Three defects the first sweep left behind, each caught by a rule the guard did not yet have. Statuses. RULES.md gives three; the tree had 25 values, and the free-text ones were carrying progress. "Implemented (wiring landed)" and "Ratified (rev. 2e). Rev. 2e adds the OpenTaskByPR lookup. Rev. 2d revised section 10..." were both changelogs living in the one field the other rules exempt. Every PRD now declares Draft, Approved or Finalised and nothing else; five that declared nothing now declare one. A revision number moves to **Revision:**, a pointer to the authoritative document to **Authority:**, and what a past revision changed to git. CLAUDE.md's citation of state-store-contract rev. 2e still resolves. Stale cross-references. "Supersedes the 2026-09-03 open question", "this bullet previously said", "an earlier draft of this document claimed" all point at text the reader cannot see, because it was replaced. A supersede marker survives only when it names a file, a section or a symbol that can be found. Temporal hedges. 100 uses of today, currently and right now across 31 files. Each dates its sentence to the day it was written: "[notify] today is one webhook URL, read in one place" stops being true the moment this design lands, while still reading as the premise it rests on. The header-field exemption was also too wide. "**Date:** 2026-09-03, Channel resolved 2026-09-22" smuggled a dated status into the one line the rules did not read; the exemption is now a Date holding only a date and a Status holding only one of the three.
9.0 KiB
task_logs pagination and bounded retrieval -- decision
Status: Approved
Date: 2026-08-26
Beads issue: archie-core-vq92
Compounds with: archie-core-p1bl (errors serialise as {})
Decision
task_logs becomes a cursor-paginated, filterable bounded read, not a
one-shot head-of-file window. The reader in internal/logging gains a
second entry point (Page) that walks the file forward from a byte
offset; the tool schema in internal/gateway/task_tools.go surfaces
level, time-range and cursor parameters; the response carries an explicit
"more available" signal so a model can either keep paging or hand the
rest to send_file when the operator wants a complete archive.
We do NOT add an unbounded read path. A single response is still capped at
MaxTailLines (2000) entries. What changes is that pagination makes the
cap reachable in practice instead of silently terminal.
Why now
archie-core-vq92 is the right place to fix this: the chat tool is the
user-visible failure mode, but the same defect propagates to every caller
of internal/logging.Tail (internal/webui/api_logs.go and the
chatTaskLogReaderAdapter in cmd/archied/main.go). Patching the
reader and the tool together is one change in two layers, not two
changes; CLAUDE.md's "package owns its format end-to-end" rule means the
reader is the right place for cursor support regardless.
The compound with p1bl matters: a truncated window of uninformative
errors leaves an operator with nothing. Cursor pagination lets the model
walk the whole file when a level filter does not narrow enough, and the
response's "more available" signal means it can honestly say "I read N of
M and stopped there" instead of presenting the head as complete.
Shape of the API
Reader (internal/logging)
Add a second entry point next to Tail:
// Page walks the log forward from cursor (a byte offset; 0 starts at
// the beginning of the scanned window), applying the same filters as
// Tail. PageResult.Cursor is the file offset just past the last line
// the walk returned; pass it back as Cursor to continue, and stop when
// MoreAvailable is false.
func Page(path string, q Query, cursor int64) (PageResult, error)
PageResult carries Entries, Truncated (the existing meaning --
older entries exist that were not examined within maxScanBytes),
Cursor (the byte offset the next page should resume from), and
MoreAvailable (true iff the scan saw more matching entries beyond the
returned page within the same call).
Cursor is a byte offset rather than an entry ID for two reasons.
First, TaskSink writes JSONL through slog.NewJSONHandler directly --
without routing through FeedHandler -- so the on-disk entries carry no
id field and a per-task attempt monotonic counter would require
either a sink-side change or a reader-side reconstruction that is
brittle against rotation. Second, the byte offset uniquely identifies
"the next unread position" in a sequential scan, which is exactly what
the call pattern needs. A cursor that points at or past the file's
current size is treated as end-of-stream (empty page, MoreAvailable
false).
The implementation is readWindow (reads the tail-most maxScanBytes
into a single in-memory buffer) plus readLines (clamps the cursor into
the window, drops a partial first line on the size-cap seek, then
delegates) plus walkWindow (walks the buffer line-by-line by exact
slice offset). Offsets are into one contiguous buffer, so they are
trivially correct -- unlike an incremental reader over the file, whose
bufio.Reader buffer-ahead would make its underlying file position
unusable as a cursor.
Tail is NOT a wrapper over Page: it keeps its legacy ring-buffer
semantics (newest Limit matches, oldest dropped, Truncated set on
ring overflow). They share readWindow/walkWindow but stay separate
entry points precisely so Tail's established behaviour is unchanged.
Query gains two optional fields
Since time.Time // inclusive lower bound on entry.Time
Until time.Time // inclusive upper bound on entry.Time
Levels, Component, Contains, Limit are unchanged. The existing
filter logic and tests move unchanged.
Tool schema (internal/gateway/task_tools.go)
ChatTaskLogQuery gains three fields:
Level []string(forwarded toQuery.Levels)Since time.Time/Until time.Time(forwarded asQuery.Since/Until)AfterID int64(the cursor; 0 to start at the beginning)
ChatTaskLogResult gains two fields:
Cursor int64-- highest ID in this page; pass back asAfterIDto continueMoreAvailable bool-- true iff the scan saw more matching entries that did not fit in this page
The tool description is updated to teach the model the pagination
protocol in one sentence: "pass the previous cursor back as after_id
until more_available is false; for an entire archive, prefer
send_file."
Production adapter (cmd/archied/main.go)
chatTaskLogReaderAdapter passes the new fields through to the reader.
TaskRegistry.Path is already attempt-scoped, so paging stays inside
one attempt.
What we deliberately do NOT do
- No time-range-only read without a cursor. Time bounds are a filter,
not a window. A model that wants only "errors in the last hour" passes
both
Levels=[ERROR]andSince=now-1h. The page is still the same size; the response'sMoreAvailabletells the caller whether the hour fit. - No "save the full log" via the tool. That is what
archie-core-absb'ssend_fileis for. The PRD description for vq92 raised this as an open question; the answer is "use the existing tool" rather than building a parallel archive endpoint. - No reversal or random access. Paging is forward-only on the byte
offset. Reverse scan would require either an index or a full file
pass; neither earns its cost given
Tailalready handles "show me the end." - No bumping
MaxTailLinespast 2000. A single response must not be able to exhaust the model's context. Paging makes the cap reachable but does not relax it. - No raising
maxScanBytes, and no walking rotated generations. See the ceiling below; lifting it is a separate decision with its own cost.
The scan-window ceiling, and what it means for "complete"
maxScanBytes (8 MiB) bounds the tail of the file any single call will
examine. Paging does not lift that bound: every call re-derives the
window from the file's current size and clamps the incoming cursor up
into it, so a paging loop walks the window exhaustively and then stops.
For a log larger than 8 MiB the oldest bytes are unreachable through
task_logs at any cursor, and rotated generations (attempt-N.jsonl.1,
.2, ...) are separate files the reader never opens.
This is a deliberate limit, not an oversight, and it is why criterion 1 below says "paging or file delivery" rather than "paging". The honest contract is:
- Within the window: paging reaches every matching entry, in order, exactly once.
- Beyond it:
Truncatedis true, and that is the only correct thing to report. An agent that pages tomore_available == falsewhiletruncatedis true has read the window, not the log, and must say so.
Lifting the ceiling would mean either buffering an unbounded file or teaching the reader to walk rotation history. Neither is needed to close vq92 — the issue's own text accepts file delivery as the answer for "save the full logs" — so both are deliberately out of scope here.
Acceptance criteria
The issue's criteria, restated against the new shape:
- An operator can obtain a task's complete logs through the agent: by
paging when the log fits the scan window, and by file delivery
(
send_file) when it does not. - A bounded response says so explicitly --
MoreAvailablefor "more matches this call did not return",Truncatedfor "the scan cap hid older entries" -- so the agent can report honestly instead of implying completeness. Levels,Since,Untilfilters exist so a bounded call can answer a specific question without paging.- The cap stays enforced: one response cannot exceed
MaxTailLines, and one call cannot examine more thanmaxScanBytes. - Tests cover a log larger than
MaxTailLinesand assert that repeated paged calls retrieve every entry without gaps or duplication, plus a log larger thanmaxScanBytesasserting the window is walked contiguously andTruncatedis set.
Files this change touches
internal/logging/reader.go-- addPage, expandQuerywithSince/Untilinternal/logging/reader_test.go-- new tests forPage,Since/Until, paging without gapsinternal/gateway/task_tools.go-- newChatTaskLogQueryandChatTaskLogResultfields, schema descriptioninternal/gateway/task_tools_test.go-- new tests for the new fields on the toolcmd/archied/main.go-- adapter passes through the new fields
The webui handler in internal/webui/api_logs.go is untouched: it is
already filter-aware (Levels, Component, Contains, Limit) and is not the
failing surface vq92 reports. If the dashboard later wants the same
pagination, it consumes the same reader.