Complete evolve status CLI

README.md

@@ -130,6 +130,7 @@ ai-agent -c config/agent_config.yaml
 | Mode | Command | Notes |
 |---|---|---|
 | Console | `uv run ai-agent -c config/agent_config.yaml` | Single agent |
+| Evolve status | `uv run ai-agent evolve status` | Organism queue, budgets, last run/PR |
 | Research operator | `uv run ai-agent -c config/agents/operator.yaml` | Interactive Research Desk |
 | One-shot brief | `uv run ai-agent brief "agent harness"` | Writes `briefs/YYYYMMDD_slug.md` |
 | Brief + approve | `uv run ai-agent brief "topic" --approve` | Console Y/n before write |

config/agents/engineer.yaml

@@ -6,8 +6,9 @@ system_prompt: >
   Workflow for each intent:
   1. Survey with workspace_search (search, or read with start_line/end_line).
      Do not open `.ai_agent/**` for doc intents.
-  2. Edit ONCE with replace_in_file (preferred for README bullets). After it
-     succeeds, NEVER call replace_in_file/write_file/apply_patch again.
+  2. Implement the FULL intent with up to 3 related edits (helper + CLI + test
+     when asked). A stub helper alone is NOT done — keep editing until usable.
+     After the feature is complete, NEVER call edit tools again.
   3. VERIFY immediately: git_diff (must be NON-EMPTY) then run_checks preset=pytest.
   4. PUBLISH in SEPARATE rounds: git_commit alone, then open_pull_request alone.
      Commit/PR are auto-approved unless the operator enabled HITL (--approve).
@@ -17,7 +18,8 @@ system_prompt: >
   Allowlist only: src/, tests/, config/, docs/, README/DECISIONS/AGENTS.
   Never edit path_policy, merge policy kernel, STOP, or .env.
   Never invent tool results. Never narrate-and-stop mid-pipeline.
-max_tool_rounds: 16
+  Survey with at most 2-3 searches, then edit — do not burn rounds on search.
+max_tool_rounds: 24
 personality:
   tone: pragmatic
   style: concise

config/evolve_backlog.yaml

@@ -1,8 +1,23 @@
-# Rotating intents used when the organism goal queue is empty.
-# Keep items small, allowlisted (src/tests/config/docs/README*), never kernel edits.
+# Rotating intents when the organism goal queue is empty.
+# Prefer concrete harness features over one-line doc tweaks.
 goals:
-  - "Add or improve one unit test under tests/ for MergePolicy day-budget behavior."
-  - "Clarify Phase 2 evolve-worker usage in README.md with one short paragraph."
-  - "Add a brief AGENTS.md note about evolve-goal add/list and the STOP kill switch."
-  - "Improve DECISIONS.md with one sentence on goal-queue vs free-form invent intents."
-  - "Add or tighten one docstring in src/ai_agent/features/evolve/organism.py without behavior changes."
+  - >-
+    Add `ai-agent evolve status` CLI: print organism id, stopped flag, goal queue,
+    evolves_today/max_evolves_per_day, merges_today, next_wake_at, last_run_id,
+    and if last run exists print its status/pr_url/intent from run.json.
+    Implement in src/ai_agent/entrypoints/cli.py (+ thin helper under
+    features/evolve if needed). Add a focused unit test. Update the Modes table
+    in README.md with one row for evolve status. Do not edit PathPolicy/STOP/.env.
+  - >-
+    Add `ai-agent ops summary` that reads OpsEvent JSONL and prints counts by
+    event name plus success rate and average latency_ms when present. Keep it
+    under src/ai_agent/harness/ops_metrics.py + CLI wiring + one test. Update
+    README Modes/ops row briefly.
+  - >-
+    Add an engineer tool `evolve_inspect` (read-only) that returns JSON for the
+    current organism + last EvolveRun summary under .ai_agent/evolve/. Register
+    it for the engineer agent YAML. Include a unit test. No writes, no git.
+  - >-
+    Fix the broken Agent harness capabilities markdown table in README.md
+    (the Evolve CLI prose currently splits the table) so all capability rows
+    render as one table again, without changing meaning.

src/ai_agent/entrypoints/cli.py

@@ -29,6 +29,7 @@
 from ai_agent.features.evolve.goals import enqueue_goal, list_goals
 from ai_agent.features.evolve.organism import ensure_organism, worker_loop, worker_tick
 from ai_agent.features.evolve.service import run_evolve, save_organism
+from ai_agent.features.evolve.status import build_status_state, render_status
 from ai_agent.features.harness_bank.bank import (
     admit_if_screened,
     list_cells,
@@ -382,6 +383,10 @@ def _run_evolve_goal(args: argparse.Namespace) -> None:
     raise ValueError("evolve-goal requires: add <text> | list")
 
 
+def _run_evolve_status() -> None:
+    print(render_status(build_status_state()))
+
+
 def _run_harness_bank(args: argparse.Namespace) -> None:
     action = args.command_arg
     if action == "list":
@@ -629,16 +634,22 @@ def main(argv: list[str] | None = None) -> None:
                 args.message = args.command_arg2 or args.topic or ""
             _run_harness_command(args)
         elif args.command == "evolve":
-            intent = args.topic or args.command_arg
-            if not intent:
-                raise ValueError('evolve requires an intent: ai-agent evolve "…"')
-            asyncio.run(
-                _run_evolve(
-                    intent,
-                    config_path=args.config,
-                    require_approval=args.approve,
+            if args.command_arg == "status" and not args.topic:
+                _run_evolve_status()
+            else:
+                intent = args.topic or args.command_arg
+                if not intent:
+                    raise ValueError(
+                        'evolve requires an intent: ai-agent evolve "…" '
+                        "(or: ai-agent evolve status)"
+                    )
+                asyncio.run(
+                    _run_evolve(
+                        intent,
+                        config_path=args.config,
+                        require_approval=args.approve,
+                    )
                 )
-            )
         elif args.command == "evolve-worker":
             _run_evolve_worker(
                 auto_merge=args.auto_merge,

src/ai_agent/features/evolve/service.py

@@ -33,19 +33,21 @@
 
 Interpret the intent carefully:
 - Doc/README intents → EDIT README.md (allowlisted). Never open `.ai_agent/**`.
-- Make ONE meaningful edit that matches the intent (not a tiny noop).
+- Feature intents must be COMPLETE before publish — not a stub helper alone.
 
 Strict pipeline — CALL TOOLS in order; do not narrate:
-1. Survey: workspace_search action=search (or read with start_line/end_line).
-2. Edit ONCE with replace_in_file (preferred). Do not call replace_in_file again
-   after it succeeds.
-3. VERIFY: git_diff (must be NON-EMPTY) then run_checks preset=pytest.
+1. Survey briefly: at most 2 workspace_search/workspace_list calls, then STOP surveying.
+2. Implement the FULL intent with up to 3 related edits (write_file/replace_in_file).
+   Example for a new CLI: helper module + CLI wiring + one test (and README row if asked).
+   A helper file with no CLI/test is NOT done — keep editing until the feature is usable.
+3. VERIFY: git_diff (must be NON-EMPTY and cover the real change) then run_checks preset=pytest.
 4. PUBLISH: git_commit alone (feature branch), then open_pull_request alone.
 5. Respond with the PR URL only after open_pull_request succeeds.
 
 If the requested text is already present and git_status is clean, respond that
 the intent is already satisfied (no PR).
 Never push to main. Never edit PathPolicy / MergePolicy / STOP / .env.
+Never treat the repo homepage URL as a PR URL.
 """
 
 
@@ -140,10 +142,16 @@ def is_stopped(*, root: Path = DEFAULT_EVOLVE_ROOT) -> bool:
     return (root / "STOP").is_file()
 
 
-def pipeline_progress(tool_results: list[dict[str, object]]) -> PipelineProgress:
-    edited = any(
-        item.get("tool_name") in EDIT_TOOLS and item.get("success") for item in tool_results
+def pipeline_progress(
+    tool_results: list[dict[str, object]],
+    *,
+    min_edits: int = 1,
+) -> PipelineProgress:
+    edit_count = sum(
+        1 for item in tool_results if item.get("tool_name") in EDIT_TOOLS and item.get("success")
     )
+    # Feature intents may need several related files before verify/publish.
+    edited = edit_count >= max(1, min_edits)
     diff_seen = False
     for item in tool_results:
         if item.get("tool_name") != "git_diff" or not item.get("success"):
@@ -161,6 +169,9 @@ def pipeline_progress(tool_results: list[dict[str, object]]) -> PipelineProgress
         item.get("tool_name") == "open_pull_request" and item.get("success")
         for item in tool_results
     )
+    # Once verify/publish starts, treat edit phase as complete.
+    if diff_seen or checks_ok or committed or pr_opened:
+        edited = True
     return PipelineProgress(
         edited=edited,
         diff_seen=diff_seen or committed or pr_opened,
@@ -170,8 +181,74 @@ def pipeline_progress(tool_results: list[dict[str, object]]) -> PipelineProgress
     )
 
 
-def continue_prompt_for(progress: PipelineProgress) -> str:
+def intent_min_edits(intent: str) -> int:
+    """Feature-style intents need more than a stub helper before publish."""
+    text = intent.lower()
+    markers = (
+        "cli",
+        "test",
+        "readme",
+        "wire",
+        "status",
+        "helper",
+        "unit test",
+        "modes table",
+    )
+    if sum(1 for m in markers if m in text) >= 2:
+        return 2
+    return 1
+
+
+def publish_block_reason(
+    intent: str,
+    *,
+    workspace: Path = Path("."),
+) -> str | None:
+    """
+    Block verify/publish when a CLI feature is still a stub.
+
+    Detects unused ``render_status`` imports and missing ``evolve status`` wiring.
+    """
+    text = intent.lower()
+    wants_cli = "cli" in text or ("evolve" in text and "status" in text)
+    if not wants_cli:
+        return None
+
+    cli_path = workspace / "src" / "ai_agent" / "entrypoints" / "cli.py"
+    if not cli_path.is_file():
+        return "cli.py missing — implement the command before publish"
+    body = cli_path.read_text(encoding="utf-8")
+
+    if "render_status" in body and "render_status(" not in body:
+        return (
+            "cli.py imports render_status but never calls it — "
+            "wire `ai-agent evolve status` before publish"
+        )
+    if "status" in text and "_run_evolve_status" not in body and '== "status"' not in body:
+        return (
+            "cli.py has no evolve status handler — "
+            "add `evolve status` command wiring before publish"
+        )
+    if "test" in text:
+        tests_dir = workspace / "tests" / "features" / "evolve"
+        has_status_test = (tests_dir / "test_status.py").is_file()
+        if not has_status_test:
+            return "missing tests/features/evolve/test_status.py — add a test before publish"
+    return None
+
+
+def continue_prompt_for(
+    progress: PipelineProgress,
+    *,
+    block_reason: str | None = None,
+) -> str:
     """State-aware nudge so evolve does not re-edit forever."""
+    if block_reason:
+        return (
+            f"CONTINUE: publish blocked — {block_reason}. "
+            "Edit the missing wiring/tests with replace_in_file or write_file NOW. "
+            "Do not git_commit or open_pull_request yet."
+        )
     action = progress.next_action
     if action == "done":
         return "PR already opened. Respond with the PR URL only."
@@ -192,12 +269,14 @@ def continue_prompt_for(progress: PipelineProgress) -> str:
         )
     if action == "git_diff":
         return (
-            "CONTINUE: edit already succeeded. Call git_diff then run_checks. "
-            "Do NOT call replace_in_file/write_file/apply_patch again."
+            "CONTINUE: if the intent still needs CLI wiring or a test, edit those "
+            "files NOW (up to 3 total edits). Otherwise call git_diff then run_checks. "
+            "Do not open a PR for a stub helper alone."
         )
     return (
-        "CONTINUE: make the intent edit with replace_in_file ONCE, then git_diff, "
-        "run_checks, git_commit, open_pull_request. Do not narrate."
+        "CONTINUE: STOP surveying. Implement the FULL intent with write_file/"
+        "replace_in_file (helper + CLI + test if the intent asks). Then git_diff, "
+        "run_checks, git_commit, open_pull_request. Do not publish stubs."
     )
 
 
@@ -254,6 +333,7 @@ async def run_evolve(
     collected: list[dict[str, object]] = []
     last_result: StepResult | None = None
     turns = max(1, max_continue_turns)
+    min_edits = intent_min_edits(cleaned)
 
     try:
         for turn in range(turns):
@@ -265,8 +345,20 @@ async def run_evolve(
                 emit_ops_event(name="evolve.stopped", run_id=rid, success=False)
                 return run
 
-            progress = pipeline_progress(collected)
-            user_input = prompt if turn == 0 else continue_prompt_for(progress)
+            progress = pipeline_progress(collected, min_edits=min_edits)
+            block = None if progress.pr_opened else publish_block_reason(cleaned)
+            if block and progress.next_action != "edit":
+                # Force another edit cycle instead of verify/publish on stubs.
+                progress = PipelineProgress(
+                    edited=False,
+                    diff_seen=False,
+                    checks_ok=False,
+                    committed=False,
+                    pr_opened=False,
+                )
+            user_input = (
+                prompt if turn == 0 else continue_prompt_for(progress, block_reason=block)
+            )
             if turn > 0:
                 status_by_next: dict[
                     str,
@@ -291,7 +383,23 @@ async def run_evolve(
             result = await agent.step(session=session, user_input=user_input)
             last_result = result
             collected.extend(result.tool_results or [])
-            progress = pipeline_progress(collected)
+            progress = pipeline_progress(collected, min_edits=min_edits)
+            block_after = None if progress.pr_opened else publish_block_reason(cleaned)
+            if block_after and progress.next_action in {
+                "git_diff",
+                "run_checks",
+                "git_commit",
+                "open_pull_request",
+                "done",
+            }:
+                # Stub still incomplete — keep evolving even if model raced to verify.
+                progress = PipelineProgress(
+                    edited=False,
+                    diff_seen=False,
+                    checks_ok=False,
+                    committed=False,
+                    pr_opened=False,
+                )
             logger.info(
                 "evolve_turn run_id=%s turn=%s kind=%s next=%s msg=%s",
                 rid,
@@ -316,9 +424,12 @@ async def run_evolve(
                 raise RuntimeError(result.message)
 
             pr_url = _extract_pr_url(result.message, collected)
-            if pr_url or progress.pr_opened:
+            # Only finish when open_pull_request actually succeeded — never
+            # treat a random github.com link from README/survey text as a PR.
+            stub = publish_block_reason(cleaned)
+            if progress.pr_opened and pr_url and not stub:
                 latency_ms = int((datetime.now(UTC) - started).total_seconds() * 1000)
-                run.pr_url = pr_url or _extract_pr_url("", collected)
+                run.pr_url = pr_url
                 run.status = "done"
                 run.error = None
                 run.last_check_log = _last_check_log(collected)
@@ -355,7 +466,7 @@ async def run_evolve(
     assert last_result is not None
     latency_ms = int((datetime.now(UTC) - started).total_seconds() * 1000)
     _write_result(artifact_dir, last_result, collected)
-    progress = pipeline_progress(collected)
+    progress = pipeline_progress(collected, min_edits=min_edits)
     if progress.committed:
         run.status = "awaiting_approval"
         run.error = "commit succeeded but open_pull_request did not run or failed"
@@ -448,14 +559,15 @@ def _tool_succeeded(tool_results: list[dict[str, object]] | None, name: str) ->
 
 
 def _extract_pr_url(message: str, tool_results: list[dict[str, object]] | None) -> str | None:
+    """Accept only URLs from a successful open_pull_request tool result."""
+    _ = message  # never scrape free-form assistant text (README links false-positive)
     for item in tool_results or []:
-        if item.get("tool_name") == "open_pull_request" and item.get("success"):
-            out = str(item.get("output") or "")
-            if out.startswith("http"):
-                return out.strip()
-    match = re.search(r"https://github\.com/[^\s)]+", message)
-    if match:
-        return match.group(0)
+        if item.get("tool_name") != "open_pull_request" or not item.get("success"):
+            continue
+        out = str(item.get("output") or "").strip()
+        if "/pull/" in out and out.startswith("http"):
+            # First URL token only
+            return out.split()[0].rstrip("\"')].>")
     return None