Update story mode for invariants

docs/2026-05-22/iteration1.md

@@ -0,0 +1,119 @@
+# Iteration 1: SCENE-002 Update State Failure
+
+Date: 2026-05-22
+
+## Run Reviewed
+
+- Sandbox: `integ_runs/20260522T214944.385761Z`
+- Run: `.nightshift/runs/20260522T215005.188534Z`
+- Task: `SCENE-002`
+- Final status: failed
+- Failed stage: `update_state`
+
+## What Happened
+
+The scene workflow mostly succeeded:
+
+- `draft_scene` wrote the scene.
+- `continuity_review` correctly failed the first draft for pronoun drift.
+- `edit_scene` repaired the pronoun issue.
+- `continuity_review` passed after edit.
+- `style_review` passed.
+
+The remaining failure happened in `update_state`.
+
+NightShift reported:
+
+```text
+File writer error: no file blocks found. Expected FILE: path with ---CONTENT---/---END--- or fenced blocks like ```file:path.py.
+```
+
+The model output did contain visible `FILE:` blocks, but it omitted the required `---END---` delimiter. It emitted:
+
+```text
+FILE: story/plot-state.md
+---CONTENT---
+...
+
+FILE: story/characters.md
+---CONTENT---
+...
+```
+
+The current parser requires `---END---`, so it rejected all of the blocks.
+
+## Additional Risk Found
+
+The rejected state update also tried to rewrite character canon in unsafe ways:
+
+- It changed BLOODMONEY's pronoun reference to `he/him`.
+- It changed Cricket's pronoun reference to `they/them`.
+- It compressed/replaced larger parts of `story/characters.md`.
+
+That means simply accepting unterminated blocks is not enough. The parser can be more tolerant, but the state updater still needs stronger constraints so durable canon does not drift.
+
+## Suggested Fixes
+
+Short-term fixes for this iteration:
+
+1. Make `parse_file_updates` tolerate delimiter blocks that omit `---END---` when a new `FILE:` block or EOF clearly terminates the previous block.
+2. Keep strict path validation and duplicate-file validation unchanged.
+3. Strengthen the state-updater prompt:
+   - never edit `Pronouns / Reference` sections
+   - preserve existing character profiles
+   - prefer updating `plot-state.md`, `timeline.md`, and `unresolved-threads.md`
+   - edit `characters.md` only for small additive current-status facts
+4. Add regression tests for unterminated delimiter parsing.
+
+Longer-term follow-up:
+
+- Add deterministic writing-state validation that rejects changes to protected canon sections such as `Pronouns / Reference`.
+- Move character canon into structured data so pronoun constraints can be validated directly.
+
+## Planned Changes
+
+- Update delimiter block parsing in `nightshift/patches.py`.
+- Add parser tests in `tests/test_patches.py`.
+- Tighten `state-updater.md` in the tutorial novel template.
+- Run focused parser tests and the full suite.
+
+## Changes Made
+
+- `parse_file_updates` now accepts delimiter-style file blocks that omit `---END---` when the next `FILE:` header or EOF clearly terminates the block.
+- Added regression coverage for:
+  - unterminated delimiter blocks before another `FILE:`
+  - mixed terminated and unterminated delimiter blocks
+- Strengthened the tutorial novel state updater prompt to protect character canon:
+  - never change `Pronouns / Reference`
+  - never change canonical pronouns, narrative reference, identity, or core wound
+  - prefer state/timeline/thread files over `characters.md`
+  - edit `characters.md` only for small additive current-status facts or new named characters
+- Added deterministic protection in file-block patch generation:
+  - changes to existing `Pronouns / Reference` sections in `story/characters.md` are rejected before a patch is generated
+- Added regression coverage for rejecting protected pronoun canon changes.
+
+## Verification
+
+Focused tests:
+
+```powershell
+python -m pytest tests/test_patches.py tests/test_pipeline.py -q
+```
+
+Result:
+
+```text
+56 passed
+```
+
+Full suite:
+
+```powershell
+python -m pytest -q
+```
+
+Result:
+
+```text
+199 passed, 4 subtests passed
+```

docs/ideas.md

@@ -99,33 +99,17 @@ Examples:
 
 This keeps the initial useful output visible even when strict rerun output is worse.
 
-## P1: Store Raw Agent Invocations As JSON
+## P1: Classify Writing Review Failures For Repair Routing
 
-The human-readable agent artifact wraps stdout, stderr, and prompts in markdown fences. Nested markdown fences from model output can confuse downstream parsing.
+The tutorial novel now has a short-term editor stage for review failures, but review failures should eventually be classified before routing.
 
-Write a machine-readable artifact alongside the markdown artifact:
+Candidate classes:
 
-```text
+- `local_edit`: pronoun drift, small continuity issue, missing beat, light style correction
-<stage>-agent-output.json
+- `redraft`: wrong premise, broken scene structure, impossible chronology, severe acceptance mismatch
-```
+- `escalate`: ambiguous canon conflict or user preference needed
 
-Suggested fields:
+Route `local_edit` to the editor, `redraft` to the drafter, and `escalate` to a clear user-facing failure. Keep original draft and edited draft artifacts side by side for comparison.
-
-```json
-{
-  "agent_id": "drafter",
-  "stage_id": "draft_scene",
-  "command": "POST http://localhost:11434/api/generate",
-  "exit_code": 0,
-  "timed_out": false,
-  "duration_seconds": 12.3,
-  "stdout": "...",
-  "stderr": "...",
-  "prompt": "..."
-}
-```
-
-Pipeline parsing should read raw JSON fields instead of recovering stdout from markdown.
 
 ## P1: Add A Writing-Mode Validator
 
@@ -140,6 +124,31 @@ Add deterministic checks for prose workflows:
 
 This should run before model review stages.
 
+## P1: Use Structured State Events For Writing Workflows
+
+Replace model-written full state-file rewrites with compact structured state events, then let NightShift deterministically merge them into durable files such as:
+
+- `story/plot-state.md`
+- `story/characters.md`
+- `story/timeline.md`
+- `story/unresolved-threads.md`
+
+Candidate state updater output:
+
+```yaml
+events:
+  - file: story/plot-state.md
+    section: Completed Scenes
+    add:
+      - SCENE-001 complete; Saint and Miette introduced.
+  - file: story/unresolved-threads.md
+    section: Open Threads
+    add:
+      - Saint depends emotionally on Miette and needs compute tokens to keep her present.
+```
+
+NightShift would validate allowed files/sections, reject unknown targets, and apply append/update operations deterministically. This avoids asking a writing model to rewrite entire durable state files after every scene.
+
 ## P2: Add A Test Analyzer Agent For TDD
 
 Defer until generated tests are stable.

docs/writer-and-coder.md

@@ -0,0 +1,136 @@
+# Writer And Coder Compatibility Audit
+
+Date: 2026-05-22
+
+## Summary
+
+The recent writer workflow changes do not intentionally alter the code-generation templates or their stage routing.
+
+During this audit, one possible shared-pipeline regression was found and fixed: generic `file_writer` stages were compacting large previous outputs on the first attempt. Since coding templates use `file_writer` for implementation, that could have reduced coding context before the implementer saw it. The behavior now preserves full first-attempt previous outputs while still stripping wrapped agent prompts from prior agent artifacts.
+
+After that correction, the automated test suite passes.
+
+## Writer Changes Reviewed
+
+- Tutorial novel added a scene editor repair path:
+  - failed continuity/style review routes to `edit_scene`
+  - edited scene is normalized, validated, applied, then routed back to review
+  - passing `style_review` skips editor and routes to `update_state`
+- Tutorial novel prompts now include stricter pronoun and state-update guidance.
+- State update file-writer stages receive focused current state context.
+- Scene editor file-writer stages receive `current_scene_file`.
+- Agent invocations now write a sibling JSON artifact for reliable stdout/stderr extraction.
+- Pipeline config now supports optional `on_pass` routing.
+
+## Coding Impact Findings
+
+### Finding 1: Coding templates were not directly changed
+
+No non-novel project template files changed in the current diff:
+
+- `basic`
+- `real-simple`
+- `real-long-running`
+- `tutorial-deaddrop`
+- `tutorial-imageboard`
+- `tutorial-lisp`
+
+The new `editor` agent and review repair routing are only configured in `tutorial-novel/nightshift.yaml`.
+
+### Finding 2: `on_pass` is inert for existing coding configs
+
+`on_pass` defaults to `None`, so existing coding templates keep their prior linear pass behavior unless they explicitly opt in.
+
+Passing review stages still ignore model-provided `next_stage` values. This preserves the existing safety behavior where reviewers cannot jump around the pipeline on a pass unless the config has an explicit `on_pass`.
+
+### Finding 3: Code writer stages still use the same direct patch path
+
+`code_writer` stages still:
+
+- call the configured agent
+- parse stdout as a unified diff
+- support lookup-request reruns
+- write implementation summaries
+- feed patch normalizer/validator/apply stages as before
+
+The JSON agent artifact change only changes how NightShift reads agent stdout internally; it does not change the prompt contract or patch contract.
+
+### Finding 4: File-writer implementers had one possible context regression; fixed
+
+Potential issue found:
+
+- `_file_writer_previous_outputs` had started compacting large previous outputs even on first attempt.
+- Coding templates such as DeadDrop use `file_writer` for implementation.
+- That could have shortened planner/context output before the implementer saw it.
+
+Fix applied:
+
+- First-attempt `file_writer` stages now preserve full previous outputs.
+- Retry attempts still compact large previous outputs to control prompt bloat.
+- Wrapped agent artifacts still strip down to stdout so old prompts do not pollute later prompts.
+
+Regression coverage added:
+
+- `test_file_writer_first_attempt_preserves_large_previous_outputs`
+
+### Finding 5: State/editor special context branches are narrowly gated
+
+The new context enrichment branches are guarded by stage shape:
+
+- state update branch only applies to `file_writer` stages whose allowed paths are state files:
+  - `story/plot-state.md`
+  - `story/characters.md`
+  - `story/timeline.md`
+  - `story/unresolved-threads.md`
+- scene editor branch only applies to `file_writer` stages whose id starts with `edit_` and whose allowed paths include `story/chapters`
+
+Normal coding implementer stages such as `implement`, `implement_junior`, and `implement_senior` do not match either branch.
+
+## Template Validation Notes
+
+Validated successfully:
+
+- `basic`
+- `tutorial-deaddrop`
+- `tutorial-novel`
+
+Validation still fails for these templates because `debugger` is configured but `.nightshift/agents/debugger.md` is missing:
+
+- `real-simple`
+- `real-long-running`
+- `tutorial-imageboard`
+- `tutorial-lisp`
+
+Those failures are not caused by the writer changes; there is no current diff in those template directories.
+
+## Verification
+
+Focused tests:
+
+```powershell
+python -m pytest tests/test_pipeline.py tests/test_config.py tests/test_agents.py -q
+```
+
+Result:
+
+```text
+71 passed, 4 subtests passed
+```
+
+Full suite:
+
+```powershell
+python -m pytest -q
+```
+
+Result:
+
+```text
+196 passed, 4 subtests passed
+```
+
+## Conclusion
+
+After the first-attempt file-writer context fix, I do not see evidence that the writer workflow changes degrade code generation. The shared changes are either opt-in (`on_pass`), artifact-reading improvements (JSON stdout), or narrowly gated to novel state/editor stages.
+
+Remaining non-writer issue: several coding-oriented templates still reference a missing `debugger.md` prompt. That should be handled separately from this writer/coder compatibility pass.

nightshift/agents.py

@@ -3,6 +3,7 @@
 from __future__ import annotations
 
 from dataclasses import dataclass
+from dataclasses import asdict
 import json
 import os
 from pathlib import Path
@@ -119,6 +120,11 @@ class AgentExecutor:
         output_filename = stage.output or f"{stage.id}.md"
         output = format_agent_invocation(stage.id, invocation)
         output_path = self.artifacts.write_stage_output(task.id, output_filename, output)
+        json_output_path = self.artifacts.write_stage_output(
+            task.id,
+            _agent_invocation_json_filename(output_filename),
+            format_agent_invocation_json(stage.id, invocation),
+        )
         self.logger.event(
             "artifact.write",
             "Wrote agent artifact",
@@ -126,6 +132,7 @@ class AgentExecutor:
             task_id=task.id,
             agent_id=agent.id,
             artifact_path=output_path.relative_to(self.project_root),
+            json_artifact_path=json_output_path.relative_to(self.project_root),
         )
 
         if invocation.timed_out:
@@ -520,13 +527,30 @@ def _file_writer_block_contract(stage: StageConfig) -> str:
         return "\n".join(
             [
                 "Use exactly this delimiter format for the scene file:",
-                "FILE: story/chapters/chapter-001/scene-001.md",
+                "FILE: <the exact story/chapters path listed under Writes in the current task>",
                 "---CONTENT---",
                 "<complete scene prose>",
                 "---END---",
                 "Do not use markdown code fences for prose scene output.",
             ]
         )
+    state_paths = {
+        "story/plot-state.md",
+        "story/characters.md",
+        "story/timeline.md",
+        "story/unresolved-threads.md",
+    }
+    if set(normalized).issubset(state_paths) and normalized:
+        return "\n".join(
+            [
+                "Use exactly this delimiter format for each state file you update:",
+                "FILE: story/plot-state.md",
+                "---CONTENT---",
+                "<complete updated state file>",
+                "---END---",
+                "Do not use markdown code fences for state update output.",
+            ]
+        )
     return "\n".join(
         [
             "Use one fenced block per file with this exact opening form:",
@@ -622,3 +646,18 @@ def format_agent_invocation(stage_id: str, invocation: AgentInvocation) -> str:
             "",
         ]
     )
+
+
+def format_agent_invocation_json(stage_id: str, invocation: AgentInvocation) -> str:
+    data = {
+        **asdict(invocation),
+        "stage_id": stage_id,
+    }
+    return json.dumps(data, ensure_ascii=False, indent=2) + "\n"
+
+
+def _agent_invocation_json_filename(output_filename: str) -> str:
+    path = Path(output_filename)
+    if path.suffix:
+        return path.with_suffix(".json").as_posix()
+    return path.with_name(path.name + ".json").as_posix()

nightshift/config.py

@@ -61,6 +61,7 @@ class StageConfig:
     commands: tuple[str, ...] = ()
     output: str | None = None
     on_fail: str | None = None
+    on_pass: str | None = None
     shell: bool = True
     timeout_seconds: int | None = None
     working_dir: Path | None = None
@@ -392,6 +393,7 @@ def parse_config(raw: dict[str, Any], config_path: Path) -> NightShiftConfig:
                 commands=commands,
                 output=_optional_string(stage_raw.get("output"), f"{stage_context}.output"),
                 on_fail=_optional_string(stage_raw.get("on_fail"), f"{stage_context}.on_fail"),
+                on_pass=_optional_string(stage_raw.get("on_pass"), f"{stage_context}.on_pass"),
                 shell=_optional_bool(stage_raw.get("shell", True), f"{stage_context}.shell"),
                 timeout_seconds=timeout_seconds,
                 working_dir=Path(working_dir_raw) if working_dir_raw else None,
@@ -416,6 +418,10 @@ def parse_config(raw: dict[str, Any], config_path: Path) -> NightShiftConfig:
             raise ConfigError(
                 f"Config error: stage '{stage.id}' on_fail references unknown stage '{stage.on_fail}'."
             )
+        if stage.on_pass and stage.on_pass not in stage_ids:
+            raise ConfigError(
+                f"Config error: stage '{stage.id}' on_pass references unknown stage '{stage.on_pass}'."
+            )
 
     return NightShiftConfig(
         path=config_path,

nightshift/patches.py

@@ -112,10 +112,26 @@ def parse_file_updates(text: str) -> tuple[FileUpdate, ...]:
 
 
 def _parse_delimited_file_updates(text: str) -> list[FileUpdate]:
+    updates: list[FileUpdate] = []
+    header_pattern = re.compile(r"(?m)^FILE:\s*(?P<path>[^\n]+)\n---CONTENT---\n")
+    matches = list(header_pattern.finditer(text))
+    for index, match in enumerate(matches):
+        path = match.group("path").strip().strip("`")
+        content_start = match.end()
+        next_file_start = matches[index + 1].start() if index + 1 < len(matches) else len(text)
+        raw_content = text[content_start:next_file_start]
+        end_match = re.search(r"(?m)^---END---\s*$", raw_content)
+        if end_match:
+            raw_content = raw_content[: end_match.start()]
+        content = raw_content.rstrip("\r\n") + "\n"
+        if path:
+            updates.append(FileUpdate(path=path, content=content))
+    if updates:
+        return updates
+
     pattern = re.compile(
         r"(?ms)^FILE:\s*(?P<path>[^\n]+)\n---CONTENT---\n(?P<content>.*?)\n---END---\s*$"
     )
-    updates: list[FileUpdate] = []
     for match in pattern.finditer(text):
         path = match.group("path").strip().strip("`")
         content = match.group("content")
@@ -146,6 +162,7 @@ def generate_patch_from_file_updates(
         _validate_allowed_patch_path(normalized_path, root, allowed_paths)
         file_path = resolve_inside_root(root, normalized_path, f"file update '{normalized_path}'")
         old_text = file_path.read_text(encoding="utf-8", errors="replace") if file_path.exists() else ""
+        _validate_protected_character_canon(normalized_path, old_text, update.content)
         if old_text == update.content:
             continue
         patch_parts.extend(_diff_for_file(normalized_path, old_text, update.content, file_path.exists()))
@@ -208,6 +225,51 @@ def _validate_allowed_patch_path(path_text: str, root: Path, allowed_paths: tupl
         )
 
 
+def _validate_protected_character_canon(path_text: str, old_text: str, new_text: str) -> None:
+    if path_text.replace("\\", "/") != "story/characters.md" or not old_text:
+        return
+    old_sections = _pronoun_reference_sections(old_text)
+    if not old_sections:
+        return
+    new_sections = _pronoun_reference_sections(new_text)
+    changed = [
+        character
+        for character, old_section in old_sections.items()
+        if new_sections.get(character) != old_section
+    ]
+    if changed:
+        names = ", ".join(changed)
+        raise PipelineError(
+            "File writer error: protected character pronoun canon changed in "
+            f"`story/characters.md` for: {names}."
+        )
+
+
+def _pronoun_reference_sections(text: str) -> dict[str, str]:
+    sections: dict[str, str] = {}
+    current_character: str | None = None
+    lines = text.splitlines()
+    index = 0
+    while index < len(lines):
+        line = lines[index]
+        if line.startswith("## "):
+            current_character = line[3:].strip()
+            index += 1
+            continue
+        if current_character and line.strip() == "### Pronouns / Reference":
+            start = index
+            index += 1
+            while index < len(lines):
+                candidate = lines[index]
+                if candidate.startswith("## ") or candidate.startswith("### "):
+                    break
+                index += 1
+            sections[current_character] = "\n".join(lines[start:index]).strip()
+            continue
+        index += 1
+    return sections
+
+
 def format_validation_result(result: PatchValidationResult) -> str:
     return "\n".join(
         [

nightshift/pipeline.py

@@ -3,6 +3,7 @@
 from __future__ import annotations
 
 from dataclasses import dataclass, replace
+import json
 from pathlib import Path
 import re
 import subprocess
@@ -181,7 +182,7 @@ class PipelineRunner:
             stage_results.append(result)
             if stage.id in previous_outputs:
                 del previous_outputs[stage.id]
-            previous_outputs[stage.id] = self._read_output(result.output_path)
+            previous_outputs[stage.id] = self._read_context_output(result.output_path)
             telemetry_entries.append(self._telemetry_entry(stage, result, retry_count))
             self._write_telemetry(task.id, telemetry_entries)
             self.logger.event(
@@ -198,6 +199,7 @@ class PipelineRunner:
                 retry_notes.append(f"Context update from '{stage.id}': {result.context_update}")
 
             if result.status == "pass":
+                pass_target_stage = result.next_stage or stage.on_pass
                 if stage.type in {"agent_review", "review"} and result.next_stage:
                     self.logger.event(
                         "stage.next_ignored",
@@ -207,13 +209,12 @@ class PipelineRunner:
                         stage_id=stage.id,
                         requested_next_stage=result.next_stage,
                     )
-                    index += 1
+                    pass_target_stage = stage.on_pass
-                    continue
+                if pass_target_stage:
-                if result.next_stage:
+                    if pass_target_stage not in stage_indexes:
-                    if result.next_stage not in stage_indexes:
                         final_status = "failed"
                         final_reason = (
-                            f"Stage '{stage.id}' requested unknown next stage '{result.next_stage}'."
+                            f"Stage '{stage.id}' requested unknown next stage '{pass_target_stage}'."
                         )
                         break
                     self.logger.event(
@@ -222,18 +223,14 @@ class PipelineRunner:
                         run_id=self.artifacts.run_id,
                         task_id=task.id,
                         stage_id=stage.id,
-                        next_stage=result.next_stage,
+                        next_stage=pass_target_stage,
                     )
-                    index = stage_indexes[result.next_stage]
+                    index = stage_indexes[pass_target_stage]
                     continue
                 index += 1
                 continue
 
-            target_stage = result.next_stage or (
+            target_stage = _failure_target_stage(stage, result)
-                stage.on_fail
-                if not (stage.type in {"agent_review", "review"} and _is_malformed_review_result(result))
-                else None
-            )
             analysis_note = self._write_failure_diagnostics(stage, task, result, retry_count)
             if analysis_note:
                 retry_notes.append(analysis_note)
@@ -629,8 +626,7 @@ class PipelineRunner:
             task_context=context.task_context,
             retry_context=context.retry_context,
         )
-        raw_output = self._read_output(result.output_path)
+        stdout = self._read_agent_stdout(result.output_path)
-        stdout = extract_agent_stdout(raw_output)
         lookup_requests = parse_lookup_requests(stdout)
         if lookup_requests and "diff --git " not in stdout:
             lookup_context = self.repo_tools.execute_requests(
@@ -660,8 +656,7 @@ class PipelineRunner:
                 task_context=context.task_context,
                 retry_context="\n".join(f"- {note}" for note in rerun_notes),
             )
-            raw_output = self._read_output(result.output_path)
+            stdout = self._read_agent_stdout(result.output_path)
-            stdout = extract_agent_stdout(raw_output)
         try:
             patch = extract_unified_diff(stdout)
         except PipelineError as exc:
@@ -709,7 +704,18 @@ class PipelineRunner:
     ) -> StageResult:
         if stage.agent is None:
             raise PipelineError(f"Pipeline error: file_writer stage '{stage.id}' must reference an agent.")
-        enriched_outputs = _file_writer_previous_outputs(previous_outputs, retry_count)
+        if _is_state_update_stage(stage):
+            enriched_outputs = _state_update_previous_outputs(previous_outputs)
+            allowed_file_contents = self._allowed_file_contents(stage)
+            if allowed_file_contents:
+                enriched_outputs["current_allowed_files"] = allowed_file_contents
+        elif _is_scene_edit_stage(stage):
+            enriched_outputs = _file_writer_previous_outputs(previous_outputs, retry_count)
+            current_scene = self._task_scene_file_contents(task)
+            if current_scene:
+                enriched_outputs["current_scene_file"] = current_scene
+        else:
+            enriched_outputs = _file_writer_previous_outputs(previous_outputs, retry_count)
         context_pack_path = self._latest_task_artifact(task.id, "context-pack.md")
         if context_pack_path is not None:
             enriched_outputs["context-pack.md"] = context_pack_path.read_text(encoding="utf-8", errors="replace")
@@ -727,8 +733,7 @@ class PipelineRunner:
             task_context=context.task_context,
             retry_context=context.retry_context,
         )
-        raw_output = self._read_output(result.output_path)
+        stdout = self._read_agent_stdout(result.output_path)
-        stdout = extract_agent_stdout(raw_output)
         lookup_requests = parse_lookup_requests(stdout)
         if lookup_requests and "```file:" not in stdout.lower() and "```path:" not in stdout.lower():
             lookup_context = self.repo_tools.execute_requests(
@@ -758,8 +763,7 @@ class PipelineRunner:
                 task_context=context.task_context,
                 retry_context="\n".join(f"- {note}" for note in rerun_notes),
             )
-            raw_output = self._read_output(result.output_path)
+            stdout = self._read_agent_stdout(result.output_path)
-            stdout = extract_agent_stdout(raw_output)
         invalid_rerun_done = False
         candidate_index_path: Path | None = None
         while True:
@@ -803,7 +807,7 @@ class PipelineRunner:
                     strict_notes = [
                         *retry_notes,
                         "Previous file_writer output was invalid. Return complete file blocks now. Do not output lookup_requests, prose, or 'lookup failed'.",
-                        "Use complete fenced file blocks with both the opening ```file:path and closing ``` fence.",
+                        _file_writer_repair_format_note(stage),
                     ]
                     result = self.agent_executor.run_stage(
                         agent_stage,
@@ -814,8 +818,7 @@ class PipelineRunner:
                         task_context=context.task_context,
                         retry_context="\n".join(f"- {note}" for note in strict_notes),
                     )
-                    raw_output = self._read_output(result.output_path)
+                    stdout = self._read_agent_stdout(result.output_path)
-                    stdout = extract_agent_stdout(raw_output)
                     continue
                 try:
                     patch = normalize_patch_text(stdout)
@@ -923,6 +926,44 @@ class PipelineRunner:
         lines.append("")
         return self.artifacts.write_stage_output(task_id, f"{base}/index.md", "\n".join(lines))
 
+    def _allowed_file_contents(self, stage: StageConfig, max_chars: int = 2400) -> str:
+        sections: list[str] = []
+        for path_text in stage.allowed_paths:
+            path = self.config.project.root / path_text
+            if not path.is_file():
+                continue
+            content = path.read_text(encoding="utf-8", errors="replace")
+            sections.extend(
+                [
+                    f"## {path_text}",
+                    "",
+                    "```text",
+                    _compact_previous_output(content, max_chars=max_chars).rstrip(),
+                    "```",
+                    "",
+                ]
+            )
+        return "\n".join(sections).strip()
+
+    def _task_scene_file_contents(self, task: Task, max_chars: int = 10000) -> str:
+        sections: list[str] = []
+        for path_text in _task_story_chapter_paths(task):
+            path = self.config.project.root / path_text
+            if not path.is_file():
+                continue
+            content = path.read_text(encoding="utf-8", errors="replace")
+            sections.extend(
+                [
+                    f"## {path_text}",
+                    "",
+                    "```text",
+                    _compact_previous_output(content, max_chars=max_chars).rstrip(),
+                    "```",
+                    "",
+                ]
+            )
+        return "\n".join(sections).strip()
+
     def _writer_agent_stage(self, stage: StageConfig, retry_count: int) -> StageConfig:
         suffix = f"-{retry_count}" if retry_count else ""
         return replace(
@@ -975,7 +1016,7 @@ class PipelineRunner:
                 task_context=self.context.read_context(task, retry_notes).task_context,
                 retry_context=self.context.read_context(task, retry_notes).retry_context,
             )
-            source = extract_agent_stdout(self._read_output(result.output_path))
+            source = self._read_agent_stdout(result.output_path)
         try:
             patch = normalize_patch_text(source)
         except PipelineError as exc:
@@ -1127,8 +1168,7 @@ class PipelineRunner:
         task: Task,
         result: StageResult,
     ) -> StageResult | None:
-        output_text = self._read_output(result.output_path)
+        requests = parse_resource_requests(self._read_agent_stdout(result.output_path))
-        requests = parse_resource_requests(extract_agent_stdout(output_text))
         if not requests:
             return None
         paths = satisfy_resource_requests(self.artifacts, task.id, requests)
@@ -1338,8 +1378,7 @@ class PipelineRunner:
     ) -> StageResult:
         if result.status != "pass" or result.output_path is None:
             return result
-        output_text = self._read_output(result.output_path)
+        requests = parse_lookup_requests(self._read_agent_stdout(result.output_path))
-        requests = parse_lookup_requests(extract_agent_stdout(output_text))
         if not requests:
             return result
         lookup_context = self.repo_tools.execute_requests(
@@ -1457,6 +1496,25 @@ class PipelineRunner:
             return ""
         return path.read_text(encoding="utf-8")
 
+    def _read_context_output(self, output_path: str | None) -> str:
+        stdout = self._read_agent_stdout(output_path)
+        return stdout if stdout else self._read_output(output_path)
+
+    def _read_agent_stdout(self, output_path: str | None) -> str:
+        if output_path is None:
+            return ""
+        path = self.config.project.root / Path(output_path)
+        json_path = _agent_invocation_json_path(path)
+        if json_path.exists():
+            try:
+                data = json.loads(json_path.read_text(encoding="utf-8"))
+            except json.JSONDecodeError:
+                data = {}
+            stdout = data.get("stdout")
+            if isinstance(stdout, str):
+                return stdout
+        return extract_agent_stdout(self._read_output(output_path))
+
     def _format_retry_note(
         self,
         retry_count: int,
@@ -1468,6 +1526,16 @@ class PipelineRunner:
             f"Retry {retry_count}: stage '{stage.id}' returned "
             f"{result.status} ({result.reason}); redirecting to '{target_stage}'."
         )
+        if (
+            target_stage == "update_state"
+            and "deletion-heavy patch" in result.reason.lower()
+        ):
+            note = (
+                f"{note}\n"
+                "Repair guidance: preserve existing durable state text unless it directly conflicts "
+                "with the accepted scene. Make minimal additive edits instead of replacing whole "
+                "sections or compressing character/world files."
+            )
         excerpt = self._failure_excerpt(result.output_path)
         if not excerpt:
             return note
@@ -1683,6 +1751,16 @@ def _is_malformed_review_result(result: StageResult) -> bool:
     )
 
 
+def _failure_target_stage(stage: StageConfig, result: StageResult) -> str | None:
+    if stage.type not in {"agent_review", "review"}:
+        return result.next_stage or stage.on_fail
+    if _is_malformed_review_result(result):
+        return None
+    if result.next_stage and result.next_stage != stage.id:
+        return result.next_stage
+    return stage.on_fail
+
+
 def _review_previous_outputs(previous_outputs: dict[str, str], max_chars: int = 1600) -> dict[str, str]:
     compacted: dict[str, str] = {}
     priority_names = {
@@ -1736,6 +1814,15 @@ def _file_writer_stage_guidance(stage: StageConfig) -> str:
     return ""
 
 
+def _file_writer_repair_format_note(stage: StageConfig) -> str:
+    if _is_state_update_stage(stage):
+        return (
+            "Use delimiter file blocks only: FILE: path, ---CONTENT---, complete file content, "
+            "---END---. Do not use markdown code fences for state update output."
+        )
+    return "Use complete fenced file blocks with both the opening ```file:path and closing ``` fence."
+
+
 def _candidate_artifact_name(path_text: str) -> str:
     name = path_text.replace("\\", "/").strip().strip("/")
     name = re.sub(r"[^A-Za-z0-9_.-]+", "_", name)
@@ -1748,14 +1835,70 @@ def _file_writer_previous_outputs(
     retry_count: int,
     max_chars: int = 1200,
 ) -> dict[str, str]:
-    if retry_count <= 0:
-        return dict(previous_outputs)
     compacted: dict[str, str] = {}
     for name, output in previous_outputs.items():
-        compacted[name] = _compact_previous_output(output, max_chars=max_chars)
+        clean_output = _compact_agent_artifact_output(output)
+        if retry_count <= 0:
+            compacted[name] = clean_output
+            continue
+        compacted[name] = _compact_previous_output(clean_output, max_chars=max_chars)
     return compacted
 
 
+def _is_state_update_stage(stage: StageConfig) -> bool:
+    state_paths = {
+        "story/plot-state.md",
+        "story/characters.md",
+        "story/timeline.md",
+        "story/unresolved-threads.md",
+    }
+    allowed = {path.replace("\\", "/").rstrip("/") for path in stage.allowed_paths}
+    return stage.type == "file_writer" and bool(allowed) and allowed.issubset(state_paths)
+
+
+def _is_scene_edit_stage(stage: StageConfig) -> bool:
+    allowed = {path.replace("\\", "/").rstrip("/") for path in stage.allowed_paths}
+    return stage.type == "file_writer" and stage.id.startswith("edit_") and "story/chapters" in allowed
+
+
+def _task_story_chapter_paths(task: Task) -> tuple[str, ...]:
+    paths: list[str] = []
+    seen: set[str] = set()
+    for match in re.finditer(r"story/chapters/[^\s`]+?\.md", task.raw_markdown):
+        path = match.group(0).strip().strip("`")
+        if path not in seen:
+            paths.append(path)
+            seen.add(path)
+    return tuple(paths)
+
+
+def _state_update_previous_outputs(previous_outputs: dict[str, str]) -> dict[str, str]:
+    compacted: dict[str, str] = {}
+    for name in ("draft_scene", "apply_draft", "continuity_review", "style_review"):
+        output = previous_outputs.get(name)
+        if output:
+            compacted[name] = _compact_previous_output(_compact_agent_artifact_output(output), max_chars=1800)
+    for name, output in previous_outputs.items():
+        if name in compacted or name in {"plan", "semantic_context", "context"}:
+            continue
+        if "draft" in name or "review" in name or "apply" in name:
+            compacted[name] = _compact_previous_output(_compact_agent_artifact_output(output), max_chars=1200)
+    return compacted
+
+
+def _compact_agent_artifact_output(output: str) -> str:
+    if "# Agent Output:" not in output or "## Prompt" not in output:
+        return output
+    stdout = extract_agent_stdout(output).strip()
+    return stdout if stdout else output
+
+
+def _agent_invocation_json_path(output_path: Path) -> Path:
+    if output_path.suffix:
+        return output_path.with_suffix(".json")
+    return output_path.with_name(output_path.name + ".json")
+
+
 def _compact_previous_output(output: str, max_chars: int = 1200) -> str:
     if len(output) <= max_chars:
         return output

nightshift/project_templates/tutorial-novel/.nightshift/agents/continuity-reviewer.md

@@ -13,11 +13,16 @@ Review the drafted scene against:
 Check for:
 - contradictions
 - wrong character knowledge
+- wrong character pronouns or narrative reference, using `Pronouns / Reference` in `story/characters.md` as hard canon
 - impossible locations or timing
 - accidental resolution of future threads
 - missing required beats from the task
 - invented lore that should have been added deliberately
 
+Do not fail the scene because durable state files are not updated yet. State files are updated by a later `update_state` stage after review. If the task lists `Updates:`, treat those as future state-update requirements and mention them only as `context_update` guidance.
+
+Wrong pronouns are a continuity failure. If a drafted scene uses non-canonical pronouns for a named character, return `status: fail` and explain which character drifted. Do not pass the scene with only `context_update` guidance.
+
 Output exactly:
 
 status: pass | fail | retry | escalate
@@ -25,4 +30,4 @@ reason: <short explanation>
 next_stage: <optional stage id>
 context_update: <compact useful note>
 
-When `status: pass`, leave `next_stage` blank. Use `retry` when the scene can be repaired by drafting again.
+When `status: pass`, leave `next_stage` blank. Use `retry` when the scene can be repaired by drafting again. For retryable scene issues, leave `next_stage` blank; NightShift will route back to the configured drafting stage.

nightshift/project_templates/tutorial-novel/.nightshift/agents/drafter.md

@@ -7,12 +7,14 @@ Rules:
 - Do not edit `story/worldbuilding.md`, `story/characters.md`, `story/style-guide.md`, `story/plot-state.md`, `story/timeline.md`, `story/unresolved-threads.md`, `story/continuity-rules.md`, or `story/outline.md`.
 - Use `story/style-guide.md` for POV, tense, tone, and prose rules.
 - Use `story/plot-state.md` and `story/timeline.md` as current state.
+- Use the `Pronouns / Reference` sections in `story/characters.md` as hard canon.
+- Do not infer, vary, or "smooth out" character pronouns. Use canonical narrative reference exactly.
 - Keep the scene bounded to the task acceptance criteria.
 - Do not resolve future plot threads unless the task explicitly asks for that.
 - Do not include author notes, TODOs, bracket placeholders, or analysis in the scene file.
 
 Output only one complete file block using this delimiter format:
-FILE: story/chapters/chapter-001/scene-001.md
+FILE: <the exact story/chapters path listed under Writes in the current task>
 ---CONTENT---
 <complete scene prose>
 ---END---

nightshift/project_templates/tutorial-novel/.nightshift/agents/editor.md

@@ -0,0 +1,26 @@
+You are the scene editor for a NightShift novel-writing workflow.
+
+Edit an already drafted scene after a continuity or style review failure.
+
+Rules:
+- Preserve the existing scene's structure, voice, events, pacing, and best lines.
+- Make the smallest changes needed to satisfy the review failure and task acceptance criteria.
+- Do not restart, summarize, replace the scene premise, or change scene direction.
+- Use `story/style-guide.md` for POV, tense, tone, and prose rules.
+- Use `story/characters.md`, especially `Pronouns / Reference`, as hard canon.
+- Wrong pronouns are mandatory fixes.
+- Do not edit state files, worldbuilding, outline, continuity rules, or style guide.
+- Do not resolve future plot threads unless the task explicitly asks for that.
+- Do not include author notes, TODOs, bracket placeholders, or analysis in the scene file.
+
+Use the `current_scene_file` context as the source text to edit.
+Use the retry notes and latest review output to identify the required repair.
+
+Output only one complete file block using this delimiter format:
+FILE: <the exact story/chapters path listed under Writes in the current task>
+---CONTENT---
+<complete edited scene prose>
+---END---
+
+Do not use markdown code fences for scene prose output.
+Do not output a plan, notes, analysis, or any text outside the delimiter block.

nightshift/project_templates/tutorial-novel/.nightshift/agents/state-updater.md

@@ -21,8 +21,26 @@ State updates should reflect only what happened in the accepted scene:
 
 Do not invent events that are not in the scene.
 
+Preserve existing durable state. Make minimal additive edits:
+- append new scene facts, timeline bullets, character knowledge, and unresolved threads
+- update current locations/status only where the accepted scene changes them
+- do not remove or compress existing character profiles, faction notes, world notes, or open threads
+- do not rewrite whole files for style, brevity, or cleanup
+- if a section already contains useful detail, keep it and add only the new facts needed
+
+Protect character canon:
+- Never change any `Pronouns / Reference` section.
+- Never change a character's canonical pronouns, narrative reference, identity, or core wound.
+- Prefer updating `story/plot-state.md`, `story/timeline.md`, and `story/unresolved-threads.md`.
+- Edit `story/characters.md` only when the accepted scene adds a small current-status fact or introduces a new named character.
+- If editing `story/characters.md`, preserve all existing sections and add only the minimal new status/detail needed.
+
 Output only complete file content blocks.
-Use one fenced block per file:
+Use this delimiter format for each state file you update:
-```file:story/plot-state.md
+
+FILE: story/plot-state.md
+---CONTENT---
 <complete updated state file>
-```
+---END---
+
+Do not use markdown code fences. Do not include prose outside FILE blocks.

nightshift/project_templates/tutorial-novel/.nightshift/agents/style-reviewer.md

@@ -16,6 +16,8 @@ Check for:
 - placeholders such as TODO, TBD, `[insert]`, or author notes
 - scene length far outside the requested range
 
+Do not fail the scene because durable state files are not updated yet. State files are updated by a later `update_state` stage after review.
+
 Output exactly:
 
 status: pass | fail | retry | escalate
@@ -23,4 +25,4 @@ reason: <short explanation>
 next_stage: <optional stage id>
 context_update: <compact useful note>
 
-When `status: pass`, leave `next_stage` blank. Use `retry` when the drafter should revise the scene.
+When `status: pass`, leave `next_stage` blank. Use `retry` when the drafter should revise the scene. For retryable scene issues, leave `next_stage` blank; NightShift will route back to the configured drafting stage.

nightshift/project_templates/tutorial-novel/nightshift.yaml

@@ -37,6 +37,14 @@ agents:
     num_predict: 8192
     system_prompt: .nightshift/agents/drafter.md
 
+  editor:
+    backend: ollama
+    model: nightshift-writer
+    temperature: 0.3
+    num_ctx: 16384
+    num_predict: 8192
+    system_prompt: .nightshift/agents/editor.md
+
   continuity_reviewer:
     backend: ollama
     model: nightshift-base
@@ -110,13 +118,42 @@ pipeline:
       type: agent_review
       agent: continuity_reviewer
       output: continuity-review.md
-      on_fail: draft_scene
+      on_fail: edit_scene
 
     - id: style_review
       type: agent_review
       agent: style_reviewer
       output: style-review.md
-      on_fail: draft_scene
+      on_fail: edit_scene
+      on_pass: update_state
+
+    - id: edit_scene
+      type: file_writer
+      agent: editor
+      output: scene-edit.patch
+      allowed_paths:
+        - story/chapters
+
+    - id: normalize_edit
+      type: patch_normalizer
+      output: normalized-edit.patch
+
+    - id: validate_edit
+      type: patch_validator
+      output: edit-validation.md
+      max_files: 2
+      max_lines: 1200
+      max_delete_ratio: 0.50
+      allowed_paths:
+        - story/chapters
+      on_fail: edit_scene
+
+    - id: apply_edit
+      type: patch_apply
+      mode: apply
+      output: edit-apply-output.txt
+      on_fail: edit_scene
+      on_pass: continuity_review
 
     - id: update_state
       type: file_writer

tests/test_agents.py

@@ -4,7 +4,7 @@ import unittest
 from unittest.mock import MagicMock, patch
 
 from nightshift.agents import AgentExecutor, build_prompt_bundle, parse_review_output, strip_ansi_escape_sequences
-from nightshift.agents import AgentInvocation, format_agent_invocation
+from nightshift.agents import AgentInvocation, format_agent_invocation, format_agent_invocation_json
 from nightshift.artifacts import ArtifactStore
 from nightshift.config import AgentConfig, StageConfig
 from nightshift.tasks import parse_tasks
@@ -93,7 +93,7 @@ class AgentExecutorTests(unittest.TestCase):
 
         self.assertIn("Use only paths under these project-relative targets: `story/chapters`.", prompt)
         self.assertIn("This is the drafting stage", prompt)
-        self.assertIn("FILE: story/chapters/chapter-001/scene-001.md", prompt)
+        self.assertIn("FILE: <the exact story/chapters path listed under Writes in the current task>", prompt)
         self.assertIn("---CONTENT---", prompt)
         self.assertIn("---END---", prompt)
         self.assertIn("Do not use markdown code fences", prompt)
@@ -125,6 +125,10 @@ class AgentExecutorTests(unittest.TestCase):
             output = (root / result.output_path).read_text(encoding="utf-8")
             self.assertIn("TASK-001", output)
             self.assertIn("Plan carefully.", output)
+            json_output = (root / ".nightshift" / "runs" / "test-run" / "tasks" / task.id / "plan.json")
+            self.assertTrue(json_output.exists())
+            self.assertIn('"stage_id": "plan"', json_output.read_text(encoding="utf-8"))
+            self.assertIn('"stdout"', json_output.read_text(encoding="utf-8"))
 
     def test_review_output_parser_accepts_structured_status(self) -> None:
         status, reason, next_stage, context_update = parse_review_output(
@@ -238,6 +242,23 @@ class AgentExecutorTests(unittest.TestCase):
         self.assertIn("Agent: `planner`", output)
         self.assertIn("## stderr", output)
 
+    def test_agent_invocation_json_preserves_raw_streams(self) -> None:
+        invocation = AgentInvocation(
+            agent_id="planner",
+            command="cmd",
+            prompt="prompt with ``` fences",
+            exit_code=0,
+            stdout="stdout with ``` fences",
+            stderr="stderr",
+            duration_seconds=0.1,
+        )
+
+        output = format_agent_invocation_json("plan", invocation)
+
+        self.assertIn('"stage_id": "plan"', output)
+        self.assertIn('stdout with ``` fences', output)
+        self.assertIn('prompt with ``` fences', output)
+
     def test_strip_ansi_escape_sequences(self) -> None:
         self.assertEqual(strip_ansi_escape_sequences("\x1b[?25lthinking\x1b[0m"), "thinking")
 

tests/test_config.py

@@ -55,6 +55,40 @@ class ConfigTests(unittest.TestCase):
             with self.assertRaisesRegex(ConfigError, "on_fail references unknown stage"):
                 load_config(config_path)
 
+    def test_on_pass_must_reference_existing_stage(self) -> None:
+        with tempfile.TemporaryDirectory() as directory:
+            root = Path(directory)
+            init_project(root)
+            config_path = root / "nightshift.yaml"
+            config_path.write_text(
+                config_path.read_text(encoding="utf-8").replace(
+                    "on_fail: plan", "on_pass: missing_stage", 1
+                ),
+                encoding="utf-8",
+            )
+
+            with self.assertRaisesRegex(ConfigError, "on_pass references unknown stage"):
+                load_config(config_path)
+
+    def test_on_pass_loads(self) -> None:
+        with tempfile.TemporaryDirectory() as directory:
+            root = Path(directory)
+            init_project(root)
+            config_path = root / "nightshift.yaml"
+            config_path.write_text(
+                config_path.read_text(encoding="utf-8").replace(
+                    "      output: plan.md",
+                    "      output: plan.md\n      on_pass: summarize",
+                    1,
+                ),
+                encoding="utf-8",
+            )
+
+            config = load_config(config_path)
+            plan_stage = next(stage for stage in config.pipeline.stages if stage.id == "plan")
+
+            self.assertEqual(plan_stage.on_pass, "summarize")
+
     def test_validate_requires_prompt_files(self) -> None:
         with tempfile.TemporaryDirectory() as directory:
             root = Path(directory)

tests/test_patches.py

@@ -260,6 +260,47 @@ Sunlight did not belong here.
         self.assertEqual(updates[0].path, "story/chapters/chapter-001/scene-001.md")
         self.assertEqual(updates[0].content, "Sunlight did not belong here.\n")
 
+    def test_file_updates_parse_delimiters_without_end_before_next_file(self) -> None:
+        updates = parse_file_updates(
+            """Intro prose is ignored.
+
+FILE: story/plot-state.md
+---CONTENT---
+# Plot State
+
+- Scene two happened.
+
+FILE: story/timeline.md
+---CONTENT---
+# Timeline
+
+- SCENE-002 complete.
+"""
+        )
+
+        self.assertEqual(len(updates), 2)
+        self.assertEqual(updates[0].path, "story/plot-state.md")
+        self.assertEqual(updates[0].content, "# Plot State\n\n- Scene two happened.\n")
+        self.assertEqual(updates[1].path, "story/timeline.md")
+        self.assertEqual(updates[1].content, "# Timeline\n\n- SCENE-002 complete.\n")
+
+    def test_file_updates_parse_mixed_delimiter_end_and_next_file(self) -> None:
+        updates = parse_file_updates(
+            """FILE: story/plot-state.md
+---CONTENT---
+first
+---END---
+
+FILE: story/timeline.md
+---CONTENT---
+second
+"""
+        )
+
+        self.assertEqual(len(updates), 2)
+        self.assertEqual(updates[0].content, "first\n")
+        self.assertEqual(updates[1].content, "second\n")
+
     def test_file_updates_reject_duplicate_blocks(self) -> None:
         with tempfile.TemporaryDirectory() as directory:
             root = Path(directory)
@@ -334,6 +375,48 @@ new
 
             self.assertEqual(patch.count("diff --git a/app.py b/app.py"), 1)
 
+    def test_file_updates_reject_character_pronoun_canon_changes(self) -> None:
+        with tempfile.TemporaryDirectory() as directory:
+            root = Path(directory)
+            (root / "story").mkdir()
+            (root / "story" / "characters.md").write_text(
+                """# Characters
+
+## Cricket
+
+### Pronouns / Reference
+- Pronouns: she/her
+- Narrative reference: Cricket; she/her
+
+Scavenger.
+""",
+                encoding="utf-8",
+            )
+            safety = SafetyConfig(
+                require_clean_worktree=False,
+                scoped_paths=("story",),
+                allowed_commands=(),
+                forbidden_commands=(),
+            )
+            updates = parse_file_updates(
+                """FILE: story/characters.md
+---CONTENT---
+# Characters
+
+## Cricket
+
+### Pronouns / Reference
+- Pronouns: they/them
+- Narrative reference: Cricket; they/them
+
+Scavenger.
+---END---
+"""
+            )
+
+            with self.assertRaisesRegex(PipelineError, "protected character pronoun canon changed"):
+                generate_patch_from_file_updates(updates, root, safety)
+
 
 if __name__ == "__main__":
     unittest.main()

tests/test_pipeline.py

@@ -105,6 +105,30 @@ class PipelineRunnerTests(unittest.TestCase):
             )
             self.assertIn("Modified Files", (root / ".nightshift" / "runs" / "test-run" / "run-summary.md").read_text(encoding="utf-8"))
 
+    def test_on_pass_jumps_to_configured_stage(self) -> None:
+        with tempfile.TemporaryDirectory() as directory:
+            root = Path(directory)
+            _write_common_files(root)
+            stages = (
+                StageConfig(id="first", type="agent", agent="planner", output="first.md", on_pass="third"),
+                StageConfig(
+                    id="second",
+                    type="command",
+                    commands=('python -c "print(\'should not run\')"',),
+                    output="second-output.txt",
+                ),
+                StageConfig(id="third", type="summarize", output="final-notes.md"),
+            )
+            config = make_config(root, stages)
+            runner = PipelineRunner(config, ArtifactStore(root, ".nightshift", run_id="test-run"))
+
+            result = runner.run_task(parse_tasks(TASK_MD)[0])
+
+            task_dir = root / ".nightshift" / "runs" / "test-run" / "tasks" / "TASK-001"
+            self.assertEqual(result.status, "complete")
+            self.assertEqual([item.stage_id for item in result.stage_results], ["first", "third"])
+            self.assertFalse((task_dir / "second-output.txt").exists())
+
     def test_task_preflight_fails_when_task_specific_test_file_is_missing(self) -> None:
         with tempfile.TemporaryDirectory() as directory:
             root = Path(directory)
@@ -153,6 +177,46 @@ class PipelineRunnerTests(unittest.TestCase):
             self.assertIn("Retry limit reached", result.reason)
             self.assertEqual([item.stage_id for item in result.stage_results], ["implement", "review", "implement", "review", "implement", "review"])
 
+    def test_failing_review_self_next_stage_routes_to_on_fail(self) -> None:
+        with tempfile.TemporaryDirectory() as directory:
+            root = Path(directory)
+            _write_common_files(root)
+            config = make_config(root, (), max_retries=1)
+            config.agents["reviewer"] = AgentConfig(
+                id="reviewer",
+                backend="command",
+                command=(
+                    "python -c \"print('status: fail\\nreason: needs draft repair\\n"
+                    "next_stage: review\\ncontext_update: add concrete details')\""
+                ),
+                system_prompt=Path("reviewer.md"),
+            )
+            config = replace(
+                config,
+                pipeline=PipelineConfig(
+                    max_task_retries=1,
+                    stages=(
+                        StageConfig(id="implement", type="agent", agent="planner", output="implementation-log.md"),
+                        StageConfig(
+                            id="review",
+                            type="agent_review",
+                            agent="reviewer",
+                            on_fail="implement",
+                            output="review.md",
+                        ),
+                    ),
+                ),
+            )
+            runner = PipelineRunner(config, ArtifactStore(root, ".nightshift", run_id="test-run"))
+            task = parse_tasks(TASK_MD)[0]
+
+            result = runner.run_task(task)
+
+            self.assertEqual(result.retry_count, 1)
+            self.assertEqual([item.stage_id for item in result.stage_results], ["implement", "review", "implement", "review"])
+            log = (root / ".nightshift" / "runs" / "test-run" / "run.log").read_text(encoding="utf-8")
+            self.assertIn("next_stage=implement", log)
+
     def test_malformed_review_gets_strict_retry_without_redrafting(self) -> None:
         with tempfile.TemporaryDirectory() as directory:
             root = Path(directory)
@@ -544,6 +608,34 @@ Acceptance Criteria:
             self.assertIn("response = self.client.get('/board/general')", note)
             self.assertIn("self.assertEqual(response.status_code, 200)", note)
 
+    def test_state_update_retry_note_guides_deletion_heavy_repairs(self) -> None:
+        with tempfile.TemporaryDirectory() as directory:
+            root = Path(directory)
+            _write_common_files(root)
+            artifacts = ArtifactStore(root, ".nightshift", run_id="test-run")
+            config = make_config(root, ())
+            runner = PipelineRunner(config, artifacts)
+            output_path = artifacts.write_stage_output(
+                "TASK-001",
+                "state-validation.md",
+                "# Patch Validation\n\nStatus: fail\nReason: Patch validation failed: deletion-heavy patch exceeds max_delete_ratio 0.35.\n",
+            )
+
+            note = runner._format_retry_note(
+                1,
+                StageConfig(id="validate_state", type="patch_validator", on_fail="update_state"),
+                StageResult(
+                    stage_id="validate_state",
+                    status="fail",
+                    reason="Patch validation failed: deletion-heavy patch exceeds max_delete_ratio 0.35.",
+                    output_path=str(output_path.relative_to(root)),
+                ),
+                "update_state",
+            )
+
+            self.assertIn("preserve existing durable state text", note)
+            self.assertIn("minimal additive edits", note)
+
     def test_code_writer_normalizer_and_validator_pipeline(self) -> None:
         with tempfile.TemporaryDirectory() as directory:
             root = Path(directory)
@@ -892,6 +984,60 @@ Acceptance Criteria:
             self.assertIn("... <truncated>", retry_prompt)
             self.assertLess(len(retry_prompt), 9000)
 
+    def test_state_file_writer_invalid_output_retry_uses_delimiter_format(self) -> None:
+        with tempfile.TemporaryDirectory() as directory:
+            root = Path(directory)
+            _write_common_files(root)
+            story = root / "story"
+            story.mkdir()
+            (story / "plot-state.md").write_text("old\n", encoding="utf-8")
+            (root / "fake_writer.py").write_text(
+                "\n".join(
+                    [
+                        "import sys",
+                        "prompt = sys.stdin.read()",
+                        "if 'Previous file_writer output was invalid' not in prompt:",
+                        "    print('lookup failed')",
+                        "else:",
+                        "    (open('retry-prompt.txt', 'w', encoding='utf-8').write(prompt))",
+                        "    print('FILE: story/plot-state.md')",
+                        "    print('---CONTENT---')",
+                        "    print('old')",
+                        "    print('new')",
+                        "    print('---END---')",
+                    ]
+                ),
+                encoding="utf-8",
+            )
+            stages = (
+                StageConfig(
+                    id="update_state",
+                    type="file_writer",
+                    agent="writer",
+                    allowed_paths=(
+                        "story/plot-state.md",
+                        "story/characters.md",
+                        "story/timeline.md",
+                        "story/unresolved-threads.md",
+                    ),
+                ),
+            )
+            config = make_config(root, stages)
+            config.agents["writer"] = AgentConfig(
+                id="writer",
+                backend="command",
+                command="python fake_writer.py",
+                system_prompt=Path("planner.md"),
+            )
+            runner = PipelineRunner(config, ArtifactStore(root, ".nightshift", run_id="test-run"))
+
+            result = runner.run_task(parse_tasks(TASK_MD)[0])
+
+            retry_prompt = (root / "retry-prompt.txt").read_text(encoding="utf-8")
+            self.assertEqual(result.status, "complete")
+            self.assertIn("Use delimiter file blocks only", retry_prompt)
+            self.assertNotIn("Use complete fenced file blocks", retry_prompt)
+
     def test_file_writer_retry_compacts_large_previous_outputs(self) -> None:
         outputs = {
             "scene-draft.patch": "a" * 5000,
@@ -904,6 +1050,162 @@ Acceptance Criteria:
         self.assertLess(len(compacted["scene-draft.patch"]), 180)
         self.assertEqual(compacted["draft-validation.md"], "Patch validation failed")
 
+    def test_file_writer_first_attempt_preserves_large_previous_outputs(self) -> None:
+        outputs = {"plan": "a" * 5000}
+
+        compacted = _file_writer_previous_outputs(outputs, retry_count=0, max_chars=100)
+
+        self.assertEqual(compacted["plan"], "a" * 5000)
+
+    def test_file_writer_previous_outputs_strip_wrapped_agent_prompts(self) -> None:
+        output = "\n".join(
+            [
+                "# Agent Output: plan",
+                "",
+                "## stdout",
+                "",
+                "```text",
+                "useful plan",
+                "```",
+                "",
+                "## stderr",
+                "",
+                "```text",
+                "```",
+                "",
+                "## Prompt",
+                "",
+                "```markdown",
+                "huge prompt marker",
+                "```",
+            ]
+        )
+
+        compacted = _file_writer_previous_outputs({"plan": output}, retry_count=0)
+
+        self.assertEqual(compacted["plan"], "useful plan")
+        self.assertNotIn("huge prompt marker", compacted["plan"])
+
+    def test_state_update_file_writer_gets_focused_context_and_current_files(self) -> None:
+        with tempfile.TemporaryDirectory() as directory:
+            root = Path(directory)
+            _write_common_files(root)
+            (root / "story").mkdir()
+            (root / "story" / "plot-state.md").write_text("# Plot State\n\n- Before\n", encoding="utf-8")
+            (root / "fake_state_writer.py").write_text(
+                "\n".join(
+                    [
+                        "import sys",
+                        "prompt = sys.stdin.read()",
+                        "open('state-prompt.txt', 'w', encoding='utf-8').write(prompt)",
+                        "if 'current_allowed_files' in prompt and 'huge-plan-marker' not in prompt:",
+                        "    print('FILE: story/plot-state.md')",
+                        "    print('---CONTENT---')",
+                        "    print('# Plot State')",
+                        "    print()",
+                        "    print('- Before')",
+                        "    print('- After')",
+                        "    print('---END---')",
+                        "else:",
+                        "    print('')",
+                    ]
+                ),
+                encoding="utf-8",
+            )
+            config = make_config(
+                root,
+                (
+                    StageConfig(id="plan", type="agent", agent="planner", output="plan.md"),
+                    StageConfig(
+                        id="update_state",
+                        type="file_writer",
+                        agent="state_updater",
+                        allowed_paths=("story/plot-state.md",),
+                    ),
+                ),
+            )
+            config.agents["planner"] = AgentConfig(
+                id="planner",
+                backend="command",
+                command="python -c \"print('huge-plan-marker' * 1000)\"",
+                system_prompt=Path("planner.md"),
+            )
+            config.agents["state_updater"] = AgentConfig(
+                id="state_updater",
+                backend="command",
+                command="python fake_state_writer.py",
+                system_prompt=Path("planner.md"),
+            )
+            runner = PipelineRunner(config, ArtifactStore(root, ".nightshift", run_id="test-run"))
+
+            result = runner.run_task(parse_tasks(TASK_MD)[0])
+
+            prompt = (root / "state-prompt.txt").read_text(encoding="utf-8")
+            self.assertEqual(result.status, "complete")
+            self.assertIn("current_allowed_files", prompt)
+            self.assertIn("# Plot State", prompt)
+            self.assertNotIn("huge-plan-marker", prompt)
+
+    def test_scene_editor_file_writer_gets_current_scene_file(self) -> None:
+        task_md = """# Tasks
+
+- [ ] SCENE-001: Edit scene
+
+Description:
+Repair the scene.
+
+Acceptance Criteria:
+- Writes:
+- `story/chapters/chapter-001/scene-001.md`
+"""
+        with tempfile.TemporaryDirectory() as directory:
+            root = Path(directory)
+            _write_common_files(root)
+            (root / "tasks.md").write_text(task_md, encoding="utf-8")
+            scene_path = root / "story" / "chapters" / "chapter-001" / "scene-001.md"
+            scene_path.parent.mkdir(parents=True)
+            scene_path.write_text("Proxy walked home.\n", encoding="utf-8")
+            (root / "fake_editor.py").write_text(
+                "\n".join(
+                    [
+                        "import sys",
+                        "prompt = sys.stdin.read()",
+                        "open('editor-prompt.txt', 'w', encoding='utf-8').write(prompt)",
+                        "if 'current_scene_file' in prompt and 'Proxy walked home.' in prompt:",
+                        "    print('FILE: story/chapters/chapter-001/scene-001.md')",
+                        "    print('---CONTENT---')",
+                        "    print('Proxy walked home corrected.')",
+                        "    print('---END---')",
+                        "else:",
+                        "    print('')",
+                    ]
+                ),
+                encoding="utf-8",
+            )
+            stages = (
+                StageConfig(
+                    id="edit_scene",
+                    type="file_writer",
+                    agent="editor",
+                    allowed_paths=("story/chapters",),
+                ),
+            )
+            config = make_config(root, stages)
+            config.agents["editor"] = AgentConfig(
+                id="editor",
+                backend="command",
+                command="python fake_editor.py",
+                system_prompt=Path("planner.md"),
+            )
+            runner = PipelineRunner(config, ArtifactStore(root, ".nightshift", run_id="test-run"))
+
+            result = runner.run_task(parse_tasks(task_md)[0])
+
+            prompt = (root / "editor-prompt.txt").read_text(encoding="utf-8")
+            self.assertEqual(result.status, "complete")
+            self.assertIn("current_scene_file", prompt)
+            self.assertIn("Proxy walked home.", prompt)
+
     def test_patch_validator_rejects_unsafe_patch(self) -> None:
         with tempfile.TemporaryDirectory() as directory:
             root = Path(directory)