LESSONS.md

Lessons from merged work

magpie's code is written, reviewed, merged and released by agents. Each night the day's commits are reviewed for what had to be fixed again, what shouldn't have merged and what was verified too thinly. What we learned is kept here. Read this before changing code. Each lesson is a rule, then why, then the evidence. "Seen" counts the days a review found it. A lesson seen on 3 or more separate days moves into docs/code-standards.md.

The day of 2026-10-05 had 220 commits and ~170 releases. About 40 of them fixed something released earlier the same day or the day before. Most fixes didn't need a second try. The ones that did fall into the patterns below.

On 2026-10-07 four lessons were seen a third day and moved into the code standards: a fix reaches every sibling, a change re-derives what is built on it, a red test on main is a bug now, and a merge records what was run at the reviewed head. They are listed under Moved to the code standards.

2026-10-08 had 192 commits and 18 releases (v0.1.1105 to v0.1.1122), each tagged after a green run. Two more lessons were seen a third day and moved: real use before building on it, and GUI tests in both engines and every language. The acceptance checks now include gofmt -l and a Linux and Windows go vet, since a cross-OS build skips _test.go.

2026-10-09 had 109 commits and 10-10 had 74 up to 13:51 (v0.1.1127 to v0.1.1154). Every tag waited for a green run, every release has its 16 assets, and every commit over 400 lines has a body that says what was run. What went wrong was earlier in the work. A privacy exemption was keyed on a field's name, which let tool arguments go out unmasked. Two sessions committing from one checkout shipped each other's files under the wrong subjects. A known DNS-rebinding hole sat documented in a subsystem page until the review reproduced it (fixed in c431225a). Three lessons were seen a second day.

From 10-10 13:51 to 10-11 05:20 there were 191 commits and 18 releases (v0.1.1155 to v0.1.1172), each tagged on a green run. Most bodies say what was run, what was broken by hand, and what wasn't tried. What went wrong was in the tests and in what "the same on main" meant. A test-isolation fix was made twice and missed a package each time (#1524). Four commits shipped past a red they called load or "the same on main" without filing it, and one of those reds was a real rate-limit gap. Two lessons were seen a third day and moved into the code standards: whatever a test starts ends when it does, and a timing test leaves room on both sides.

From 10-11 05:20 to 10-14 04:00 there were 602 commits and 56 tags (v0.1.1173 to v0.1.1228). Every tag the review checked was on a green run, but v0.1.1213 and v0.1.1218 were never published (#1846), and two sessions tagged one commit twice. What went wrong was mostly running less than the suite: four commits on 10-13 ran chosen packages and turned main red through guard tests in packages they didn't touch, and two left GUI files red. Seven lessons were seen a third day and moved into the code standards: a failed read means unknown, nothing blocks under a lock its reader needs, another app's values come from that app, a test named for a guarantee fails without it, a test never reaches the network, a tag waits for a green run and is not a release, and a large commit has a body with what was run.

Reading the report

Reproduce the reporter's exact case, on the surface they use, before you call it fixed. Use their provider, plan, agent and layout, and their GUI, TUI or CLI. Check their goal (it drags, the title shows, the reason is readable), not a step on the way to it.

• #834 came back 4 times (v0.1.920 → .925 → .931 → .938). Each round fixed only the agent in the screenshot. #792, #743, #791 and #933 were closed and reopened within hours.
• 4788cf2b put gnayiab's Cursor list error in the GUI. The reporter uses the TUI in WSL. Fixed in f3c05b5e.
• 60b910b7 made a warm /api/state faster. The report was the cold start (~8.5s of SyncCatalog). Fixed in 2a4042a9.
• 10-13: def3a41e3 built on a request its reporter (Fallmonkey) had put on hold ("先别做"), and was reverted in 2c30e5cee. Read the whole thread, to its last comment, before starting.
• Seen 2× (2026-10-05, 2026-10-13).

Build fixtures from the reporter's literal bytes, not a similar sample you made up.

• #823: 878a8489 "proved" the Trae case with an invented <tool_call>{json}. The screenshot had no JSON. The reporter came back (c353f7bd), and the parser has grown 2 more formats since.
• 3f76105b's test used a short check-in reason. The real one was ellipsized (#821, eb5edced).
• 26706c33 used a different skills layout from #791's.
• For model-output parsers, collect several real failing turns first. Truncated and empty bodies go in the tests.
• Seen 1× (2026-10-05).

When the vendor refuses magpie but serves its own client on the same account, diff every header of the failing request against the official client's. Do that before adding retries or fallbacks.

• #256 Copilot Auto took four releases (719a6820, 026158b0, 251ef218, then 4660fd69). Each added another fallback. The cause was a missing X-GitHub-Api-Version beside Copilot-Session-Token.
• Seen 1× (2026-10-05).

Decide whether a refusal is about the account, the model or the request content before you rest an account.

• 5353ed54 counted WorkBuddy's "unapproved channel" (an agent's system prompt) as an account fault. It rested every WorkBuddy AI account for 5 days (2033df0f).
• Test that the same request fails the same way on a second account.
• Done right on 10-07: bed306ae (#1062) told Google's 400 for an unsupported thinking level apart from thinking_config being turned away. It now remembers that the model refused reasoning off, not that the provider refused thoughts.
• Seen 1× (2026-10-05).

Fix what the error is about, not only how it is worded.

• e63fbb0e (kkgg on Discord, "GLM 直接429了"): Zhipu answered a GLM Coding Plan key sent to its pay-as-you-go endpoint with 429 1113 "余额不足". The commit relabelled the 429 as out of credit, and nothing in magpie tells the user the key belongs on the Coding Plan endpoint, which is what would make their requests work.
• When a report shows a wrong label, also ask why the request failed. Fix that, or point the user to the fix in the message itself.
• 10-10: 1134c7a8 reworded the "restart codex" advice. luci hit the same stale app-server daemon on the next release, and e611fc65 then restarted it from magpie.
• Seen 2× (2026-10-06, 2026-10-10).

Fix the class, not the sample

Scope a vendor rule to that vendor.

• e950eb08 capped every model at 272K over one DeepSeek report. Claude [1m] models were capped too (531dc0aa).
• 71d2da1f priced Anthropic's 1-hour cache write onto OpenAI models (3a30ac87).
• Check the claim about the vendor's behaviour against its source or docs. Add a test that another vendor's model is unchanged.
• Seen 1× (2026-10-05).

A restriction covers every route that reaches the thing it guards.

• 46f03154 (#882) held keys to some models. count_tokens and System One's decision model weren't covered until d7a8ddc8.
• Done right on 10-07: 6b6113aa let a gateway key take the user's own value. It found that a local request with such a value would skip key limits, and moved managedKey onto access.Named, with a test that fails on the old check. c83c0d5f (#1112) kept a pinned MAGPIE_ADDR host while sharing, and checked that requests behind Tailscale Serve still count as remote.
• Seen 1× (2026-10-05).

Scope a privacy or security exemption by where the value sits in the protocol, never by a key's name.

• 7137feb4 kept any object holding a "signature" or "encrypted_content" key, and any *_id key, unmasked. A tool call's input is written by the model, so {"signature": …, "body": "DB_PASSWORD=… 13800138000"} or a user_id in tool_use input went to the vendor unmasked. 1965e1f0 limited the exemption to thinking/reasoning blocks and Gemini's thoughtSignature, and never keeps anything under a tool call's arguments.
• Test the exemption with the protected key placed where the user's data goes: tool arguments, message text, a nested object.
• Seen 1× (2026-10-09).

A security gap a doc names is an open bug, not a description.

• docs/subsystems/gateway-routing.md said "There is no Host check, so a page whose own name resolves to 127.0.0.1 (DNS rebinding) counts as same-origin". On 10-10 the review reproduced it on main. A rebound page got 200 on /v1/models, /v1/chat/completions and /v1/magpie/quotas without a key. Fixed in c431225a.
• A same-origin exemption needs the Host checked too. A rebound page's GET carries no Origin at all.
• When writing such a sentence, file an issue in the same commit.
• Seen 1× (2026-10-10).

Text magpie adds to a prompt that goes to a vendor names neither magpie nor the vendor's limits.

• 5d33f060b (10-12) headed the models it added to Codex's spawn_agent description "(Codex names only five above; magpie serves these too)". That text goes upstream, to OpenAI too. 3c4f8f9ae made it neutral, and its test fails when the text names magpie, Codex or "five". The cost is ban risk, so a prompt change gets the same care as a header change.
• Seen 1× (2026-10-12).

A bit or enum value has one owner.

• d4c2512d2 (#1874) gave a usage row's Flags bit 256 to rowOwnCost while a rebase of #1463 gave it to rowReasoningHidden, so every reported cost read back as hidden reasoning. 8179ce768's TestRowFlagsApart fails when a bit is used twice. After a rebase, grep the values both sides claim.
• Seen 1× (2026-10-13).

A command never takes a flag, or a value it doesn't know, as data to write.

• magpie snow --help took "--help" as a model: it deleted profiles/magpie.json and wrote "--help" into the user's profile. 76db4c6c2 (#1457) made --help print usage on every command, and Snow and Gemini CLI refuse an unknown value. Try each command with --help and a bad value in a sandbox HOME.
• Seen 1× (2026-10-11).

The user's own files and settings

Never let magpie's default override a value the user set themselves.

• CLAUDE_CODE_AUTO_COMPACT_WINDOW overrode the user's own autoCompactWindow for ~150 releases (602d4ec6).
• magpie claude model default deleted the user's own ANTHROPIC_BASE_URL and dropped the stash for 12 days (fe2603ff).
• Before writing a key, list everywhere the agent reads that setting from. Test with user-owned values present.
• 10-09/10-10: four fixes for magpie writing over a user's own choice.
  • 603dc0c4: dsh's picked model was lost on Disconnect.
  • 5aca7d09: Codex's effort left at Default was written over with medium.
  • c65d9fb4: magpie-models.json dropped keys added by hand.
  • d9bf7c1c: ZCode's API format was put back on every start.
  • Each was written first without a test that holds a user-set value.
• Seen 2× (2026-10-05, 2026-10-10).

Before deleting or overwriting a live credential, find every record that could match the same key.

• 43a65950 matched Codex accounts by display name. Two Team seats of one email share it, so removing one deleted auth.json for both (c13f3bfb). The sign-out rule was then rewritten again (4c35c1b8).
• Seen 1× (2026-10-05).

When you hand users a writable copy, test "user edits it, then sync".

• 3c7ceb68's "Give skills as Copies" undid the user's own edits on the next start (20bdf1e9).
• Seen 1× (2026-10-05).

Verification

Windows CI only compiles the tests, so a Windows-only red shows nowhere. Run the package on the Windows box before calling a Windows failure "the same on main".

• TestCursor was red on Windows for 3 days, from 585f7f7d until #1248.
• 5d0c1dca's 16 and 76 Windows failures were called the same as main. Neither was filed. #923 is the open work to run more of the suite there.
• Seen 1× (2026-10-08).

What telemetry sends is an allow-list, and a change to what is sent re-derives every place that discloses it.

• 169c23c1 sent omp# profile ids and the user's own model ids, which aren't in the public catalog (fixed in 2cdae80a: only known ids are sent).
• 7d663a27 sent each partner's daily counts, while the READMEs' Privacy and the site's "What does magpie send about me?" still listed what was sent before (b4aa62af).
• Grep the READMEs (en, zh) and site/public/docs (en, zh, ja) for the privacy text in the same commit.
• Seen 1× (2026-10-08).

A test that reads a dependency's source needs the dependency on disk.

• 884dbdb2's TestTrayMenuOpenFrameIsWails read Wails' source through go list -m Dir. CI's fresh module cache hasn't unpacked it, so main was red on macOS and ubuntu for ~25 minutes (9b6b9966).
• Run go mod download for it in the test, or keep the bytes it checks in testdata.
• Seen 1× (2026-10-10).

A check that a value is gone looks past text magpie generates.

• 6c099edd's redact test checked that "ops" was gone by substring. The placeholder's random suffix spelled it ({{HOSTNAME_mlopsiy3}}) on macOS CI. f0453192 strips placeholders before the check, in its sibling TestMask too.
• Seen 1× (2026-10-10).

Playwright's WebKit is not WebKitGTK. A GUI change with motion or a new browser API is tried in the real Linux app too.

• 18e82adc6 (10-11) traded the layouts in a view transition. The GUI suite passed in WebKit, but under Wayland magpie turns WebKitGTK's compositing off, and WebKitGTK 2.52 crashed in the transition and froze the window (f5dd4d604, a reporter on X).
• Seen 1× (2026-10-12).

A path that differs by OS is tested with each OS's value.

• ee7a9b24a (10-11) read Grok App's folder from $XDG_DATA_HOME on Linux. The test sandbox set HOME, XDG_CONFIG_HOME and XDG_CACHE_HOME but not XDG_DATA_HOME, so the drift test was red on ubuntu CI (f38f65030). Force the OS switch in a test, and give the sandbox every variable the code reads.
• Seen 1× (2026-10-11).

Reproduce on CI's Go version.

• #1184 couldn't be reproduced on go1.27.1 locally. CI ran 1.26.x, and the cause was a Go 1.26.8 segfault. Docker golang:1.26.8 with GOTOOLCHAIN=local matches CI.
• Seen 1× (2026-10-08).

Concurrency and tests

A fake or child process a test starts ends when the test does, and a timing test leaves room on both sides, are in the code standards since 10-11. See Moved to the code standards. Two habits from them:

• After a test times out, check ps for its binary. An orphaned gateway.test spun at 96% CPU for 30 minutes. Size -count to -timeout; a timeout panic is not a hang until its stacks show one.
• Reproduce a timing flake by injecting the delay, not by rerunning until it fails (0f8e4ebc, 229ce19a, 9e35eef3).

Red tests and releases

Flaky tests: find the cause. "A test that fails on main is a bug to fix now" is in the code standards. What the reviews found about finding the cause:

• TestPluginSOCKSProxy timed out at 60s on ubuntu CI and was rerun as "load jitter". It takes 3s on a Mac every time, and a dead local proxy should refuse in milliseconds, so something on that path waits. Its cause isn't found yet.
• TestUnreadLoginsNotWrittenOver needed two writes inside one wall-clock second (a .bad-<second> name) and failed on Linux CI (237d216b). Test a name made from time.Now() across a second boundary.
• Before fixing a red CI run, git log origin/main for a commit that names that run or test. 85fd35ab and e34e53a1 fixed the same lost tail in two sessions at once.
• 10-07: 07e246e5 named 24 GUI failures as "the same on origin/main". Run alone on main, all 8 of their files pass, and so do the two the full run failed (api-key-theme, fold-show-grip). They fail only under the full run, and CI never runs the GUI suite, so nothing shows it. e95852e3 (#948) named TestCursor (Windows) and TestOTelRetryAfterBound the same way. Both pass on macOS main and on CI. Neither is filed.
• 10-08: 6ab4c16f fixed the run 0f8e4ebc had already fixed. bc7c62dc, #1216, a2a7fb11, 5d0c1dca and #950 each shipped past "fails the same on main".
• Done right on 10-08: #1184 found the Go 1.26.8 segfault behind its flake, and 30a8ce29 fixed its flake at the cause.
• 10-11: four commits named a red and shipped past it unfiled. b1396d23 reran TestMaxRPM as load; it was a real gap (the vendor counted a late first request in the next request's minute, ce1ecaa4). 2305ab1f's dsh test and 20582866a's queued-stream test "pass alone"; both had causes (9c377231, 3f4287e6). db965824 said panel-usage "fails the same on origin/main 0a11b183", but 0a11b183 was hours old and f823a53d had fixed that test; on the real main it passed 3 of 3.
  • "The same on main" means on origin/main just fetched: git fetch, then name the sha you ran.
  • A red under load is a timing assumption a user will meet on a busy machine. File it, with its output, in the same commit.
• 10-11: two sessions again fixed one flake in the same minute (4caae1c8 and 3f4287e6), and on 10-10 3a1ade97 landed while another fix of the same test was being written. Checking git log isn't enough while both are working; say on an issue, or file one, that you are on it before you start.
• 10-12/10-13: #1794 got four "seen again" comments before anyone took it (fixed in 6ae6b3402), and #1543 was seen three times with no owner (097479dab). A comment that a red came back is not work on it: take it or say who has it. 31c1e3ab1 named 8 GUI files that "fail the same way on an origin/main 552e974f2 copy" and filed none of them; they were fixed one by one over the next day (searcher-pick, profile-in-use, library-mcp-tools among them).
• Done right: 2bd49cc5's race was fixed at its cause in 25 minutes (28049bd2, -race -count=40 -cpu 1,2). 60f87223, b1cba654, 443f1cbc and 395c4bc5 fixed flakes at their cause instead of retrying. 06420a71 fixed the macOS first-run check at its cause.
• Seen 3× (2026-10-05, 2026-10-06, 2026-10-07); the rule itself is in the code standards.

Merging and closing

Never merge a PR whose own new test fails, or with known problems a review listed. (Merging only the reviewed head, and writing what was run before the merge, are in the code standards.)

• #885 and #826 merged with their own new test failing. #846 merged with four known-wrong translations. Fix or file the problems first.
• Seen 1× (2026-10-05).

Keep the issue open until the reporter's case works. Reopen when they say it doesn't.

• On #876 the reporter couldn't reopen it themselves and had to file #888.
• Don't close as not-planned what the owner hasn't ruled on. #769 and #871 were reversed.
• Done right on 10-07: #1050, #1053, #1063 and #958 shipped a fix and stayed open for the reporter.
• Seen 1× (2026-10-05).

One commit, one goal.

• 653cb8ee bundled ZCode sign-in with WSL session deletion.
• 5464a362 capped six pages at 1200px when #860 asked for a speed display.
• A fix found on the way goes in its own commit. 1e0829b1 hid the ja/de fix inside a Usage change, beside two issues (#925, #928).
• 76dc5aed bundled a Routing animation fix, a scrollbar fix and bringing the GUI tests up to date.
• Seen 2× (2026-10-05, 2026-10-06).

A squash merge's message is written, not GitHub's default. Give gh pr merge --squash a --subject and a --body-file holding the PR's own change and what was run at the reviewed head.

• 708ae4144 (#852) and 81b7016e6 (#1028) each carry a ~520KB body. Their branches had main merged in (100 commits on #1028), and the default squash message lists every commit's message. #1028's own run record isn't in git at all: the body after its subject is other commits.
• Before pushing, check that git log -1 --format=%B | wc -c is a few KB at most.
• Seen 1× (2026-10-13).

Two sessions never commit from the same checkout.

• 10-09 12:10: 2018882e carries the davsync subject but holds the GTK font-DPI fix (#1371, 7 files). b86001d4, "#1371 GTK", is empty. be203728 is the real davsync change under the same subject. One session committed the other's staged files, because they shared the main checkout's index.
• Work in a worktree of your own (git worktree add --detach). After committing, git show --stat HEAD must list your files and only yours. An empty commit means another session took your index.
• 10-12 21:25: 3e88fc09d carries the auto-mode subject but holds the Get started card; the auto-mode change is 24350a372. A review reads git show --stat against the subject, not the subject alone (#1825).
• Seen 2× (2026-10-09, 2026-10-12).

Moved to the code standards

These were seen on three or more days. The rules and their evidence are in docs/code-standards.md; a review still checks each one.

• A fix reaches every sibling of the bug (10-05, 10-06, 10-07). Seen six times on 10-05. On 10-07 cd8d64a9's sub2api balance URL was left out of the hint shown after a failed read.
• A change re-derives everything built on what it changes (10-05, 10-06, 10-07). On 10-07: click-scroll red since 5d871723, four GUI tests left red by page-only test runs, and ee2588c2's test race (3e38bd76).
• A test that fails on main is a bug to fix now (10-05, 10-06, 10-07). On 10-07 dcbae2ef, 76d82528, b2345078, 07e246e5 and e95852e3 each shipped past "fails the same on origin/main".
• Merge only the reviewed head, and write what was run before the merge (10-05, 10-06, 10-07). On 10-07 #1037 merged with no record.
• Real use before building on it, or telling users it works (10-05, 10-07, 10-08). Acceptance criterion 5. On 10-08 bbf99143 "not tried on Linux" broke Linux vet with the WSL2 distro there, and #1185 was reviewed on macOS only.
• GUI tests in both engines, at narrow widths, in every language (10-05, 10-06, 10-08). Acceptance criterion 4. On 10-08 aff4f9f2's zh-TW missed later strings, and usage-ledger was red ~7h after 586f2bce.
• A fix reaches every sibling was missed again on 10-08: #1239 fixed Air 9h after a8290d27, and 173ab71e fixed Gemini bodies after #950. Done right on 10-08: a8290d27 itself, 9fcd332b (cmd.exe quoting at one choke point, tried on the Windows box), 74a809dc with 84f37ac9 (re-derived account keys), #1271/#1275 (built-in and plugin together), 3aeb7d1b (a refusal scoped to the model) and #1288.
• 10-09/10-10, criteria still missed.
  • Sibling: 12c0550a said "Only aBlock had this" while #1445 (open for an hour) listed the class. 64179e73, 13356b9d and d7186e51 followed.
  • The suite: 2da8fbe4 ran six packages, not go test ./.... Its testenv exec broke internal/proc's TestNoCommandBypassesProc, red ~50 min until b934d64a.
  • Real use: a3a8e257 shipped GitHub plugin icons without loading the real list, and the owner found none showing (52f6b3a3). aa82288f's Volcengine plan windows went out "not tried against the real API". It was said honestly, and #1427 stays open.
• Whatever a test starts ends when the test does (10-06, 10-08, 10-11). On 10-11 a CLI-identity ask outliving its test wrote into the next test's folder (#1524). 4ee3a76b fixed it in the gateway's TestMain only; 9c377231 fixed it for every binary through testenv's marker, but the gateway isolates itself without testenv and never set the marker, so #1524 came back on CI (a349de82). A guard test now fails if the gateway's TestMain doesn't set it.
• A timing test leaves room on both sides (10-08, 10-09, 10-11). On 10-10 8424a004 assumed two concurrent handlers let go in order (3a1ade97). On 10-11 TestMaxRPM and the queued-stream tests (above).
• A failed read, parse or hash means "unknown" (10-05, 10-11, 10-12, 10-13). On 10-12 7a9490638's IP allow list let every address in when settings.json couldn't be read (7d9382db2). On 10-13 #2017's patch read an unreadable codex-autoreset.json as "no reset spent" (e297a8ab2). When a shared reader is wrong, fix it, not a strict copy beside it.
• Nothing that can block runs under a lock its reader needs (10-06, 10-08, 10-11, 10-13). 7cba316b9's timer went red under -race (a2afc1828); a fake Codex appended from two goroutines (a75de4c21).
• Another app's values come from that app (10-05, 10-06, 10-13). accc8c090 sent 127.0.0.1 where Claude Code sends localhost (e0ecf0161).
• A test named for a guarantee fails when it breaks (10-08, 10-12, 10-13). bbe8825c0 (9dd5917a4) and ggbdpq's 44b5b2b96 and e21b1dbc1 (cc93b566e) passed on the old code. Break it by hand, watch it go red.
• A test never reaches the network (10-08, 10-11, 10-13). TestPiPackage tested npm's latest pi (f230fa23b); tests waited 30s on models.dev (bde11c778).
• A tag waits for a green run, and a tag is not a release (10-05, 10-08, 10-12, 10-13). v0.1.1192/1193 tagged one commit; v0.1.1213 and v0.1.1218 never published (#1846). npm view --prefer-online a plugin version before naming it.
• A large commit has a body with what was run (10-08, 10-12, 10-13). ea3d480ed and 1cd814004 (community, no record); 3ccbe6bc1's "push notes" hid two red GUI tests (2dcc51417, ba0712888).
• 10-11 to 10-13, criteria still missed.
  • The suite: 6fa3109ea, 43d8a5633, 5617281f5 and cdb7b5ff7 ran chosen packages and broke guard tests elsewhere (494609866, 24146c031, c6de0d2a5).
  • GUI: 18e82adc6 left 8 GUI files red after a run cut short (054bdb8bb); e3a497d7a left gui-ja red ~3h (28aac86bc).
  • Sibling: 4ed379f11 left three deciders on the raw body (0c6752dbd); 90ac1f038 mended StepCode only (cb003877f).
  • Timing: #1874's TestUsageIngestTouchesNoLimit stamped its calls a minute ago and went red a minute after midnight UTC (1acdf3a8f).
  • Whatever a test starts: Devin's held installer (2ff4ecb42) and a Claude sign-in's claude auth login run (ebbe76420, #1802) outlived their tests.
• Done right on 10-11 to 10-13. 2dcc51417 found 3ccbe6bc1's reds the same evening. 5241de909 guards a caller's image URL at the one place all three draw paths download it. 86b2f4d05 fixed revoked sign-ins for every built-in at once. 8179ce768 says what it broke by hand. 097479dab fixed #1543 at its cause, and ebbe76420 forced each failure by hand instead of rerunning.
• Done right on 10-10/10-11.
  • 9e35eef3 fixed #1469 half an hour after it was filed, reproducing CI's numbers with an injected pause.
  • ede1e32a (#1508) audited all six pickers of the same kind and added a guard test that every one can be narrowed.
  • 64262a19 (Prime Agent), 9c11abbd (jcode) and b72e7d4a (#1539) were tried with the real binary in a sandbox, and each says what was broken by hand.
  • #1197 scoped Kimi's sampling rule to Kimi Code's endpoints and models.
• Done right on 10-09/10-10.
  • 229ce19a, 570b8fdb and ab72c8eb reproduced CI flakes by injecting a delay and fixed their cause.
  • 7f62dbf3 named the siblings the reporter's patch missed (Cursor, Muse, GUI, CLI) and covered them.
  • 9b51428e found the same hold-out in Factory's list.
  • #1454 stayed open until the reporter confirmed.
  • #1213 and #1282 merged at the reviewed head with the record seconds before.