pi-sleep-agent 0.8.13 → 0.8.15
This diff represents the content of publicly available package versions that have been released to one of the supported registries. The information contained in this diff is provided for informational purposes only and reflects changes between package versions as they appear in their respective public registries.
- package/README.md +9 -0
- package/package.json +1 -1
- package/prompts/consolidate.md +10 -0
- package/sleepkit/cli.py +10 -1
- package/sleepkit/consolidate.py +12 -4
- package/sleepkit/store.py +82 -4
package/README.md
CHANGED
|
@@ -34,6 +34,14 @@ Sleep adds a reviewable memory cycle to nine selectable coding-agent targets. It
|
|
|
34
34
|
- **Project isolation.** A project's provisional memory is injected only into that project.
|
|
35
35
|
- **Reviewable permanence.** Each approved memory in your index carries its review date and source, so you can tell when a saved instruction was pinned.
|
|
36
36
|
- **Relevance-ranked injection.** With a query, provisional memory is ordered by overlap with the live request and each line carries its age, so the most relevant record survives the `--max` cut and stale facts read as stale.
|
|
37
|
+
- **Two clocks per record.** `ts` is when the record was written; `event_at` is when
|
|
38
|
+
the thing it describes actually happened, and it is written **only when the record
|
|
39
|
+
states a date** (a date equal to the capture day is not a second clock). An empty
|
|
40
|
+
`event_at` means the record does not say, never a guess, so `show` can answer
|
|
41
|
+
"when was this true" without inventing an event time.
|
|
42
|
+
- **Tagged records.** Consolidation writes 1-4 lowercase topic tags per record (`deploy|aws`),
|
|
43
|
+
normalized and capped in `store`, and `inject --topic` selects on them. A tag is a
|
|
44
|
+
hint a model proposed, so a wrong tag costs a missing filter hit, never a wrong fact.
|
|
37
45
|
- **Served-record accounting.** Every injection stamps the records it actually printed (`injected`, `injected_at` on the record itself), so `doctor` can name a provisional that **no session has ever asked for** in 8+ weeks (`unused-candidate-never-served`). That is a different fact from "nobody re-learned it", and only the record can tell the two apart; report only, nothing is dropped.
|
|
38
46
|
- **Indexed by default.** A long observation is injected as its opening claim plus a pointer (`[+714 chars: sleep.py show <id>]`) instead of its whole narrative; `inject --full` prints the complete bodies. Measured on a live pool of 11 records: 5581 -> 3377 characters per injection.
|
|
39
47
|
- **Stale-value scan.** `doctor --stale <value>` finds which approved records still hold a value you know has changed (for example a deployment URL). Unreviewed provisional candidates are also flagged: a fact the agent proposed once and never rediscovered for 4+ weeks keeps getting injected into every session's context, so `doctor` surfaces it as `stale-candidate-unrediscovered` (a report line it never acts on, nothing is dropped).
|
|
@@ -338,6 +346,7 @@ name: it routes through its own config, so only `SLEEP_MODEL` reaches it.
|
|
|
338
346
|
| `./sleep.py inject --project "$PWD"` | Print provisional context for the current project, newest first, each line dated |
|
|
339
347
|
| `./sleep.py inject --project "$PWD" --query "..."` | Same, but rank records by overlap with the live request so the most relevant survive `--max` |
|
|
340
348
|
| `./sleep.py inject --project "$PWD" --full` | Print each record's whole body instead of the index line |
|
|
349
|
+
| `./sleep.py inject --project "$PWD" --topic deploy` | Only records whose `tags` carry that topic (an index lookup, not a keyword search) |
|
|
341
350
|
|
|
342
351
|
Run `./sleep.py <command> --help` for command-specific options.
|
|
343
352
|
|
package/package.json
CHANGED
package/prompts/consolidate.md
CHANGED
|
@@ -18,6 +18,9 @@ Each proposal is header fields, a blank line, then the observation:
|
|
|
18
18
|
source: user | agent | inferred
|
|
19
19
|
kind: fact | preference | procedure | warning
|
|
20
20
|
trust: verified | observed | speculative
|
|
21
|
+
tags: <1-4 pipe-separated lowercase topics, e.g. deploy|aws or gitlab|auth>
|
|
22
|
+
event_at: <optional YYYY-MM-DD, ONLY when the observation itself states a date for
|
|
23
|
+
the thing you are describing; omit the line when it does not>
|
|
21
24
|
evidence: <optional, comma separated file paths or session refs>
|
|
22
25
|
|
|
23
26
|
<one observation, 1-3 sentences, written so a reader in a future session can
|
|
@@ -52,5 +55,12 @@ here.
|
|
|
52
55
|
|
|
53
56
|
- You have read access only, and you do not need it: the buffer and the existing
|
|
54
57
|
pool are already in your prompt. Do not run commands.
|
|
58
|
+
- The `tags` line is the only index a future session can filter on: use the
|
|
59
|
+
topic words a person would search for (deploy, gitlab, auth), not the
|
|
60
|
+
sentence's verbs. Never repeat the kind as a tag.
|
|
61
|
+
- `event_at` is only for a date the observation already states about the thing it
|
|
62
|
+
describes ("on 2026-10-14 the staging host moved"). Never derive it from what you
|
|
63
|
+
know, never write the session's own date, and omit the line otherwise: the record
|
|
64
|
+
keeps two separate clocks and an empty `event_at` means "the record does not say".
|
|
55
65
|
- Proposals are cheap and refusals are cheap; a poisoned memory store is not. When
|
|
56
66
|
in doubt, emit fewer and mark the remainder `speculative`.
|
package/sleepkit/cli.py
CHANGED
|
@@ -155,6 +155,10 @@ def cmd_show(a) -> int:
|
|
|
155
155
|
return 2
|
|
156
156
|
meta, body = got
|
|
157
157
|
_p(store.emit(meta, body))
|
|
158
|
+
event_at = str(meta.get("event_at") or "").strip()
|
|
159
|
+
if event_at:
|
|
160
|
+
_p("# two clocks: recorded %s, the recorded event happened %s"
|
|
161
|
+
% (str(meta.get("ts", "?"))[:10], event_at))
|
|
158
162
|
# Show the supersede chain: with only the two pointer fields visible, answering
|
|
159
163
|
# "which record is current, and what did it replace?" means reading other files.
|
|
160
164
|
chain = store.supersession_chain(a.id)
|
|
@@ -535,8 +539,10 @@ def _index_line(body: str, cid: str) -> str:
|
|
|
535
539
|
|
|
536
540
|
def cmd_inject(a) -> int:
|
|
537
541
|
"""Print the provisional-memory block for a session's context. Framed as context."""
|
|
542
|
+
topic = getattr(a, "topic", "") or ""
|
|
538
543
|
pool = [c for c in store.list_candidates(status="provisional", with_body=True)
|
|
539
|
-
if _within_project(c.get("cwd", ""), a.project)
|
|
544
|
+
if _within_project(c.get("cwd", ""), a.project)
|
|
545
|
+
and store.matches_topic(c, topic)]
|
|
540
546
|
full = bool(getattr(a, "full", False))
|
|
541
547
|
recs = []
|
|
542
548
|
served = []
|
|
@@ -721,6 +727,9 @@ def main(argv=None) -> int:
|
|
|
721
727
|
ij.add_argument("--query", default="",
|
|
722
728
|
help="rank provisional records by overlap with this text "
|
|
723
729
|
"(the live request); newest first when omitted")
|
|
730
|
+
ij.add_argument("--topic", default="",
|
|
731
|
+
help="only records tagged with this topic (matches the record's "
|
|
732
|
+
"tags field; records with no tags never match)")
|
|
724
733
|
ij.add_argument("--full", action="store_true",
|
|
725
734
|
help="print each record's whole body instead of the index line")
|
|
726
735
|
ij.set_defaults(fn=cmd_inject)
|
package/sleepkit/consolidate.py
CHANGED
|
@@ -15,7 +15,7 @@ from datetime import datetime
|
|
|
15
15
|
|
|
16
16
|
from . import llm, paths, safety, store
|
|
17
17
|
|
|
18
|
-
_HDR = re.compile(r"^(source|kind|trust|evidence)\s*:\s*(.+)$", re.M)
|
|
18
|
+
_HDR = re.compile(r"^(source|kind|trust|tags|event_at|evidence)\s*:\s*(.+)$", re.M)
|
|
19
19
|
|
|
20
20
|
|
|
21
21
|
def pending_dir() -> str:
|
|
@@ -190,7 +190,8 @@ def distill(session: str | None = None, model: str | None = None,
|
|
|
190
190
|
|
|
191
191
|
def parse_proposal(text: str) -> dict:
|
|
192
192
|
"""Split one proposal into header fields + body. Headers are optional."""
|
|
193
|
-
fields = {"source": "agent", "kind": "fact", "trust": "observed",
|
|
193
|
+
fields = {"source": "agent", "kind": "fact", "trust": "observed",
|
|
194
|
+
"tags": "", "event_at": "", "evidence": ""}
|
|
194
195
|
for m in _HDR.finditer(text):
|
|
195
196
|
fields[m.group(1)] = m.group(2).strip()
|
|
196
197
|
body = _HDR.sub("", text).strip()
|
|
@@ -204,7 +205,7 @@ def parse_proposal(text: str) -> dict:
|
|
|
204
205
|
# fallback. Neither decides what becomes memory: every block still goes through
|
|
205
206
|
# ingest_records(), which is where the gates live.
|
|
206
207
|
_SEP_LINE = re.compile(r"^[^A-Za-z0-9\n]*PROPOSAL[^A-Za-z0-9\n]*$", re.I | re.M)
|
|
207
|
-
_HDR_LINE = re.compile(r"^(source|kind|trust|evidence)\s*:.*$", re.M)
|
|
208
|
+
_HDR_LINE = re.compile(r"^(source|kind|trust|tags|event_at|evidence)\s*:.*$", re.M)
|
|
208
209
|
|
|
209
210
|
|
|
210
211
|
def split_blocks(text: str) -> tuple:
|
|
@@ -306,7 +307,14 @@ def ingest_records(parsed: list, session: str = "", cwd: str = "",
|
|
|
306
307
|
item["body"], kind=f.get("kind", "fact"), source=source, trust=trust,
|
|
307
308
|
session=session or os.environ.get("SLEEP_SESSION", ""),
|
|
308
309
|
cwd=cwd or os.getcwd(),
|
|
309
|
-
evidence=[x for x in f.get("evidence", "").split(",") if x]
|
|
310
|
+
evidence=[x for x in f.get("evidence", "").split(",") if x],
|
|
311
|
+
# The proposal's own tag line is a hint, never authority: store.add_candidate
|
|
312
|
+
# normalizes and caps it, and a wrong tag only costs a bad topic filter.
|
|
313
|
+
tags=[x for x in str(f.get("tags", "")).split("|") if x.strip()],
|
|
314
|
+
# The model may state when the described thing happened. add_candidate
|
|
315
|
+
# validates it and keeps the two clocks apart; a date the record does not
|
|
316
|
+
# state stays empty rather than being backfilled from the capture time.
|
|
317
|
+
event_at=f.get("event_at", ""))
|
|
310
318
|
(accepted if ok else refused).append({"body": item["body"][:60], "result": out})
|
|
311
319
|
return {"proposals": len(parsed), "accepted": len(accepted), "refused": refused,
|
|
312
320
|
"accepted_ids": [a["result"] for a in accepted]}
|
package/sleepkit/store.py
CHANGED
|
@@ -53,7 +53,7 @@ IDX_END = "<!-- sleep:end -->"
|
|
|
53
53
|
|
|
54
54
|
# ------------------------------------------------------------- frontmatter --
|
|
55
55
|
def emit(record: dict, body: str = "") -> str:
|
|
56
|
-
order = ["id", "ts", "source", "trust", "kind", "status", "session", "cwd",
|
|
56
|
+
order = ["id", "ts", "event_at", "source", "trust", "kind", "status", "session", "cwd",
|
|
57
57
|
"tags", "evidence", "supersedes", "superseded_by",
|
|
58
58
|
"valid_from", "valid_until", "redacted", "notes"]
|
|
59
59
|
# Known fields first in a stable order, then anything else the record carries.
|
|
@@ -324,7 +324,8 @@ def record_injection(ids: list[str]) -> int:
|
|
|
324
324
|
def add_candidate(text: str, kind: str = "fact", source: str = "agent",
|
|
325
325
|
trust: str = "observed", session: str = "", cwd: str = "",
|
|
326
326
|
evidence: list[str] | None = None, supersedes: list[str] | None = None,
|
|
327
|
-
notes: str = ""
|
|
327
|
+
notes: str = "", tags: list[str] | None = None,
|
|
328
|
+
event_at: str = "") -> tuple[bool, str]:
|
|
328
329
|
"""T1: distilled, provisional. Never injected as a directive."""
|
|
329
330
|
text = (text or "").strip()
|
|
330
331
|
if not text:
|
|
@@ -360,16 +361,22 @@ def add_candidate(text: str, kind: str = "fact", source: str = "agent",
|
|
|
360
361
|
return False, "duplicate of earlier refusal %s (%.2f); this idea was judged and rejected" % (
|
|
361
362
|
ref_id, ref_score)
|
|
362
363
|
|
|
364
|
+
ts = paths.now()
|
|
365
|
+
# Trust the caller's own statement when it gives one; otherwise read a date out of
|
|
366
|
+
# the record's own wording. Never invent: an empty event_at means the record does
|
|
367
|
+
# not say when the described thing happened, which is a different fact from ts.
|
|
368
|
+
event_at = _valid_event_at(event_at, ts) or stated_event_at(red, ts)
|
|
363
369
|
rec = {
|
|
364
370
|
"id": paths.new_id("c"),
|
|
365
|
-
"ts":
|
|
371
|
+
"ts": ts,
|
|
366
372
|
"source": source if source in safety.SOURCES else "inferred",
|
|
367
373
|
"trust": safety.cap_trust(source, trust),
|
|
368
374
|
"kind": kind if kind in KINDS else "fact",
|
|
369
375
|
"status": "provisional",
|
|
376
|
+
"event_at": event_at,
|
|
370
377
|
"session": session or os.environ.get("SLEEP_SESSION", ""),
|
|
371
378
|
"cwd": cwd or os.getcwd(),
|
|
372
|
-
"tags": "|".join(
|
|
379
|
+
"tags": "|".join(_clean_tags(tags)),
|
|
373
380
|
"evidence": "|".join(evidence or []),
|
|
374
381
|
"supersedes": "|".join(supersedes or []),
|
|
375
382
|
"superseded_by": "",
|
|
@@ -817,6 +824,77 @@ def prune_buffer(older_than_hours: float = 0.0, session: str | None = None) -> d
|
|
|
817
824
|
return {"moved": moved, "remaining": buffer_counts()}
|
|
818
825
|
|
|
819
826
|
|
|
827
|
+
# Tags are the only index a record carries besides its project, so they are kept
|
|
828
|
+
# machine-shaped: lowercase, no spaces, bounded in number and length. A model may
|
|
829
|
+
# propose them, which is exactly why they are normalized here rather than trusted.
|
|
830
|
+
MAX_TAGS = 6
|
|
831
|
+
MAX_TAG_CHARS = 32
|
|
832
|
+
|
|
833
|
+
|
|
834
|
+
# The two clocks are separate: `ts` is when the record was written (always set), and
|
|
835
|
+
# `event_at` is when the thing being described actually happened (only set when the
|
|
836
|
+
# record states it). Backfilling event_at from ts would manufacture a fact nobody
|
|
837
|
+
# recorded, and every later "where was X in October" question would answer with the
|
|
838
|
+
# capture date. Empty means unknown, and unknown is reported as unknown.
|
|
839
|
+
_DATE_RE = re.compile(r"\b(\d{4}-\d{2}-\d{2})\b")
|
|
840
|
+
|
|
841
|
+
|
|
842
|
+
def _valid_event_at(value: str, ts: str = "") -> str:
|
|
843
|
+
"""Accept a caller-stated event date only in YYYY-MM-DD form, never the capture day."""
|
|
844
|
+
m = _DATE_RE.fullmatch(str(value or "").strip())
|
|
845
|
+
if not m:
|
|
846
|
+
return ""
|
|
847
|
+
if ts and m.group(1) in ts[:10]:
|
|
848
|
+
return ""
|
|
849
|
+
return m.group(1)
|
|
850
|
+
|
|
851
|
+
|
|
852
|
+
def stated_event_at(text: str, fallback_ts: str = "") -> str:
|
|
853
|
+
"""Return a date the record's own wording states, else "".
|
|
854
|
+
|
|
855
|
+
Deliberately a literal match, not an inference: a relative phrasing ("last week")
|
|
856
|
+
is left to the model that reads the record, and this only accepts a date the text
|
|
857
|
+
spells out. A date equal to the capture day is not an event time either - it is
|
|
858
|
+
the same clock, and writing it twice would invent a distinction.
|
|
859
|
+
"""
|
|
860
|
+
m = _DATE_RE.search(text or "")
|
|
861
|
+
if not m:
|
|
862
|
+
return ""
|
|
863
|
+
value = m.group(1)
|
|
864
|
+
if fallback_ts and value in fallback_ts[:10]:
|
|
865
|
+
return ""
|
|
866
|
+
return value
|
|
867
|
+
|
|
868
|
+
|
|
869
|
+
def _clean_tags(tags) -> list:
|
|
870
|
+
out, seen = [], set()
|
|
871
|
+
for raw in tags or []:
|
|
872
|
+
tag = re.sub(r"[^a-z0-9._:+-]+", "-", str(raw).strip().lower()).strip("-")
|
|
873
|
+
if not tag or tag in seen:
|
|
874
|
+
continue
|
|
875
|
+
seen.add(tag)
|
|
876
|
+
out.append(tag[:MAX_TAG_CHARS])
|
|
877
|
+
if len(out) >= MAX_TAGS:
|
|
878
|
+
break
|
|
879
|
+
return out
|
|
880
|
+
|
|
881
|
+
|
|
882
|
+
def matches_topic(meta: dict, topic: str) -> bool:
|
|
883
|
+
"""True when a record carries the topic as one of its tags (substring, either way).
|
|
884
|
+
|
|
885
|
+
Deliberately tag-only: matching the body would turn a topic filter into a keyword
|
|
886
|
+
search, which is what --query already does. Substring in both directions so
|
|
887
|
+
`deploy` selects `deploy` and `deployment`, not `predeployed`."""
|
|
888
|
+
want = str(topic or "").strip().lower()
|
|
889
|
+
if not want:
|
|
890
|
+
return True
|
|
891
|
+
for tag in meta.get("tags") or []:
|
|
892
|
+
tag = str(tag).strip().lower()
|
|
893
|
+
if want == tag or want in tag or tag in want:
|
|
894
|
+
return True
|
|
895
|
+
return False
|
|
896
|
+
|
|
897
|
+
|
|
820
898
|
# A cheap pre-filter for the LLM adjudicator. Token Jaccard is a poor *decider* on
|
|
821
899
|
# short observations (measured: no real pair in a live 11-record pool cleared 0.5),
|
|
822
900
|
# but it is a fine *shortlist*: pairs that share at least a handful of content words
|