Files
FabledScribe/src/scribe/routes/plugin.py
T
bvandeusenandClaude Opus 5 91bc0fb01e
CI & Build / Python lint (push) Successful in 3s
CI & Build / Plugin hooks (push) Successful in 13s
CI & Build / TypeScript typecheck (push) Successful in 52s
CI & Build / integration (push) Successful in 59s
CI & Build / Python tests (push) Failing after 1m13s
CI & Build / Build & push image (push) Skipped
feat(plugin): a high-confidence rule is put in front of a command, not beside its result (#4214)
Milestone 419 step 3, the pre-act checkpoint. Every rule surface in this
plugin returns `additionalContext`, which Claude Code delivers alongside the
tool RESULT — so the rule is read after the call is written and lands as
commentary on a decision already made. That is the milestone's central
finding, measured over a session with seven misses, three caught by the
operator and none by this system.

The action arm can now return a `deny` instead. The act does not run, the
rule's text can be read before the call exists, and the remedy is one
`get_rule` call after which the act may be re-submitted unchanged. Nothing
reaches the operator: a deny is a message to the model.

"CONSEQUENTIAL" IS DERIVED, NOT ENUMERATED. The obvious implementation lists
act kinds — a write to product code, a schema change, a bulk classification, a
merge. Every one of those is consequential because THIS operator wrote rules
about it, and shipping that list is this instance's corpus hard-coded into the
product (rule 115). So the corpus decides: an act is consequential when the
install's own rules speak to it above the checkpoint bar. A fresh install with
no rules never stops anything.

FOUR CONDITIONS, EACH PREVENTING A DIFFERENT WRONG. Above the bar; a rule and
never a preference (which claims no such force); the band's top hit only (the
ranker's confidence claim attaches to its first element); and only a rule the
session has NOT opened — `held` is observable from the get_rule PostToolUse
hook (#4100), not self-report.

WHY "NOT OPENED" RATHER THAN "NO OUTCOME RECORDED". An outcome can be
satisfied with one cheap call asserting compliance without producing any, and
a checkpoint dismissible that way manufactures exactly the compliance data
step 2 was built to measure. Reading a rule cannot be faked in that direction:
after `get_rule` the statement is in context, which is the whole of what was
wanted.

THE BAR IS MEASURED. `retrieval_telemetry(days=30)`: write_path_rule p90
0.7628 max 0.8817; pre_tool_rule p90 0.7373 max 0.8293. 0.80 is above p90 on
both and below max on both, so it selects from the top decile of an already
selective arm and is still reachable. It ships as a setting with a Settings
card, because a cosine distance in one model's geometry over one corpus cannot
transfer.

TWO GUARDS ON THE WORST CASE: at most one hold per rule and five per session,
so a mis-set floor degrades to a noisy session rather than one that cannot
proceed. The ledger lives in the swept directory and is named `.ids`, so the
existing compaction-clear guards cover it.

WRITES ARE NOT HELD, AND THAT IS THE OPERATOR'S DECISION RATHER THAN MINE.
`scribe_prior_art.sh` carries a tested property that it never returns a
permissionDecision — a recall aid may not stand in the way of a write. Three
of the milestone's seven misses were file edits and none are reachable from
the command side, so there is a live argument for extending this; that
argument is exactly why the boundary is now asserted by a test rather than
left to memory. The write-path arm computes and returns the same block so the
decision can be revisited with evidence; the hook ignores it, and a change of
mind is a hook edit rather than a feature.

Verified by lifting `checkpoint_for` and `_rule_band` out of source with `ast`
and exercising the shipped functions over 17 populations, by running the
ledger and deny envelope in bash (10 cases, including that a refused hold is
not written and that a garbled rule id fails closed), and by
scripts/check_plugin.py — which caught the unminted plugin version, 0300 ->
0426.

Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01821k5B3Ysecp9fNYs92Kuy
2026-09-21 00:26:54 -04:00

383 lines
19 KiB
Python

"""Scribe-plugin support endpoints.
Consumed by the Scribe Claude Code plugin (not the web UI). `GET /api/plugin/
context` is curled by the plugin's SessionStart hook to push the operator's
always-on rules + active-project context into the session — Scribe's own push
channel. Auth is the standard `@login_required` path, which accepts a `Bearer
fmcp_<key>` API key (a read-scoped key suffices for this GET).
"""
from __future__ import annotations
from quart import Blueprint, g, jsonify, request
from scribe.auth import admin_required, get_current_user_id, login_required
from scribe.config import Config
from scribe.services import plugin_context as plugin_ctx_svc
from scribe.services import repo_bindings as repo_bindings_svc
from scribe.services import report_check as report_check_svc
from scribe.services.settings import get_admin_setting, set_setting
plugin_bp = Blueprint("plugin", __name__, url_prefix="/api/plugin")
# Setting key for the git URL of the marketplace that serves this plugin (the
# Scribe app's own repo). Instance-global (stored on an admin account) so every
# user's Settings page shows the real, copyable install command.
_MARKETPLACE_KEY = "plugin_marketplace_url"
def _int_list(raw: str | None) -> list[int]:
"""A comma-separated id list from the query string; non-ints dropped."""
return [int(p) for p in (raw or "").split(",") if p.strip().isdigit()]
async def _project_scope() -> tuple[int, str, str]:
"""(project_id, repo, unbound_repo) from the request's `project_id` /
`repo` query args — the one resolution every plugin endpoint shares.
An explicit project_id wins; otherwise the repo remote is resolved
through the caller's bindings, and a remote nobody bound comes back as
`unbound_repo` (normalised) so /context can say "bind this repo".
"""
try:
project_id = int(request.args.get("project_id", 0) or 0)
except (TypeError, ValueError):
project_id = 0
repo = (request.args.get("repo") or "").strip()
unbound_repo = ""
if repo and not project_id:
resolved = await repo_bindings_svc.resolve_project(g.user.id, repo)
if resolved:
project_id = resolved
else:
unbound_repo = repo_bindings_svc.normalize_repo_key(repo)
return project_id, repo, unbound_repo
@plugin_bp.get("/context")
@login_required
async def session_context():
"""Return SessionStart context (always-on rules + active project).
Query:
repo (optional str) — the working repo's git remote. The server
resolves it to the bound project (see services/repo_bindings); an
unbound repo yields a "bind this repo" hint instead. This is how the
active project is determined — the plugin never pins a project id.
project_id (optional int) — the project named directly, and the only
key a caller outside a git repo has (#4085): the plugin's hooks
send it when a `.scribe` marker file names a project. Takes
precedence over `repo`. Access-checked like any other read — an id
this account cannot read loads no project rather than failing.
"""
project_id, _repo, unbound_repo = await _project_scope()
result = await plugin_ctx_svc.build_session_context(
g.user.id, project_id, unbound_repo=unbound_repo
)
return jsonify(result)
@plugin_bp.get("/retrieve")
@login_required
async def autoinject_retrieve():
"""Title-first knowledge auto-inject for the plugin's UserPromptSubmit hook.
Given the user's prompt (`q`), returns a compact awareness hint — note
titles + scores only, never bodies — for the top hits that clear the
per-user gates (see services.plugin_context.build_autoinject_hint). Returns
empty context (most of the time) when disabled or nothing is relevant.
Query:
q (str) — the user's prompt to retrieve against.
repo (optional) — working repo remote; resolved to the bound project
to scope the search (mirrors /context). Unbound or
absent → searches all the user's notes.
project_id (opt) — explicit project scope override (ad-hoc/testing).
exclude_ids (opt) — comma-separated note ids already injected this
session; skipped so each note injects at most once.
exclude_rule_ids — comma-separated rule ids already surfaced this
(opt) session. SHARED with /prior-art and /tool-rules on
purpose: one session keeps ONE rule ledger, so a
rule named by any arm is not re-announced by
another. Ages out (#3751), so salience decays.
held_rule_ids — comma-separated rule ids the session actually
(opt) OPENED, observed from the `get_rule` call itself
rather than claimed. A rule on exclude_rule_ids
was NAMED; a rule here was READ, and the two say
different things about what the reader holds —
so they get different lines (#4100). Shares the
ledger directory and the same clear-on-compact.
TWO ARMS, TWO SETS OF GATES. Rules ride the same hook and the same query
but nothing else: the notes menu can be disabled, thresholded and top-k'd
by the operator without touching whether a rule reaches them. Composed
here rather than inside either builder so neither one's early return can
silently suppress the other.
Rules come FIRST in the payload. A rule or preference governing the answer
is more consequential than a menu of things that might be worth reading,
and a reader who stops after the first block should have stopped after
the right one.
"""
q = (request.args.get("q") or "").strip()
project_id, _repo, _unbound = await _project_scope()
exclude_ids = _int_list(request.args.get("exclude_ids"))
exclude_rule_ids = _int_list(request.args.get("exclude_rule_ids"))
held_rule_ids = _int_list(request.args.get("held_rule_ids"))
rules = await plugin_ctx_svc.build_prompt_rule_hint(
g.user.id, q, project_id=project_id, exclude_rule_ids=exclude_rule_ids, held_rule_ids=held_rule_ids
)
result = await plugin_ctx_svc.build_autoinject_hint(
g.user.id, q, project_id=project_id, exclude_ids=exclude_ids
)
blocks = [b for b in (rules["context"], result["context"]) if b]
result["context"] = "\n\n".join(blocks)
result["rule_ids"] = rules["rule_ids"]
return jsonify(result)
@plugin_bp.get("/tool-rules")
@login_required
async def pre_tool_rules():
"""Standing rules for the plugin's PreToolUse hook on ACTIONS (#3476).
Answers "does a recorded rule speak to the command about to be run?" — the
sibling of /prior-art, which can only answer that question about a code
write. Rules about which tool to reach for (don't curl the forge, don't
stand up a stack, don't run the suite locally) had no retrieval surface at
all before this, which is why they all had to live in the resident preload.
Titles + trigger only, never the statement: the hint says a rule may apply
and hands over `get_rule(id)`. One rule at most (RULEHINT_LIMIT), and empty
most of the time.
Query:
tool (str) — the tool about to run, e.g. `Bash`. Used in
the hint's wording, not in the search: a
rule is about the action, not the harness.
command (str) — the command about to run; the semantic query.
Absent or blank → empty, no search.
repo (optional) — working repo remote, resolved to the bound
project exactly as /retrieve and /prior-art.
exclude_rule_ids (opt) — comma-separated rule ids already surfaced
this session. SHARED with /prior-art's
ledger on purpose: one session keeps one
list, so a rule named by either arm is not
re-offered by the other.
held_rule_ids (opt) — rule ids the session actually OPENED, as
against merely named. Named and read are
different claims about the reader's
context, so they get different lines
(#4100).
Returns `context`, `rule_ids`, and `checkpoint` (#4214, milestone 419).
`checkpoint` IS THE ONE PART OF THIS RESPONSE THAT IS NOT A HINT. It is
empty on almost every call. When present it carries `rule_id`, `title`,
`trigger`, `score` and a rendered `reason`, and it means the hook should
DENY the call rather than annotate it — because every other line this
endpoint returns is delivered by Claude Code alongside the tool RESULT,
so it reaches the reader after the act is composed and reads as
commentary on a decision already made.
It is raised only for a rule (never a preference, which claims no such
force), only for the ranker's top hit, only above the checkpoint
threshold — well above this arm's own floor — and only when the session
has NOT opened that rule. The remedy is one `get_rule` call, and the act
may then be re-submitted unchanged; the hook caps stops at one per rule
and five per session so a mis-set floor degrades to noise rather than to
a session that cannot proceed.
"""
tool = (request.args.get("tool") or "tool").strip()
command = request.args.get("command") or ""
project_id, _repo, _unbound = await _project_scope()
exclude_rule_ids = _int_list(request.args.get("exclude_rule_ids"))
held_rule_ids = _int_list(request.args.get("held_rule_ids"))
result = await plugin_ctx_svc.build_tool_rule_hint(
g.user.id, tool, command,
project_id=project_id, exclude_rule_ids=exclude_rule_ids, held_rule_ids=held_rule_ids,
)
return jsonify(result)
@plugin_bp.get("/prior-art")
@login_required
async def write_path_prior_art():
"""Prior-art hint for the plugin's PreToolUse hook on Write/Edit.
Answers "what is already recorded for the file about to be written?" — by
place (a snippet at this path or in its directory) and by meaning (snippets
resembling the code about to be written). Titles only, gated, and empty most
of the time. See services.plugin_context.build_write_path_hint.
Query:
path (str) — the target file, REPO-RELATIVE, matching the
convention snippet locations are recorded in. An
absolute path simply won't match anything.
code (optional) — the code about to be written; the semantic query.
Omit it and only the location arm runs.
repo (optional) — working repo remote; resolved to the bound project
to scope the search, exactly as /retrieve does. It
is NOT used as the location `repo` filter — see the
service docstring for why those two differ.
project_id (opt) — explicit project scope override (ad-hoc/testing).
exclude_ids (opt) — comma-separated REUSE-class ids already surfaced
this session.
exclude_sync_ids (opt) — comma-separated SYNC-class ids (snippets
recorded AT the edited file, #2708) already
surfaced. A separate channel on purpose: a reuse
hint shown early must not suppress the record-sync
nudge when the recorded file is edited later.
exclude_rule_ids (opt) — comma-separated RULE ids already surfaced
this session. Its own channel like the three
above, and for the same reason: a rule named
twenty turns ago should not be re-offered on
every subsequent write.
held_rule_ids (opt) — RULE ids the session actually OPENED, as
against merely named; drives the third
reference wording (#4100).
exclude_derive (opt) — comma-separated derive keys (a derive group id
or `canon:<snippet_id>`) already named this
session by the ledger arm (#2900); its own
channel, like the two above.
(Returns a `checkpoint` block on the same contract as /tool-rules —
see that endpoint. The write-path HOOK deliberately does not act on
it: scribe_prior_art.sh carries a tested property that it never
returns a permissionDecision, on the operator's decision that a recall
aid may not stand in the way of a write. The block is computed and
returned so that decision can be revisited with evidence rather than
re-argued, and so a change of mind is a hook edit and not a feature.)
shapes (opt) — comma-separated `kind:name` definitions the hook
found in (or enclosing) the payload, kind being
css|sym. The shape ledger's write-path feed
(#2791): when the session recently PULLED a
snippet this payload references or resembles,
these land as instance rows (classified_by=hook).
Honoured only for a caller allowed to write — a
read-scoped key still gets the hint, and never
changes accounting on a GET.
"""
path = (request.args.get("path") or "").strip()
code = request.args.get("code") or ""
project_id, repo, _unbound = await _project_scope()
exclude_ids = _int_list(request.args.get("exclude_ids"))
exclude_sync_ids = _int_list(request.args.get("exclude_sync_ids"))
exclude_derive = [
p.strip() for p in (request.args.get("exclude_derive") or "").split(",") if p.strip()
]
exclude_rule_ids = _int_list(request.args.get("exclude_rule_ids"))
held_rule_ids = _int_list(request.args.get("held_rule_ids"))
shapes = _parse_shapes(request.args.get("shapes") or "")
api_key = getattr(g, "api_key", None)
may_stamp = api_key is None or getattr(api_key, "scope", "") == "write"
result = await plugin_ctx_svc.build_write_path_hint(
g.user.id, path, code=code, project_id=project_id,
exclude_ids=exclude_ids, exclude_sync_ids=exclude_sync_ids,
stamp_shapes=shapes if may_stamp else None,
repo_key=repo_bindings_svc.normalize_repo_key(repo) if repo else "",
exclude_derive=exclude_derive,
exclude_rule_ids=exclude_rule_ids, held_rule_ids=held_rule_ids,
)
return jsonify(result)
# The hook names at most a dozen definitions per write; anything past that is
# a generated file, not a shape being instantiated.
_SHAPES_CAP = 12
def _parse_shapes(raw: str) -> list[tuple[str, str]]:
"""`css:btn-primary,sym:onTrash` → [("css", "btn-primary"), ("sym", "onTrash")].
Unknown kinds and empty names are dropped, duplicates collapse, and the
list is capped — the hook's own cap, re-applied so the contract holds
for any caller."""
out: list[tuple[str, str]] = []
for part in raw.split(","):
kind, _sep, name = part.strip().partition(":")
kind, name = kind.strip(), name.strip()
if kind in ("css", "sym") and name and (kind, name) not in out:
out.append((kind, name))
if len(out) >= _SHAPES_CAP:
break
return out
@plugin_bp.get("/report-check")
@login_required
async def report_check():
"""Record what a Stop hook found in a reply that closed a task (milestone 409 step 5).
The hook decides which completion sections the reply lacks — a local check
of text it can read — and reports the outcome here. For `blocked` the
response carries the `reason` to send the agent back with: the words are
the server's, so every client's hook says the same thing (plugin/PACKAGING.md).
A hook blocks only on a `reason` it received, which means only on a block
that was recorded.
A GET for the reason every plugin endpoint is one: a read-scoped key must
be enough to run the plugin, and this records telemetry the way /retrieve
records a retrieval log.
Query:
outcome (str) — passed | blocked | passed_after_rewrite |
missing_after_rewrite. Anything else is a 400.
missing (opt) — comma-separated sections the reply lacked:
"where it sits", "needs you", "next".
task_ids (opt) — comma-separated ids of the tasks the turn closed.
repo (opt) — working repo remote, resolved like the other arms.
"""
outcome = (request.args.get("outcome") or "").strip()
if outcome not in report_check_svc.OUTCOMES:
return jsonify({"error": f"outcome must be one of {list(report_check_svc.OUTCOMES)}"}), 400
missing = [m for m in (request.args.get("missing") or "").split(",") if m.strip()]
task_ids = _int_list(request.args.get("task_ids"))[:20]
project_id, _repo, _unbound = await _project_scope()
await report_check_svc.record_report_check(
g.user.id, outcome, missing=missing, task_ids=task_ids, project_id=project_id or None,
)
body: dict = {"status": "ok"}
if outcome == "blocked":
body["reason"] = report_check_svc.block_reason(missing)
return jsonify(body)
@plugin_bp.get("/processes")
@login_required
async def process_manifest():
"""Stored Processes as skill-stub specs for the plugin's sync script.
The plugin's `scribe_sync_processes.sh` (run at SessionStart and via the
`/scribe:sync` command) curls this and writes one auto-surfacing local skill
per Process into ~/.claude/skills/. See services/plugin_context.
build_process_manifest. A read-scoped API key suffices.
"""
result = await plugin_ctx_svc.build_process_manifest(g.user.id)
return jsonify(result)
@plugin_bp.get("/marketplace-url")
@login_required
async def get_marketplace_url():
"""The plugin marketplace git URL, readable by any logged-in user so their
Settings page can show a copyable install command. Falls back to the
config default (this deployment's own repo) when no admin override is set."""
url = await get_admin_setting(_MARKETPLACE_KEY, default=Config.PLUGIN_MARKETPLACE_URL)
return jsonify({"marketplace_url": url})
@plugin_bp.put("/marketplace-url")
@admin_required
async def set_marketplace_url():
"""Admin-set the marketplace git URL (e.g. this app's own repo)."""
data = await request.get_json() or {}
url = (data.get("marketplace_url") or "").strip()
if url:
scheme = url.split("://")[0].lower() if "://" in url else ""
if scheme not in ("http", "https"):
return jsonify({"error": "Marketplace URL must use http or https"}), 400
await set_setting(get_current_user_id(), _MARKETPLACE_KEY, url)
return jsonify({"status": "ok"})