User report: YOLO mode was still routing shell commands through the
WorkspaceWrite sandbox, which intercepted legitimate outside-workspace
writes (package installs, sub-agent workspaces, package-manager state
under ~/.cache, brew, npm install -g, pipx, …) and forced approval
round-trips. That contradicts the YOLO contract — the user opted into
"no guardrails" and instead got a guardrail.
YOLO already auto-approves all tools and enables trust mode. The
sandbox was the last residual restriction. Drop it.
Change in `Engine::build_tool_context`: split the previously-merged
`AppMode::Agent | AppMode::Yolo` arm into two:
* **Agent** keeps `WorkspaceWrite { writable_roots, network_access:
true, … }` — interactive mode with explicit per-tool approval, so
the sandbox plus the approval flow form a defense-in-depth layer.
* **Yolo** uses `DangerFullAccess` — no sandbox. The user has
opted into auto-approval + trust mode + no sandbox as one
consistent posture.
Plan mode unchanged (read-only, no shell tool registered).
Updated `agent_and_yolo_modes_elevate_shell_sandbox_to_allow_network`
to pin the new YOLO contract: `DangerFullAccess` specifically, not
just "has network access."
Verified locally:
* `cargo fmt --all -- --check` clean.
* `cargo clippy --workspace --all-targets --all-features --locked
-- -D warnings` clean.
* `cargo test --workspace --all-features --locked` — green
(the snapshot::repo flake still flakes in batch but passes in
isolation; unrelated).
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Live repro: in a session producing content rapidly (sub-agent
running, multiple tool calls), the user scrolls up to read earlier
output. Their scroll position briefly takes effect, then snaps back
to the live tail when the next stream chunk arrives. Symptom is
"scrolling is broken / takes over instead of the transcript".
Root cause in `crates/tui/src/tui/widgets/mod.rs:188-210`:
* The user's mouse-scroll-up sets `transcript_scroll = at_line(N)`
and `user_scrolled_during_stream = true`.
* During render, `resolve_top` clamps the state against
`max_start = total_lines.saturating_sub(visible_lines)`. If
`max_start < N` (transcript shrunk between scrolls and render —
e.g., a sub-agent in-progress card collapsed into a smaller
finished card, or the content briefly fits in one screen),
`resolve_top` returns `Self::to_bottom()` (TAIL_SENTINEL).
* `is_at_tail()` on the post-resolve state returns `true`.
* The auto-clear at line 208 fires →
`user_scrolled_during_stream = false`.
* Next `add_message` / sub-agent envelope sees `is_at_tail() &&
!user_scrolled_during_stream` and calls `scroll_to_bottom()`. The
user is yanked off their position mid-read.
`scrolled_by` has the same trapdoor: when `total_lines <= visible_
lines` it returns `to_bottom()` regardless of scroll direction
(line 145-148 in scrolling.rs). A user scroll-up while content
fits in one screen produces `to_bottom()` → `is_at_tail()` true →
auto-clear → next chunk yanks.
The fix
=======
Snapshot whether the user's PRIOR state was deliberately tail
(`is_at_tail()` BEFORE `resolve_top`), and only clear the lock
when:
1. Prior state was already TAIL_SENTINEL (deliberate, set by
`scrolled_by` reaching `max_start` while scrolling DOWN, or by
`scroll_to_bottom()`).
2. AND `total_lines > visible_lines` (so "tail" is meaningful —
if the whole transcript fits, "is_at_tail" is trivially true
and clearing the lock would yank the user back to bottom on
the next chunk despite their explicit scroll-up).
This preserves all the legitimate clear paths:
* `TurnComplete` event clears the lock at the per-turn boundary
(`ui.rs:879`).
* User invokes `scroll_to_bottom()` explicitly via key/menu
(`app.rs:2459`).
* User scrolls down enough that `scrolled_by` reaches `max_start`
in a transcript with real scroll room — state goes through
`to_bottom()` BEFORE resolve, so `was_explicit_tail = true` and
the lock clears.
What it stops:
* Render-time resolve clamping `at_line(N)` to tail when content
shrunk doesn't quietly revoke the user's intent.
* `scrolled_by` collapsing a scroll-up to `to_bottom()` when
content briefly fits in one screen no longer triggers the
auto-clear (the prior state wasn't tail).
Verified locally:
* `cargo fmt --all -- --check` clean.
* `cargo clippy --workspace --all-targets --all-features --locked
-- -D warnings` clean.
* `cargo test --workspace --all-features --locked` — 2038 passed,
2 ignored, 0 failed (a snapshot::repo flake unrelated to scroll;
passes in isolation).
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
User-facing repro:
* In YOLO mode at low context utilisation (~5%), the engine briefly
showed `resetting plan` in the footer and the transcript area went
mostly black. Tools kept running (Plan panel + sidebar still
rendered), but the chat history above the latest turn was gone.
Root cause: the capacity controller's `VerifyAndReplan` action
(`crates/tui/src/core/engine/capacity_flow.rs::apply_verify_and_replan`)
runs `self.session.messages.clear()` and rebuilds from the canonical
state. The capacity controller fires this when its slack-based
`p_fail` calculation crosses the high-severe band — independently of
the `auto_compact` setting, independently of token utilisation.
The user opted out of auto-compaction in v0.8.11 (default
`auto_compact = false`, #665), explicitly trusting the model with
the full 1M-token V4 window. Auto-managing the prefix on their
behalf via the capacity controller contradicts that posture and
silently destroys both the user-visible transcript and V4's prefix
cache.
The fix
=======
Flip `CapacityControllerConfig::default().enabled` from `true` to
`false`. The controller's `observe_*` and `decide` methods already
short-circuit when `enabled` is false (`capacity.rs:255`,
`capacity.rs:396`), so the existing wiring becomes a no-op for the
default config — no need for defensive gating in
`capacity_flow.rs`.
Power users who want the controller can opt in via
`capacity.enabled = true` in `~/.deepseek/config.toml`. The slack
heuristics, model priors, cooldowns, and intervention paths all
remain in the codebase, ready to re-engage on opt-in. Nothing
deleted.
Tests
=====
* `default_controller_is_disabled_and_skips_observations` — pins
the new default; `observe_pre_turn` returns `None`.
* `opt_in_controller_observes_and_decides` — confirms `enabled =
true` rearms the controller end-to-end.
* `app_config_without_capacity_uses_default_disabled` — pins that
loading a config with no `[capacity]` section produces
`enabled = false`.
* `capacity_disabled_by_default_keeps_messages_intact` — direct
regression for the user-reported symptom: with default config,
even a forced error-escalation checkpoint cannot trigger
`messages.clear()`. Asserts the transcript length is preserved.
Verified locally:
* `cargo fmt --all -- --check` clean.
* `cargo clippy --workspace --all-targets --all-features --locked
-- -D warnings` clean.
* `cargo test --workspace --all-features --locked` — 2039 passed,
2 ignored, 0 failed (one flake on `snapshot::repo::tests::
restore_removes_files_added_after_target_snapshot` was filesystem-
timing-dependent, passes on isolation re-run; unrelated to this
change).
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Final step in the v0.8.11 patch release. Bumps the workspace
`Cargo.toml`, all 9 internal path-dep version pins, and
`npm/deepseek-tui/package.json` to **0.8.11**. `Cargo.lock`
regenerated alongside.
The v0.8.11 CHANGELOG entry already landed on `main` via the
cache-maxing overhaul PR (#684). This commit only stamps the
version. Together they ship:
* **Cache-maxing for V4 1M context** — engine no longer rebuilds the
system prompt on every turn (#684's `Session::last_system_prompt_hash`),
the volatile working-set summary moved out of the system prompt
into per-turn `<turn_meta>` on the latest user message, the tool
array is anchored with `cache_control: ephemeral`, and the
`messages_with_turn_metadata` injection skips tool-result
messages so the assistant→tool_result invariant stays intact.
* **500K compaction floor** — automatic compaction refuses below
500K tokens via `MINIMUM_AUTO_COMPACTION_TOKENS`. Manual
`/compact` bypasses (explicit user agency).
* **Token-only compaction trigger** — dropped
`CompactionConfig::message_threshold` and the message-count
branch in `should_compact`; that 128K-era heuristic only fired
on long sessions of small messages, exactly the case where
rewriting the V4 prefix cache is most wasteful.
* **Legacy 128K naming** — `DEFAULT_CONTEXT_WINDOW_TOKENS` →
`LEGACY_DEEPSEEK_CONTEXT_WINDOW_TOKENS`.
* **`npm install` resilience** — `install.js` now retries with
exponential backoff, enforces per-attempt timeout + 30 s stall
detector, honors `HTTPS_PROXY` / `HTTP_PROXY` / `NO_PROXY` (pure
Node, no new dependencies), and prints download progress to
stderr. Driven by a community report that `npm install` took 18
minutes through a CN npm mirror; the GitHub Releases binary
fetch was the bottleneck and CN mirrors don't proxy GitHub.
Verified locally:
* cargo fmt --all -- --check ✓
* cargo clippy --workspace --all-targets --all-features
--locked -- -D warnings ✓
* cargo test --workspace --all-features --locked ✓
* parity gates (snapshot, parity_protocol, parity_state) ✓
* bash scripts/release/check-versions.sh ✓
(workspace=0.8.11, npm=0.8.11, lockfile in sync)
* node scripts/release/npm-wrapper-smoke.js ✓
Reminder for the maintainer at release time: the npm publish is
manual and requires 2FA OTP on every publish. After this PR
merges and the GitHub Release is fully drafted by `auto-tag.yml`,
publish from a developer machine:
cd npm/deepseek-tui
npm publish --access public
The `prepublishOnly` hook checks all eight binaries plus the
SHA256 manifest are present on the GitHub Release before letting
`npm publish` proceed, so this must happen *after* the GitHub
Release is finalized.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
CI on PR #684 caught two real issues that local checks missed:
**Lint failure (cargo fmt).** A regression test landed with a multi-line
`let ContentBlock::Text { text, .. } = real_user.content...` pattern
that local rustfmt accepted but CI's pinned toolchain collapsed onto a
single line. Reformatted to match.
**npm wrapper smoke failure ("Checksum manifest is missing
deepseek-<platform>").** Subtle Node.js streams interaction in
`install.js` introduced by the network-resilience cluster:
* `httpRequest` attaches a `data` event listener on the response to
re-arm the stall timer.
* Attaching a `data` listener on a `Readable` puts the stream into
flowing mode immediately.
* `downloadText` then ran `for await (const chunk of response)` to
collect the body — the async iterator expects paused-mode and
silently misses chunks that flow before / between iteration ticks.
* For small bodies (the ~100-byte SHA256 manifest), the entire
response could flow through the stall listener before the async
iterator's `read()` calls landed, leaving the joined body empty.
* Result: `parseChecksumManifest("")` returned an empty Map →
`verifyChecksum` saw no entries → "manifest is missing X" after
the actual binary download succeeded.
Binary downloads were unaffected because `download()` uses
`response.pipe(sink)` plus a `data` listener for progress — both
consume chunks via `data` events, no async iterator involved.
Fix: collect the response body in `downloadText` via direct `data`/
`end` event subscription. `data` listeners stack — both the stall
re-arm and the body collector fire on every chunk, no flowing-vs-
paused conflict. Stall detection still works.
Verified locally: `node scripts/release/npm-wrapper-smoke.js`
"npm wrapper smoke passed with local assets from <url>".
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Live-test repro: typing a single user message in the TUI triggered a
tool call (read_file Cargo.toml), and the *next* request to DeepSeek's
API returned HTTP 400:
"An assistant message with 'tool_calls' must be followed by tool
messages responding to each 'tool_call_id'. (insufficient tool
messages following tool_calls message)"
Root cause: `messages_with_turn_metadata` walked the message list from
the tail and prepended a `<turn_meta>` Text block to the *last* message
with role="user". But tool-result messages also use role="user"
internally (they serialize to role="tool" on the wire). Inserting a
Text content block at index 0 of a tool-result message changed the
shape from `[ToolResult(...)]` to `[Text("turn_meta..."), ToolResult(...)]`,
which on the wire becomes a role="user" message with text instead of
the role="tool" message the API needs to satisfy the assistant's
prior tool_call. Hence the 400.
The fix:
* Restrict the injection target to messages that have at least one
Text content block AND no ToolResult blocks. This identifies actual
user-typed messages and skips tool-result envelopes.
* When the trailing slice has no eligible user message (e.g. mid-turn
when a tool result is the most recent message), skip injection
entirely. The working_set will surface again on the next genuine
user prompt; we don't retroactively prepend onto an earlier user
message because that would also confuse the API's tool-call
continuity checks.
Two regression tests pin the contract:
* `turn_metadata_skips_tool_result_messages` — assistant tool_call +
tool_result + earlier user message: only the user message gets the
prefix, the tool_result message stays a single-block ToolResult.
* `turn_metadata_skips_when_only_tool_results_trail` — the corner
case where the trailing user-role message is solely a tool result
(no real user message in the slice): no injection happens, the
message returns unchanged.
Verified locally:
* 2038 tests passed in TUI bin (2 ignored, was 2036 — these are the
+2 new regressions).
* `cargo fmt`, `cargo clippy --locked -D warnings`, parity gates all
clean.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
The `message_threshold` field on `CompactionConfig` was a 128K-era
heuristic that fired compaction on long sessions of small messages
— exactly the case where rewriting V4's prefix cache is most wasteful.
Token budget is the only signal that maps to actual model context
pressure; counting messages adds nothing.
Changes:
* Remove `CompactionConfig::message_threshold` field.
* Remove the message-count branch in `should_compact` — token
threshold + 500K floor is now the sole compaction trigger.
* Remove `compaction_message_threshold_for_model`,
`DEFAULT_COMPACTION_MESSAGE_THRESHOLD`,
`COMPACTION_MESSAGE_DIVISOR`, `MAX_COMPACTION_MESSAGE_THRESHOLD`
from `models.rs`.
* Drop the `forced_config.message_threshold` tweak in the engine's
capacity-guardrail forced-compaction path; that path now also
bypasses the floor (`auto_floor_tokens = 0`) because we're at a
hard ceiling and have to free budget regardless of cache cost.
* Update production constructors (`main.rs`, `runtime_threads.rs`,
`app.rs::compaction_config`) to drop the field.
* Update tests: keep the floor + token-threshold assertions, delete
the two tests that specifically validated message-count
triggering, replace `should_compact_respects_message_threshold`
with `message_count_no_longer_triggers_compaction` pinning the
new contract.
Verified locally:
* `cargo fmt --all -- --check` clean.
* `cargo clippy --workspace --all-targets --all-features --locked
-- -D warnings` clean.
* `cargo test --workspace --all-features --locked` — 2036 passed
in TUI bin (2 ignored), all other crates green.
* parity gates: snapshot, parity_protocol, parity_state — all pass.
* `git diff --exit-code -- Cargo.lock` — clean.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Maintainer-reported (handoff): after spawning a sub-agent in YOLO, the
transcript renders solid black and scroll keys go dead, but footer +
sidebar still render fine. The shape (black + dead input together)
strongly suggests a `View` is on the stack that returns empty layout
AND intercepts key events at the top level. The fix wants a tighter
repro than we have today.
Add `tracing::debug!` to every push / push_boxed / pop on `ViewStack`
and to the implicit pops in `apply_action` (Close + EmitAndClose).
Each line carries the `ModalKind` and post-action depth, so a future
`RUST_LOG=deepseek_tui::view_stack=debug` capture will show exactly
which view stayed pushed when the symptom recurred.
No behavior change. The handoff explicitly suggested this as the
first-look diagnostic step; we ship the diagnostic now so the next
report comes with evidence.
Refs the unresolved sub-agent black-transcript symptom captured in
session-3 handoff. Will surface to a tracking issue once we have a
concrete repro from the maintainer.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Two coordinated changes that stop the engine from routinely rewriting the
prompt prefix and burning DeepSeek V4's prefix-cache discount:
1. `Settings::default().auto_compact` flips from `true` to `false`. The
`auto_compact = on` opt-in and the explicit `/compact` slash command
stay available for users / agents that decide their workload benefits
from compaction more than from cache stability. With V4's 1M-token
window the user has plenty of headroom to run long sessions without
auto-trimming, and aggressive compaction has been the dominant
cost-spike vector in long sessions (the rewritten prefix invalidates
~90% of the cache discount on every compaction event).
2. `DEFAULT_COMPACTION_TOKEN_THRESHOLD` raised from `50_000` to
`102_400` (80% of `DEFAULT_CONTEXT_WINDOW_TOKENS = 128_000`). This is
the last-resort threshold used when `context_window_for_model` returns
`None` — i.e. an unrecognised model id. Pre-v0.8.11 the fallback
compacted at ~5% of a V4 window when model detection silently fell
through. Now the fallback inherits the same late-trigger discipline as
the V4 path, so model-detection drift doesn't quietly burn cache.
Together: the two changes mean compaction never fires automatically by
default, and even when explicitly opted in (or when the runtime-thread /
capacity-flow paths invoke compaction with their own `enabled = true`
config), the threshold is anchored at 80% of the model's context window
(or 80% of the 128K default if the model is unknown), never below.
Tests
=====
- `default_settings_disable_auto_compact_to_protect_v4_prefix_cache` —
pins the new default and explains the rationale inline.
- `auto_compact_remains_explicitly_configurable` — unchanged; still
asserts the `set("auto_compact", "on" | "off")` round-trip works.
- `compaction_threshold_scales_with_context_window` — updated to assert
`compaction_threshold_for_model("unknown-model") == 102_400`.
- `v4_soft_caps_only_apply_to_v4_models` — updated to assert the
unknown-model + reasoning-effort path also lands on the new floor.
Verification
============
- `cargo fmt --all -- --check` clean.
- `cargo clippy -p deepseek-tui --bin deepseek-tui --all-features
--locked -- -D warnings` clean.
- `cargo test -p deepseek-tui --bin deepseek-tui --locked` →
2028 passed, 2 ignored.
Refs #664 (handoff-instead-of-compact pattern, full implementation
deferred). Behaviour-only change for v0.8.11; the larger
agent-aware-handoff mechanism is its own design surface.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Three sources share the /foo namespace with clear precedence:
1. Native built-ins (match block in mod.rs)
2. User-config commands (~/.deepseek/commands/*.md) — checked first
3. Skills (~/.deepseek/skills/) — new fallback in the _ arm
Template substitution: $1, $2, $ARGUMENTS are replaced in user-command
and skill content before the message is sent. Existing exact-match and
alias behavior is unchanged.
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
Add resident_file parameter to agent_spawn. When set, the child's prompt
is prefixed with the full file contents under a stable <!-- resident_file -->
comment block so DeepSeek's prefix cache stays warm across follow-up
send_input calls. File ownership is tracked via a process-scoped OnceLock
table; a second spawn on the same path receives a resident_conflict warning.
- input_schema: new resident_file string property with description
- SpawnRequest: resident_file: Option<String> field
- execute(): reads file, builds byte-stable prefix, checks ownership table
- Backward compatible: absent resident_file follows existing code path
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
Tool outputs (read_file, grep_files, exec_shell, fetch_url, web_search) that
exceed a configurable token threshold are now intercepted before they reach
the parent context. A structured synthesis header replaces the raw blob; the
full content is stored in the workshop variable `last_tool_result` for later
`promote_to_context` retrieval.
Key changes:
- New `crates/tui/src/tools/large_output_router.rs`: `LargeOutputRouter`,
`WorkshopConfig`, `WorkshopVariables`, `RouteDecision`, token estimator,
synthesis-prompt builder, and wrap_synthesis helper. Full unit-test suite.
- `ToolContext` gains `large_output_router` and `workshop_vars` fields plus
the `with_large_output_router` builder; constructor defaults are `None` so
sub-agents and test contexts are unaffected.
- `ToolRegistry::execute_full_with_context` applies routing after every tool
call; `raw=true` in the tool input bypasses routing for that invocation.
- `EngineConfig` gains a `workshop` field; `Engine::new` creates the shared
`WorkshopVariables` Arc when the field is present and wires it into every
`build_tool_context` call.
- `Config` gains `[workshop]` table deserialization; `merge_config` propagates
it like other optional tables.
- `config.example.toml` documents `[workshop]`, `large_output_threshold_tokens`
(default 4096), and per-tool threshold overrides.
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
Add `TuiPrefs` struct and `KeybindPrefs` sub-struct to `settings.rs`
to decouple TUI-specific preferences (theme, font_size, keybinds) from
the agent/project `config.toml` so they survive project switches.
- `TuiPrefs::path()` resolves to `~/.deepseek/tui.toml`; honours
`DEEPSEEK_CONFIG_PATH` env-var redirect used by tests and CI.
- `TuiPrefs::load()` falls back to struct defaults when the file is
absent — no error, backwards-compatible.
- `TuiPrefs::save()` creates `~/.deepseek/` if necessary.
- `TuiPrefs::validate()` normalises theme case and rejects unknown
values with a helpful message.
- 8 new unit tests cover defaults, round-trip TOML serde, validation,
absent-file fallback, and the save→load cycle; all pass.
- `cargo check` clean (0 errors, 0 new warnings beyond expected
dead-code on the new public API).
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
After every successful apply_patch / write_file / edit_file the engine
queries LspManager for diagnostics on touched files and appends a
<diagnostics file="…"> block to the tool result.
- ToolContext gains optional lsp_manager field + with_lsp_manager() builder
- lsp_diagnostics_for_paths() helper handles async fan-out and rendering
- All three edit tools append the block when non-empty
- Gated behind existing [lsp] enabled config flag; silent fallback
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
Add `sync_registry` to `skills/install.rs` that pulls `index.json` from the
configured `[skills] registry_url`, resolves each entry to a download URL,
checks ETag + SHA-256 for freshness, and writes SKILL.md (or the full
unpacked tarball) into `~/.deepseek/cache/skills/<name>/`. A `.cache-meta.json`
sidecar records the ETag and hash so subsequent syncs skip unchanged skills in
one round-trip.
Wire the new `/skills sync` slash-command in `commands/skills.rs` (dispatched
from `list_skills`) and update the `COMMANDS` usage string in `mod.rs` to
`/skills [--remote|sync]`. The per-skill failure model is non-fatal: the
command prints a per-entry `[+]`/`[=]`/`[!]` summary and returns a final
tally. `default_cache_skills_dir` and the new outcome types are re-exported
from `skills/mod.rs` for downstream consumers.
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
Hunter Bown