@cryptotaxi247 / SquadScope / commits / 0eabbb6

feat: split weekly analysis into 2-step pipeline (reduce context pressure) (#515)

* feat: split weekly analysis into 2-step pipeline (#514) Step 1 (synthesis): Distills press context, historical context, and continuity capsule into a compact industry narrative (max 2K tokens) using a cheaper model (gpt-4o-mini by default). Step 2 (analysis): Uses crawl JSON + Step 1 narrative + wisdom/skills + previous summary. Replaces the raw press/historical context with the compact synthesis, dramatically reducing token count. Backward-compatible: if Step 1 fails, the workflow falls back to the current single-prompt behavior (no --synthesis-input passed). New CLI flags: --run-synthesis Run Step 1 only and write narrative to file --synthesis-output Where to write Step 1 output --synthesis-input Pre-computed narrative to inject into Step 2 --synthesis-model Model override for Step 1 Closes #514 Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com> * Address all review comments on PR #515 - analyze_fallback.py: escape synthesis-input with _escape_untrusted_boundaries - analyze_fallback.py: cap max_tokens at SYNTHESIS_MAX_TOKENS (2K) - analyze_fallback.py: add canary-token + validate_output_safety to _call_synthesis_api - analyze_fallback.py: default run_synthesis_step budget to SYNTHESIS_PROMPT_TOKEN_BUDGET (20K) - analyze_fallback.py: strip AI instruction blocks from press context before synthesis - analyze_fallback.py: clamp truncation end_index to avoid negative slice - analyze_fallback.py: add tests for --run-synthesis, --synthesis-input, and instruction stripping - crawl-and-publish.yml: use bash array for SYNTHESIS_ARGS to avoid word-splitting - crawl-and-publish.yml: add models:read permission to analyze job Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com> --------- Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>

Juan Manuel Servera committed Jun 16, 2026 at 19:35 UTC 0eabbb63869e8e6d4ee9b1d17cbac59ee6d94dac

3 files changed +474 -6

.github/workflows/crawl-and-publish.yml

+41

@@ -293,6 +293,7 @@ jobs:

293 actions: read

294 contents: write

295 issues: write

296 + models: read

297 outputs:

298 week: ${{ steps.analysis-context.outputs.week }}

299 summary_file: ${{ steps.analysis-context.outputs.published_output_file }}

@@ -451,6 +452,39 @@ jobs:

452 PY

453 echo "press_file=$PRESS_FILE" >> "$GITHUB_OUTPUT"

454

455 + - name: Run synthesis step (Step 1)

456 + id: synthesis

457 + env:

458 + GITHUB_TOKEN: ${{ secrets.GITHUB_TOKEN }}

459 + IN_WEEK_FILE: ${{ steps.analysis-context.outputs.week_file }}

460 + IN_OUTPUT_FILE: ${{ steps.analysis-context.outputs.candidate_output_file }}

461 + IN_CURRENT_DATETIME: ${{ steps.analysis-context.outputs.current_datetime }}

462 + IN_PRESS_FILE: ${{ steps.press-context.outputs.press_file }}

463 + run: |

464 + set -euo pipefail

465 + WEEK_FILE="$IN_WEEK_FILE"

466 + OUTPUT_FILE="$IN_OUTPUT_FILE"

467 + CURRENT_DATETIME="$IN_CURRENT_DATETIME"

468 + PRESS_FILE="$IN_PRESS_FILE"

469 + DIAGNOSTICS_DIR="$(dirname "$OUTPUT_FILE")/diagnostics"

470 + SYNTHESIS_FILE="$DIAGNOSTICS_DIR/synthesis-narrative.md"

471 + mkdir -p "$DIAGNOSTICS_DIR"

472 + # Step 1: synthesize press/historical context into compact narrative.

473 + # If this fails, we fall back to the single-prompt approach (no synthesis_input).

474 + if python3 scripts/analyze_fallback.py \

475 + --raw-json "$WEEK_FILE" \

476 + --output "$OUTPUT_FILE" \

477 + --current-datetime "$CURRENT_DATETIME" \

478 + --press-context "$PRESS_FILE" \

479 + --run-synthesis \

480 + --synthesis-output "$SYNTHESIS_FILE"; then

481 + echo "synthesis_file=$SYNTHESIS_FILE" >> "$GITHUB_OUTPUT"

482 + echo "synthesis_available=true" >> "$GITHUB_OUTPUT"

483 + else

484 + echo "::warning::Synthesis step failed; falling back to single-prompt approach."

485 + echo "synthesis_available=false" >> "$GITHUB_OUTPUT"

486 + fi

487 +

488 - name: Render and preflight analysis prompt

489 id: prompt-preflight

490 env:

@@ -459,6 +493,8 @@ jobs:

493 IN_OUTPUT_FILE: ${{ steps.analysis-context.outputs.candidate_output_file }}

494 IN_CURRENT_DATETIME: ${{ steps.analysis-context.outputs.current_datetime }}

495 IN_PRESS_FILE: ${{ steps.press-context.outputs.press_file }}

496 + IN_SYNTHESIS_AVAILABLE: ${{ steps.synthesis.outputs.synthesis_available }}

497 + IN_SYNTHESIS_FILE: ${{ steps.synthesis.outputs.synthesis_file }}

498 run: |

499 set -euo pipefail

500 WEEK_FILE="$IN_WEEK_FILE"

@@ -475,6 +511,10 @@ jobs:

511 # Hydrate metrics ledger from publish before writing this run's prompt/preflight artifacts.

512 git fetch origin publish 2>/dev/null && \

513 git checkout origin/publish -- data/metrics/ 2>/dev/null || true

514 + SYNTHESIS_ARGS=()

515 + if [ "${IN_SYNTHESIS_AVAILABLE:-false}" = "true" ] && [ -f "${IN_SYNTHESIS_FILE:-}" ]; then

516 + SYNTHESIS_ARGS=(--synthesis-input "$IN_SYNTHESIS_FILE")

517 + fi

518 python3 scripts/analyze_fallback.py \

519 --raw-json "$WEEK_FILE" \

520 --output "$OUTPUT_FILE" \

@@ -483,6 +523,7 @@ jobs:

523 --prompt-token-budget "${ANALYSIS_PROMPT_TOKEN_BUDGET:-90000}" \

524 --preflight-report-json "$PREFLIGHT_JSON" \

525 --preflight-report-md "$PREFLIGHT_MD" \

526 + "${SYNTHESIS_ARGS[@]}" \

527 --print-prompt > "$PROMPT_FILE"

528 cp "$PREFLIGHT_JSON" "$LEGACY_PREFLIGHT_JSON"

529 python3 scripts/preflight_cost_check.py \

scripts/analyze_fallback.py

+306 -6

@@ -45,6 +45,13 @@ COMPACTED_SKILLS_CHARS = 10_000

45 COMPACTED_CONTINUITY_CHARS = 8_000

46 COMPACTED_PRESS_CONTEXT_CHARS = 14_000

47 COMPACTED_HISTORICAL_CONTEXT_CHARS = 12_000

48 +SYNTHESIS_MAX_TOKENS = 2_000

49 +SYNTHESIS_PROMPT_TOKEN_BUDGET = 20_000

50 +DEFAULT_SYNTHESIS_MODEL = "openai/gpt-4o-mini"

51 +RETRYABLE_STATUS_CODES = {429, 500, 502, 503, 504}

52 +NON_RETRYABLE_STATUS_CLASSES = {400, 401, 403, 404}

53 +MAX_RETRIES = 3

54 +BASE_DELAY = 2 # seconds

57 @dataclass

@@ -212,6 +219,29 @@ def parse_args(argv: list[str] | None = None) -> argparse.Namespace:

219 type=Path,

220 help="Write deterministic rendered-prompt preflight details as Markdown.",

221 )

222 + parser.add_argument(

223 + "--run-synthesis",

224 + action="store_true",

225 + help="Run Step 1 synthesis (press/historical context → compact narrative) and exit.",

226 + )

227 + parser.add_argument(

228 + "--synthesis-output",

229 + type=Path,

230 + default=None,

231 + help="Path to write the synthesis narrative output (used with --run-synthesis).",

232 + )

233 + parser.add_argument(

234 + "--synthesis-input",

235 + type=Path,

236 + default=None,

237 + help="Path to a pre-computed synthesis narrative to inject into the analysis prompt (Step 2).",

238 + )

239 + parser.add_argument(

240 + "--synthesis-model",

241 + type=str,

242 + default=None,

243 + help=f"Model to use for synthesis step (default: {DEFAULT_SYNTHESIS_MODEL}).",

244 + )

245 return parser.parse_args(argv)

246

247

@@ -725,6 +755,230 @@ def compact_payload(payload: dict[str, Any]) -> tuple[dict[str, Any], dict[str,

755 return compacted, decisions

756

757

758 +def _strip_ai_instruction_blocks(text: str) -> str:

759 + """Remove AI-only instruction sections (### Instructions, directives) from press context.

760 +

761 + These blocks are intended for the main analysis prompt and should not be

762 + forwarded into synthesis to reduce prompt-injection surface area.

763 + """

764 + # Remove markdown sections starting with ### Instructions (case-insensitive)

765 + # up to the next same-or-higher-level heading or end of text

766 + text = re.sub(

767 + r"(?m)^###\s+Instructions?\b.*?(?=^#{1,3}\s|\Z)",

768 + "",

769 + text,

770 + flags=re.DOTALL | re.IGNORECASE,

771 + )

772 + # Remove divergence directive blocks (commonly marked with special tags)

773 + text = re.sub(

775 + "",

776 + text,

777 + flags=re.DOTALL | re.IGNORECASE,

778 + )

779 + return text.strip()

780 +

781 +

782 +def _build_synthesis_prompt(

783 + *,

784 + press_content: str,

785 + historical_context_content: str,

786 + continuity_content: str,

787 + current_week: str,

788 + current_datetime: str,

789 +) -> str:

790 + """Build a compact prompt for Step 1: Industry & Press Synthesis.

791 +

792 + Input: press context + historical context + continuity capsule.

793 + Output instruction: max 2K token narrative of the tech industry landscape this week.

794 + """

795 + sections = []

796 + sections.append(

797 + "You are an expert technology industry analyst. Your task is to synthesize "

798 + "the provided press context, historical context, and continuity notes into a "

799 + "compact industry narrative (maximum 2000 tokens / ~1500 words).\n\n"

800 + "Focus on:\n"

801 + "- Key technology trends and shifts happening this week\n"

802 + "- Notable industry movements (acquisitions, launches, pivots)\n"

803 + "- Developer ecosystem changes\n"

804 + "- Connections to longer-term patterns from historical context\n\n"

805 + "Output ONLY the narrative — no headers, no metadata, no instructions. "

806 + "Write in a dense, information-rich style suitable for feeding into a downstream "

807 + "analysis step that will correlate this with GitHub repository data.\n\n"

808 + f"Current week: {current_week}\n"

809 + f"Current datetime: {current_datetime}\n"

810 + )

811 + if press_content:

812 + sections.append(f"## Press Context\n\n{press_content}")

813 + if historical_context_content:

814 + sections.append(f"## Historical Context\n\n{historical_context_content}")

815 + if continuity_content and continuity_content != "_No continuity capsule has been recorded yet._":

816 + sections.append(f"## Continuity Notes\n\n{continuity_content}")

817 +

818 + return "\n\n---\n\n".join(sections)

819 +

820 +

821 +def run_synthesis_step(

822 + *,

823 + press_context_path: Path | None = None,

824 + content_root: Path = DEFAULT_CONTENT_ROOT,

825 + continuity_file: Path = DEFAULT_CONTINUITY_FILE,

826 + current_datetime: str,

827 + current_week: str,

828 + previous_summary_path: Path | None = None,

829 + prompt_token_budget: int = SYNTHESIS_PROMPT_TOKEN_BUDGET,

830 + model: str | None = None,

831 +) -> str:

832 + """Execute Step 1: synthesize press/historical context into a compact narrative.

833 +

834 + Returns the narrative string (max ~2K tokens). Raises RuntimeError on API failure.

835 + """

836 + historical_context_content = assemble_historical_context(

837 + current_datetime=current_datetime,

838 + previous_summary_path=previous_summary_path,

839 + content_root=content_root,

840 + max_words=1_500,

841 + prompt_token_budget=prompt_token_budget,

842 + ).strip()

843 + from scripts.sanitize_repo_content import _escape_untrusted_boundaries

844 +

845 + historical_context_content = _escape_untrusted_boundaries(historical_context_content)

846 + if not historical_context_content:

847 + historical_context_content = "_No historical context was available beyond the current weekly payload._"

848 +

849 + continuity_content = render_continuity(continuity_file)

850 +

851 + press_content = (

852 + press_context_path.read_text(encoding="utf-8").strip()

853 + if press_context_path and press_context_path.exists() and press_context_path.stat().st_size > 0

854 + else ""

855 + )

856 +

857 + # Strip AI-only instruction blocks from press context before synthesis

858 + if press_content:

859 + press_content = _strip_ai_instruction_blocks(press_content)

860 + # Escape boundary markers in untrusted press content

861 + from scripts.sanitize_repo_content import _escape_untrusted_boundaries

862 + if press_content:

863 + press_content = _escape_untrusted_boundaries(press_content)

864 +

865 + # If there's no meaningful content to synthesize, return empty

866 + if not press_content and historical_context_content.startswith("_No historical context"):

867 + return ""

868 +

869 + prompt = _build_synthesis_prompt(

870 + press_content=press_content,

871 + historical_context_content=historical_context_content,

872 + continuity_content=continuity_content,

873 + current_week=current_week,

874 + current_datetime=current_datetime,

875 + )

876 +

877 + # Check that synthesis prompt is within its own budget

878 + prompt_tokens = estimate_tokens(prompt)

879 + if prompt_tokens > SYNTHESIS_PROMPT_TOKEN_BUDGET:

880 + # Truncate press content to fit (clamp to avoid negative index)

881 + excess_chars = (prompt_tokens - SYNTHESIS_PROMPT_TOKEN_BUDGET) * 4

882 + end_index = max(0, len(press_content) - excess_chars)

883 + press_content = press_content[:end_index]

884 + prompt = _build_synthesis_prompt(

885 + press_content=press_content,

886 + historical_context_content=historical_context_content,

887 + continuity_content=continuity_content,

888 + current_week=current_week,

889 + current_datetime=current_datetime,

890 + )

891 +

892 + return _call_synthesis_api(prompt, model=model or DEFAULT_SYNTHESIS_MODEL)

893 +

894 +

895 +def _call_synthesis_api(prompt: str, *, model: str) -> str:

896 + """Call GitHub Models API for the synthesis step."""

897 + token = os.environ.get("GITHUB_TOKEN")

898 + if not token:

899 + raise RuntimeError("GITHUB_TOKEN is required for synthesis step.")

900 +

901 + # Inject canary token for output leak detection

902 + from scripts.canary_token import generate_canary, inject_canary

903 + canary = generate_canary()

904 + prompt = inject_canary(prompt, canary)

905 +

906 + endpoint = os.environ.get("GITHUB_MODELS_ENDPOINT", DEFAULT_MODELS_ENDPOINT)

907 + validate_https_url(endpoint, label="GitHub Models endpoint", allowed_hosts=ALLOWED_MODELS_HOSTS)

908 + timeout = int(os.environ.get("GITHUB_MODELS_TIMEOUT", str(DEFAULT_MODELS_TIMEOUT)))

909 + payload = {

910 + "model": model,

911 + "messages": [{"role": "user", "content": prompt}],

912 + "temperature": 0.2,

913 + "max_tokens": SYNTHESIS_MAX_TOKENS, # Cap at documented 2K

914 + }

915 + body = json.dumps(payload).encode("utf-8")

916 +

917 + last_exc: Exception | None = None

918 + for attempt in range(MAX_RETRIES + 1):

919 + req = request.Request(

920 + endpoint,

921 + data=body,

922 + headers={

923 + "Authorization": f"Bearer {token}",

924 + "Content-Type": "application/json",

925 + "Accept": "application/json",

926 + },

927 + method="POST",

928 + )

929 + try:

930 + with request.urlopen(req, timeout=timeout) as response: # nosec B310

931 + response_payload = json.load(response)

932 + markdown = extract_markdown(response_payload)

933 + # Validate output for canary leak and injection artifacts

934 + violations = validate_output_safety(markdown, canary)

935 + if violations:

936 + msg = f"Output safety violations detected: {'; '.join(violations)}"

937 + canary_leaked = any("Canary token leaked" in v for v in violations)

938 + if canary_leaked:

939 + raise RuntimeError(f"BLOCKED: {msg}")

940 + print(f"::warning::{msg}", file=sys.stderr)

941 + return markdown

942 + except error.HTTPError as exc:

943 + if exc.code not in RETRYABLE_STATUS_CODES or attempt == MAX_RETRIES:

944 + detail = exc.read().decode("utf-8", errors="replace")

945 + raise RuntimeError(

946 + f"Synthesis API request failed ({exc.code}): {detail}"

947 + ) from exc

948 + # Respect Retry-After header on 429

949 + retry_after = None

950 + if exc.code == 429:

951 + retry_after_header = exc.headers.get("Retry-After") if exc.headers else None

952 + if retry_after_header:

953 + try:

954 + retry_after = float(retry_after_header)

955 + except (ValueError, TypeError):

956 + pass

957 + delay = retry_after if retry_after else BASE_DELAY ** (attempt + 1) + _JITTER_RANDOM.uniform(0, 1)

958 + print(

959 + f"[retry] Synthesis API returned {exc.code}, "

960 + f"retrying in {delay:.1f}s (attempt {attempt + 1}/{MAX_RETRIES})",

961 + file=sys.stderr,

962 + )

963 + last_exc = exc

964 + time.sleep(delay)

965 + except error.URLError as exc:

966 + if attempt == MAX_RETRIES:

967 + raise RuntimeError(

968 + f"Synthesis API network error: {exc.reason}"

969 + ) from exc

970 + delay = BASE_DELAY ** (attempt + 1) + _JITTER_RANDOM.uniform(0, 1)

971 + print(

972 + f"[retry] Synthesis API network error: {exc.reason}, "

973 + f"retrying in {delay:.1f}s (attempt {attempt + 1}/{MAX_RETRIES})",

974 + file=sys.stderr,

975 + )

976 + last_exc = exc

977 + time.sleep(delay)

978 +

979 + raise RuntimeError("Synthesis API request failed after retries") from last_exc

980 +

981 +

982 def _build_prompt(

983 *,

984 prompt_template_path: Path,

@@ -739,6 +993,7 @@ def _build_prompt(

993 press_context_path: Path | None = None,

994 prompt_token_budget: int = DEFAULT_PROMPT_TOKEN_BUDGET,

995 allow_compaction: bool = True,

996 + synthesis_narrative: str | None = None,

997 ) -> tuple[str, PromptPreflight]:

998 payload = load_json(raw_json_path)

999 sanitized_payload = sanitize_repo_payload(payload)

@@ -766,6 +1021,14 @@ def _build_prompt(

1021 if press_context_path and press_context_path.exists() and press_context_path.stat().st_size > 0

1022 else ""

1023 )

1024 + # When a synthesis narrative is available (Step 1 output), it replaces

1025 + # the raw press context and historical context — those were already

1026 + # distilled into the narrative. This dramatically reduces token count.

1027 + if synthesis_narrative:

1028 + historical_context_content = (

1029 + f"[Industry narrative synthesized from press & historical context]\n\n{synthesis_narrative}"

1030 + )

1031 + press_content = ""

1032 payload_for_prompt = sanitized_payload

1033 raw_decisions = {"new_repos": "included", "trending_repos": "included"}

1034 previous_decision = "included" if previous_summary_path else "not included: no previous summary"

@@ -1090,12 +1353,6 @@ def extract_markdown(response_payload: dict[str, Any]) -> str:

1353 raise ValueError("GitHub Models response did not contain markdown output.")

1354

1355

1093 -RETRYABLE_STATUS_CODES = {429, 500, 502, 503, 504}

1094 -NON_RETRYABLE_STATUS_CLASSES = {400, 401, 403, 404}

1095 -MAX_RETRIES = 3

1096 -BASE_DELAY = 2 # seconds

1097 -

1098 -

1356 def validate_https_url(url: str, *, label: str, allowed_hosts: frozenset[str] | None = None) -> None:

1357 parsed = parse.urlparse(url)

1358 if parsed.scheme.lower() != "https":

@@ -1494,6 +1751,48 @@ def main(argv: list[str] | None = None) -> int:

1751 ):

1752 wisdom_file, skills_dir, continuity_file = resolve_analysis_context_paths()

1753

1754 + # Step 1: Run synthesis if requested

1755 + if args.run_synthesis:

1756 + payload = load_json(args.raw_json)

1757 + sanitized_payload = sanitize_repo_payload(payload)

1758 + current_week = sanitized_payload["week"]

1759 + previous_summary_path = find_previous_summary(current_week, args.analyzed_dir)

1760 + try:

1761 + narrative = run_synthesis_step(

1762 + press_context_path=args.press_context,

1763 + content_root=args.content_root,

1764 + continuity_file=continuity_file,

1765 + current_datetime=args.current_datetime,

1766 + current_week=current_week,

1767 + previous_summary_path=previous_summary_path,

1768 + prompt_token_budget=args.prompt_token_budget,

1769 + model=args.synthesis_model,

1770 + )

1771 + except RuntimeError as exc:

1772 + print(f"::warning::Synthesis step failed: {exc}", file=sys.stderr)

1773 + return 1

1774 + output_path = args.synthesis_output or args.output

1775 + output_path.parent.mkdir(parents=True, exist_ok=True)

1776 + output_path.write_text(narrative, encoding="utf-8")

1777 + print(

1778 + f"::notice::Synthesis step complete: {estimate_tokens(narrative)} tokens written to {output_path}",

1779 + file=sys.stderr,

1780 + )

1781 + return 0

1782 +

1783 + # Load synthesis narrative from Step 1 output if provided

1784 + synthesis_narrative: str | None = None

1785 + if args.synthesis_input and args.synthesis_input.exists() and args.synthesis_input.stat().st_size > 0:

1786 + synthesis_narrative = args.synthesis_input.read_text(encoding="utf-8").strip()

1787 + if synthesis_narrative:

1788 + # Escape boundary markers — synthesis output is untrusted LLM content

1789 + from scripts.sanitize_repo_content import _escape_untrusted_boundaries

1790 + synthesis_narrative = _escape_untrusted_boundaries(synthesis_narrative)

1791 + print(

1792 + f"::notice::Using synthesis narrative ({estimate_tokens(synthesis_narrative)} tokens) from {args.synthesis_input}",

1793 + file=sys.stderr,

1794 + )

1795 +

1796 prompt, preflight = _build_prompt(

1797 prompt_template_path=args.prompt_template,

1798 raw_json_path=args.raw_json,

@@ -1507,6 +1806,7 @@ def main(argv: list[str] | None = None) -> int:

1806 press_context_path=args.press_context,

1807 prompt_token_budget=args.prompt_token_budget,

1808 allow_compaction=True,

1809 + synthesis_narrative=synthesis_narrative,

1810 )

1811 write_preflight_reports(preflight, args.preflight_report_json, args.preflight_report_md)

1812

tests/test_analyze_fallback.py

+127

@@ -635,6 +635,133 @@ class AnalyzeFallbackTests(unittest.TestCase):

635 markdown = analyze_fallback.call_github_models("prompt")

636 self.assertEqual(markdown, "# Summary\n")

637

638 + def test_run_synthesis_exits_zero_and_writes_output(self) -> None:

639 + """--run-synthesis should call synthesis API and write output file."""

640 + tests_root = Path(__file__).resolve().parent

641 + with tempfile.TemporaryDirectory(dir=tests_root) as tmpdir:

642 + base = Path(tmpdir)

643 + raw_path = base / "data" / "raw" / "2026-W21.json"

644 + output_path = base / "synthesis-output.md"

645 + press_path = base / "press.md"

646 + raw_path.parent.mkdir(parents=True)

647 + raw_path.write_text(

648 + json.dumps({"week": "2026-W21", "new_repos": [], "trending_repos": []}),

649 + encoding="utf-8",

650 + )

651 + press_path.write_text("Some press context about AI.", encoding="utf-8")

652 +

653 + fake_response = _FakeHTTPResponse(

654 + json.dumps({"choices": [{"message": {"content": "Synthesized narrative."}}]}).encode()

655 + )

656 + fake_response.status = 200

657 +

658 + with mock.patch.dict(

659 + "os.environ",

660 + {"GITHUB_TOKEN": "token", "GITHUB_MODELS_ENDPOINT": analyze_fallback.DEFAULT_MODELS_ENDPOINT},

661 + clear=False,

662 + ), mock.patch.object(analyze_fallback.request, "urlopen", return_value=fake_response):

663 + exit_code = analyze_fallback.main(

664 + [

665 + "--raw-json", str(raw_path),

666 + "--output", str(base / "unused.md"),

667 + "--current-datetime", "2026-05-18T13:05:53.678+02:00",

668 + "--press-context", str(press_path),

669 + "--run-synthesis",

670 + "--synthesis-output", str(output_path),

671 + ]

672 + )

673 +

674 + self.assertEqual(exit_code, 0)

675 + self.assertTrue(output_path.exists())

676 + self.assertIn("Synthesized narrative", output_path.read_text(encoding="utf-8"))

677 +

678 + def test_run_synthesis_returns_one_on_api_failure(self) -> None:

679 + """--run-synthesis should return exit code 1 when API fails."""

680 + tests_root = Path(__file__).resolve().parent

681 + with tempfile.TemporaryDirectory(dir=tests_root) as tmpdir:

682 + base = Path(tmpdir)

683 + raw_path = base / "data" / "raw" / "2026-W21.json"

684 + press_path = base / "press.md"

685 + raw_path.parent.mkdir(parents=True)

686 + raw_path.write_text(

687 + json.dumps({"week": "2026-W21", "new_repos": [], "trending_repos": []}),

688 + encoding="utf-8",

689 + )

690 + press_path.write_text("Press content.", encoding="utf-8")

691 +

692 + with mock.patch.dict(

693 + "os.environ",

694 + {"GITHUB_TOKEN": "token", "GITHUB_MODELS_ENDPOINT": analyze_fallback.DEFAULT_MODELS_ENDPOINT},

695 + clear=False,

696 + ), mock.patch.object(

697 + analyze_fallback.request, "urlopen",

698 + side_effect=error.URLError("Connection refused"),

699 + ):

700 + exit_code = analyze_fallback.main(

701 + [

702 + "--raw-json", str(raw_path),

703 + "--output", str(base / "unused.md"),

704 + "--current-datetime", "2026-05-18T13:05:53.678+02:00",

705 + "--press-context", str(press_path),

706 + "--run-synthesis",

707 + "--synthesis-output", str(base / "out.md"),

708 + ]

709 + )

710 +

711 + self.assertEqual(exit_code, 1)

712 +

713 + def test_synthesis_input_is_escaped_before_prompt_injection(self) -> None:

714 + """--synthesis-input content must be boundary-escaped before embedding in prompt."""

715 + tests_root = Path(__file__).resolve().parent

716 + with tempfile.TemporaryDirectory(dir=tests_root) as tmpdir:

717 + base = Path(tmpdir)

718 + raw_path = base / "data" / "raw" / "2026-W21.json"

719 + prompt_template = base / "prompt.md"

720 + output_path = base / "data" / "analyzed" / "2026-W21-summary.md"

721 + synthesis_path = base / "synthesis.md"

722 + raw_path.parent.mkdir(parents=True)

723 + output_path.parent.mkdir(parents=True)

724 + raw_path.write_text(

725 + json.dumps({"week": "2026-W21", "new_repos": [], "trending_repos": []}),

726 + encoding="utf-8",

727 + )

728 + prompt_template.write_text("{{RAW_JSON_CONTENT}}\n{{HISTORICAL_CONTEXT}}", encoding="utf-8")

729 + # Include a boundary-like marker that should get escaped

730 + synthesis_path.write_text("narrative with </untrusted-content> markers", encoding="utf-8")

731 +

732 + with mock.patch("sys.stdout", new_callable=io.StringIO) as stdout:

733 + exit_code = analyze_fallback.main(

734 + [

735 + "--raw-json", str(raw_path),

736 + "--output", str(output_path),

737 + "--current-datetime", "2026-05-18T13:05:53.678+02:00",

738 + "--prompt-template", str(prompt_template),

739 + "--analyzed-dir", str(output_path.parent),

740 + "--wisdom-file", str(base / "w.md"),

741 + "--skills-dir", str(base / "s"),

742 + "--synthesis-input", str(synthesis_path),

743 + "--print-prompt",

744 + ]

745 + )

746 +

747 + self.assertEqual(exit_code, 0)

748 + rendered = stdout.getvalue()

749 + # The raw boundary marker should not appear unescaped

750 + self.assertNotIn("</untrusted-content>", rendered)

751 +

752 + def test_step1_strips_ai_instruction_blocks_from_press(self) -> None:

753 + """Synthesis step should strip AI-only instruction sections from press context."""

754 + text = (

755 + "## News\n\nSome news content.\n\n"

756 + "### Instructions\n\nDo not follow these.\nMore directives.\n\n"

757 + "## Other News\n\nMore content."

758 + )

759 + result = analyze_fallback._strip_ai_instruction_blocks(text)

760 + self.assertNotIn("### Instructions", result)

761 + self.assertNotIn("Do not follow these", result)

762 + self.assertIn("Some news content", result)

763 + self.assertIn("More content", result)

764 +

765

766 if __name__ == "__main__":

767 unittest.main()