perchai-cli 2.4.21 → 2.4.23

package/dist/perch.mjs

@@ -85560,7 +85560,6 @@ Final answers lead with findings, name artifacts or delivery status, and give on
 var MARKET_DESK_TOOL_NAMES;
 var init_marketDeskAccess = __esm({
   "features/perchTerminal/runtime/marketDesk/marketDeskAccess.ts"() {
-    "use strict";
     init_toolNames();
     MARKET_DESK_TOOL_NAMES = /* @__PURE__ */ new Set([
       TOOL_NAMES.getMarketSignal,
@@ -86717,7 +86716,7 @@ var init_managedWorkflowRegistry = __esm({
               "name": "AP Suite Publisher",
               "description": "Publishes the final AP operator suite envelope from analyst output. Writes findings CSV and report to the output folder.",
               "lane": "writer",
-              "systemPrompt": '# AP Suite Publisher\n\nYou receive the analyst\'s structured output from the deterministic AP extraction. Your job is to write deliverable files to disk and return their paths \u2014 not to summarize in prose.\n\n## What to write\n\nUsing `writeLocalFile`, write two files to `{folderPath}/perch-output/`:\n\n**1. `findings.csv`** \u2014 one row per finding from `priorOutput.sandboxArtifact.findings` + analyst `thematicBlocks`:\n```\nid,severity,category,title,financial_exposure,evidence\nfinding-1,critical,duplicate_invoice,"INV-88421 matches INV-77209 same vendor same amount",12400,invoices/INV-88421.csv\n```\n\n**2. `report.md`** \u2014 a structured markdown report:\n```markdown\n# AP Audit Report \u2014 {folderPath leaf}\nGenerated: {ISO date}\n\n## Executive Summary\n{analyst executiveSummary \u2014 2\u20134 sentences, verbatim from analyst output}\n\n## Risk Assessment\n**Level:** {level} \u2014 {justification}\n\n## Findings ({N} total)\n{For each thematic block: ### {title}, severity badge, narrative, finding IDs}\n\n## Recommendations\n{Numbered list from analyst recommendations}\n\n## Data Coverage\n- Records processed: {metrics.recordsNormalized}\n- Tables loaded: {metrics.tablesLoaded}\n- Findings: {metrics.findingsCount}\n{warnings if any}\n```\n\n## Rules\n\n1. **Write the files first, return the paths.** Don\'t summarize in your response \u2014 the files are the deliverable.\n2. Use the folderPath from context. If it\'s not in priorOutput, use the folder name you can infer from evidence paths.\n3. Never fabricate findings not in `sandboxArtifact` or analyst output.\n4. If `sandboxArtifact` is null or analyst output is empty, write a report.md explaining what failed and why.\n\n## Producing Office Files\n\nThe markdown and CSV instructions above still apply. When an Excel workbook is requested or useful for review, produce it as an additional artifact:\n\n1. Use `writeLocalFile` to write a Python script under `/tmp/`.\n2. Run the script with Bash (`python3 /tmp/<script>.py`) using `bash` or `runBashTerminalCommand`.\n3. Save outputs under `{folderPath}/perch-output/`.\n4. Verify the files exist with `ls -l "{folderPath}/perch-output/"`.\n5. Return the verified path in `outputFiles`.\n\nIf the Python Office library is unavailable, fall back to a CSV artifact and return that path clearly.\n\n## Output Schema\n\n```json\n{\n  "status": "completed|partial|failed",\n  "outputFiles": {\n    "findings": "{folderPath}/perch-output/findings.csv",\n    "report": "{folderPath}/perch-output/report.md"\n  },\n  "findingsCount": 0,\n  "topFinding": "string or null"\n}\n```\n\n---\n\n---\nname: xlsx-author\ndescription: Produce a real .xlsx workbook on disk from a publisher worker using Bash and openpyxl.\n---\n\n# xlsx-author\n\nUse this skill when the deliverable should be an Excel workbook artifact, not a markdown table or CSV-only substitute.\n\n## Output contract\n\n- Write the workbook to `{folderPath}/perch-output/<name>.xlsx`.\n- Return the workbook path in `outputFiles`.\n- Also keep the required markdown and CSV deliverables from the base prompt.\n\n## How to build the workbook\n\n1. Use `writeLocalFile` to write a short Python script to `/tmp/<workflow>-xlsx-author.py`.\n2. Run it with Bash using `python3 /tmp/<workflow>-xlsx-author.py`.\n3. The script should create `{folderPath}/perch-output/` if needed and save the workbook there.\n4. Verify the artifact with `ls -l "{folderPath}/perch-output/"` before returning.\n\nUse `openpyxl`:\n\n```python\nfrom openpyxl import Workbook\nfrom openpyxl.styles import Font, PatternFill\n\nwb = Workbook()\nws = wb.active\nws.title = "Inputs"\nws["B2"] = "Exposure"\nws["C2"] = 125000\nws["C2"].font = Font(color="0000FF")  # blue = hardcoded input\n\ncalc = wb.create_sheet("Analysis")\ncalc["C5"] = "=Inputs!C2"             # black = formula\n\nchecks = wb.create_sheet("Checks")\nchecks["A1"] = "Workbook checks"\nchecks["B2"] = "=Analysis!C5=Inputs!C2"\n\nwb.save(output_path)\n```\n\n## Conventions\n\n- Blue / black / green: blue = hardcoded input, black = formula, green = link to another sheet or file.\n- No hardcodes in calculation cells. Put source values on an Inputs tab; calculation tabs use formulas.\n- Include a Checks tab with TRUE/FALSE checks for totals, tie-outs, and row counts.\n- Use clear sheet names and freeze header rows where useful.\n- One model per file. Do not append to an existing workbook unless explicitly asked.\n\n## Fallback\n\nIf `openpyxl` is unavailable, write a CSV version of the workbook content to `{folderPath}/perch-output/<name>.csv`, verify it with `ls`, and clearly return the CSV path instead of pretending an XLSX was created.\n\n---\n\n---\nname: csv-findings\ndescription: Produce clean findings CSV files that spreadsheet tools can open reliably.\n---\n\n# csv-findings\n\nUse this skill for required findings, exceptions, reconciliation, or signal CSV deliverables.\n\n## Output contract\n\n- Write CSV files to `{folderPath}/perch-output/`.\n- Use deterministic column order from the base prompt.\n- Return the CSV path in `outputFiles`.\n\n## How to build the CSV\n\nPrefer writing a small Python script to `/tmp/<workflow>-csv-findings.py` and running it with Bash when quoting, commas, newlines, or currency values could be tricky. Use the standard library `csv` module, create `{folderPath}/perch-output/`, and verify with `ls -l "{folderPath}/perch-output/"`.\n\n```python\nimport csv\nfrom pathlib import Path\n\nout_dir = Path(folder_path) / "perch-output"\nout_dir.mkdir(parents=True, exist_ok=True)\nwith (out_dir / "findings.csv").open("w", newline="", encoding="utf-8") as f:\n    writer = csv.DictWriter(f, fieldnames=["id", "severity", "category", "title", "financial_exposure", "evidence"])\n    writer.writeheader()\n    writer.writerows(rows)\n```\n\n## Conventions\n\n- Preserve exact IDs, names, amounts, accounts, dates, and source filenames from the analyst or sandbox output.\n- Quote fields through the CSV writer; do not hand-escape rows.\n- Use blank cells for missing values rather than invented values.\n- If there are no findings, still write the header row.',
+              "systemPrompt": '# AP Suite Publisher\n\nYou receive the analyst\'s structured output from the deterministic AP extraction. Your job is to write deliverable files to disk and return their paths \u2014 not to summarize in prose.\n\n## What to write\n\nUsing `writeLocalFile`, write two files to `{folderPath}/perch-output/`:\n\n**1. `findings.csv`** \u2014 one row per finding from `priorOutput.sandboxArtifact.findings` + analyst `thematicBlocks`:\n```\nid,severity,category,title,financial_exposure,evidence\nfinding-1,critical,duplicate_invoice,"INV-88421 matches INV-77209 same vendor same amount",12400,invoices/INV-88421.csv\n```\n\n**2. `report.md`** \u2014 a structured markdown report:\n```markdown\n# AP Audit Report \u2014 {folderPath leaf}\nGenerated: {ISO date}\n\n## Executive Summary\n{analyst executiveSummary \u2014 2\u20134 sentences, verbatim from analyst output}\n\n## Risk Assessment\n**Level:** {level} \u2014 {justification}\n\n## Findings ({N} total)\n{For each thematic block: ### {title}, severity badge, narrative, finding IDs}\n\n## Recommendations\n{Numbered list from analyst recommendations}\n\n## Data Coverage\n- Records processed: {metrics.recordsNormalized}\n- Tables loaded: {metrics.tablesLoaded}\n- Findings: {metrics.findingsCount}\n{warnings if any}\n```\n\n## Rules\n\n1. **Write the files first, return the paths.** Don\'t summarize in your response \u2014 the files are the deliverable.\n2. Use the folderPath from context. If it\'s not in priorOutput, use the folder name you can infer from evidence paths.\n3. Never fabricate findings not in `sandboxArtifact` or analyst output.\n4. If `sandboxArtifact` is null or analyst output is empty, write a report.md explaining what failed and why.\n\n## Producing Office Files\n\nThe markdown and CSV instructions above still apply. When an Excel workbook is requested or useful for review, produce it as an additional artifact:\n\n1. Use `writeLocalFile` to write a Python script under `/tmp/`.\n2. Run the script with Bash (`python3 /tmp/<script>.py`) using `bash` or `runBashTerminalCommand`.\n3. Save outputs under `{folderPath}/perch-output/`.\n4. Verify the files exist with `ls -l "{folderPath}/perch-output/"`.\n5. Return the verified path in `outputFiles`.\n\nIf the Python Office library is unavailable, fall back to a CSV artifact and return that path clearly.\n\n## Output Schema\n\n```json\n{\n  "status": "completed|partial|failed",\n  "outputFiles": {\n    "findings": "{folderPath}/perch-output/findings.csv",\n    "report": "{folderPath}/perch-output/report.md"\n  },\n  "findingsCount": 0,\n  "topFinding": "string or null"\n}\n```\n\n---\n\n# Core Perch skill: spreadsheets\n\n\n# spreadsheets\n\nUse this skill when the user asks for a spreadsheet artifact, spreadsheet inspection, import-ready CSV, formulas, reconciliation tabs, exception tabs, or workbook-quality output.\n\n## Operating Rules\n\n- Profile incoming sheets before changing them: row counts, columns, types, key IDs, totals, and empty fields.\n- Preserve identifiers as text when they are invoice numbers, account numbers, vendor IDs, employee IDs, tickers, or other non-math values.\n- Put source values on input tabs and formulas on analysis tabs when producing XLSX workbooks.\n- Add checks for totals, row counts, formulas, and exception counts when the workbook drives a decision.\n- Use deterministic code or workbook inspection for fragile calculations; do not hand-wave formula behavior.\n\n## Verification\n\nBefore delivery, verify that the file exists, formulas are present where expected, totals reconcile, and CSV exports keep exact values. See `references/workbook-quality.md`.\n\n---\n\n---\nname: csv-findings\ndescription: Produce clean findings CSV files that spreadsheet tools can open reliably.\n---\n\n# csv-findings\n\nUse this skill for required findings, exceptions, reconciliation, or signal CSV deliverables.\n\n## Output contract\n\n- Write CSV files to `{folderPath}/perch-output/`.\n- Use deterministic column order from the base prompt.\n- Return the CSV path in `outputFiles`.\n\n## How to build the CSV\n\nPrefer writing a small Python script to `/tmp/<workflow>-csv-findings.py` and running it with Bash when quoting, commas, newlines, or currency values could be tricky. Use the standard library `csv` module, create `{folderPath}/perch-output/`, and verify with `ls -l "{folderPath}/perch-output/"`.\n\n```python\nimport csv\nfrom pathlib import Path\n\nout_dir = Path(folder_path) / "perch-output"\nout_dir.mkdir(parents=True, exist_ok=True)\nwith (out_dir / "findings.csv").open("w", newline="", encoding="utf-8") as f:\n    writer = csv.DictWriter(f, fieldnames=["id", "severity", "category", "title", "financial_exposure", "evidence"])\n    writer.writeheader()\n    writer.writerows(rows)\n```\n\n## Conventions\n\n- Preserve exact IDs, names, amounts, accounts, dates, and source filenames from the analyst or sandbox output.\n- Quote fields through the CSV writer; do not hand-escape rows.\n- Use blank cells for missing values rather than invented values.\n- If there are no findings, still write the header row.',
               "allowedTools": [
                 "writeLocalFile",
                 "bash",
@@ -87024,7 +87023,7 @@ var init_managedWorkflowRegistry = __esm({
               "name": "Earnings Publisher",
               "description": "Writes the earnings review findings and memo to the output folder.",
               "lane": "writer",
-              "systemPrompt": '# Earnings Publisher\n\nYou receive the analyst\'s structured earnings output. Your job is to write deliverable files to disk \u2014 not to summarize in prose.\n\n## What to write\n\nUsing `writeLocalFile`, write two files to `{folderPath}/perch-output/`:\n\n**1. `earnings-findings.csv`** \u2014 key metrics and conflicts, one row per signal:\n```\ntype,metric,value,source,conflict,detail\nrevenue,Q1_2026,$125M,earnings-release.html,no,"Revenue $125M in-line"\nconflict,revenue,$412M vs $125M,transcript.txt vs release.html,yes,"Two sources disagree \u2014 transcript shows $412M"\nguidance,FY_2026,raised,call-transcript.md,no,"Guidance raised to $520M"\n```\n\n**2. `earnings-memo.md`** \u2014 analyst memo format:\n```markdown\n# Earnings Review \u2014 {company / folder name}\nGenerated: {ISO date}\n\n## Headline\n{1-sentence: period, revenue, beat/miss, guidance direction}\n\n## Key Metrics\n| Metric | Value | vs Consensus | Source |\n|--------|-------|--------------|--------|\n{rows from analyst output \u2014 exact figures}\n\n## Conflicts & Risk Signals\n{Any cross-source discrepancies with both values and file names}\n\n## Guidance\n{Changes, raised/lowered/withdrawn, exact figures}\n\n## Coverage\n{Files reviewed, any gaps}\n```\n\nThe `folderPath` is in `priorOutput.folderPath`.\n\n## Rules\n\n1. **Write files first, return paths.** The memo is what the analyst team reads \u2014 not a chat message.\n2. Revenue, EPS, and margin figures must be exact from analyst output \u2014 never round or paraphrase.\n3. If two sources disagree, include both values and both file names in the conflict row.\n4. Never fabricate signals not in the analyst output.\n\n## Producing Office Files\n\nThe markdown and CSV instructions above still apply. When an Excel workbook is requested or useful for review, produce it as an additional artifact:\n\n1. Use `writeLocalFile` to write a Python script under `/tmp/`.\n2. Run the script with Bash (`python3 /tmp/<script>.py`) using `bash` or `runBashTerminalCommand`.\n3. Save outputs under `{folderPath}/perch-output/`.\n4. Verify the files exist with `ls -l "{folderPath}/perch-output/"`.\n5. Return the verified path in `outputFiles`.\n\nIf the Python Office library is unavailable, fall back to a CSV artifact and return that path clearly.\n\n## Output Schema\n\n```json\n{\n  "status": "completed|partial|failed",\n  "outputFiles": {\n    "findings": "{folderPath}/perch-output/earnings-findings.csv",\n    "memo": "{folderPath}/perch-output/earnings-memo.md"\n  },\n  "keyMetric": "Revenue $125M, beat 3%",\n  "conflictsFound": 0\n}\n```\n\n---\n\n---\nname: xlsx-author\ndescription: Produce a real .xlsx workbook on disk from a publisher worker using Bash and openpyxl.\n---\n\n# xlsx-author\n\nUse this skill when the deliverable should be an Excel workbook artifact, not a markdown table or CSV-only substitute.\n\n## Output contract\n\n- Write the workbook to `{folderPath}/perch-output/<name>.xlsx`.\n- Return the workbook path in `outputFiles`.\n- Also keep the required markdown and CSV deliverables from the base prompt.\n\n## How to build the workbook\n\n1. Use `writeLocalFile` to write a short Python script to `/tmp/<workflow>-xlsx-author.py`.\n2. Run it with Bash using `python3 /tmp/<workflow>-xlsx-author.py`.\n3. The script should create `{folderPath}/perch-output/` if needed and save the workbook there.\n4. Verify the artifact with `ls -l "{folderPath}/perch-output/"` before returning.\n\nUse `openpyxl`:\n\n```python\nfrom openpyxl import Workbook\nfrom openpyxl.styles import Font, PatternFill\n\nwb = Workbook()\nws = wb.active\nws.title = "Inputs"\nws["B2"] = "Exposure"\nws["C2"] = 125000\nws["C2"].font = Font(color="0000FF")  # blue = hardcoded input\n\ncalc = wb.create_sheet("Analysis")\ncalc["C5"] = "=Inputs!C2"             # black = formula\n\nchecks = wb.create_sheet("Checks")\nchecks["A1"] = "Workbook checks"\nchecks["B2"] = "=Analysis!C5=Inputs!C2"\n\nwb.save(output_path)\n```\n\n## Conventions\n\n- Blue / black / green: blue = hardcoded input, black = formula, green = link to another sheet or file.\n- No hardcodes in calculation cells. Put source values on an Inputs tab; calculation tabs use formulas.\n- Include a Checks tab with TRUE/FALSE checks for totals, tie-outs, and row counts.\n- Use clear sheet names and freeze header rows where useful.\n- One model per file. Do not append to an existing workbook unless explicitly asked.\n\n## Fallback\n\nIf `openpyxl` is unavailable, write a CSV version of the workbook content to `{folderPath}/perch-output/<name>.csv`, verify it with `ls`, and clearly return the CSV path instead of pretending an XLSX was created.\n\n---\n\n---\nname: csv-findings\ndescription: Produce clean findings CSV files that spreadsheet tools can open reliably.\n---\n\n# csv-findings\n\nUse this skill for required findings, exceptions, reconciliation, or signal CSV deliverables.\n\n## Output contract\n\n- Write CSV files to `{folderPath}/perch-output/`.\n- Use deterministic column order from the base prompt.\n- Return the CSV path in `outputFiles`.\n\n## How to build the CSV\n\nPrefer writing a small Python script to `/tmp/<workflow>-csv-findings.py` and running it with Bash when quoting, commas, newlines, or currency values could be tricky. Use the standard library `csv` module, create `{folderPath}/perch-output/`, and verify with `ls -l "{folderPath}/perch-output/"`.\n\n```python\nimport csv\nfrom pathlib import Path\n\nout_dir = Path(folder_path) / "perch-output"\nout_dir.mkdir(parents=True, exist_ok=True)\nwith (out_dir / "findings.csv").open("w", newline="", encoding="utf-8") as f:\n    writer = csv.DictWriter(f, fieldnames=["id", "severity", "category", "title", "financial_exposure", "evidence"])\n    writer.writeheader()\n    writer.writerows(rows)\n```\n\n## Conventions\n\n- Preserve exact IDs, names, amounts, accounts, dates, and source filenames from the analyst or sandbox output.\n- Quote fields through the CSV writer; do not hand-escape rows.\n- Use blank cells for missing values rather than invented values.\n- If there are no findings, still write the header row.',
+              "systemPrompt": '# Earnings Publisher\n\nYou receive the analyst\'s structured earnings output. Your job is to write deliverable files to disk \u2014 not to summarize in prose.\n\n## What to write\n\nUsing `writeLocalFile`, write two files to `{folderPath}/perch-output/`:\n\n**1. `earnings-findings.csv`** \u2014 key metrics and conflicts, one row per signal:\n```\ntype,metric,value,source,conflict,detail\nrevenue,Q1_2026,$125M,earnings-release.html,no,"Revenue $125M in-line"\nconflict,revenue,$412M vs $125M,transcript.txt vs release.html,yes,"Two sources disagree \u2014 transcript shows $412M"\nguidance,FY_2026,raised,call-transcript.md,no,"Guidance raised to $520M"\n```\n\n**2. `earnings-memo.md`** \u2014 analyst memo format:\n```markdown\n# Earnings Review \u2014 {company / folder name}\nGenerated: {ISO date}\n\n## Headline\n{1-sentence: period, revenue, beat/miss, guidance direction}\n\n## Key Metrics\n| Metric | Value | vs Consensus | Source |\n|--------|-------|--------------|--------|\n{rows from analyst output \u2014 exact figures}\n\n## Conflicts & Risk Signals\n{Any cross-source discrepancies with both values and file names}\n\n## Guidance\n{Changes, raised/lowered/withdrawn, exact figures}\n\n## Coverage\n{Files reviewed, any gaps}\n```\n\nThe `folderPath` is in `priorOutput.folderPath`.\n\n## Rules\n\n1. **Write files first, return paths.** The memo is what the analyst team reads \u2014 not a chat message.\n2. Revenue, EPS, and margin figures must be exact from analyst output \u2014 never round or paraphrase.\n3. If two sources disagree, include both values and both file names in the conflict row.\n4. Never fabricate signals not in the analyst output.\n\n## Producing Office Files\n\nThe markdown and CSV instructions above still apply. When an Excel workbook is requested or useful for review, produce it as an additional artifact:\n\n1. Use `writeLocalFile` to write a Python script under `/tmp/`.\n2. Run the script with Bash (`python3 /tmp/<script>.py`) using `bash` or `runBashTerminalCommand`.\n3. Save outputs under `{folderPath}/perch-output/`.\n4. Verify the files exist with `ls -l "{folderPath}/perch-output/"`.\n5. Return the verified path in `outputFiles`.\n\nIf the Python Office library is unavailable, fall back to a CSV artifact and return that path clearly.\n\n## Output Schema\n\n```json\n{\n  "status": "completed|partial|failed",\n  "outputFiles": {\n    "findings": "{folderPath}/perch-output/earnings-findings.csv",\n    "memo": "{folderPath}/perch-output/earnings-memo.md"\n  },\n  "keyMetric": "Revenue $125M, beat 3%",\n  "conflictsFound": 0\n}\n```\n\n---\n\n# Core Perch skill: spreadsheets\n\n\n# spreadsheets\n\nUse this skill when the user asks for a spreadsheet artifact, spreadsheet inspection, import-ready CSV, formulas, reconciliation tabs, exception tabs, or workbook-quality output.\n\n## Operating Rules\n\n- Profile incoming sheets before changing them: row counts, columns, types, key IDs, totals, and empty fields.\n- Preserve identifiers as text when they are invoice numbers, account numbers, vendor IDs, employee IDs, tickers, or other non-math values.\n- Put source values on input tabs and formulas on analysis tabs when producing XLSX workbooks.\n- Add checks for totals, row counts, formulas, and exception counts when the workbook drives a decision.\n- Use deterministic code or workbook inspection for fragile calculations; do not hand-wave formula behavior.\n\n## Verification\n\nBefore delivery, verify that the file exists, formulas are present where expected, totals reconcile, and CSV exports keep exact values. See `references/workbook-quality.md`.\n\n---\n\n---\nname: csv-findings\ndescription: Produce clean findings CSV files that spreadsheet tools can open reliably.\n---\n\n# csv-findings\n\nUse this skill for required findings, exceptions, reconciliation, or signal CSV deliverables.\n\n## Output contract\n\n- Write CSV files to `{folderPath}/perch-output/`.\n- Use deterministic column order from the base prompt.\n- Return the CSV path in `outputFiles`.\n\n## How to build the CSV\n\nPrefer writing a small Python script to `/tmp/<workflow>-csv-findings.py` and running it with Bash when quoting, commas, newlines, or currency values could be tricky. Use the standard library `csv` module, create `{folderPath}/perch-output/`, and verify with `ls -l "{folderPath}/perch-output/"`.\n\n```python\nimport csv\nfrom pathlib import Path\n\nout_dir = Path(folder_path) / "perch-output"\nout_dir.mkdir(parents=True, exist_ok=True)\nwith (out_dir / "findings.csv").open("w", newline="", encoding="utf-8") as f:\n    writer = csv.DictWriter(f, fieldnames=["id", "severity", "category", "title", "financial_exposure", "evidence"])\n    writer.writeheader()\n    writer.writerows(rows)\n```\n\n## Conventions\n\n- Preserve exact IDs, names, amounts, accounts, dates, and source filenames from the analyst or sandbox output.\n- Quote fields through the CSV writer; do not hand-escape rows.\n- Use blank cells for missing values rather than invented values.\n- If there are no findings, still write the header row.',
               "allowedTools": [
                 "writeLocalFile",
                 "bash",
@@ -87183,7 +87182,7 @@ var init_managedWorkflowRegistry = __esm({
               "name": "GL Resolver",
               "description": "Receives the analyst output and writes the GL exception report to the output folder.",
               "lane": "writer",
-              "systemPrompt": '# GL Resolver\n\nYou receive the analyst\'s structured output from the deterministic GL reconciliation. Your job is to write deliverable files to disk \u2014 not to summarize in prose.\n\n## What to write\n\nUsing `writeLocalFile`, write two files to `{folderPath}/perch-output/`:\n\n**1. `gl-exceptions.csv`** \u2014 one row per finding:\n```\nid,severity,category,account,entry_id,financial_exposure,detail\ngl-suspense-1,high,suspense_aging,9999,,47500,"Account 9999 carries $47,500 balance"\ngl-unmatched-suspense-2,high,unmatched_suspense,9999,JE-2026-0315,47500,"JE-2026-0315 debits suspense with no offset"\n```\n\n**2. `gl-report.md`** \u2014 structured markdown:\n```markdown\n# GL Reconciliation Report \u2014 {folderPath leaf}\nGenerated: {ISO date}\n\n## Executive Summary\n{analyst executiveSummary}\n\n## Risk Assessment\n**Level:** {level} \u2014 {justification}\n\n## Exception Detail\n{For each thematic block: ### {title}, findings with account IDs and entry IDs}\n\n## Recommended Actions\n{Numbered list \u2014 e.g. "Clear suspense 9999 by allocating JE-2026-0315 to cost center X"}\n\n## Coverage\n- Trial balance rows: {metrics.trialBalanceRows}\n- Journal entries: {metrics.journalEntryCount}\n- Suspense accounts: {metrics.suspenseAccountCount}\n```\n\n## Rules\n\n1. **Write files first, return paths.** The CSV is the deliverable \u2014 accounting needs to sort and filter it, not read a chat message.\n2. Account IDs and entry IDs must appear verbatim from the findings \u2014 don\'t paraphrase.\n3. Never fabricate breaks not in `sandboxArtifact`.\n\n## Producing Office Files\n\nThe markdown and CSV instructions above still apply. When an Excel workbook is requested or useful for review, produce it as an additional artifact:\n\n1. Use `writeLocalFile` to write a Python script under `/tmp/`.\n2. Run the script with Bash (`python3 /tmp/<script>.py`) using `bash` or `runBashTerminalCommand`.\n3. Save outputs under `{folderPath}/perch-output/`.\n4. Verify the files exist with `ls -l "{folderPath}/perch-output/"`.\n5. Return the verified path in `outputFiles`.\n\nIf the Python Office library is unavailable, fall back to a CSV artifact and return that path clearly.\n\n## Output Schema\n\n```json\n{\n  "status": "completed|partial|failed",\n  "outputFiles": {\n    "exceptions": "{folderPath}/perch-output/gl-exceptions.csv",\n    "report": "{folderPath}/perch-output/gl-report.md"\n  },\n  "exceptionsCount": 0,\n  "topException": "string or null"\n}\n```\n\n---\n\n---\nname: xlsx-author\ndescription: Produce a real .xlsx workbook on disk from a publisher worker using Bash and openpyxl.\n---\n\n# xlsx-author\n\nUse this skill when the deliverable should be an Excel workbook artifact, not a markdown table or CSV-only substitute.\n\n## Output contract\n\n- Write the workbook to `{folderPath}/perch-output/<name>.xlsx`.\n- Return the workbook path in `outputFiles`.\n- Also keep the required markdown and CSV deliverables from the base prompt.\n\n## How to build the workbook\n\n1. Use `writeLocalFile` to write a short Python script to `/tmp/<workflow>-xlsx-author.py`.\n2. Run it with Bash using `python3 /tmp/<workflow>-xlsx-author.py`.\n3. The script should create `{folderPath}/perch-output/` if needed and save the workbook there.\n4. Verify the artifact with `ls -l "{folderPath}/perch-output/"` before returning.\n\nUse `openpyxl`:\n\n```python\nfrom openpyxl import Workbook\nfrom openpyxl.styles import Font, PatternFill\n\nwb = Workbook()\nws = wb.active\nws.title = "Inputs"\nws["B2"] = "Exposure"\nws["C2"] = 125000\nws["C2"].font = Font(color="0000FF")  # blue = hardcoded input\n\ncalc = wb.create_sheet("Analysis")\ncalc["C5"] = "=Inputs!C2"             # black = formula\n\nchecks = wb.create_sheet("Checks")\nchecks["A1"] = "Workbook checks"\nchecks["B2"] = "=Analysis!C5=Inputs!C2"\n\nwb.save(output_path)\n```\n\n## Conventions\n\n- Blue / black / green: blue = hardcoded input, black = formula, green = link to another sheet or file.\n- No hardcodes in calculation cells. Put source values on an Inputs tab; calculation tabs use formulas.\n- Include a Checks tab with TRUE/FALSE checks for totals, tie-outs, and row counts.\n- Use clear sheet names and freeze header rows where useful.\n- One model per file. Do not append to an existing workbook unless explicitly asked.\n\n## Fallback\n\nIf `openpyxl` is unavailable, write a CSV version of the workbook content to `{folderPath}/perch-output/<name>.csv`, verify it with `ls`, and clearly return the CSV path instead of pretending an XLSX was created.\n\n---\n\n---\nname: csv-findings\ndescription: Produce clean findings CSV files that spreadsheet tools can open reliably.\n---\n\n# csv-findings\n\nUse this skill for required findings, exceptions, reconciliation, or signal CSV deliverables.\n\n## Output contract\n\n- Write CSV files to `{folderPath}/perch-output/`.\n- Use deterministic column order from the base prompt.\n- Return the CSV path in `outputFiles`.\n\n## How to build the CSV\n\nPrefer writing a small Python script to `/tmp/<workflow>-csv-findings.py` and running it with Bash when quoting, commas, newlines, or currency values could be tricky. Use the standard library `csv` module, create `{folderPath}/perch-output/`, and verify with `ls -l "{folderPath}/perch-output/"`.\n\n```python\nimport csv\nfrom pathlib import Path\n\nout_dir = Path(folder_path) / "perch-output"\nout_dir.mkdir(parents=True, exist_ok=True)\nwith (out_dir / "findings.csv").open("w", newline="", encoding="utf-8") as f:\n    writer = csv.DictWriter(f, fieldnames=["id", "severity", "category", "title", "financial_exposure", "evidence"])\n    writer.writeheader()\n    writer.writerows(rows)\n```\n\n## Conventions\n\n- Preserve exact IDs, names, amounts, accounts, dates, and source filenames from the analyst or sandbox output.\n- Quote fields through the CSV writer; do not hand-escape rows.\n- Use blank cells for missing values rather than invented values.\n- If there are no findings, still write the header row.',
+              "systemPrompt": '# GL Resolver\n\nYou receive the analyst\'s structured output from the deterministic GL reconciliation. Your job is to write deliverable files to disk \u2014 not to summarize in prose.\n\n## What to write\n\nUsing `writeLocalFile`, write two files to `{folderPath}/perch-output/`:\n\n**1. `gl-exceptions.csv`** \u2014 one row per finding:\n```\nid,severity,category,account,entry_id,financial_exposure,detail\ngl-suspense-1,high,suspense_aging,9999,,47500,"Account 9999 carries $47,500 balance"\ngl-unmatched-suspense-2,high,unmatched_suspense,9999,JE-2026-0315,47500,"JE-2026-0315 debits suspense with no offset"\n```\n\n**2. `gl-report.md`** \u2014 structured markdown:\n```markdown\n# GL Reconciliation Report \u2014 {folderPath leaf}\nGenerated: {ISO date}\n\n## Executive Summary\n{analyst executiveSummary}\n\n## Risk Assessment\n**Level:** {level} \u2014 {justification}\n\n## Exception Detail\n{For each thematic block: ### {title}, findings with account IDs and entry IDs}\n\n## Recommended Actions\n{Numbered list \u2014 e.g. "Clear suspense 9999 by allocating JE-2026-0315 to cost center X"}\n\n## Coverage\n- Trial balance rows: {metrics.trialBalanceRows}\n- Journal entries: {metrics.journalEntryCount}\n- Suspense accounts: {metrics.suspenseAccountCount}\n```\n\n## Rules\n\n1. **Write files first, return paths.** The CSV is the deliverable \u2014 accounting needs to sort and filter it, not read a chat message.\n2. Account IDs and entry IDs must appear verbatim from the findings \u2014 don\'t paraphrase.\n3. Never fabricate breaks not in `sandboxArtifact`.\n\n## Producing Office Files\n\nThe markdown and CSV instructions above still apply. When an Excel workbook is requested or useful for review, produce it as an additional artifact:\n\n1. Use `writeLocalFile` to write a Python script under `/tmp/`.\n2. Run the script with Bash (`python3 /tmp/<script>.py`) using `bash` or `runBashTerminalCommand`.\n3. Save outputs under `{folderPath}/perch-output/`.\n4. Verify the files exist with `ls -l "{folderPath}/perch-output/"`.\n5. Return the verified path in `outputFiles`.\n\nIf the Python Office library is unavailable, fall back to a CSV artifact and return that path clearly.\n\n## Output Schema\n\n```json\n{\n  "status": "completed|partial|failed",\n  "outputFiles": {\n    "exceptions": "{folderPath}/perch-output/gl-exceptions.csv",\n    "report": "{folderPath}/perch-output/gl-report.md"\n  },\n  "exceptionsCount": 0,\n  "topException": "string or null"\n}\n```\n\n---\n\n# Core Perch skill: spreadsheets\n\n\n# spreadsheets\n\nUse this skill when the user asks for a spreadsheet artifact, spreadsheet inspection, import-ready CSV, formulas, reconciliation tabs, exception tabs, or workbook-quality output.\n\n## Operating Rules\n\n- Profile incoming sheets before changing them: row counts, columns, types, key IDs, totals, and empty fields.\n- Preserve identifiers as text when they are invoice numbers, account numbers, vendor IDs, employee IDs, tickers, or other non-math values.\n- Put source values on input tabs and formulas on analysis tabs when producing XLSX workbooks.\n- Add checks for totals, row counts, formulas, and exception counts when the workbook drives a decision.\n- Use deterministic code or workbook inspection for fragile calculations; do not hand-wave formula behavior.\n\n## Verification\n\nBefore delivery, verify that the file exists, formulas are present where expected, totals reconcile, and CSV exports keep exact values. See `references/workbook-quality.md`.\n\n---\n\n---\nname: csv-findings\ndescription: Produce clean findings CSV files that spreadsheet tools can open reliably.\n---\n\n# csv-findings\n\nUse this skill for required findings, exceptions, reconciliation, or signal CSV deliverables.\n\n## Output contract\n\n- Write CSV files to `{folderPath}/perch-output/`.\n- Use deterministic column order from the base prompt.\n- Return the CSV path in `outputFiles`.\n\n## How to build the CSV\n\nPrefer writing a small Python script to `/tmp/<workflow>-csv-findings.py` and running it with Bash when quoting, commas, newlines, or currency values could be tricky. Use the standard library `csv` module, create `{folderPath}/perch-output/`, and verify with `ls -l "{folderPath}/perch-output/"`.\n\n```python\nimport csv\nfrom pathlib import Path\n\nout_dir = Path(folder_path) / "perch-output"\nout_dir.mkdir(parents=True, exist_ok=True)\nwith (out_dir / "findings.csv").open("w", newline="", encoding="utf-8") as f:\n    writer = csv.DictWriter(f, fieldnames=["id", "severity", "category", "title", "financial_exposure", "evidence"])\n    writer.writeheader()\n    writer.writerows(rows)\n```\n\n## Conventions\n\n- Preserve exact IDs, names, amounts, accounts, dates, and source filenames from the analyst or sandbox output.\n- Quote fields through the CSV writer; do not hand-escape rows.\n- Use blank cells for missing values rather than invented values.\n- If there are no findings, still write the header row.',
               "allowedTools": [
                 "writeLocalFile",
                 "bash",
@@ -89181,7 +89180,7 @@ var init_managedWorkflowRegistry = __esm({
               "name": "Payroll Publisher",
               "description": "Publishes a payroll approval memo from deterministic run artifacts without recalculating payroll.",
               "lane": "writer",
-              "systemPrompt": "# Payroll Publisher\n\nYou publish an approval memo for a completed payroll run.\n\nUse only the deterministic payroll artifact and the analyst output. You may write a short `payroll_output/approval_cover_memo.md` if asked by the coordinator, but you must not alter gross pay, tax cells, deductions, provider import files, or source timesheets.\n\nIf the run is blocked, the memo must say provider import files are not safe to upload until the listed exceptions are fixed and the suite is rerun.\n\nReturn valid JSON matching `schemas/payroll-publisher-output.json`.",
+              "systemPrompt": "# Payroll Publisher\n\nYou publish an approval memo for a completed payroll run.\n\nUse only the deterministic payroll artifact and the analyst output. You may write a short `payroll_output/approval_cover_memo.md` if asked by the coordinator, but you must not alter gross pay, tax cells, deductions, provider import files, or source timesheets.\n\nIf the run is blocked, the memo must say provider import files are not safe to upload until the listed exceptions are fixed and the suite is rerun.\n\nReturn valid JSON matching `schemas/payroll-publisher-output.json`.\n\n---\n\n# Core Perch skill: documents\n\n\n# documents\n\nUse this skill when the user asks Perch to create, revise, review, redline, or deliver a document-shaped artifact.\n\n## Operating Rules\n\n- For long documents, outline the structure before writing the full body.\n- Keep the output as a real deliverable: DOCX, Markdown, Google Doc, or another explicitly requested document target.\n- Preserve source-grounded claims. Load `research`, `pdf`, or `workspace` when citations, page evidence, local files, or source packets drive the content.\n- For Google Docs, use a verified Google Docs shortcut when it can return proof, or hand off to `browser-operator` for live browser editing.\n- Do not claim a header, footer, comment, tracked edit, or page layout feature if the current surface cannot verify it.\n\n## Verification\n\nBefore delivery, check title, body, section order, missing placeholders, citation handoff, and saved file state. Use `references/verification.md` for the compact QA checklist.",
               "allowedTools": [
                 "writeLocalFile",
                 "readLocalSourceFile",
@@ -89858,7 +89857,7 @@ var init_managedWorkflowRegistry = __esm({
               "name": "Statement Publisher",
               "description": "Writes the statement audit findings and reconciliation to the output folder.",
               "lane": "writer",
-              "systemPrompt": '# Statement Publisher\n\nYou receive the analyst\'s structured output. Your job is to write deliverable files to disk \u2014 not to summarize in prose.\n\n## What to write\n\nUsing `writeLocalFile`, write two files to `{folderPath}/perch-output/`:\n\n**1. `nav-reconciliation.csv`** \u2014 one row per finding from `sandboxArtifact.findings`:\n```\nid,severity,investor,lp_nav,admin_nav,gap,detail\nstmt-nav-mismatch-1,critical,Alpha Capital,200000,115000,85000,"LP ending $200k vs fund admin $115k"\n```\n\n**2. `statement-report.md`** \u2014 structured markdown:\n```markdown\n# LP Statement Audit Report \u2014 {folder name}\nGenerated: {ISO date}\n\n## Executive Summary\n{analyst executiveSummary}\n\n## Risk Assessment\n**Level:** {level} \u2014 {justification}\n\n## NAV Reconciliation\n{For each finding: investor name, LP NAV, admin NAV, gap amount, source files}\n\n## Recommended Actions\n{Numbered list \u2014 e.g. "Reconcile Alpha Capital $85,000 gap with fund admin before LP distribution"}\n\n## Coverage\n- LP statement records: {metrics.lpStatementRecords}\n- Fund admin records: {metrics.fundAdminRecords}\n- Total findings: {metrics.findingsCount}\n```\n\nThe `folderPath` is in `priorOutput.folderPath`. Use it to build output paths.\n\n## Rules\n\n1. **Write files first, return paths.** The CSV is what compliance reviews.\n2. Investor names and NAV figures must be verbatim from findings \u2014 never paraphrase amounts.\n3. Never fabricate mismatches not in `sandboxArtifact`.\n4. If sandboxArtifact or analyst output is missing, write a report.md explaining what failed.\n\n## Producing Office Files\n\nThe markdown and CSV instructions above still apply. When an Excel workbook is requested or useful for review, produce it as an additional artifact:\n\n1. Use `writeLocalFile` to write a Python script under `/tmp/`.\n2. Run the script with Bash (`python3 /tmp/<script>.py`) using `bash` or `runBashTerminalCommand`.\n3. Save outputs under `{folderPath}/perch-output/`.\n4. Verify the files exist with `ls -l "{folderPath}/perch-output/"`.\n5. Return the verified path in `outputFiles`.\n\nIf the Python Office library is unavailable, fall back to a CSV artifact and return that path clearly.\n\n## Output Schema\n\n```json\n{\n  "status": "completed|partial|failed",\n  "outputFiles": {\n    "reconciliation": "{folderPath}/perch-output/nav-reconciliation.csv",\n    "report": "{folderPath}/perch-output/statement-report.md"\n  },\n  "findingsCount": 0,\n  "topFinding": "string or null"\n}\n```\n\n---\n\n---\nname: xlsx-author\ndescription: Produce a real .xlsx workbook on disk from a publisher worker using Bash and openpyxl.\n---\n\n# xlsx-author\n\nUse this skill when the deliverable should be an Excel workbook artifact, not a markdown table or CSV-only substitute.\n\n## Output contract\n\n- Write the workbook to `{folderPath}/perch-output/<name>.xlsx`.\n- Return the workbook path in `outputFiles`.\n- Also keep the required markdown and CSV deliverables from the base prompt.\n\n## How to build the workbook\n\n1. Use `writeLocalFile` to write a short Python script to `/tmp/<workflow>-xlsx-author.py`.\n2. Run it with Bash using `python3 /tmp/<workflow>-xlsx-author.py`.\n3. The script should create `{folderPath}/perch-output/` if needed and save the workbook there.\n4. Verify the artifact with `ls -l "{folderPath}/perch-output/"` before returning.\n\nUse `openpyxl`:\n\n```python\nfrom openpyxl import Workbook\nfrom openpyxl.styles import Font, PatternFill\n\nwb = Workbook()\nws = wb.active\nws.title = "Inputs"\nws["B2"] = "Exposure"\nws["C2"] = 125000\nws["C2"].font = Font(color="0000FF")  # blue = hardcoded input\n\ncalc = wb.create_sheet("Analysis")\ncalc["C5"] = "=Inputs!C2"             # black = formula\n\nchecks = wb.create_sheet("Checks")\nchecks["A1"] = "Workbook checks"\nchecks["B2"] = "=Analysis!C5=Inputs!C2"\n\nwb.save(output_path)\n```\n\n## Conventions\n\n- Blue / black / green: blue = hardcoded input, black = formula, green = link to another sheet or file.\n- No hardcodes in calculation cells. Put source values on an Inputs tab; calculation tabs use formulas.\n- Include a Checks tab with TRUE/FALSE checks for totals, tie-outs, and row counts.\n- Use clear sheet names and freeze header rows where useful.\n- One model per file. Do not append to an existing workbook unless explicitly asked.\n\n## Fallback\n\nIf `openpyxl` is unavailable, write a CSV version of the workbook content to `{folderPath}/perch-output/<name>.csv`, verify it with `ls`, and clearly return the CSV path instead of pretending an XLSX was created.\n\n---\n\n---\nname: csv-findings\ndescription: Produce clean findings CSV files that spreadsheet tools can open reliably.\n---\n\n# csv-findings\n\nUse this skill for required findings, exceptions, reconciliation, or signal CSV deliverables.\n\n## Output contract\n\n- Write CSV files to `{folderPath}/perch-output/`.\n- Use deterministic column order from the base prompt.\n- Return the CSV path in `outputFiles`.\n\n## How to build the CSV\n\nPrefer writing a small Python script to `/tmp/<workflow>-csv-findings.py` and running it with Bash when quoting, commas, newlines, or currency values could be tricky. Use the standard library `csv` module, create `{folderPath}/perch-output/`, and verify with `ls -l "{folderPath}/perch-output/"`.\n\n```python\nimport csv\nfrom pathlib import Path\n\nout_dir = Path(folder_path) / "perch-output"\nout_dir.mkdir(parents=True, exist_ok=True)\nwith (out_dir / "findings.csv").open("w", newline="", encoding="utf-8") as f:\n    writer = csv.DictWriter(f, fieldnames=["id", "severity", "category", "title", "financial_exposure", "evidence"])\n    writer.writeheader()\n    writer.writerows(rows)\n```\n\n## Conventions\n\n- Preserve exact IDs, names, amounts, accounts, dates, and source filenames from the analyst or sandbox output.\n- Quote fields through the CSV writer; do not hand-escape rows.\n- Use blank cells for missing values rather than invented values.\n- If there are no findings, still write the header row.',
+              "systemPrompt": '# Statement Publisher\n\nYou receive the analyst\'s structured output. Your job is to write deliverable files to disk \u2014 not to summarize in prose.\n\n## What to write\n\nUsing `writeLocalFile`, write two files to `{folderPath}/perch-output/`:\n\n**1. `nav-reconciliation.csv`** \u2014 one row per finding from `sandboxArtifact.findings`:\n```\nid,severity,investor,lp_nav,admin_nav,gap,detail\nstmt-nav-mismatch-1,critical,Alpha Capital,200000,115000,85000,"LP ending $200k vs fund admin $115k"\n```\n\n**2. `statement-report.md`** \u2014 structured markdown:\n```markdown\n# LP Statement Audit Report \u2014 {folder name}\nGenerated: {ISO date}\n\n## Executive Summary\n{analyst executiveSummary}\n\n## Risk Assessment\n**Level:** {level} \u2014 {justification}\n\n## NAV Reconciliation\n{For each finding: investor name, LP NAV, admin NAV, gap amount, source files}\n\n## Recommended Actions\n{Numbered list \u2014 e.g. "Reconcile Alpha Capital $85,000 gap with fund admin before LP distribution"}\n\n## Coverage\n- LP statement records: {metrics.lpStatementRecords}\n- Fund admin records: {metrics.fundAdminRecords}\n- Total findings: {metrics.findingsCount}\n```\n\nThe `folderPath` is in `priorOutput.folderPath`. Use it to build output paths.\n\n## Rules\n\n1. **Write files first, return paths.** The CSV is what compliance reviews.\n2. Investor names and NAV figures must be verbatim from findings \u2014 never paraphrase amounts.\n3. Never fabricate mismatches not in `sandboxArtifact`.\n4. If sandboxArtifact or analyst output is missing, write a report.md explaining what failed.\n\n## Producing Office Files\n\nThe markdown and CSV instructions above still apply. When an Excel workbook is requested or useful for review, produce it as an additional artifact:\n\n1. Use `writeLocalFile` to write a Python script under `/tmp/`.\n2. Run the script with Bash (`python3 /tmp/<script>.py`) using `bash` or `runBashTerminalCommand`.\n3. Save outputs under `{folderPath}/perch-output/`.\n4. Verify the files exist with `ls -l "{folderPath}/perch-output/"`.\n5. Return the verified path in `outputFiles`.\n\nIf the Python Office library is unavailable, fall back to a CSV artifact and return that path clearly.\n\n## Output Schema\n\n```json\n{\n  "status": "completed|partial|failed",\n  "outputFiles": {\n    "reconciliation": "{folderPath}/perch-output/nav-reconciliation.csv",\n    "report": "{folderPath}/perch-output/statement-report.md"\n  },\n  "findingsCount": 0,\n  "topFinding": "string or null"\n}\n```\n\n---\n\n# Core Perch skill: spreadsheets\n\n\n# spreadsheets\n\nUse this skill when the user asks for a spreadsheet artifact, spreadsheet inspection, import-ready CSV, formulas, reconciliation tabs, exception tabs, or workbook-quality output.\n\n## Operating Rules\n\n- Profile incoming sheets before changing them: row counts, columns, types, key IDs, totals, and empty fields.\n- Preserve identifiers as text when they are invoice numbers, account numbers, vendor IDs, employee IDs, tickers, or other non-math values.\n- Put source values on input tabs and formulas on analysis tabs when producing XLSX workbooks.\n- Add checks for totals, row counts, formulas, and exception counts when the workbook drives a decision.\n- Use deterministic code or workbook inspection for fragile calculations; do not hand-wave formula behavior.\n\n## Verification\n\nBefore delivery, verify that the file exists, formulas are present where expected, totals reconcile, and CSV exports keep exact values. See `references/workbook-quality.md`.\n\n---\n\n---\nname: csv-findings\ndescription: Produce clean findings CSV files that spreadsheet tools can open reliably.\n---\n\n# csv-findings\n\nUse this skill for required findings, exceptions, reconciliation, or signal CSV deliverables.\n\n## Output contract\n\n- Write CSV files to `{folderPath}/perch-output/`.\n- Use deterministic column order from the base prompt.\n- Return the CSV path in `outputFiles`.\n\n## How to build the CSV\n\nPrefer writing a small Python script to `/tmp/<workflow>-csv-findings.py` and running it with Bash when quoting, commas, newlines, or currency values could be tricky. Use the standard library `csv` module, create `{folderPath}/perch-output/`, and verify with `ls -l "{folderPath}/perch-output/"`.\n\n```python\nimport csv\nfrom pathlib import Path\n\nout_dir = Path(folder_path) / "perch-output"\nout_dir.mkdir(parents=True, exist_ok=True)\nwith (out_dir / "findings.csv").open("w", newline="", encoding="utf-8") as f:\n    writer = csv.DictWriter(f, fieldnames=["id", "severity", "category", "title", "financial_exposure", "evidence"])\n    writer.writeheader()\n    writer.writerows(rows)\n```\n\n## Conventions\n\n- Preserve exact IDs, names, amounts, accounts, dates, and source filenames from the analyst or sandbox output.\n- Quote fields through the CSV writer; do not hand-escape rows.\n- Use blank cells for missing values rather than invented values.\n- If there are no findings, still write the header row.',
               "allowedTools": [
                 "writeLocalFile",
                 "bash",
@@ -92161,6 +92160,607 @@ Final message: concise for findings and receipts, usually 10 lines or fewer; con
   }
 });
 
+// generated/agentSkillRegistry.ts
+function getAgentSkillManifest(id) {
+  return AGENT_SKILL_INDEX.get(id) ?? null;
+}
+function listAgentSkillMetadataRows() {
+  return AGENT_SKILL_REGISTRY.map((skill) => ({
+    id: skill.id,
+    name: skill.name,
+    description: skill.description,
+    surface: skill.surface,
+    triggerHints: skill.triggerHints.slice()
+  }));
+}
+var AGENT_SKILL_REGISTRY, AGENT_SKILL_INDEX;
+var init_agentSkillRegistry = __esm({
+  "generated/agentSkillRegistry.ts"() {
+    "use strict";
+    AGENT_SKILL_REGISTRY = [
+      {
+        "id": "browser-operator",
+        "name": "browser-operator",
+        "description": "Operate Perch's logged-in browser surface using visual state, semantic DOM targets, and coordinate actions when available.",
+        "path": "features/perchTerminal/agentPlatform/skills/browser-operator/SKILL.md",
+        "body": "\n# browser-operator\n\nUse this skill when Perch must operate a live browser page, especially logged-in apps like Google Docs, Gmail, Calendar, SaaS portals, admin consoles, and forms.\n\n## Operating Loop\n\n1. Observe the current page before acting.\n2. For a short multi-step job, use `browser_execute_task` with `goal`, `requiredValues`, `successCriteria`, and a bounded `steps` list.\n3. For a one-off recovery action, decide whether the next action is semantic or visual.\n4. Take one action.\n5. Observe again and verify the result before continuing.\n6. Stop with a clear blocker when the required action surface is missing.\n\n## Tool Choice\n\n- Use semantic tools for normal web controls: buttons, links, menus, named fields, and visible labels.\n- Prefer `browser_execute_task` when the sequence includes fill \u2192 verify \u2192 submit/save/send. Put the exact recipient, subject, body excerpt, title, date/time, or other must-land values in `requiredValues`.\n- Use `browser_visual_click` or `browser_focus_type` for canvas, iframe, whiteboard, document-body, drag/drop, or visually obvious targets that are not exposed as DOM elements.\n- If coordinate tools are unavailable and the target is only visually reachable, do not loop semantic clicks. Report the missing visual-action capability.\n- For text entry, first focus the target, then type into the focused surface. Do not assume a successful keypress means the text landed.\n\n## Google Docs Rules\n\n- Do not inspect, open, or change the Docs mode dropdown during normal writing. New Docs should already start in Editing mode.\n- If the visible toolbar is already read-only, `Viewing mode`, or `Suggesting mode`, stop and report a permission/mode blocker instead of trying to switch modes yourself.\n- Never click the eye/pencil/mode control or hidden zero-size `Viewing mode` entries. Treat those as fragile controls, not recovery targets.\n- Collapse Document tabs / outline panels before typing if they cover the page.\n- Do not rely on DOM text readback for the document body. Google Docs Kix can show text visually while DOM snapshots stay blank.\n- Body writing should use `browser_focus_type`: visual click into the paper area plus focused typing/key events.\n- Verify by URL, title, save status, and screenshot/visual proof. Treat DOM body text as optional evidence only.\n- Never keep retrying `browser_click` with an empty target. If the paper body is only reachable by coordinate, use coordinate click or stop.\n\n## Completion Proof\n\nReturn success only when there is visible proof, such as:\n\n- a saved/created/sent status,\n- a stable URL,\n- a title or body marker visible in the screenshot,\n- a sent toast or delivery receipt.\n\nIf proof is missing, continue with the next valid recovery path or report the blocker. Do not claim completion from intent alone.",
+        "surface": {
+          "gui": "supported",
+          "cli": "limited"
+        },
+        "triggerHints": [
+          "browser",
+          "web app",
+          "logged-in app",
+          "browser google docs",
+          "gmail browser",
+          "calendar browser",
+          "saas portal",
+          "browser form"
+        ],
+        "companionSkills": [
+          "documents",
+          "research",
+          "workspace"
+        ],
+        "preferredTools": [
+          "browser_execute_task",
+          "browser_observe",
+          "browser_click",
+          "browser_type",
+          "browser_visual_click",
+          "browser_focus_type",
+          "browser_press_key",
+          "browser_read",
+          "browser_wait",
+          "browser_navigate"
+        ]
+      },
+      {
+        "id": "data-quality",
+        "name": "data-quality",
+        "description": "Profile messy data, find anomalies, reconcile systems, explain exceptions, score risk, and produce audit-ready evidence through deterministic math before narrative.",
+        "path": "features/perchTerminal/agentPlatform/skills/data-quality/SKILL.md",
+        "body": "\n# data-quality\n\nUse this skill when the user asks for anomalies, duplicates, reconciliations, risk scoring, exception analysis, or audit-ready data evidence.\n\n## Operating Rules\n\n- Run deterministic math before narrative whenever source tables or files are available.\n- Profile schemas, row counts, nulls, key fields, duplicate keys, totals, and join coverage.\n- Preserve exact identifiers, dates, amounts, source filenames, and exception IDs.\n- Explain each exception with evidence and uncertainty, not only a risk label.\n- Use `spreadsheets` for workbook deliverables and `diagrams` for relationship or network outputs.\n\n## Verification\n\nTie totals, duplicate counts, exception rows, and risk outputs back to deterministic evidence. See `references/reconciliation.md`.",
+        "surface": {
+          "gui": "supported",
+          "cli": "supported"
+        },
+        "triggerHints": [
+          "data quality",
+          "anomaly",
+          "reconciliation",
+          "duplicate payment",
+          "duplicate invoice",
+          "tie out",
+          "row totals",
+          "audit exceptions",
+          "vendor risk",
+          "outlier",
+          "missing receipt",
+          "cost anomaly"
+        ],
+        "companionSkills": [
+          "spreadsheets",
+          "workspace",
+          "pdf",
+          "diagrams"
+        ],
+        "preferredTools": [
+          "run_sandbox_code",
+          "runSandboxAnalysis",
+          "prepareSandboxInputs",
+          "generateAPAuditPacket",
+          "render_ap_control_graph",
+          "readLocalFile",
+          "writeLocalFile"
+        ]
+      },
+      {
+        "id": "diagrams",
+        "name": "diagrams",
+        "description": "Create, edit, and verify diagrams, graphs, workflows, org charts, entity networks, control graphs, architecture maps, and evidence-backed visual structures.",
+        "path": "features/perchTerminal/agentPlatform/skills/diagrams/SKILL.md",
+        "body": "\n# diagrams\n\nUse this skill when the user asks for a diagram, graph, workflow, architecture map, network, org chart, or visual relationship model.\n\n## Operating Rules\n\n- Choose diagram format by use: Mermaid for simple structural diagrams, richer rendered artifacts for shareable or visually complex output.\n- Keep charts and diagrams distinct. Charts quantify data; diagrams explain structure, flow, responsibility, or relationships.\n- Tie graph nodes and edges to evidence when the diagram represents real entities, controls, risks, or findings.\n- Use `data-quality` for anomaly and reconciliation networks, and `presentations` or `documents` when the diagram needs to land in a deck or memo.\n\n## Verification\n\nRender or inspect complex diagrams before delivery. Confirm labels are readable, layout is not tangled, and evidence-backed nodes are traceable. See `references/diagram-quality.md`.",
+        "surface": {
+          "gui": "supported",
+          "cli": "supported"
+        },
+        "triggerHints": [
+          "diagram",
+          "mermaid",
+          "drawio",
+          "workflow graph",
+          "org chart",
+          "entity network",
+          "control graph",
+          "architecture map",
+          "vendor-risk graph",
+          "node edge"
+        ],
+        "companionSkills": [
+          "data-quality",
+          "documents",
+          "presentations",
+          "workspace"
+        ],
+        "preferredTools": [
+          "render_ap_control_graph",
+          "createTextArtifact",
+          "writeLocalFile",
+          "run_sandbox_code",
+          "visionInspect"
+        ]
+      },
+      {
+        "id": "documents",
+        "name": "documents",
+        "description": "Create, edit, verify, and deliver professional document artifacts such as DOCX, Google Docs, memos, letters, reports, redlines, and formal written deliverables.",
+        "path": "features/perchTerminal/agentPlatform/skills/documents/SKILL.md",
+        "body": "\n# documents\n\nUse this skill when the user asks Perch to create, revise, review, redline, or deliver a document-shaped artifact.\n\n## Operating Rules\n\n- For long documents, outline the structure before writing the full body.\n- Keep the output as a real deliverable: DOCX, Markdown, Google Doc, or another explicitly requested document target.\n- Preserve source-grounded claims. Load `research`, `pdf`, or `workspace` when citations, page evidence, local files, or source packets drive the content.\n- For Google Docs, use a verified Google Docs shortcut when it can return proof, or hand off to `browser-operator` for live browser editing.\n- Do not claim a header, footer, comment, tracked edit, or page layout feature if the current surface cannot verify it.\n\n## Verification\n\nBefore delivery, check title, body, section order, missing placeholders, citation handoff, and saved file state. Use `references/verification.md` for the compact QA checklist.",
+        "surface": {
+          "gui": "supported",
+          "cli": "supported"
+        },
+        "triggerHints": [
+          "document",
+          "docx",
+          "word file",
+          "google doc",
+          "memo",
+          "letter",
+          "report",
+          "redline",
+          "tracked changes",
+          "policy",
+          "formal artifact"
+        ],
+        "companionSkills": [
+          "research",
+          "pdf",
+          "workspace",
+          "browser-operator"
+        ],
+        "preferredTools": [
+          "createDocumentArtifact",
+          "writeLocalFile",
+          "google_docs_create_verified",
+          "google_docs_create",
+          "browser_execute_task",
+          "readLocalFile",
+          "visionInspect"
+        ]
+      },
+      {
+        "id": "pdf",
+        "name": "pdf",
+        "description": "Extract, summarize, cite, split, merge, create, inspect, and verify PDF content, including text PDFs, scanned PDFs, tables, forms, page citations, and generated reports.",
+        "path": "features/perchTerminal/agentPlatform/skills/pdf/SKILL.md",
+        "body": "\n# pdf\n\nUse this skill when PDF content is the source, destination, or verification target.\n\n## Operating Rules\n\n- Preserve page awareness. Cite exact pages when answering from a PDF.\n- Treat PDF-to-Markdown and OCR as intermediate views, not final truth.\n- Use table-aware extraction or deterministic code when numbers, rows, or column alignment matter.\n- For scanned PDFs, use visual or OCR fallback and report uncertainty when text is unreadable.\n- Do not invent citations, fields, signatures, or form values.\n\n## Verification\n\nCheck that cited pages support the answer, table values are not flattened incorrectly, and generated PDFs render as expected. See `references/page-citations.md`.",
+        "surface": {
+          "gui": "supported",
+          "cli": "supported"
+        },
+        "triggerHints": [
+          "pdf",
+          "annual report pdf",
+          "scanned pdf",
+          "page citation",
+          "pdf table",
+          "split pdf",
+          "merge pdf",
+          "form pdf",
+          "ocr",
+          "pdf report"
+        ],
+        "companionSkills": [
+          "research",
+          "documents",
+          "workspace",
+          "data-quality"
+        ],
+        "preferredTools": [
+          "extractDocument",
+          "extractDocumentSection",
+          "visionInspect",
+          "readLocalFile",
+          "run_sandbox_code",
+          "webFetch"
+        ]
+      },
+      {
+        "id": "presentations",
+        "name": "presentations",
+        "description": "Build executive decks, board decks, sales decks, research summaries, and visual narratives as PPTX or shareable slide artifacts with verified layout and source-backed content.",
+        "path": "features/perchTerminal/agentPlatform/skills/presentations/SKILL.md",
+        "body": "\n# presentations\n\nUse this skill when the user asks for slides, a deck, a board packet, a pitch, or a visual narrative.\n\n## Operating Rules\n\n- Start with audience, decision, and slide outline before building a long deck.\n- Each slide should carry one takeaway; supporting charts, tables, and bullets must serve that takeaway.\n- Source every important number or claim from provided evidence, workspace files, research, or companion skills.\n- Use `spreadsheets` for chart data, `diagrams` for networks and workflows, and `pdf` when source packets are page-cited.\n- Prefer real deck files when requested. HTML slide experiences are acceptable only when design iteration or sharing calls for them.\n\n## Verification\n\nCheck slide count, readable text, non-overflowing layout, chart/image rendering, and file existence before delivery. See `references/deck-quality.md`.",
+        "surface": {
+          "gui": "supported",
+          "cli": "supported"
+        },
+        "triggerHints": [
+          "deck",
+          "pptx",
+          "powerpoint",
+          "board deck",
+          "slide deck",
+          "presentation",
+          "executive deck",
+          "pitch deck",
+          "sales deck",
+          "speaker notes",
+          "chart slide"
+        ],
+        "companionSkills": [
+          "documents",
+          "spreadsheets",
+          "research",
+          "diagrams",
+          "pdf",
+          "workspace"
+        ],
+        "preferredTools": [
+          "writeLocalFile",
+          "run_sandbox_code",
+          "createDocumentArtifact",
+          "visionInspect",
+          "readLocalFile"
+        ]
+      },
+      {
+        "id": "research",
+        "name": "research",
+        "description": "Gather, compare, rank, and cite evidence from public web sources, scholarly sources, workspace files, source packets, Qdrant knowledge, and attached local folders.",
+        "path": "features/perchTerminal/agentPlatform/skills/research/SKILL.md",
+        "body": "\n# research\n\nUse this skill when the user asks Perch to gather, compare, verify, or cite evidence.\n\n## Operating Rules\n\n- Choose the source lane intentionally: current public questions use web; attached folders use workspace or local sources; internal product questions use repo/workspace evidence; stored examples use knowledge.\n- Do not use knowledge retrieval as a catch-all when a domain or workspace tool is more direct.\n- Rank sources by authority, recency, relevance, and proximity to the claim.\n- Surface contradictions instead of averaging them away.\n- Keep citations concise and visibly tied to titles, URLs, files, pages, or evidence IDs.\n\n## Verification\n\nCheck freshness requirements, cite the actual sources used, and label uncertainty when sources are missing or disagree. See `references/source-quality.md`.",
+        "surface": {
+          "gui": "supported",
+          "cli": "supported"
+        },
+        "triggerHints": [
+          "research",
+          "web research",
+          "credible sources",
+          "current source",
+          "citation",
+          "source comparison",
+          "filing",
+          "scholarly source",
+          "public web",
+          "fact check",
+          "latest"
+        ],
+        "companionSkills": [
+          "pdf",
+          "documents",
+          "workspace",
+          "data-quality"
+        ],
+        "preferredTools": [
+          "webSearch",
+          "webFetch",
+          "scholarSearch",
+          "searchKnowledge",
+          "searchWorkspaceSources",
+          "readLocalSourceFile",
+          "listLocalSources"
+        ]
+      },
+      {
+        "id": "skill-creator",
+        "name": "skill-creator",
+        "description": "Create, validate, package, and test Perch skills with frontmatter checks, trigger collision detection, reference validation, surface metadata, and acceptance tasks.",
+        "path": "features/perchTerminal/agentPlatform/skills/skill-creator/SKILL.md",
+        "body": "\n# skill-creator\n\nUse this skill when creating, updating, validating, or evaluating Perch skills.\n\n## Operating Rules\n\n- Keep each skill as one reusable workflow unit with `SKILL.md`, optional `references/`, optional `scripts/`, and optional `assets/`.\n- Frontmatter must include `name`, `description`, and Perch metadata for GUI and CLI surface support.\n- Trigger hints should be precise enough to activate the skill without colliding with unrelated skills.\n- Put detailed reusable rules in reference files and fragile deterministic checks in scripts.\n- Add at least one acceptance task for each supported surface when the skill is productized.\n\n## Verification\n\nRun skill codegen and registry tests after edits. Use `references/validation-checklist.md` before declaring a skill ready.",
+        "surface": {
+          "gui": "supported",
+          "cli": "supported"
+        },
+        "triggerHints": [
+          "skill creator",
+          "create skill",
+          "skill.md",
+          "agent skill",
+          "skill validation",
+          "trigger collision",
+          "skill test",
+          "skill authoring",
+          "perch skill"
+        ],
+        "companionSkills": [
+          "workspace",
+          "data-quality"
+        ],
+        "preferredTools": [
+          "glob",
+          "grep",
+          "readLocalFile",
+          "writeLocalFile",
+          "bash"
+        ]
+      },
+      {
+        "id": "spreadsheets",
+        "name": "spreadsheets",
+        "description": "Create, inspect, clean, calculate, format, and verify spreadsheet artifacts such as XLSX, CSV, TSV, Google Sheets, formulas, pivots, charts, and reconciliation workbooks.",
+        "path": "features/perchTerminal/agentPlatform/skills/spreadsheets/SKILL.md",
+        "body": "\n# spreadsheets\n\nUse this skill when the user asks for a spreadsheet artifact, spreadsheet inspection, import-ready CSV, formulas, reconciliation tabs, exception tabs, or workbook-quality output.\n\n## Operating Rules\n\n- Profile incoming sheets before changing them: row counts, columns, types, key IDs, totals, and empty fields.\n- Preserve identifiers as text when they are invoice numbers, account numbers, vendor IDs, employee IDs, tickers, or other non-math values.\n- Put source values on input tabs and formulas on analysis tabs when producing XLSX workbooks.\n- Add checks for totals, row counts, formulas, and exception counts when the workbook drives a decision.\n- Use deterministic code or workbook inspection for fragile calculations; do not hand-wave formula behavior.\n\n## Verification\n\nBefore delivery, verify that the file exists, formulas are present where expected, totals reconcile, and CSV exports keep exact values. See `references/workbook-quality.md`.",
+        "surface": {
+          "gui": "supported",
+          "cli": "supported"
+        },
+        "triggerHints": [
+          "spreadsheet",
+          "xlsx",
+          "excel workbook",
+          "csv",
+          "tsv",
+          "google sheet",
+          "formula",
+          "pivot",
+          "reconciliation tab",
+          "exception tab",
+          "import template",
+          "row count",
+          "workbook check"
+        ],
+        "companionSkills": [
+          "data-quality",
+          "workspace",
+          "pdf",
+          "research"
+        ],
+        "preferredTools": [
+          "inspectWorkbookSheets",
+          "google_sheets_create",
+          "google_sheets_append_rows",
+          "writeLocalFile",
+          "run_sandbox_code",
+          "bash",
+          "readLocalFile"
+        ]
+      },
+      {
+        "id": "workspace",
+        "name": "workspace",
+        "description": "Navigate, inspect, search, summarize, and safely edit local workspaces, attached folders, source inventories, project files, and codebases with exact file-grounded evidence.",
+        "path": "features/perchTerminal/agentPlatform/skills/workspace/SKILL.md",
+        "body": "\n# workspace\n\nUse this skill when the task requires local files, attached folders, repo search, safe edits, or source inventory.\n\n## Operating Rules\n\n- Start with inventory or targeted search before reading many files.\n- Cite exact files when answering codebase or workspace questions.\n- Distinguish direct filesystem reads, indexed local source retrieval, Supabase source packets, and Qdrant knowledge.\n- Do not say Perch cannot access a file path when the local runtime and permissions make it accessible.\n- For edits, keep changes scoped and preserve unrelated user work.\n\n## Verification\n\nCheck file existence, paths, and write results before reporting completion. Use `references/source-boundaries.md` to keep source lanes clear.",
+        "surface": {
+          "gui": "supported",
+          "cli": "supported"
+        },
+        "triggerHints": [
+          "repo",
+          "codebase",
+          "local folder",
+          "workspace",
+          "find file",
+          "inspect folder",
+          "search files",
+          "edit file",
+          "source inventory",
+          "project files"
+        ],
+        "companionSkills": [
+          "research",
+          "pdf",
+          "data-quality",
+          "skill-creator"
+        ],
+        "preferredTools": [
+          "glob",
+          "grep",
+          "readLocalFile",
+          "writeLocalFile",
+          "statPath",
+          "listLocalSources",
+          "readLocalSourceFile",
+          "bash"
+        ]
+      }
+    ];
+    AGENT_SKILL_INDEX = /* @__PURE__ */ new Map();
+    for (const skill of AGENT_SKILL_REGISTRY) {
+      if (AGENT_SKILL_INDEX.has(skill.id)) {
+        throw new Error(`generated registry duplicate skill id: ${skill.id}`);
+      }
+      AGENT_SKILL_INDEX.set(skill.id, skill);
+    }
+  }
+});
+
+// features/perchTerminal/agentPlatform/skills/skillSelector.ts
+function normalizeText(input) {
+  return input.toLowerCase().replace(/\s+/g, " ").trim();
+}
+function escapeRegExp2(input) {
+  return input.replace(/[.*+?^${}()|[\]\\]/g, "\\$&");
+}
+function triggerMatches(text, trigger) {
+  const normalized = normalizeText(trigger);
+  if (!normalized) return false;
+  return new RegExp(`(^|[^a-z0-9])${escapeRegExp2(normalized)}([^a-z0-9]|$)`, "i").test(text);
+}
+function skillMatches(skill, text) {
+  if (skill.triggerHints.some((trigger) => triggerMatches(text, trigger))) {
+    return true;
+  }
+  return (INTENT_PATTERNS[skill.id] ?? []).some(
+    (pattern) => pattern.test(text)
+  );
+}
+function isMarketDeskSignalTask(text, availableToolNames) {
+  const hasMarketDeskTool = [...MARKET_DESK_TOOL_NAMES].some(
+    (toolName) => availableToolNames.has(toolName)
+  );
+  return hasMarketDeskTool && /\b(market\s+desk|signal|strategy|backtest|track\s+record)\b/i.test(text) && /\b[A-Z]{1,5}\b/.test(text);
+}
+function selectionText(input) {
+  return [
+    input.userMessage,
+    input.currentMode ?? "",
+    input.personaId ?? "",
+    input.selectedFolder ?? "",
+    ...input.selectedFiles ?? [],
+    input.sourceSummary ?? ""
+  ].filter(Boolean).join("\n");
+}
+function canUseSkillOnSurface(skill, input, availableToolNames) {
+  if (skill.surface[input.surface] === "unsupported") {
+    return `${input.surface.toUpperCase()} surface is unsupported`;
+  }
+  if (skill.id === "browser-operator" && ![...availableToolNames].some((toolName) => BROWSER_TOOL_NAMES.has(toolName))) {
+    return "browser operator tools are not available this turn";
+  }
+  return null;
+}
+function priorityIndex(skillId) {
+  const index = CORE_PRIORITY.indexOf(skillId);
+  return index === -1 ? CORE_PRIORITY.length : index;
+}
+function sortSkillIds(ids) {
+  return [...ids].sort(
+    (a, b2) => priorityIndex(a) - priorityIndex(b2) || a.localeCompare(b2)
+  );
+}
+function selectAgentSkills(input) {
+  const maxSelected = Math.max(
+    0,
+    input.maxSelectedSkills ?? DEFAULT_MAX_SELECTED_SKILLS
+  );
+  const availableToolNames = new Set(input.availableToolNames ?? []);
+  const text = selectionText(input);
+  const normalizedText = normalizeText(text);
+  const byId = new Map(AGENT_SKILL_REGISTRY.map((skill) => [skill.id, skill]));
+  const metadataRows = AGENT_SKILL_REGISTRY.map((skill) => ({
+    id: skill.id,
+    name: skill.name,
+    description: skill.description
+  }));
+  const omitted = [];
+  const explicit = /* @__PURE__ */ new Set();
+  for (const skill of AGENT_SKILL_REGISTRY) {
+    if (skillMatches(skill, normalizedText)) explicit.add(skill.id);
+  }
+  if (/\bgoogle\s+docs?\b/i.test(text) && availableToolNames.has("browser_execute_task")) {
+    explicit.add("documents");
+    explicit.add("browser-operator");
+  }
+  if (/\b(this|attached|local)\s+folder\b/i.test(text) || Boolean(input.selectedFolder) && /\bfolder\b/i.test(text)) {
+    explicit.add("workspace");
+  }
+  if (/\b(ap|accounts?\s+payable)\b/i.test(text)) {
+    explicit.add("data-quality");
+  }
+  if (isMarketDeskSignalTask(text, availableToolNames)) {
+    explicit.delete("research");
+    omitted.push({
+      id: "research",
+      reason: "market-desk signal tasks should use Market Desk tools before public research"
+    });
+  }
+  const companionExpanded = /* @__PURE__ */ new Set();
+  for (const id of sortSkillIds(explicit)) {
+    const skill = byId.get(id);
+    if (!skill) continue;
+    for (const companionId of skill.companionSkills) {
+      if (!explicit.has(companionId)) companionExpanded.add(companionId);
+    }
+  }
+  const selected = [];
+  for (const id of [
+    ...sortSkillIds(explicit),
+    ...sortSkillIds(companionExpanded)
+  ]) {
+    const skill = byId.get(id);
+    if (!skill) continue;
+    const omittedReason = canUseSkillOnSurface(skill, input, availableToolNames);
+    if (omittedReason) {
+      omitted.push({ id, reason: omittedReason });
+      continue;
+    }
+    if (selected.length >= maxSelected) {
+      omitted.push({ id, reason: `max selected skills reached (${maxSelected})` });
+      continue;
+    }
+    selected.push(skill);
+  }
+  return { metadataRows, selected, omitted };
+}
+var DEFAULT_MAX_SELECTED_SKILLS, BROWSER_TOOL_NAMES, CORE_PRIORITY, INTENT_PATTERNS;
+var init_skillSelector = __esm({
+  "features/perchTerminal/agentPlatform/skills/skillSelector.ts"() {
+    init_agentSkillRegistry();
+    init_marketDeskAccess();
+    DEFAULT_MAX_SELECTED_SKILLS = 4;
+    BROWSER_TOOL_NAMES = /* @__PURE__ */ new Set([
+      "browser_execute_task",
+      "browser_observe",
+      "browser_click",
+      "browser_type",
+      "browser_visual_click",
+      "browser_focus_type",
+      "browser_press_key",
+      "browser_read",
+      "browser_wait",
+      "browser_navigate"
+    ]);
+    CORE_PRIORITY = [
+      "data-quality",
+      "presentations",
+      "spreadsheets",
+      "documents",
+      "pdf",
+      "research",
+      "diagrams",
+      "workspace",
+      "browser-operator",
+      "skill-creator"
+    ];
+    INTENT_PATTERNS = {
+      "browser-operator": [
+        /\b(browser|web\s+app|logged[-\s]?in|saas|portal|form)\b/i,
+        /\b(gmail|google\s+docs?|google\s+calendar|notion|salesforce)\b/i
+      ],
+      documents: [
+        /\b(docx|word\s+file|google\s+docs?|memo|letter|report|redline|tracked\s+changes?|policy|formal\s+artifact)\b/i,
+        /\b(create|write|draft|revise|edit|publish|deliver)\b.{0,48}\b(document|doc|memo|report|letter)\b/i
+      ],
+      spreadsheets: [
+        /\b(spreadsheet|xlsx|excel|workbook|csv|tsv|google\s+sheet|formula|pivot|reconciliation\s+tab|exception\s+tab|import\s+template)\b/i,
+        /\b(row\s+count|totals?|tie[-\s]?out|sheet)\b/i
+      ],
+      presentations: [
+        /\b(deck|pptx|powerpoint|slides?|presentation|board\s+deck|executive\s+deck|pitch\s+deck|sales\s+deck|speaker\s+notes?)\b/i
+      ],
+      pdf: [
+        /\b(pdf|scanned\s+pdf|page\s+citation|pdf\s+table|split\s+pdf|merge\s+pdf|ocr|annual\s+report)\b/i,
+        /\.pdf\b/i
+      ],
+      research: [
+        /\b(research|credible\s+sources?|current\s+source|citations?|source\s+comparison|scholarly|public\s+web|fact[-\s]?check|latest|recent|filing)\b/i,
+        /\b(find|gather|compare|verify)\b.{0,48}\b(sources?|evidence|filings?)\b/i
+      ],
+      diagrams: [
+        /\b(diagram|mermaid|drawio|workflow\s+graph|org\s+chart|entity\s+network|control\s+graph|architecture\s+map|vendor[-\s]?risk\s+graph|nodes?\s+and\s+edges?)\b/i,
+        /\b(graph|map)\b.{0,32}\b(risk|workflow|architecture|network|relationship)\b/i
+      ],
+      workspace: [
+        /\b(repo|codebase|local\s+folder|workspace|find\s+file|inspect\s+folder|search\s+files?|edit\s+file|source\s+inventory|project\s+files?)\b/i,
+        /\b(this\s+repo|this\s+folder|attached\s+folder|in\s+this\s+workspace)\b/i
+      ],
+      "data-quality": [
+        /\b(data\s+quality|anomal(?:y|ies)|reconciliation|duplicate\s+(?:payment|invoice)|tie[-\s]?out|row\s+totals?|audit\s+exceptions?|vendor\s+risk|outliers?|missing\s+receipt|cost\s+anomal(?:y|ies))\b/i,
+        /\b(audit|reconcile|profile|clean)\b.{0,56}\b(ap|accounts?\s+payable|payments?|invoices?|vendors?|data|folder)\b/i
+      ],
+      "skill-creator": [
+        /\b(skill\s+creator|create\s+(?:a\s+)?skill|skill\.md|agent\s+skill|skill\s+validation|trigger\s+collision|skill\s+test|skill\s+authoring|perch\s+skill)\b/i
+      ]
+    };
+  }
+});
+
 // features/perchTerminal/runtime/context/tokenEstimates.ts
 function estimateTokensForLane(text, lane, modelTokenize) {
   return countTokensForLane(
@@ -92237,6 +92837,26 @@ function assembleContext(input) {
   const contextCompaction = input.threadId ? getThreadSessionFromMemory(input.threadId)?.contextCompaction ?? null : null;
   const threadSession = input.threadId ? getThreadSessionFromMemory(input.threadId) : null;
   const coreSystem = buildCoreSystemSection(input);
+  const selectedSkillPack = selectAgentSkills({
+    userMessage: input.trimmedInput,
+    currentMode: input.chatMode,
+    personaId: input.personaId ?? null,
+    selectedFiles: [
+      ...input.workspaceFileResolution?.matchedFiles.map(
+        (file) => file.relativePath
+      ) ?? [],
+      ...(input.localSources ?? []).slice(0, 12).map((source) => source.relativePath || source.fileName)
+    ],
+    selectedFolder: input.activeRootPath,
+    sourceSummary: [
+      input.selectedSourceId ? `selected source ${input.selectedSourceId}` : "",
+      input.folderIndexSummary ? `folder indexed ${input.folderIndexSummary.totalFiles} files` : "",
+      input.selectedSourceBundle?.fileName ?? ""
+    ].filter(Boolean).join("\n"),
+    availableToolNames: input.enabledToolNames ?? [],
+    surface: input.skillSelectionSurface ?? (input.cliLocalTools === true && !input.desktopConnected ? "cli" : "gui")
+  });
+  const selectedSkillsSection = buildSelectedSkillsSection(selectedSkillPack);
   const selectedModelSection = buildSelectedModelSection(
     input.selectedModel ?? null
   );
@@ -92300,6 +92920,7 @@ function assembleContext(input) {
       content: coreSystem,
       reason: "Core operator instructions for the current turn."
     },
+    selectedSkillsSection,
     ...toolsJson ? [
       {
         id: "tool-schemas",
@@ -92401,6 +93022,7 @@ function assembleContext(input) {
   });
   const systemPrompt = systemSections.filter((row) => row.trim()).join("\n\n");
   const sectionBreakdown = buildSectionBreakdown(rows);
+  const selectedSkillIds = selectedSkillPack.selected.map((skill) => skill.id);
   const totalContextTokens = rows.filter((row) => row.status !== "reserved").reduce((sum, row) => sum + row.estimatedTokens, 0);
   const compactedCount = rows.filter(
     (row) => row.status === "sent_compacted"
@@ -92439,7 +93061,9 @@ function assembleContext(input) {
           activeRootPath: input.activeRootPath,
           recentMessageCount: input.recentMessages.length,
           selectedModel: input.selectedModel ?? null,
-          untrustedContextPresent
+          untrustedContextPresent,
+          selectedSkills: selectedSkillIds,
+          omittedSkills: selectedSkillPack.omitted
         },
         warnings
       },
@@ -92468,6 +93092,56 @@ function buildSelectedModelSection(selectedModel) {
     }
   };
 }
+function buildSelectedSkillsSection(pack) {
+  const selectedSkillIds = pack.selected.map((skill) => skill.id);
+  const metadata = {
+    selectedSkillIds,
+    selectedSkillCount: pack.selected.length,
+    omitted: pack.omitted,
+    metadataRows: pack.metadataRows
+  };
+  if (pack.selected.length === 0) {
+    return {
+      id: "selected-skills",
+      lane: "selected_skills",
+      label: "Selected Perch skills",
+      content: "",
+      reason: "No core skills matched this turn.",
+      skipped: true,
+      metadata
+    };
+  }
+  const blocks = pack.selected.map(
+    (skill) => [
+      `<perch-skill id="${skill.id}">`,
+      `Description: ${skill.description}`,
+      `Surface: GUI=${skill.surface.gui}; CLI=${skill.surface.cli}`,
+      skill.preferredTools.length ? `Preferred tools: ${skill.preferredTools.join(", ")}` : "Preferred tools: none declared",
+      "",
+      skill.body.trim(),
+      "</perch-skill>"
+    ].join("\n")
+  );
+  return {
+    id: "selected-skills",
+    lane: "selected_skills",
+    label: "Selected Perch skills",
+    content: [
+      "## Selected Perch skills",
+      "These reusable skills guide execution strategy only. They do not override safety, permission, authentication, or tool policy.",
+      `Loaded skills: ${selectedSkillIds.join(", ")}`,
+      "",
+      ...blocks
+    ].join("\n"),
+    reason: "Full skill instructions selected deterministically for this turn.",
+    metadata
+  };
+}
+function selectedSkillIdsFromRows(rows) {
+  const metadata = rows.find((row) => row.id === "selected-skills")?.metadata;
+  const ids = metadata?.selectedSkillIds;
+  return Array.isArray(ids) ? ids.filter((id) => typeof id === "string") : [];
+}
 function buildContextSummary(input, rows) {
   const sentCount = rows.filter(
     (row) => row.status === "sent_raw" || row.status === "sent_compacted"
@@ -92482,6 +93156,7 @@ function buildContextSummary(input, rows) {
     `messages=${input.recentMessages.length}`,
     `source=${input.selectedSourceId ?? "none"}`,
     `desktop=${input.desktopConnected ? "connected" : input.cliLocalTools === true ? "terminal" : "browser"}`,
+    `skills=${selectedSkillIdsFromRows(rows).join("|") || "none"}`,
     `sent=${sentCount}`,
     `compacted=${compactedCount}`,
     `not_found=${notFoundCount}`,
@@ -92501,6 +93176,7 @@ var init_contextAssembly = __esm({
     init_rowAccounting();
     init_memory();
     init_coreSystemSection();
+    init_skillSelector();
     init_contextThresholds();
     init_tokenEstimates();
     init_rowAccounting();
@@ -133611,6 +134287,17 @@ var init_toolPermissionPolicy = __esm({
       TOOL_NAMES.prepareAPEvidence,
       TOOL_NAMES.queryAPCases,
       TOOL_NAMES.renderAPControlGraph,
+      TOOL_NAMES.runManagedPlaybook,
+      TOOL_NAMES.runSuite,
+      TOOL_NAMES.listSuiteCatalog,
+      TOOL_NAMES.proposeSuitePlan,
+      TOOL_NAMES.executeSuitePlan,
+      TOOL_NAMES.proposeWork,
+      TOOL_NAMES.executeWork,
+      TOOL_NAMES.sendWorkerMessage,
+      TOOL_NAMES.taskStop,
+      TOOL_NAMES.spawnWorker,
+      TOOL_NAMES.dispatchAgent,
       TOOL_NAMES.runSandboxCode
     ];
     CLI_LOCAL_TOOL_NAME_SET = new Set(CLI_LOCAL_TOOL_NAMES);
@@ -168151,7 +168838,7 @@ function template(text, settings, oldSettings) {
   var index = 0;
   var source = "__p+='";
   text.replace(matcher2, function(match, escape4, interpolate, evaluate, offset) {
-    source += text.slice(index, offset).replace(escapeRegExp2, escapeChar);
+    source += text.slice(index, offset).replace(escapeRegExp3, escapeChar);
     index = offset + match.length;
     if (escape4) {
       source += "'+\n((__t=(" + escape4 + "))==null?'':_.escape(__t))+\n'";
@@ -168186,7 +168873,7 @@ function template(text, settings, oldSettings) {
   template2.source = "function(" + argument + "){\n" + source + "}";
   return template2;
 }
-var noMatch, escapes, escapeRegExp2, bareIdentifier;
+var noMatch, escapes, escapeRegExp3, bareIdentifier;
 var init_template = __esm({
   "node_modules/underscore/modules/template.js"() {
     init_defaults();
@@ -168201,7 +168888,7 @@ var init_template = __esm({
       "\u2028": "u2028",
       "\u2029": "u2029"
     };
-    escapeRegExp2 = /\\|'|\r|\n|\u2028|\u2029/g;
+    escapeRegExp3 = /\\|'|\r|\n|\u2028|\u2029/g;
     bareIdentifier = /^\s*(\w|\$)+\s*$/;
   }
 });
@@ -199949,7 +200636,6 @@ function containsBrowserDeliveryTask(tasks) {
 var BROWSER_DELIVERY_ROLE_IDS;
 var init_browserDeliveryLock = __esm({
   "features/perchTerminal/agentPlatform/browserDeliveryLock.ts"() {
-    "use strict";
     BROWSER_DELIVERY_ROLE_IDS = /* @__PURE__ */ new Set([
       "doc_writer",
       "email_sender",
@@ -201862,7 +202548,7 @@ function lifecycleState(ctx) {
   }
   return state;
 }
-function normalizeText(value) {
+function normalizeText2(value) {
   if (typeof value === "string") return value.toLowerCase().replace(/\s+/g, " ").trim();
   if (value === null || value === void 0) return "";
   try {
@@ -201873,7 +202559,7 @@ function normalizeText(value) {
 }
 function explicitRetryRequested(tasks) {
   const text = tasks.map((task) => `${task.objective}
-${normalizeText(task.context)}`).join("\n");
+${normalizeText2(task.context)}`).join("\n");
   return /\b(retry|re-run|rerun|run again|fresh pass|fresh read|repeat|redo)\b/i.test(text);
 }
 function isSourceReaderLifecycleRole(roleId) {
@@ -201884,7 +202570,7 @@ function isSourceReaderLifecycleRole(roleId) {
 }
 function extractSourceScopes(task) {
   const text = `${task.objective}
-${normalizeText(task.context)}`;
+${normalizeText2(task.context)}`;
   const scopes = /* @__PURE__ */ new Set();
   const pathPattern = /\b(?:\d{2}-[a-z0-9_-]+|[a-z0-9_-]+\/[a-z0-9_ .&/-]+)(?:\/[a-z0-9_ .&-]+)*/gi;
   for (const match of text.matchAll(pathPattern)) {
@@ -201894,14 +202580,14 @@ ${normalizeText(task.context)}`;
     }
     if (raw.length >= 4) scopes.add(raw.toLowerCase());
   }
-  if (scopes.size === 0) scopes.add(normalizeText(task.objective).slice(0, 120));
+  if (scopes.size === 0) scopes.add(normalizeText2(task.objective).slice(0, 120));
   return [...scopes].sort();
 }
 function taskSignature(task) {
   return JSON.stringify({
     roleId: task.roleId,
-    objective: normalizeText(task.objective),
-    context: normalizeText(task.context)
+    objective: normalizeText2(task.objective),
+    context: normalizeText2(task.context)
   });
 }
 function batchSignature(tasks) {
@@ -217660,7 +218346,7 @@ function isBrowserSnapshotToolResult(messages, index) {
   const toolName = toolNameForMessage(messages, index);
   const text = messageContentText(msg);
   if (!text.includes("snapshot")) return false;
-  if (toolName && BROWSER_TOOL_NAMES.has(toolName)) return true;
+  if (toolName && BROWSER_TOOL_NAMES2.has(toolName)) return true;
   try {
     const parsed = JSON.parse(text);
     return typeof parsed.snapshot === "string" && parsed.snapshot.length > 200;
@@ -218135,7 +218821,7 @@ async function prepareLoopMessagesForSend(input) {
     microcompactedToolResults
   };
 }
-var OPERATOR_SCREENSHOT_MARKER2, KEPT_SNAPSHOT_MAX_CHARS, KEPT_SNAPSHOT_TRUNCATED_MARKER, BROWSER_TOOL_NAMES, RECENT_MESSAGES_KEEP_COUNT, RECENT_TOOL_RESULTS_KEEP_COUNT, POST_COMPACT_TARGET_TOKENS, POST_COMPACT_MIN_TARGET_TOKENS, POST_COMPACT_MAX_TARGET_TOKENS, POST_COMPACT_RECENT_TAIL_MIN_TOKENS, POST_COMPACT_RECENT_TAIL_MAX_TOKENS, MICROCOMPACT_CLEARED_MARKER, COMPACT_SUMMARY_RESERVE_TOKENS, COMPACT_SYSTEM_PROMPT;
+var OPERATOR_SCREENSHOT_MARKER2, KEPT_SNAPSHOT_MAX_CHARS, KEPT_SNAPSHOT_TRUNCATED_MARKER, BROWSER_TOOL_NAMES2, RECENT_MESSAGES_KEEP_COUNT, RECENT_TOOL_RESULTS_KEEP_COUNT, POST_COMPACT_TARGET_TOKENS, POST_COMPACT_MIN_TARGET_TOKENS, POST_COMPACT_MAX_TARGET_TOKENS, POST_COMPACT_RECENT_TAIL_MIN_TOKENS, POST_COMPACT_RECENT_TAIL_MAX_TOKENS, MICROCOMPACT_CLEARED_MARKER, COMPACT_SUMMARY_RESERVE_TOKENS, COMPACT_SYSTEM_PROMPT;
 var init_contextLoopCompaction = __esm({
   "features/perchTerminal/runtime/services/contextLoopCompaction.ts"() {
     "use strict";
@@ -218150,7 +218836,7 @@ var init_contextLoopCompaction = __esm({
     OPERATOR_SCREENSHOT_MARKER2 = "[operator-screen]";
     KEPT_SNAPSHOT_MAX_CHARS = 12e3;
     KEPT_SNAPSHOT_TRUNCATED_MARKER = "[snapshot truncated]";
-    BROWSER_TOOL_NAMES = /* @__PURE__ */ new Set([
+    BROWSER_TOOL_NAMES2 = /* @__PURE__ */ new Set([
       TOOL_NAMES.browserNavigate,
       TOOL_NAMES.browserObserve,
       TOOL_NAMES.browserClick,
@@ -221650,7 +222336,7 @@ var init_preTurnRetrieval = __esm({
 // features/perchTerminal/rag/chunking.ts
 import { createHash as createHash2 } from "crypto";
 function buildRetrievalChunks(text, metadata, options = {}) {
-  const normalized = normalizeText2(text);
+  const normalized = normalizeText3(text);
   if (!normalized) return [];
   const targetChars = Math.max(400, (options.targetTokens ?? DEFAULT_TARGET_TOKENS) * APPROX_CHARS_PER_TOKEN);
   const overlapChars = Math.max(0, Math.min(targetChars / 2, (options.overlapTokens ?? DEFAULT_OVERLAP_TOKENS) * APPROX_CHARS_PER_TOKEN));
@@ -221680,7 +222366,7 @@ function buildRetrievalChunks(text, metadata, options = {}) {
     return chunk2;
   });
 }
-function normalizeText2(text) {
+function normalizeText3(text) {
   return text.replace(/\r\n/g, "\n").replace(/[ \t]+/g, " ").replace(/\n[ \t]+/g, "\n").replace(/\n{3,}/g, "\n\n").trim();
 }
 function stableContentHash(value) {
@@ -224323,6 +225009,8 @@ function buildClientContextSnapshot(input) {
     compacted: context.snapshot.compacted,
     chatMode: input.input.chatMode,
     selectedSourceId: input.input.selectedSourceId,
+    selectedSkills: extractSelectedSkills(context),
+    omittedSkills: extractOmittedSkills(context),
     desktopConnected: input.input.desktopConnected,
     cliLocalTools: input.input.cliLocalTools === true,
     sectionBreakdown: context.snapshot.sectionBreakdown,
@@ -224361,6 +225049,23 @@ function buildClientContextSnapshot(input) {
     } : {}
   };
 }
+function extractSelectedSkills(context) {
+  const metadata = context.snapshot.rows.find(
+    (row) => row.id === "selected-skills"
+  )?.metadata;
+  const ids = metadata?.selectedSkillIds;
+  return Array.isArray(ids) ? ids.filter((id) => typeof id === "string") : [];
+}
+function extractOmittedSkills(context) {
+  const metadata = context.snapshot.rows.find(
+    (row) => row.id === "selected-skills"
+  )?.metadata;
+  const omitted = metadata?.omitted;
+  if (!Array.isArray(omitted)) return [];
+  return omitted.filter(
+    (item) => typeof item === "object" && item !== null && typeof item.id === "string" && typeof item.reason === "string"
+  );
+}
 var init_turnContextSnapshot = __esm({
   "features/perchTerminal/runtime/turn/turnContextSnapshot.ts"() {
     "use strict";
@@ -259813,7 +260518,7 @@ var require_websocket = __commonJS({
     var http2 = __require("http");
     var net = __require("net");
     var tls = __require("tls");
-    var { randomBytes, createHash: createHash4 } = __require("crypto");
+    var { randomBytes, createHash: createHash5 } = __require("crypto");
     var { Duplex, Readable } = __require("stream");
     var { URL: URL2 } = __require("url");
     var PerMessageDeflate2 = require_permessage_deflate();
@@ -260481,7 +261186,7 @@ var require_websocket = __commonJS({
           abortHandshake(websocket, socket, "Invalid Upgrade header");
           return;
         }
-        const digest = createHash4("sha1").update(key + GUID).digest("base64");
+        const digest = createHash5("sha1").update(key + GUID).digest("base64");
         if (res.headers["sec-websocket-accept"] !== digest) {
           abortHandshake(websocket, socket, "Invalid Sec-WebSocket-Accept header");
           return;
@@ -260850,7 +261555,7 @@ var require_websocket_server = __commonJS({
     var EventEmitter3 = __require("events");
     var http2 = __require("http");
     var { Duplex } = __require("stream");
-    var { createHash: createHash4 } = __require("crypto");
+    var { createHash: createHash5 } = __require("crypto");
     var extension2 = require_extension();
     var PerMessageDeflate2 = require_permessage_deflate();
     var subprotocol2 = require_subprotocol();
@@ -261157,7 +261862,7 @@ var require_websocket_server = __commonJS({
           );
         }
         if (this._state > RUNNING) return abortHandshake(socket, 503);
-        const digest = createHash4("sha1").update(key + GUID).digest("base64");
+        const digest = createHash5("sha1").update(key + GUID).digest("base64");
         const headers = [
           "HTTP/1.1 101 Switching Protocols",
           "Upgrade: websocket",
@@ -283093,6 +283798,7 @@ __export(perch_cli_exports, {
   runPerchCli: () => runPerchCli
 });
 import fs14 from "node:fs";
+import { createHash as createHash4 } from "node:crypto";
 import os5 from "node:os";
 import path14 from "node:path";
 import readline from "node:readline/promises";
@@ -283118,12 +283824,17 @@ ${HELP_TEXT}`);
     if (parsed.domain === "auth") {
       return runAuthCommand(parsed, writer);
     }
+    if (parsed.domain === "skills") {
+      writer.stdout(renderCliSkills(parsed.name));
+      return 0;
+    }
     if (parsed.domain === "run") {
       const connectModelProxy = deps.connectModelProxy ?? connectCliModelProxy;
       const connection = await connectModelProxy({ appUrl: parsed.appUrl });
       try {
         const threadId = parsed.threadId?.trim() || "cli-default";
-        const persisted = await hydrateCliThreadState(threadId);
+        const threadScopeKey = await resolveCliThreadScopeKey(connection);
+        const persisted = await hydrateCliThreadState(threadId, threadScopeKey);
         const hostedContext = await fetchCliHostedContext(connection, {
           query: parsed.prompt,
           threadId
@@ -283165,6 +283876,7 @@ ${HELP_TEXT}`);
         });
         await persistCliThreadState({
           threadId: result3.threadId,
+          threadScopeKey,
           recentMessages: nextRecentMessages,
           contextSnapshot: result3.contextSnapshot ?? persisted?.contextSnapshot ?? null
         });
@@ -283222,6 +283934,7 @@ function parsePerchCli(argv) {
   if (domain === "login") return { ok: true, domain: "auth", action: "login", ...global2.appUrl ? { appUrl: global2.appUrl } : {} };
   if (domain === "logout") return { ok: true, domain: "auth", action: "logout" };
   if (domain === "status") return { ok: true, domain: "auth", action: "status" };
+  if (domain === "skills") return { ok: true, domain: "skills", ...action ? { name: action } : {} };
   if (domain !== "ap" && domain !== "test") return { ok: false, error: `Unknown domain: ${domain ?? "(missing)"}` };
   if (domain === "test") {
     if (action !== "ap") return { ok: false, error: `Unknown test command: ${action ?? "(missing)"}` };
@@ -283439,14 +284152,15 @@ async function runReadlineInteractivePerchCli(writer, deps, options) {
     terminal: true
   });
   const state = createInteractiveCliState(options);
-  await hydrateInteractiveCliState(state);
   const connectModelProxy = deps.connectModelProxy ?? connectCliModelProxy;
   let connection = await connectModelProxy({ appUrl: state.appUrl });
   if (!state.appUrl && connection.appUrl) state.appUrl = connection.appUrl;
+  await syncInteractiveCliThreadScope(state, connection, { force: true });
   const reconnect = async () => {
     connection.restore();
     connection = await connectModelProxy({ appUrl: state.appUrl });
     if (!state.appUrl && connection.appUrl) state.appUrl = connection.appUrl;
+    await syncInteractiveCliThreadScope(state, connection);
   };
   writer.stdout(renderInteractiveStartup(state, connection));
   try {
@@ -283532,14 +284246,15 @@ async function runInkInteractivePerchCli(writer, deps, options) {
       init_build2().then(() => build_exports)
     ]);
     const state = createInteractiveCliState(options);
-    await hydrateInteractiveCliState(state);
     const connectModelProxy = deps.connectModelProxy ?? connectCliModelProxy;
     let connection = await connectModelProxy({ appUrl: state.appUrl });
     if (!state.appUrl && connection.appUrl) state.appUrl = connection.appUrl;
+    await syncInteractiveCliThreadScope(state, connection, { force: true });
     const reconnect = async () => {
       connection.restore();
       connection = await connectModelProxy({ appUrl: state.appUrl });
       if (!state.appUrl && connection.appUrl) state.appUrl = connection.appUrl;
+      await syncInteractiveCliThreadScope(state, connection);
     };
     const runTurn = deps.runCliTurn ?? runPerchCliTurn;
     const instance = Ink2.render(
@@ -284254,6 +284969,9 @@ async function runInteractiveSlashCommand(input) {
       ));
       return "continue";
     }
+    case "skills":
+      input.writer.stdout(renderCliSkills(parsed.value));
+      return "continue";
     case "context":
       input.writer.stdout(renderCliContextDetails(input.state.contextSnapshot));
       return "continue";
@@ -284340,7 +285058,7 @@ async function runInteractiveSlashCommand(input) {
       input.state.contextSnapshot = null;
       input.state.persistedThreadUpdatedAt = null;
       clearThreadSession(input.state.threadId);
-      await deleteCliThreadState(input.state.threadId);
+      await deleteCliThreadState(input.state.threadId, input.state.threadScopeKey);
       writeModeLine(input.writer, "memory", "cleared");
       return "continue";
     case "login":
@@ -284373,6 +285091,8 @@ function parseInteractiveSlashCommand(input) {
       return { ok: true, kind: "help" };
     case "status":
       return { ok: true, kind: "status" };
+    case "skills":
+      return { ok: true, kind: "skills", ...value ? { value } : {} };
     case "context":
       return { ok: true, kind: "context" };
     case "memory-audit":
@@ -284428,6 +285148,7 @@ function renderInteractiveStatus(state, connection, session, workspaceId) {
     ["connection", connectionStatus],
     ["memory", renderCliMemoryAvailability(state.cwd, workspaceId ?? null, signedIn)],
     ["tools", renderCliCapabilityCount(state, workspaceId ?? null)],
+    ["skills", renderCliLoadedSkills(state.contextSnapshot)],
     ["permission", state.permissionMode],
     ["mode", state.chatMode],
     ["persona", state.personaId],
@@ -284458,6 +285179,43 @@ function renderCliCapabilityCount(state, workspaceId) {
   }).length;
   return `${count} visible`;
 }
+function renderCliLoadedSkills(snapshot) {
+  const skills = snapshot?.selectedSkills ?? [];
+  return skills.length > 0 ? skills.join(", ") : "none loaded";
+}
+function renderCliSkills(name) {
+  const color = shouldUseCliColor();
+  if (name) {
+    const skill = getAgentSkillManifest(name);
+    if (!skill) {
+      return `Unknown skill: ${name}
+Run /skills to list available skills.
+`;
+    }
+    return [
+      `${paint(skill.name, "accent", color)}`,
+      skill.description,
+      `surface: GUI=${skill.surface.gui}; CLI=${skill.surface.cli}`,
+      skill.triggerHints.length ? `triggers: ${skill.triggerHints.join(", ")}` : "triggers: none",
+      skill.companionSkills.length ? `companions: ${skill.companionSkills.join(", ")}` : "companions: none",
+      "",
+      skill.body.trim(),
+      ""
+    ].join("\n");
+  }
+  const rows = listAgentSkillMetadataRows();
+  return [
+    "Perch core skills",
+    ...rows.map(
+      (skill) => [
+        `- ${skill.id}`,
+        `  ${skill.description}`,
+        `  surface: GUI=${skill.surface.gui}; CLI=${skill.surface.cli}`
+      ].join("\n")
+    ),
+    ""
+  ].join("\n");
+}
 function renderInteractiveStartup(state, connection) {
   const color = shouldUseCliColor();
   const auth = renderCliAuthSummary(connection);
@@ -284539,6 +285297,7 @@ function createInteractiveCliState(options) {
     personaId: options.personaId ?? "saffron",
     permissionMode: options.permissionMode ?? "default",
     threadId: options.threadId?.trim() || "cli-default",
+    threadScopeKey: null,
     desktopConnected: options.desktopConnected ?? false,
     cliLocalTools: options.cliLocalTools ?? true,
     appUrl: resolveCliAppUrl(options.appUrl ?? null, null),
@@ -284548,14 +285307,25 @@ function createInteractiveCliState(options) {
   };
 }
 async function hydrateInteractiveCliState(state) {
-  const persisted = await hydrateCliThreadState(state.threadId);
+  const persisted = await hydrateCliThreadState(state.threadId, state.threadScopeKey);
   if (!persisted) return;
   state.recentMessages = persisted.recentMessages;
   state.contextSnapshot = persisted.contextSnapshot;
   state.persistedThreadUpdatedAt = persisted.updatedAt;
 }
-async function hydrateCliThreadState(threadId) {
-  const persisted = await readCliThreadState(threadId);
+async function syncInteractiveCliThreadScope(state, connection, options = {}) {
+  const nextScopeKey = await resolveCliThreadScopeKey(connection);
+  if (!options.force && state.threadScopeKey === nextScopeKey) return;
+  state.threadScopeKey = nextScopeKey;
+  state.recentMessages = [];
+  state.contextSnapshot = null;
+  state.persistedThreadUpdatedAt = null;
+  clearThreadSession(state.threadId);
+  await hydrateInteractiveCliState(state);
+}
+async function hydrateCliThreadState(threadId, threadScopeKey) {
+  const persisted = await readCliThreadState(threadId, threadScopeKey);
+  clearThreadSession(threadId);
   if (!persisted) return null;
   if (persisted.threadSession) {
     await saveThreadSession(persisted.threadSession);
@@ -284565,6 +285335,7 @@ async function hydrateCliThreadState(threadId) {
 async function persistInteractiveCliState(state) {
   await persistCliThreadState({
     threadId: state.threadId,
+    threadScopeKey: state.threadScopeKey,
     recentMessages: state.recentMessages,
     contextSnapshot: state.contextSnapshot
   });
@@ -284580,13 +285351,13 @@ async function persistCliThreadState(input) {
     threadSession,
     updatedAt: (/* @__PURE__ */ new Date()).toISOString()
   };
-  const filePath = cliThreadStatePath(input.threadId);
+  const filePath = cliThreadStatePath(input.threadId, input.threadScopeKey);
   await fs14.promises.mkdir(path14.dirname(filePath), { recursive: true, mode: 448 });
   await fs14.promises.writeFile(filePath, JSON.stringify(payload, null, 2), { mode: 384 });
 }
-async function readCliThreadState(threadId) {
+async function readCliThreadState(threadId, threadScopeKey) {
   try {
-    const raw = await fs14.promises.readFile(cliThreadStatePath(threadId), "utf8");
+    const raw = await fs14.promises.readFile(cliThreadStatePath(threadId, threadScopeKey), "utf8");
     const parsed = JSON.parse(raw);
     if (parsed.version !== 1 || parsed.threadId !== threadId) return null;
     return {
@@ -284601,12 +285372,28 @@ async function readCliThreadState(threadId) {
     return null;
   }
 }
-async function deleteCliThreadState(threadId) {
-  await fs14.promises.rm(cliThreadStatePath(threadId), { force: true }).catch(() => void 0);
+async function deleteCliThreadState(threadId, threadScopeKey) {
+  await fs14.promises.rm(cliThreadStatePath(threadId, threadScopeKey), { force: true }).catch(() => void 0);
 }
-function cliThreadStatePath(threadId) {
+function cliThreadStatePath(threadId, threadScopeKey) {
   const base = process.env.PERCH_CLI_STATE_DIR?.trim() || path14.join(os5.homedir(), ".perch", "threads");
-  return path14.join(base, `${safeCliThreadId(threadId)}.json`);
+  return path14.join(base, safeCliThreadScopeKey(threadScopeKey), `${safeCliThreadId(threadId)}.json`);
+}
+async function resolveCliThreadScopeKey(connection) {
+  const session = await readStoredCliAuthSession().catch(() => null);
+  const sessionUsable = isStoredCliAuthSessionUsable(session);
+  const authenticated = Boolean(connection?.authenticated || sessionUsable);
+  const appUrl = connection?.appUrl ?? session?.appUrl ?? "local";
+  const userId = connection?.userId ?? session?.userId ?? null;
+  const email = connection?.email ?? session?.email ?? null;
+  const fallbackToken = authenticated && !userId && !email ? session?.accessToken ?? null : null;
+  const authKey = authenticated ? `user:${userId ?? ""}|email:${email ?? ""}|token:${fallbackToken ? createHash4("sha256").update(fallbackToken).digest("hex").slice(0, 20) : ""}` : "anonymous";
+  const raw = `app:${appUrl}|${authKey}`;
+  const prefix = authenticated ? "user" : "anon";
+  return `${prefix}-${createHash4("sha256").update(raw).digest("hex").slice(0, 20)}`;
+}
+function safeCliThreadScopeKey(threadScopeKey) {
+  return (threadScopeKey?.trim() || "anon-local").replace(/[^a-zA-Z0-9_.-]+/g, "_").slice(0, 96);
 }
 function safeCliThreadId(threadId) {
   return threadId.trim().replace(/[^a-zA-Z0-9_.-]+/g, "_").slice(0, 96) || "cli-default";
@@ -285480,6 +286267,7 @@ var init_perch_cli = __esm({
     init_learningMemory();
     init_toolDefinitions();
     init_nodeLocalBridge();
+    init_agentSkillRegistry();
     execFileAsync3 = promisify3(execFile3);
     DEFAULT_CLI_LOGIN_APP_URL = "https://app.perchai.app";
     CLI_PACKAGE_VERSION = readCliPackageVersion();
@@ -285513,6 +286301,7 @@ Usage:
   perch login
   perch status
   perch logout
+  perch skills [name]
   perch run "<task>" [--json] [--thread <id>] [--cwd <dir>] [--mode ask|agents|plan] [--persona saffron|quill]
   perch ap evidence <folder> [--json] [--out <dir>] [--timestamp <id>]
   perch ap packet   <folder> [--json] [--out <dir>] [--timestamp <id>]
@@ -285523,6 +286312,7 @@ Commands:
   login         Open browser sign-in and save this terminal's Perch session.
   status        Show the saved CLI auth/app connection status.
   logout        Clear the saved CLI auth session.
+  skills        List Perch core skills, or show one skill body.
   run           Run a natural-language Perch turn through the shared runtime.
   ap evidence   Prepare AP evidence, cases, metrics, and graph artifacts.
   ap packet     Generate the full AP audit packet and report artifacts.
@@ -285532,6 +286322,7 @@ Commands:
   While Perch is working: type steering text and press Enter; Ctrl-C stops; Ctrl-E expands details.
   /help                         Show this list.
   /status                       Show cwd, auth, mode, persona, permission, and thread.
+  /skills [name]                 List Perch core skills, or show one skill body.
   /cwd [dir]                    Show or change the working directory.
   /permission [mode]            Show or set default, auto_review, take_the_wheel, or plan.
   /permissions [mode]           Alias for /permission.
@@ -285582,6 +286373,7 @@ Commands:
     ];
     PERCH_SPLASH_COMMANDS = [
       ["/status", "show auth, route, tools, and thread"],
+      ["/skills", "show reusable operator skills"],
       ["/persona", "swap saffron or quill"],
       ["/permission", "change autonomy for the next turns"],
       ["/login", "connect your Perch account"]