Keep next-page command payloads free of advisory prose and emit page-size or budget advice on a separate hint line. Normalize literal percent characters and valid path escapes consistently. Document and test quoting the full negative-filter query without relaxing unknown-option rejection. Validated by paging continuation, percent-path, quoted argument, and strict-option regressions in the 593 default and 617 CUDA-enabled test suites. Co-authored-by: Codex <[email protected]>
11 KiB
CLI usage
Install the Linux binary with cargo install --path . --locked. Put Cargo's
binary directory ($CARGO_HOME/bin, default $HOME/.cargo/bin) on PATH;
command -v trufflepig verifies discovery.
--root defaults to the current directory. Without a workspace, root discovery
does not walk up to Git metadata; use the same root and cache for follow-up reads; configured workspaces route reads to their recorded member.
trufflepig --help -b 2000
trufflepig --root . --no-daemon index
trufflepig --root . search 'sym:TokenBucket'
trufflepig --root . search 'sym:User lang:php file:src/'
trufflepig --root . search 're:quarantine file:script/host -file:tests'
trufflepig --root . search 're:fn .*helper lang:rust'
trufflepig --root . --no-daemon search 'file:src/ kind:function'
trufflepig --root . semantic prepare --wait
trufflepig --root . semantic status
trufflepig semantic worker status
Use an explicit search verb for all queries; unknown commands exit 2 with an
error naming the likely command (next and page point to more CURSOR). No
command means status; help prints the --help text. sym: selects exact,
case-sensitive symbol occurrences, declarations before modules, members, locals,
and imports; re: scans live bytes and returns one occurrence per matching line,
with ^ and $ matching at line boundaries as in grep. file: (repeatable),
-file:, lang:, and kind: filter as the
retrieval contract defines; an empty filtered
search explains empty filters on coverage. Quote the whole query containing
-file:, such as trufflepig search 'needle -file:tests'; a standalone -file:
is parsed as an option. Quote c# in shell commands.
One JSON object plus newline is the default output; --json accepts the same
format. --format lines renders search, refs, map, more, and show as
tab-separated lines for agents that read output directly: one
HANDLE<TAB>[MEMBER/]PATH:START-END[<TAB>NAME] line per hit, each followed by
an indented LINE: TEXT snippet, then coverage:. Continuing pages print
next: more CURSOR; the first adds a hint: line for -n N or -b N.
truncated: true appears when present. show prints a PATH [(MEMBER)] lines FIRST-LAST header,
LINE<TAB>text rows, next: show CURSOR when the budget cut the read, and a
verified: footer (verified: current file for explicit path reads, which JSON
marks "source": "current_file") that flags encoding: byte-escaped once.
Errors and every other verb stay JSON. show defaults to a 1500-token budget,
never below a workspace's [output].budget; -b/--budget otherwise defaults to
600 o200k_base tokens for the serialized stdout response, measured on the
rendered text of the selected format, or to the workspace's [output].budget
when one is set. Search ranks files first and emits one compact representative
per file before the -n/--limit page cap (20 files by default); the budget may
fit fewer. Each hit includes a file URI, line span, and immutable handle, and
a snippet (line, text of at most 120 characters): the line in the span
that mentions the most distinct query terms, else the hit's name, the earliest
on ties; a code line beats a comment line with one more term.
A page keeps snippets while at least eight hits (or all remaining hits) fit,
otherwise it drops them for more locators. Snippets locate evidence; show
remains the verified read. Increase the budget when a hit or source line cannot fit.
Budget failures include a retry hint on stderr even when no JSON error fits.
--help, --version, and option parsing errors print plain, unbudgeted text
(parsing errors also on stderr); help ends with a query, navigation, and page summary.
Workspace search
trufflepig --workspace evaluation/workspaces/space.toml ws show
trufflepig --workspace evaluation/workspaces/space.toml search 'airlock'
trufflepig search 'crayon in:tgstation'
trufflepig search 'airlock ws:home'
trufflepig --member tales-from-space hist path:README.md
trufflepig ws status
trufflepig ws discover ~/Source/lunatic ~/Source/tales-from-space ~/Source/tgstation
An explicit --workspace FILE selects named local checkouts. Otherwise, discovery
checks the nearest ancestor trufflepig.workspace.toml, then the global registry.
--no-workspace retains singleton operation. An explicit subtree --root never
silently expands to a configured member root. Unselected workspace queries search
the home member and widen to all members when home has no hits; see the
workspace contract.
in:NAME or --member NAME selects a member; ws:home selects the invocation
checkout, which may be a linked Git worktree standing in for its member, and
ws:all searches the complete workspace. File-first ranking fuses
member lanes with provenance; compact hits carry a percent-encoded file URI,
lines, and owner-qualified handle. Pagination freezes captured member snapshots;
missing members are reported as incomplete coverage. Use show, more, and ctx
from any member of the same workspace; persisted handles retain their original
owner. Explicit paths, history revisions, sessions, and other owner commands use
home or --member NAME. ws discover returns only a proposed configuration and
does not scan unrelated directories. See the workspace contract
for schema, cache behavior, routing guarantees, and current limits.
Follow-up reads and navigation
Copy hits[].handle (the first column in lines format) into show or ctx.
The next: label introduces more CURSOR or show CURSOR; invoke as
trufflepig more CURSOR or trufflepig show CURSOR. Widening advice is separate. JSON
next is the bare cursor. Search pages use compact
file-first entries, whose name appears only for symbol hits; show supplies
source lines for a selected handle.
trufflepig show SET:ORDINAL
trufflepig more SET@OFFSET
trufflepig ctx SET:ORDINAL
trufflepig refs refill_tokens
trufflepig map src/
trufflepig show path:src/main.rs:1-20
trufflepig show 'sym:refill_tokens file:src/'
show 'sym:NAME' (optional file:, lang:, kind:, quoted or not) reads the
best-ranked definition as a verified handle read. refs lists occurrence sites and
map outlines a file or prefix; ranking, filters, and JSON fields are in the
navigation contract.
A handle is a 32-character result-set ID plus a one-based ordinal. A pagination
cursor uses the same set ID and a zero-based next offset. Handles survive restart
within their retention limits; they never name the latest unrelated query. In a
workspace, the result set freezes each member owner and generation, so later ranking
or configuration changes cannot retarget a handle. show continuations retain
source identity and remaining bytes; pass their next value unchanged to show.
A handle whose file changed answers stale_source: run a fresh search or an
explicitly current path read. show 'sym:NAME' on a changed file instead returns
the definition from the current bytes, marked verified: false and
freshness: unpublished, once a serving daemon has had 3 s to republish.
stale_result means the graph generation changed and ctx needs a fresh handle.
Paths in responses percent-encode raw filename bytes. Paste the encoded path
unchanged into show, including %20 for a space, %25 for %, and %3A for
:, so filename punctuation cannot become a line-range separator. Traversal,
absolute paths, and symlinks are rejected.
Historical navigation
trufflepig hist-index --wait
trufflepig hist-status
trufflepig hist path:src/main.rs
trufflepig hist sym:TokenBucket
trufflepig since HEAD~3 path:src/main.rs
trufflepig since HEAD --uncommitted
trufflepig diff HEAD --target path:src/main.rs
trufflepig show SET:ORDINAL --side before
trufflepig blame path:src/main.rs
trufflepig blame path:src/main.rs --raw
trufflepig blame path:src/main.rs --ignore-revs-file .git-blame-ignore-revs
History requires Git 2.55 or newer and local objects. Ordinary search remains
available when history is unavailable. hist uses captured first-parent history;
since compares the exact resolved revision directly with captured HEAD, or with
one published working-tree generation under --uncommitted. diff compares a
commit with its first parent; only explicit targets permit budgeted source hunks.
Historical changes require show HANDLE --side before|after for exact source.
path: selectors are root-relative. sym: uses exact live symbol occurrences;
ambiguous matches return selectable handles. The history contract
defines correspondence limits, immutable reads, resource exclusions, and coverage.
Diagnostics and sessions
trufflepig session start
trufflepig --session SESSION_ID search 'refill tokens'
trufflepig --session SESSION_ID show SET:ORDINAL
trufflepig session end SESSION_ID
trufflepig audit SESSION_ID
trufflepig forget-logs
Pass --session ID on each assigned request, or set TRUFFLEPIG_SESSION in the
client environment. A session is not inferred from a process or daemon. Optional
--client NAME labels the caller. Diagnostics default to bounded metadata;
--diagnostics detailed additionally records bounded raw search queries, and
--diagnostics off disables request recording. audit reports retained summaries.
See diagnostics for delivery evidence, retention,
overlap labels, and incomplete-window limits.
Cache and daemon
The runtime contract defines cache ownership, daemon lifecycle, systemd upgrades, and sandbox spool transport.
Optional semantic retrieval
Default builds provide lexical/structural retrieval. Install with
cargo install --path . --locked --features semantic for CPU inference or
--features semantic-cuda for CUDA support. Configure the pinned model, ONNX Runtime
library, provider, GPU UUID, and ORT arena in $HOME/.config/trufflepig/inference.toml;
model and runtime paths may also use TRUFFLEPIG_MODEL_DIR and ORT_DYLIB_PATH. Run
semantic-check MODEL_DIRECTORY to verify assets and provider execution before
preparing a root. semantic prepare schedules missing source-region vectors for
background preparation; semantic prepare --wait waits for the requested generation.
semantic status reports preparation coverage. semantic worker status and
semantic worker stop inspect or stop the shared per-user worker. Missing assets,
provider failures, pending vectors, and query timeouts remain explicit statuses
while lexical results stay available. --no-sem disables semantic retrieval,
including a workspace's persistent opt-in; --rerank/--no-rerank also toggles
rerank. Source regions are never embedded synchronously during search; search uses
the cache and has a 500 ms semantic query deadline before returning lexical fallback.
The semantic contract defines worker, cache, snapshot, and
status behavior; evaluation usage defines the reproducible
fixture and development-task measurements.