| name | backtick |
| description | Local task compiler: prepare_task before exploring, work from the evidence packet, verify_patch before done. |
backtick
You drive an evidence engine, not a search party. backtick compiles the
request plus repository evidence into a minimal task packet so you read less
and verify more. Cloud tokens per verified task is the number this workflow
exists to cut — and a short packet that produces an unverified patch has cut
nothing.
Persistence
ACTIVE EVERY RESPONSE while coding in a repository with the backtick tools
attached. Still active if unsure. Off only: "stop backtick" / "normal mode".
Default: full. Switch: /backtick lite|full|ultra.
The loop
Before broad repository exploration, in order:
- Call
prepare_task with the user's request verbatim and your level's
token budget.
- Work from the returned evidence packet. Its FACTs — symbols, exact source
spans, structural facts, tests, diagnostics — are mechanically extracted;
trust them. Its INTERPRETED GOAL is a hypothesis; confirm it against the
code before building on it.
- Missing evidence? Call
expand_context with a pointed question and a
small additional budget. Do not fall back to repo-wide scanning.
- Need one symbol's neighborhood?
inspect_symbol returns its definition,
callers, callees, types, and related tests.
- After editing, call
verify_patch. Revise from the compact failure
packet it returns — failing test names and implicated symbols, not raw
logs.
- Do not claim completion until verification succeeds.
Rules
- The packet's OMITTED CONTEXT section names what was cut. If the task needs
it, ask for it by name via
expand_context; never re-read the whole repo
to check.
- Prefer the packet's exact source spans over re-opening files it already
quoted.
- Keep edits inside the evidence neighborhood. Wandering past it usually
means the packet is missing something — expand, then edit.
- If the same test fails twice,
inspect_symbol the implicated symbol
before a third attempt.
- If the evidence contradicts the interpreted goal, say so and expand;
never force a patch to match a wrong goal.
Levels
Levels set the prepare_task token budget.
| Level | What changes |
|---|
| lite | 1000-token packets. Small fixes in code you already know. |
| full | 2000-token packets. Default for most tasks. |
| ultra | 4000-token packets. Cross-file work or unfamiliar repos. |
Example: "Fix users being logged out after token expiry"
- lite: "prepare_task, budget 1000 — symbols and the one hot span."
- full: "prepare_task, budget 2000 — spans plus structure, tests, diagnostics."
- ultra: "prepare_task, budget 4000 — the full neighborhood: callers, config, related tests."
When NOT to use backtick
Skip the loop for questions that need no repository code. If the backtick
tools are not attached or prepare_task errors, say so and proceed with
normal exploration — never fake a packet. A user request for a full-repo
audit or survey overrides the packet discipline; give it in full.
Never let a small packet shrink correctness. Compression limits what you
read, never what you check: a patch that passes on a misread goal is still
the wrong patch.
Boundaries
backtick governs how you gather context and how you verify, not what you
build. "stop backtick" / "normal mode": revert. Level persists until changed
or session end.
The smallest packet that produces a verified patch is the right packet.