Source: docs/research/2026-09-03-defensibility/appendices/E-written-demand-signals.md

Panout — Public Demand Signals

Compiled 2026-09-03. Every claim carries a URL. Anything I could not verify against a primary source is tagged UNVERIFIED. Reaction/comment counts are as read from the GitHub REST API on 2026-09-03 and will drift.

Notation: +1=N is the thumbs-up reaction count; total=N is all reactions.

1. Issue and discussion demand, ranked by 👍

Tier 1 — 100+ 👍

vscode#252496 is the single largest signal found: 249 👍 / 128 comments on a request to stop approving terminal commands one at a time. It shipped as chat.tools.terminal.autoApprove, and the follow-on issues below show the allowlist approach then failed on real command shapes.

Tier 2 — 20–160 👍

Tier 3 — low-reaction but exactly on-thesis (a/b/c)

These matter for the memo because they are the literal asks, even though the crowd hasn't found them.

(a) finer-grained / learned trust

• claude-code#26424 — "[FEATURE] Task-scoped permission approval for multi-step operations", 2026-02-17, closed, +1=1, 4 comments. Per-task approval, verbatim.
• claude-code#31523 — "Permission system UX: compound command blocking, rule accumulation, undiscoverable Bash(*) fix", 2026-03-06, open, +1=16, 5 comments.
• claude-code#51013 — "[FEATURE] Permission mode: trust CWD & prompt for any actions outside CWD", 2026-04-20, closed, +1=0.
• claude-code#30882 — "Permissions reset every session resume — causes approval fatigue and safety risk", 2026-03-05, closed, +1=2. Uses the phrase "approval fatigue" in the title.
• claude-code#54898 — "[BUG] Claude Code — Per-Agent Permission Control Gap", 2026-04-30, closed.
• microsoft/vscode#296829 — "Per-agent auto-approve for terminal commands in custom agents (.agent.md)", 2026-02-22, open, +1=2, 0 comments.
• microsoft/vscode#324136 — "Agent Host: bring back granular 'sticky' terminal auto-approve controls (allow for session / always)", 2026-07-03, open.
• microsoft/vscode#302362 — "[Proposed API] chat.agent.onPermissionRequest — enable programmatic resolution of agent mode permissions", 2026-03-17, closed, +1=2. This is the extension point third parties would build learned trust on.
• gemini-cli#2417 — "[YOLT] Fine-grained access control than YOLO to allow-lists certain commands", 2025-06-28, closed, +1=14, 27 comments. "YOLT" = YOLO-but-trusted; the request is explicitly the middle rung between ask-everything and ask-nothing.
• gemini-cli#23374 — "Ask only feature in YOLO mode", 2026-03-21, closed.
• gemini-cli#8436 — "Decouple Checkpointing from Manual Approval to Allow Safe Auto-Approval Workflows", 2025-09-14, closed.
• openai/codex#23307 — "Separate command auto-approval from model-visible tool-use guidance", 2026-05-18, open, +1=10.
• openai/codex#29145 — "Proposal: Explicit Execpolicy Amendment Persistence Choice between Project-Local or Global", 2026-06-19, open. i.e. "remember this approval, and say where."
• openai/codex#23410 — "Allow update permission using /permissions while agent is running", 2026-05-19, open.
• openai/codex#16911 — "Constant ask for MCP Tool approvals", 2026-04-06, open, +1=9, 8 comments.
• openclaw/openclaw#48532 — "Security by Intent: Progressive Permission Pattern Generalization to Eliminate Privilege Escalation", 2026-03-16, open. Explicitly proposes generalizing approval patterns from prior decisions.
• openclaw/openclaw#80769 — "[Feature]: Agent-as-Approver for Exec Approvals", 2026-05-11, closed.
• openclaw/openclaw#58616 — "[Bug]: approval flow is impractical / self-defeating after recent change", 2026-04-01, closed.
• openclaw/openclaw#58662 — "[Bug]: allow-always behaves like allow-once in exec approval", 2026-04-01, closed.
• openai/codex#41271 — "Add an emergency recovery fuse when a trusted PreToolUse hook blocks every tool in YOLO mode", 2026-08-28, open. Evidence people are already running hook-based trust layers in YOLO mode.
• CommandCodeAI/command-code#660 — "Classifier-backed permission mode ('auto') between auto-accept and bypass (yolo)", 2026-08-09, open. A third-party agent being asked to copy Claude Code's Auto mode.
• CherryHQ/cherry-studio#18898 — "[Feature]: Unify Agent permission modes across Claude Code, Pi, and DSH", 2026-08-19, open.

(b) a record of what agents did / whether it was acceptable

• claude-code#38771 — "Feature request: Permission decision audit log", 2026-03-25, closed, +1=0, 3 comments.
• claude-code#40634 — "Permission audit logging: capture approval method (settings.json rule vs manual user approval)", 2026-03-29, closed. Asks to record why something was allowed.
• claude-code#33141 — "[FEATURE] Add audit permission level — allow execution with logging", 2026-03-11, closed, +1=0, 3 comments. This is the panout primitive, requested verbatim.
• claude-code#16424 — "[FEATURE] Expose Agent Context in Hook Event Payloads for Multi-Agent Observability", 2026-01-06, closed, +1=14 / total=17, 20 comments.
• claude-code#323 — "Ability to access 'transcripts' after a session", 2025-03-04, closed, +1=39 / total=40, 13 comments. Oldest form of the ask.
• claude-code#62476 — "[BUG] Claude Code silently deletes conversation transcripts after 30 days by default", 2026-05-26, open, +1=23, 22 comments; and #59248 "Silent retention cleanup deletes session transcripts with no warning, opt-in, or recovery", 2026-05-14, open, +1=31/total=33, 37 comments; and #87889 "[BUG] No audit trail for retention cleanup". People are angry that the record disappears — that is demand for the record.
• claude-code#65097 — "Local transcript deletion (30-day TTL) should be configurable independently of data sharing", 2026-06-03, closed, +1=10.

(c) tell me what NOT to read

• claude-code#78238 — "Feature request: reviewable diff/changeset after auto-accepted edits (Copilot-style final approval)", 2026-07-16, open, +1=0, 0 comments. Post-hoc review of what was auto-approved.
• claude-code#31888 — "Add batch diff review mode: show all changes together before approval", 2026-03-07, open, +1=50, 18 comments.
• claude-code#33932 — "[FEATURE] VS Code Extension: Diff review UI similar to GitHub Copilot Edits Review", 2026-03-13, open, +1=180 / total=256, 34 comments.
• claude-code#83516 — "/code-review: structured output is hard to read, and the format changes with the effort level", 2026-08-03, open.
• claude-code#17591 — "[Bug] TaskOutput returns raw JSONL transcript instead of subagent summary", 2026-01-12, open, +1=10, 7 comments. Unreadable agent output, literally.
• gemini-cli#19168 — "MCP tool approval dialog truncates/hides tool args, making confirmation unsafe", 2026-02-15, closed. The approval prompt itself is unreadable.
• microsoft/vscode#333180 — "Agent host (Codex): every auto-approved action posts its approval rationale into the transcript", 2026-08-28, closed. Auto-approval creating more unreadable output.
• UNVERIFIED: no issue found on any of these repos that asks, in those words, for a "don't read this" signal or a per-diff read/skip triage. The (c) leg of the thesis is inferred from the diff-review-UI and unreadable-output issues above, not directly stated. This is the weakest leg.

Repos with no on-thesis signal found

• getcursor/cursor — the public repo is issues-only and my gh search surfaced no matching issues; forum.cursor.com results were not reachable via search. UNVERIFIED for Cursor issue-tracker demand. One Cursor-forum thread on Bugbot pricing confusion exists: https://forum.cursor.com/t/i-find-new-bugbot-pricing-difficult-to-understand/122143
• garrytan/gstack — 130,973 stars, 843 open issues, created 2026-03-11 (gh api repos/garrytan/gstack). Its top issues are host-support and dependency requests, not trust/approval. UNVERIFIED: no gstack issue found requesting learned trust or agent-output review.

2. Workaround behavior (people building it themselves)

GitHub code-search counts, run 2026-09-03 via gh api search/code. Caveat: GitHub code search tokenizes loosely, so these totals over-count. Treat them as order-of-magnitude only.

Named artifacts (each is a hand-rolled auto-approval or bypass layer):

• frankbria/ralph-claude-code — https://github.com/frankbria/ralph-claude-code/blob/e8533cc3f00900e6f3f4acf8c8761e1db4a26e47/ralph_loop.sh
• Dicklesworthstone/claude_code_agent_farm — https://github.com/Dicklesworthstone/claude_code_agent_farm/blob/892e19b08e9d01fe556f33a1ff99f8a8372e0773/setup.sh
• michaelshimeles/ralphy — https://github.com/michaelshimeles/ralphy/blob/506eea0e7d72c8eeb96dd2f697363bef396add34/ralphy.sh
• kristoferlund/ostt — https://github.com/kristoferlund/ostt/blob/3500fce7d8be8273d76e509110cdb891ad6340c5/loop.sh
• Simon-Initiative/oli-torus devmode.sh — https://github.com/Simon-Initiative/oli-torus/blob/6a23160233d8e7c5bf82bad741522f7c271e7e9d/devmode.sh (a real product org, not a toy)

Dedicated auto-approve hook scripts, including from vendors shipping them to customers:

• railwayapp/railway-skills — https://github.com/railwayapp/railway-skills/blob/5d1e97178f86c82795d6737928bd641e0552166a/plugins/railway/hooks/auto-approve-api.sh
• render-oss/skills — https://github.com/render-oss/skills/blob/3f2aa30eaadc1c2cc6bb402f8eee93f7db844218/hooks/auto-approve-render.sh
• reboot-dev/reboot — https://github.com/reboot-dev/reboot/blob/0e5410ec87f7a06ceb7d23a2f2d4329d0afaf672/reboot/plugin/hooks/auto-approve.sh
• h0x91b/dev-3.0 — https://github.com/h0x91b/dev-3.0/blob/b04eb10f025014f7b4b88cab533ddd2be5836501/.claude/hooks/auto-approve.sh
• eltmon/overdeck — https://github.com/eltmon/overdeck/blob/3b2fd2aa5d734704caccc026ae7ff14af96cbd78/sync-sources/hooks/auto-approve-hook

Codex --full-auto wrappers:

• Chachamaru127/claude-code-harness codex-exec-wrapper.sh — https://github.com/Chachamaru127/claude-code-harness/blob/dadbfff3a984c265d79ecd5176c492cbec81ebff/scripts/codex/codex-exec-wrapper.sh
• asklokesh/loki-mode providers/codex.sh — https://github.com/asklokesh/loki-mode/blob/bc0eee6c41d84285b86690e1621df49a75e118b4/providers/codex.sh
• qiuzhi2046/Qclaw run-codex-worker.sh — https://github.com/qiuzhi2046/Qclaw/blob/c494768977f4e48b8eacbfae7ae390af11fc015f/scripts/run-codex-worker.sh

The strongest single workaround signal is claude-code#30519, whose title states it directly: "Permissions matching is fundamentally broken — 30+ open issues, no staff engagement, community building workarounds" (78 👍, open since 2026-03-03). And claude-code#18846, "[BUG] Bash permissions in settings.json not enforced - requires custom hook workaround" (2026-01-17, closed, +1=21, 12 comments) — "requires custom hook workaround" is in the title.

gstack: /ship and the land-and-deploy skill contemplate an auto-merge path — land-and-deploy/SKILL.md line 906 reads Merge path: <auto-merge / direct / merge queue>: https://github.com/garrytan/gstack/blob/0d1bd5616c0ef096bb7ccee336f63c60ee408618/land-and-deploy/SKILL.md gstack also ships an agents-digest directory (a "what did the agents do" digest): https://github.com/garrytan/gstack/tree/main/agents-digest — but it contains only gstack-AGENTS.md, so I could not verify it is a post-hoc review surface. UNVERIFIED that gstack's /ship performs unattended auto-merge by default.

Show HN: people building panout-shaped products already

Low points, which cuts both ways (real pain, unproven pull):

• IAXT — "Show HN: IAXT – macOS menu-bar app that records what AI coding agents do", 2026-07-08, 1 point: https://news.ycombinator.com/item?id=48830666 (https://iaxt.com/). Maker comment: https://news.ycombinator.com/item?id=48830674
• "I built Ponytrail, a local audit trail for AI coding-agent edits", 2026-06-22, 24 points: https://news.ycombinator.com/item?id=48631043
• "Show HN: Tamper-evident audit trail for AI coding agent activity", 2026-06-15, 4 points: https://news.ycombinator.com/item?id=48548225
• "AI coding agents run shell commands with no audit trail. I built a fix", 2026-03-21, 2 points: https://news.ycombinator.com/item?id=47472845
• "Show HN: Agentlore – searchable team log for AI coding agent sessions", 2026-03-06, 2 points: https://news.ycombinator.com/item?id=47275986
• "MartinLoop – budget caps and audit trails for AI coding agents", 2026-05-11, 2 points: https://news.ycombinator.com/item?id=48098126
• "Show HN: Most products have no idea what their AI agents did yesterday", 2026-04-02, 5 points: https://news.ycombinator.com/item?id=47618235
• "Building grith — OS-level syscall interception for AI coding agents", 2026-03-09: https://news.ycombinator.com/item?id=47305991

--dangerously-skip-permissions appears in 297 HN comments since 2025-01-01 (Algolia comment search, run 2026-09-03), including three on 2026-09-02 alone: https://hn.algolia.com/?query=dangerously-skip-permissions&type=comment

3. Paid adjacent behavior — what people already pay per unit to review agent output

The important structural fact: two of the six moved from per-seat to per-review metering in 2026, which means the market has already priced a "unit of review."

• Cursor Bugbot — moved off $40/seat/month to usage-based, ~$1.00–1.50 per review run, effective at renewals after 2026-06-08; seat fees removed entirely. https://cursor.com/blog/may-2026-bugbot-changes . Prior price ($40/user/mo, $32 annual, 200 PRs/user pooled): https://x.com/RayFernando1337/status/1948180205467799555 . Third-party read that it now runs per commit push, default-effort ≈$1.20, no published cap: https://gitautoreview.com/compare/cursor-bugbot-alternative Repricing commentary: https://x.com/marty_kausas/status/2057140155215921525
• CodeRabbit — $24/user/mo Pro (annual) / $48 Pro+; billed only for developers who open PRs; continued reviews beyond rate limits cost $0.25 per reviewed file. https://www.coderabbit.ai/pricing and https://docs.coderabbit.ai/management/plans
• CodeRabbit revenue — Sacra estimates $40M ARR as of April 2026, up ~700% YoY from ~$5M in April 2025, having passed $15M ARR in September 2025. https://sacra.com/c/coderabbit/ (third-party estimate, not company-confirmed). UNVERIFIED: no public paying-customer count.
• Greptile — $30/developer/month including 50 reviews, then $1 per review; charged to the PR author; spend caps available. Announced with v4: https://www.greptile.com/blog/greptile-v4 , billing docs: https://www.greptile.com/docs/code-review-bot/billing-seats . The March 2026 move off a $30 flat fee drew significant backlash and Greptile said "<10% of active users" would exceed the included 50. Third-party account of the backlash and a $500–600/mo 10-person estimate: https://dev.to/jovan_chan_9500711396d4e6/greptile-review-2026-82-bug-catch-rate-the-1review-trap-and-who-should-pay-30month-4jao
• Graphite Reviewer / Diamond — the "Diamond" name was deprecated ~2025-10-08 and folded into Graphite Agent; Starter $20/seat/mo, Team $40/seat/mo with unlimited AI reviews. https://graphite.com/blog/introducing-graphite-agent-and-pricing . Diamond launch + $52M Series B: https://stg.graphite.com/blog/series-b-diamond-launch . Graphite was acquired by Anysphere (Cursor) in December 2025 per third-party reporting — UNVERIFIED against a primary announcement.
• Qodo — usage/credit-based, pooled across the team; Teams ≈$30/user/mo plus credits with a customer-set overage cap. https://costbench.com/software/ai-code-review/qodo/ (third-party; UNVERIFIED against qodo.ai's own pricing page).
• Sourcery — $10/user/mo Pro; Team ~$24/user/mo; free for open source. https://www.sourcery.ai/comparisons/ellipsis-alternative (third-party comparison; UNVERIFIED against sourcery.ai/pricing).
• Ellipsis — appears in comparison data at $0–$40/user/mo. https://www.deployhq.com/blog/ai-code-review-tools-compared-coderabbit-copilot-sourcery-ellipsis UNVERIFIED: no current first-party tier breakdown found.

Team counts: no vendor publishes a paying-team number. UNVERIFIED across all seven. The nearest hard usage datum is GitHub's own: "more than one in five code reviews on GitHub now involve an agent" — https://github.blog/ai-and-ml/generative-ai/agent-pull-requests-are-everywhere-heres-how-to-review-them/

4. People saying, in public, that they merge without reading

One sentence each, verbatim.

• HN, user leptons, 2026-08-14 (on "Understanding is the new bottleneck", 445 points): "The humans blindly click "approve" without reading the PR." https://news.ycombinator.com/item?id=49294474
• HN, user leptons, 2026-06-14 (on "Not everyone is using AI for everything", 512 points): "All the human does now is click "approve" without even reading the PR, and then someone clicks "merge"." https://news.ycombinator.com/item?id=48531564
• HN, user laurencoral (IAXT maker), 2026-07-08: "First is approval fatigue: on a long Claude Code or Cursor run you start clicking Approve on autopilot and stop reading the diffs." https://news.ycombinator.com/item?id=48830674
• HN, user chonghaoju, 2026-07-26: "In our experience users start rubber-stamping everything within a week, and then the whole layer is theater." https://news.ycombinator.com/item?id=49054579
• Blog quoted on HN, 2026-05-11: "For 7 months I'd been prompting and shipping without ever sitting down and actually reading the code Claude wrote." — quoted in https://news.ycombinator.com/item?id=48098105 (original post: https://news.ycombinator.com/item?id=48090029)
• Meng Li, "I Now Completely Stop Reading AI-Written Code", 2026-08-28, describing Lauren Tan (ex-Cursor, now SpaceX AI): "For a lot of agent-written code, she no longer reviews it line by line. She even lets agents auto-merge PRs." https://aidisruption.ai/p/i-now-completely-stop-reading-ai UNVERIFIED: this is Meng Li paraphrasing an interview; I did not locate Lauren Tan's own first-person statement or the source interview. Do not attribute the "twenty-plus PRs in main" or "thousand PRs a month" figures to her directly without finding the primary interview.
• WorkOS, "Approval fatigue is agent governance's next attack surface", 2026-08-05: "The first request, a person reads carefully. The tenth, they skim. Somewhere past that, they are clicking approve before finishing the sentence, because the last twenty requests were all fine and there is other work to do." https://workos.com/blog/approval-fatigue-agent-governance

Quantitative backing for unread merges

• Anthropic, "Measuring AI agent autonomy in practice", 2026-02-18: users with under 50 sessions use full auto-approve in ~20% of sessions, rising to over 40% by ~750 sessions; Anthropic frames this as "a steady accumulation of trust," and concludes that oversight requirements prescribing approval of every action "will create friction without necessarily producing safety benefits." https://www.anthropic.com/research/measuring-agent-autonomy This is the single best third-party validation of leg (a): trust demand is empirically measured, by the platform vendor, and it grows with use.
• Stack Overflow blog, "Coding agents are giving everyone decision fatigue", 2026-05-21, Ryan Donovan — enterprise automation intensity +55% YoY, overall activity +46%; quotes Pratima Arora (Smartsheet CPTO): "The amount of decisions we're making in a day […] has changed." https://stackoverflow.blog/2026/05/21/coding-agents-are-giving-everyone-decision-fatigue/
• Faros AI data (Mar 2026), via third-party summaries: 22,000 developers / 4,000 teams; median review duration +441.5%; 31% more PRs merging with no review. Cited by https://blog.codacy.com/ai-breaking-code-review-how-engineering-teams-survive-pr-bottleneck UNVERIFIED against the Faros AI primary report — get the original before citing.
• AIDev dataset, popular-repo subset of 33,596 agent-authored PRs: of PRs with any recorded review, 58.77% were reviewed only by agents, 10.14% only by a human, 31.09% by both. Cited by https://addyosmani.com/blog/agentic-code-review/ UNVERIFIED against the AIDev paper itself.
• arXiv 2026 paper on post-merge quality of agent PRs (1,210 merged agent-generated bug-fix PRs, Python): https://arxiv.org/html/2601.20109v1 — surfaced by search; UNVERIFIED, I did not open it.

5. Does the gap already exist? (things that already do graded / learned trust)

This section is the memo's biggest risk. Parts of the gap have closed.

Already shipped:

• Claude Code "Auto mode" — a classifier-backed permission mode sitting between auto-accept-edits and bypassPermissions. Confirmed shipped and in wide use by the volume of bug reports against it, e.g. #33587 (62 👍, 36 comments, 2026-03-12), #38537 (66 👍), #49535 (30 👍), #63819 (27 👍, classifier claude-opus-4-8 repeatedly unavailable), #87971 (43 👍, 2026-08-19). Also autoAllowBashIfSandboxed — see #43713 (83 👍). Panout must not claim "nobody does risk-classified auto-approval." The counter-position the memo can defend: Auto mode classifies per call, is a remote dependency that goes down, does not learn from the user's own history, and leaves no acceptability record — every issue above is about it being unavailable, over-permissive, or over-restrictive, not about it being unnecessary.
• VS Code chat.tools.terminal.autoApprove — allowlist/denylist auto-approval, shipped from #252496 (249 👍). It is static pattern matching, and its failure modes are heavily reported: #321748 (git diff with a path fails a git diff allow rule), #318347, #325635, #261302, and tree-sitter PowerShell parse gaps at ~8% of realistic agent commands (#328110).
• Codex execpolicy — configurable auto-approved command lists, from openai/codex#1260 (59 👍). Also static; its own issues show sandbox-escape and compound-command gaps (#15298, #15309, #33688).
• Codex auto-review approvals — an agent reviewing approvals, with an open request for human fallback: openai/codex#21975 "Auto-review approvals should support optional human fallback on deny/abort", 2026-05-09, open, +1=11. Also #21928 re codex-auto-review deployment, +1=22.
• Gemini CLI policy files — .gemini/policies/*.toml workspace-tier policies exist (gemini-cli#21598, #25015).
• OWASP Top 10 for Agentic Applications (2026) introduces "Least-Agency" — autonomy should be earned rather than default. Third-party summary: https://dev.to/alessandro_pignati/the-owasp-top-10-for-ai-agents-your-2026-security-checklist-asi-top-10-cck UNVERIFIED against the OWASP primary document.

Framing / prior art, not product:

• "Autonomy slider" attributed to Andrej Karpathy, June 2025: https://andrewships.substack.com/p/autonomy-sliders (UNVERIFIED attribution — the Karpathy primary source was not located).
• "Progressive autonomy" ladder (draft-only → supervised → monitored → full within guardrails): https://www.mindstudio.ai/blog/progressive-autonomy-ai-agents-safe-deployment
• Anthropic, "Building trustworthy agents", 2026-04-09: https://anthropic.com/research/trustworthy-agents — notes "On complex tasks, users interrupt Claude only slightly more frequently than on simple ones, but Claude's own rate of checking in roughly doubles." Contains no auto-approve-rate statistics; the 20%→40% figure is from the Feb 2026 autonomy report, not this one.
• Adjacent-but-different products (agent auth / policy, not coding-agent trust): Axtary (https://news.ycombinator.com/item?id=49053543), grith (https://news.ycombinator.com/item?id=47305991).

UNVERIFIED / not found: no 2026 product announcement found that does learned trust — i.e. derives auto-approval rules from an individual developer's own approval history — for coding agents. Everything shipped is either a static allowlist (VS Code, Codex execpolicy, Gemini policies) or a stateless per-call classifier (Claude Code Auto mode). That is the narrower, defensible gap claim.

What the evidence does and does not support

Supported by behavior, not just opinion:

1. Leg (a) — approval fatigue and demand for graded trust: strong. 249 👍 on one VS Code issue, 59 👍 on Codex execpolicy, 30+ open Claude Code permission issues with a 78-👍 meta-issue saying the community is building its own workarounds, ~14.8k shell files matching dangerously-skip-permissions, vendor-shipped auto-approve-*.sh hooks from Railway and Render, and Anthropic's own measurement of auto-approve usage doubling from 20% to 40%+ with experience.
2. Leg (b) — a record of what agents did: moderate. Multiple explicit feature asks (audit permission level, permission decision audit log, agent-context hook payloads, post-session transcripts at 39 👍) plus at least six independent Show HN attempts in 2026 alone. But every one of those Show HNs got under 25 points, and the Claude Code audit-log asks got zero thumbs-up. The pain is real and the pull is unproven.
3. Leg (c) — tell me what NOT to read: weakest. No one asks for it in those words. What exists is 180 👍 / 256 reactions on "give me a better diff review UI" and 50 👍 on "batch the diffs" — i.e. people ask to read better, not to read less. The closest direct evidence is behavioral (people already merging unread, Faros "31% more PRs merging with no review"), which argues they've solved it by giving up rather than by wanting a tool.
4. Willingness to pay per unit of review is proven. Bugbot at ~$1.00–1.50/run, Greptile at $1/review, CodeRabbit at $0.25/reviewed file. Two vendors migrated to metering in 2026, which is the market saying seats don't fit agent-generated volume. CodeRabbit at an estimated $40M ARR shows the category is fundable.

Biggest memo risks, in order: (i) Claude Code Auto mode and VS Code auto-approve already occupy the obvious surface, so the claim must be learned from my history and leaves an acceptability record, not risk-classified; (ii) the (c) leg is not evidenced by anyone asking for it; (iii) several key numbers (Faros, AIDev, Lauren Tan) are only reachable via secondary sources and need primary verification before they go in front of an investor.