory files).\n- Use insights to feed the recursive improvement loop: identify high-cost patterns → register improvement in harness-improvement queue.\n","# hermes-insights\n\nAnalyze your Claude Code usage patterns across sessions and produce a structured insight report: what topics you work on most, which skills you use, what types of projects dominate your workflow, and where your productivity is strongest.\n\n## Invocation\n\n```\n\u002Fhermes-insights\n\u002Fhermes-insights --days \u003CN>\n```\n\n---\n\n## Subcommands\n\n### `\u002Fhermes-insights` — Full Analysis\n\nAnalyzes all available memory files and session history with no time constraint.\n\n### `\u002Fhermes-insights --days \u003CN>` — Time-Bounded Analysis\n\nRestricts analysis to session files created in the last N days.\n\n**How Claude filters by date:**\n\n```bash\nfind ~\u002F.claude\u002Fprojects\u002F*\u002Fmemory\u002Fsession_*.md \\\n     -mtime -\u003CN> -type f\n```\n\n---\n\n## Data Sources and What Claude Analyzes\n\n### Source Priority\n\nThis project collects precise telemetry via Claude Code hooks. When JSONL log files exist in `_workspace\u002F`, they are the **primary data source** — they contain exact API-reported numbers, not estimates. Memory files are the fallback for projects without instrumentation.\n\n```bash\n# Check which sources are available\nls _workspace\u002Ftoken-log.jsonl _workspace\u002Fskill-usage-log.jsonl _workspace\u002Fcommand-log.jsonl 2>\u002Fdev\u002Fnull\n```\n\n### 1. Token Usage — `_workspace\u002Ftoken-log.jsonl` (primary)\n\nEach line is a JSON object with exact API usage per session checkpoint:\n\n| Field | Meaning |\n|-------|---------|\n| `input_tokens` | API-reported input tokens |\n| `output_tokens` | API-reported output tokens |\n| `cache_creation_input_tokens` | KV cache creation cost |\n| `cache_read_input_tokens` | KV cache hits (discounted rate) |\n| `total_billed_tokens` | `input_tokens + output_tokens` |\n| `turns` | Number of assistant turns in session |\n| `models` | Models used in session |\n| `session_id` | Session identifier (deduplicate by taking max-turns entry per session) |\n\n**Analysis to produce:**\n- Daily\u002Fweekly billed token totals and trends\n- Average tokens per session and per turn\n- Cache hit ratio: `cache_read \u002F (cache_read + cache_creation + input)` — higher is better\n- Model distribution across sessions\n- Cost efficiency: `total_billed_tokens \u002F changed_lines` (cross-reference with `git log --numstat`)\n- Use `python tools\u002Fshow_token_usage.py --days N` for quick pre-formatted reports\n\n**Fallback:** If this file does not exist, skip token analysis and note \"no token instrumentation found\".\n\n### 2. Skill & Agent Utilization — `_workspace\u002Fskill-usage-log.jsonl` (primary)\n\nEach line records an exact Skill or Agent tool invocation:\n\n| Field | Meaning |\n|-------|---------|\n| `tool` | `\"Skill\"` or `\"Agent\"` |\n| `name` | Skill name or subagent_type |\n| `args` \u002F `description` | Invocation context (truncated to 120 chars) |\n\n**Analysis to produce:**\n- Exact invocation counts per skill\u002Fagent, ranked by frequency\n- Last-used date per skill\n- Skills with zero invocations (trigger problem candidates)\n- Agent type distribution (which specialist agents are most used)\n- Skill usage trend over time (daily\u002Fweekly)\n\n**Fallback:** If this file does not exist, infer skill usage from memory files (mark as `inferred`).\n\n### 3. Command Patterns — `_workspace\u002Fcommand-log.jsonl` (primary)\n\nEach line records a Bash tool call:\n\n| Field | Meaning |\n|-------|---------|\n| `command` | Executed command (max 300 chars) |\n| `purpose` | Agent-written description of intent |\n| `script` | Extracted script path (`.py`, `.sh`, etc.) |\n\n**Analysis to produce:**\n- Most-used scripts\u002Ftools ranked by frequency\n- Purpose categories (build, test, deploy, query, etc.)\n- Commands without purpose descriptions (quality signal for agent prompting)\n- **Repeated command sequences** — same command appearing 3+ times in a session indicates a retry loop or missing automation\n- **Failed commands** — if `exit_code` is non-zero, count failure rate per script\u002Fcommand\n\n**Fallback:** If this file does not exist, skip command analysis.\n\n### 4. Waste & Repetition Detection (cross-source)\n\nThis is the core analysis for the recursive improvement loop. Cross-reference all three JSONL logs to find token waste candidates:\n\n| Signal | How to detect | Improvement action |\n|--------|--------------|-------------------|\n| **Retry loops** | Same command ≥3 times in one session in command-log | Diagnose root cause; add error handling to skill or create a script |\n| **High-turn sessions** | Sessions with turns ≫ median in token-log | Check if scope was too broad or agent got stuck; tighten exec-plan scoping |\n| **Repeated skill+fail** | Same skill invoked multiple times with similar args in skill-usage-log | Fix skill instructions or add precondition checks |\n| **Low cache ratio sessions** | `cache_read \u002F (cache_read + cache_creation + input) \u003C 0.8` | Investigate context instability — frequent file switches or large tool outputs |\n| **Outlier sessions** | Billed tokens > 2× median | Identify what drove the cost — large file reads, excessive tool calls, or retry loops |\n| **Manual work patterns** | Same purpose description appearing across multiple sessions | Candidate for new skill or script automation |\n\n**Output:** A \"Waste Candidates\" table listing the top 3–5 patterns with estimated token savings if automated\u002Ffixed, linked to harness-improvement queue actions.\n\n### 5. Memory Files — Topic and Theme Extraction (fallback)\n\n```bash\nls ~\u002F.claude\u002Fprojects\u002F*\u002Fmemory\u002Fsession_*.md\n```\n\nFor each session file, Claude reads the `decisions`, `facts_learned`, `artifacts_created`, and `open_issues` sections and extracts noun phrases as topic signals.\n\n**Output:** Top 10 topics, each with session count and a representative example decision or fact.\n\n### 5. Data Hygiene and Confidence Rules\n\n- **Deduplication:** For token-log, keep only the entry with the highest `turns` per `session_id` (later checkpoints are cumulative).\n- Normalize topic strings by lowercasing, trimming punctuation, and folding obvious singular\u002Fplural variants before counting.\n- Prefer JSONL data over memory-file inference whenever both are available.\n- Downgrade trend claims to `low confidence` when fewer than 5 sessions match the selected date range.\n- Report `no data matched the filter` instead of fabricating empty charts when `--days \u003CN>` returns zero entries.\n- Separate unresolved carry-over work from newly opened issues so the resolution rate is not overstated.\n\n---\n\n## Full Output Structure\n\n```\n## Claude Code Usage Insights\nGenerated: 2026-04-08 | Sessions analyzed: 12 | Date range: 2026-04-05 – 2026-04-08\n\n### Token Usage (from _workspace\u002Ftoken-log.jsonl)\n| Date       | Sessions | Turns | Billed     | Cache Read | Cache Ratio |\n|------------|----------|-------|------------|------------|-------------|\n| 2026-04-08 |        3 |    45 |     52.1k  |     8.2M   |       94%   |\n| 2026-04-07 |        5 |   120 |    118.3k  |    22.1M   |       91%   |\n| 2026-04-05 |        4 |   219 |    109.6k  |    13.1M   |       89%   |\n\nTotals: 280.0k billed | 43.4M cache read | avg 23.3k\u002Fsession | avg 1.2k\u002Fturn\n\n### Skill & Agent Utilization (from _workspace\u002Fskill-usage-log.jsonl)\n| Name              | Type   | Invocations | Last Used  |\n|-------------------|--------|-------------|------------|\n| opencode-review   | Skill  |           2 | 2026-04-08 |\n| codex-delegate    | Skill  |           1 | 2026-04-08 |\n| opencode-reviewer | Agent  |           2 | 2026-04-08 |\n\nZero-invocation skills (potential trigger issues): hermes-persona, hermes-traj\n\n### Command Patterns (from _workspace\u002Fcommand-log.jsonl)\n| Script\u002FCommand         | Count | Top Purpose                    |\n|------------------------|-------|--------------------------------|\n| tools\u002Fhi_done.py       |     2 | Mark queue items as done       |\n| git status             |     1 | Show working tree status       |\n| git log                |     1 | Show recent commits            |\n\n### Efficiency Metrics\n- Billed tokens per changed line: 142 tokens\u002Fline (lower is better)\n- Cache hit ratio trend: 89% → 91% → 94% (improving)\n- Median session tokens: 18.2k | Outlier threshold (2×): 36.4k\n\n### Waste Candidates (recursive improvement targets)\n| Pattern | Evidence | Est. Savings | Action |\n|---------|----------|-------------|--------|\n| `git status` retry loop | 4× in session abc123 | ~2k tokens\u002Fsession | Add pre-check to commit skill |\n| codex process timeout | `taskkill \u002FF \u002FIM codex.exe` in 3 sessions | ~5k tokens\u002Fsession | Add timeout guard to codex-delegate |\n| Outlier session 7e58b64a | 47k billed (2.5× median) | — | Scope was too broad; split exec-plan |\n\n→ Register top candidates in `docs\u002Fworkflow\u002Fharness-improvement-queue.md`\n\n### Recommendations\n- **Cache efficiency improving** — recent sessions show 94% cache hit ratio, keep using stable context patterns.\n- **Zero-invocation skills** — hermes-persona, hermes-traj have never triggered. Review trigger descriptions or remove.\n- **Retry loops are the #1 waste source** — automate or add error handling to eliminate re-runs.\n\n### Key Takeaway\n12 sessions consumed 280k billed tokens with 92% average cache hit ratio. Top waste candidate: retry loops (~2–5k tokens\u002Fsession recoverable).\n```\n\n---\n\n## How Claude Gathers the Data\n\n1. Check for instrumented JSONL logs in `_workspace\u002F`:\n   - `token-log.jsonl` → token usage analysis (deduplicate by max turns per session_id)\n   - `skill-usage-log.jsonl` → exact skill\u002Fagent invocation counts\n   - `command-log.jsonl` → command pattern analysis\n2. If `--days N` is specified, filter JSONL entries by `date` field.\n3. For token efficiency, cross-reference with `git log --numstat --since=\u003Cdate>` for changed-line counts.\n4. If JSONL logs are absent, fall back to memory files: `ls -t ~\u002F.claude\u002Fprojects\u002F*\u002Fmemory\u002Fsession_*.md`\n5. For quick token reports, run `python tools\u002Fshow_token_usage.py --days N` directly.\n6. Aggregate and format the insight report.\n\nClaude does not write any files during `\u002Fhermes-insights`. The command is read-only.\n\n---\n\n## Relationship to Hermes `\u002Finsights` Command\n\nThis skill mirrors the `\u002Finsights` command in Hermes Agent (NousResearch), which produces usage analytics from the SQLite session store. In hermes-CCC, the equivalent data source is the Markdown session memory files written by `hermes-compress`. The output categories (topics, skill utilization, productivity trends, recommendations) match the Hermes insights contract.\n\n---\n\n## Notes\n\n- When `_workspace\u002Ftoken-log.jsonl` exists, token analysis is precise (API-reported). No hermes-compress runs needed for quantitative insights.\n- Qualitative analysis (topics, decisions) still benefits from hermes-compress session files.\n- `\u002Fhermes-insights` works best after at least 5 sessions of data accumulation (from either JSONL or memory files).\n- Use insights to feed the recursive improvement loop: identify high-cost patterns → register improvement in harness-improvement queue.\n",{"name":5,"author":44,"license":15,"version":45,"metadata":46,"description":7},"hermes-CCC (ported from Hermes Agent by NousResearch)","1.0.0",{"hermes":47},{"tags":48,"related_skills":53},[49,50,51,52],"Analytics","Insights","Usage","Productivity",[54,35],"hermes-search","master","f3c953d73d9891c823b3b0a94a07d95ff71f9a01",".claude\u002Fskills\u002Fhermes-insights","https:\u002F\u002Fgithub.com\u002Fpshpjr\u002Fcoding-agent-test","https:\u002F\u002Fgithub.com\u002Fpshpjr","2026-04-17 17:14:56",{"source":8,"skillName":5,"commands":62,"prompt":72,"zipAvailable":73},[63,66,69],{"tool":64,"command":65},"npx","npx skills add pshpjr\u002Fcoding-agent-test --skill hermes-insights",{"tool":67,"command":68},"bunx","bunx skills add pshpjr\u002Fcoding-agent-test --skill hermes-insights",{"tool":70,"command":71},"pnpm","pnpm dlx skills add pshpjr\u002Fcoding-agent-test --skill hermes-insights","安装技能，技能地址：https:\u002F\u002Fgithub.com\u002Fpshpjr\u002Fcoding-agent-test\u002Fblob\u002Fmaster\u002F.claude\u002Fskills\u002Fhermes-insights\u002FSKILL.md",true,[],[76,101,117,135,150],{"skillNo":77,"name":78,"nameZh":6,"description":79,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":15,"licenseFileUrl":6,"sourceHtmlUrl":80,"categories":81,"keywords":82,"publishedAt":39},"01M6DZB1C493J2N4BA1HYY3SK7","hermes-traj","Capture Claude Code interaction trajectories in training-friendly formats. Use when saving successful runs, failed runs, review samples, evaluation traces, or compact datasets for later QA and fine-tuning analysis.","https:\u002F\u002Fgithub.com\u002Fpshpjr\u002Fcoding-agent-test\u002Fblob\u002Fmaster\u002F.claude\u002Fskills\u002Fhermes-traj\u002FSKILL.md",[],[24,83,84,85,86,87,88,89,90,91,92,93,94,95,96,28,97,98,99,100],"hermes traj","analysis","interaction","dataset","openai codex","compact","later","datasets","use","filesystem","json","Fine-Tuning","beta","qa","training","telemetry","codex","review",{"skillNo":102,"name":54,"nameZh":6,"description":103,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":15,"licenseFileUrl":6,"sourceHtmlUrl":104,"categories":105,"keywords":106,"publishedAt":39},"01M6DZAY7AWKYDH7WXX36348E5","Search across past conversation sessions, memory files, and project history to recall what was discussed, decided, or built in previous sessions.","https:\u002F\u002Fgithub.com\u002Fpshpjr\u002Fcoding-agent-test\u002Fblob\u002Fmaster\u002F.claude\u002Fskills\u002Fhermes-search\u002FSKILL.md",[],[107,24,29,108,109,110,111,32,112,113,114,115,116,38],"hermes search","recall","compress","sessions","history","files","conversation","across","memory","project",{"skillNo":118,"name":119,"nameZh":6,"description":120,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":15,"licenseFileUrl":6,"sourceHtmlUrl":121,"categories":122,"keywords":123,"publishedAt":39},"01M6DZAVD72X1W04HHYCNZJX5Z","hermes-persona","Switch Claude's personality\u002Fpersona between specialized modes — researcher, coder, analyst, creative, advisor.","https:\u002F\u002Fgithub.com\u002Fpshpjr\u002Fcoding-agent-test\u002Fblob\u002Fmaster\u002F.claude\u002Fskills\u002Fhermes-persona\u002FSKILL.md",[],[124,24,125,38,126,127,128,129,35,130,131,132,32,133,134],"persona","hermes persona","switch","analyst","personality","specialized","creative","customization","coder","researcher","advisor",{"skillNo":136,"name":137,"nameZh":6,"description":138,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":15,"licenseFileUrl":6,"sourceHtmlUrl":139,"categories":140,"keywords":141,"publishedAt":39},"01M6DZARHABNCNCC6P8PBMQENG","hermes-memory","Manage durable project memory for Claude Code. Use when prefetching relevant history before work, syncing stable decisions after work, nudging for unsaved lessons, compressing noisy notes, or auditing memory quality.","https:\u002F\u002Fgithub.com\u002Fpshpjr\u002Fcoding-agent-test\u002Fblob\u002Fmaster\u002F.claude\u002Fskills\u002Fhermes-memory\u002FSKILL.md",[],[24,115,142,95,143,28,144,99,145,146,87,116,91,147,34,92,32,111,148,149],"hermes memory","openai","notes","shell","manage","decisions","quality","markdown",{"skillNo":151,"name":152,"nameZh":6,"description":153,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":15,"licenseFileUrl":6,"sourceHtmlUrl":154,"categories":155,"keywords":156,"publishedAt":39},"01M6DZAHPNV1YJKYXRJ4F17CMD","hermes-compress","Compress conversation context — summarize the current session, extract key decisions and facts, then compact history to free up context window.","https:\u002F\u002Fgithub.com\u002Fpshpjr\u002Fcoding-agent-test\u002Fblob\u002Fmaster\u002F.claude\u002Fskills\u002Fhermes-compress\u002FSKILL.md",[],[157,109,24,158,38,159,88,113,160,161,162,147,32,111,163,29,115,164],"hermes compress","up","free","session","extract","context","compression","summarize"]