agentic-dev-loop 0.15.2__tar.gz → 0.15.4__tar.gz

{agentic_dev_loop-0.15.2/src/agentic_dev_loop.egg-info → agentic_dev_loop-0.15.4}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: agentic-dev-loop
-Version: 0.15.2
+Version: 0.15.4
 Summary: Turn an LLM into a gated, sandboxed, resumable construction loop over your codebase.
 Author: Luke Wilson
 License-Expression: Apache-2.0
@@ -29,6 +29,8 @@ Requires-Dist: pydantic>=2.0
 Requires-Dist: PyYAML>=6.0
 Requires-Dist: openhands-sdk~=1.17.0
 Requires-Dist: openhands-tools~=1.17.0
+Requires-Dist: browser-use==0.11.13
+Requires-Dist: pillow>=13.0.0
 Requires-Dist: agent-client-protocol<1.0.0
 Requires-Dist: langgraph>=1.2
 Requires-Dist: langgraph-checkpoint-sqlite>=3.0
@@ -560,10 +562,11 @@ Conventions that make a good agent file:
 - **dev enumerates its impossible set and stops at it.** List the commands
   the sandbox cannot run (offline package fetches, browser downloads), one
   attempt per command - keeps the productive loop and kills doomed retries.
-- **Failure output returns as *Previous attempt feedback*.** Steps route
-  `on_fail: { goto: dev, retry: true }`, so gate failure output lands in the
-  dev agent's next context; the dev prompt makes every listed issue a
-  required fix.
+- **Failure output is a required fix in dev's next context.** Agent verdict
+  steps land their failing output under *Previous attempt feedback*; command
+  gates (test, restore) land it under *Mechanical gate output*, with the exit
+  code under *Verified state*. Steps route `on_fail: { goto: dev, retry: true }`,
+  and the dev prompt makes every listed failure a required fix.
 - **No git mutations, ever.** The `commit` step is the single commit point.
 
 ### The pipeline file

{agentic_dev_loop-0.15.2 → agentic_dev_loop-0.15.4}/README.md

@@ -515,10 +515,11 @@ Conventions that make a good agent file:
 - **dev enumerates its impossible set and stops at it.** List the commands
   the sandbox cannot run (offline package fetches, browser downloads), one
   attempt per command - keeps the productive loop and kills doomed retries.
-- **Failure output returns as *Previous attempt feedback*.** Steps route
-  `on_fail: { goto: dev, retry: true }`, so gate failure output lands in the
-  dev agent's next context; the dev prompt makes every listed issue a
-  required fix.
+- **Failure output is a required fix in dev's next context.** Agent verdict
+  steps land their failing output under *Previous attempt feedback*; command
+  gates (test, restore) land it under *Mechanical gate output*, with the exit
+  code under *Verified state*. Steps route `on_fail: { goto: dev, retry: true }`,
+  and the dev prompt makes every listed failure a required fix.
 - **No git mutations, ever.** The `commit` step is the single commit point.
 
 ### The pipeline file

{agentic_dev_loop-0.15.2 → agentic_dev_loop-0.15.4}/pyproject.toml

@@ -4,7 +4,7 @@ build-backend = "setuptools.build_meta"
 
 [project]
 name = "agentic-dev-loop"
-version = "0.15.2"
+version = "0.15.4"
 description = "Turn an LLM into a gated, sandboxed, resumable construction loop over your codebase."
 readme = "README.md"
 requires-python = ">=3.12"
@@ -36,6 +36,17 @@ dependencies = [
     "PyYAML>=6.0",
     "openhands-sdk ~=1.17.0",
     "openhands-tools ~=1.17.0",
+    # browser-use is a transitive dep (openhands-tools declares
+    # browser-use>=0.8.0, unpinned). 0.12.0+ hard-pin pillow==12.x, and
+    # pillow 12.3.0 carries 39 CVEs (fixed in 13.0.0); pin the last release
+    # with an open pillow constraint (>=11.2.1) so a fresh pip resolve lands
+    # on pillow 13.0.0 and the restore step's pip-audit stays green.
+    "browser-use==0.11.13",
+    # pillow is a transitive dep (browser-use, reportlab). 12.3.0 carries 39
+    # CVEs (fixed in 13.0.0); pin a floor so the restore step's pip-audit
+    # stays green, and so a future browser-use pin back to ==12.x fails the
+    # install loudly instead of silently reintroducing the CVEs.
+    "pillow>=13.0.0",
     # openhands-sdk declares agent-client-protocol>=0.8.1 with NO upper bound;
     # the 1.0.0rc pre-releases drop acp.helpers (breaks sdk 1.17.0 imports).
     # Without this cap, --pre installs resolve to the broken rc. See AGENTS.md

{agentic_dev_loop-0.15.2 → agentic_dev_loop-0.15.4}/src/agentic_dev_loop/engine/brief_coverage.py

@@ -11,11 +11,26 @@ and file-agnostic.
 from __future__ import annotations
 
 import re
+import unicodedata
+
+
+_GLYPH_FOLD = str.maketrans({
+    "\u2010": "-", "\u2011": "-", "\u2012": "-",
+    "\u2013": "-", "\u2014": "-", "\u2015": "-",
+    "\u2018": "'", "\u2019": "'",
+    "\u201c": '"', "\u201d": '"',
+})
+
+
+def _fold_glyphs(text: str) -> str:
+    """Fold compatibility glyphs and collapse whitespace, preserving case."""
+    folded = unicodedata.normalize("NFKC", text).translate(_GLYPH_FOLD)
+    return " ".join(folded.split())
 
 
 def _normalize(text: str) -> str:
-    """Case-fold and collapse whitespace - the matching normal form."""
-    return " ".join(text.split()).casefold()
+    """Case-folded, glyph-folded, whitespace-collapsed matching form."""
+    return _fold_glyphs(text).casefold()
 
 
 def _normalize_line(line: str) -> str:
@@ -23,7 +38,7 @@ def _normalize_line(line: str) -> str:
 
     A compliant digest writes anchors as headings on their own line, wrapped
     in ``**`` and (for the four section headings) prefixed with ``##``. List
-    bullets are tolerated; the whole-line comparison is what makes a mid-line
+    bullets are tolerated; the line-prefix comparison is what makes a mid-line
     prose mention fail."""
     stripped = line.strip().replace("**", "")
     stripped = re.sub(r"^#+\s*", "", stripped)
@@ -67,19 +82,27 @@ def missing_anchors(brief_text: str, anchors: list[str],
     """Return every required item the captured brief fails to cover.
 
     Required section headings (from the configured ``brief_inputs.sections``)
-    and anchors are matched line-anchored (case-insensitive,
-    whitespace-collapsed, ``**`` decoration stripped). Env facts are matched
-    verbatim, exactly as handed to the brief agent. An empty list means the
+    and anchors are matched as line prefixes (case-insensitive,
+    glyph-folded, whitespace-collapsed, decoration stripped): an item is
+    covered when it starts a line, so content may follow on that line. A
+    mid-line mention does not count. Env facts are matched as glyph-folded,
+    whitespace-collapsed, case-sensitive substrings. An empty list means the
     digest passes coverage."""
     missing: list[str] = []
-    lines = {_normalize_line(line) for line in brief_text.splitlines()}
+    lines = [_normalize_line(line) for line in brief_text.splitlines()]
+    lines = [line for line in lines if line]
+
+    def covered(item: str) -> bool:
+        return any(line.startswith(item) for line in lines)
+
     for section in sections:
-        if _normalize(section) not in lines:
+        if not covered(_normalize(section)):
             missing.append(section)
     for anchor in anchors:
-        if anchor not in lines:
+        if not covered(anchor):
             missing.append(anchor)
+    folded_brief = _fold_glyphs(brief_text)
     for fact in env_facts:
-        if fact not in brief_text:
+        if _fold_glyphs(fact) not in folded_brief:
             missing.append(fact)
     return missing

{agentic_dev_loop-0.15.2 → agentic_dev_loop-0.15.4}/src/agentic_dev_loop/engine/brief_prompt.md

@@ -26,21 +26,24 @@ a compact brief that the dev and review agents reuse on every attempt.
   attempt. Do not invent them and do not re-derive them.
 - The repo-rules digest is structural: one line per top-level section of each
   repo-rule input. Markdown anchors are the `##` headings: keep each heading
-  verbatim, then compress that section's rules into the one line that follows
-  it. YAML anchors are each `- name:` entry under `categories:` plus the
+  verbatim at the start of its own line, then compress that section's rules
+  into the text that follows on that line (after a colon) or on the next
+  line. YAML anchors are each `- name:` entry under `categories:` plus the
   `enforcement_levels` preamble: keep each name verbatim. No top-level section
   or category may be dropped; merging is allowed only between a section and
-  its own subsections. The anchor line holds the heading and nothing else; the
-  compressed rules go on the following line:
+  its own subsections:
 
   - general_principles
     Simplest thing that works; defer decisions to the last responsible moment.
 
-  A line like `general_principles: Simplest thing that works; ...` does not
-  satisfy coverage; the heading must be the only text on its line.
+  A line like `general_principles: Simplest thing that works; ...` passes too:
+  the heading must start its line.
 - Keep the environment-facts entries verbatim: each fact and the behavior it
   implies travel together.
-- ASCII only. Do not use em dash characters.
+- ASCII only, except that a required heading, env fact, or configured section
+  heading must be reproduced character-exact: when such an item contains a
+  non-ASCII character, keep that exact character, never transliterate it. Do
+  not introduce em dashes or other non-ASCII characters of your own.
 - If the task message carries a `## Coverage retry` section, regenerate the
   brief so every listed item appears, verbatim where the contract requires
   verbatim. No exploration: the listed items are the only correction signal.

{agentic_dev_loop-0.15.2 → agentic_dev_loop-0.15.4}/src/agentic_dev_loop/engine/brief_prompt.py

@@ -15,7 +15,7 @@ from __future__ import annotations
 from importlib.resources import files
 
 BRIEF_PROMPT_RESOURCE = "brief_prompt.md"
-BRIEF_PROMPT_VERSION = "1"
+BRIEF_PROMPT_VERSION = "2"
 
 
 def load_template() -> str:
@@ -36,9 +36,9 @@ def _output_block(sections: list[str]) -> str:
         "compressed content for that section (write `None` when a section has "
         "nothing):\n\n"
         f"{headings}\n\n"
-        f"The section headings must appear exactly as written above: {names}. "
-        "Do not shorten or rename them, and do not merge a section with "
-        "another section."
+        "The section headings must appear at the start of their own lines, as "
+        f"written above: {names}. Do not shorten or rename them, and do not "
+        "merge a section with another section."
     )
 
 

{agentic_dev_loop-0.15.2 → agentic_dev_loop-0.15.4}/src/agentic_dev_loop/engine/onboarding/templates/dev.agent.md

@@ -10,10 +10,11 @@ opinionated LangGraph + OpenHands SDK pipeline engine written in Python.
 ## WORKFLOW
 
 1. Read the plan documents and README.md in the context below.
-2. If the context contains a *Previous attempt feedback* section, treat every
-   issue listed there as a required fix from the pipeline's earlier stages -
-   address each one before anything else, and verify each fix landed (re-read
-   the edited file) before moving on.
+2. If the context contains a *Previous attempt feedback* section or a
+   *Mechanical gate output* section, treat every issue listed there as a
+   required fix from the pipeline's earlier stages - address each one before
+   anything else, and verify each fix landed (re-read the edited file) before
+   moving on. A mechanical entry names the attempt that produced it.
 3. Check for prior-attempt residue before implementing. The pipeline syncs
    every lane's edits back into this repo's working tree after each terminal
    action, so an earlier attempt that died at a later gate can leave the full
@@ -66,10 +67,23 @@ opinionated LangGraph + OpenHands SDK pipeline engine written in Python.
   exit code alone.
 - You do not need to prove anything here: fill in the verification matrix,
   write your final response, and stop.
-- If the gate fails, you are retried with the failing output in your notes
-  under *Previous attempt feedback*. Find the root cause - a regression in
-  your change, a stale test pinning old behavior, or a genuine library bug -
-  and fix it, never the symptom.
+- If the gate fails, you are retried with the failing output under
+  *Mechanical gate output* and the exit code under *Verified state* (command
+  gates do not write *Previous attempt feedback*). Find the root cause - a
+  regression in your change, a stale test pinning old behavior, or a genuine
+  library bug - and fix it, never the symptom.
+
+## AUDIT GATE
+
+- The restore step re-runs the install and a dependency audit on every pass.
+  A failing audit means a pin in the dependency manifest carries known CVEs;
+  it is a required fix, never "pre-existing" or "unrelated".
+- Fix it with a manifest edit, never a package-manager command (agent steps
+  are offline). Follow the floor-pin pattern: pin the vulnerable resolved
+  package to a CVE-free version, or pin the intermediate package to the last
+  release that admits it, so a fresh resolve lands on the fixed version.
+- Restore re-audits the bumped pins on the next pass, so the loop
+  self-corrects; report the audit as fixed only after that gate passes.
 
 ## TEST-FIRST
 
@@ -81,7 +95,8 @@ opinionated LangGraph + OpenHands SDK pipeline engine written in Python.
 
 - The moment your final response is written AND the full suite has passed
   locally (`.venv-sandbox/bin/python -m pytest tests/ -q` exits 0) AND every
-  *Previous attempt feedback* item is addressed, the step is COMPLETE. Stop.
+  *Previous attempt feedback* and *Mechanical gate output* item is addressed,
+  the step is COMPLETE. Stop.
 - Do not re-read files, re-run checks, or `cat`/`tail`/`grep` files you just
   wrote to confirm what already passed. A green suite is a green suite:
   re-running it buys nothing, since the mechanical test gate re-runs the

{agentic_dev_loop-0.15.2 → agentic_dev_loop-0.15.4}/src/agentic_dev_loop/engine/stages.py

@@ -446,7 +446,10 @@ def build_context(runtime: Runtime, docs: dict[str, str], subject: str,
     own mechanical section (process fact, never an agent claim), and only
     engine-verified facts (command exit codes, brief coverage, commit outcome)
     appear in the verified section. Nothing in the claims section is ever
-    routed on."""
+    routed on. The mechanical heading binds every failure it carries as a
+    required fix: command gates (test, restore) never write the notes channel,
+    so without that binding a retried dev agent has no required-fix trigger
+    (plan 065)."""
     parts = [f"You are working on {subject}.\n", guide_section()]
     for name, content in docs.items():
         parts.append(f"## {name}\n\n{content}")
@@ -470,7 +473,8 @@ def build_context(runtime: Runtime, docs: dict[str, str], subject: str,
         parts.extend(claims)
     if mechanical_notes:
         parts.append("## Mechanical gate output "
-                     "(pipeline-captured process output, not agent claims)")
+                     "(pipeline-captured process output, not agent claims; "
+                     "each failure here is a required fix)")
         parts.extend(mechanical_notes)
     verified_lines = [f"- {name}: {fact}" for name, fact in (verified or {}).items()]
     parts.append(
@@ -678,16 +682,11 @@ def build_brief_context(runtime: Runtime, step: StepConfig,
         parts.append(
             "## Coverage retry\n\n"
             "The captured brief is missing the following required items. "
-            "Regenerate the digest so every listed item occupies its own "
-            "line, exactly as listed: the item alone on the line, no "
-            "trailing colon and no content. Put the compressed content on "
-            "the following line:\n\n"
+            "Regenerate the digest so every listed item starts its own line; "
+            "compressed content may follow on that line or the next line.\n\n"
             "- general_principles\n"
             "  Simplest thing that works; defer decisions to the last "
             "responsible moment.\n\n"
-            "A line like `general_principles: Simplest thing that works; "
-            "...` does not satisfy coverage; the item must be the only "
-            "text on its line:\n\n"
             + coverage_note)
     return "\n\n".join(parts)
 

{agentic_dev_loop-0.15.2 → agentic_dev_loop-0.15.4/src/agentic_dev_loop.egg-info}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: agentic-dev-loop
-Version: 0.15.2
+Version: 0.15.4
 Summary: Turn an LLM into a gated, sandboxed, resumable construction loop over your codebase.
 Author: Luke Wilson
 License-Expression: Apache-2.0
@@ -29,6 +29,8 @@ Requires-Dist: pydantic>=2.0
 Requires-Dist: PyYAML>=6.0
 Requires-Dist: openhands-sdk~=1.17.0
 Requires-Dist: openhands-tools~=1.17.0
+Requires-Dist: browser-use==0.11.13
+Requires-Dist: pillow>=13.0.0
 Requires-Dist: agent-client-protocol<1.0.0
 Requires-Dist: langgraph>=1.2
 Requires-Dist: langgraph-checkpoint-sqlite>=3.0
@@ -560,10 +562,11 @@ Conventions that make a good agent file:
 - **dev enumerates its impossible set and stops at it.** List the commands
   the sandbox cannot run (offline package fetches, browser downloads), one
   attempt per command - keeps the productive loop and kills doomed retries.
-- **Failure output returns as *Previous attempt feedback*.** Steps route
-  `on_fail: { goto: dev, retry: true }`, so gate failure output lands in the
-  dev agent's next context; the dev prompt makes every listed issue a
-  required fix.
+- **Failure output is a required fix in dev's next context.** Agent verdict
+  steps land their failing output under *Previous attempt feedback*; command
+  gates (test, restore) land it under *Mechanical gate output*, with the exit
+  code under *Verified state*. Steps route `on_fail: { goto: dev, retry: true }`,
+  and the dev prompt makes every listed failure a required fix.
 - **No git mutations, ever.** The `commit` step is the single commit point.
 
 ### The pipeline file

{agentic_dev_loop-0.15.2 → agentic_dev_loop-0.15.4}/src/agentic_dev_loop.egg-info/requires.txt

@@ -2,6 +2,8 @@ pydantic>=2.0
 PyYAML>=6.0
 openhands-sdk~=1.17.0
 openhands-tools~=1.17.0
+browser-use==0.11.13
+pillow>=13.0.0
 agent-client-protocol<1.0.0
 langgraph>=1.2
 langgraph-checkpoint-sqlite>=3.0

{agentic_dev_loop-0.15.2 → agentic_dev_loop-0.15.4}/tests/test_041_brief_coverage.py

@@ -142,6 +142,46 @@ class TestMissingAnchors:
                                ["offline; never fetch"], SECTIONS) == []
 
 
+    def test_inline_anchor_content_line_passes(self):
+        text = _brief_ok().replace(
+            "**Error handling**\nNever fall back silently.\n\n",
+            "**Error handling**: Never fall back silently.\n\n")
+        assert missing_anchors(text, ["error handling", "mock policy"],
+                               ["offline; never fetch"], SECTIONS) == []
+
+    def test_em_dash_heading_transliterated_line_passes(self):
+        anchors = extract_anchors(
+            "## Default development mode \u2014 ask before coding\n")
+        assert anchors == ["default development mode - ask before coding"]
+        text = _brief_ok().replace(
+            "**Error handling**\nNever fall back silently.\n\n",
+            "**Default development mode - ask before coding**\n"
+            "Ask before writing code.\n\n")
+        assert missing_anchors(text, anchors, ["offline; never fetch"],
+                               SECTIONS) == []
+
+    def test_curly_quote_heading_folded(self):
+        anchors = extract_anchors("## User\u2019s rules\n")
+        assert anchors == ["user's rules"]
+        text = _brief_ok().replace(
+            "**Error handling**\nNever fall back silently.\n\n",
+            "**User's rules**\nApply them.\n\n")
+        assert missing_anchors(text, anchors, ["offline; never fetch"],
+                               SECTIONS) == []
+
+    def test_env_fact_em_dash_tolerated(self):
+        text = _brief_ok().replace("offline; never fetch",
+                                   "offline - never fetch")
+        assert missing_anchors(text, ["error handling", "mock policy"],
+                               ["offline \u2014 never fetch"], SECTIONS) == []
+
+    def test_env_fact_case_still_matters(self):
+        text = _brief_ok().replace("offline; never fetch",
+                                   "the build is efficient")
+        assert "CI" in missing_anchors(
+            text, ["error handling", "mock policy"], ["CI"], SECTIONS)
+
+
 class TestCoverageNoteWording:
     """The retry note must state the line rule unambiguously (043): a
     correction that the model can misread as "the item starts its own line"
@@ -154,11 +194,12 @@ class TestCoverageNoteWording:
             runtime, runtime.cfg.steps[0], {"plan.md": "THE PLAN"},
             coverage_note=note)
         assert "## Coverage retry" in message
-        assert ("the item alone on the line, no trailing colon and no "
-                "content") in message
+        assert ("Regenerate the digest so every listed item starts its own "
+                "line; compressed content may follow on that line or the next "
+                "line.") in message
         assert "- general_principles\n  Simplest thing that works" in message
-        assert "does not satisfy coverage" in message
-        assert "the item must be the only text on its line" in message
+        assert "the item alone on the line" not in message
+        assert "the only text on its line" not in message
         assert note in message
 
     def test_note_keeps_missing_items_verbatim(self, runtime):

{agentic_dev_loop-0.15.2 → agentic_dev_loop-0.15.4}/tests/test_046_brief_prompt.py

@@ -29,3 +29,11 @@ class TestRenderBriefPrompt:
         assert "`## Scope`" in prompt
         assert "`## Risks`" in prompt
         assert "Do not shorten or rename them" in prompt
+
+    def test_relaxed_contract_is_stated(self):
+        prompt = render_brief_prompt(SECTIONS)
+        assert "at the start of its own line" in prompt
+        assert "must be reproduced character-exact" in prompt
+        assert "never transliterate it" in prompt
+        assert "must appear at the start of their own lines" in prompt
+        assert "the only text on its line" not in prompt

{agentic_dev_loop-0.15.2 → agentic_dev_loop-0.15.4}/tests/test_outputs_pbt.py

@@ -56,8 +56,12 @@ def text_with_single_decision(draw):
 
 @st.composite
 def text_with_conflicting_decisions(draw):
-    first = draw(verdict_line.filter(lambda v: "PASS" in v.split(":", 1)[1]))
-    second = draw(verdict_line.filter(lambda v: "FAIL" in v.split(":", 1)[1]))
+    first = draw(verdict_line.filter(
+        lambda v: "PASS" in v.split(":", 1)[1]
+        and "FAIL" not in v.split(":", 1)[1]))
+    second = draw(verdict_line.filter(
+        lambda v: "FAIL" in v.split(":", 1)[1]
+        and "PASS" not in v.split(":", 1)[1]))
     return f"{first}\n{second}"
 
 

{agentic_dev_loop-0.15.2 → agentic_dev_loop-0.15.4}/tests/test_unit11.py

@@ -8,6 +8,7 @@ instead of reading as a clean pass.
 
 import dataclasses
 import json
+from pathlib import Path
 from types import SimpleNamespace
 
 import pytest
@@ -30,6 +31,7 @@ from agentic_dev_loop.engine.completion import (classify_completion,
                                                 count_actions)
 from agentic_dev_loop.engine.dryrun import _validate_agents
 from agentic_dev_loop.engine.graph import build_graph
+from agentic_dev_loop.engine.onboarding.templates import load_template
 from agentic_dev_loop.engine.provider_wire import (WireShapeError,
                                                    merge_choices,
                                                    validate_finish_shape,
@@ -450,7 +452,8 @@ class TestMechanicalOutputChannel:
                             verified=update["verified_state"])
         claims = "## Stage artifacts (each agent's own claims, not verified)"
         mechanical = ("## Mechanical gate output "
-                      "(pipeline-captured process output, not agent claims)")
+                      "(pipeline-captured process output, not agent claims; "
+                      "each failure here is a required fix)")
         verified = "## Verified state (mechanically checked by the pipeline)"
         assert claims in ctx and mechanical in ctx
         claims_section = ctx[ctx.index(claims):ctx.index(mechanical)]
@@ -463,6 +466,40 @@ class TestMechanicalOutputChannel:
         ctx = build_context(runtime, {}, "s", notes=["agent prose note"])
         assert "## Mechanical gate output" not in ctx
 
+    def test_mechanical_heading_binds_failures_as_required_fixes(self, runtime):
+        _rt, update = self._run_failing_command(runtime)
+        ctx = build_context(runtime, {}, "s",
+                            mechanical_notes=update["mechanical_notes"],
+                            verified=update["verified_state"])
+        heading = next(line for line in ctx.splitlines()
+                       if line.startswith("## Mechanical gate output"))
+        assert "required fix" in heading
+
+
+class TestDevAuditContract:
+    """Plan 065: the dev prompt binds a failing mechanical or audit gate to
+    dev, and the two copies of the contract stay in sync."""
+
+    @staticmethod
+    def _texts():
+        repo = Path(__file__).parent.parent
+        return [
+            (repo / "agentic-dev-loop" / "sdk_agents" / "dev.agent.md")
+            .read_text(encoding="utf-8"),
+            load_template("dev.agent.md"),
+        ]
+
+    def test_contract_names_both_required_fix_channels(self):
+        for text in self._texts():
+            assert "*Mechanical gate output*" in text
+            assert "*Previous attempt feedback*" in text
+            assert "failing output in your notes" not in text
+
+    def test_contract_has_an_audit_gate_section(self):
+        for text in self._texts():
+            assert "## AUDIT GATE" in text
+            assert "floor-pin" in text.lower()
+
 
 class TestVerifiedStateAge:
     """m4: a verified fact carries the attempt that produced it, so attempt 1's