| name | self-patch |
| description | How I notice a recurring failure of my own, root-cause it, fix what I can myself (behavioral rules, config, ignore files), and — for anything needing a kernel change I can't apply — draft a verified proposal for the operator to review. The capstone that turns my mistakes into durable fixes. Use when I hit a repeated mistake, get corrected by the operator, see a failed VERIFY, or notice a pattern worth fixing. |
self-patch — fixing myself
When I fail the same way twice, I fix it — I don't just apologize. This is the loop. It's
REACTIVE (invoked by the operator, or by me noticing mid-session) and TIERED: I fix the
non-kernel layers myself and verify; kernel changes I can only PROPOSE (my src/ is locked —
that's the safety boundary, not a limitation to route around).
The loop: notice → root-cause → classify → act → record
1. Notice
A repeated mistake, operator feedback ("you did it again"), a failed VERIFY, or a pattern in
memory/work-journal.md / bounded logs.
2. Root-cause — read, don't guess
Read the actual code/logs. Confirm the mechanism before fixing. Use gitignore-aware search
(bare rg works via .ignore; ls/find for existence). Every good fix I've made came
from reading source, not assuming.
3. Classify the layer
- behavioral — a habit/instruction gap (I stall, I gaslight, I over-write). Fix = a rule
in the PINNED hot set.
- config — an env var, rc file, or ignore file (rg blindness →
.ignore; nudge count →
LAZAR_STALL_NUDGES). Fix = change the file/var.
- kernel — needs Rust (max_tokens=8192, reset wipes memory/). I CANNOT edit the locked
src/. Fix = draft a proposal (step 4c).
4. Act
4a. Behavioral fix (autonomous). Add ONE tight dated line to the PINNED hot set (see
below). Then PROVE it: end my turn with
[VERIFY]\ncontains=memory/PINNED.md:<distinctive phrase from the rule>\n[/VERIFY].
Never claim a behavioral fix without a passing VERIFY.
4b. Config fix (autonomous). Change the file/var, then VERIFY the effect
(contains=, or a command whose output confirms it). Record it.
4c. Kernel fix (PROPOSAL ONLY — never apply).
staged=$(bash skills/_meta/self-patch/stage-proposal.sh <name>) — stages a writable
src/ copy under workspace/proposals//src/.- Delegate the mechanical edit+build to a worker (isolates it from my context) by writing a
PRECISE spawn prompt — exact file, exact old→new text, and an
EXPECT-VERIFY: block
requiring a build-ok marker — then:
bash "$LAZAR_HOME/bin/lazar-spawn" <name> <prompt-file>.
The worker edits $staged, runs
cd $staged && CARGO_TARGET_DIR=/tmp/sp-$$ cargo build --release and, on success, writes
$staged/../build-ok + a short DESIGN.md. (Or, for a tiny edit, I do this directly.)
- Gate on the envelope (
matched:true = it compiled). NEVER run kernel-apply.sh myself.
- Surface to the operator, verbatim:
"Kernel proposal
<name> is staged and verified-building. Review the diff and run
bash scripts/kernel-apply.sh <name> to apply."
5. Record
bash skills/_meta/self-patch/ledger.sh <layer> <symptom> <cause> <fix> <verified> →
appends to memory/distilled/self-patch.md. The lesson persists.
PINNED hot-set discipline (keep the every-session injection lean)
- Behavioral rules live under
## Behavioral rules (hot) in PINNED, ≤6 lines, each dated.
- Promote a rule when a failure recurs; retire the coldest to
memory/distilled/behavioral-history.md when the hot set overflows. Full history is never
lost — only the always-injected copy is bounded.
Boundaries (never cross)
- Never edit
src/ directly (locked). Never run kernel-apply.sh myself — kernel changes are
proposal-only, operator-applied.
- Never touch
workspace/tui (owner-managed).
- Never pattern-delete under
memory/distilled/.
- Never call a fix done without a passing VERIFY.
Self-test
bash skills/_meta/self-patch/selftest.sh
