From ad02febf6a5a75b9ddb894c9c8231358289f686f Mon Sep 17 00:00:00 2001
From: namos2502 <nam.tr.ha02@gmail.com>
Date: Mon, 23 Mar 2026 18:57:59 -0400
Subject: [PATCH 1/5] fix: correct Copilot CLI invocation patterns and sandbox
 handling

- Remove -s flag (causes exit code 1)
- Fix sandbox: cd to repo root instead of nonexistent --cwd flag
- Fix --allow-tool format: no spaces (write,read)
- Update model: claude-sonnet-4-5 for real tasks, haiku for pings
- Add --cwd to all Claude CLI invocation patterns
- Fix orchestration detection command: remove -s, add cd to root
- Bump version to 0.5.6

Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>
---
 .claude-plugin/plugin.json         |  2 +-
 CHANGELOG.md                       | 10 ++++++
 skills/agents/claude-cli/SKILL.md  | 21 +++++++++--
 skills/agents/copilot-cli/SKILL.md | 58 ++++++++++++++++++++----------
 skills/orchestration/SKILL.md      |  3 +-
 5 files changed, 71 insertions(+), 23 deletions(-)

diff --git a/.claude-plugin/plugin.json b/.claude-plugin/plugin.json
index ba227c1..3f24424 100644
--- a/.claude-plugin/plugin.json
+++ b/.claude-plugin/plugin.json
@@ -1,7 +1,7 @@
 {
   "name": "cortexlink",
   "description": "Cross-agent workflow plugin — gives your agent the full programmatic reference for every CLI tool you use. One plugin. Every CLI. Stay in flow.",
-  "version": "0.5.5",
+  "version": "0.5.6",
   "author": {
     "name": "namos2502"
   },
diff --git a/CHANGELOG.md b/CHANGELOG.md
index a7f5449..20059f0 100644
--- a/CHANGELOG.md
+++ b/CHANGELOG.md
@@ -2,6 +2,16 @@
 
 All notable changes to CortexLink will be documented here.
 
+## [0.5.6] — 2026-03-23
+
+### Fixed
+- Removed `-s` flag from all Copilot CLI invocation patterns — causes exit code 1, not a valid flag for programmatic use
+- Copilot CLI sandbox fix: no `--cwd` flag exists; all invocations now `cd` to repo root first (`cd $(git rev-parse --show-toplevel) &&`) — prevents silent file access failures under `--no-ask-user` when invoked from a subdirectory
+- `--allow-tool` values corrected to no-space format (`'write,read'`, `'write,shell(git:*),read'`)
+- Model updated to `--model=claude-sonnet-4-5` for real tasks (was "omit"); `--model=claude-haiku-4.5` kept for availability checks only
+- Claude CLI: added `--cwd $(git rev-parse --show-toplevel)` to all invocation patterns — same sandbox restriction applies
+- Orchestration detection command: removed `-s`, added `cd` to repo root
+
 ## [0.5.5] — 2026-03-23
 
 ### Changed
diff --git a/skills/agents/claude-cli/SKILL.md b/skills/agents/claude-cli/SKILL.md
index b6723b2..82d4e76 100644
--- a/skills/agents/claude-cli/SKILL.md
+++ b/skills/agents/claude-cli/SKILL.md
@@ -66,11 +66,22 @@ Shell access (`Bash(...)`) is a separate, deliberate decision. Use `--disallowed
 
 Use short aliases (`sonnet`, `opus`) for the latest version, or full IDs (e.g. `claude-sonnet-4-6`) to pin a specific model.
 
+## Working Directory
+
+Claude CLI is sandboxed to its working directory — the same restriction as Copilot CLI.
+
+⛔ **Always invoke from the repo root.** Use `--cwd` or `cd` before invoking.
+
+```bash
+claude -p "..." --cwd $(git rev-parse --show-toplevel) --output-format text ...
+```
+
 ## Invocation Patterns
 
 **Read-only delegation (question, analysis):**
 ```bash
 claude -p "[delegation prompt]" --output-format text \
+  --cwd $(git rev-parse --show-toplevel) \
   --allowedTools "Read" --model claude-haiku-4-5 \
   --no-session-persistence --max-turns 3
 ```
@@ -78,6 +89,7 @@ claude -p "[delegation prompt]" --output-format text \
 **Write delegation (fix, implement):**
 ```bash
 claude -p "[delegation prompt]" --output-format text \
+  --cwd $(git rev-parse --show-toplevel) \
   --allowedTools "Read" "Edit" "Write" \
   --no-session-persistence
 ```
@@ -85,6 +97,7 @@ claude -p "[delegation prompt]" --output-format text \
 **Write delegation + shell (runs commands):**
 ```bash
 claude -p "[delegation prompt]" --output-format text \
+  --cwd $(git rev-parse --show-toplevel) \
   --allowedTools "Read" "Edit" "Bash(git *)" \
   --no-session-persistence
 ```
@@ -92,12 +105,14 @@ claude -p "[delegation prompt]" --output-format text \
 **Planning / analysis only (no writes):**
 ```bash
 claude -p "[delegation prompt]" --output-format text \
+  --cwd $(git rev-parse --show-toplevel) \
   --permission-mode plan --no-session-persistence --max-turns 5
 ```
 
 **Piped input:**
 ```bash
 cat file.ts | claude -p "[delegation prompt]" --output-format text \
+  --cwd $(git rev-parse --show-toplevel) \
   --allowedTools "Read" --no-session-persistence
 ```
 
@@ -116,7 +131,9 @@ Follow the template from `skills/orchestration/SKILL.md`. Include the structured
 
 The agent's stdout is its report. Capture it directly:
 ```bash
-REPORT=$(claude -p "[prompt]" --output-format text --allowedTools "Read" --no-session-persistence 2>/dev/null)
+REPORT=$(claude -p "[prompt]" --output-format text \
+  --cwd $(git rev-parse --show-toplevel) \
+  --allowedTools "Read" --no-session-persistence 2>/dev/null)
 ```
 
 Read STATUS first. If ⚠️ or ❌, read ISSUES before deciding next action.
@@ -126,8 +143,8 @@ Read STATUS first. If ⚠️ or ❌, read ISSUES before deciding next action.
 | | Copilot CLI | Claude CLI |
 |--|-------------|------------|
 | Tool permissions | `--allow-tool='write, read'` | `--allowedTools "Read" "Edit"` |
-| Silence | `-s` | `--output-format text` |
 | Prevent questions | `--no-ask-user` | implied by `-p` |
+| Working directory | `--cwd PATH` | `--cwd PATH` |
 | Model flag | `--model=claude-haiku-4.5` | `--model claude-haiku-4-5` |
 
 ## Error Handling
diff --git a/skills/agents/copilot-cli/SKILL.md b/skills/agents/copilot-cli/SKILL.md
index 1a9b4d5..f10b638 100644
--- a/skills/agents/copilot-cli/SKILL.md
+++ b/skills/agents/copilot-cli/SKILL.md
@@ -23,12 +23,26 @@ description: Behavioral reference for GitHub Copilot CLI. Use when delegating Gi
 | Flag | Purpose |
 |------|---------|
 | `-p "PROMPT"` | Programmatic mode — executes prompt and exits |
-| `-s` | Silent — output only the agent response (no usage stats) |
-| `--no-ask-user` | Agent works autonomously, no questions |
-| `--no-auto-update` | Suppress update checks |
-| `--no-color` | Plain text output |
+| `--no-ask-user` | Suppresses interactive permission prompts — required for non-interactive use |
+| `--no-auto-update` | Prevents Copilot from updating itself mid-run |
+| `--no-color` | Clean output, no ANSI escape codes |
+| `-s` | ❌ Causes exit code 1 — do not use |
 
-Always combine `-s --no-ask-user --no-auto-update --no-color` for clean programmatic output.
+Always combine `--no-ask-user --no-auto-update --no-color` for clean programmatic output.
+
+## Sandbox Limitation
+
+Copilot CLI is sandboxed to its working directory. With `--no-ask-user`, any file access outside that directory silently fails — no error, no output. There is no `--cwd` flag; the only fix is to `cd` to the repo root before invoking.
+
+⛔ **Always `cd` to the repo root before invoking.** Never call from a subdirectory.
+
+```bash
+# ❌ Sandboxed to client/ — reads outside it silently fail
+copilot -p "..." --no-ask-user ...
+
+# ✅ Correct — full repo accessible
+cd $(git rev-parse --show-toplevel) && copilot -p "..." --no-ask-user ...
+```
 
 ## Tool Permissions (`--allow-tool`)
 
@@ -37,15 +51,15 @@ Pre-approves tools so the agent doesn't pause to prompt.
 | Use case | Flag |
 |----------|------|
 | Questions, analysis, review | `--allow-tool='read'` |
-| Modify or create files | `--allow-tool='write, read'` |
-| Modify files + run git commands | `--allow-tool='write, shell(git:*), read'` |
+| Modify or create files | `--allow-tool='write,read'` |
+| Modify files + run git commands | `--allow-tool='write,shell(git:*),read'` |
 
 Shell access (`shell(...)`) is a separate, deliberate decision — not an automatic addition to write access. Only grant it when the task genuinely requires running commands.
 
 Use `--deny-tool` to block specific commands within an allowed scope:
 ```bash
 # Allow git reads but block pushes
---allow-tool='shell(git:*), read' --deny-tool='shell(git push)'
+--allow-tool='shell(git:*),read' --deny-tool='shell(git push)'
 ```
 Deny rules always override allow rules.
 
@@ -53,33 +67,37 @@ Deny rules always override allow rules.
 
 | Task | Flag |
 |------|-------|
-| Quick question, analysis | `--model=claude-haiku-4.5` |
-| Complex fix, multi-step | omit (uses session default) |
+| Availability check / quick question | `--model=claude-haiku-4.5` |
+| Real tasks — analysis, fix, review | `--model=claude-sonnet-4-5` |
 
 ## Invocation Patterns
 
 **Read-only delegation (question, analysis):**
 ```bash
-copilot -p "[delegation prompt]" -s --no-ask-user --no-auto-update --no-color \
-  --allow-tool='read' --model=claude-haiku-4.5
+cd $(git rev-parse --show-toplevel) && \
+copilot -p "[delegation prompt]" --no-ask-user --no-auto-update --no-color \
+  --allow-tool='read' --model=claude-sonnet-4-5
 ```
 
 **Write delegation (fix, implement):**
 ```bash
-copilot -p "[delegation prompt]" -s --no-ask-user --no-auto-update --no-color \
-  --allow-tool='write, read'
+cd $(git rev-parse --show-toplevel) && \
+copilot -p "[delegation prompt]" --no-ask-user --no-auto-update --no-color \
+  --allow-tool='write,read' --model=claude-sonnet-4-5
 ```
 
 **Write delegation + git (runs git commands):**
 ```bash
-copilot -p "[delegation prompt]" -s --no-ask-user --no-auto-update --no-color \
-  --allow-tool='write, shell(git:*), read'
+cd $(git rev-parse --show-toplevel) && \
+copilot -p "[delegation prompt]" --no-ask-user --no-auto-update --no-color \
+  --allow-tool='write,shell(git:*),read' --model=claude-sonnet-4-5
 ```
 
 **Code review (built-in /review agent):**
 ```bash
-copilot -p "/review [scope]" -s --no-ask-user --no-auto-update --no-color \
-  --allow-tool='shell(git:*), read'
+cd $(git rev-parse --show-toplevel) && \
+copilot -p "/review [scope]" --no-ask-user --no-auto-update --no-color \
+  --allow-tool='shell(git:*),read' --model=claude-sonnet-4-5
 ```
 
 Redirect stderr if needed: add `2>/dev/null`
@@ -95,7 +113,9 @@ Follow the template from `skills/orchestration/SKILL.md`. Include the structured
 
 The agent's stdout is its report. Capture it directly:
 ```bash
-REPORT=$(copilot -p "[prompt]" -s --no-ask-user --no-auto-update --no-color --allow-tool='read' 2>/dev/null)
+REPORT=$(cd $(git rev-parse --show-toplevel) && \
+  copilot -p "[prompt]" --no-ask-user --no-auto-update --no-color \
+  --allow-tool='read' --model=claude-sonnet-4-5 2>/dev/null)
 ```
 
 Read STATUS first. If ⚠️ or ❌, read ISSUES before deciding next action.
diff --git a/skills/orchestration/SKILL.md b/skills/orchestration/SKILL.md
index 241caac..b136695 100644
--- a/skills/orchestration/SKILL.md
+++ b/skills/orchestration/SKILL.md
@@ -147,7 +147,8 @@ Execute → Verify → (fix if needed) → Report.
 
 ```bash
 # Copilot CLI
-which copilot 2>/dev/null && copilot -p "ping" -s --no-ask-user --no-auto-update --no-color --allow-tool='read' --model=claude-haiku-4.5 2>/dev/null | head -1
+which copilot 2>/dev/null && cd $(git rev-parse --show-toplevel) && \
+  copilot -p "ping" --no-ask-user --no-auto-update --no-color --allow-tool='read' --model=claude-haiku-4.5 2>/dev/null | head -1
 
 # Claude CLI
 which claude 2>/dev/null && claude -p "ping" --output-format text --allowedTools "Read" --max-turns 1 --no-session-persistence 2>/dev/null | head -1

From 2ec87f05c3edf411fcf1446e890161b50ace22f4 Mon Sep 17 00:00:00 2001
From: namos2502 <nam.tr.ha02@gmail.com>
Date: Tue, 24 Mar 2026 14:57:58 -0400
Subject: [PATCH 2/5] feat: apply plugin-dev eval improvements (v0.5.7)

Commands:
- Add allowed-tools frontmatter to setup (Read, Edit, Bash) and cleanup (Read, Edit)
- Fix help directive style to direct output instruction

Skills:
- Rewrite all 3 skill descriptions to third-person trigger format
- Add user-invocable: false to orchestration, copilot-cli, claude-cli skills
- Split orchestration/SKILL.md (8.8KB -> 5.3KB): extract delegation template
  and report format into references/ for progressive disclosure

Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>
---
 .claude-plugin/plugin.json                    |   2 +-
 CHANGELOG.md                                  |   9 ++
 commands/cleanup.md                           |   1 +
 commands/help.md                              |   2 +-
 commands/setup.md                             |   1 +
 skills/agents/claude-cli/SKILL.md             |   4 +-
 skills/agents/copilot-cli/SKILL.md            |   4 +-
 skills/orchestration/SKILL.md                 | 115 +++---------------
 .../references/delegation-template.md         |  36 ++++++
 .../orchestration/references/report-format.md |  72 +++++++++++
 10 files changed, 139 insertions(+), 107 deletions(-)
 create mode 100644 skills/orchestration/references/delegation-template.md
 create mode 100644 skills/orchestration/references/report-format.md

diff --git a/.claude-plugin/plugin.json b/.claude-plugin/plugin.json
index 3f24424..06fbb39 100644
--- a/.claude-plugin/plugin.json
+++ b/.claude-plugin/plugin.json
@@ -1,7 +1,7 @@
 {
   "name": "cortexlink",
   "description": "Cross-agent workflow plugin — gives your agent the full programmatic reference for every CLI tool you use. One plugin. Every CLI. Stay in flow.",
-  "version": "0.5.6",
+  "version": "0.5.7",
   "author": {
     "name": "namos2502"
   },
diff --git a/CHANGELOG.md b/CHANGELOG.md
index 20059f0..350589f 100644
--- a/CHANGELOG.md
+++ b/CHANGELOG.md
@@ -2,6 +2,15 @@
 
 All notable changes to CortexLink will be documented here.
 
+## [0.5.7] — 2026-03-24
+
+### Improved
+- **Commands:** Added `allowed-tools` frontmatter to `setup` (Read, Edit, Bash) and `cleanup` (Read, Edit) — prevents unexpected permission prompts during the most critical user interactions
+- **Commands:** Fixed `help` command directive style — replaced indirect "Reply to the user with..." with direct output instruction
+- **Skills:** All three skill descriptions rewritten to third-person trigger format ("This skill should be used when...") for reliable auto-activation
+- **Skills:** Added `user-invocable: false` to `orchestration`, `copilot-cli`, and `claude-cli` skills — these are internal protocol skills, not user-invocable actions
+- **Skills:** Split `orchestration/SKILL.md` (8.8 KB → 5.3 KB) — delegation prompt template and report format details extracted to `references/delegation-template.md` and `references/report-format.md` for progressive disclosure
+
 ## [0.5.6] — 2026-03-23
 
 ### Fixed
diff --git a/commands/cleanup.md b/commands/cleanup.md
index 04f95f3..5836f08 100644
--- a/commands/cleanup.md
+++ b/commands/cleanup.md
@@ -1,5 +1,6 @@
 ---
 description: "Remove CortexLink configuration added by /cortexlink:setup"
+allowed-tools: Read, Edit
 ---
 
 Remove the CortexLink sections that were added during setup. Run the following steps in order.
diff --git a/commands/help.md b/commands/help.md
index ab7a76f..486d22f 100644
--- a/commands/help.md
+++ b/commands/help.md
@@ -2,7 +2,7 @@
 description: "Show CortexLink plugin information and how to get started"
 ---
 
-Reply to the user with the following information:
+Output the following to the user:
 
 **CortexLink** — One plugin. Every CLI. Stay in flow.
 
diff --git a/commands/setup.md b/commands/setup.md
index 9784632..2712475 100644
--- a/commands/setup.md
+++ b/commands/setup.md
@@ -1,5 +1,6 @@
 ---
 description: "One-time setup — detects installed CLI agents, authenticates, and registers CortexLink as always-on in ~/.claude/CLAUDE.md and ~/.copilot/copilot-instructions.md"
+allowed-tools: Read, Edit, Bash
 ---
 
 Run the following steps in order and report the results clearly to the user.
diff --git a/skills/agents/claude-cli/SKILL.md b/skills/agents/claude-cli/SKILL.md
index 82d4e76..799e929 100644
--- a/skills/agents/claude-cli/SKILL.md
+++ b/skills/agents/claude-cli/SKILL.md
@@ -1,7 +1,7 @@
 ---
 name: claude-cli
-description: Behavioral reference for Claude CLI. Use when delegating general code tasks, analysis, or tasks requiring Anthropic-native reasoning via `claude -p`.
-
+description: This skill should be used when delegating general code tasks — explanations, refactors, fixes, analysis — via `claude -p`. Also use when context isolation or a specific Anthropic model is needed for the subtask.
+user-invocable: false
 ---
 
 # Claude CLI Agent
diff --git a/skills/agents/copilot-cli/SKILL.md b/skills/agents/copilot-cli/SKILL.md
index f10b638..32fd6c5 100644
--- a/skills/agents/copilot-cli/SKILL.md
+++ b/skills/agents/copilot-cli/SKILL.md
@@ -1,7 +1,7 @@
 ---
 name: copilot-cli
-description: Behavioral reference for GitHub Copilot CLI. Use when delegating GitHub-specific or GitHub-adjacent tasks via `copilot -p`.
-
+description: This skill should be used when delegating GitHub-specific tasks — PRs, issues, repos, Actions, branches, code review — via `copilot -p`. Use it when Copilot CLI's GitHub-native tooling is an advantage over the host agent.
+user-invocable: false
 ---
 
 # Copilot CLI Agent
diff --git a/skills/orchestration/SKILL.md b/skills/orchestration/SKILL.md
index b136695..de82c6d 100644
--- a/skills/orchestration/SKILL.md
+++ b/skills/orchestration/SKILL.md
@@ -1,7 +1,7 @@
 ---
 name: orchestration
-description: Cross-agent orchestration protocol. Teaches the host AI agent to act as a persistent control center — decomposing tasks, fanning out to peer CLI agents, receiving verified reports, and synthesizing results.
-
+description: This skill should be used when a multi-step task can benefit from cross-CLI delegation — decomposing work into subtasks, routing them to Copilot CLI or Claude CLI agents, and synthesizing their structured reports back to the user.
+user-invocable: false
 ---
 
 # CortexLink Orchestration
@@ -31,7 +31,7 @@ CONTROL CENTER (your AI agent — decomposes, routes, reviews)
         └── [own tools & sub-agents — internal, platform-native]
 ```
 
-The control center IS your active AI agent — it holds the plan and directs the work. Each Agent is a full CLI agent in its own right, not a dumb executor. It can use its own platform-native tools and sub-agents internally to complete the task. The tree is one level deep: Agents are peers, they don't chain to each other. Width scales as you add agents; depth stays fixed.
+The control center IS your active AI agent — it holds the plan and directs the work. Each Agent is a full CLI agent in its own right, not a dumb executor. The tree is one level deep: Agents are peers, they don't chain to each other. Width scales as you add agents; depth stays fixed.
 
 ## Task Complexity
 
@@ -39,17 +39,17 @@ Classify every task before routing. This determines how much spec detail to writ
 
 | Level | Signals | Delegation style |
 |-------|---------|-----------------|
-| **Simple** | Single operation, read-only, unambiguous output (a list, a status, a count) | Handle inline — no cross-CLI |
+| **Simple** | Single operation, read-only, unambiguous output | Handle inline — no cross-CLI |
 | **Standard** | 2–4 operations, may write, clear success criteria, low rework cost | Full spec: problem + acceptance criteria |
-| **Complex** | 5+ operations, writes/commits/PRs, judgment calls, OR high rework cost if misunderstood | Full spec + Q&A turn before execution |
+| **Complex** | 5+ operations, writes/commits/PRs, judgment calls, OR high rework cost | Full spec + Q&A turn before execution |
 
-**Key signal for Complex:** rework cost. If the agent misunderstands and you must redo the work — how expensive is that? Tasks with irreversible steps (PRs, commits, deploys) or required judgment calls always qualify.
+**Key signal for Complex:** rework cost. Tasks with irreversible steps (PRs, commits, deploys) or required judgment calls always qualify.
 
 ## When to Delegate Cross-CLI
 
 **Delegate when:**
 1. **Platform-specific** — GitHub ops (PRs, repos, Actions) → Copilot; Anthropic reasoning or Claude-specific model → Claude CLI
-2. **Context isolation** — offload a long subtask so its intermediate work never enters your context (the final result DOES return via stdout)
+2. **Context isolation** — offload a long subtask so its intermediate work never enters your context
 3. **Different model needed** — the target CLI runs a model the host cannot
 4. Task is **Standard or Complex** — the work inside the agent justifies the delegation overhead
 
@@ -58,7 +58,7 @@ Classify every task before routing. This determines how much spec detail to writ
 2. The host has a native subagent that can handle it
 3. The target CLI is not installed or not authenticated
 4. You have no clear reason — "big task" is not a reason
-5. Task is **Simple** — handle inline or use a native subagent; a full agent session costs more than the task itself
+5. Task is **Simple** — handle inline; a full agent session costs more than the task itself
 
 **Decision tree:**
 ```
@@ -78,101 +78,11 @@ New task →
 ## Control Center Protocol
 
 1. **Decompose** — Break into scoped, independently executable subtasks. Each must be verifiable by the agent itself.
-2. **Route** — Apply decision tree. Check agent availability (see below).
-3. **Dispatch** — Use delegation prompt template. Always include scope, success criteria, report format.
+2. **Route** — Apply decision tree. Check agent availability (see `references/report-format.md`).
+3. **Dispatch** — Use delegation prompt template (see `references/delegation-template.md`). Always include scope, success criteria, report format.
 4. **Review** — Read STATUS first. Spot-check if needed (`git diff`, tests). Decide: proceed, re-assign, or adjust.
 5. **Track** — Update state (done / pending / failed). Never skip to the next subtask without reviewing the current report.
-6. **Synthesize** — When all subtasks are done, consolidate into one output for the user. Never surface raw agent output. Lead with issues (🔴 blocker / 🟠 should fix / 🟡 minor), then a one-sentence verdict on what to do next. If any subtask is ❌, hold the verdict until resolved.
-
-## Delegation Prompt Template
-
-```
-[Task]: <one sentence — what needs to be done>
-[Context]: <only what the agent strictly needs>
-[Success criteria]: <how the agent knows it's done>
-[Report format]: Return ONLY this when finished. Plain text labels only — no **bold**, no # headers:
-  STATUS: ✅ Verified / ⚠️ Partial / ❌ Failed
-  SUMMARY: <1-2 sentences>
-  STEPS: <bullets>
-  FILES: <changed files or "none">
-  ISSUES: <notes for the control center or "none">
-  Max 150 words.
-```
-
-Keep prompts lean. No project dumps.
-
-**For Complex tasks only**, append this block to the delegation prompt before the report format:
-
-```
-[Before executing]: This task is complex. In your first response, list any questions
-or ambiguities — 1 turn only. Do not perform any actions until I confirm.
-```
-
-The agent surfaces questions, you refine the spec if needed, then issue the execution follow-up.
-
-## Structured Report Format
-
-Every agent, every tier, every time:
-
-```
-STATUS: ✅ Verified / ⚠️ Partial / ❌ Failed
-SUMMARY: <1-2 sentences>
-STEPS:
-  - <step 1>
-  - <step 2>
-FILES: <list or "none">
-ISSUES: <notes or "none">
-```
-
-**Never return raw output or tool call dumps to the control center.**
-
-## Self-Verify Before Reporting (when YOU are the agent)
-
-Execute → Verify → (fix if needed) → Report.
-
-1. Run the task
-2. Check your own work — tests, file contents, expected state
-3. If verification fails → fix, then verify again
-4. Only then → write the structured report
-
-| Status | Meaning |
-|--------|---------|
-| ✅ Verified | You actively checked. It works. |
-| ⚠️ Partial | Some parts verified, some could not be checked. Explain in ISSUES. |
-| ❌ Failed | Describe what was attempted and what failed. |
-
-**Never mark ✅ without actually checking.**
-
-## Agent Detection
-
-```bash
-# Copilot CLI
-which copilot 2>/dev/null && cd $(git rev-parse --show-toplevel) && \
-  copilot -p "ping" --no-ask-user --no-auto-update --no-color --allow-tool='read' --model=claude-haiku-4.5 2>/dev/null | head -1
-
-# Claude CLI
-which claude 2>/dev/null && claude -p "ping" --output-format text --allowedTools "Read" --max-turns 1 --no-session-persistence 2>/dev/null | head -1
-```
-
-If unavailable: fall back to another agent, handle natively, or tell the user to run `/cortexlink:setup`.
-
-## Error Handling
-
-| STATUS | Control center action |
-|--------|-----------------------|
-| ✅ | Accept, proceed |
-| ⚠️ | Read ISSUES, accept partial or re-assign unverified parts |
-| ❌ | Read ISSUES. Re-assign, adjust scope, or handle natively |
-| No output | Retry once. If still failing, handle natively. |
-| Auth failure | Tell user: `/cortexlink:setup` |
-
-**After 2 consecutive ❌ on the same subtask — stop and discuss with the user before retrying.**
-
-## Cross-Agent Chaining
-
-Fan out to multiple Agents in parallel when tasks are independent. Pass a report excerpt (SUMMARY + STEPS) as context into the next delegation prompt — never raw output. Agents do not chain to each other; all coordination happens at the control center.
-
-**Avoid duplicate fetches:** When multiple agents need the same source data (e.g. a PR diff, a file's contents), fetch it once in the control center and pass it as `[Context]` in each delegation prompt. Do not let each agent re-fetch the same data independently.
+6. **Synthesize** — Consolidate into one output for the user. Lead with issues (🔴 blocker / 🟠 should fix / 🟡 minor), then a one-sentence verdict. If any subtask is ❌, hold the verdict until resolved.
 
 ## Red Flags — STOP
 
@@ -190,6 +100,9 @@ Fan out to multiple Agents in parallel when tasks are independent. Pass a report
 |------------|---------|
 | GitHub task | Copilot — see `skills/agents/copilot-cli/SKILL.md` |
 | General code task | Claude CLI — see `skills/agents/claude-cli/SKILL.md` |
+| Delegation prompt | See `references/delegation-template.md` |
+| Report format / self-verify | See `references/report-format.md` |
 | Add a new agent | Drop `skills/agents/<name>/SKILL.md` — follow existing format |
 | Handle auth failure | Tell user: `/cortexlink:setup` |
 | Handle ❌ report | Read ISSUES. Re-assign or handle natively. Never retry blindly. |
+
diff --git a/skills/orchestration/references/delegation-template.md b/skills/orchestration/references/delegation-template.md
new file mode 100644
index 0000000..b841439
--- /dev/null
+++ b/skills/orchestration/references/delegation-template.md
@@ -0,0 +1,36 @@
+# Delegation Prompt Template
+
+Use this template for every cross-CLI delegation. Keep prompts lean — no project dumps.
+
+```
+[Task]: <one sentence — what needs to be done>
+[Context]: <only what the agent strictly needs>
+[Success criteria]: <how the agent knows it's done>
+[Report format]: Return ONLY this when finished. Plain text labels only — no **bold**, no # headers:
+  STATUS: ✅ Verified / ⚠️ Partial / ❌ Failed
+  SUMMARY: <1-2 sentences>
+  STEPS: <bullets>
+  FILES: <changed files or "none">
+  ISSUES: <notes for the control center or "none">
+  Max 150 words.
+```
+
+**For Complex tasks only**, append this block before the report format:
+
+```
+[Before executing]: This task is complex. In your first response, list any questions
+or ambiguities — 1 turn only. Do not perform any actions until I confirm.
+```
+
+The agent surfaces questions, you refine the spec if needed, then issue the execution follow-up.
+
+⛔ **Critical (Copilot CLI only):** Always include this line in the prompt body:
+*"Return ONLY the structured report. No reasoning steps, no 'Let me...' output before the report."*
+
+---
+
+# Cross-Agent Chaining
+
+Fan out to multiple Agents in parallel when tasks are independent. Pass a report excerpt (SUMMARY + STEPS) as context into the next delegation prompt — never raw output. Agents do not chain to each other; all coordination happens at the control center.
+
+**Avoid duplicate fetches:** When multiple agents need the same source data (e.g. a PR diff, a file's contents), fetch it once in the control center and pass it as `[Context]` in each delegation prompt. Do not let each agent re-fetch the same data independently.
diff --git a/skills/orchestration/references/report-format.md b/skills/orchestration/references/report-format.md
new file mode 100644
index 0000000..09dc3f5
--- /dev/null
+++ b/skills/orchestration/references/report-format.md
@@ -0,0 +1,72 @@
+# Structured Report Format
+
+Every agent, every tier, every time:
+
+```
+STATUS: ✅ Verified / ⚠️ Partial / ❌ Failed
+SUMMARY: <1-2 sentences>
+STEPS:
+  - <step 1>
+  - <step 2>
+FILES: <list or "none">
+ISSUES: <notes or "none">
+```
+
+**Never return raw output or tool call dumps to the control center.**
+
+---
+
+# Self-Verify Before Reporting (when YOU are the agent)
+
+Execute → Verify → (fix if needed) → Report.
+
+1. Run the task
+2. Check your own work — tests, file contents, expected state
+3. If verification fails → fix, then verify again
+4. Only then → write the structured report
+
+| Status | Meaning |
+|--------|---------|
+| ✅ Verified | You actively checked. It works. |
+| ⚠️ Partial | Some parts verified, some could not be checked. Explain in ISSUES. |
+| ❌ Failed | Describe what was attempted and what failed. |
+
+**Never mark ✅ without actually checking.**
+
+**Status semantics for analysis tasks:**
+- ✅ Verified = you read the code and completed the analysis — even if you cannot *run* it
+- ⚠️ Partial = you could not access required files, or the diff was truncated
+- ❌ Failed = the command failed or the data was inaccessible
+
+⛔ Do NOT use ⚠️ just because you cannot execute the code being reviewed. Completed static analysis = ✅.
+
+---
+
+# Agent Detection
+
+Run these before delegating to verify agents are available:
+
+```bash
+# Copilot CLI
+which copilot 2>/dev/null && cd $(git rev-parse --show-toplevel) && \
+  copilot -p "ping" --no-ask-user --no-auto-update --no-color --allow-tool='read' --model=claude-haiku-4.5 2>/dev/null | head -1
+
+# Claude CLI
+which claude 2>/dev/null && claude -p "ping" --output-format text --allowedTools "Read" --max-turns 1 --no-session-persistence 2>/dev/null | head -1
+```
+
+If unavailable: fall back to another agent, handle natively, or tell the user to run `/cortexlink:setup`.
+
+---
+
+# Error Handling
+
+| STATUS | Control center action |
+|--------|-----------------------|
+| ✅ | Accept, proceed |
+| ⚠️ | Read ISSUES, accept partial or re-assign unverified parts |
+| ❌ | Read ISSUES. Re-assign, adjust scope, or handle natively |
+| No output | Retry once. If still failing, handle natively. |
+| Auth failure | Tell user: `/cortexlink:setup` |
+
+**After 2 consecutive ❌ on the same subtask — stop and discuss with the user before retrying.**

From a9256d222480c1b4f9509f65c7c63a044e82b231 Mon Sep 17 00:00:00 2001
From: namos2502 <nam.tr.ha02@gmail.com>
Date: Wed, 25 Mar 2026 10:25:42 -0400
Subject: [PATCH 3/5] fix: correct claude-cli SKILL.md invocation example and
 comparison table

- Add missing 'Write' to 'Write delegation + shell' allowedTools example
  (was inconsistent with the Tool Permissions table above it)
- Fix Key Differences table: Copilot CLI has no --cwd flag; correct to
  'cd $(git rev-parse --show-toplevel) &&' pattern

Fixes CodeRabbit findings on PR #6.

Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>
---
 skills/agents/claude-cli/SKILL.md | 4 ++--
 1 file changed, 2 insertions(+), 2 deletions(-)

diff --git a/skills/agents/claude-cli/SKILL.md b/skills/agents/claude-cli/SKILL.md
index 799e929..b5b3926 100644
--- a/skills/agents/claude-cli/SKILL.md
+++ b/skills/agents/claude-cli/SKILL.md
@@ -98,7 +98,7 @@ claude -p "[delegation prompt]" --output-format text \
 ```bash
 claude -p "[delegation prompt]" --output-format text \
   --cwd $(git rev-parse --show-toplevel) \
-  --allowedTools "Read" "Edit" "Bash(git *)" \
+  --allowedTools "Read" "Edit" "Write" "Bash(git *)" \
   --no-session-persistence
 ```
 
@@ -144,7 +144,7 @@ Read STATUS first. If ⚠️ or ❌, read ISSUES before deciding next action.
 |--|-------------|------------|
 | Tool permissions | `--allow-tool='write, read'` | `--allowedTools "Read" "Edit"` |
 | Prevent questions | `--no-ask-user` | implied by `-p` |
-| Working directory | `--cwd PATH` | `--cwd PATH` |
+| Working directory | `cd $(git rev-parse --show-toplevel) &&` | `--cwd PATH` |
 | Model flag | `--model=claude-haiku-4.5` | `--model claude-haiku-4-5` |
 
 ## Error Handling

From d4de4a217d74c8c842d40eb4b95973871a997f04 Mon Sep 17 00:00:00 2001
From: namos2502 <nam.tr.ha02@gmail.com>
Date: Mon, 30 Mar 2026 19:42:55 -0400
Subject: [PATCH 4/5] feat: add SessionStart hook and agent context injection

- Add hooks/hooks.json (Claude Code) and hooks/hooks-cursor.json (Cursor)
  to register a SessionStart hook that fires on startup|clear|compact
- Add hooks/run-hook.cmd cross-platform polyglot shim (Windows + Unix)
- Add hooks/session-start: injects orchestration/SKILL.md as additionalContext
  into every host session from turn 1 via hookSpecificOutput (Claude Code)
  or additional_context (Cursor)
- Add skills/orchestration/references/agent-context.md: compact CortexLink
  agent protocol (report format + self-verify) for delegated agents
- Update claude-cli/SKILL.md: all invocation patterns now include
  --append-system-prompt with inlined agent context (system prompt layer)
- Update copilot-cli/SKILL.md: delegation prompt now opens with CortexLink
  agent context preamble (no --append-system-prompt equivalent available)

Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>
---
 hooks/hooks-cursor.json                       | 10 ++++
 hooks/hooks.json                              | 16 +++++++
 hooks/run-hook.cmd                            | 46 +++++++++++++++++++
 hooks/session-start                           | 43 +++++++++++++++++
 skills/agents/claude-cli/SKILL.md             | 21 +++++++--
 skills/agents/copilot-cli/SKILL.md            | 15 ++++--
 .../orchestration/references/agent-context.md | 44 ++++++++++++++++++
 7 files changed, 187 insertions(+), 8 deletions(-)
 create mode 100644 hooks/hooks-cursor.json
 create mode 100644 hooks/hooks.json
 create mode 100755 hooks/run-hook.cmd
 create mode 100755 hooks/session-start
 create mode 100644 skills/orchestration/references/agent-context.md

diff --git a/hooks/hooks-cursor.json b/hooks/hooks-cursor.json
new file mode 100644
index 0000000..6df4461
--- /dev/null
+++ b/hooks/hooks-cursor.json
@@ -0,0 +1,10 @@
+{
+  "version": 1,
+  "hooks": {
+    "sessionStart": [
+      {
+        "command": "./hooks/session-start"
+      }
+    ]
+  }
+}
diff --git a/hooks/hooks.json b/hooks/hooks.json
new file mode 100644
index 0000000..79d8cee
--- /dev/null
+++ b/hooks/hooks.json
@@ -0,0 +1,16 @@
+{
+  "hooks": {
+    "SessionStart": [
+      {
+        "matcher": "startup|clear|compact",
+        "hooks": [
+          {
+            "type": "command",
+            "command": "\"${CLAUDE_PLUGIN_ROOT}/hooks/run-hook.cmd\" session-start",
+            "async": false
+          }
+        ]
+      }
+    ]
+  }
+}
diff --git a/hooks/run-hook.cmd b/hooks/run-hook.cmd
new file mode 100755
index 0000000..ceec3a7
--- /dev/null
+++ b/hooks/run-hook.cmd
@@ -0,0 +1,46 @@
+: << 'CMDBLOCK'
+@echo off
+REM Cross-platform polyglot wrapper for hook scripts.
+REM On Windows: cmd.exe runs the batch portion, which finds and calls bash.
+REM On Unix: the shell interprets this as a script (: is a no-op in bash).
+REM
+REM Hook scripts use extensionless filenames (e.g. "session-start" not
+REM "session-start.sh") so Claude Code's Windows auto-detection -- which
+REM prepends "bash" to any command containing .sh -- doesn't interfere.
+REM
+REM Usage: run-hook.cmd <script-name> [args...]
+
+if "%~1"=="" (
+    echo run-hook.cmd: missing script name >&2
+    exit /b 1
+)
+
+set "HOOK_DIR=%~dp0"
+
+REM Try Git for Windows bash in standard locations
+if exist "C:\Program Files\Git\bin\bash.exe" (
+    "C:\Program Files\Git\bin\bash.exe" "%HOOK_DIR%%~1" %2 %3 %4 %5 %6 %7 %8 %9
+    exit /b %ERRORLEVEL%
+)
+if exist "C:\Program Files (x86)\Git\bin\bash.exe" (
+    "C:\Program Files (x86)\Git\bin\bash.exe" "%HOOK_DIR%%~1" %2 %3 %4 %5 %6 %7 %8 %9
+    exit /b %ERRORLEVEL%
+)
+
+REM Try bash on PATH (e.g. user-installed Git Bash, MSYS2, Cygwin)
+where bash >nul 2>nul
+if %ERRORLEVEL% equ 0 (
+    bash "%HOOK_DIR%%~1" %2 %3 %4 %5 %6 %7 %8 %9
+    exit /b %ERRORLEVEL%
+)
+
+REM No bash found - exit silently rather than error
+REM (plugin still works, just without SessionStart context injection)
+exit /b 0
+CMDBLOCK
+
+# Unix: run the named script directly
+SCRIPT_DIR="$(cd "$(dirname "$0")" && pwd)"
+SCRIPT_NAME="$1"
+shift
+exec bash "${SCRIPT_DIR}/${SCRIPT_NAME}" "$@"
diff --git a/hooks/session-start b/hooks/session-start
new file mode 100755
index 0000000..c689bc6
--- /dev/null
+++ b/hooks/session-start
@@ -0,0 +1,43 @@
+#!/usr/bin/env bash
+# SessionStart hook for CortexLink plugin
+
+set -euo pipefail
+
+SCRIPT_DIR="$(cd "$(dirname "$0")" && pwd)"
+PLUGIN_ROOT="$(cd "${SCRIPT_DIR}/.." && pwd)"
+
+# Read orchestration skill — the CortexLink meta skill
+orchestration_content=$(cat "${PLUGIN_ROOT}/skills/orchestration/SKILL.md" 2>&1 || echo "Error reading orchestration skill")
+
+# Escape string for JSON embedding.
+# Each ${s//old/new} is a single C-level pass — fast and heredoc-safe.
+escape_for_json() {
+    local s="$1"
+    s="${s//\\/\\\\}"
+    s="${s//\"/\\\"}"
+    s="${s//$'\n'/\\n}"
+    s="${s//$'\r'/\\r}"
+    s="${s//$'\t'/\\t}"
+    printf '%s' "$s"
+}
+
+orchestration_escaped=$(escape_for_json "$orchestration_content")
+
+session_context="<EXTREMELY_IMPORTANT>\nYou have CortexLink — a cross-agent workflow plugin.\n\nBelow is your orchestration skill. It defines when and how to delegate tasks across CLI agents (Copilot CLI, Claude CLI). Follow it whenever multi-agent delegation could apply.\n\n${orchestration_escaped}\n</EXTREMELY_IMPORTANT>"
+
+# Platform detection: emit the field consumed by the current platform only.
+# Cursor sets CURSOR_PLUGIN_ROOT (may also set CLAUDE_PLUGIN_ROOT).
+# Claude Code sets CLAUDE_PLUGIN_ROOT only.
+# Emitting both fields would cause double injection in Claude Code.
+#
+# Uses printf instead of heredoc to avoid a bash 5.3+ bug where heredoc
+# variable expansion hangs when content exceeds ~512 bytes.
+if [ -n "${CURSOR_PLUGIN_ROOT:-}" ]; then
+    printf '{\n  "additional_context": "%s"\n}\n' "$session_context"
+elif [ -n "${CLAUDE_PLUGIN_ROOT:-}" ]; then
+    printf '{\n  "hookSpecificOutput": {\n    "hookEventName": "SessionStart",\n    "additionalContext": "%s"\n  }\n}\n' "$session_context"
+else
+    printf '{\n  "additional_context": "%s"\n}\n' "$session_context"
+fi
+
+exit 0
diff --git a/skills/agents/claude-cli/SKILL.md b/skills/agents/claude-cli/SKILL.md
index b5b3926..ae9ac68 100644
--- a/skills/agents/claude-cli/SKILL.md
+++ b/skills/agents/claude-cli/SKILL.md
@@ -78,12 +78,19 @@ claude -p "..." --cwd $(git rev-parse --show-toplevel) --output-format text ...
 
 ## Invocation Patterns
 
+Always include `--append-system-prompt` with the CortexLink agent context below. This teaches the agent the report format and self-verify protocol via the system prompt — separate from the task prompt.
+
+```
+CORTEXLINK_AGENT_CONTEXT="You are operating as a CortexLink agent. Your output is consumed directly by a control center. Execute the task, verify your own work (Execute → Verify → fix if needed → Report), then return ONLY this structured report — plain text labels, no bold, no headers:\nSTATUS: ✅ Verified / ⚠️ Partial / ❌ Failed\nSUMMARY: <1-2 sentences>\nSTEPS:\n  - <step>\nFILES: <changed files or none>\nISSUES: <notes or none>\nMax 150 words. Never mark ✅ without actually checking. For analysis tasks, ✅ = completed the analysis even if you cannot run the code."
+```
+
 **Read-only delegation (question, analysis):**
 ```bash
 claude -p "[delegation prompt]" --output-format text \
   --cwd $(git rev-parse --show-toplevel) \
   --allowedTools "Read" --model claude-haiku-4-5 \
-  --no-session-persistence --max-turns 3
+  --no-session-persistence --max-turns 3 \
+  --append-system-prompt "$CORTEXLINK_AGENT_CONTEXT"
 ```
 
 **Write delegation (fix, implement):**
@@ -91,7 +98,8 @@ claude -p "[delegation prompt]" --output-format text \
 claude -p "[delegation prompt]" --output-format text \
   --cwd $(git rev-parse --show-toplevel) \
   --allowedTools "Read" "Edit" "Write" \
-  --no-session-persistence
+  --no-session-persistence \
+  --append-system-prompt "$CORTEXLINK_AGENT_CONTEXT"
 ```
 
 **Write delegation + shell (runs commands):**
@@ -99,21 +107,24 @@ claude -p "[delegation prompt]" --output-format text \
 claude -p "[delegation prompt]" --output-format text \
   --cwd $(git rev-parse --show-toplevel) \
   --allowedTools "Read" "Edit" "Write" "Bash(git *)" \
-  --no-session-persistence
+  --no-session-persistence \
+  --append-system-prompt "$CORTEXLINK_AGENT_CONTEXT"
 ```
 
 **Planning / analysis only (no writes):**
 ```bash
 claude -p "[delegation prompt]" --output-format text \
   --cwd $(git rev-parse --show-toplevel) \
-  --permission-mode plan --no-session-persistence --max-turns 5
+  --permission-mode plan --no-session-persistence --max-turns 5 \
+  --append-system-prompt "$CORTEXLINK_AGENT_CONTEXT"
 ```
 
 **Piped input:**
 ```bash
 cat file.ts | claude -p "[delegation prompt]" --output-format text \
   --cwd $(git rev-parse --show-toplevel) \
-  --allowedTools "Read" --no-session-persistence
+  --allowedTools "Read" --no-session-persistence \
+  --append-system-prompt "$CORTEXLINK_AGENT_CONTEXT"
 ```
 
 ## Delegation Prompt
diff --git a/skills/agents/copilot-cli/SKILL.md b/skills/agents/copilot-cli/SKILL.md
index 32fd6c5..153235e 100644
--- a/skills/agents/copilot-cli/SKILL.md
+++ b/skills/agents/copilot-cli/SKILL.md
@@ -104,10 +104,19 @@ Redirect stderr if needed: add `2>/dev/null`
 
 ## Delegation Prompt
 
-Follow the template from `skills/orchestration/SKILL.md`. Include the structured report format instructions at the end of every prompt.
+Copilot CLI has no `--append-system-prompt` flag, so CortexLink agent context must be prepended to every delegation prompt. **Always open every delegation prompt with this block:**
 
-⛔ **Critical:** Always include this exact line in your delegation prompt:
-*"Return ONLY the structured report. No reasoning steps, no 'Let me...' output before the report."*
+```
+You are operating as a CortexLink agent. Your output is consumed directly by a control center — not displayed to a user. Execute the task, verify your own work (Execute → Verify → fix if needed → Report), then return ONLY the structured report below. No reasoning steps, no "Let me..." output before the report.
+
+---
+
+[Task]: ...
+```
+
+The report format goes at the end (from the delegation template). The opening block teaches the agent the protocol — the closing block specifies this task's expected output.
+
+⛔ **Critical:** Never omit the opening CortexLink agent context block. Without it, Copilot CLI has no way to learn the report format or self-verify protocol.
 
 ## Handling the Report
 
diff --git a/skills/orchestration/references/agent-context.md b/skills/orchestration/references/agent-context.md
new file mode 100644
index 0000000..fa6845c
--- /dev/null
+++ b/skills/orchestration/references/agent-context.md
@@ -0,0 +1,44 @@
+# CortexLink Agent Context
+
+You are operating as a **CortexLink agent**. Your output is consumed directly by a control center — not displayed to a user. Follow this protocol exactly.
+
+## Your Job
+
+Execute the task. Verify your own work. Return the structured report below. Nothing else.
+
+## Self-Verify Before Reporting
+
+Execute → Verify → (fix if needed) → Report.
+
+- Run the task
+- Check your own work (tests, file contents, expected state)
+- If verification fails → fix, then verify again
+- Only then → write the report
+
+## Report Format
+
+Return ONLY this. Plain text labels — no **bold**, no # headers:
+
+```
+STATUS: ✅ Verified / ⚠️ Partial / ❌ Failed
+SUMMARY: <1-2 sentences>
+STEPS:
+  - <step 1>
+  - <step 2>
+FILES: <changed files or "none">
+ISSUES: <notes for the control center or "none">
+```
+
+Max 150 words.
+
+## Status Semantics
+
+| Status | Meaning |
+|--------|---------|
+| ✅ Verified | You actively checked. It works. |
+| ⚠️ Partial | Some parts verified, some could not be checked. Explain in ISSUES. |
+| ❌ Failed | Describe what was attempted and what failed in ISSUES. |
+
+For analysis tasks: ✅ = you read the code and completed the analysis (even if you cannot *run* it). Do NOT use ⚠️ just because you cannot execute the code being reviewed.
+
+**Never mark ✅ without actually checking.**

From bc5ec72f248b8ab022cd075f432434d27af39694 Mon Sep 17 00:00:00 2001
From: namos2502 <nam.tr.ha02@gmail.com>
Date: Mon, 30 Mar 2026 22:28:59 -0400
Subject: [PATCH 5/5] fix: enforce one-agent-per-subtask and peer direction
 constraints
MIME-Version: 1.0
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 8bit

- Delete hooks/hooks-cursor.json (Cursor IDE format — not a supported
  CortexLink platform; Copilot CLI uses hooks/hooks.json via CLAUDE_PLUGIN_ROOT)
- Update hooks/session-start: inject PEER DIRECTION RULES block — Claude Code
  delegates to Copilot CLI (GitHub) or Claude CLI (last resort); Copilot CLI
  delegates to Claude CLI; ONE agent per subtask; bash -p only
- Update orchestration/SKILL.md:
  - Iron Law: add ONE AGENT PER SUBTASK rule + peer direction by host identity
  - When to Delegate: add peer direction preamble and routing exclusivity
  - Dispatch: add one-agent-per-subtask rule + programmatic-mode requirement
  - Red Flags: add 3 new entries (same subtask to >1 agent, interactive
    terminal, Claude→Claude-CLI when native tools suffice)
- Update claude-cli/SKILL.md: add ⛔ last-resort note when host is Claude Code

Fixes: model spawning both Copilot CLI and Claude CLI for a single task

Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>
---
 hooks/hooks-cursor.json           | 10 ----------
 hooks/session-start               |  8 +++++---
 skills/agents/claude-cli/SKILL.md |  4 ++++
 skills/orchestration/SKILL.md     | 14 +++++++++++++-
 4 files changed, 22 insertions(+), 14 deletions(-)
 delete mode 100644 hooks/hooks-cursor.json

diff --git a/hooks/hooks-cursor.json b/hooks/hooks-cursor.json
deleted file mode 100644
index 6df4461..0000000
--- a/hooks/hooks-cursor.json
+++ /dev/null
@@ -1,10 +0,0 @@
-{
-  "version": 1,
-  "hooks": {
-    "sessionStart": [
-      {
-        "command": "./hooks/session-start"
-      }
-    ]
-  }
-}
diff --git a/hooks/session-start b/hooks/session-start
index c689bc6..81bf513 100755
--- a/hooks/session-start
+++ b/hooks/session-start
@@ -23,11 +23,13 @@ escape_for_json() {
 
 orchestration_escaped=$(escape_for_json "$orchestration_content")
 
-session_context="<EXTREMELY_IMPORTANT>\nYou have CortexLink — a cross-agent workflow plugin.\n\nBelow is your orchestration skill. It defines when and how to delegate tasks across CLI agents (Copilot CLI, Claude CLI). Follow it whenever multi-agent delegation could apply.\n\n${orchestration_escaped}\n</EXTREMELY_IMPORTANT>"
+peer_direction="PEER DIRECTION RULES (apply these every time you delegate):\n- If you are Claude Code (Claude): GitHub tasks → Copilot CLI. Isolated code/analysis tasks → Claude CLI (last resort only — prefer native tools first).\n- If you are Copilot CLI: code tasks, analysis, and refactors → Claude CLI.\n- ONE agent per subtask. Never route the same subtask to more than one agent.\n- Always use bash with -p flag (programmatic mode). Never open an interactive terminal session.\n- Violating any of these rules violates the spirit of CortexLink."
+
+session_context="<EXTREMELY_IMPORTANT>\nYou have CortexLink — a cross-agent workflow plugin.\n\nBelow is your orchestration skill. It defines when and how to delegate tasks across CLI agents (Copilot CLI, Claude CLI). Follow it whenever multi-agent delegation could apply.\n\n${orchestration_escaped}\n\n${peer_direction}\n</EXTREMELY_IMPORTANT>"
 
 # Platform detection: emit the field consumed by the current platform only.
-# Cursor sets CURSOR_PLUGIN_ROOT (may also set CLAUDE_PLUGIN_ROOT).
-# Claude Code sets CLAUDE_PLUGIN_ROOT only.
+# Claude Code and Copilot CLI both set CLAUDE_PLUGIN_ROOT (shared plugin architecture).
+# Cursor IDE sets CURSOR_PLUGIN_ROOT (not a supported CortexLink platform — kept as fallback).
 # Emitting both fields would cause double injection in Claude Code.
 #
 # Uses printf instead of heredoc to avoid a bash 5.3+ bug where heredoc
diff --git a/skills/agents/claude-cli/SKILL.md b/skills/agents/claude-cli/SKILL.md
index ae9ac68..c86f2f4 100644
--- a/skills/agents/claude-cli/SKILL.md
+++ b/skills/agents/claude-cli/SKILL.md
@@ -8,6 +8,10 @@ user-invocable: false
 
 ## When to Use
 
+⛔ If the host is **Claude Code**: use Claude CLI only when you specifically need context isolation or a different Anthropic model AND native tools (Task tool, inline work) are insufficient. Claude Code → Claude CLI is last resort. Prefer Copilot CLI for GitHub tasks; prefer Claude Code's native Task tool for general code tasks.
+
+---
+
 **Use Claude CLI for:**
 - General code tasks — explanations, refactors, fixes, analysis
 - Tasks needing Anthropic-specific models (Sonnet, Opus, Haiku)
diff --git a/skills/orchestration/SKILL.md b/skills/orchestration/SKILL.md
index de82c6d..3ef9b0f 100644
--- a/skills/orchestration/SKILL.md
+++ b/skills/orchestration/SKILL.md
@@ -19,6 +19,9 @@ CortexLink turns your active AI agent into a **control center** that fans out ta
 ```
 NO CROSS-CLI DELEGATION WITHOUT A CLEAR REASON.
 NO PROCEEDING WITHOUT REVIEWING THE REPORT.
+ONE AGENT PER SUBTASK — never delegate the same subtask to more than one agent.
+Claude Code → Copilot CLI (GitHub tasks) or Claude CLI (last resort for isolated tasks).
+Copilot CLI → Claude CLI.
 ```
 
 ## The Two-Tier Model
@@ -47,6 +50,12 @@ Classify every task before routing. This determines how much spec detail to writ
 
 ## When to Delegate Cross-CLI
 
+Each subtask goes to exactly **ONE agent** — the decision tree picks which one, then stop. Never route the same subtask to both agents simultaneously.
+
+**Peer direction by host:**
+- **Claude Code** → Copilot CLI for GitHub tasks; Claude CLI only as last resort (when context isolation or a specific Anthropic model is needed AND native tools are insufficient)
+- **Copilot CLI** → Claude CLI for code tasks, analysis, and refactors
+
 **Delegate when:**
 1. **Platform-specific** — GitHub ops (PRs, repos, Actions) → Copilot; Anthropic reasoning or Claude-specific model → Claude CLI
 2. **Context isolation** — offload a long subtask so its intermediate work never enters your context
@@ -79,7 +88,7 @@ New task →
 
 1. **Decompose** — Break into scoped, independently executable subtasks. Each must be verifiable by the agent itself.
 2. **Route** — Apply decision tree. Check agent availability (see `references/report-format.md`).
-3. **Dispatch** — Use delegation prompt template (see `references/delegation-template.md`). Always include scope, success criteria, report format.
+3. **Dispatch** — Use delegation prompt template (see `references/delegation-template.md`). Always include scope, success criteria, report format. Each subtask goes to ONE agent — never dispatch the same work to two agents. Always use bash with `-p` (programmatic/print mode); never open an interactive terminal session.
 4. **Review** — Read STATUS first. Spot-check if needed (`git diff`, tests). Decide: proceed, re-assign, or adjust.
 5. **Track** — Update state (done / pending / failed). Never skip to the next subtask without reviewing the current report.
 6. **Synthesize** — Consolidate into one output for the user. Lead with issues (🔴 blocker / 🟠 should fix / 🟡 minor), then a one-sentence verdict. If any subtask is ❌, hold the verdict until resolved.
@@ -93,6 +102,9 @@ New task →
 - Retrying a ❌ task without reading ISSUES first
 - Delegating a task that needs current session context
 - Agents chaining to each other (all coordination goes through the control center)
+- Delegating the same subtask to more than one agent (one subtask → one agent, always)
+- Opening an interactive terminal session for delegation (always use `-p` programmatic mode)
+- Claude Code delegating to Claude CLI when native tools (Task tool, inline work) would suffice
 
 ## Quick Reference