Skip to content

feat(autofix): escalate a non-converging diff to a maintainer handoff #9104

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Merged
wenshao merged 1 commit into from
Aug 14, 2026
Merged

feat(autofix): escalate a non-converging diff to a maintainer handoff #9104

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
1 commit
Select commit
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
131 changes: 131 additions & 0 deletions .github/workflows/qwen-autofix.yml
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -165,6 +165,15 @@ env:
# value falls back to its default at the read site.
GROWTH_BUDGET_SRC_LINES: '${{ vars.QWEN_AUTOFIX_GROWTH_BUDGET_SRC_LINES || 400 }}'
GROWTH_BUDGET_TEST_LINES: '${{ vars.QWEN_AUTOFIX_GROWTH_BUDGET_TEST_LINES || 400 }}'
# Non-convergence handoff: once the growth brake has been over budget for
# this many PRIOR rounds in the window AND the diff has not shrunk from the
# most recent over-budget round, the round is DIVERGING (the fixes keep
# growing the diff, so
# Critical-only — which only trims non-Criticals — cannot help). At that
# point the round escalates to a maintainer-decision handoff (split / accept
# core + track the tail / redesign) instead of patching again. Same tunable
# contract as the budgets above (malformed → default at the read site).
GROWTH_DIVERGENCE_ROUNDS: '${{ vars.QWEN_AUTOFIX_GROWTH_DIVERGENCE_ROUNDS || 2 }}'
# An auth/access model error (401/402/403, "no access"/"does not exist")
# never self-heals - only a maintainer can fix the key - and every retry
# costs an agent run AND a PR comment. Cap those attempts far below
Expand Down Expand Up @@ -4463,6 +4472,11 @@ jobs:
# round's pushed growth would escape the budget for the rest of
# the live window.
echo "growth_base_win=${LIVE_REARM_KEY}"
# The round's own growth + over-budget flag, so the report step can
# write this round's autofix-growth-now marker (the per-round
# history the divergence read above consumes).
echo "growth_src=${GROWTH_SRC}"
echo "growth_test=${GROWTH_TEST}"
} >> "${GITHUB_OUTPUT}"
echo "📏 net diff src ${NET_SRC} / test ${NET_TEST} lines (window baseline ${BASE_SRC}/${BASE_TEST}, growth ${GROWTH_SRC}/${GROWTH_TEST}, budgets ${GROWTH_BUDGET_SRC_LINES}/${GROWTH_BUDGET_TEST_LINES})"

Expand All @@ -4477,6 +4491,74 @@ jobs:
CRITICAL_ONLY='true'
CRITICAL_ONLY_GROWTH='true'
fi
# Divergence: Critical-only only trims non-Criticals, so when the
# GROWTH that trips the brake is Critical-driven the diff keeps
# climbing anyway. Read this window's prior per-round growth markers
# (written by the report step): count the rounds that were over
# budget, and take the MOST RECENT prior over-budget run's growth
# SUM (highest run_id — see below; NOT the window-wide max, which a
# one-off spike would raise forever). The round is DIVERGING when it
# is over budget now, the brake has already fired for
# >= GROWTH_DIVERGENCE_ROUNDS prior rounds, and the diff has NOT
# shrunk from that most-recent sum — the fixes are not converging, so
# the round must escalate to a human decision instead of patching
# again. A diff that is over budget but SHRINKING (agent removing
# code) or a one-off overshoot stays in ordinary Critical-only.
if [[ ! "${GROWTH_DIVERGENCE_ROUNDS}" =~ ^([1-9][0-9]{0,3})$ ]]; then
echo "::warning::GROWTH_DIVERGENCE_ROUNDS='${GROWTH_DIVERGENCE_ROUNDS}' is not a positive count; using 2"
GROWTH_DIVERGENCE_ROUNDS=2
Comment thread
wenshao marked this conversation as resolved.
fi
Comment thread
wenshao marked this conversation as resolved.
# Count runs whenever the net is measured (not only over budget), so
# the trajectory clause below is accurate even on a round that pulled
# back under budget. markers:
# <!-- autofix-growth-now src=N test=N over=BOOL round=N run=ID key=W -->
# Deduped and ordered by run=GITHUB_RUN_ID, the per-workflow-run id:
# the report post's bounded retry re-posts one run's marker (same
# run_id → collapses to the LATEST by created_at, so a re-run's
# fresher attempt wins over its own stale first post), while every
# distinct address run has a fresh, monotonically increasing run_id. round=/eval-watermark are NOT a
# safe identity — a state-triggered lane (a persistent merge conflict
# selects the PR every scan with no new evaluable feedback) freezes
# both NEWEST and ROUND, so distinct over-budget runs would share them
# and collapse, stalling the count. Filtered to markers AFTER any
# mid-window base update (BASE_UPD_AT) — the same re-anchoring the
# growth-base read guards against. The "not shrinking" test compares
# against the MOST RECENT prior over-budget run's sum (highest
# run_id), not the window-wide max: a single transient spike would
# otherwise raise the bar forever and a genuine plateau-over-budget
# runaway (the exact case to escalate) would never clear it. The
# CURRENT run's own markers are excluded (run != GITHUB_RUN_ID): a
# re-run of a failed job keeps the same run id and its failed
# attempt already posted a marker, so counting it would over-report
# the round's own attempt as a PRIOR one.
GROWTH_DIVERGED='false'
OVER_ROUNDS_PRIOR=0
PREV_SUM=0
if [[ "${NET_MEASURED}" == 'true' ]]; then
read -r OVER_ROUNDS_PRIOR PREV_SUM < <(jq -r --arg ab "${AUTOFIX_BOT}" --arg key "${LIVE_REARM_KEY}" --arg baseupd "${BASE_UPD_AT}" --arg curr "${GITHUB_RUN_ID}" '
[ .[] | select((.user.login // "") == $ab) | . as $c | ($c.body // "")
| [ scan("<!-- autofix-growth-now src=(-?[0-9]+) test=(-?[0-9]+) over=([a-z]+) round=([0-9]+) run=([0-9]+) key=([^ ]+) -->") ] | .[]
| {sum: ((.[0] | tonumber) + (.[1] | tonumber)), over: .[2], round: (.[3] | tonumber), run: (.[4] | tonumber), win: .[5], at: ($c.created_at // "")} ]

@wenshao wenshao Aug 14, 2026

Copy link
Copy Markdown
Collaborator Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

[Critical] R6-2: Source and test growth have separate budgets, but the detector compares only their signed sum. Failure scenario: prior source/test growth is 500/500 and current growth is 900/0; source worsens over its own budget, but the sum falls 1000→900 and suppresses the handoff. Preserve both dimensions and compare the currently over-budget dimension(s) independently.

Witness:

C2 prior_sum=1000 current=900/0 over=true sum_not_shrinking=false
中文说明source/test 有独立预算，但检测器只比较有符号总和。上一轮 500/500、本轮 900/0 时，source 已恶化并超预算，总和却从 1000 降至 900，错误抑制 handoff。请保留两个维度并分别判断当前超预算维度。

— gpt-5.6-sol via Qwen Code /review (v0.21.11)

| map(select(.win == $key and .over == "true"))
Comment thread
wenshao marked this conversation as resolved.

@qwen-code-ci-bot qwen-code-ci-bot Aug 13, 2026

Copy link
Copy Markdown
Collaborator

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

[Suggestion] R2-8: the comparability guard covers mid-window BASE UPDATES but not AUTHOR PUSHES — and an author push shifts the net-diff measurement while nothing re-anchors. Verified at HEAD: LIVE_REARM_KEY advances only on takeover-ack/rearm comments (~4274-4280), BASE_UPD_AT only on the bot's own update-branch merge (~4436-4439), the pr-force-push marker appears only in new-feedback comment filters, and the growth-now marker carries no head sha; growth is measured against the persisted baseline, which survives author pushes. So prior sums measured on the old head are compared against this round's sum on the new head. Probe on the real block: rounds 1-3 over budget (sums 500/550/600), author pushes +300 lines of their own (ordinary mid-review behavior), next round over budget with sum 900 → GROWTH_DIVERGED=true OVER_ROUNDS_PRIOR=3 PREV_SUM=600: a handoff asserting "the Criticals ARE the growth" when the growth was the author's push; the stopped agent can no longer shrink the diff, so every later over-budget round re-diverges, stalling the loop until /retry re-arms. Mirror direction: an author deletion makes a genuinely diverging spiral read as shrinking, delaying a real escalation. The block comment below states the comparability contract this violates. Fix direction: fold the newest author-push signal into the cutoff alongside base updates — e.g. BASE_UPD_AT = max(newest autofix-base-updated, newest pr-force-push-reminder comment created_at; that comment already exists via pr-force-push-reminder.yml) — or give the marker a head identity (the autofix-redcheck head= marker rides in the same comment) and invalidate prior sums when the head moved without a bot eval marker.

中文说明

R2-8：可比性守卫覆盖窗口内的 base 更新，但不覆盖作者 push——而作者 push 会改变净 diff 测量，同时没有任何东西重新锚定。在 HEAD 上核实：LIVE_REARM_KEY 只在 takeover-ack/rearm 评论上推进（~4274-4280），BASE_UPD_AT 只在 bot 自己的 update-branch 合并上推进（~4436-4439），pr-force-push marker 只出现在新反馈评论过滤器中，growth-now marker 不带 head sha；增长按持久化基线测量，而该基线在作者 push 后依然存留。因此按旧 head 测得的先前总和会与本轮按新 head 测得的总和比较。对真实块的探针：第 1-3 轮超预算（总和 500/550/600），作者自己 push 了约 +300 行（评审中途的常见行为），下一轮超预算、总和 900 → GROWTH_DIVERGED=true OVER_ROUNDS_PRIOR=3 PREV_SUM=600：handoff 断言"Critical 就是增长本身"，而增长其实来自作者的 push；停下的 agent 无法再缩小 diff，于是之后每个超预算轮都会再次判定发散，循环停摆直到 /retry 重新 arm。镜像方向：作者删行会让真正发散的螺旋读起来像在缩小，延迟真正的升级。下方块注释声明的可比性契约正被此违反。修复方向：把最新的作者 push 信号与 base 更新一起折入截断——例如 BASE_UPD_AT = max(最新 autofix-base-updated、最新 pr-force-push-reminder 评论的 created_at；该评论已由 pr-force-push-reminder.yml 存在)——或给 marker 一个 head 身份（autofix-redcheck head= marker 就在同一评论里），当 head 在没有 bot eval marker 的情况下移动时使先前总和失效。

— qwen3.8-max via Qwen Code /review (v0.21.11)

@wenshao wenshao Aug 14, 2026

Copy link
Copy Markdown
Collaborator Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Acknowledged, deferring with rationale. Correct that an author push mid-window can make prior sums (old head) incomparable to this round's (new head) and produce a false handoff — but the outcome is fail-safe: the handoff stops and asks a maintainer, who is looking at a genuinely large diff anyway, and /retry re-arms. The clean fix (fold the newest author-push signal / a head-identity into the comparability cutoff, invalidating prior sums when the head moved without a bot marker) is a moderate change; I'd rather land the core feature + the correctness fixes (R2-1/R2-2) now and do the author-push + measured_at precision (R2-6) as a focused follow-up rather than grow this PR again. Will open a tracking note.

Comment thread
wenshao marked this conversation as resolved.

@wenshao wenshao Aug 14, 2026

Copy link
Copy Markdown
Collaborator Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

[Critical] R6-4: Filtering to over=true before ordering means recovery rounds never reset the divergence streak. Failure scenario: over-budget, recovery, over-budget, recovery, then over-budget again retains both old episodes and reaches the threshold despite repeated recovery. Count only the trailing consecutive over=true suffix after the latest recovery marker.

Witness:

C4 over_count=2 latest_over_sum=500 (recovery rows ignored)
中文说明排序前过滤为 `over=true`，会导致恢复轮永远无法重置发散计数。多次“超预算→恢复”后再次超预算，旧记录仍会达到阈值。请只统计最近一次恢复后的尾部连续超预算区间。

— gpt-5.6-sol via Qwen Code /review (v0.21.11)

| map(select(.run != ($curr | tonumber)))
| map(select($baseupd == "" or (.at > $baseupd)))

@qwen-code-ci-bot qwen-code-ci-bot Aug 13, 2026

Copy link
Copy Markdown
Collaborator

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

[Suggestion] R2-6: this comparability filter uses the marker comment's created_at as a proxy for when the growth was MEASURED, but the marker carries no measurement time: GROWTH_SRC/GROWTH_TEST are measured in prepare, and the report writes the marker only after the agent run and push — the file itself documents that gap as "the agent's ~120-minute window" (the PUSH_RACE_MERGED salvage exists because mid-run branch movement was "observed twice in one day"). A round in flight when a concurrent run lands a base update therefore posts its marker AFTER the autofix-base-updated comment with sums measured against the OLD merge base, yet its post-update timestamp passes this filter — admitting an incomparable sum into OVER_ROUNDS_PRIOR/PREV_SUM. The idiom is inherited from the growth-base read, but this diff newly spends that data on an escalation decision, and the comment below ("so prior sums stay comparable to this round's") states the contract this violates. Probe on the real block: base-update comment 12:00; stale marker (round=2, over=true, measured pre-update, posted 12:30); genuine post-update marker (round=3, 13:00); current round over budget → WITH the stale marker true 2 (handoff after only ONE genuine prior over-budget round), WITHOUT it false 1. The premature arm is bounded to at most one round early and fails safe (a false handoff waits for a maintainer; /retry resumes). Fix direction: stamp the marker with the prepare-time measurement instant (e.g. a measured_at= prepare output) and filter on that instead of comment created_at.

中文说明

R2-6：这个可比性过滤器把 marker 评论的 created_at 当作增长被测量时刻的代理，但 marker 本身不带测量时刻：GROWTH_SRC/GROWTH_TEST 在 prepare 中测量，而 report 直到 agent 运行并 push 之后才写 marker——文件本身把这个间隙记为"agent 的约 120 分钟窗口"（PUSH_RACE_MERGED 补救逻辑的存在正是因为运行中的分支变动"一天内观察到两次"）。因此，当某轮运行中并发运行落地了一次 base 更新时，该轮会在 autofix-base-updated 评论之后发布 marker，其总和却是按旧 merge base 测量的，而其更新后的时间戳能通过本过滤器——把一个不可比的总和放进 OVER_ROUNDS_PRIOR/PREV_SUM。这个写法继承自 growth-base 读取，但本 diff 首次把该数据用于升级决策，且下方注释（"so prior sums stay comparable to this round's"）声明的契约正被此违反。对真实块的探针：base-update 评论 12:00；过期 marker（round=2、over=true、更新前测量、12:30 发布）；真正的更新后 marker（round=3、13:00）；当前轮超预算→有过期 marker 时 true 2（仅有一个真实先前超预算轮就 handoff），没有时 false 1。提前分支至多提前一轮且失败安全（误 handoff 会等待 maintainer；/retry 恢复）。修复方向：把 prepare 时刻的测量时刻打进 marker（如 measured_at= prepare 输出），按它过滤而非评论 created_at。

— qwen3.8-max via Qwen Code /review (v0.21.11)

@wenshao wenshao Aug 14, 2026

Copy link
Copy Markdown
Collaborator Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Acknowledged, deferring with rationale. This is real but self-bounded: it needs a round in flight during a concurrent base update, and you note it's 'at most one round early and fails safe' (a false handoff waits for a maintainer; /retry resumes). The precise fix (a prepare-time measured_at= output stamped into the marker, filtered instead of comment created_at) is a clean follow-up but adds a new output + marker field + filter; I'd rather not expand this PR further for a fail-safe one-round-early edge. Tracking it for a follow-up alongside R2-8.

Comment thread
wenshao marked this conversation as resolved.
| group_by(.run) | map(max_by(.at))
| sort_by(.run)
| "\(length) \((last.sum) // 0)"' "${WORKDIR}/ic.json" 2> /dev/null || echo "0 0")
[[ "${OVER_ROUNDS_PRIOR}" =~ ^[0-9]+$ ]] || OVER_ROUNDS_PRIOR=0
[[ "${PREV_SUM}" =~ ^-?[0-9]+$ ]] || PREV_SUM=0
if [[ "${CRITICAL_ONLY_GROWTH}" == 'true' \
&& "${OVER_ROUNDS_PRIOR}" -ge "${GROWTH_DIVERGENCE_ROUNDS}" \
&& $(( GROWTH_SRC + GROWTH_TEST )) -ge "${PREV_SUM}" ]]; then
GROWTH_DIVERGED='true'
fi
fi
# The over-budget flag feeds the report's per-round marker; the
# handoff itself is enforced by the feedback.md text below, so
# GROWTH_DIVERGED needs no step output.
echo "critical_only_growth=${CRITICAL_ONLY_GROWTH}" >> "${GITHUB_OUTPUT}"
Comment thread
wenshao marked this conversation as resolved.
[[ "${GROWTH_DIVERGED}" == 'true' ]] &&
echo "🛑 diff not converging: over budget now, ${OVER_ROUNDS_PRIOR} prior over-budget round(s) in this window, growth not shrinking — escalating to a maintainer decision instead of patching."
# Which trusted humans have exhausted their per-window regular
# feedback budget (see CRITICAL_ONLY_HUMAN_BATCHES). A batch is
# COUNTED only when a Critical-only round actually consumed it:
Expand Down Expand Up @@ -4655,6 +4737,26 @@ jobs:
echo "Only feedback newer than the last evaluation (${WATERMARK}) from"
echo "trusted maintainers or the automated reviewer is listed."
echo
# Diff-growth trajectory: the agent sees how big this PR has grown
# so it can prefer minimal/subtractive fixes and recognise a
# non-converging spiral, rather than reflexively adding code for
# every finding.
if [[ "${NET_MEASURED}" == 'true' ]]; then
echo "## Diff growth this window"
Comment thread
wenshao marked this conversation as resolved.
echo
echo "Net diff vs this counting window's baseline: source ${GROWTH_SRC} / test ${GROWTH_TEST} lines (budgets ${GROWTH_BUDGET_SRC_LINES}/${GROWTH_BUDGET_TEST_LINES}; ${OVER_ROUNDS_PRIOR} prior round(s) already over budget). Prefer minimal, root-cause, subtractive fixes. If closing a finding would grow the diff materially AND the same class of gap keeps reappearing on code earlier rounds added, that is a signal to escalate for a split — not to keep adding guards."
Comment thread
wenshao marked this conversation as resolved.
echo
fi
# Non-convergence handoff (the diff keeps growing past budget across
# rounds): a maintainer-decision item the agent must NOT settle by
# patching. Framed as defer-to-human so the address run stops with a
# handoff (SKILL: "Stop BLOCKED when any defer-to-human item remains").
if [[ "${GROWTH_DIVERGED}" == 'true' ]]; then
Comment thread
wenshao marked this conversation as resolved.
echo "## Needs a maintainer's decision — this PR is not converging"
Comment thread
wenshao marked this conversation as resolved.
Comment thread
wenshao marked this conversation as resolved.
echo
echo "This PR's diff has stayed over the growth budget for ${OVER_ROUNDS_PRIOR}+ rounds in this window and is still not shrinking (source ${GROWTH_SRC} / test ${GROWTH_TEST} net lines vs budgets ${GROWTH_BUDGET_SRC_LINES}/${GROWTH_BUDGET_TEST_LINES}). The review findings are themselves driving the growth, so continuing to patch will not converge — Critical-only mode cannot help because the Criticals are the growth. Do NOT apply more code fixes this round. Treat this as a \`defer-to-human\` item: STOP with a handoff that names the decision and lays out the options — split the PR (land the core, track the remaining findings as follow-up issues), redesign the approach, or accept the current state with the tail deferred — plus your recommendation. Leave the call to the maintainer; declining or implementing one direction yourself IS deciding."
echo
fi
echo "## Reviews"
jq -r --arg wm "${WATERMARK}" --arg rb "${REVIEW_BOT}" --arg ab "${AUTOFIX_BOT}" \
--argjson critical_only "${CRITICAL_ONLY}" --argjson trust "${TRUSTED_ASSOC}" \
Expand Down Expand Up @@ -5251,6 +5353,12 @@ jobs:
# can report under a stale WINDOW after a re-arm; the marker must
# land under the key later reads will use.
GROWTH_BASE_WIN: '${{ steps.prepare.outputs.growth_base_win }}'
# This round's own growth + over-budget flag, written as the
# per-round autofix-growth-now marker so later rounds can measure
# divergence (growth still climbing over budget = not converging).
GROWTH_SRC: '${{ steps.prepare.outputs.growth_src }}'
GROWTH_TEST: '${{ steps.prepare.outputs.growth_test }}'
CRITICAL_ONLY_GROWTH: '${{ steps.prepare.outputs.critical_only_growth }}'
Comment thread
wenshao marked this conversation as resolved.
Comment thread
wenshao marked this conversation as resolved.
Comment thread
wenshao marked this conversation as resolved.
run: |-
# gh has its own $GITHUB_ENV-injectable channels: pin the host and
# drop any planted token BEFORE the identity check below, so a
Expand Down Expand Up @@ -5620,6 +5728,10 @@ jobs:
if [[ "${GROWTH_BASE_NEW}" == 'true' ]]; then
echo "<!-- autofix-growth-base src=${GROWTH_BASE_SRC} test=${GROWTH_BASE_TEST} key=${GROWTH_BASE_WIN:-${WINDOW:-none}} -->"
fi
# Per-round growth history the next round's divergence read
# counts; run=GITHUB_RUN_ID is the dedup+order identity (a retry
# re-posts the same run, distinct runs get fresh run ids).
echo "<!-- autofix-growth-now src=${GROWTH_SRC:-0} test=${GROWTH_TEST:-0} over=${CRITICAL_ONLY_GROWTH:-false} round=${NEXT_ROUND} run=${GITHUB_RUN_ID} key=${GROWTH_BASE_WIN:-${WINDOW:-none}} -->"

@wenshao wenshao Aug 14, 2026

Copy link
Copy Markdown
Collaborator Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

[Critical] R6-6: Historical markers store only over=true, not the effective budgets that produced it. Failure scenario: two 500-line rounds exceed an old 400-line budget; after the budget is raised to 1000, the first 1100-line round immediately counts those old rows as prior rounds over the new budget and triggers a premature handoff. Include effective budgets in the marker and match them, or make budget configuration part of the window identity.

Witness:

C6 old_markers_count=2 prior=500 current=1100 diverged=true
中文说明历史 marker 只保存 `over=true`，没有保存当时的有效预算。旧预算 400 下的两次 500 行记录，在预算提高到 1000 后仍会被当作新预算下的先前超标轮，导致首次 1100 行就过早 handoff。请记录并匹配有效预算，或把预算配置纳入窗口身份。

— gpt-5.6-sol via Qwen Code /review (v0.21.11)

} > "${WORKDIR}/report.md"
STATUS="pushed (round ${NEXT_ROUND}/${MAX_ROUNDS})"
else
Expand All @@ -5644,6 +5756,10 @@ jobs:
if [[ "${GROWTH_BASE_NEW}" == 'true' ]]; then
echo "<!-- autofix-growth-base src=${GROWTH_BASE_SRC} test=${GROWTH_BASE_TEST} key=${GROWTH_BASE_WIN:-${WINDOW:-none}} -->"
fi
# Per-round growth history the next round's divergence read
# counts; run=GITHUB_RUN_ID is the dedup+order identity (a retry
# re-posts the same run, distinct runs get fresh run ids).
echo "<!-- autofix-growth-now src=${GROWTH_SRC:-0} test=${GROWTH_TEST:-0} over=${CRITICAL_ONLY_GROWTH:-false} round=${ROUND} run=${GITHUB_RUN_ID} key=${GROWTH_BASE_WIN:-${WINDOW:-none}} -->"
} > "${WORKDIR}/report.md"
STATUS="no action needed"
fi
Expand Down Expand Up @@ -5790,6 +5906,16 @@ jobs:
EFFECTIVE_ROUND: '${{ steps.prepare.outputs.effective_round }}'
MODEL: '${{ vars.QWEN_AUTOFIX_MODEL || vars.QWEN_PR_REVIEW_MODEL }}'
CHECKED_OUT_HEAD: '${{ steps.prepare.outputs.checked_out_head }}'
# This step also posts a round report (timeout / gate-rejection /
# abort), so it writes the per-round growth-now marker too — else an
# over-budget round that never reaches 'Push and report' leaves a
# history gap and the divergence count under-reports. Empty outputs
# (prepare never ran) fall through the :-0/:-false marker fallbacks
# to an inert over=false entry.
GROWTH_SRC: '${{ steps.prepare.outputs.growth_src }}'
GROWTH_TEST: '${{ steps.prepare.outputs.growth_test }}'
CRITICAL_ONLY_GROWTH: '${{ steps.prepare.outputs.critical_only_growth }}'
GROWTH_BASE_WIN: '${{ steps.prepare.outputs.growth_base_win }}'
run: |-
# The head the agent actually evaluated — captured in prepare before
# any mutation, not the report-time remote head (which can move
Expand Down Expand Up @@ -6295,6 +6421,11 @@ jobs:
echo "🧠 Handled by **Qwen Code** · model/模型 \`${MODEL_DISPLAY}\`"
echo
echo "<!-- autofix-eval ts=${MARK_TS} acted=false round=${MARK_ROUND} win=${WINDOW:-none} -->"
# Per-round growth history the divergence read counts — same
# marker the push/no-op report paths write, so an over-budget
# round that timed out or was gate-rejected is not a gap. run=
Comment on lines +6424 to +6426

@wenshao wenshao Aug 14, 2026

Copy link
Copy Markdown
Collaborator Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

[Critical] R6-3: The stale-base retry can update the branch after growth was measured, but this report posts the pre-update sample without an autofix-base-updated marker. Failure scenario: prepare measures against old main, update-branch succeeds, and later rounds attribute the base merge’s size shift to PR/autofix growth. Emit the existing re-anchor marker on successful stale-base retry, or omit this sample and force a fresh baseline.

Witness: not run — settling the complete path requires a live GitHub update-branch mutation; the exact writer/reader trace shows this path emits no marker recognized by BASE_UPD_AT.

中文说明stale-base retry 会在增长测量后更新分支，但报告仍写入更新前样本，且没有 `autofix-base-updated` marker。后续轮次会把 base merge 的体积变化误归因于 PR/autofix 增长。请在成功更新 base 时写入重锚 marker，或省略该样本并强制重建基线。

— gpt-5.6-sol via Qwen Code /review (v0.21.11)

# (per-workflow-run) is the dedup/order identity.
echo "<!-- autofix-growth-now src=${GROWTH_SRC:-0} test=${GROWTH_TEST:-0} over=${CRITICAL_ONLY_GROWTH:-false} round=${MARK_ROUND} run=${GITHUB_RUN_ID} key=${GROWTH_BASE_WIN:-${WINDOW:-none}} -->"

@wenshao wenshao Aug 14, 2026

Copy link
Copy Markdown
Collaborator Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

[Critical] R6-1: Timeout, gate-rejection, abort, and no-op reports count as evidence that fixes keep growing the diff even when they changed no branch state. Failure scenario: two failed address runs push nothing but write distinct unchanged over=true samples; the next run reaches the threshold and forces a maintainer handoff although no fix grew the diff. Use only successful change-producing samples for divergence; retain non-change samples only for trajectory diagnostics.

Witness:

C1 count=2 prev=500 now=500 diverged=true
中文说明timeout、gate-rejection、abort 与 no-op 即使没有改变分支，也会被当作“修复持续扩大 diff”的证据。两次失败运行未 push 却写入不同的 `over=true` 样本，下一轮便达到阈值并 handoff。发散判断应只使用成功产生变更的样本；无变更样本仅用于轨迹诊断。

— gpt-5.6-sol via Qwen Code /review (v0.21.11)

# A sentinel ts means the agent evaluated NOTHING (crash, API
# error, gate crash) and the next scan must retry. Recording a
# judged head here would make RED_HEAD == LIVE_HEAD, so the
Expand Down
18 changes: 18 additions & 0 deletions .qwen/skills/autofix/SKILL.md
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -319,6 +319,24 @@ implement — satisfying a nit is never a reason to bloat the code.
identity). A maintainer writing "fix X before merge" after round five
means exactly that when it reaches you — plus failed checks and the
requested base-conflict resolution.
- Diff-growth trajectory: `feedback.md` opens with a `Diff growth this window`
section (source/test net lines vs budget, and how many prior rounds were
already over budget) whenever growth is measured. Use it: prefer minimal,
root-cause, subtractive fixes over additive guards, and read a rising
trajectory as a signal — if closing a finding would grow the diff materially
AND the same class of gap keeps reappearing on code an earlier round added,
the right response is to escalate for a split, not to add another guard.
- Not converging (the diff keeps growing past budget): when `feedback.md`
contains a `Needs a maintainer's decision — this PR is not converging`
section, the growth brake has been over budget across rounds and the diff is
still not shrinking — the findings themselves are driving the growth, so
Critical-only cannot help (the Criticals ARE the growth). Do NOT apply more
code fixes this round. This is a `defer-to-human` item: STOP `BLOCKED` with a
handoff that names the decision and lays out the options — split the PR (land
the core, track the remaining findings as follow-up issues), redesign, or
accept the current state with the tail deferred — plus your recommendation.
Continuing to patch, or deciding the split yourself, is exactly the wrong
move; the call is the maintainer's.
- Needs a maintainer's decision: a finding that turns on a judgment that is
NOT yours to make — a product or scope tradeoff (is this acceptable for v1?
should the PR be split?), two reviewers asking for opposite things, or whether
Expand Down
Loading
Loading