Skip to content

Commit 4e7ca1b

Browse files
pythonbytepythonbyte
authored
Merge pull request #5 from pythonbyte/agent/evolve-status-cli
Complete evolve status CLI
2 parents 461bf40 + 3310613 commit 4e7ca1b

9 files changed

Lines changed: 393 additions & 54 deletions

File tree

‎README.md‎

Lines changed: 1 addition & 0 deletions
Original file line numberDiff line numberDiff line change
@@ -130,6 +130,7 @@ ai-agent -c config/agent_config.yaml
130130
| Mode | Command | Notes |
131131
|---|---|---|
132132
| Console | `uv run ai-agent -c config/agent_config.yaml` | Single agent |
133+
| Evolve status | `uv run ai-agent evolve status` | Organism queue, budgets, last run/PR |
133134
| Research operator | `uv run ai-agent -c config/agents/operator.yaml` | Interactive Research Desk |
134135
| One-shot brief | `uv run ai-agent brief "agent harness"` | Writes `briefs/YYYYMMDD_slug.md` |
135136
| Brief + approve | `uv run ai-agent brief "topic" --approve` | Console Y/n before write |

‎config/agents/engineer.yaml‎

Lines changed: 5 additions & 3 deletions
Original file line numberDiff line numberDiff line change
@@ -6,8 +6,9 @@ system_prompt: >
66
Workflow for each intent:
77
1. Survey with workspace_search (search, or read with start_line/end_line).
88
Do not open `.ai_agent/**` for doc intents.
9-
2. Edit ONCE with replace_in_file (preferred for README bullets). After it
10-
succeeds, NEVER call replace_in_file/write_file/apply_patch again.
9+
2. Implement the FULL intent with up to 3 related edits (helper + CLI + test
10+
when asked). A stub helper alone is NOT done — keep editing until usable.
11+
After the feature is complete, NEVER call edit tools again.
1112
3. VERIFY immediately: git_diff (must be NON-EMPTY) then run_checks preset=pytest.
1213
4. PUBLISH in SEPARATE rounds: git_commit alone, then open_pull_request alone.
1314
Commit/PR are auto-approved unless the operator enabled HITL (--approve).
@@ -17,7 +18,8 @@ system_prompt: >
1718
Allowlist only: src/, tests/, config/, docs/, README/DECISIONS/AGENTS.
1819
Never edit path_policy, merge policy kernel, STOP, or .env.
1920
Never invent tool results. Never narrate-and-stop mid-pipeline.
20-
max_tool_rounds: 16
21+
Survey with at most 2-3 searches, then edit — do not burn rounds on search.
22+
max_tool_rounds: 24
2123
personality:
2224
tone: pragmatic
2325
style: concise

‎config/evolve_backlog.yaml‎

Lines changed: 22 additions & 7 deletions
Original file line numberDiff line numberDiff line change
@@ -1,8 +1,23 @@
1-
# Rotating intents used when the organism goal queue is empty.
2-
# Keep items small, allowlisted (src/tests/config/docs/README*), never kernel edits.
1+
# Rotating intents when the organism goal queue is empty.
2+
# Prefer concrete harness features over one-line doc tweaks.
33
goals:
4-
- "Add or improve one unit test under tests/ for MergePolicy day-budget behavior."
5-
- "Clarify Phase 2 evolve-worker usage in README.md with one short paragraph."
6-
- "Add a brief AGENTS.md note about evolve-goal add/list and the STOP kill switch."
7-
- "Improve DECISIONS.md with one sentence on goal-queue vs free-form invent intents."
8-
- "Add or tighten one docstring in src/ai_agent/features/evolve/organism.py without behavior changes."
4+
- >-
5+
Add `ai-agent evolve status` CLI: print organism id, stopped flag, goal queue,
6+
evolves_today/max_evolves_per_day, merges_today, next_wake_at, last_run_id,
7+
and if last run exists print its status/pr_url/intent from run.json.
8+
Implement in src/ai_agent/entrypoints/cli.py (+ thin helper under
9+
features/evolve if needed). Add a focused unit test. Update the Modes table
10+
in README.md with one row for evolve status. Do not edit PathPolicy/STOP/.env.
11+
- >-
12+
Add `ai-agent ops summary` that reads OpsEvent JSONL and prints counts by
13+
event name plus success rate and average latency_ms when present. Keep it
14+
under src/ai_agent/harness/ops_metrics.py + CLI wiring + one test. Update
15+
README Modes/ops row briefly.
16+
- >-
17+
Add an engineer tool `evolve_inspect` (read-only) that returns JSON for the
18+
current organism + last EvolveRun summary under .ai_agent/evolve/. Register
19+
it for the engineer agent YAML. Include a unit test. No writes, no git.
20+
- >-
21+
Fix the broken Agent harness capabilities markdown table in README.md
22+
(the Evolve CLI prose currently splits the table) so all capability rows
23+
render as one table again, without changing meaning.

‎src/ai_agent/entrypoints/cli.py‎

Lines changed: 20 additions & 9 deletions
Original file line numberDiff line numberDiff line change
@@ -29,6 +29,7 @@
2929
from ai_agent.features.evolve.goals import enqueue_goal, list_goals
3030
from ai_agent.features.evolve.organism import ensure_organism, worker_loop, worker_tick
3131
from ai_agent.features.evolve.service import run_evolve, save_organism
32+
from ai_agent.features.evolve.status import build_status_state, render_status
3233
from ai_agent.features.harness_bank.bank import (
3334
admit_if_screened,
3435
list_cells,
@@ -382,6 +383,10 @@ def _run_evolve_goal(args: argparse.Namespace) -> None:
382383
raise ValueError("evolve-goal requires: add <text> | list")
383384

384385

386+
def _run_evolve_status() -> None:
387+
print(render_status(build_status_state()))
388+
389+
385390
def _run_harness_bank(args: argparse.Namespace) -> None:
386391
action = args.command_arg
387392
if action == "list":
@@ -629,16 +634,22 @@ def main(argv: list[str] | None = None) -> None:
629634
args.message = args.command_arg2 or args.topic or ""
630635
_run_harness_command(args)
631636
elif args.command == "evolve":
632-
intent = args.topic or args.command_arg
633-
if not intent:
634-
raise ValueError('evolve requires an intent: ai-agent evolve "…"')
635-
asyncio.run(
636-
_run_evolve(
637-
intent,
638-
config_path=args.config,
639-
require_approval=args.approve,
637+
if args.command_arg == "status" and not args.topic:
638+
_run_evolve_status()
639+
else:
640+
intent = args.topic or args.command_arg
641+
if not intent:
642+
raise ValueError(
643+
'evolve requires an intent: ai-agent evolve "…" '
644+
"(or: ai-agent evolve status)"
645+
)
646+
asyncio.run(
647+
_run_evolve(
648+
intent,
649+
config_path=args.config,
650+
require_approval=args.approve,
651+
)
640652
)
641-
)
642653
elif args.command == "evolve-worker":
643654
_run_evolve_worker(
644655
auto_merge=args.auto_merge,

‎src/ai_agent/features/evolve/service.py‎

Lines changed: 138 additions & 26 deletions
Original file line numberDiff line numberDiff line change
@@ -33,19 +33,21 @@
3333
3434
Interpret the intent carefully:
3535
- Doc/README intents → EDIT README.md (allowlisted). Never open `.ai_agent/**`.
36-
- Make ONE meaningful edit that matches the intent (not a tiny noop).
36+
- Feature intents must be COMPLETE before publish — not a stub helper alone.
3737
3838
Strict pipeline — CALL TOOLS in order; do not narrate:
39-
1. Survey: workspace_search action=search (or read with start_line/end_line).
40-
2. Edit ONCE with replace_in_file (preferred). Do not call replace_in_file again
41-
after it succeeds.
42-
3. VERIFY: git_diff (must be NON-EMPTY) then run_checks preset=pytest.
39+
1. Survey briefly: at most 2 workspace_search/workspace_list calls, then STOP surveying.
40+
2. Implement the FULL intent with up to 3 related edits (write_file/replace_in_file).
41+
Example for a new CLI: helper module + CLI wiring + one test (and README row if asked).
42+
A helper file with no CLI/test is NOT done — keep editing until the feature is usable.
43+
3. VERIFY: git_diff (must be NON-EMPTY and cover the real change) then run_checks preset=pytest.
4344
4. PUBLISH: git_commit alone (feature branch), then open_pull_request alone.
4445
5. Respond with the PR URL only after open_pull_request succeeds.
4546
4647
If the requested text is already present and git_status is clean, respond that
4748
the intent is already satisfied (no PR).
4849
Never push to main. Never edit PathPolicy / MergePolicy / STOP / .env.
50+
Never treat the repo homepage URL as a PR URL.
4951
"""
5052

5153

@@ -140,10 +142,16 @@ def is_stopped(*, root: Path = DEFAULT_EVOLVE_ROOT) -> bool:
140142
return (root / "STOP").is_file()
141143

142144

143-
def pipeline_progress(tool_results: list[dict[str, object]]) -> PipelineProgress:
144-
edited = any(
145-
item.get("tool_name") in EDIT_TOOLS and item.get("success") for item in tool_results
145+
def pipeline_progress(
146+
tool_results: list[dict[str, object]],
147+
*,
148+
min_edits: int = 1,
149+
) -> PipelineProgress:
150+
edit_count = sum(
151+
1 for item in tool_results if item.get("tool_name") in EDIT_TOOLS and item.get("success")
146152
)
153+
# Feature intents may need several related files before verify/publish.
154+
edited = edit_count >= max(1, min_edits)
147155
diff_seen = False
148156
for item in tool_results:
149157
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
161169
item.get("tool_name") == "open_pull_request" and item.get("success")
162170
for item in tool_results
163171
)
172+
# Once verify/publish starts, treat edit phase as complete.
173+
if diff_seen or checks_ok or committed or pr_opened:
174+
edited = True
164175
return PipelineProgress(
165176
edited=edited,
166177
diff_seen=diff_seen or committed or pr_opened,
@@ -170,8 +181,74 @@ def pipeline_progress(tool_results: list[dict[str, object]]) -> PipelineProgress
170181
)
171182

172183

173-
def continue_prompt_for(progress: PipelineProgress) -> str:
184+
def intent_min_edits(intent: str) -> int:
185+
"""Feature-style intents need more than a stub helper before publish."""
186+
text = intent.lower()
187+
markers = (
188+
"cli",
189+
"test",
190+
"readme",
191+
"wire",
192+
"status",
193+
"helper",
194+
"unit test",
195+
"modes table",
196+
)
197+
if sum(1 for m in markers if m in text) >= 2:
198+
return 2
199+
return 1
200+
201+
202+
def publish_block_reason(
203+
intent: str,
204+
*,
205+
workspace: Path = Path("."),
206+
) -> str | None:
207+
"""
208+
Block verify/publish when a CLI feature is still a stub.
209+
210+
Detects unused ``render_status`` imports and missing ``evolve status`` wiring.
211+
"""
212+
text = intent.lower()
213+
wants_cli = "cli" in text or ("evolve" in text and "status" in text)
214+
if not wants_cli:
215+
return None
216+
217+
cli_path = workspace / "src" / "ai_agent" / "entrypoints" / "cli.py"
218+
if not cli_path.is_file():
219+
return "cli.py missing — implement the command before publish"
220+
body = cli_path.read_text(encoding="utf-8")
221+
222+
if "render_status" in body and "render_status(" not in body:
223+
return (
224+
"cli.py imports render_status but never calls it — "
225+
"wire `ai-agent evolve status` before publish"
226+
)
227+
if "status" in text and "_run_evolve_status" not in body and '== "status"' not in body:
228+
return (
229+
"cli.py has no evolve status handler — "
230+
"add `evolve status` command wiring before publish"
231+
)
232+
if "test" in text:
233+
tests_dir = workspace / "tests" / "features" / "evolve"
234+
has_status_test = (tests_dir / "test_status.py").is_file()
235+
if not has_status_test:
236+
return "missing tests/features/evolve/test_status.py — add a test before publish"
237+
return None
238+
239+
240+
def continue_prompt_for(
241+
progress: PipelineProgress,
242+
*,
243+
block_reason: str | None = None,
244+
) -> str:
174245
"""State-aware nudge so evolve does not re-edit forever."""
246+
if block_reason:
247+
return (
248+
f"CONTINUE: publish blocked — {block_reason}. "
249+
"Edit the missing wiring/tests with replace_in_file or write_file NOW. "
250+
"Do not git_commit or open_pull_request yet."
251+
)
175252
action = progress.next_action
176253
if action == "done":
177254
return "PR already opened. Respond with the PR URL only."
@@ -192,12 +269,14 @@ def continue_prompt_for(progress: PipelineProgress) -> str:
192269
)
193270
if action == "git_diff":
194271
return (
195-
"CONTINUE: edit already succeeded. Call git_diff then run_checks. "
196-
"Do NOT call replace_in_file/write_file/apply_patch again."
272+
"CONTINUE: if the intent still needs CLI wiring or a test, edit those "
273+
"files NOW (up to 3 total edits). Otherwise call git_diff then run_checks. "
274+
"Do not open a PR for a stub helper alone."
197275
)
198276
return (
199-
"CONTINUE: make the intent edit with replace_in_file ONCE, then git_diff, "
200-
"run_checks, git_commit, open_pull_request. Do not narrate."
277+
"CONTINUE: STOP surveying. Implement the FULL intent with write_file/"
278+
"replace_in_file (helper + CLI + test if the intent asks). Then git_diff, "
279+
"run_checks, git_commit, open_pull_request. Do not publish stubs."
201280
)
202281

203282

@@ -254,6 +333,7 @@ async def run_evolve(
254333
collected: list[dict[str, object]] = []
255334
last_result: StepResult | None = None
256335
turns = max(1, max_continue_turns)
336+
min_edits = intent_min_edits(cleaned)
257337

258338
try:
259339
for turn in range(turns):
@@ -265,8 +345,20 @@ async def run_evolve(
265345
emit_ops_event(name="evolve.stopped", run_id=rid, success=False)
266346
return run
267347

268-
progress = pipeline_progress(collected)
269-
user_input = prompt if turn == 0 else continue_prompt_for(progress)
348+
progress = pipeline_progress(collected, min_edits=min_edits)
349+
block = None if progress.pr_opened else publish_block_reason(cleaned)
350+
if block and progress.next_action != "edit":
351+
# Force another edit cycle instead of verify/publish on stubs.
352+
progress = PipelineProgress(
353+
edited=False,
354+
diff_seen=False,
355+
checks_ok=False,
356+
committed=False,
357+
pr_opened=False,
358+
)
359+
user_input = (
360+
prompt if turn == 0 else continue_prompt_for(progress, block_reason=block)
361+
)
270362
if turn > 0:
271363
status_by_next: dict[
272364
str,
@@ -291,7 +383,23 @@ async def run_evolve(
291383
result = await agent.step(session=session, user_input=user_input)
292384
last_result = result
293385
collected.extend(result.tool_results or [])
294-
progress = pipeline_progress(collected)
386+
progress = pipeline_progress(collected, min_edits=min_edits)
387+
block_after = None if progress.pr_opened else publish_block_reason(cleaned)
388+
if block_after and progress.next_action in {
389+
"git_diff",
390+
"run_checks",
391+
"git_commit",
392+
"open_pull_request",
393+
"done",
394+
}:
395+
# Stub still incomplete — keep evolving even if model raced to verify.
396+
progress = PipelineProgress(
397+
edited=False,
398+
diff_seen=False,
399+
checks_ok=False,
400+
committed=False,
401+
pr_opened=False,
402+
)
295403
logger.info(
296404
"evolve_turn run_id=%s turn=%s kind=%s next=%s msg=%s",
297405
rid,
@@ -316,9 +424,12 @@ async def run_evolve(
316424
raise RuntimeError(result.message)
317425

318426
pr_url = _extract_pr_url(result.message, collected)
319-
if pr_url or progress.pr_opened:
427+
# Only finish when open_pull_request actually succeeded — never
428+
# treat a random github.com link from README/survey text as a PR.
429+
stub = publish_block_reason(cleaned)
430+
if progress.pr_opened and pr_url and not stub:
320431
latency_ms = int((datetime.now(UTC) - started).total_seconds() * 1000)
321-
run.pr_url = pr_url or _extract_pr_url("", collected)
432+
run.pr_url = pr_url
322433
run.status = "done"
323434
run.error = None
324435
run.last_check_log = _last_check_log(collected)
@@ -355,7 +466,7 @@ async def run_evolve(
355466
assert last_result is not None
356467
latency_ms = int((datetime.now(UTC) - started).total_seconds() * 1000)
357468
_write_result(artifact_dir, last_result, collected)
358-
progress = pipeline_progress(collected)
469+
progress = pipeline_progress(collected, min_edits=min_edits)
359470
if progress.committed:
360471
run.status = "awaiting_approval"
361472
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) ->
448559

449560

450561
def _extract_pr_url(message: str, tool_results: list[dict[str, object]] | None) -> str | None:
562+
"""Accept only URLs from a successful open_pull_request tool result."""
563+
_ = message # never scrape free-form assistant text (README links false-positive)
451564
for item in tool_results or []:
452-
if item.get("tool_name") == "open_pull_request" and item.get("success"):
453-
out = str(item.get("output") or "")
454-
if out.startswith("http"):
455-
return out.strip()
456-
match = re.search(r"https://github\.com/[^\s)]+", message)
457-
if match:
458-
return match.group(0)
565+
if item.get("tool_name") != "open_pull_request" or not item.get("success"):
566+
continue
567+
out = str(item.get("output") or "").strip()
568+
if "/pull/" in out and out.startswith("http"):
569+
# First URL token only
570+
return out.split()[0].rstrip("\"')].>")
459571
return None
460572

461573

0 commit comments

Comments
 (0)