mirror of
https://github.com/QwenLM/qwen-code.git
synced 2026-07-10 01:29:17 +00:00
refactor(core)!: replace tail-preservation compaction with summary + restoration attachments (#4599)
* refactor(core): rewrite compression prompt to 9-section claude-code-style format
Replaces the <state_snapshot> XML template with a numbered 9-section
structure that mandates verbatim preservation of user messages, including
the historical chronological list (section 6). The new format is
designed to pair with post-compact file/image restoration (separate work)
so the agent can resume long single-turn tasks without losing intent.
* refactor(core): align compaction trigger string with new 9-section prompt
The user-turn trigger injected after the system prompt still said
'generate the <state_snapshot>' from the old XML prompt era. Updated to
'produce the 9-section summary' to match Task 1's new prompt format.
Also tightens the prompt test to assert the specific user-message
verbatim mandate (not just the word 'verbatim' anywhere) so a future
regression that drops the mandate won't silently pass.
* feat(core): add postCompactAttachments module with file path extractor
extractRecentFilePaths walks history newest-first and returns the top N
unique file paths touched by read_file/write_file/edit/replace tool calls.
Pure function, no side effects, no state cache — readiness for the next
compaction-rewrite tasks.
* refactor(core): simplify extractRecentFilePaths internals
Three small cleanups from code review:
- Map<string, number> -> Set<string> (the index value was never read)
- Guard against maxFiles <= 0 explicitly (avoids returning 1 result
when caller passes 0 as a 'disable' sentinel)
- Document 'replace' as a legacy alias for 'edit' so a future cleanup
pass does not delete it as apparent dead code
Adds one test covering the maxFiles=0 path.
* feat(core): add image extractor with source-tool metadata
extractRecentImages walks history newest-first, collects up to N image
inlineData parts, and attributes each one to the model+functionCall that
preceded it (when one exists). Returns chronological order so callers
can render a meaningful 'last visual state ends here' strip.
* feat(core): add size-adaptive file reader for post-compact restore
readFileSizeAdaptive reads a file and returns one of: embed (full content
for files ≤ maxTokens × 4 chars), reference (path-only for large files),
missing (deleted since last touch), or binary (non-text content). The
embed/reference distinction mirrors claude-code's compact_file_reference
vs file attachment behavior, but without introducing new message types.
* refactor(core): harden readFileSizeAdaptive size accounting
Three corrections from code review:
- Import CHARS_PER_TOKEN from tokenEstimation.ts (canonical) instead of
redeclaring locally, preventing silent drift between modules.
- Compare decoded character length, not raw byte length, against the
cap. Otherwise a 10k-char Chinese file would be ~30k bytes and would
be mis-classified as 'reference' despite fitting the budget.
- Rename FileReadResult -> FileEmbedResult to avoid a name collision
with the unrelated FileReadResult interface in fileUtils.ts.
Adds a CJK-text test that catches the byte/char regression.
* feat(core): add file restoration block composer
buildFileRestorationBlocks reads each candidate file, classifies it as
embed/reference/missing/binary, and emits one consolidated reference
block (path-only list) plus one user message per embedded small file.
Total embed size is capped at POST_COMPACT_TOKEN_BUDGET; over-budget
files downgrade to reference.
* test(core): make budget test actually exercise the downgrade path
The previous version of this test wrote 3 files totalling 9k chars
against a 200k char budget. The assertions trivially passed regardless
of whether the budget check existed in the implementation.
The new version writes 11 files of 20k chars (each at the per-file cap)
so the budget is exhausted by the 10th and the 11th must downgrade
from embed to reference. Asserts both: file 11 appears in the reference
block, and file 11's content does NOT appear in any embed block.
* feat(core): add image restoration block composer
buildImageRestorationBlock emits a single user message whose first part
is a metadata header (turn index + source tool name + args per image),
followed by the inlineData parts themselves. Handles user-paste images
(no source tool) by labeling them as 'user-provided'.
* feat(core): add composePostCompactHistory orchestrator
Assembles the full post-compact history in order:
summary → model ack → file references → file embeds → image block.
Each section is built by the per-concern extractors and builders added
in previous tasks. This is the single integration point that
chatCompressionService.compress() will call once the wire-up task lands.
* feat(core)!: rewrite compress() to claude-code-style full-history model
Replaces the split-point + tail-preservation model with full-history
compression + composePostCompactHistory. The entire curated history is
sent to the summary side-query, and the post-compact history is
assembled by the new composer (summary + ack + file restores + image
restore).
BREAKING: the previously-exported findCompressSplitPoint,
splitPointRetainingTrailingPairs, COMPRESSION_PRESERVE_THRESHOLD, and
TOOL_ROUND_RETAIN_COUNT will be removed in the next commit. Tests that
exercise them remain failing temporarily.
* chore(core): remove obsolete split-point compression infrastructure
Deletes findCompressSplitPoint, splitPointRetainingTrailingPairs,
COMPRESSION_PRESERVE_THRESHOLD, MIN_COMPRESSION_FRACTION, and
TOOL_ROUND_RETAIN_COUNT, plus the tests that exercised them. The new
behavior is covered by composePostCompactHistory and its unit tests.
Also cleans up:
- Stale orphan-strip comment in compress() that described the deleted
manual-trigger orphan-funcCall handling.
- TEST_ONLY.COMPRESSION_PRESERVE_THRESHOLD hatch in client.ts.
- Docstring references in config.ts and compactionInputSlimming.ts.
* test(core): add single-turn computer-use compaction regression
Reproduces the scenario the rewrite targets: one user prompt kicks off
many screenshot tool calls. Asserts that (a) the user prompt is carried
into the summary verbatim and (b) the 3 most recent screenshots are
restored as an image block with source-tool metadata. This is the canary
test for the computer-use UX claim made in the design discussion.
* docs(core): remove stale "split point" references in tokenEstimation comments
Aligns the docstrings with the new compose-based compression flow. The
"split point" and "splitter" concepts no longer exist after the rewrite.
* fix(core): iterate parts reverse so parallel tool calls keep the last N
Real-session E2E surfaced a bug: a model that issues N parallel ReadFile
calls puts all N functionCall parts in ONE model+fc content. The
extractor's outer history walk is newest-first, but the inner parts
walk was forward — so for a 6-parallel batch hitting the cap of 5,
the FIRST 5 parts won and the actually-most-recent (last-listed) file
was dropped.
Fix: walk parts in reverse within each content. Applied symmetrically
to extractRecentImages (same shape, even rarer trigger).
Adds a regression test that hits a 6-parallel batch.
* fix(core): code-review fixes — fence escape, path sanitize, alias removal
- CommonMark-safe fence in file embed blocks. The old 3-backtick fence
closed prematurely when a file's content contained a triple-backtick
run (Markdown, CLAUDE.md, JSDoc with code examples) — leaking the
remainder as unfenced text. Now uses a fence one longer than the
longest backtick run in the content.
- Strip control characters (\r, \n, \t) from file paths before
rendering into attachment markdown. Paths come from model-controlled
history; a \n could inject markdown structure. The actual path stays
intact for tool calls — only the displayed string is sanitized.
- Remove the historyForCompression alias for curatedHistory in
compress(). The alias was added as a comment anchor during the
rewrite but didn't carry semantic information.
* refactor(core): rewrite compression prompt to <state_snapshot> XML with 9 claude-aligned sections
Replaces the 9-section numbered-text prompt with qwen-code's original
<state_snapshot> XML envelope, but with the 9 inner section tags
content-aligned to claude-code:
<primary_request_and_intent>
<key_technical_concepts>
<files_and_code_sections>
<errors_and_fixes>
<problem_solving>
<all_user_messages>
<pending_tasks>
<current_work>
<next_step>
Also:
- <scratchpad> -> <analysis>, stripped by postProcessSummary (saves
~600-800 tokens of CoT noise per compaction).
- "Resume directly..." trailer moved out of the prompt body and into
postProcessSummary (no longer re-generated by the model every
compaction; lives once in code with our own wording).
- Section 6 verbatim-policed mandate relaxed to "chronological, include
short messages like 'ok' / 'continue'" — matches claude-code intent
without forcing the model to literally copy long user messages.
E2E (qwen3.6-plus, 6 substantial .ts files + thorough analysis):
raw history 6508 -> summary 1513 (after strip ~947), 38% history
compression. Overall context 24642 -> 20647 reported (-16%), with
another ~664 tokens actually saved by the post-strip but not
reflected in the conservative token-math heuristic.
* docs(core): code-review polish on XML prompt rewrite
Four small follow-ups from review of 641a0eadd:
- prompts.ts: rewrite getCompressionPrompt's stale JSDoc — it still
described the deleted 9-section numbered-text format and the
verbatim mandate that was relaxed.
- chatCompressionService.ts: clarify the token-math comment so it's
obvious the ~1000 token deduction covers the full compression
system prompt + kick-off user turn (not any single instruction)
and that newTokenCount slightly over-counts because <analysis>
gets stripped by postProcessSummary downstream.
- postCompactAttachments.ts: add a NOTE comment on the <analysis>
strip regex covering its strict-tag-match assumption and
multi-block / non-greedy semantics.
- postCompactAttachments.test.ts: replace the four lazy
`await import('./postCompactAttachments.js')` calls inside the
postProcessSummary describe block with one top-level static import
— consistent with how every other describe in the file imports.
* docs(core): drop stale duplicate sentence left in token-math comment
* fix(core): address wenshao review on PR #4599 (correctness + security + ergonomics)
Seven follow-ups from wenshao's review of the compaction rewrite.
Critical:
- newTokenCount now includes restoration-block tokens via
estimateContentChars over extraHistory[2..]. Previously the formula
only counted side-query output, so up to 5 × 5K (files) + 3 × image
tokens were missing — letting the inflation guard miss and the
cheap-gate under-estimate the next prompt size (Finding 1).
- composePostCompactHistory now merges every file restoration block
and the image block into a single user Content following the model
ack. The previous output had consecutive user roles, which
geminiChat.test.ts:6289 enforces against and Gemini providers
reject with 400 "consecutive same-role content" (Finding 2).
- Preserve a trailing model+functionCall through compaction so a
pending functionResponse (sitting in sendMessageStream's
pendingUserMessage) has a matching call. Without this, hard-rescue
auto-compaction mid tool-use loop produces a user+functionResponse
with no preceding model+functionCall → API 400. This restores the
protection the split-point in-flight fallback used to provide.
When the funcCall lands without attachments it folds into the
ack's own model Content to avoid model→model adjacency (Finding 3).
- composePostCompactHistory now takes an optional workspaceRoot and
silently skips file paths that resolve outside it.
extractRecentFilePaths picks up paths from model functionCall args
regardless of whether the tool execution succeeded; without a
boundary check, an adversarial model that issued
read_file('/etc/passwd') — denied by the permission system —
would still have its path extracted and re-read into the next
prompt. compress() passes config.getTargetDir() as the boundary
(Finding 4).
Suggestions:
- composePostCompactHistory + buildFileRestorationBlocks +
readFileSizeAdaptive all take optional AbortSignal and short-
circuit / pass it to readFile's { signal } option. Cancelled
compactions stop on the next file read (Finding 5).
- postProcessSummary fallback no longer re-injects the raw
<analysis> block when the strip leaves nothing. The new
stripAnalysisBlock helper runs the closed-tag strip AND an
unclosed-tag strip (handles 'model ran out of output tokens
before closing'). If both leave nothing, postProcessSummary
emits '[Summary unavailable]' rather than leaking scratchpad
(Finding 6).
- firePostCompactEvent now receives stripAnalysisBlock(summary) so
hook consumers see the same text that lands in history. The
resume trailer stays out of the hook payload — that's wrapper
decoration for the next agent turn, not state for consumers
(Finding 8a).
Docs:
- Update the geminiChat.ts comment around `trigger: 'auto'` to
describe what the trigger actually does post-refactor (hook event
categorization) rather than the deleted manual-only orphan-strip
it used to guard against (Finding 8b).
Regression tests cover all six fixable code-path changes
(role alternation, trailing funcCall preservation, workspace
boundary, abort propagation, closed-tag fallback strip, unclosed-tag
fallback strip).
* fix(core): add getTargetDir to geminiChat auto-compression test mock
The R3.4 end-to-end auto-compression test drives the real
ChatCompressionService, which reads config.getTargetDir() for the
post-compact file-restoration workspace boundary. The geminiChat mock
config lacked getTargetDir, so the test threw "config.getTargetDir is
not a function" on CI. Add the mock to unblock the failing Test jobs.
* feat(core): configurable compaction retention + computer-use screenshot trigger
Add four env-overridable chatCompression settings (priority env >
settings > default):
- maxRecentFilesToRetain (QWEN_COMPACT_MAX_RECENT_FILES, default 5)
- maxRecentImagesToRetain (QWEN_COMPACT_MAX_RECENT_IMAGES, default 3)
- enableScreenshotTrigger (QWEN_COMPACT_SCREENSHOT_TRIGGER, default true)
- screenshotTriggerThreshold(QWEN_COMPACT_SCREENSHOT_THRESHOLD, default 50)
The screenshot trigger fires auto-compaction once tool-returned images
accumulate to the threshold even when token usage is below the auto tier,
so computer-use sessions don't drown the model in stale screenshots. It
counts only images nested in functionResponse.parts (tool results), not
user pastes, and runs only in the would-be-NOOP path when enabled.
Fix a latent bug surfaced while wiring the trigger: extractRecentImages
only inspected top-level inlineData parts, but convertToFunctionResponse
nests tool media under functionResponse.parts — so post-compact
restoration recovered ZERO tool screenshots in real sessions, while unit
tests stayed green against a fabricated top-level shape. It now walks both
shapes; the image counter and tests use the real nested shape.
Remove the now-defunct contextPercentageThreshold deprecation warning (the
field was already dropped from ChatCompressionSettings) and its tests, and
document the four new settings.
* test(core): assert screenshot trigger can't re-fire post-compaction; fix misleading docs
Code-review follow-up. The screenshot trigger counts only images nested in
functionResponse.parts. Compaction replaces those with the summary and
re-embeds survivors as TOP-LEVEL parts in the restoration block, which the
counter ignores — so the tool-image count always resets to ~0 and the
trigger cannot immediately re-fire, independent of maxRecentImages.
The resolveCompactionTuning JSDoc and the settings.md note previously warned
of a non-existent "maxRecentImages near threshold => compact every turn"
loop. Correct both, and add a regression test asserting
countToolResponseImages() is 0 on composePostCompactHistory output.
* fix(core): guard readFileSizeAdaptive against multi-GB reads; cover composer 4-entry branch
wenshao review round 2 on PR #4599.
- readFileSizeAdaptive now stats the file first and short-circuits to a
reference when its byte size exceeds maxChars*4 (the safe UTF-8 upper
bound — a file larger than that cannot fit within maxChars chars). This
stops a multi-GB file the agent previously touched from being slurped
into a Buffer and exhausting the heap mid-compaction, exactly when we're
trying to reduce memory. A large binary file now references rather than
reading to binary-detect.
- Add a test for composePostCompactHistory's 4-entry branch (attachments +
trailing model+functionCall) producing [user(summary), model(ack),
user(attachments), model(fc)]. This is the common mid-tool-loop
compaction case; a model->model adjacency here is a provider 400. Prior
tests only covered the 2-entry fold (no attachments) and 3-entry (no
trailing fc) shapes.
* fix(core): resolve symlinks in workspace boundary; guard compose against throws
wenshao review round 3 on PR #4599 (two Criticals).
- isInsideWorkspace now resolves symlinks via realpathSync (safeRealpath,
with a lexical fallback for non-existent paths). A symlink living inside
the workspace but pointing outside (e.g. workspace/.env -> ~/.ssh/id_rsa)
previously passed the lexical boundary check and had its target read and
embedded into the post-compact history sent to the provider. Added a
RED-verified security regression test (secret embedded under the old
lexical check; rejected under realpath).
- Wrap composePostCompactHistory in try/catch inside compress(). The
summary side-query has already succeeded at that point, so a
restoration-assembly throw (disk I/O / malformed history) previously
escaped to sendMessageStream, crashing the active turn AND bypassing the
COMPRESSION_FAILED breaker. It now degrades to summary + ack.
* fix(core): close 4 compaction Criticals from review round 4
wenshao review round 4 on PR #4599.
- isSummaryEmpty now checks the STRIPPED summary: a response that is only an
<analysis> block (no <state_snapshot>) strips to empty, so it takes the
COMPRESSION_FAILED_EMPTY_SUMMARY path instead of "succeeding" with
`[Summary unavailable]` as the agent's only context (silent amnesia).
- Manual /compress strips a trailing ORPHANED model+functionCall before
composing — it has no pending functionResponse, so preserving it would
emit model[fc] then the next user text turn -> API 400. Auto-compaction
still keeps it (the pending response pairs with it).
- The restoration-failure catch fallback now folds a trailing
model+functionCall into the ack turn, so a pending functionResponse
(auto mid-tool-loop) keeps its matching call even on the degraded path.
- extractRecentFilePaths skips file paths whose tool call FAILED (an error
functionResponse), so a denied read_file is never re-read off disk during
compaction — closing a permission-bypass side channel.
RED-verified regression tests for the empty-summary, orphan-strip, and
permission-bypass fixes. Corrected the postProcessSummary comment.
* test(core): cover composePostCompactHistory catch-fallback; document fold text drop
wenshao review round 5 on PR #4599.
- Regression test for the restoration-failure catch fallback: mock
composePostCompactHistory to reject and assert compaction still returns
COMPRESSED (no escape to sendMessageStream / breaker bypass) with the
trailing functionCall folded into the ack and the trailing text dropped.
- Document that the fold branch intentionally keeps only functionCall parts
(the trailing turn's text is already captured in the summary); the
asymmetry with the with-attachments branch is deliberate.
This commit is contained in:
顾盼
GitHub
parent
39cc9b3e6f
commit
365409366d
16 changed files with 2946 additions and 1227 deletions
|
|
@ -146,7 +146,11 @@ Settings are organized into categories. Most settings should be placed within th
|
|||
| `model.maxWallTimeSeconds` | number | Wall-clock budget for headless / unattended runs, in seconds. `-1` means unlimited. Overridable per-invocation via `--max-wall-time`, which requires a positive duration (`90`, `30s`, `5m`, `1h`, `1.5h`); the minimum is 1 second — sub-second values (`500ms`, `0.5`) are rejected as typos. Omit the flag to fall back to this setting. Aborts with exit code 55 when exceeded. | `-1` |
|
||||
| `model.maxToolCalls` | number | Cumulative tool-call budget for a run (counts every executed tool, success or failure; `structured_output` under `--json-schema` is exempt). `-1` means unlimited; `0` means "no tool calls allowed". Capped at 1,000,000 to catch typos. Overridable via `--max-tool-calls`. Aborts with exit code 55 when exceeded. | `-1` |
|
||||
| `model.generationConfig` | object | Advanced overrides passed to the underlying content generator. Supports request controls such as `timeout`, `maxRetries`, `enableCacheControl`, `splitToolMedia` (set `true` for strict OpenAI-compatible servers like LM Studio that reject non-text content on `role: "tool"` messages — splits media into a follow-up user message), `contextWindowSize` (override model's context window size), `modalities` (override auto-detected input modalities), `customHeaders` (custom HTTP headers for API requests), and `extra_body` (additional body parameters for OpenAI-compatible API requests only), along with fine-tuning knobs under `samplingParams` (for example `temperature`, `top_p`, `max_tokens`). Leave unset to rely on provider defaults. | `undefined` |
|
||||
| `model.chatCompression.contextPercentageThreshold` | number | **REMOVED.** Auto-compaction now uses a three-tier threshold ladder (warn / auto / hard) computed internally from the model's context window via the `computeThresholds()` function — no longer user-configurable. Setting this field in `settings.json` is silently ignored, and a one-line deprecation warning is emitted to stderr at startup. There is currently no replacement for "disable compression entirely" — reactive overflow recovery remains the safety net at the API layer if compression itself fails. (See PR #4345 / `docs/design/auto-compaction-threshold-redesign.md` for the redesign rationale.) | `N/A` |
|
||||
| `model.chatCompression.contextPercentageThreshold` | number | **REMOVED.** Auto-compaction now uses a three-tier threshold ladder (warn / auto / hard) computed internally from the model's context window via the `computeThresholds()` function — no longer user-configurable. Setting this field in `settings.json` is silently ignored (no startup warning). There is currently no replacement for "disable compression entirely" — reactive overflow recovery remains the safety net at the API layer if compression itself fails. (See PR #4345 / `docs/design/auto-compaction-threshold-redesign.md` for the redesign rationale.) | `N/A` |
|
||||
| `model.chatCompression.maxRecentFilesToRetain` | number | Number of most-recently-touched files whose current content is restored (embedded if small, otherwise referenced by path) into history after auto-compaction. `0` restores none. Env override: `QWEN_COMPACT_MAX_RECENT_FILES`. | `5` |
|
||||
| `model.chatCompression.maxRecentImagesToRetain` | number | Number of most-recent images (tool screenshots / user pastes) restored into history after auto-compaction. `0` restores none. Env override: `QWEN_COMPACT_MAX_RECENT_IMAGES`. | `3` |
|
||||
| `model.chatCompression.enableScreenshotTrigger` | boolean | When `true`, auto-compaction also fires once the number of tool-returned images accumulated in history reaches `screenshotTriggerThreshold`, independent of token usage — aimed at computer-use sessions where frequent screenshots dilute model attention. Counts only images returned inside tool results, not user-pasted images. Env override: `QWEN_COMPACT_SCREENSHOT_TRIGGER` (`1`/`true`/`0`/`false`). | `true` |
|
||||
| `model.chatCompression.screenshotTriggerThreshold` | number | Tool-returned image count at or above which the screenshot trigger fires (only when `enableScreenshotTrigger`). Compaction resets the count — surviving images are re-embedded as top-level parts, which the trigger doesn't count — so it won't immediately re-fire. Env override: `QWEN_COMPACT_SCREENSHOT_THRESHOLD`. | `50` |
|
||||
| `model.skipNextSpeakerCheck` | boolean | Skip the next speaker check. | `false` |
|
||||
| `model.skipLoopDetection` | boolean | Disables streaming loop detection checks. Defaults to `true` (loop detection is skipped) to avoid false positives interrupting legitimate workflows. Set to `false` to re-enable streaming loop detection — useful as a guardrail in headless / non-interactive runs where stuck repetition can otherwise waste budget. | `true` |
|
||||
| `model.skipStartupContext` | boolean | Skips sending the startup workspace context (environment summary and acknowledgement) at the beginning of each session. Enable this if you prefer to provide context manually or want to save tokens on startup. | `false` |
|
||||
|
|
|
|||
Loading…
Add table
Add a link
Reference in a new issue