<- all tokdocs
2026.10.18@dailyaisignal

The Codex "deny rules only" catch repeats a README note that context-mode's own code and Codex's live docs have outgrown

AI
Watch on TikTok

View on TikTok ->

The artifact is a 43.70-second vertical clip at 720x1280 in an MP4 container, video encoded as HEVC Main in yuv420p at 30 fps and 291,613 bps, audio as HE-AACv2 at 44,100 Hz stereo and 48,150 bps, overall bitrate 346,423 bps, file size 1,892,426 bytes. The extracted audio.mp3 runs 43.73 seconds. The post timestamp is 1792332011, which is 2026-10-18 14:00:11 UTC. The uploader handle is dailyaisignal, the channel nickname is "Daily AI Signal", and the audio is credited as "original sound" by "Daily AI Signal". The description is 518 characters: "Claude Code can stop drowning in tool output, with a catch. Does Context Mode work with Cursor and Codex? Not fully. It's a free add-on with public code that keeps raw tool output out of your context window, taking 315 KB down to 5.4 KB, and it brings back your file edits and tasks after the chat compacts. On Cursor the context its hook adds never reaches the model, and on Codex the routing only supports deny rules, so Claude Code is where you get the whole thing. #ClaudeCode #ContextMode #AICoding #Cursor #Codex". At capture on 2026-10-19 the metadata recorded 787 views, 25 likes, 4 comments, 4 reposts, and 18 saves. I read 12 of the 22 extracted frames (001, 004, 007, 008, 009, 010, 013, 016, 017, 018, 019, 022). No person appears in any of them. Frame 001 is a screenshot of the context-mode README under a cream card reading "CLAUDE CODE STOPS DROWNING IN TOOL OUTPUT", a red stamp "THERE'S A CATCH", and a box reading "99% SAVINGS". Frames 004 through 019 use a trading-card layout headed "CONTEXT MODE No. 037", labeled "SOURCE-AVAILABLE CODING ADD-ON", one stat per card: "56 KB" "What one Playwright snapshot costs your context" (004, 007), "40%" "Of your context gone after thirty minutes" (008, 009), "315 KB becomes 5.4 KB. 98% reduction." over "github.com/mksglu/context-mode" (010, 013), "SESSION MEMORY" (016), "CURSOR AND CODEX GET LESS" "Cursor's hook context never reaches the model, and Codex gets deny rules only." (017, 018), and "DENY RULES ONLY" (019). Frame 004's screenshot shows README badges "Hacker News #1 • 570+ points" and "Used across teams at Microsoft Google Meta Amazon IBM NVIDIA". The transcript is 145 words in 14 segments from 0.00 s to 42.76 s: a single narrator reads the description nearly verbatim and closes with "I'll find one of these every day."

Every figure in the audio is copied from the README, and the repo's benchmark file disagrees with one of them

The narration's numbers all sit in the first forty lines of the context-mode README as I fetched it on 2026-10-19: "A Playwright snapshot costs 56 KB. Twenty GitHub issues cost 59 KB." "After 30 minutes, 40% of your context is gone." and "315 KB becomes 5.4 KB. 98% reduction." Frames 010 and 013 even quote that last sentence over the repo URL.

The provenance is thinner than the delivery. The repo's BENCHMARK.md describes 21 scenarios built from "real outputs from popular Claude Code MCP servers, Skills, and dev tools". Its Part 1 subtotal reads "315 KB raw → 5.5 KB context (98% savings)", so the README's 5.4 KB and the benchmark's 5.5 KB differ by a rounding step. The number for all 21 scenarios is 376 KB of raw output to 16.5 KB of context, stated as 96%. Frame 001's "99% SAVINGS" and the audio's "98% reduction" are different rows of the same table: 99% is the Playwright snapshot line (56.2 KB to 299 B), 98% is the Part 1 subtotal.

Two figures have no method behind them. "After 30 minutes, 40% of your context is gone" appears in the README and in the maintainer's February 26, 2026 launch post with no measurement described. The "Used across teams at" strip in frame 004 is a row of shields.io badges whose links all point to "#". The "697k+" users badge reads from a stats.json file inside the repo. The Hacker News badge is partly checkable: item 47193064 showed 570 points and 107 comments when I fetched it; the "#1" rank is the maintainer's claim.

The Claude Code half checks out against Anthropic's live docs and changelog

The whole-thing claim rests on hooks, and every hook the README says the plugin registers (PreToolUse, PostToolUse, UserPromptSubmit, PreCompact, SessionStart, Stop) is in the hooks reference at code.claude.com as I read it on 2026-10-19. On the mechanism the video describes, the doc says a PreToolUse permissionDecision can be "allow", "deny", "ask", or "defer", that updatedInput "replaces a tool's arguments before it runs", that PostToolUse's updatedToolOutput "replaces the tool's result", and that additionalContext "passes a string from your hook into Claude's context window. Claude Code wraps the string in a system reminder and inserts it into the conversation at the point where the hook fired."

The CHANGELOG.md in anthropics/claude-code, last committed 2026-10-18 and topping out at 2.1.303, dates the pieces (dates are npm publish times): hooks released in 1.0.38 (2025-06-30), PreCompact in 1.0.48 (2025-07-10), SessionStart in 1.0.62 (2025-07-28), and "Plugin System Released" with /plugin install and /plugin marketplace in 2.0.12 (2025-10-09). One README detail fails this check: its Claude Code prerequisite reads "Claude Code v1.0.33+", a version that predates both hooks and the /plugin command the install step depends on.

The Cursor catch is confirmed by Cursor staff through June, and Cursor's docs still say the opposite

The audio says "on Cursor the context its hook adds never reaches the model." The README's Cursor section states it as "Known limitation: Cursor accepts additional_context in hook responses but does not surface it to the model" and cites a forum thread. Forum post 155689 was opened by Jim_Beno on 2026-03-23 against Cursor 2.6.20. Cursor's Mohit Jain replied on 2026-03-24 that the field "is accepted and validated by the hook runner" but "the value is never actually injected into the model's conversation context." The last update, on 2026-06-15, reports the same failure on 3.7.27 and 3.7.36.

Against that, Cursor's own hooks documentation, fetched 2026-10-19, lists additional_context for postToolUse, postToolUseFailure, and sessionStart as "Extra context injected into the conversation after the tool result." So the video's "never" rests on a March bug report that still reproduced in June, contradicted by current docs, with four months of silence since; I treat it as supported through June 2026 and unverified for today's Cursor builds.

The README's Cursor section also still says Cursor's sessionStart hook is "currently rejected by their validator", citing thread 149566, where Cursor's Colin answered on 2026-03-30: "This was fixed in v2.4." The video does not repeat that line, but it dates the text the video is reading from.

The Codex catch is the line that has aged out, and the repo's own code says so

The audio says "on Codex the routing only supports deny rules." It comes from a README note: "Codex PreToolUse routing currently supports deny rules only (blocks dangerous commands). It still needs upstream updatedInput support before context-mode can rewrite tool input; track openai/codex#18491." Three sources say the note is behind.

First, openai/codex#18491, filed 2026-04-18, is still open. But the issue author's last status update, on 2026-07-09, says "Both blockers from the original issue appear resolved upstream" and names the pull requests. Per the GitHub API, "Support PreToolUse updatedInput rewrites" (#20527) merged 2026-05-12, and "Default function tools into tool hooks" (#23757) merged 2026-05-23.

Second, Codex's live hooks documentation, fetched 2026-10-19, says PreToolUse "can intercept Bash, file edits performed through apply_patch, MCP tool calls, and other local function tools", that "To rewrite a supported tool call without blocking, return permissionDecision: "allow" with updatedInput", and that "To add model-visible context without blocking, return hookSpecificOutput.additionalContext." The docs also say "Hooks are enabled by default", while the README still tells Codex users to enable them behind a [features].hooks flag.

Third, context-mode's shipped Codex hook contradicts its own README. The header of hooks/codex/pretooluse.mjs on main reads: "Codex PreToolUse honors permissionDecision:"deny" on all builds, and permissionDecision:"allow" + updatedInput / additionalContext on codex-cli >= 0.141.0 (#845). Capability is detected at runtime by codex-caps.mjs; older builds fail closed (redirect → deny)." Codex 0.141.0 was published on 2026-06-18, four months before this video. So "deny rules only" describes Codex builds older than 0.141.0; on anything newer the plugin's own capability check switches to rewrites and context injection.

"Free" and "public code" are accurate, and the card's "source-available" label is the right word

The narration calls Context Mode "a free add-on" whose "code is public", and the card label reads "SOURCE-AVAILABLE CODING ADD-ON" in frames 004 through 022. All three hold. The repo's LICENSE file is the Elastic License 2.0, copyright "2026 Mert Koseoglu", which grants a "royalty-free" license to use, copy, modify, and distribute, but bars providing the software "to third parties as a hosted or managed service", circumventing license keys, or removing notices. ELv2 is a source-available license, which is the term the card uses. It was not always: the commit history shows "chore: switch license from MIT to Elastic License 2.0 (ELv2)" on 2026-03-03, and the first Show HN post from February described the project as MIT licensed.

"Free" holds for the plugin. The README's privacy section says "No telemetry, no cloud sync, no usage tracking, no account required" and argues that context optimization "should happen at the source, not in a dashboard behind a per-seat subscription." The homepage tells a different story. context-mode.com/pricing, fetched 2026-10-19, lists Free, Starter at "$20 /month or $200 a year", Pro at "$49 /month or $490 a year", Team, and Enterprise, where "A request is one model request, subagents included", for a hosted console that the README's /context-mode:ctx-insight command opens. The add-on costs nothing; the maintainer sells something else under the same name.

Claude Code already does part of this on its own, which the February framing predates

The problem statement the video reads is the README's, written for the February 2026 launch. Claude Code has changed since. The changelog entry for 2.1.261 (published 2026-09-04) adds "bashOutputMaxChars and taskOutputMaxChars settings to raise how much command and background-task output Claude receives inline before it is saved to a file, up to 128K characters". Version 2.1.265 (2026-09-08) adds "a 1 GB cap on tool results saved to disk", and 2.1.287 (2026-10-01) lists "Improved handling of large MCP tool results". I hit this behavior in my own session today: a 39.1 KB metadata dump came back as "Output too large (39.1KB). Full output saved to" a file path, with a 2 KB preview.

The "SESSION MEMORY" card in frame 016 ("File edits and tasks come back after the chat compacts") also overlaps native behavior. Anthropic's context window page describes the compaction summary as keeping "your requests and intent, key technical concepts, files examined or modified with important code snippets, errors and how they were fixed, pending tasks, and current work".

What context-mode adds, per its README, is sandboxed execution of code against tool output so only stdout enters context, plus an FTS5 index with BM25 retrieval over past session events, and Claude Code ships neither. But "drowning" describes the baseline as it stood in February, and a viewer on 2.1.303 starts from a drier place than the card implies.

Key Takeaways

  • Verified: The artifact facts are confirmed by ffprobe and the post metadata: 43.70 seconds, 720x1280, HEVC Main at 30 fps, 1,892,426 bytes, posted 2026-10-18 14:00:11 UTC, 787 views at capture on 2026-10-19.
  • Verified: Every figure in the audio (56 KB, 59 KB, 40% after 30 minutes, 315 KB to 5.4 KB, 98%) appears verbatim in the context-mode README, and the Hacker News post behind the badge showed 570 points and 107 comments on 2026-10-19.
  • Verified: The six hooks the plugin registers on Claude Code are all in Anthropic's live hooks reference (checked 2026-10-19); the changelog places hooks at 1.0.38 (2025-06-30) and the plugin system at 2.0.12 (2025-10-09).
  • Correction: The video says "on Codex the routing only supports deny rules." The README note it quotes is stale: Codex merged updatedInput rewrites on 2026-05-12, its live docs now document updatedInput and additionalContext for PreToolUse, and context-mode's own hooks/codex/pretooluse.mjs honors both on codex-cli 0.141.0 or newer (released 2026-06-18). Deny-only applies to older Codex builds.
  • Partial correction: "On Cursor the context its hook adds never reaches the model" is confirmed by a Cursor employee on 2026-03-24 and still reported broken on 2026-06-15 (Cursor 3.7.36), but Cursor's current docs say additional_context is "injected into the conversation". Supported through June 2026, unverified today.
  • Partial correction: "315 KB becomes 5.4 KB" is the Part 1 subtotal of the maintainer's own benchmark, which BENCHMARK.md states as 5.5 KB; the full 21-scenario result is 376 KB to 16.5 KB (96%). Frame 001's "99% SAVINGS" is the Playwright row, and the audio's 98% is the subtotal.
  • Unverified: "After 30 minutes, 40% of your context is gone" has no published method; the "Used across teams at Microsoft, Google, Meta" strip links to "#"; "697k+" users is a JSON file in the repo; the Hacker News "#1" rank.
  • Clarification: "Free add-on" and "the code is public" are accurate. The license is Elastic License 2.0 (switched from MIT on 2026-03-03), which the on-screen card correctly calls "SOURCE-AVAILABLE". The plugin needs no account; the same domain sells a hosted console at $20 and $49 a month.
  • Context: Claude Code now saves oversized tool output to a file by default (bashOutputMaxChars, 2.1.261, 2026-09-04) and its compaction summary keeps modified files and pending tasks; the "drowning" premise is the README's February 2026 framing.

Resources

Published October 18, 2026. Writeup generated from a favorited TikTok.