bvandeusenandClaude Opus 5
44e0b0541f
CI & Build / Plugin hooks (push) Successful in 9s
CI & Build / Python lint (push) Successful in 3s
CI & Build / TypeScript typecheck (push) Successful in 53s
CI & Build / integration (push) Successful in 1m1s
CI & Build / Python tests (push) Successful in 1m35s
CI & Build / Build & push image (push) Successful in 38s
The third rule arm, and the one the other two cannot reach. `write_path_rule` is keyed on code, `pre_tool_rule` on a command — both things the session is about to DO. A rule that governs what to SAY has no such trigger: extract intent from loose phrasing, raise a conflict before acting, hand off an action with its reason, end a finding with an offer all bind on a RESPONSE, and no tool call precedes one. The operator's message is the only query that exists before a response is composed. That hook searched notes alone, so no rule had ever been retrieved against a thing the operator actually said — and residency was the only surface those rules had, which is what milestone 394 removes. A SEPARATE FUNCTION, not a branch in build_autoinject_hint, because of its early returns. That arm bails when auto-inject is disabled, when the query is blank, when nothing clears the note bar — every one a statement about NOTES. Folded in, an operator who turned the awareness menu off would silently lose their rules, a coupling with no symptom since both look like a quiet hook. Two functions, two sets of gates, composed in the route. Guarded as "the rule arm never asks the notes arm's config", which is the structural fact. Joins _ARMS rather than getting its own test file. #3497's history is that the pre-tool arm inherited a defect from its sibling by being MODELLED on it instead of sharing with it, and a third arm modelled on two is two chances to repeat that. Repeat rendering, fresh-only counting, log-before-bailout, the kind register and the two-recorders identity are properties of every arm or of none. The bar is INHERITED and says so. 0.72 was tuned against code and commands; prose is a different query shape against the same documents, and triggers are written in the vocabulary of the moment — which for most rules is act vocabulary. Starting at the only number with evidence behind it and logging every call from the first deploy is what makes it settleable; guessing lower would put an unmeasured bar in front of a corpus that binds. k=3, anchored on this hook's own budget rather than the act arms'. RULEHINT_LIMIT is 1 because that arm fires before every Bash call; this one fires once per turn, beside a notes menu already spending three slots. And a prompt genuinely contains more than one act — "merge to main and then start on X" is two — where a command is one thing. `prompt_rule` added to RANKED_SOURCES: a ranker picked it, and a ranked source missing from that tuple is silently counted as bulk delivery and drops out of the pull-through denominator. The hook reads and writes the SHARED rule ledger under scribe-priorart, not a private one — one session keeps one list, aged (#3751), so a rule named here is not re-announced before the next Bash call. Co-Authored-By: Claude Opus 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_011cPyzNnegXHr5iRMzzy5KJ
132 lines
6.3 KiB
Bash
Executable File
132 lines
6.3 KiB
Bash
Executable File
#!/usr/bin/env bash
|
|
# Scribe plugin — UserPromptSubmit push channel (knowledge auto-inject, Path A).
|
|
#
|
|
# On each user prompt, asks the operator's Scribe instance for a TITLE-FIRST
|
|
# awareness hint: the few notes that clear the per-user auto-inject gates
|
|
# (high-confidence threshold, margin gate, session dedup, top-k). Titles + ids
|
|
# only — never bodies; the agent calls get_note(id) to pull anything it judges
|
|
# relevant. Most turns inject nothing.
|
|
#
|
|
# TWO ARMS SINCE #3852, on one request. Rules and preferences are retrieved
|
|
# against the same prompt and returned in the same payload, ahead of the notes
|
|
# menu. That arm exists because the two act arms are keyed on a file write or
|
|
# a command, so a rule governing what to SAY — extract intent from loose
|
|
# phrasing, raise a conflict before acting, end a finding with an offer — had
|
|
# no moment to fire at. The operator's message is the only query that exists
|
|
# before a response is composed.
|
|
#
|
|
# The two arms are gated separately server-side: turning the notes menu off
|
|
# leaves rules arriving, because they are different claims with different
|
|
# costs of being missed.
|
|
#
|
|
# Best-effort enrichment ONLY: unlike the SessionStart channel there is no
|
|
# static floor here. If the instance is unconfigured/unreachable, or anything
|
|
# fails, the hook stays SILENT and exits 0 — it must never block a prompt.
|
|
#
|
|
# Config (same as scribe_session_context.sh), exported to the hook by Claude Code
|
|
# with the userConfig key UPPERCASED (see #2198 — reading the lowercase spelling
|
|
# silently disables this hook, and silence is indistinguishable from "nothing
|
|
# cleared the threshold"):
|
|
# CLAUDE_PLUGIN_OPTION_API_ENDPOINT base URL, no trailing slash
|
|
# CLAUDE_PLUGIN_OPTION_API_TOKEN fmcp_ API key (sensitive)
|
|
# SCRIBE_URL / SCRIBE_TOKEN override for the settings.json dogfooding path.
|
|
#
|
|
# Session dedup: each surfaced note id is remembered in a per-session file so a
|
|
# note is injected at most once per session. Passed back as exclude_ids.
|
|
set -uo pipefail
|
|
|
|
# shellcheck source=plugin/hooks/scribe_defs.sh
|
|
. "$(dirname "${BASH_SOURCE[0]}")/scribe_defs.sh"
|
|
|
|
command -v jq >/dev/null 2>&1 || exit 0
|
|
command -v curl >/dev/null 2>&1 || exit 0
|
|
|
|
# UserPromptSubmit delivers a JSON event on stdin: { prompt, session_id, cwd, ... }
|
|
event=$(cat 2>/dev/null || true)
|
|
prompt=$(printf '%s' "$event" | jq -r '.prompt // empty' 2>/dev/null) || prompt=""
|
|
session_id=$(printf '%s' "$event" | jq -r '.session_id // empty' 2>/dev/null) || session_id=""
|
|
event_cwd=$(printf '%s' "$event" | jq -r '.cwd // empty' 2>/dev/null) || event_cwd=""
|
|
|
|
# Nothing to retrieve against.
|
|
[ -n "$prompt" ] || exit 0
|
|
|
|
# Unconfigured install → silent (auto-inject is pure enrichment).
|
|
scribe_config || exit 0
|
|
|
|
# Cap the query length — a giant prompt makes a giant URL for no extra signal.
|
|
# `head -c`, not `cut -c1-2000`: cut is line-oriented and caps EACH LINE, so a
|
|
# long multi-line prompt sailed past the budget entirely. Same defect as the
|
|
# prior-art hook's code cap; this copy was missed when that one was fixed, and
|
|
# scripts/check_plugin.py caught it.
|
|
q=$(printf '%s' "$prompt" | head -c 2000)
|
|
# `-sRr`, not `-rR`: jq -R reads LINE BY LINE, so a multi-line prompt encoded as
|
|
# several lines joined by raw newlines and the request died. Single-line prompts
|
|
# worked, which is why this looked healthy — the long, substantial prompts most
|
|
# worth retrieving against were exactly the ones silently dropped. -s slurps.
|
|
q_enc=$(printf '%s' "$q" | jq -sRr '@uri' 2>/dev/null) || exit 0
|
|
|
|
# Resolve the working repo's remote so the server can scope to the bound project.
|
|
repo_dir=${event_cwd:-${CLAUDE_PROJECT_DIR:-$PWD}}
|
|
repo=$(git -C "$repo_dir" remote get-url origin 2>/dev/null || true)
|
|
repo_q=""
|
|
if [ -n "$repo" ]; then
|
|
enc=$(printf '%s' "$repo" | jq -sRr '@uri' 2>/dev/null) || enc=""
|
|
[ -n "$enc" ] && repo_q="&repo=${enc}"
|
|
fi
|
|
|
|
# Per-session dedup: ids already injected this session are skipped.
|
|
state_dir="${TMPDIR:-/tmp}/scribe-autoinject"
|
|
mkdir -p "$state_dir" 2>/dev/null || true
|
|
# RULES DEDUP IN A DIFFERENT DIRECTORY, and it has to be this one. The rule
|
|
# ledger is SHARED by every arm that can name a rule — the two PreToolUse
|
|
# hooks already keep it under scribe-priorart — so that one session keeps ONE
|
|
# list and a rule named here is not re-announced before the next Bash call.
|
|
# A private copy here would make each arm's "already seen" mean something
|
|
# different, which is the state #3749/#3750 exist to keep coherent. The
|
|
# directory name is the prior-art hook's history, not a scope claim.
|
|
rule_state_dir="${TMPDIR:-/tmp}/scribe-priorart"
|
|
mkdir -p "$rule_state_dir" 2>/dev/null || true
|
|
idfile=""
|
|
rulefile=""
|
|
exclude_q=""
|
|
if [ -n "$session_id" ]; then
|
|
# session_id is an opaque token from Claude Code; keep only filename-safe chars.
|
|
safe_sid=$(printf '%s' "$session_id" | tr -c 'A-Za-z0-9._-' '_')
|
|
idfile="$state_dir/${safe_sid}.ids"
|
|
rulefile="$rule_state_dir/${safe_sid}.rules.ids"
|
|
if [ -f "$idfile" ]; then
|
|
seen=$(tr '\n' ',' < "$idfile" 2>/dev/null | sed 's/,$//')
|
|
[ -n "$seen" ] && exclude_q="&exclude_ids=${seen}"
|
|
fi
|
|
# AGED, not read flat: an exclusion that never expires means a rule surfaced
|
|
# once in a long session is silenced for the rest of it, even as the session
|
|
# stops holding what it was told. scribe_rules_live carries the reasoning.
|
|
rule_seen=$(scribe_rules_live "$rulefile")
|
|
[ -n "$rule_seen" ] && exclude_q="${exclude_q}&exclude_rule_ids=${rule_seen}"
|
|
fi
|
|
|
|
body=$(curl -fsS --max-time 5 \
|
|
-H "Authorization: Bearer ${token}" \
|
|
"${url%/}/api/plugin/retrieve?q=${q_enc}${repo_q}${exclude_q}" 2>/dev/null) || exit 0
|
|
[ -n "$body" ] || exit 0
|
|
|
|
context=$(printf '%s' "$body" | jq -r '.context // empty' 2>/dev/null) || exit 0
|
|
[ -n "$context" ] || exit 0
|
|
|
|
# Remember the surfaced ids so they aren't injected again this session.
|
|
if [ -n "$idfile" ]; then
|
|
printf '%s' "$body" | jq -r '.note_ids[]? // empty' 2>/dev/null >> "$idfile" || true
|
|
fi
|
|
# Rules onto the SHARED ledger, stamped so they can age out. Only FRESH ids
|
|
# come back in rule_ids (#3752) — a rule rendered as a repeat is already on
|
|
# the ledger, and re-appending it would keep pushing its stamp forward so it
|
|
# never aged at all.
|
|
if [ -n "$rulefile" ]; then
|
|
printf '%s' "$body" | jq -r '.rule_ids[]? // empty' 2>/dev/null \
|
|
| scribe_rules_append "$rulefile"
|
|
fi
|
|
|
|
jq -n --arg c "$context" \
|
|
'{hookSpecificOutput: {hookEventName: "UserPromptSubmit", additionalContext: $c}}'
|
|
exit 0
|