Inventory (offline; paths and parsed labels only):
  excluded roots: 0
  git-ignored roots: 0
  nested git repositories: 0
  observed starts: {"adapter_coverage": {"claude-code": {"broken_lines": 0, "duplicate_events": 0, "files_scanned": 0, "inherited_session_meta": 0, "invalid_timestamps": 0, "invalid_token_records": 0, "lines_read": 0, "unknown_record_types": {}, "unreadable_files": 0}, "codex": {"broken_lines": 0, "duplicate_events": 0, "files_scanned": 0, "inherited_session_meta": 0, "invalid_timestamps": 0, "invalid_token_records": 0, "lines_read": 0, "unknown_record_types": {}, "unreadable_files": 0}}, "agents": [], "coverage": {}, "paths": [], "placement_sessions": 0, "roles": {"top-level": 0, "worker": 0}, "sessions": 0, "start_evidence": {}, "status": "no-starts-found", "window": {"since": "<START_WINDOW_SINCE>", "until": "<START_WINDOW_UNTIL>"}}
  exclude **/.git: 0
  exclude **/.sumbi: 0
  exclude **/node_modules: 0
  exclude **/vendor: 0
  exclude **/.venv: 0
  exclude **/venv: 0
  exclude **/dist: 0
  exclude **/build: 0
  exclude **/__pycache__: 0
  exclude **/.tmp: 0
  exclude **/test/**/fixtures: 0
  exclude **/test/**/testdata: 0
  exclude **/tests/**/fixtures: 0
  exclude **/tests/**/testdata: 0
  capabilities: claude_skills=0, agent_skills=0, commands=0, agents=0
  workflows=0, codeowners=0, test_sources=0
  codeowners: 0
  pr_templates: 0
  git_hooks: 0
  lint: 0
  format: 0
  type_check: 0
  required checks: unknown (offline); rulesets: unknown (offline)
  estimate: script-aware-v1: ceil(other_characters / 4) + Hangul/Han/Hiragana/Katakana characters
  instruction tokens (root/largest scope): codex=0/0, claude=0/0, gemini=0/0, copilot=0/0, cursor=0/0
  conditional instruction tokens claude: 0 (0 rules)
  convention review_gate: absent []
  convention handoff: absent []
  convention friction_line: absent []
  convention parallel_worktree: absent []
  convention plan_approval: absent []
  budget=2000; warnings=1
  warning: {"kind": "root-not-versioned"}
Gaps: missing-agents, missing-test-command, missing-risk-review, missing-handoff, missing-friction-line, missing-worktree-rule, missing-plan-approval
  missing-agents: {"expected_path": "AGENTS.md", "paths": []}; candidates=instruction-map
  missing-test-command: {"paths": []}; candidates=test-and-verify
  missing-risk-review: {"paths": [], "searched_paths": []}; candidates=review-gates
  missing-handoff: {"paths": [], "searched_paths": []}; candidates=handoff
  missing-friction-line: {"paths": [], "searched_paths": []}; candidates=friction-line
  missing-worktree-rule: {"paths": [], "searched_paths": []}; candidates=parallel-worktrees
  missing-plan-approval: {"paths": [], "searched_paths": []}; candidates=plan-and-approval
Practices: instruction-map, test-and-verify, review-gates, plan-and-approval, handoff, friction-line, parallel-worktrees
baseline: available (runs before apply)
--- /dev/null
+++ b/AGENTS.md
@@ -0,0 +1,27 @@
+<!-- sumbi:begin instruction-map -->
+Keep this entry file brief. Open detailed guidance only when the task needs it. See [instruction maintenance](docs/sumbi/instructions.md).
+<!-- sumbi:end instruction-map -->
+
+<!-- sumbi:begin test-and-verify -->
+Before claiming completion, run the task's relevant checks and report their results. See [test evidence](docs/sumbi/testing.md).
+<!-- sumbi:end test-and-verify -->
+
+<!-- sumbi:begin review-gates -->
+Risky paths need review before release. Preserve existing approvals and require independent review for security, permissions and money paths. See [review levels](docs/sumbi/review.md).
+<!-- sumbi:end review-gates -->
+
+<!-- sumbi:begin plan-and-approval -->
+For substantial changes, resolve missing requirements, write a plan with acceptance criteria, and critique it before implementation. Follow the existing approval policy. See [planning](docs/sumbi/planning.md).
+<!-- sumbi:end plan-and-approval -->
+
+<!-- sumbi:begin handoff -->
+Leave a handoff document when work changes hands: acceptance criteria, evidence, remaining work and the next action. See [handoffs](docs/sumbi/handoff.md).
+<!-- sumbi:end handoff -->
+
+<!-- sumbi:begin friction-line -->
+End a work report with `Harness friction: none` or one short description of the harness obstacle. Keep sensitive details in local notes. See [friction feedback](docs/sumbi/friction.md).
+<!-- sumbi:end friction-line -->
+
+<!-- sumbi:begin parallel-worktrees -->
+Parallel agents use separate worktrees and explicit file ownership. Do not edit another agent's working copy. See [parallel work](docs/sumbi/parallel.md).
+<!-- sumbi:end parallel-worktrees -->
--- /dev/null
+++ b/docs/sumbi/friction.md
@@ -0,0 +1,7 @@
+<!-- sumbi:begin friction-line -->
+# Friction feedback
+
+Use the final friction line to identify an instruction, tool or check that blocked or repeated work. Describe the failing layer and its consequence briefly; do not include prompts, transcripts, secrets or machine identifiers in shared output.
+
+Review repeated obstacles before adding more instructions. Prefer a reliable check or a simpler workflow when it addresses the cause. A report is evidence to investigate, not proof that the proposed fix saves cost.
+<!-- sumbi:end friction-line -->
--- /dev/null
+++ b/docs/sumbi/handoff.md
@@ -0,0 +1,7 @@
+<!-- sumbi:begin handoff -->
+# Handoff convention
+
+Keep each live handoff in a local location approved by the repository owner. Do not commit session transcripts, secrets or machine paths.
+
+Record the deliverable ID and acceptance criteria, current state, files changed, checks actually run, unresolved risks, and the next action. Separate observations from assumptions. A successor should be able to resume without repeating the investigation.
+<!-- sumbi:end handoff -->
--- /dev/null
+++ b/docs/sumbi/instructions.md
@@ -0,0 +1,7 @@
+<!-- sumbi:begin instruction-map -->
+# Instruction maintenance
+
+Keep the entry file as a table of contents plus the rules that always apply. Put examples, background and specialized procedures in documents opened on demand.
+
+An owner should review the token estimate, move optional detail, and confirm that approvals and review gates still reach every affected agent. sumbi only adds this guidance; it never trims existing instructions automatically.
+<!-- sumbi:end instruction-map -->
--- /dev/null
+++ b/docs/sumbi/parallel.md
@@ -0,0 +1,7 @@
+<!-- sumbi:begin parallel-worktrees -->
+# Parallel work
+
+Assign each parallel agent a separate branch and worktree, plus an explicit set of files it owns. Agree on shared interfaces before editing. Check the working copy and branch before making changes.
+
+Integrate through reviewed changes. Preserve another agent's uncommitted work, and resolve overlaps with its owner. Never remove a worktree or branch merely because a task appears finished.
+<!-- sumbi:end parallel-worktrees -->
--- /dev/null
+++ b/docs/sumbi/planning.md
@@ -0,0 +1,7 @@
+<!-- sumbi:begin plan-and-approval -->
+# Planning
+
+Ask only for information that blocks a correct implementation. Record the deliverable, acceptance criteria, constraints and likely failure cases. Critique the plan against those criteria before making a substantial change. Keep small reversible edits proportionate.
+
+Prepare a concrete reviewable result before requesting a final approval. Existing approvals still govern destructive or external actions. Do not treat silence as approval, and do not ask again when the requested scope is already authorized.
+<!-- sumbi:end plan-and-approval -->
--- /dev/null
+++ b/docs/sumbi/review.md
@@ -0,0 +1,11 @@
+<!-- sumbi:begin review-gates -->
+# Review levels
+
+| Work size or risk | Evidence and review |
+| --- | --- |
+| Small wording or local configuration | Read the resulting change and check the affected behavior. |
+| Behavior across several components | Ask an independent engineering reviewer; include targeted test evidence. |
+| Security, permissions, money, or irreversible operations | Require owner approval and a reviewer from a different model family; include the applicable security and release checks. |
+
+The owner must map risky repository paths to these roles. A role is a review responsibility, not a claim that a tool enforces it. Existing stronger gates remain in force. If review is unavailable, leave the risky action pending.
+<!-- sumbi:end review-gates -->
--- /dev/null
+++ b/docs/sumbi/testing.md
@@ -0,0 +1,7 @@
+<!-- sumbi:begin test-and-verify -->
+# Test evidence
+
+Owner setup: record the repository's exact test command here, together with any required environment. Until then, the missing-command gap remains unresolved.
+
+For a behavior change, first show a focused check that exposes the problem, then make it pass. Run the required checks for the final change. State the command, result and any untested boundary in the local handoff. Do not turn an intended check into a claim that it ran.
+<!-- sumbi:end test-and-verify -->
