git diff

# git-diff.toml — reduce a full unified diff to a per-file change summary
#
# Raw: full unified diff (potentially thousands of lines)
# Filtered: one line per file with insertion/deletion counts, plus a total
#
# This filter used to `run = "git diff --stat {args}"`. The patch content was
# then destroyed before tokf ever saw it — a user asking for a diff got a line
# summary, and `tokf raw <id>` could not give the patch back (issue #430).
# The summary is now computed here, over the real diff, so the patch is
# captured in history and recoverable.
#
# Escape hatches: when the model needs the actual diff content (or an
# alternative summary format), it can pass any of the flags in
# `passthrough_args` and tokf will pass `git diff` straight through. The list
# below was chosen from observed agent retry patterns: forcing a summary with
# no escape caused models to thrash through ~50 retries trying every variation
# of `git diff` to get the patch content out.

command = "git diff"
description = "Summarise a diff as per-file insertion/deletion counts"

# Skip the filter entirely when the user explicitly asks for diff content,
# specific context-line counts, or an alternative summary format they would
# rather have than ours.
passthrough_args = [
  "-p",            # patch (full diff content)
  "--patch",       # also matches --patch-with-stat, --patch-with-raw, --patch-id
  "--stat",        # git's own stat rendering
  "-U",            # -U3, -U10, ... custom unified context line counts
  "--unified",
  "--numstat",     # alternative summary: machine-readable
  "--shortstat",   # alternative summary: one line
  "--raw",         # alternative summary: raw format
]

# Anchored to column 0 on purpose. `contains = "fatal:"` would fire on any
# patch whose *content* mentions "fatal:", collapsing a legitimate diff to a
# single error line — a real hazard now that the full patch flows through this
# filter rather than a `--stat` summary. In a unified diff every content line
# is prefixed with `+`, `-`, or a space, so only git's own errors start at
# column 0.
match_output = [
  { pattern = "(?m)^fatal: ", output = "✗ {line_containing}" },
]

# NOTE: the counting sections, the per-file [[chunk]] and the totals
# aggregates below are duplicated in git/show.toml. tokf filters have no
# include mechanism (config resolution is first-match-wins, by design),
# so this is copied on purpose — but it is logic, not boilerplate.
# Fix the sign-prefix patterns in both files or neither.
# Whole-output counts for the totals line.
#
# Stateful on purpose: they only collect inside a hunk body (between an `@@`
# header and the next file's `diff --git`), so nothing outside a real diff can
# be mistaken for changed content.
#
# The patterns exclude the `+++ b/…` / `--- a/…` file headers — and nothing
# else. An earlier, simpler `^\+([^+]|$)` also swallowed every content line
# whose own first character was `+` or `-` (markdown bullets, YAML list items,
# `--flag` in a script), silently undercounting very ordinary diffs.
#
# Read them as "starts with the sign, but is not the literal `+++ `/`--- `
# header": the alternation spells out the 1-, 2- and 3-character cases because
# Rust's regex crate has no lookahead.
[[section]]
name = "file-headers"
match = '^diff --git '
collect_as = "file_header_lines"

[[section]]
name = "added"
enter = '^@@ '
exit = '^diff --git '
match = '^\+([^+]|$|\+([^+]|$)|\+\+[^ ])'
collect_as = "added_lines"

[[section]]
name = "removed"
enter = '^@@ '
exit = '^diff --git '
match = '^-([^-]|$|-([^-]|$)|--[^ ])'
collect_as = "removed_lines"

# One chunk per file, counting its own added/removed lines.
[[chunk]]
split_on = '^diff --git '
collect_as = "files"

# Use the `b/` (destination) path so renames report where the file ended up.
[chunk.extract]
pattern = '^diff --git a/.* b/(.*)$'
as = "path"

[[chunk.aggregate]]
pattern = '^\+([^+]|$|\+([^+]|$)|\+\+[^ ])'
count_as = "added"

[[chunk.aggregate]]
pattern = '^-([^-]|$|-([^-]|$)|--[^ ])'
count_as = "removed"

[on_success]
output = """{files | each: "{path} | +{added} -{removed}" | join: "\n"}
{file_count} files changed, {insertions} insertions(+), {deletions} deletions(-)"""

[[on_success.aggregates]]
from = "file_header_lines"
pattern = '^diff --git '
count_as = "file_count"

[[on_success.aggregates]]
from = "added_lines"
pattern = '^\+'
count_as = "insertions"

[[on_success.aggregates]]
from = "removed_lines"
pattern = '^-'
count_as = "deletions"

[on_failure]
tail = 5

# Route --name-only / --name-status to a tree-structured child filter
# instead of passing through unfiltered.
[[variant]]
name = "name-list"
detect.args_pattern = '--(name-only|name-status)'
filter = "git/diff-name-list"