Why: relaxAndRemove cleared the per-exec id directory but left the -ws
parent behind, so every conversation accumulated an empty husk - the
live probe caught it in the workspace listing.
Evidence: go test ./... green; TestOverlayCleansShellDir pins the
behavior with a real bwrap roundtrip.
Impact: One extra os.Remove of the (now empty) parent; failures are
best-effort as before.
Refs: Dev task feature/agent-s0-workspace-overlay (fix-up of #135).
Why: The sandbox tier wrote the session directory directly, so a buggy
or runaway script could leave the workspace half-mutated and nothing
bounded what a timeout had already written. The plan's first landing
shape (section 4.4) swaps the session bind for a frozen-base overlay:
each exec freezes the session dir (hardlink clone + sha256 baseline),
runs on a private upper, and only a clean exit promotes the change set
back via CAS (baseline hash match -> atomic replace/remove; externally
changed targets double-store as *.conflict-* instead of clobbering).
Timeouts and kills discard the upper entirely - half-written state
never lands.
Evidence: go test ./... green including bwrap integration roundtrips
(TestSandboxWorkspaceOverlayRoundtrip: add/modify/delete through a real
overlayfs upper with whiteout handling and ws cleanup;
TestSandboxWorkspaceOverlayTimeoutDiscards: partial writes discarded on
timeout) and unit anchors for freeze manifest (.agentx excluded), CAS
add/modify, conflict double-store, and the change summary.
Impact: Opt-in via --exec-workspace overlay / AGENT_SYSTEM_EXEC_WORKSPACE
(only with --exec-backend sandbox; fail-closed on unknown modes). The
model sees a one-line change summary appended to the tool output, so
discards and conflict files are visible. Default behavior is unchanged.
Refs: proposal 2026-10-10-agent-delegation-routing-sandbox.md section
4.4; Dev task feature/agent-s0-workspace-overlay.
Why: Real runs were fully serial: the direct path never set the kernel's
parallel tool mode, delegate_task handed the parent only the child's
160-char summary (the full text was dropped), MiniMax streamed its
reasoning inside content <think> blocks, and the EXT-16 sandbox hid
~/.local so python-docx/openpyxl/pptx could not even import. All four
were measured defects (F1/F2/F3/F5) in the delegation/routing proposal.
Why: Serial-only mutual exclusion is not enough for order-sensitive
tools - the card-stream test failed with out-of-order appends - so the
ordering moved into the kernel (SerialTools partition, Agent-Native
#44); process events pair by CallID (#42) and tool results write back
through the injected call context (#43).
Evidence: go test ./... green, including new regression anchors:
TestDelegateToolReturnsFullText, TestToolEventsPairByCallID,
TestToolExecModeAndSerialSet, TestOpenLLMClientReasoningSplitAuto,
TestSandboxArgvDepsLayer, TestExecEnvJoinsPythonPath.
Impact: Direct runs now execute independent same-turn calls
(searches, reads, delegate_task, review_findings) concurrently by
default (runtime.parallelTools=false restores sequential); card flow,
plans, workflows, exec primitives and file edits keep model-declared
order. delegate_task returns the capped full child text. MiniMax-*
models get reasoning_split=true so thinking streams via
reasoning_content. --exec-deps-dir/AGENT_SYSTEM_EXEC_DEPS_DIR mounts a
read-only dependency layer into the sandbox and merges PYTHONPATH.
Bumps deps/agent-native to c348441.
Refs: proposal 2026-10-10-agent-delegation-routing-sandbox.md (E1,
F1-F5); Agent-Native #42/#43/#44; Dev task feature/agent-e1-parallel-fixes.
Why: User rulings 4/5 plus the P2-P4 live evidence. Keyword scoring dimensions
(readability, actionability) kept failing valid documents - normal Chinese
section headings flagged as column titles, loops of 4 corrections - because
they grade writing preferences, not verifiable facts. Unverified numbers
(~10-13 per run) remained the largest quality leak through every phase.
And the verifier runs on the same model as the worker (lightweight falls back
to chat), so "independent" verification was self-review. Separately, svc
seeding after boot never registered the tavily tools until a restart because
keys were snapshotted at startup.
Changes: quality.HardFacts (lint actually ran + zero errors) replaces the
full ScoreBytes projection in CheckArtifacts - keyword dimensions remain in
ScoreBytes for eval visibility only. New fact-consistency check
(runtime.factCheck, default on): extract percentages/dates/large numbers from
deliverable text, trace each against the material corpus (attachment text +
tool evidence); misses get exactly one correction round (fix, drop, or mark
as target/estimate), never a loop. runVerifier skips same-source model
verification unless runtime.verifierSameModel is set (hard facts + fact check
still run). Search tools now register lazily, reading tavily keys per call
(settings/seeding take effect immediately).
Evidence: SELF - TestHardFactsIgnoresKeywordDimensions,
TestHardFactsKeepsLintUnavailableBlocking, TestUnverifiedFactsAndNotice,
TestFactCheckContinuationOnce (fact note fires, corrected run passes);
go vet + go test ./... pass; live before/after on the golden task set
recorded on the phase05 task card per the policy clause.
Impact: Gate stops rejecting valid documents on stylistic grounds; invented
figures get one forced correction round; same-model verification no longer
pretends to be independent; tavily config is hot. No contract or storage
changes.
Why: Q2 ruled exec tools stay disabled for now but code execution must
return later; the recovery path treats model-generated code as
untrusted. This abstracts the interpreter launch point in
exec_interpreter.go into an ExecBackend port (direct = byte-identical
status quo; sandbox = bwrap isolation) so downstream products can
re-enable execution behind isolation without further base surgery.
Evidence: go build/vet clean; go test ./... 32/32 packages green;
sandbox escape negatives all blocked locally (bwrap 0.11.1): home
secret unreadable (tmpfs /home), sibling session unreadable
(session-only bind), network blocked (--unshare-net), 400-process
spawn storm capped by ulimit -u 256 (fork: retry), 60s sleep killed at
3s timeout, orphans die with the namespace; session-dir write visible
on host; direct backend regression (exit=7 / exit=timeout texts)
unchanged.
Impact: run_shell/run_python route through the backend; workflow
ScriptCommand and the squad kernel ExecConfig stay direct (follow-up
scope); config key execBackend + env AGENT_SYSTEM_EXEC_BACKEND + CLI
--exec-backend; extension.Options.ExecBackend (SurfaceVersion
2.6.0-experimental); DisableExecTools remains the top-priority
fail-closed switch.
Refs: proposal 2026-10-09-ext-g1-b1-b3-exec-backend.md section 4
(EXT-16, G2-8 prototype); Backend #110 (DisableExecTools).
Why: The DOCX parser only read w:r runs that were direct children of w:p, so
text inside w:ins, w:hyperlink and w:fldSimple was dropped with no warning,
violating the no-silent-degradation rule. Separately, runtime.exec is a
user-writable setting, so multi-user deployments could not reliably turn the
exec primitives off.
Evidence: SELF - red tests TestParseDocxTrackedChangesAndHyperlinks and
TestExecHardDisabledOverridesSettings failed before and pass after; backend
go test ./... (32 packages) passes; a mock-mode binary started with
--disable-exec-tools hides run_shell/run_python/register_artifact from
/api/tools even after the user sets runtime.exec=true.
Impact: DOCX paragraphs are decoded from the token stream at any depth;
deleted/moved-from text, field codes and mc:Fallback are excluded and
revised documents carry an explicit warning. New Options.DisableExecTools
(SurfaceVersion 2.5.0-experimental), CLI --disable-exec-tools, config key
disableExecTools and env AGENT_SYSTEM_DISABLE_EXEC_TOOLS; zero value keeps
current behavior.
Refs: Agent-System-RDM docs/proposals/2026-10-09-V22-review-and-implementation-plan.md (G0-7, B-5)
Why: the library now pins eligibility, guards symlinks, reports partial
restores, and bounds the log (Agent-Native#33); the product wires
retention after successful rewinds and needed the HTTP-layer and
eligibility evidence.
Evidence: go test -count=1 ./... green; TestRewindFilesHTTPHappy and
TestRewindFilesHTTPBusy409 (HTTP happy + conflict),
TestBackendCheckpointEligibilityIsWriteFileOnly (the product surface
checkpoints exactly {fs_write_file}).
Impact: checkpoints.jsonl is compacted to the latest 200 anchors after
each successful rewind.
Refs: Agent-Native#33, Agent-Squad#37, N3 plan WP3
Why: workflow scripts and run_tool_script should share one Python host, while a slow MCP server must not block the first prompt.
Evidence: GOWORK=off go test ./internal/runtime/ ./internal/app/ ./internal/settings/ ./internal/tools/ ./internal/runmgr/
Impact: runtime.codeOrchestration defaults to false, so the tool stays hidden. Workflow script confirmation and its longer timeout cap stay as they were.
Refs: feature/code-orchestration
Co-authored-by: Cursor <cursoragent@cursor.com>
Why: workspace writes had no checkpoint, and a 5xx before any output stopped the turn instead of trying the next configured model.
Evidence: GOWORK=off go test ./internal/runtime/ ./internal/runmgr/ ./internal/tools/ ./internal/settings/ ./internal/server/ ./internal/app/
Impact: models.chat.fallbacks defaults to empty, so existing runs keep their model. Rewind restores only tracked writes and appends a system note. Usage records the model that served.
Refs: feature/durable-rewind
Co-authored-by: Cursor <cursoragent@cursor.com>
Why: Plan mode has to block writes, confirmations dropped the argument preview, and direct chat did not share its stop gate with squad members.
Evidence: GOWORK=off go test ./internal/store/ ./internal/tools/ ./internal/runmgr/ ./internal/runtime/ ./internal/server/ passed. Default mode still runs write_artifact; plan mode rejects it.
Impact: Empty permission mode keeps the profile allow, deny, and confirm lists. High-risk shell commands now ask. There is no bypass-all mode.
Refs: feature/permission-hooks
Co-authored-by: Cursor <cursoragent@cursor.com>
Why: Byte truncation split UTF-8, subagent results could return unbounded text, and a write could overwrite a file that changed after it was read.
Evidence: cutRunes keeps a two-character Chinese prefix; tool selection honors -name; checklist reminders go through the reminder bus.
Impact: Tool output limits count characters. A subagent result over 32000 characters is stored as an artifact and the parent sees the path. fs_write_file rejects an existing file that was not read in this run.
Refs: feature/context-engineering
Co-authored-by: Cursor <cursoragent@cursor.com>
Why: Tool calls stayed invisible until execution finished, and turn, tool, timeout, and output caps stopped long tasks early.
Evidence: go test ./internal/runtime/ ./internal/tools/ ./internal/prompts/ ./internal/squadadapter/ ./internal/runmgr/ ./internal/app/ ./internal/settings/
Impact: tool.call.delta and tool.output reach the client; default profile, subagents, and squad no longer cap turns or tool calls; exec timeout defaults to 600s with no hard ceiling. A profile or subagent maxTurns above 0 still applies.
Refs: feature/stream-display-overhaul
Co-authored-by: Cursor <cursoragent@cursor.com>
Why: Plugin scripts and model-written scripts need the same step runner, without giving the script a way to call interpreters or start another workflow.
Evidence: go test ./... passed. Callbacks, denied tools, the 50-call cap, a 1MB line, and cancel-kills-the-process are covered. A model script does not run until the switch is on and the user confirms.
Impact: runtime.workflowScripts stays off by default. Plugin script bodies are copied into the launch. Squad entry still has no run_workflow.
Refs: feature/agent-harness-p4
Co-authored-by: Cursor <cursoragent@cursor.com>
Why: Layout rules alone do not produce conclusion-first documents, extra views, or infographics, and nothing scored a deliverable before it was returned.
Evidence: go test ./internal/prompts/ ./internal/quality/ ./internal/tools/ ./internal/squadadapter/ ./eval/ -count=1 and the house-style unittest passed in the previous session.
Impact: Every task inherits content rules. Meetings, reports, and comparisons can name a view family. Long documents without a visual summary warn. quality_review and the squad checker share one scorecard.
Refs: deliverable-quality-system A1-A5
Co-authored-by: Cursor <cursoragent@cursor.com>
Why: Generated files kept Word's default theme, US Letter pages,
hand-typed bullets and empty sections. Recipes only showed library
calls, the prompt had no craft standard, and seeded skills could not
be upgraded once written.
Evidence: go test ./... passed. The house-style suite covers three
themes across docx, pptx and xlsx, and rejects a document that repeats
the diagnosed layout failures. lint on the sample minutes exits 1.
Impact: Every run gets a resident quality standard. Office files use
house-style (switchable themes, templates, imitation) and must pass
lint. Skill seeding upgrades untouched builtins and keeps user edits.
Uploads land in inputs/, and the default style is a setting.
Refs: feature/agent-professional-deliverables
Co-authored-by: Cursor <cursoragent@cursor.com>
Why: the full-agent batch (Agent-Native PR #3, Agent-Squad PR #3) made
roles capable of declaring model/tools/skills/bash execution, but the
product composition root had no path to register the kernel backend or
inject search/exec/skills config.
Evidence: go vet ./... clean; go test ./... 19 packages ok. New cases:
buildBackends registers agent-native only when composition-root options
are enabled (default backend stays agent-system, per the parity
progressive-enablement clause) and resolves explicit
Backend:agent-native roles; InterpreterExecConfigForKernel validates
absolute interpreter paths and the env whitelist reused from the
product exec tools.
Impact: deps/agent-native pinned to f9ab2ef and deps/agent-squad to
9a402e4; app.go resolves settings.models.chat, tools.tavily.*,
runtime.exec, and <data>/skills into squad.AgentNativeConfig per run;
squadadapter exposes the AgentNativeOptions hook and buildBackends.
Refs: Dev task feature/full-agent-r1 batch B10.
Why: string writes of .docx/.xlsx/.pdf produce files that will not open,
and generated binary artifacts were skipped by search entirely.
Evidence: go test ./... 19 packages ok. Office extensions are refused
with a pointer to run_python plus register_artifact. A docgen docx is
searchable after WriteArtifact; unparseable bytes are not indexed.
Impact: write_artifact and fs_write_file no longer accept those
extensions. Text artifacts index as before. Merge after the interpreter
PR so the error text names tools that exist.
Refs: task feature/binary-doc-index; supersedes Dev #99 / Backend #27
Co-authored-by: Cursor <cursoragent@cursor.com>
Why: deliverable prompts already tell the model to generate Office files
with run_python and register them, but the executors were never
registered, so calls missed the registry and fell back to text-only
write_artifact.
Evidence: go test ./... — 19 packages ok, including echo/script-retention,
empty-script and runtime.exec gate, process-group timeout, python ZIP
magic, and ExecuteTool backfill of register_artifact after run_python.
Scripts stay under the session .agent/ directory; bash/python3 are
resolved to absolute paths; timeout kills the process group.
Impact: models can execute workspace scripts and register binary
deliverables. runtime.exec=false still fail-closes the whole family.
Office extensions now get real MIME types so binary bytes are not
indexed as text.
Refs: task feature/exec-interpreters
Co-authored-by: Cursor <cursoragent@cursor.com>
Why: two field reports against the live build — deliverable cards never
appeared in chat (only model-written filenames) and the artifacts panel
felt dead. Root cause found in runmgr.ExecuteTool: registry tools that
write artifacts via rc.WriteArtifact never had res.Artifact set (only
the oversized-result path did), so llm.go never emitted artifact_ref
parts for real tool output; the unit suite passed via fakes that set
it directly. Architecturally, document generation moves back to the
agent (user directive): system provides execution primitives and the
delivery bridge, not domain tools like write_docx.
Evidence:
- runmgr: toolCtx tracks artifacts written during a call and
ExecuteTool backfills res.Artifact (new runmgr test proves
write_artifact now surfaces one; negative test: no phantom artifact)
- new tools/register_artifact: bridge from workspace files (binary
docx/xlsx/pptx/pdf/images, <=10MB) to deliverable artifacts;
run_shell/run_python interpreter executors are designed
(script-to-disk + whitelisted bash/python3, sandbox cwd, caps) and
land in a follow-up batch behind the security gate decision
- removed builtin write_docx + its prompt; write_artifact prompt now
text-only with explicit no-binary guidance; new always-on base
prompt base.deliverables teaches the generate->register->summarize
workflow with python-docx/openpyxl/python-pptx recipes;
task.doc-minutes rewritten to the workflow
- final-output contract: inline <think> blocks in streamed content are
split into thinking (process) parts so the final reply stays clean
(unit + runner tests)
- go test ./... 19 packages ok; vet clean
Impact: every tool artifact now lands as an in-chat deliverable card;
mock #doc unchanged as demo fixture; binary office mimes stay out of
the text index; write_docx callers see the tool gone (prompt reroutes
them).
Refs: task feature/agent-native-deliverables-r1
Why: docx deliverables were plain text saved with a .docx name (no
headings/tables/styles), and the artifact_ref part was marked Process
so the deliverable vanished into the collapsed process section — the
user saw only model-written text like "📄 xxx.docx 工件 ID …" instead
of an openable, formatted document delivered in the conversation.
Evidence:
- new internal/docgen: Markdown -> styled OOXML (Title/Heading1-4,
numbered+bullet lists with numbering.xml, GFM tables with shaded
header, quotes, code blocks, bold/italic/strike, real hyperlinks);
zero third-party deps; tests cover zip structure, XML well-formedness,
numbering refs, hyperlink rels, escaping, plain-text fallback
- ingest round-trip test proves generated docx parses back to blocks
- new write_docx builtin tool (markdown in -> styled .docx artifact);
write_artifact prompt now routes Word docs to it; task.doc-minutes
output contract requires write_docx delivery
- llm.go: artifact refs collected as deliverables and appended as
content parts AFTER the final text (pinned tail, not process); mock
#dogfood path generates a real styled docx; runmgr skips string
indexing for binary office mimes
- go test ./... 19 packages ok (incl. new e2e asserting <w:tbl> and
title text inside the artifact served by the API)
Impact: model-written documents open in Word/WPS with real layout and
preview with formatting (mammoth maps Heading1-4 to h1-h4); every tool
artifact now lands as a visible in-chat card; no behavior change for
text artifacts.
Refs: task feature/deliverable-docx-r1
Why: the adjudicated preview plan needs server-side OG fetching for link
cards, and the proposal's reserved-capability items call for the fs tool
family to exist behind a default-off gate (implemented, not offered).
Evidence: GET /api/url-preview fetches and parses og:title/description/
site_name/image/favicon with a 5s timeout, 256KB body cap, text/html
content check, TTL cache, and SSRF guards enforced at dial time (loopback
/private/link-local/multicast/reserved/CGNAT rejected per hop, anti DNS-
rebind; AS_URL_PREVIEW_ALLOW_PRIVATE=1 exists for local testing only,
default fail-closed; errors return sanitized cards, never internal
addresses). fs_read_file/fs_write_file/fs_list_dir register as Hidden
tools sandboxed under <root>/<owner>/<conversationId>: traversal,
absolute paths, backslashes, and symlink escape rejected; 1MB/2MB size
caps; writes register an artifact so workspace files stay visible in the
panel. ToolDesc.Hidden gates both model visibility (runctx.Tools) and
/api/tools listing on runtime.fsTools, and registry.Execute re-checks
fail-closed. ToolExecCtx gains OwnerID/WorkspaceRoot/SettingBool.
go test ./... green (fs negative matrix, SSRF negatives, endpoint
fail-closed, cache).
Impact: link previews work end to end once the frontend wires the
fetcher; fs tools are inert until runtime.fsTools is enabled.
Refs: task feature/file-preview-caps; core 4d29301.
Why: conversation transcripts were the only project channel not
readable across sessions (artifacts/knowledge are already shared),
which caused duplicate parallel work; see proposal
2026-09-30-cross-session-reading.
Evidence: runmgr/crossread_test.go covers the full gate matrix
(project gate on/off x outbound flag x private target x cross-project
x cross-tenant x self-read x archived x limit clamping) against a real
store + settings chain; store tests cover migration v11, flag
persistence and tenant isolation; tools/peers_test.go covers args
validation, transcript rendering and the untrusted-content banner;
go test ./... green.
Impact: setting crossSession.read (default off, schema-driven);
conversation columns allow_cross_read (default true) and cross_private
(default false) with PATCH support; builtin tools list_conversations
and read_conversation gated fail-closed in runmgr toolCtx; nested
agent-system-core pin bumped to fd44179 (Core PR #6).
Refs: Agent-System-Documents proposals/2026-09-30-cross-session-reading.md
(Dev task feature/cross-session-reading).
Why: the user ruled out self-registration (2026-09-30): accounts are
created only by admins who set an initial password; first login (and
first login after any admin reset) must force a password change.
Evidence:
- migration v10 auth_accounts (users password/must_change/is_admin +
sessions table; upgrade test walks v1<->v10 both directions).
- argon2id hash/verify with PHC encoding; store tests cover hash
roundtrip, duplicate-owner conflict, must-change lifecycle, session
keep/except semantics.
- e2e TestAdminAccountFlow: static login page reachable without auth,
wrong password 401, login sets HttpOnly cookie + mustChangePassword,
admin creates user (409 on duplicate), non-admin create 403, forced
change clears flag and old password dies, admin reset revokes sessions
and re-arms forced change, API-key channel untouched, 5-fail backoff
429, logout kills session. Full go test ./... green.
Impact: middleware now accepts Bearer key OR session cookie and exempts
static + /api/auth/login (also fixes the latent case where the login
page itself could 401). Embedded no-auth mode unchanged.
Refs: proposal 2026-09-30-ui-redesign-and-auth.md section 5 (admin-issued
variant, user decision 2026-09-30)
Why: Agent-Native ships a Tavily integration (key-pool rotation,
markdown output, fallback copy) that the backend never registered, so
the catalog's WebSearch prompt had no carrier; wiring the tool gives it
one and gives the agent live-web reach.
Evidence: 4 new tool tests (registration, markdown formatting, empty
query, failure fallback carried inside the error to match registry
replay semantics); go test ./... 16 packages ok; live boot check - no
keys means not registered, two keys log registration with both tools
in /api/tools and 46 library modules in /api/prompts.
Impact: tools.tavily.apiKeys/endpoint settings (restart effect; empty
pool means the tool is absent, no silent degradation); tool.tavily_search
and tool.web_search prompt modules inject citation discipline (cite
sources, keep inference separate, fall back to workspace materials on
failure); library floor raised to 46.
Refs: feature/tavily-search-tool task card; guide disposition table in
Documents.
weichen
