| name | context-loader |
| description | Auto-loads the bencium-harness project memory at the start of every session. Triggers when the cwd contains a .harness/ folder. Reads .harness/memory.md, .harness/rules.md, and the last 3 entries in .harness/archive/ to restore project context across /clear and session boundaries. |
context-loader
When to invoke
At the start of any session (or when explicitly asked to refresh context) IF the current working directory contains a .harness/ folder.
You should also invoke this skill whenever:
- The user runs
/clear and you need to re-establish project context.
- The user asks "what was the context?" or "where were we?" and a
.harness/ exists.
- You're about to make an architectural recommendation and want to check existing decisions first.
What to do
-
Verify .harness/ exists in the project root. If not, do nothing — this is not a bencium-harness project.
-
Read these files (in this order):
CLAUDE.md (project root) — pointer file. Skim only; the detailed files it points to are the authoritative source. Skip if absent.
.harness/memory.md — hot context. Read in full.
.harness/conventions.md — code-quality rules the agent must follow on every task. Read in full. If missing, note "no conventions.md — running without quality guardrails" internally.
.harness/rules.md — non-negotiables. Read in full.
.harness/constraints.md — hard limits. Read if non-empty.
- The most recent 3 entries in
.harness/archive/ — find by ls .harness/archive/ | sort -r | head -3. Read each in full.
tasks.md (project root) — read the ## Now section only.
ACCEPTANCE.md (project root) — read all unchecked items.
-
Internally summarize what you learned. Do NOT print the summary to the user unless they ask — this is context for YOU. If .harness/memory.md contains a ## Session handoff block, treat it as the most recent state — that's where the last session said it left off (active task, phase, up next, blockers, uncommitted files). Start there.
-
If you find anything in memory or recent archive that directly contradicts what the user just asked, surface it: "Heads up: a decision in archive/NNNN says X — does this still apply, or has it changed?"
What NOT to do
- Do not load this if
.harness/ doesn't exist. Stay silent.
- Do not load this on every prompt. Once per session is enough unless the user explicitly asks to refresh.
- Do not dump the memory file contents into the chat. Use it as context.
- Do not modify any files when loading context. This is read-only.