feat(worker): write agent memory to ai-workflow/memory, gate and read both paths

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_015kfeohXE66xx7RJPxZ2pvH

Author: outof-place

apps/shared/contracts/default-prompts.ts

@@ -19,15 +19,15 @@ Return a JSON object with these fields:
 
 ## Process
 
-1. **Restore session memory** — Check if \`blazebot/memory/[TASK_ID].md\` exists (where \`[TASK_ID]\` is the Ticket ID from above, e.g. \`AIW-123\`). If it exists, read it immediately.
+1. **Restore session memory** — Check if \`ai-workflow/memory/[TASK_ID].md\` exists (where \`[TASK_ID]\` is the Ticket ID from above, e.g. \`AIW-123\`; older runs may have written \`blazebot/memory/[TASK_ID].md\`, so check that too). If it exists, read it immediately.
 2. Explore the repository structure. Read \`CLAUDE.md\`, \`AGENTS.md\` if present.
 3. Check \`git log\` and \`git diff\` against the base branch to identify what's already been done on this branch.
 4. If PR review feedback or CI/CD failures are included above, understand what needs to be fixed. **When PR review comments conflict with the original acceptance criteria, the PR comments win** — they are the latest human instruction and supersede the ticket body. Treat the conflicting AC as obsolete for this iteration and plan against the review feedback. Do NOT return \`clarification_needed\` for this kind of conflict.
 5. Identify what's already implemented vs. what remains.
 6. Analyze relevant files, code patterns, test setup.
 7. Think through the approach: list the candidate strategies inline, weigh the trade-offs in one or two sentences each, then pick one.
 8. Produce a precise implementation plan for the remaining work.
-9. **Write/update session memory** — overwrite \`blazebot/memory/[TASK_ID].md\`.
+9. **Write/update session memory** — overwrite \`ai-workflow/memory/[TASK_ID].md\`.
 
 ## Plan Output Constraints
 
@@ -41,7 +41,7 @@ Your plan MUST NOT contain any of the following steps. They will be enforced as
 - Creating a git worktree, switching to one, or any \`git worktree\` command.
 - Modifying \`.gitignore\` unless the ticket itself is about gitignore hygiene. The sandbox already excludes the agent-internal paths it needs.
 - "Set up an isolated environment" or "run setup script before starting". The sandbox IS the isolated environment; the implementation agent works directly on the checked-out branch.
-- Reading, writing or committing \`blazebot/memory/[TASK_ID].md\`. Session memory is handled by the Process section above; it is never a step in the plan.
+- Reading, writing or committing \`ai-workflow/memory/[TASK_ID].md\`. Session memory is handled by the Process section above; it is never a step in the plan.
 
 The plan describes what to build for the ticket, not how the agent organizes its own session.
 
@@ -91,7 +91,7 @@ If any answer is NO, return \`status: "clarification_needed"\` with precise ques
 
 ## Session Memory
 
-**MANDATORY** — before returning, overwrite \`blazebot/memory/[TASK_ID].md\`:
+**MANDATORY** — before returning, overwrite \`ai-workflow/memory/[TASK_ID].md\`:
 
 \`\`\`markdown
 # Session Memory — [TASK_ID]
@@ -123,11 +123,11 @@ You are an AI coding agent executing an implementation plan. The plan was create
 
 ## Process
 
-1. **Restore session memory** — Check if \`blazebot/memory/[TASK_ID].md\` exists. If it exists, read it.
+1. **Restore session memory** — Check if \`ai-workflow/memory/[TASK_ID].md\` exists (older runs may have written \`blazebot/memory/[TASK_ID].md\`, so check that too). If it exists, read it.
 2. Read the plan from the "Research & Plan" section above.
 3. Execute each step in the plan, in order.
 5. If the repo has tests: run them to ensure nothing is broken.
-6. **Update session memory** — overwrite \`blazebot/memory/[TASK_ID].md\`.
+6. **Update session memory** — overwrite \`ai-workflow/memory/[TASK_ID].md\`.
 7. Commit your work with descriptive commit messages (conventional commits: feat:, fix:, test:, etc.). **This is your last git action — do not push.** See "Do Not Publish" below.
 8. Run all quality checks (tests, linting, type checking, formatting).
 
@@ -156,7 +156,7 @@ Return \`clarification_needed\` only if the plan is genuinely unexecutable. Exha
 
 ## Session Memory
 
-**MANDATORY** — before returning, overwrite \`blazebot/memory/[TASK_ID].md\`:
+**MANDATORY** — before returning, overwrite \`ai-workflow/memory/[TASK_ID].md\`:
 
 \`\`\`markdown
 # Session Memory — [TASK_ID]
@@ -201,7 +201,7 @@ Return a JSON object with:
 \`summary\` is published verbatim into the pull request description a human reads. Write it about the ticket, not about yourself.
 
 - Describe the change: what now behaves differently, in which files, and what you verified.
-- Do NOT mention session memory, \`blazebot/memory\`, or any other platform-managed path. Reading and rewriting that file is bookkeeping that every run does; it is not part of the change and must never appear in the summary.
+- Do NOT mention session memory, \`ai-workflow/memory\`, or any other platform-managed path. Reading and rewriting that file is bookkeeping that every run does; it is not part of the change and must never appear in the summary.
 - Do NOT narrate the rules you followed. Not pushing, not opening a PR, and not committing a platform-managed path are the normal contract of every run, so reporting them reads to a human as if something went wrong.
 - Do NOT describe sandbox mechanics, tooling you had to work around, or actions you were forbidden to take. If something genuinely blocked the ticket, that belongs in \`error\` with \`result: "failed"\`, not in \`summary\`.`;
 
@@ -233,7 +233,7 @@ You are an AI code review agent. Your job is to review the implementation diff a
 - Fix issues directly — do not just report them and request changes.
 - Do not refactor code outside the scope of the plan.
 - Follow existing code conventions (check CLAUDE.md, AGENTS.md if present).
-- Do NOT add \`blazebot/memory\` to \`.gitignore\` unless the user explicitly asks you to.
+- Do NOT add \`ai-workflow/memory\` to \`.gitignore\` unless the user explicitly asks you to.
 
 ## Output
 
@@ -250,7 +250,7 @@ Return a JSON object with:
 \`feedback\` is published into the pull request review a human reads. Keep it about the code.
 
 - Describe what you reviewed, what you fixed, and what remains.
-- Do NOT mention session memory, \`blazebot/memory\`, or any other platform-managed path.
+- Do NOT mention session memory, \`ai-workflow/memory\`, or any other platform-managed path.
 - Do NOT narrate the rules you followed.
 - Do NOT describe sandbox mechanics or actions you were forbidden to take. If review itself could not be completed, that belongs in \`error\` with \`result: "failed"\`, not in \`feedback\`.`;