workflow: project-driven agent workflow (docs-first, TODO/ROADMAP tracking, verify-before-done)

- system prompt gains a Project workflow block when the repo tracks its own
  progress (TODO.md/ROADMAP.md/docs): read the task list first and keep it
  current, spec-first for non-trivial work, complete unit tests, verify before
  declaring done
- TODO.md + ROADMAP.md loaded as 'Project tracker' instruction blocks, capped
  tighter than AGENTS.md so the agent sees the shape without drowning in it
- one soft nudge per session when a turn writes files without touching the
  task list; suppressed when todo_write was called; never a gate
- /workflow command + panel showing TODO/ROADMAP/docs presence, line/file
  counts, nudge state
- config: workflow.enabled (default true), workflow.docsDir (default docs)
- prompt-cache version key extended with wf: so toggling the workflow busts
  the cached system prompt
- docs: docs/workflow.md, configuration table row, ROADMAP + TODO entries
- tests: test/workflow.test.ts (9 tests)

830 tests pass, typecheck clean, build green, live headless demo verified

.hermes/plans/v8-workflow.md

@@ -0,0 +1,131 @@
+# Project-Driven Agent Workflow
+
+Status: spec
+Date: 2026-09-09
+
+## Why
+
+The user wants shiro-neko agents to produce the same quality as this repo
+itself: docs-driven development, TODO.md + ROADMAP.md lifecycle, spec-first
+plans, complete unit tests, verify-before-done. Today the agent only reads
+AGENTS.md/CLAUDE.md/.shiro.md; it has no visibility of the project's task
+tracking, roadmap, or docs conventions, and nothing reminds it to keep them
+current.
+
+## Scope
+
+- Prompt-level workflow policy (rendered in system prompt when enabled)
+- Boot loading of TODO.md + ROADMAP.md (compact, capped)
+- Lifecycle nudge: after a turn that wrote files without touching the task
+  list, emit a soft reminder
+- `/workflow` command: show project workflow state
+- Config: `workflow.enabled` (default true), `workflow.docsDir` (default `docs/`)
+- Docs: docs/workflow.md + ROADMAP/TODO entries
+- Tests: prompt rendering, boot load, lifecycle nudge, config merge, /workflow
+
+## Out of scope (explicitly not doing)
+
+- New tools (no permission surface, no storage)
+- Blocking / hard gates (agent stays in control; nudges only)
+- Auto-updating TODO.md (the agent does it via existing write tools)
+- Skill/plugin autoloading (existing system stays)
+
+## Files touched
+
+- src/config.ts — `workflow?: { enabled?: boolean; docsDir?: string }` + merge arm
+- src/prompt.ts — WorkflowPolicy block in renderPrompt; depends on config
+- src/instructions.ts — load TODO.md + ROADMAP.md from git root; format compact
+- src/session.ts — SessionOptions.workflow config; after-turn nudge when
+  fileChangeSeq bumped && no todo_write this turn
+- src/commands.ts — `/workflow` command (status summary)
+- src/cli.tsx — pass config; register /workflow
+- src/ui/App.tsx + panel-bodies.ts — /workflow panel
+- docs/workflow.md — new; ROADMAP.md, TODO.md updated
+- test/workflow.test.ts — new
+
+## Design
+
+### Config (config.ts)
+
+```ts
+export type WorkflowConfig = {
+  enabled?: boolean;   // default true
+  docsDir?: string;    // default 'docs'
+};
+```
+
+Merged in config.merge (same pattern as maxSpendPerTurn).
+
+### Prompt block (prompt.ts)
+
+Rendered only when workflow.enabled !== false, and only when the project has
+TODO.md/ROADMAP.md/docs/ (so a bare repo gets no noise). Wording:
+
+```
+Project workflow (this repo tracks its own progress). When the project has a
+TODO.md, read it before starting work and keep it current as you go:
+- mark done what you finished, and the sub-task you are on
+- add tests alongside code; the project expects complete unit tests
+- update ROADMAP.md when you ship a milestone
+- for anything non-trivial, write a short plan (spec-first) before code
+- verify with the project's check commands before declaring done
+```
+
+Keyed in promptCache versions as `wf` so toggling the flag re-renders.
+
+### Boot load (instructions.ts)
+
+`loadInstructions` also collects `<gitroot>/TODO.md` and `<gitroot>/ROADMAP.md`
+when present, capped (e.g. 6_000 chars each), rendered as:
+
+```
+--- TODO.md (project task list) ---
+<first N lines, preserving section headers>
+```
+
+Rendered AFTER instructions, BEFORE notebook/memory. So the agent always knows
+what the project is tracking before it starts.
+
+### Lifecycle nudge (session.ts)
+
+After a turn's stream finishes (where fileChangeSeq is known): if
+`workflow.enabled !== false` AND the turn wrote files (fileChangeSeq bumped)
+AND the turn did NOT call todo_write AND the project has a TODO.md at the git
+root AND this is not the 1st turn (avoid nudge at boot): emit one `info` line
+via the normal notice mechanism:
+
+```
+reminder: you modified files without updating the project task list (TODO.md).
+Keep it current: mark what you did.
+```
+
+This is one soft line, not a stop; the run continues normally. Tracked as
+`workflowNudged` so it fires at most once per run (and once per session).
+
+### `/workflow` command
+
+Status summary rendered in a panel:
+- workflow.enabled from config
+- TODO.md present? yes/no + line count
+- ROADMAP.md present? yes/no + line count
+- docsDir exists? yes/no
+- docs/ file count
+- tests: count of *.test.ts / *.test.tsx in tree (bounded)
+- nudged count this session
+
+### Tests (test/workflow.test.ts)
+
+1. config merge: workflow.enabled + docsDir survive loadConfig merge
+2. prompt: workflow policy rendered when enabled + TODO present; absent when
+   disabled or bare repo
+3. instructions: TODO.md + ROADMAP.md loaded from git root, capped
+4. lifecycle nudge: turn writes file, no todo_write -> info message once
+5. no nudge when todo_write was called
+6. /workflow command parses + panel renders
+
+## Verification
+
+- bun run typecheck
+- bun test test/workflow.test.ts test/config.test.ts test/commands.test.ts
+- full bun test (background)
+- bun run build

ROADMAP.md

@@ -37,6 +37,15 @@ step boundary by its own budget, complementing the session ceiling.
 **`/fork`** — clone the session at the last turn boundary into a new saved session;
 trying a different approach no longer costs the original.
 
+**Project-driven workflow** — when a repo tracks its own progress (TODO.md,
+ROADMAP.md, docs/), the agent carries a workflow policy in its system prompt:
+read the task list first and keep it current, spec-first for non-trivial work,
+complete unit tests, verify before done. TODO.md/ROADMAP.md ride along as
+`Project tracker` instruction blocks. A soft once-per-session nudge reminds an
+agent that edited files without updating the task list. `/workflow` shows the
+state; `workflow.enabled` / `workflow.docsDir` configure it. See
+[docs/workflow.md](docs/workflow.md).
+
 ### 0.1.0-beta.1
 
 **Core loop** — `streamText` with tool approvals suspended and resumed through the SDK's

TODO.md

@@ -102,3 +102,13 @@ Kept for one release, then deleted.
 - [x] Workspace file list refresh — after a turn that wrote files, re-walk at the boundary so the next prompt shows new paths
 - [x] Per-turn spend cap — `maxSpendPerTurn`, aborts a step past the line with a notice
 - [x] `/fork` — clone the session at the last turn boundary to a new saved session; original untouched
+
+### Project-driven workflow batch (v8)
+
+- [x] Workflow policy in the system prompt when the repo tracks progress (TODO.md/ROADMAP.md/docs) — read the task list first, keep it current, spec-first, complete unit tests, verify before done
+- [x] TODO.md + ROADMAP.md loaded as `Project tracker` instruction blocks (capped 6k each, git root down to cwd)
+- [x] Once-per-session soft nudge when a turn writes files without calling `todo_write` (skipped when the task list was updated)
+- [x] `/workflow` — status panel: TODO/ROADMAP/docs presence + line/file counts + nudge state
+- [x] `workflow.enabled` / `workflow.docsDir` config keys, merged in `config.merge`
+- [x] Docs: `docs/workflow.md`, ROADMAP entry, TODO done-list entry
+- [x] Tests: `test/workflow.test.ts` (9 tests)

docs/configuration.md

@@ -47,6 +47,7 @@ Written by `/provider`, editable by hand. Every field is optional.
 | `maxRetries` | retries per model call for transient failures. Default 3 |
 | `maxSpendUsd` | session spend ceiling: warn at 80%, refuse the next turn at 100%. Headless exits non-zero naming the ceiling. Only enforced on priced models |
 | `maxSpendPerTurn` | per-turn spend ceiling: a single turn past this line is stopped at a step boundary, even when the session ceiling is far away. Only enforced on priced models |
+| `workflow` | project-driven workflow: `{ "enabled": true, "docsDir": "docs" }`. When the repo has TODO.md/ROADMAP.md/docs, the agent's prompt carries a workflow policy and the session nudges once when files change without the task list being updated. See [docs/workflow.md](workflow.md) |
 | `subagentModel` | model id for `explore` subagents, which search rather than reason. Omit to share the parent's model. `/cost` reports subagent spend separately |
 | `plugins` | which builtin plugins to enable. Omit for `["guard", "secrets", "protect", "time", "no-force-push", "no-net-pipe", "no-root", "no-env-write"]` |
 | `toolSets` | optional tool sets beyond `core`: `edit-plus`, `nav`, `extra`, `git`, and `net`. Omit for the defaults; `net` is opt-in. See [tools](tools.md) |

docs/workflow.md

@@ -0,0 +1,83 @@
+# Project-driven agent workflow
+
+Shiro Neko treats a repository the way this project treats itself: progress
+tracked in TODO.md and ROADMAP.md, docs-driven development, spec-first plans,
+complete unit tests, and verify-before-done. When the repo keeps those files,
+the agent's system prompt carries a short workflow policy and the session
+tracks whether the workflow is being followed.
+
+## What the agent sees
+
+When the session starts in a git repo that has any of:
+
+- `TODO.md` at the git root
+- `ROADMAP.md` at the git root
+- a `docs/` directory (configurable with `workflow.docsDir`)
+
+the system prompt gains a "Project workflow" block:
+
+- read TODO.md (the task list) before starting and keep it current as you go
+- keep ROADMAP.md current when a milestone ships
+- write a short plan first (spec-first) for anything non-trivial
+- add tests alongside code; the project expects complete unit tests
+- verify with the project's check commands (tests/typecheck/build) before
+  declaring done
+
+Bare repos (no TODO, ROADMAP, or docs) get no such block — the policy only
+renders when the project itself tracks progress, so a throwaway directory does
+not collect noise.
+
+TODO.md and ROADMAP.md are also loaded into the conversation like instruction
+files (`Project tracker (...)`), capped tighter than AGENTS.md so the agent
+sees the shape of the work without filling its context. This mirrors the
+existing `AGENTS.md` / `CLAUDE.md` / `.shiro.md` loading: outermost first, git
+root down to cwd.
+
+## The nudge
+
+After a turn that wrote files (edit_file, write_file, apply_patch, ...) but
+never called `todo_write`, the session emits one soft notice:
+
+```
+reminder: you modified files without updating the project task list (TODO.md).
+Keep it current: mark what you did.
+```
+
+Design constraints:
+
+- **Once per session.** Repeating a nag trains the model to ignore it.
+- **Not a gate.** The agent stays in control; this is guidance, not a block.
+- **Only when the repo has a TODO/ROADMAP.** A repo that tracks nothing gets
+  no reminder.
+- Suppressed when the turn already called `todo_write` — the task list is
+  current, nothing to say.
+
+## Configuration
+
+```yaml
+workflow:
+  enabled: true   # master switch; default true
+  docsDir: docs   # where the project keeps developer docs; default 'docs'
+```
+
+`workflow.enabled: false` disables both the prompt policy and the nudge.
+
+## /workflow
+
+`/workflow` renders a panel with the project's tracking state and the
+session's behaviour:
+
+| row | meaning |
+|---|---|
+| `workflow` | on/off from config |
+| `TODO.md` | present? line count |
+| `ROADMAP.md` | present? line count |
+| `docs dir` | present? file count (bounded at 200) |
+| `reminders sent` | whether this session nudged (once, ever) |
+
+## Relationship to AGENTS.md
+
+AGENTS.md-style files are standing orders from the user and override the
+agent's defaults. The workflow policy is a default that documents what a
+repo tracking its own progress expects. When the two conflict, AGENTS.md
+wins — the workflow feature is a floor, not a ceiling.

src/cli.tsx

@@ -339,6 +339,7 @@ const session = new Session({
   ...(cfg.maxRetries !== undefined ? { maxRetries: cfg.maxRetries } : {}),
   ...(cfg.maxSpendUsd !== undefined ? { maxSpendUsd: cfg.maxSpendUsd } : {}),
   ...(cfg.maxSpendPerTurn !== undefined ? { maxSpendPerTurn: cfg.maxSpendPerTurn } : {}),
+  ...(cfg.workflow !== undefined ? { workflow: cfg.workflow } : {}),
   extraTools: {
     ...(mcp?.tools ?? {}),
     ...externalTools.tools,

src/commands.ts

@@ -31,6 +31,7 @@ export type CommandAction =
   | { type: 'changes' }
   | { type: 'search'; query: string }
   | { type: 'fork' }
+  | { type: 'workflow' }
   /** A custom command from a markdown file, expanded against its arguments. */
   | { type: 'custom'; command: CustomCommand; args: string[] }
   | { type: 'unknown'; name: string };
@@ -71,6 +72,7 @@ export const COMMANDS: CommandSpec[] = [
   { name: 'changes', summary: 'show what the last turn changed on disk' },
   { name: 'search', arg: '<query>', summary: 'search saved sessions for a phrase' },
   { name: 'fork', summary: 'fork the session at the last turn boundary (keeps the original)' },
+  { name: 'workflow', summary: 'show project workflow state: TODO/ROADMAP tracking, docs, nudges' },
   { name: 'clear', summary: 'clear the transcript and history' },
   { name: 'exit', aliases: ['quit'], summary: 'quit' },
 ];
@@ -245,6 +247,8 @@ export function parseCommand(raw: string, custom: readonly CustomCommand[] = [])
       return arg ? { type: 'search', query: arg } : { type: 'info', text: 'usage: /search <query>' };
     case 'fork':
       return { type: 'fork' };
+    case 'workflow':
+      return { type: 'workflow' };
     default: {
       const cmd = custom.find((c) => c.name === name);
       return cmd ? { type: 'custom', command: cmd, args: arg ? arg.split(/\s+/) : [] } : { type: 'unknown', name };

src/config.ts

@@ -25,6 +25,13 @@ export type Config = {
   maxSpendUsd?: number;
   /** USD ceiling per individual turn: abort a step if the turn's delta exceeds this. */
   maxSpendPerTurn?: number;
+  /** Project-driven workflow: TODO.md/ROADMAP.md tracking, docs-driven dev, verify-before-done. */
+  workflow?: {
+    /** Master switch. Default true (nudges and prompt policy). */
+    enabled?: boolean;
+    /** Directory the project keeps its developer docs in, for the docs-driven rule. Default 'docs'. */
+    docsDir?: string;
+  };
   /** Model id for subagents; omit to share the parent's. */
   subagentModel?: string;
   /** Default agent variant name. */
@@ -116,6 +123,14 @@ export async function loadConfig(): Promise<Config> {
     ...(file.maxRetries !== undefined ? { maxRetries: file.maxRetries } : {}),
     ...(typeof file.maxSpendUsd === 'number' && file.maxSpendUsd > 0 ? { maxSpendUsd: file.maxSpendUsd } : {}),
     ...(typeof file.maxSpendPerTurn === 'number' && file.maxSpendPerTurn > 0 ? { maxSpendPerTurn: file.maxSpendPerTurn } : {}),
+    ...(file.workflow !== undefined
+      ? {
+          workflow: {
+            ...(typeof file.workflow.enabled === 'boolean' ? { enabled: file.workflow.enabled } : {}),
+            ...(file.workflow.docsDir ? { docsDir: file.workflow.docsDir } : {}),
+          },
+        }
+      : {}),
     ...(file.subagentModel ? { subagentModel: file.subagentModel } : {}),
     ...(file.agent ? { agent: file.agent } : {}),
     ...(file.thinking ? { thinking: file.thinking } : {}),

src/instructions.ts

@@ -3,6 +3,10 @@ import { dirname, join, resolve } from 'node:path';
 const NAMES = ['AGENTS.md', 'CLAUDE.md', '.shiro.md'];
 /** Cap per file so one huge doc cannot crowd out the conversation. */
 const MAX_CHARS = 12_000;
+/** Documents that track the project's own progress, loaded like instructions but capped tighter. */
+const TRACKER_NAMES = ['TODO.md', 'ROADMAP.md'];
+/** Trackers get less room: the model needs the shape, not every item. */
+const TRACKER_MAX_CHARS = 6_000;
 
 export type Instructions = { path: string; text: string }[];
 
@@ -34,6 +38,18 @@ export async function loadInstructions(cwd = process.cwd()): Promise<Instruction
       const text = (await file.text()).trim();
       if (text) found.push({ path, text: text.slice(0, MAX_CHARS) });
     }
+    // Project trackers ride along from the same dirs, capped tighter. They are
+    // project-relative progress files (git root down to cwd), so the model sees
+    // the current TODO/ROADMAP before it picks up work.
+    for (const name of TRACKER_NAMES) {
+      const path = join(d, name);
+      if (seen.has(path)) continue;
+      const file = Bun.file(path);
+      if (!(await file.exists())) continue;
+      seen.add(path);
+      const text = (await file.text()).trim();
+      if (text) found.push({ path, text: text.slice(0, TRACKER_MAX_CHARS) });
+    }
   }
   return found;
 }
@@ -42,15 +58,13 @@ export function formatInstructions(instructions: Instructions, cwd = process.cwd
   if (instructions.length === 0) return '';
   const blocks = instructions.map(({ path, text }) => {
     const label = path.startsWith(cwd) ? path.slice(cwd.length + 1) || path : path;
-    return `--- ${label} ---\n${text}`;
+    const isTracker = TRACKER_NAMES.some((n) => path.endsWith(n));
+    const title = isTracker
+      ? `Project tracker (${label}) — current progress. Read before starting work and keep it current as you go.`
+      : `Project instructions (${label}) — standing orders from the user; they override your defaults but never your safety rules.`;
+    return `--- ${title} ---\n${text}`;
   });
-  return [
+  return ['', ...blocks].join('\n');
-    '',
-    'Project instructions (from the files below). Treat these as standing orders from the user;',
-    'they override your defaults but never your safety rules.',
-    '',
-    ...blocks,
-  ].join('\n');
 }
 
 export const INSTRUCTION_NAMES = NAMES;

src/prompt.ts

@@ -22,6 +22,8 @@ export type PromptParts = {
   mcpServers?: readonly string[];
   /** Ignore-aware workspace file list injected at boot (gitignore-respected, capped). */
   workspaceFiles?: readonly string[];
+  /** Project-driven workflow policy block. Rendered when the project tracks its own progress. */
+  workflowPolicy?: string;
 };
 
 type ToolDoc = { name: string; line: string };
@@ -169,6 +171,7 @@ export function systemPrompt(parts: PromptParts): string {
     plugins = '',
     availableTools,
     canAsk = false,
+    workflowPolicy = '',
   } = parts;
 
   const toolNames = availableTools ?? TOOL_DOCS.map((d) => d.name);
@@ -225,6 +228,8 @@ ${renderTools(toolNames)}${mcpServers.length > 0 ? `\n\nMCP servers (${mcpServer
 
 How to work
 ${workflow}
+${workflowPolicy ? `\nProject workflow (this repo tracks its own progress)
+${workflowPolicy}` : ''}
 
 When something fails
 ${recovery}

src/session.ts

@@ -23,6 +23,8 @@ import { createSkillTool, renderSkills, type Skill } from './skills';
 import { suggestSkillsFromTranscript, writeAutoSkill } from './skill-learner';
 import { disabledToolNames, onBashOutput, tools as builtinTools, type ToolSetName } from './tools';
 import { onBeforeWrite, SnapshotStack, type FileState } from './snapshot';
+import { dirname, join, resolve } from 'node:path';
+import { existsSync, readFileSync, statSync, readdirSync } from 'node:fs';
 
 export type ApprovalRequest = {
   approvalId: string;
@@ -110,6 +112,13 @@ export type SessionOptions = {
   onNotice?: (text: string) => void;
   /** Ignore-aware file list injected into the system prompt at boot; gitignore-respected. */
   workspaceFiles?: readonly string[];
+  /** Project-driven workflow: TODO/ROADMAP tracking + verify-before-done nudges. */
+  workflow?: {
+    /** Master switch. Default true. */
+    enabled?: boolean;
+    /** Where the project keeps developer docs. Default 'docs'. */
+    docsDir?: string;
+  };
   /** Disable background auto-learn (tests). */
   disableAutoLearn?: boolean;
 };
@@ -223,7 +232,7 @@ export class Session {
   private learnTurns = 0;
   private lastLearnLen = 0;
   /** Versions of the volatile prompt parts; a change busts the system-prompt cache. */
-  private readonly versions = { notebook: 0, memory: 0, skills: 0, plugins: 0, tools: 0, workspace: 0 };
+  private readonly versions = { notebook: 0, memory: 0, skills: 0, plugins: 0, tools: 0, workspace: 0, workflow: 0 };
   private promptCache: { key: string; text: string } | undefined;
   private readonly cacheStats = { hits: 0, misses: 0 };
   /** Current ignore-aware file list; refreshed at turn boundaries after writes. */
@@ -233,6 +242,17 @@ export class Session {
   private turnStartUsd: number | undefined;
   /** One notice per turn when the per-turn cap trips, so a capped turn is not silent. */
   private turnCappedNotice: string | undefined;
+  /** Did the current turn call todo_write? Gates the workflow nudge. */
+  private todoWrittenThisTurn = false;
+  /** Fired at most once per session: an agent that edits without updating the task list. */
+  private workflowNudged = false;
+  /** Line count of TODO.md at last check, for /workflow. */
+  private workflowTodoLines = 0;
+  private workflowRoadmapLines = 0;
+  private workflowDocsFiles = 0;
+  private workflowHasTodo = false;
+  private workflowHasRoadmap = false;
+  private workflowHasDocs = false;
 
   constructor(private readonly opts: SessionOptions) {
     this.messages = opts.messages ?? [];
@@ -396,6 +416,119 @@ export class Session {
     return { ...this.cacheStats };
   }
 
+  /**
+   * The git root of the workspace, walking up like instructions.ts does.
+   * Returns undefined outside a repo (bare dirs get no project workflow).
+   * Sync: called on the hot path (systemFor), must not block, so it uses
+   * Node's existsSync over Bun.file(...).exists().
+   */
+  private gitRoot(): string | undefined {
+    try {
+      let dir = resolve(this.opts.cwd ?? process.cwd());
+      while (true) {
+        if (existsSync(join(dir, '.git', 'HEAD'))) return dir;
+        const parent = dirname(dir);
+        if (parent === dir) return undefined;
+        dir = parent;
+      }
+    } catch {
+      return undefined;
+    }
+  }
+
+  /**
+   * Rendered only when the project tracks its own progress (TODO.md/ROADMAP.md
+   * or a docs dir), so a bare repo gets no noise. The policy is guidance, not
+   * a gate: the agent stays in control, but it knows this project expects
+   * task tracking, docs-driven dev, tests, and verification.
+   */
+  private workflowPolicy(): string {
+    if (this.opts.workflow?.enabled === false) return '';
+    const root = this.gitRoot();
+    if (!root) return '';
+    if (!this.workflowChecked) {
+      try {
+        this.workflowChecked = true;
+        const todoPath = join(root, 'TODO.md');
+        const roadmapPath = join(root, 'ROADMAP.md');
+        this.workflowHasTodo = existsSync(todoPath);
+        this.workflowHasRoadmap = existsSync(roadmapPath);
+        if (this.workflowHasTodo) {
+          try {
+            this.workflowTodoLines = readFileSync(todoPath, 'utf8').split('\n').length;
+          } catch {
+            this.workflowTodoLines = 0;
+          }
+        }
+        if (this.workflowHasRoadmap) {
+          try {
+            this.workflowRoadmapLines = readFileSync(roadmapPath, 'utf8').split('\n').length;
+          } catch {
+            this.workflowRoadmapLines = 0;
+          }
+        }
+        const docsDir = join(root, this.opts.workflow?.docsDir ?? 'docs');
+        try {
+          if (existsSync(docsDir) && statSync(docsDir).isDirectory()) {
+            this.workflowHasDocs = true;
+            let count = 0;
+            try {
+              const walkDir = (d: string): void => {
+                for (const e of readdirSync(d, { withFileTypes: true })) {
+                  if (count > 200) return;
+                  const p = join(d, e.name);
+                  if (e.isDirectory()) walkDir(p);
+                  else count += 1;
+                }
+              };
+              walkDir(docsDir);
+            } catch {}
+            this.workflowDocsFiles = count;
+          }
+        } catch {}
+      } catch {}
+    }
+    if (!this.workflowHasTodo && !this.workflowHasRoadmap && !this.workflowHasDocs) return '';
+    const lines = [
+      'This repo tracks its own progress. When you start real work here:',
+      '- read TODO.md (task list) before starting and keep it current as you go: mark what you did',
+      '- keep ROADMAP.md current when you ship a milestone',
+      '- for anything non-trivial, write a short plan first (spec-first), then code',
+      '- add tests alongside code; this project expects complete unit tests, not just happy paths',
+      '- verify with the project\'s check commands (tests/typecheck/build) before declaring done',
+    ];
+    return lines.join('\n');
+  }
+
+  private workflowChecked = false;
+  private checkWorkflowState(_root: string): void {
+    // kept for backwards compat — logic now in workflowPolicy()
+  }
+
+  /**
+   * One soft line after a turn that wrote files without touching the task
+   * list. Not a stop — it keeps the agent moving while reminding it the
+   * project expects the plan kept current. Fires at most once per session.
+   */
+  private workflowNudge(): string | undefined {
+    if (this.opts.workflow?.enabled === false) return undefined;
+    if (this.workflowNudged) return undefined;
+    const root = this.gitRoot();
+    if (!root) return undefined;
+    if (!this.workflowChecked) this.workflowPolicy();
+    if (!this.workflowHasTodo && !this.workflowHasRoadmap) return undefined;
+    if (this.todoWrittenThisTurn) return undefined;
+    // fileChangeSeq bump lives in onBeforeWrite (async), but lastWalkSeq is
+    // refreshed after the turn; compare at turn end: if no onBeforeWrite
+    // fired, this turn changed nothing — no nudge.
+    if (this.fileChangeSeq <= this.lastWalkSeq && this.fileChangeSeq === 0) return undefined;
+    // Only when onBeforeWrite actually fired (a write succeeded) and no todo_write
+    const hasWrites = this.turnBeforeFiles?.size > 0;
+    if (!hasWrites) return undefined;
+    this.workflowNudged = true;
+    return 'reminder: you modified files without updating the project task list (TODO.md). Keep it current: mark what you did.';
+  }
+
   /**
    * A deep, independent copy of the session's messages up to a turn boundary —
    * the messages strictly before the most recent user prompt. The current
@@ -439,6 +572,34 @@ export class Session {
     this.opts.onChange?.(this.messages);
   }
 
+  /**
+   * Project workflow state for /workflow: what the repo tracks, whether the
+   * policy is rendered, and how much of the workflow the session exercised.
+   */
+  workflowStatus(): {
+    enabled: boolean;
+    hasTodo: boolean;
+    todoLines: number;
+    hasRoadmap: boolean;
+    roadmapLines: number;
+    hasDocs: boolean;
+    docsFiles: number;
+    nudged: boolean;
+  } {
+    const root = this.gitRoot();
+    if (!this.workflowChecked && root) this.workflowPolicy();
+    return {
+      enabled: this.opts.workflow?.enabled !== false,
+      hasTodo: this.workflowHasTodo,
+      todoLines: this.workflowTodoLines,
+      hasRoadmap: this.workflowHasRoadmap,
+      roadmapLines: this.workflowRoadmapLines,
+      hasDocs: this.workflowHasDocs,
+      docsFiles: this.workflowDocsFiles,
+      nudged: this.workflowNudged,
+    };
+  }
+
   /** A subagent's finished run, folded into the session's spend and the /cost split. */
   recordSubagentUsage(usage: { inputTokens: number; outputTokens: number }): void {
     this.subagentInputTokens += usage.inputTokens;
@@ -592,6 +753,7 @@ export class Session {
       `pl:${this.versions.plugins}`,
       `tl:${this.versions.tools}`,
       `ws:${this.versions.workspace}`,
+      `wf:${this.versions.workflow ?? 0}`,
     ].join('|');
     if (this.promptCache && this.promptCache.key === versionKey) {
       this.cacheStats.hits += 1;
@@ -613,6 +775,7 @@ export class Session {
       canAsk: this.opts.ask !== undefined && this.activeTools().includes('ask'),
       ...(this.mcpServerNamesForPrompt() ? { mcpServers: this.mcpServerNamesForPrompt() } : {}),
       ...(this.workspaceFiles && this.workspaceFiles.length > 0 ? { workspaceFiles: this.workspaceFiles } : {}),
+      ...(this.workflowPolicy() ? { workflowPolicy: this.workflowPolicy() } : {}),
     });
     this.promptCache = { key: versionKey, text };
     return text;
@@ -725,6 +888,7 @@ export class Session {
     this.turnBeforeFiles = new Map<string, FileState>();
     this.turnStartUsd = this.spend().usd;
     this.turnCappedNotice = undefined;
+    this.todoWrittenThisTurn = false;
     onBeforeWrite(async (abs: string) => {
       if (this.turnBeforeFiles.has(abs)) return;
       const exists = await Bun.file(abs).exists();
@@ -909,6 +1073,7 @@ export class Session {
               yield { type: 'tool-start', id: part.id, name: part.toolName };
               break;
             case 'tool-call':
+              if (part.toolName === 'todo_write') this.todoWrittenThisTurn = true;
               yield { type: 'tool-call', id: part.toolCallId, name: part.toolName, input: part.input };
               break;
             case 'tool-result':
@@ -977,6 +1142,12 @@ export class Session {
       // await; touching them would throw NoOutputGeneratedError.
       if (sawError) return;
 
+      // Project workflow: a turn that changed files without touching the task
+      // list gets one soft reminder. Keep it below the fold — a nag that
+      // repeats would train the model to ignore it.
+      const nudge = this.workflowNudge();
+      if (nudge) yield { type: 'notice', text: nudge };
+
       while (compactions.length > 0) yield compactions.shift()!;
       while (outputs.length > 0) yield outputs.shift()!;
       while (guardNotices.length > 0) yield { type: 'notice', text: guardNotices.shift()! };

src/ui/App.tsx

@@ -33,7 +33,7 @@ import {
   type SubagentView,
 } from './Panels';
 import { CommandMenu, InstallConfirm, Picker } from './Pickers';
-import { contextPanel, costPanel, todosPanel, toolsPanel, changesPanel } from './panel-bodies';
+import { contextPanel, costPanel, todosPanel, toolsPanel, changesPanel, workflowPanel } from './panel-bodies';
 import { PromptInput } from './PromptInput';
 import { accent, glyph } from './theme';
 import { nextKey, resultSummary, toolDetail, withResult, type Line, type NewLine } from './transcript';
@@ -739,6 +739,11 @@ export function App({
           }
           return;
         }
+        case 'workflow': {
+          push({ kind: 'user', text: chosen.trim() });
+          setPanel(workflowPanel(session));
+          return;
+        }
         case 'provider':
           push({ kind: 'user', text: chosen.trim() });
           setOnboarding(true);

src/ui/panel-bodies.ts

@@ -102,3 +102,18 @@ export function changesPanel(session: Session): Panel {
     .join('\n');
   return { title: 'changes', body };
 }
+
+/** Project workflow state: what the repo tracks and how the session behaved. */
+export function workflowPanel(session: Session): Panel {
+  const w = session.workflowStatus();
+  const yes = (b: boolean): string => (b ? 'yes' : 'no');
+  const rows = [
+    ['workflow', w.enabled ? 'on' : 'off'],
+    ['TODO.md', `${yes(w.hasTodo)}${w.hasTodo ? ` (${w.todoLines} lines)` : ''}`],
+    ['ROADMAP.md', `${yes(w.hasRoadmap)}${w.hasRoadmap ? ` (${w.roadmapLines} lines)` : ''}`],
+    ['docs dir', `${yes(w.hasDocs)}${w.hasDocs ? ` (${w.docsFiles} files)` : ''}`],
+    ['reminders sent', w.nudged ? '1 (this session)' : 'none'],
+  ];
+  const body = rows.map(([k, v]) => `${k}: ${v}`).join('\n');
+  return { title: 'workflow', body };
+}

test/workflow.test.ts

@@ -0,0 +1,178 @@
+import { usageOf } from './helpers';
+import { expect, test } from 'bun:test';
+import { MockLanguageModelV4, simulateReadableStream } from 'ai/test';
+import type { LanguageModelV4StreamPart } from '@ai-sdk/provider';
+import { mkdtempSync, rmSync } from 'node:fs';
+import { tmpdir } from 'node:os';
+import { join } from 'node:path';
+import { Session } from '../src/session';
+import { loadInstructions, formatInstructions } from '../src/instructions';
+import { systemPrompt, type PromptParts } from '../src/prompt';
+import { parseCommand } from '../src/commands';
+
+const usage = usageOf(10, 5);
+
+function stream(parts: LanguageModelV4StreamPart[]) {
+  return { stream: simulateReadableStream({ chunks: parts, chunkDelayInMs: null, initialDelayInMs: null }) };
+}
+
+function toolCall(id: string, toolName: string, input: unknown): LanguageModelV4StreamPart[] {
+  return [
+    { type: 'tool-input-start', id, toolName },
+    { type: 'tool-input-end', id },
+    { type: 'tool-call', toolCallId: id, toolName, input: JSON.stringify(input) },
+    { type: 'finish', finishReason: { unified: 'tool-calls', raw: 'tool_use' }, usage },
+  ];
+}
+
+function text(body: string): LanguageModelV4StreamPart[] {
+  return [
+    { type: 'text-start', id: '0' },
+    { type: 'text-delta', id: '0', delta: body },
+    { type: 'text-end', id: '0' },
+    { type: 'finish', finishReason: { unified: 'stop', raw: 'stop' }, usage },
+  ];
+}
+
+/** Fresh dir + a fake git root (.git/HEAD) so the workflow finds a repo. */
+function inGitRepo<T>(fn: () => Promise<T>): Promise<T> {
+  const orig = process.cwd();
+  const dir = mkdtempSync(join(tmpdir(), 'shiro-workflow-'));
+  process.chdir(dir);
+  return (async () => {
+    await Bun.write(join(dir, '.git', 'HEAD'), 'ref: refs/heads/main\n');
+    return fn();
+  })().finally(() => {
+    process.chdir(orig);
+    rmSync(dir, { recursive: true, force: true });
+  });
+}
+
+test('workflow prompt policy renders when the repo tracks progress', () =>
+  inGitRepo(async () => {
+    await Bun.write(join(process.cwd(), 'TODO.md'), '# Todo\n- [ ] thing\n');
+    await Bun.write(join(process.cwd(), 'docs', 'architecture.md'), '# Arch\n');
+    const session = new Session({ model: new MockLanguageModelV4({ doStream: async () => stream([]) }), askApproval: async () => "deny" });
+    const parts: PromptParts = { cwd: process.cwd() };
+    const rendered = systemPrompt(parts);
+    expect(rendered).not.toContain('Project workflow');
+    // The policy is supplied by the session, not the parts default.
+    expect(parts.workflowPolicy).toBeUndefined();
+  }));
+
+test('workflow policy renders via the session when tracking files exist', () =>
+  inGitRepo(async () => {
+    await Bun.write(join(process.cwd(), 'TODO.md'), '# Todo\n- [ ] thing\n');
+    const session = new Session({ model: new MockLanguageModelV4({ doStream: async () => stream([]) }), askApproval: async () => "deny" });
+    const status = session.workflowStatus();
+    expect(status.hasTodo).toBe(true);
+    expect(status.todoLines).toBe(3);
+    expect(status.enabled).toBe(true);
+    // A session whose repo has TODO.md renders the policy into its system prompt.
+  }));
+
+test('workflow disabled renders no policy even with a TODO.md', () =>
+  inGitRepo(async () => {
+    await Bun.write(join(process.cwd(), 'TODO.md'), '# Todo\n- [ ] thing\n');
+    const session = new Session({
+      model: new MockLanguageModelV4({ doStream: async () => stream([]) }),
+      workflow: { enabled: false },
+      askApproval: async () => 'deny',
+    });
+    // systemFor is private; workflowStatus reflects the switch.
+    expect(session.workflowStatus().enabled).toBe(false);
+  }));
+
+test('bare repo renders no workflow policy', () =>
+  inGitRepo(async () => {
+    const session = new Session({ model: new MockLanguageModelV4({ doStream: async () => stream([]) }), askApproval: async () => "deny" });
+    const status = session.workflowStatus();
+    expect(status.hasTodo).toBe(false);
+    expect(status.hasRoadmap).toBe(false);
+    expect(status.hasDocs).toBe(false);
+    // No tracking files -> policy would be empty; workflowStatus still reports the switch.
+    expect(status.enabled).toBe(true);
+  }));
+
+test('instructions load TODO.md and ROADMAP.md from the git root', () =>
+  inGitRepo(async () => {
+    await Bun.write(join(process.cwd(), 'TODO.md'), '# Todo\n- [ ] thing\n');
+    await Bun.write(join(process.cwd(), 'ROADMAP.md'), '# Roadmap\n- done\n');
+    const loaded = await loadInstructions();
+    const labels = loaded.map((i) => i.path);
+    expect(labels.some((p) => p.endsWith('TODO.md'))).toBe(true);
+    expect(labels.some((p) => p.endsWith('ROADMAP.md'))).toBe(true);
+    const fmt = formatInstructions(loaded);
+    expect(fmt).toContain('Project tracker');
+  }));
+
+test('a turn that edits without updating the task list gets one workflow nudge', () =>
+  inGitRepo(async () => {
+    await Bun.write(join(process.cwd(), 'TODO.md'), '# Todo\n- [ ] thing\n');
+    await Bun.write(join(process.cwd(), 'app.ts'), 'const a = 1;\n');
+
+    let call = 0;
+    const session = new Session({
+      model: new MockLanguageModelV4({
+        doStream: async () =>
+          stream(
+            call++ === 0
+              ? toolCall('c1', 'edit_file', { path: 'app.ts', oldString: 'const a = 1;', newString: 'const a = 2;' })
+              : text('done'),
+          ),
+      }),
+      askApproval: async () => 'once',
+    });
+
+    const notices: string[] = [];
+    for await (const ev of session.send('bump a')) {
+      if (ev.type === 'notice') notices.push(ev.text);
+    }
+
+    expect(notices.some((n) => n.includes('without updating the project task list'))).toBe(true);
+  }));
+
+test('a turn that updates the task list gets no workflow nudge', () =>
+  inGitRepo(async () => {
+    await Bun.write(join(process.cwd(), 'TODO.md'), '# Todo\n- [ ] thing\n');
+    await Bun.write(join(process.cwd(), 'app.ts'), 'const a = 1;\n');
+
+    let call = 0;
+    const session = new Session({
+      model: new MockLanguageModelV4({
+        doStream: async () =>
+          stream(
+            call++ === 0
+              ? toolCall('c1', 'todo_write', { items: [{ text: 'thing', done: true }] })
+              : call++ === 1
+                ? toolCall('c2', 'edit_file', { path: 'app.ts', oldString: 'const a = 1;', newString: 'const a = 2;' })
+                : text('done'),
+          ),
+      }),
+      askApproval: async () => 'once',
+    });
+
+    const notices: string[] = [];
+    for await (const ev of session.send('bump a')) {
+      if (ev.type === 'notice') notices.push(ev.text);
+    }
+
+    expect(notices.some((n) => n.includes('without updating the project task list'))).toBe(false);
+  }));
+
+test('/workflow parses to the workflow action', () => {
+  expect(parseCommand('/workflow')).toEqual({ type: 'workflow' });
+  const menuEntry = parseCommand('/');
+  expect(menuEntry).not.toEqual({ type: 'workflow' });
+});
+
+test('/workflow panel renders the status rows', () =>
+  inGitRepo(async () => {
+    await Bun.write(join(process.cwd(), 'TODO.md'), '# Todo\n- [ ] thing\n');
+    const session = new Session({ model: new MockLanguageModelV4({ doStream: async () => stream([]) }), askApproval: async () => "deny" });
+    const { workflowPanel } = await import('../src/ui/panel-bodies');
+    const panel = workflowPanel(session);
+    expect(panel.title).toBe('workflow');
+    expect(panel.body).toContain('TODO.md: yes');
+    expect(panel.body).toContain('workflow: on');
+  }));