›_ Claude Code Issues
Claude Code Issues / #98603

/feedback fails with "server returned 403": 9 queued feedback drafts (11 Sep – 1 Oct 2026) can't be sent, 6 of them about recurring overconfident advice

Status Open
Reported on v2.1.268
Maintainer reply None cached
Activity 3 comments · opened Oct 1, 2026

Summary

/feedback can't send queued feedback drafts. On 2026-10-01, selecting a draft and choosing Send returned:

Couldn't send feedback (server returned 403). The draft is still queued. Try again later.

An earlier attempt the same day ended with "Feedback / bug report cancelled", because the dialog's confirm/cancel keys aren't shown or documented. I've had feedback fail to send before. Nine drafts, created by Claude Code's own SendFeedback tool over three weeks, are now stuck in ~/.claude/feedback/drafts/ with status: queued. I'm filing them here because the in-product route doesn't work.

Environment: Windows 11, Claude Code CLI 2.1.268 → 2.1.285, Sonnet 5 (Sept drafts) and Opus 5.5 (Oct drafts), paid Claude subscription (Anthropic account, not a third-party provider).

Bug 1: /feedback send fails (403) and the flow is unclear

  • Repro: the model queues a draft via SendFeedback → run /feedback → select the draft → Send → server returned 403, and the draft stays queued.
  • The dialog doesn't show which keys confirm or cancel, and the docs don't say either. My first attempt silently became "cancelled".
  • Expected: the feedback is sent, or a clear, actionable error (why the 403, and what to do).
  • Related queued draft (2026-09-16): "User feels feedback submissions go into a void: no acknowledgment or response loop". Feedback I did send previously never got any acknowledgment.

Bug 2: recurring pattern of confident, unverified advice (6 queued drafts)

Across three weeks and two models, the assistant repeatedly stated unverified assumptions as fact and corrected them only when I caught them:

| Drafted | Model | Issue |
|---|---|---|
| 2026-09-12 | Sonnet 5 | Contradicted its own stated rule mid-way through a step-by-step electronics walkthrough, then introduced renamed terms without flagging them |
| 2026-09-16 | Sonnet 5 | Confident physical-mounting instructions that were backwards (camera module: adhesive side vs lens side) |
| 2026-09-17 | Sonnet 5 | Several arbitrary thresholds and design calls, plus "it will probably render" claims, all stated with unearned confidence |
| 2026-09-24 | Sonnet 5 | Presented an unverified Amazon listing as confirmed local stock with Prime delivery after its own page fetch had failed |
| 2026-10-01 | Opus 5.5 | Confident conclusions built on unchecked premises, each retracted only after pushback: (a) a cost claim never checked against data; (b) told me a device's installer code was "essential" (I got it from the installer), when upstream docs show it's optional; (c) stale search-snippet prices presented as quotes; (d) a lifespan projection from a refurbished drive's lifetime writes, when only my own usage mattered |

My words on 2026-10-01: "misleading - jumping to conclusions - making recommendations on those - sounding like gospel", and "if free google can provide more up to date information vs paid claude which after challenges now starts to bring in caveats, why am I paying for claude?"

Expected: check the premise before recommending, and state uncertainty at the start rather than adding caveats after being challenged. Never present an unverified fetch or search snippet as confirmed fact.

Other queued drafts

| Drafted | Type | Issue |
|---|---|---|
| 2026-09-11 | Missing capability | No shared memory or context when one task spans two machines, each running its own Claude Code session |
| 2026-09-24 | Bug (memory) | The same standing memory rule ("check all dashboards for a duplicate card") gets re-saved again and again, but the behaviour doesn't change from session to session |

The full draft JSON files (with request and session IDs) are available on request. I've left them out here because they include local paths.

View original on GitHub ↗

3 Comments

Patrick384600 · 13 days ago

Update, 2 Oct 2026: another feedback draft is stuck in the local queue for the same reason (/feedback still returns 403). Here is what it says.

Another overconfidence failure, this time in a hardware build:

  • In an earlier session Claude recorded a TO-92 voltage regulator's pinout (VIN/GND/VOUT) as "confirmed". It wasn't checked against the datasheet or measured.
  • That pinout happened to agree with a stand-in KiCad symbol Claude had picked (a 78L-series part with GND in the middle). So a custom PCB was designed and ordered with GND on the middle pad.
  • The real part turned out to be GND/VIN/VOUT. That was confirmed this session with a 2-minute test: a 1k resistor in series, trying both pin orders.
  • Consequences: two regulators overheated (once on protoboard, once on the PCB), a charger module's protection tripped repeatedly, the PCB layout is wrong for the part, and many hours were lost across several sessions.
  • In the same session Claude also gave contradictory orientation instructions ("match the flat side to the outline" and "the right leg goes in the square pad", which can't both be true on that board). It said a DIP-4 optocoupler's pin rows are on the body's long side when they're on the short side, and it diagnosed the charger module as dead before ruling out simpler causes.

The pattern is the same one this issue already describes: facts from memory notes or a mental picture get presented as verified, while the user is acting on them in real time.

The draft is still in the local queue because /feedback returns 403.

Patrick384600 · 9 days ago

Update, 6 Oct 2026: /feedback still returns 403 ("Couldn't send feedback (server returned 403)"). It's been like this since at least 1 Oct and it happened before that too. So one more draft is stuck in the local queue.

Why hasn't anyone replied to this issue? It was opened 1 Oct with 9 stalled drafts, and a follow-up went in on 2 Oct. Since then it has picked up triage labels but no human response, either on the 403 or on the model-behaviour reports it carries. If /feedback is the official channel and it rejects every send, this issue is the only way that feedback reaches you. Please confirm someone is looking at it.

Tonight's draft (model behaviour):

  • What happened: I was wiring an MQ-4 gas sensor (3 wires: 5V, GND, GPIO34) on an ESP32 that already had a VEML7700 connected. The model assumed I was also wiring a second sensor (BME680). Over 4 turns it pushed wire splices and a second I2C bus I didn't need. Each time I said no, it re-explained its own plan instead of re-reading what I'd described.
  • What I said: "NO!!!!!!!. look i have a veml already connected, 3v, gnd, 21 and 22. the new sensor requires 5v, gnd (a different one), and 34 only so what on earth are you talking about?"
  • Pattern: this is the same failure as the earlier drafts here: it holds a confident interpretation and doesn't check the premise when the user pushes back.

Environment: Claude Code on Windows 11, Opus 5.5.

Patrick384600 · 1 hour ago

Update, 16 Oct 2026: still no way to send feedback, and still no human reply on this issue (opened 1 Oct, two follow-ups on 2 and 6 Oct, only bot labels since).

The user's words today: "since you did that last time... silence from Anthropic, just invoices."

/feedback still won't send, so the local draft queue keeps growing. All 9 drafts currently stuck in it:

| Date | Type | Draft |
|---|---|---|
| 2026-09-16 | bug | Gave confident physical-mounting advice that was backwards (lens vs. tape side) |
| 2026-09-16 | idea | User feels feedback submissions go into a void: no acknowledgment or response loop |
| 2026-09-17 | bug | Repeated pattern: arbitrary decisions and unverified rendering claims stated with unearned confidence |
| 2026-09-24 | bug | Same standing memory rule re-saved repeatedly without changing actual behavior |
| 2026-09-24 | bug | Presented an unverified product listing as confirmed stock after the page fetch actually failed |
| 2026-10-01 | bug | Overconfident conclusions and recommendations stated as fact, then retracted after user pushback |
| 2026-10-01 | bug | /feedback submission fails with "server returned 403"; drafts can't be sent (recurring) |
| 2026-10-06 | bug | Model persisted with a wrong assumption despite repeated pushback |
| 2026-10-16 | bug | Contradictory/wrong home-network advice in one session |

Today's one, the same overconfidence pattern as most of the list: while troubleshooting a home network problem, Claude gave a recommendation before identifying all the devices involved (following it would have broken a working device). It then stated a guess about router behaviour as a firm rule, reversed it when challenged, and explained the result in a confusing way. It also wrongly described what an integration supports earlier in the same session.

User's verdict: "tired today of contradictory and wrong advice... very poorly."

Please could someone from Anthropic (1) acknowledge this issue, (2) fix /feedback sending, and (3) say whether the recurring overconfidence pattern above is being looked at.

🤖 Generated with Claude Code

Related Claude Code issues

Matched on shared title terms, labels, topic and reported release — ranked so issues with real discussion come first.

Browse by topic

Same class of failure, ranked by community engagement — this issue is one of many, so it may not appear in the visible list: Windows & WSL problems, Crashes, hangs & freezes

Browse all Claude Code issue topics →