agentic-swe 1.0.0

package/.claude/templates/audit.log

@@ -0,0 +1,30 @@
+# Audit Log
+
+Use append-only records for operations that change durable state or external side effects.
+
+Format:
+
+```text
+<timestamp> actor=<agent> action=<verb> target=<file-or-resource> result=<ok|blocked|failed> note="<short explanation>"
+```
+
+Example:
+
+```text
+<timestamp> actor=claude action=create target=.claude/.work/<id>/state.json result=ok note="initialized from template"
+```
+
+## Actor Naming Convention
+
+Use these standardized actor names in all audit log entries:
+
+`orchestrator` | `developer` | `git-ops` | `pr-manager` | `panel-architect` | `panel-security` | `panel-adversarial` | `user`
+
+## Delegation Examples
+
+```text
+<timestamp> actor=orchestrator action=delegate target=agents/developer.md result=ok note="implementation slice 1: auth middleware"
+<timestamp> actor=developer action=complete target=agents/developer.md result=ok note="3 files changed, 2 tests added"
+<timestamp> actor=orchestrator action=integrate target=agents/developer.md result=ok note="findings merged into implementation.md"
+<timestamp> actor=orchestrator action=escalate target=escalate-code result=ok note="code review loop exhausted after 5 iterations"
+```

package/.claude/templates/evidence-standard.md

@@ -0,0 +1,19 @@
+# Evidence Standard
+
+Cite this template in phase and agent output to ensure consistent evidence quality.
+
+## Four-Point Framework
+
+Every substantive finding or recommendation must address:
+
+1. **Observed** — what was directly seen in the repository, command output, or documentation
+2. **Inferred** — what conclusion follows from the observations
+3. **Evidence** — what specific facts support the inference
+4. **Uncertain** — what remains unknown or unverified
+
+## Separation Rules
+
+- separate facts from inferences from speculative concerns
+- label low-evidence concerns as speculative rather than presenting them as certainty
+- when uncertainty is high and behavior would materially change, stop instead of guessing
+- delegated findings should be integrated with evidence attribution, not copied blindly

package/.claude/templates/phase-checklist.md

@@ -0,0 +1,62 @@
+# Phase Checklist
+
+Use this checklist before every state transition.
+
+## Structural Gate
+
+- required files for the destination state exist
+- required files are non-empty
+- `state.json` `current_state` matches the intended transition
+- `history` includes the most recent reasoned transition
+- artifact filenames in `state.json.artifacts` match actual files when applicable
+- `updated_at` reflects the latest change
+- `progress.md` has an entry for the current phase work
+- `audit.log` contains durable or external actions
+- loop counters are updated when the phase is budgeted or retry-driven
+
+## Evidence Gate
+
+- each artifact distinguishes observation from inference
+- important claims are supported by repository evidence, command output, or primary documentation
+- open uncertainties are named explicitly
+- delegated findings are integrated, not copied blindly
+
+## Quality Gate
+
+- self-review rubric scores recorded before code-review (full path) or within fast-implementation (fast path)
+- reflection-log read before rework when prior rejection exists
+- test adequacy assessment completed before implementation begins (full path)
+- changed behavior has at least one decisive verification path
+- tests pass before or during `validation`, or the lack of runnable tests is precisely documented
+- reviewable output exists before `pr-created`
+- risky git or GitHub actions use official workflow semantics instead of assumptions
+- design feedback is resolved before moving from `design-review` to `verification`
+- code review rejection results in explicit `review-feedback.md`
+- blocked validation records why escalation is necessary
+
+## Reliability Gate
+
+- likely rollback or repair path is understood for high-risk changes
+- destructive or privileged actions have explicit approval when required
+- external blockers are recorded rather than worked around unsafely
+- resumability is preserved if the work pauses now
+
+## Convergence Gate
+
+- if reflection-log has 2+ entries with the same failure category, escalate rather than retry
+- if fast-path self-review scores a dimension as 1 that the developer cannot resolve, escalate rather than consuming the next iteration
+- if design review is failing on fundamentally different criteria each iteration (thrashing), escalate after iteration 2
+
+## Context Condensation Gate
+
+- if this is the 3rd+ transition, progress.md contains a current Context Summary
+- subsequent phase reads prioritize context summary over full early-phase artifacts
+
+## Stop Conditions
+
+- `ambiguity-wait`
+- `approval-wait`
+- `escalate-code`
+- `escalate-validation`
+- `completed`
+- `failed`

package/.claude/templates/progress.md

@@ -0,0 +1,15 @@
+# Progress Log
+
+Use one line per meaningful change in execution state.
+
+Format:
+
+```md
+- <timestamp> [state=<state>] <action taken> | evidence: <short evidence> | next: <next step>
+```
+
+Example:
+
+```md
+- <timestamp> [state=initialized] work item created | evidence: state.json written from template | next: feasibility
+```

package/.claude/templates/state.json

@@ -0,0 +1,108 @@
+{
+  "schema_version": 2,
+  "work_id": "<work-id>",
+  "task": "<user task>",
+  "current_state": "initialized",
+  "created_at": "<ISO-8601 timestamp>",
+  "updated_at": "<ISO-8601 timestamp>",
+  "timeout_at": "<ISO-8601 timestamp>",
+  "owner": "claude",
+  "mode": "full",
+  "resume": {
+    "detected": false,
+    "last_session_id": null,
+    "detected_at": null,
+    "branch_validated": false,
+    "rebase_performed": false
+  },
+  "budget": {
+    "iteration_budget": 10,
+    "budget_remaining": 10,
+    "cost_budget_usd": 3.0,
+    "cost_used": 0.0
+  },
+  "counters": {
+    "fast_iter": 0,
+    "design_iter": 0,
+    "code_iter": 0,
+    "panel_runs": 0,
+    "merge_iter": 0,
+    "approval_iter": 0,
+    "self_review_iter": 0,
+    "test_adequacy_iter": 0,
+    "subagent_spawns": 0
+  },
+  "ambiguity": {
+    "detected": false,
+    "notes": [],
+    "resolved": false
+  },
+  "approvals": {
+    "pr_approved": false,
+    "changes_requested": false
+  },
+  "metrics": {
+    "tests_passed": false
+  },
+  "risk": {
+    "level": "unknown",
+    "score": null,
+    "top_items": []
+  },
+  "artifacts": {
+    "feasibility": null,
+    "ambiguity-report": null,
+    "fast-path-check": null,
+    "design": null,
+    "design-review": null,
+    "design-feedback": null,
+    "design-panel-review": null,
+    "verification-results": null,
+    "test-stubs": null,
+    "test-results": null,
+    "implementation": null,
+    "self-review": null,
+    "reflection-log": null,
+    "review-feedback": null,
+    "review-pass": null,
+    "permissions-changes": null,
+    "validation-results": null,
+    "cicd": null,
+    "pr-link": null,
+    "approval-feedback": null
+  },
+  "validation": {
+    "structural_passed": false,
+    "quality_passed": false,
+    "last_checked_at": null,
+    "notes": []
+  },
+  "git": {
+    "base_branch": null,
+    "working_branch": null,
+    "remote": "origin",
+    "pr_url": null,
+    "merge_sha": null,
+    "merge_strategy": null
+  },
+  "pipeline": {
+    "complexity_score": null,
+    "fast_path_eligible": false,
+    "fast_path_decision": null,
+    "design_review_status": null,
+    "review_status": null,
+    "verification_status": null,
+    "permissions_status": null,
+    "validation_status": null,
+    "self_review_status": null,
+    "pr_status": null,
+    "subagent_auto_select": true,
+    "subagent_mode": null,
+    "subagent_selections": []
+  },
+  "convergence": {
+    "last_failure_category": null,
+    "consecutive_same_category": 0
+  },
+  "history": []
+}

package/.claude/tools/subagent-catalog/README.md

@@ -0,0 +1,58 @@
+# subagent-catalog
+
+A Claude Code skill for browsing and fetching subagents from the [awesome-claude-code-subagents](https://github.com/VoltAgent/awesome-claude-code-subagents) catalog.
+
+## Installation
+
+Copy the `subagent-catalog/` folder to `~/.claude/commands/`:
+
+```bash
+cp -r tools/subagent-catalog ~/.claude/commands/
+```
+
+## Usage
+
+| Command | Description |
+|---------|-------------|
+| `/subagent-catalog:search <query>` | Find agents by name, description, or category |
+| `/subagent-catalog:fetch <name>` | Get full agent definition |
+| `/subagent-catalog:list` | Browse all categories |
+| `/subagent-catalog:invalidate` | Clear cache (add `--fetch` to refresh immediately) |
+
+## Examples
+
+**Find security-related agents:**
+```
+/subagent-catalog:search security
+```
+
+**Get the code-reviewer definition:**
+```
+/subagent-catalog:fetch code-reviewer
+```
+
+**Browse all available agents:**
+```
+/subagent-catalog:list
+```
+
+## Features
+
+- **Smart caching**: 12-hour TTL with graceful fallback on network failure
+- **Atomic updates**: Uses tmp file + mv pattern to prevent partial writes
+- **Cross-platform**: Works on macOS and Linux
+- **Best practices**: Follows Anthropic skill authoring guidelines
+
+## Cache
+
+- **Location**: `~/.claude/cache/subagent-catalog.md`
+- **TTL**: 12 hours (configurable in `config.sh`)
+- **Behavior**: Auto-refreshes when stale, falls back to old cache on network failure
+
+## Troubleshooting
+
+| Issue | Fix |
+|-------|-----|
+| Stale results | `/subagent-catalog:invalidate --fetch` |
+| Network error | Check connection, retry |
+| Agent not found | `/subagent-catalog:search <partial-name>` first |

package/.claude/tools/subagent-catalog/config.sh

@@ -0,0 +1,88 @@
+#!/usr/bin/env bash
+# subagent-catalog configuration
+# Works with local agents/subagents/ directory
+
+set -euo pipefail
+
+# --- CONFIG ---
+# Resolve the pipeline root (source repo = repo root, target repo = .claude/)
+_resolve_subagent_dir() {
+  local script_dir
+  script_dir="$(cd "$(dirname "${BASH_SOURCE[0]:-${(%):-%x}}")" && pwd 2>/dev/null || pwd)"
+
+  # Walk up to find the pipeline root
+  local candidate="$script_dir"
+  while [ "$candidate" != "/" ]; do
+    if [ -d "$candidate/agents/subagents" ]; then
+      echo "$candidate/agents/subagents"
+      return 0
+    fi
+    candidate="$(dirname "$candidate")"
+  done
+
+  # Fallback: check .claude/ in git root
+  local git_root
+  git_root="$(git rev-parse --show-toplevel 2>/dev/null || echo "")"
+  if [ -n "$git_root" ] && [ -d "$git_root/.claude/agents/subagents" ]; then
+    echo "$git_root/.claude/agents/subagents"
+    return 0
+  fi
+
+  echo ""
+  return 1
+}
+
+readonly SUBAGENT_DIR="$(_resolve_subagent_dir)"
+export SUBAGENT_DIR
+
+# --- HELPERS ---
+
+subagent_catalog_log_info() { echo "$1"; }
+subagent_catalog_log_error() { echo "ERROR: $1" >&2; }
+
+# List all categories (directory names under subagents/)
+subagent_catalog_list_categories() {
+  if [ -z "$SUBAGENT_DIR" ] || [ ! -d "$SUBAGENT_DIR" ]; then
+    subagent_catalog_log_error "subagent directory not found"
+    return 1
+  fi
+  ls -1d "$SUBAGENT_DIR"/*/ 2>/dev/null | xargs -I{} basename {} | sort
+}
+
+# List all agent files in a category
+subagent_catalog_list_agents_in() {
+  local category="$1"
+  if [ -d "$SUBAGENT_DIR/$category" ]; then
+    ls -1 "$SUBAGENT_DIR/$category"/*.md 2>/dev/null | xargs -I{} basename {} .md | sort
+  fi
+}
+
+# Find agent file by name (returns full path)
+subagent_catalog_find() {
+  local name="$1"
+  local result
+  result=$(find "$SUBAGENT_DIR" -name "${name}.md" -type f 2>/dev/null | head -1)
+  if [ -n "$result" ]; then
+    echo "$result"
+    return 0
+  fi
+  # Try partial match
+  result=$(find "$SUBAGENT_DIR" -name "*${name}*.md" -type f 2>/dev/null)
+  echo "$result"
+}
+
+# Search agents by keyword in name or frontmatter description
+subagent_catalog_search() {
+  local query="$1"
+  find "$SUBAGENT_DIR" -name "*.md" -type f 2>/dev/null | while read -r f; do
+    if grep -qil "$query" "$f" 2>/dev/null; then
+      local category
+      category=$(basename "$(dirname "$f")")
+      local agent
+      agent=$(basename "$f" .md)
+      local desc
+      desc=$(grep -m1 '^description:' "$f" 2>/dev/null | sed 's/^description: *"*//;s/"*$//' || echo "")
+      echo "$category/$agent|$desc"
+    fi
+  done
+}

package/.claude/tools/subagent-catalog/fetch.md

@@ -0,0 +1,54 @@
+---
+name: fetch
+description: "Read and display the full definition of a specific subagent from the local catalog."
+---
+
+# Subagent Catalog - Fetch
+
+Get the full definition of a specific agent from local storage.
+
+## Input: $ARGUMENTS
+
+Accepts: agent name (e.g., `code-reviewer`, `python-pro`)
+
+## Instructions
+
+### Step 1: Find agent file
+
+```bash
+source tools/subagent-catalog/config.sh
+subagent_catalog_find "{{NAME}}"
+```
+
+### Step 2: Read and display
+
+Read the agent file and display its contents with parsed frontmatter:
+
+```
+## code-reviewer
+
+**Category**: 04-quality-security
+**Model**: opus
+**Tools**: Read, Write, Edit, Bash, Glob, Grep
+
+[full definition follows]
+```
+
+### Step 3: Show invocation options
+
+After displaying the definition, show how to use it:
+
+```
+**How to invoke:**
+- As orchestrator subagent: Agent(prompt="agents/subagents/04-quality-security/code-reviewer.md")
+- With model override: Agent(prompt="...", model="opus")
+- In background: Agent(prompt="...", run_in_background=true)
+- In isolation: Agent(prompt="...", isolation="worktree")
+```
+
+### Error handling
+
+| error | suggestion |
+|-------|------------|
+| not found | run `/subagent-catalog:search <partial>` |
+| multiple matches | list them, ask user to specify |

package/.claude/tools/subagent-catalog/invalidate.md

@@ -0,0 +1,47 @@
+---
+name: invalidate
+description: "Invalidate the subagent-catalog cache. Use when results seem stale or user explicitly asks to refresh or clear cache."
+---
+
+# Subagent Catalog - Invalidate
+
+Force-refresh the cached catalog by deleting the local cache file. The next `search` or `fetch` call will pull fresh data from the repository.
+
+## Input: $ARGUMENTS
+
+No arguments required. Optional: pass `--fetch` to immediately refresh after invalidation.
+
+## Instructions
+
+### Step 1: Source config
+
+```bash
+source ~/.claude/commands/subagent-catalog/config.sh
+```
+
+### Step 2: Invalidate (and optionally refresh)
+
+**Invalidate only** (default):
+
+```bash
+subagent_catalog_invalidate_cache
+```
+
+**Invalidate and refresh** (if `$ARGUMENTS` contains `--fetch` or user explicitly asks to refresh):
+
+```bash
+subagent_catalog_invalidate_cache
+subagent_catalog_refresh_cache
+```
+
+### Step 3: Confirm
+
+Report the result:
+- If invalidated only: "cache invalidated. next search/fetch will pull fresh data."
+- If refreshed: "cache invalidated and refreshed with latest catalog."
+
+### When to use
+
+- After the upstream repo has been updated with new subagents
+- If you suspect the cache is corrupted
+- To troubleshoot stale results

package/.claude/tools/subagent-catalog/list.md

@@ -0,0 +1,48 @@
+---
+name: list
+description: "List all categories and agents in the local subagent catalog. Use when user wants to browse available subagents."
+---
+
+# Subagent Catalog - List
+
+Browse all available categories and agents from the local subagent catalog.
+
+## Input: $ARGUMENTS
+
+No arguments required.
+
+## Instructions
+
+### Step 1: List categories and agents
+
+```bash
+source tools/subagent-catalog/config.sh
+
+for category in $(subagent_catalog_list_categories); do
+  echo "### $category"
+  subagent_catalog_list_agents_in "$category" | tr '\n' ', ' | sed 's/,$/\n/'
+  echo ""
+done
+```
+
+### Step 2: Format output
+
+Display as a scannable list:
+
+```
+## Subagent Catalog (Local)
+
+### 01-core-development
+api-designer, backend-developer, frontend-developer, fullstack-developer, ...
+
+### 02-language-specialists
+typescript-pro, python-pro, rust-engineer, golang-pro, ...
+
+[...continue for all 10 categories...]
+```
+
+### Tips
+
+- use `/subagent-catalog:search <query>` to filter by keyword
+- use `/subagent-catalog:fetch <name>` to get full definition
+- agents are stored locally in `agents/subagents/`

package/.claude/tools/subagent-catalog/search.md

@@ -0,0 +1,41 @@
+---
+name: search
+description: "Search the local subagent catalog by name, description, or capability keyword."
+---
+
+# Subagent Catalog - Search
+
+Find agents by name, description, or category.
+
+## Input: $ARGUMENTS
+
+## Instructions
+
+### Step 1: Search local agents
+
+```bash
+source tools/subagent-catalog/config.sh
+subagent_catalog_search "{{QUERY}}"
+```
+
+### Step 2: Format results
+
+Display matches as a table:
+
+```
+## Results for "kubernetes"
+
+| Agent | Category | Description |
+|-------|----------|-------------|
+| kubernetes-specialist | 03-infrastructure | Container orchestration master |
+| devops-engineer | 03-infrastructure | CI/CD and automation expert |
+
+> use `/subagent-catalog:fetch <name>` to get full definition
+> invoke with: Agent(prompt=agents/subagents/<category>/<name>.md)
+```
+
+### Step 3: Handle edge cases
+
+- **no results**: suggest related terms or `/subagent-catalog:list`
+- **too many results**: ask user to narrow the query
+- **category match**: show all agents in that category