A user-provided "research budget" now defines the required search volume:
it is binding and must be spent in full, overriding the default "stop at
saturation" rule.
- STEP 0: split the budget into user-set (binding, must be used up) vs
self-estimated (when the user gave no number)
- VOLUME: limit "stop at saturation" to the no-budget case; add a
MANDATORY BUDGET block requiring the full budget be spent on genuine
broadening/lateral/primary-source/verification searches, not padding
- Apply identically to bundles/research/ru.yaml and en.yaml
- Bump researcher version 2 -> 3 in index.yaml; refresh content-hashes.json
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Rework the "WHERE TO WRITE THE RESULT" section of the researcher role so
the report document is created at the very beginning of the run (right
after the plan, before any searches) and filled dynamically after each
finding, instead of being dumped in one pass at the end.
- Rewrite the section identically in bundles/research/ru.yaml and en.yaml
- Bump researcher version 1 -> 2 in index.yaml
- Refresh scripts/content-hashes.json via check.mjs --update-hashes
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Editorial roles (Corrector/Factchecker) brute-forced `get_node` block-by-block to
find occurrences (unquoted «ё», straight quotes, «т.е.»), burning tokens. New
`search_in_page(pageId, query, {regex?, caseSensitive?, limit?})` reads the page's
ProseMirror JSON via the existing getPageRaw and searches it IN MEMORY — no server
endpoint, no DB/schema change, no touch to the packages/mcp/src/lib schema mirror.
New pure `searchInDoc(doc, query, opts)` (packages/mcp/src/lib/page-search.ts):
recursive descent to each TEXT CONTAINER (paragraph/heading/table-cell paragraph),
glues its inline text via `blockPlainText` (a match survives inline-mark
boundaries — e.g. «т.е.» split across bold/italic), searches literal (indexOf) or
regex, and returns `{ total, truncated, matches:[{ nodeId, blockIndex, type,
before, match, after }] }`. `nodeId` is the container's attrs.id or the
`#<topLevelIndex>` of the enclosing top-level block — the SAME ref format
get_node/patch_node/comment-anchoring accept (verified identical to getNodeByRef),
so the agent goes straight from a hit to a targeted comment; `before`/`after` are
~40-char windows for a unique selection. `total`/`truncated` always reported (never
silent truncation). Lives in the SHARED_TOOL_SPECS registry → exposed in BOTH
transports (external /mcp + in-app AI-chat), with a SERVER_INSTRUCTIONS line and a
DocmostClientLike signature + contract-test entry. Corrector/Factchecker prompts
get a one-line "use search_in_page first" hint (versions bumped, catalog hash lock
refreshed).
Guards: empty/whitespace query → clear error; invalid regex → clear error (not a
generic 500); zero-length regex matches (`\b`, `a*`) skipped with lastIndex
advanced (no loop/flood); MAX_PATTERN_LENGTH=1000, MAX_CONTAINER_TEXT=100k bound
each exec; limit clamped [1,200] (default 50).
Tests: new page-search.test.mjs (17) — literal+regex, case-sensitivity,
mark-boundary glue, nodeId for paragraph/heading (attrs.id) and table-cell
(#<index> fallback), context bounds, limit/total/truncated + clamp, invalid
regex/empty/over-long errors, zero-length skip, empty-doc null-safety.
mcp: tsc clean; node --test 467 passed (+17). apps/server: tsc --noEmit clean
(DocmostClientLike + wiring). catalog check.mjs OK.
Known limitations (from internal review, non-blocking):
- Residual ReDoS: a crafted catastrophic-backtracking pattern (e.g. `(a+)+$`)
against a large single container can hang the event loop — JS regex is not
interruptible, so the length caps bound the base but not the backtracking.
Realistic exposure is low (containers are small; the pattern is supplied by the
authenticated model). Candidate for a follow-up hardening (safe-regex validation
or a worker+timeout) if it matters.
- Case-insensitive LITERAL search folds via toLowerCase; a char whose lowercase
differs in length (e.g. Turkish İ) BEFORE a match could shift the context
window — negligible for the RU/EN editorial scenario.
- On a `#<index>` table-cell fallback, `type` is the inline container ("paragraph")
while nodeId addresses the top-level block — addressing is correct; the field is
documented as the container's type.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Chat now shows the agent's name in the comment header, so the '[Copyedit]' /
'[Structure]' / '[Style]' / '[Facts]' prefix each role prepended just
duplicated the visible author.
- Remove the 'open the comment with the label [Role]' instruction from all four
labelled roles (structural-editor, line-editor, fact-checker, proofreader),
ru + en; the narrator was already label-free.
- Severity tags ([Critical]/[Major]/[Minor]) and the fact-checker's verdicts
([Incorrect]/[Unverified]/…) are kept — they carry meaning, not the role name.
- Versions bumped: structural-editor 3->4, line-editor 3->4, fact-checker 4->5,
proofreader 6->7; content-hash lock refreshed.
Check: agent-roles-catalog check.mjs OK.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
The copyeditor had to be re-run several times to surface all issues: it has no
'work the whole document' instruction (unlike the developmental editor and the
narrator), and the severity labels nudge it toward reporting only the salient
few.
- Add a HOW TO WORK section (ru + en): one pass over the whole text start to
finish; flag EVERY violation including all repeat occurrences and [Minor]
items; don't summarize instead of marking up; one run covers the whole text,
not just 'the most important'.
- proofreader version 5 -> 6, content-hash lock refreshed.
Check: agent-roles-catalog check.mjs OK.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
The copyeditor was emitting summary comments ('throughout, unify KB'; 'switch
the decimal separator everywhere') that carry no applicable suggestedText — a
human can't one-click any of them. This was actively instructed: the TONE
section told it to 'group repeated fixes so you don't spawn dozens of identical
comments'.
- Invert that TONE guidance: spread repeated fixes across the specific spots;
ten targeted comments each with a ready replacement beat one blanket note;
'spawning' comments is normal for a copyeditor.
- HOW-TO: explicitly forbid summary/consistency notes and require a separate
targeted comment with its own suggestedText on EVERY occurrence; the only
exception is a note that genuinely can't be a fragment replacement.
- ru + en mirrored; proofreader version 4 -> 5, content-hash lock refreshed.
Check: agent-roles-catalog check.mjs OK.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
- editorial roles (ru/en): proofreader and line editor attach suggestedText
replacements to targeted fixes; fact-checker ALWAYS attaches the ready
correction for [Incorrect] verdicts; structural editor and narrator get a
light-touch rule for in-place rewordings; role versions bumped and the
content-hash lock refreshed
- MCP SERVER_INSTRUCTIONS: route 'propose a concrete text fix for one-click
human approval' to create_comment with suggestedText (unique-selection
reminder); build/ artifacts rebuilt
- AI-chat SAFETY_FRAMEWORK: mention the comment-suggestion capability so the
default assistant offers ready fixes instead of only describing changes
Checks: catalog check.mjs OK; @docmost/mcp tests 448/448; server
ai-chat.prompt spec 28/28.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Rework the fact-checker editorial role prompt so it stops commenting on
correct facts and only flags problems (errors, doubtful, unverifiable).
- Add the directive "don't write/comment that a fact is right or confirmed:
your job is to find errors, not confirm facts" to both RU and EN bundles.
- Remove the [Подтверждено]/[Verified] verdict; reframe the verdict list as
"for problem claims only".
- Reword the role description (no longer "confirms") and the
comment-on-every-claim rule to "problem claims only".
- Bump fact-checker role version 2 -> 3 and refresh the content-hash lock.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
The proofreader role content was changed (STYLE SHEET block removed) without
bumping its catalog version, so clients never saw an update. Bump proofreader
2 -> 3, and add a content-hash guard so this can't happen silently again.
- index.json: proofreader version 2 -> 3
- scripts/check.mjs: new content-hash guard. A scripts/content-hashes.json lock
maps slug -> { version, hash } (sha256 over emoji/autoStart/name/description/
instructions/launchMessage across all languages). check.mjs now fails when a
role's content changed without bumping its version; the new --update-hashes
(alias --fix) refreshes the lock but refuses to write when a bump is missing.
- check.mjs: also require every index.json role to carry a finite numeric
version (matches the server's catalog validation), with defense-in-depth so a
missing version can't bypass the bump guard.
- scripts/content-hashes.json: new lock artifact (not part of the served catalog).
- README.md: document the guard, the lockfile, --update-hashes, and the
prune-then-readd limitation.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
agent_vscode
claude code agent 227