| name | smaqit.session-start |
| description | Start a new chat with full project context. Use when beginning a session to load README, recent history, and task planning. |
| metadata | {"version":"0.8.1"} |
Session Start
Start a new chat with full project context. Execute these steps IN ORDER:
Steps
-
Read core project files from start to finish (in parallel, if they exist):
README.md
CONTRIBUTING.md
.github/copilot-instructions.md (GitHub Copilot), CLAUDE.md (Claude Code), or AGENTS.md (Codex)
- Project documentation directories (e.g.,
docs/, documentation/) — scan for index files like README.md, index.md, architecture.md, or ADRs in adr/ subdirectories
- Build/test entrypoints (whichever exist):
Makefile, package.json, pyproject.toml, go.mod, Cargo.toml
-
Load recent session context (use both sources if available; memory takes priority for cross-branch continuity):
- From memory (primary, if available): If a persistent memory/notes capability is available in this environment, check it for stored entries with subjects
"session history" and "next steps". These are written by session-finish and are available across all branches, making them the most reliable source when working in parallel or on a new branch.
- From files (fallback/supplement, always available): Read the most recent history entry from
.smaqit/history/ for full detail. If no entries exist yet, continue without file history.
- If both sources exist, memory provides the freshest cross-branch context; the history file provides the full narrative. If no memory capability is available, the history file is the sole source.
-
Load task planning:
- Read
.smaqit/tasks/PLANNING.md (NOT individual task files).
- If a persistent memory/notes capability is available, supplement with any stored entries with subject
"task state" — these are written by task skills (task-create, task-start, task-complete) and reflect the most recent state across all branches. PLANNING.md remains sufficient on its own when memory is unavailable.
- Note: Task workflow rules (autonomous vs assisted modes) are loaded via
task-list skill when working on tasks.
-
Load project glossary (conditional — only if .smaqit/glossary.md exists):
- Read
.smaqit/glossary.md in full.
- Surface the terms in context so they are available to the agent throughout the session without requiring explicit invocation.
- If the file does not exist, skip this step silently.
-
Load project compendium (conditional — only if .smaqit/compendium.md exists):
- Read
.smaqit/compendium.md in full.
- Surface all Q&A entries in context so they are available throughout the session for reference and consistency.
- If the file does not exist, skip this step silently.
-
Read the codebase for the next unblocked task:
- Identify the next unblocked task from PLANNING.md.
- Read the source areas it would touch: relevant interfaces, abstractions, factories, pools, and existing implementations.
- This step is MANDATORY before presenting tasks. Do not skip it because the task description appears complete.
-
Synthesize and present a summary covering:
- Current project state (from READMEs)
- Recent changes and decisions (from memory and/or history)
- Open tasks sorted by priority, with a brief assessment of each task's approach against the codebase
- Suggested next steps for the user to take (e.g., which task to start, what information to provide, or what questions to ask).
Critical Requirements
CRITICAL: Read complete markdown (.md) files without line limits. Do NOT truncate at any arbitrary limit.
Note: Only read individual task files (.smaqit/tasks/NNN_*.md) when actively working on that specific task.