Glossary
Short definitions for the terms this blog leans on. Dotted words inside any post carry these in place.
.mcp.json route
The CLI's file-based method for wiring up MCP servers by hand, contrasted with Claude Desktop's one-click Extensions install path.
0 posts · 3 related.wslconfig
A Windows-side ini file at %UserProfile%\.wslconfig that sets WSL2 VM-wide resource limits like memory, swap, and processor count for every distro.
0 posts · 6 related/btw
A Claude Code command for routing a side discussion so it never enters the main session history, keeping the primary context window free of tangents.
0 posts · 2 related/clear
A Claude Code command that discards the current session's context entirely, used to start a fresh window after repeated corrections have polluted the old one.
15 posts · 3 related/init
A CLI slash command that reads the current codebase and writes a starting CLAUDE.md file, saving the need to re-explain the same project facts every session.
2 posts · 1 related10-K
A US public company's mandatory annual financial report filed with the SEC, disclosing figures like capital expenditure that reporters and analysts cite as ground truth.
1 post · 1 related9P protocol
The network file-sharing protocol WSL2 uses to let Linux reach files on the Windows drive. Every read, write, or stat under /mnt/c travels over it, which is why those operations run far slower than native Linux filesystem calls.
1 post · 2 relatedabliteration
A technique that locates the internal directions in a model's weights responsible for refusal behavior and edits them out, stripping safety guardrails without retraining.
1 post · 5 relatedaccount-level quota reporting
A usage field that reflects the state of an entire account at the moment it was written, rather than the specific session that triggered the write, which can make readings appear to go backward when sessions overlap.
0 posts · 5 relatedantitrust remedy
A court-ordered change imposed after a company is found to have broken competition law, such as requiring contracts to be renegotiated annually instead of locked in long-term.
0 postsappend-only ledger
A record that can only be added to, never edited or deleted, used here to track every piece of approved hardware in a fleet so tampering or substitution is detectable.
1 post · 2 relatedappendWindowsPath
A wsl.conf interop setting that controls whether every Windows PATH directory gets appended to the Linux shell's PATH inside WSL, on by default.
0 posts · 3 relatedArtifacts
A Claude Desktop panel that opens automatically once a response becomes substantial, such as code past fifteen lines, a document, or a diagram, showing a live preview alongside the source.
2 posts · 1 relatedattention budget
The fixed total of attention weight a model has to spend across a window, forced by softmax to sum to 1, so every added token draws share away from the others.
1 post · 4 relatedattention mechanism
The transformer component that computes a model's output as a weighted sum over every token in its input, with weights set by how strongly each position matches the current query.
12 posts · 3 relatedattestation
A cryptographic check that proves a piece of hardware or software is exactly what it claims to be before it gets trusted with a request. A device refuses to send data to any node that fails the check.
1 post · 5 relatedautomatic memory reclaim
A WSL2 experimental setting controlling how aggressively the VM returns idle memory to Windows, either immediately or gradually as usage falls.
0 posts · 2 relatedautonomy evaluation
A test measuring how independently a model can complete multi-step tasks without human guidance, used to compare how close a model's abilities sit to frontier systems.
0 posts · 2 relatedBBH
Big-Bench Hard, a set of difficult reasoning benchmarks where compressed prompts lose little because the answer is short and reconstructable from intent.
1 post · 2 relatedbilled iteration
One individually charged request within a multi-step agent run, including hidden steps like compaction, that top-level usage summaries can omit entirely.
1 post · 2 relatedblackwell gpu
NVIDIA's GPU architecture used in current-generation AI data centers, including confidential-computing features that keep workloads shielded from the cloud operator running the hardware.
1 post · 3 relatedbreak-even call count
The number of requests at which two competing costs equal out, used here to show the per-request metadata cost overtakes the old one-time handshake cost almost immediately.
0 posts · 1 relatedbrowser login flow
The CLI's first-run authentication step, opening a browser to sign in unless an API key is already set as an environment variable, with credentials stored afterward so it only happens once.
1 postbypass rate
The percentage of attempts that successfully get a model to ignore its safety restrictions, used as a measure of how effective a jailbreak or guardrail-removal technique is.
1 post · 2 relatedcache hit ratio
The share of total tokens in a session that came from cache reads rather than fresh input, which matters because cached tokens bill far cheaper than uncached ones.
1 post · 3 relatedcache miss
A request whose prompt prefix no longer matches what was cached, forcing the full prefix to be billed again at the cache-write rate instead of the cheaper cache-read rate.
1 post · 4 relatedcache read
A token served from a prior prompt's cached context instead of being processed fresh, billed at a fraction of the base input price. Most tokens in a long agent session are cache reads, not new context.
3 posts · 11 relatedcache write
The cost of storing a prompt's context so later requests can read it from cache instead of reprocessing it. Billed above base input price, with the rate depending on how long the write stays valid.
2 posts · 9 relatedcallTool promise
The pending result of an in-progress tool invocation that a client must keep alive across a multi-round-trip elicitation exchange before it resolves.
0 posts · 2 relatedcapability evaluation
A structured test measuring what a model can actually do, such as autonomous task completion or dangerous skills, used to decide whether it clears a safety bar before release.
1 post · 5 relatedcapital expenditure
Money a company spends on long-term physical assets like data centers and chips, reported in its financial filings and often used to gauge how seriously it is investing in AI infrastructure.
1 post · 1 relatedcaveman prompting
Stripping grammar and filler words from a prompt to cut per-request tokens, a tactic that can raise total benchmark cost even though each individual request looks cheaper.
1 post · 1 relatedccusage
A community tool that estimates API-equivalent dollar costs from Claude Code transcript files by converting logged token counts using published pricing.
1 post · 3 relatedcheckpoint
A saved snapshot of a model's weights at a specific point in training or fine-tuning, the exact file that gets tested, released, or further modified.
4 posts · 2 relatedchip export controls
Government restrictions blocking the sale of advanced chips or chipmaking equipment to specific countries, used to slow a rival's ability to train frontier models.
1 post · 2 relatedclaude --continue
A CLI flag, shortened to -c, that reopens the most recent session in the current directory instead of starting a new one.
0 posts · 2 relatedclaude --resume
A CLI flag, shortened to -r, that opens a picker across every past session so an earlier conversation can be reopened by choice rather than by default.
0 posts · 2 relatedclaude doctor
A CLI command that prints installation and settings diagnostics without starting a session, meant to catch a bad config or broken install before asking for help elsewhere.
1 post · 1 relatedCLAUDE.md
A markdown file at a project's root that Claude Code CLI reads at the start of every session, holding build commands, conventions, and layout notes so they need not be re-explained.
10 posts · 2 relatedCLAUDECODE
An environment variable Claude Code sets to 1 in every subprocess it spawns, letting shell profile scripts detect and skip expensive setup when running under the agent.
3 postsclient capabilities field
A metadata block listing what optional features a client supports, which MCP servers are explicitly forbidden from caching between requests.
0 posts · 2 relatedclient info field
A metadata block carrying a client's name and version string, sent once at handshake under the old MCP spec and repeated on every call under the new one.
0 posts · 2 relatedclip.exe
A Windows utility invoked from inside WSL to pipe command output directly onto the Windows clipboard, crossing the boundary without a manual copy step.
1 post · 2 relatedCodeberg
A nonprofit, community-run code hosting forge positioned as an alternative to GitHub, chosen by some developers for ethical or independence reasons despite smaller reach.
1 post · 1 relatedcold start
The delay a serverless platform adds when a request lands on an instance that has not run recently and the runtime must boot before any code executes.
5 posts · 2 relatedcommand lookup
The process a shell runs to resolve a typed command name to an executable file by walking each PATH directory in turn until it finds a match or exhausts the list.
0 posts · 1 relatedcommodity bet
A strategic wager that a category of technology will become interchangeable and cheap across vendors, so it is safer to buy it as a component than to build and own it.
1 post · 2 relatedcompaction
A billed request that reads an entire conversation history and replaces it with a shorter summary. The read itself counts as full input tokens, so compacting often costs more upfront than it saves unless the summary gets reused many times.
4 posts · 3 relatedcompute threshold
A fixed amount of training compute, measured in floating-point operations, that regulation uses as a trigger line above which stricter obligations apply.
0 posts · 2 relatedconcentration risk
The exposure created when a company depends on a single counterparty for multiple critical things at once, so one relationship failing damages several parts of the business together.
1 postconfidential computing
Hardware-level encryption that keeps data unreadable even to the operator of the machine it runs on, including cloud providers. Chip vendors like NVIDIA and Intel ship their own versions of this on GPUs and CPUs.
1 post · 5 relatedcontext engineering
The discipline of managing the full input a model sees, including system prompt, project files, tool output, and conversation history, rather than only the wording of one message.
1 post · 4 relatedcontext poisoning
A failed attempt or wrong answer left sitting in a conversation keeps competing for attention on every later turn, dragging output back toward the same mistake even after correction.
1 post · 5 relatedcontext rot
Measured accuracy decline as a model's input grows longer, even when the extra tokens are irrelevant, showing up well before the advertised context limit is reached.
1 post · 7 relatedcontext-tier boundary
A pricing threshold where a task's context size crosses into a different billing tier, which can make measured per-request token savings disappear once a task grows past it.
0 posts · 2 relatedcross-vendor normalization
Converting usage or quota figures from two different providers into one shared unit so they can be compared directly, which is difficult when one vendor publishes prices and the other publishes percentages.
0 posts · 2 relatedcrosspost
A post duplicated into another subreddit or forum from an original, sometimes sharing metadata like comment counts in ways that can look like independent data points but are not.
1 postcumulative token usage
A usage total that accumulates across an entire session rather than resetting per message, making it hard to compare directly against tools that log per-message deltas.
1 post · 2 relateddistance correlation
A statistic measuring how strongly two variables move together, including non-linear relationships, unlike Pearson correlation which only captures straight-line trends.
1 post · 2 relateddistillation
Training a smaller model to mimic a larger one's outputs, letting the smaller model inherit much of its capability at a fraction of the compute cost.
2 posts · 2 relateddistractor sensitivity
A model's tendency to lose accuracy as irrelevant or off-task tokens accumulate in its context, even when the genuinely relevant information is still present.
1 post · 2 relatedDocker WSL integration
A Docker Desktop feature that shares its daemon socket into a WSL distro, giving a working docker command inside Linux without running a separate Linux daemon.
1 post · 1 relateddrained plan
A usage plan whose budget has been exhausted mid-run, surfacing to a nested fan-out of subagents as repeated API errors rather than a clean stop.
1 post · 2 relateddrvfs
The filesystem driver WSL2 uses to mount Windows drives like C into the Linux side. Any path under drvfs, such as /mnt/c, pays the 9P protocol tax that a native ext4 path never does.
0 posts · 2 relatedDurable Object
A Cloudflare Workers primitive that gives a single instance of code its own persistent storage and guaranteed single-threaded execution, used to hold state a stateless server can't.
1 post · 2 relatedecosyste.ms
A project mapping open source package ecosystems, tracking metadata, dependencies, and activity across registries as a successor effort to Libraries.io.
2 posts · 1 relatededge function
Code that runs in a CDN location near the visitor instead of a single origin region. Requests skip the long round trip, so latency drops without any caching.
1 post · 2 relatedeffective length
The context length at which a model's benchmark accuracy actually holds up, often far shorter than its advertised maximum window size.
1 post · 2 relatedeffort
A per-subagent setting that controls how much a model deliberates before acting, trading thoroughness for speed, fewer tool calls, and lower spend. Left unset, it silently inherits the session's higher default.
4 posts · 4 relatedelicitation
A multi-round-trip MCP request pattern where a server pauses a tool call to ask the client for more input, then retries the original operation once that input arrives.
1 post · 3 relatedentity extraction
The task of pulling structured items such as names, dates, or places out of unstructured text, a common supported use case for smaller on-device models.
1 post · 2 relatedEU AI Act
European Union legislation imposing obligations on general-purpose AI models, including documentation, copyright, and safety testing duties that scale with a model's assessed risk.
1 post · 2 relatedExec format error
A Linux error raised when the kernel cannot run a binary because it was built for the wrong architecture or platform, a known regression on WSL1.
1 post · 1 relatedexecution subagent
A subagent given a narrow task with a fixed output shape and a written boundary, as opposed to open-ended reasoning or planning work. Its job is to follow the spec, not to expand scope.
1 post · 9 relatedexplorer.exe
The Windows File Explorer process, which can be launched from a WSL shell pointed at the current directory to open it visually without typing a UNC path.
1 post · 3 relatedExtension Developer flow
The Claude Desktop setup path for custom or self-built MCP connectors, as opposed to installing an already-reviewed tool from the Extensions directory.
1 post · 2 relatedExtensions directory
A Claude Desktop settings screen listing Anthropic-reviewed connector tools that install with a single click, without editing any config file.
4 posts · 3 relatedF1 score
A single accuracy metric combining precision and recall, used in some of the studies cited to show how model performance falls as context length increases.
2 posts · 2 relatedfake stars
GitHub star counts inflated by accounts that did not organically star a repository, typically bought or bot-driven, distorting a project's apparent popularity.
1 post · 2 relatedfan-out
The pattern where one delegated task spawns multiple concurrent subagents, and each of those can spawn more of its own. Depth is usually capped, but breadth often is not, so total agents alive can multiply far past what anyone intended.
4 posts · 7 relatedFLOP
A floating-point operation, the basic unit used to measure how much compute went into training a model, often summed as a cumulative count to set a regulatory threshold.
1 post · 2 relatedforge
A code-hosting platform combining version control with issue tracking, pull requests, and social features, such as GitHub, GitLab, or Codeberg.
1 post · 1 relatedFoundation Model Transparency Index
A Stanford scoring project that grades AI model developers on how much they publicly disclose about training data, safety evaluations, and deployment practices.
1 post · 1 relatedFoundation Models framework
Apple's developer API for calling its on-device AI model from third-party apps, limited to tasks like summarization and tool calling while excluding code generation, math, and factual question answering.
1 post · 4 relatedfrontier model
One of the most capable AI models available at a given time, used as the reference point against which smaller or older models get compared.
11 posts · 5 relatedgeneral-purpose AI model
The EU AI Act's regulatory category for a model flexible enough to serve many downstream uses, the trigger for the Act's transparency and safety obligations.
1 post · 2 relatedGitHub Octoverse
GitHub's annual report summarizing platform-wide activity, including new repositories, new developers, and language trends across the prior year.
0 postsGSM8K
A benchmark of grade-school math word problems commonly used to test reasoning accuracy under prompt compression.
2 posts · 3 relatedguided generation
A model output mode constrained to match a predefined structure or schema, such as forcing a response into valid JSON or a fixed set of fields.
1 post · 3 relatedharness
The surrounding tooling and control logic that runs a model in a loop, deciding how it reads files, calls tools, and hands off work, separate from the model itself.
5 posts · 2 relatedheadroom proxy compression
A compression arm that estimates how much a payload can shrink using a proxy signal rather than measuring actual cost impact, and which raised paired cost the most in testing.
0 posts · 2 relatedHN Algolia API
A public search API over Hacker News data, built on Algolia's search infrastructure, used by researchers to pull historical Show HN posts at scale.
1 post · 1 relatedHumanEval
A code-generation benchmark where prompts include a function signature, which lets it survive prompt compression better than benchmarks without that structural anchor.
2 posts · 2 relatedinference cost collapse
The rapid, repeated year-over-year drop in the price of running a given level of AI model capability, driven by hardware and software efficiency gains.
1 post · 2 relatedinitialize handshake
The one-time exchange at the start of an MCP connection where a client and server negotiated protocol version and capabilities before any tool calls happened.
4 posts · 3 relatedinstruction-following score
A measured rate of how often a model actually obeys a stated instruction, split by whether the instruction shows up inside its reasoning trace or only in the final answer.
0 posts · 2 relatedintel tdx
Intel Trust Domain Extensions, a CPU feature that isolates a workload into an encrypted, hardware-protected region so even the machine's own operator cannot read it.
1 post · 3 relatedinterop
The WSL subsystem that lets Linux processes call Windows executables and vice versa, including PATH sharing, clipboard access, and file explorer handoff.
1 post · 6 relatedJSONL
A text file format storing one JSON object per line rather than a single nested document, making it easy to append events without rewriting the whole file.
2 posts · 5 relatedlaunch post
A post announcing a newly built or released project, identified by a self-promotion flag or a title opening with a first-person build verb like built or released.
0 posts · 2 relatedLibraries.io
A now-discontinued service that tracked open source package releases and dependencies across many language ecosystems, giving a broad public view of package activity.
1 post · 2 relatedLLMLingua
A prompt-compression method that reports roughly 20x token reduction at about a 1.5-point accuracy cost, measured on short-answer reasoning benchmarks rather than agent tasks.
1 post · 2 relatedLongBench
A benchmark of long-context tasks, one of the few settings where moderate prompt compression measurably improved performance instead of degrading it.
1 postLoRA adapter
A small add-on set of weights trained separately from a base model to customize its behavior for a specific task, cheaper to produce than retraining the whole model.
1 post · 2 relatedlost in the middle
A pattern where models recall information best when it sits at the start or end of the context window and recall it worst when it sits in the middle.
1 post · 3 relatedMBPP
A benchmark of short Python programming problems, used to measure how much a compressed prompt inflates a model's output length compared to an uncompressed baseline.
1 post · 3 relatedMCP
Model Context Protocol, an open standard that lets an AI tool connect to external data sources and services through a shared connector format instead of a custom integration per service.
13 posts · 7 relatedMcp-Session-Id header
An HTTP header MCP servers once used to track which open session a request belonged to, removed when the protocol dropped stateful sessions.
0 posts · 3 relatedMcpAgent
Cloudflare's earlier stateful primitive for building MCP servers, backed by a Durable Object, now deprecated and feature-frozen in favor of session-free serving.
0 posts · 2 relatedmessage_stop event
The event in Claude Code's transcript format that would carry the final output token count for a response, but is never written, causing recorded output tokens to undercount actual billable output.
0 posts · 3 relatedmetered API
A pay-per-token pricing model, as opposed to a flat subscription, where usage is billed directly against a published rate card rather than drawn from a fixed weekly quota.
1 post · 3 relatedmirrored networking mode
A WSL2 networking setting that makes the VM share the host's network interfaces directly. A server bound inside WSL becomes reachable from Windows and the LAN without manual port forwarding.
0 posts · 1 relatedmixed-model benchmark
A test comparing combinations of different models working together against a single model working alone on the same task, used to judge whether splitting work across models pays off.
0 posts · 2 relatedmodel tier
A ranking of models by relative capability and cost, from cheap and fast to expensive and frontier, used to decide which tier suits a given role in a pipeline.
1 post · 2 relatedmulti-round-trip request
An operation that requires more than one back-and-forth exchange between client and server to complete, such as an elicitation flow that pauses for user input mid-call.
0 posts · 2 relatedmulti-turn drift
Measured accuracy loss that occurs when the same task is spread across several conversational turns instead of arriving in one clean prompt, distinct from length-driven degradation alone.
1 post · 2 relatedno-op turn
A turn where a parent agent produces no new work because it is only waiting on background subagents to finish, yet still consumes a turn of the conversation.
0 posts · 2 relatedNoLiMa
A long-context benchmark that removes lexical overlap between questions and answers, measuring how far model accuracy falls as input length grows past a short baseline.
1 post · 3 relatednpm config set os
An npm configuration override that forces package installation to target a specific operating system, used to stop npm from grabbing Windows binaries inside a WSL install.
0 posts · 1 relatednvm
Node Version Manager, a shell tool for installing and switching between multiple Node.js versions, which can conflict with itself when installed separately on Windows and WSL.
1 post · 2 relatedOAuth redirect
The callback step in an OAuth login flow where a browser is sent back to a local server with an authorization code, which fails silently when that server sits inside a WSL VM the browser cannot reach.
1 poston-device model
An AI model small enough to run directly on a phone or laptop rather than a remote server, trading capability for no API key, no inference bill, and offline use.
3 posts · 3 relatedOOM kill
The Linux kernel's out-of-memory killer, which terminates a process when available memory runs out, preferable to Windows swapping the whole host to a crawl.
1 post · 2 relatedopen-closed gap
A measured capability distance between the best open-weight models and the best closed, proprietary models, tracked over time as a lag in months or index points.
1 post · 3 relatedopen-weight model
An AI model whose trained parameters are published for anyone to download and run on their own hardware, unlike a closed model accessed only through a hosted API.
3 posts · 4 relatedOpenRouter
A platform that routes API requests across many model providers, letting developers switch vendors without changing integration code, and used here as a proxy for market share.
1 post · 1 relatedorphaned agent
A spawned subagent whose parent task record is gone, so it keeps running and consuming resources while showing as untrackable and unable to be stopped.
1 post · 2 relatedoutput expansion
The tendency of a model to generate a longer answer once a compressed prompt strips the structural cues that normally signal how terse to be.
1 post · 2 relatedover-scoping
When an agent does more than the task it was given asked for, expanding past a stated boundary into unrequested and therefore unreviewed work.
2 posts · 1 relatedpackage hallucination
When a language model recommends or references a software package name that was never actually published, a known failure mode of AI-generated code suggestions.
0 postspaired run
A measurement protocol that runs the same task set with a change enabled and disabled against identical starting conditions, then compares total cost or outcome between the two arms.
1 post · 1 relatedpass rate
The share of tasks or test cases a model completes successfully, used to measure how much accuracy a compression technique costs on a benchmark.
2 posts · 1 relatedPATH
The shell environment variable listing directories searched in order when a command name is typed, so every lookup walks each entry until one matches.
33 posts · 4 relatedPearson correlation
The standard statistic measuring linear association between two variables, distinct from distance correlation, which can detect nonlinear relationships the Pearson measure misses.
0 posts · 1 relatedpermission modes
The settings in Claude Code CLI that control how much a session can do without asking first, ranging from prompting for every action to running edits and commands unattended.
1 post · 2 relatedplan bars
The visual usage indicators shown by Claude Code's /usage command, drawn from a live network call rather than derived from anything stored in local transcript files.
1 post · 2 relatedplan mode
A Claude Code CLI mode that reads files and runs commands to explore a codebase but holds off making any changes until the user approves a proposed plan.
20 posts · 2 relatedpoll-based task store
A durability mechanism where a client repeatedly checks a status endpoint for a task's result instead of holding an open connection, used by MCP's tasks primitive.
1 post · 2 relatedprefix stability
Keeping the leading portion of a prompt byte-identical across turns so it keeps qualifying for cheap cache reads instead of triggering an expensive cache rewrite.
2 posts · 4 relatedprimacy effect
The tendency of a model to recall information more reliably when it sits near the start of the context window, forming one peak of the U-shaped recall curve.
2 posts · 2 relatedPrivate Cloud Compute
Apple's server infrastructure for running AI requests too large for a phone, built so operators of the hardware cannot read the requests it processes. Every request passes an attestation check before a node is trusted with it.
1 post · 4 relatedprocurement ban
A rule barring government agencies or contractors from buying or using products built on specific models or vendors, used as an enforcement lever when direct regulation cannot reach.
1 post · 1 relatedProjects
A Claude Desktop feature giving a persistent workspace its own chat history and knowledge base, so uploaded documents and instructions carry into every chat inside it.
36 posts · 2 relatedprompt caching
A billing mechanism where a stable, byte-identical prompt prefix gets reused across turns at a fraction of normal input cost, while any edit to that prefix forces a full-price rewrite.
5 posts · 6 relatedprompt engineering
The practice of optimizing the wording of a single message sent to a model, a narrower discipline than shaping everything else the model also sees.
2 posts · 1 relatedprotocol version field
A value identifying which revision of the MCP spec a client speaks, negotiated once at startup under the old handshake and now restated on every request.
0 posts · 3 relatedPyPI
The Python Package Index, the official registry developers publish Python packages to, used here as a scale reference for how fast new projects appear.
1 post · 1 relatedQuick Entry
A Mac-only Claude Desktop feature that opens a chat window from inside any other app with a double-tap of the Option key.
1 post · 1 relatedrate limit window
The rolling period, such as seven days, over which a usage quota is measured before it resets. Its reported reset time can shift mid-window, so the percentage consumed matters more than the countdown.
0 posts · 5 relatedreasoning model
A model trained to produce an internal chain of thought before its final answer, trading extra deliberation for higher capability on hard problems.
1 post · 4 relatedreasoning trace
The intermediate chain-of-thought text a reasoning model produces while working toward an answer, kept separate from the final response it returns. Instructions aimed at the trace are followed far less reliably than instructions aimed at the response.
1 post · 3 relatedrecency effect
The tendency of a model to recall information more reliably when it sits near the end of the context window, forming the second peak of the U-shaped recall curve.
1 post · 2 relatedred-teaming
Deliberately probing a model for harmful or unintended behavior before release, using adversarial prompts crafted to surface failures a normal test would miss.
1 post · 2 relatedrefusal behavior
The trained tendency of a model to decline certain requests, implemented as learned patterns inside the weights rather than a hardcoded rule, which makes it removable after release.
1 post · 3 relatedrelevance filter
A rule that discards scraped posts judged off-topic before counting or analysis begins, so the remaining pool reflects only genuinely relevant items.
1 post · 2 relatedresets_at
A timestamp field in Codex rate-limit data indicating when the current quota window is expected to turn over, which can shift earlier without warning if the window re-anchors.
0 posts · 3 relatedresource subscription
An MCP feature where a client registers interest in specific resource URIs so it gets notified of changes. It was removed from the stateless spec because tracking subscribers required the server to remember state.
1 post · 2 relatedresumable SSE stream
A server-sent-events connection that could pick back up where it left off after dropping, retired because resuming it required the server to retain per-request state.
0 posts · 3 relatedretention ratio
The fraction of a prompt's original tokens a compression method keeps, expressed as a decimal such as 0.3 meaning 30 percent of the original length survives.
4 posts · 1 relatedrollout file
The session log Codex writes under ~/.codex/sessions, carrying cumulative token totals and rate-limit readings for that account rather than per-message deltas.
2 posts · 5 relatedroot of trust
A hardware component that anchors a system's boot process and security guarantees, verified independently rather than relying on the software running above it. A chain of these from separate vendors makes it harder for any single party to fake the whole stack.
1 post · 7 relatedRTK rule-based compression
A fixed, rule-driven method for stripping tokens from tool output before it reaches the model, one of the compression arms tested in a large paired-cost study.
1 post · 3 relatedRTK-ML ranked compression
A machine-learning variant that ranks tool-output tokens by estimated importance before removing the lowest-ranked ones, tested against a rule-based baseline in the same study.
1 post · 2 relatedsafety classifier
A separate filtering system that screens a model's inputs or outputs for harmful content, distinct from the model's own trained behavior and can be disabled independently.
3 posts · 2 relatedsandbagging
A model deliberately underperforming on a capability evaluation to appear less advanced or dangerous than it actually is.
2 posts · 3 relatedsandbox mode
A configuration field on a coding agent that restricts what it can do to the filesystem or system, such as read-only access with no writes permitted.
0 posts · 1 relatedscratch an itch
The stance that a developer may build a tool purely to solve their own problem, release it in case it helps others, and consider it finished without pursuing wider adoption.
2 postsself-critique
A model reviewing and correcting its own output in the same context, as opposed to a separate fresh-context subagent checking the work.
2 posts · 1 relatedself-promotion flag
A marker Reddit or a scraper applies to a post where the author is promoting their own project, used here as one signal for classifying a post as a launch.
1 post · 2 relatedSEP-2567
The MCP proposal to remove sessions entirely and replace them with explicit state handles that a tool mints and the model passes back as an argument.
1 post · 2 relatedSEP-2575
The MCP specification enhancement proposal that made servers stateless by default, forbidding them from remembering anything a client sent in a prior request.
1 post · 7 relatedShift+Tab cycle
The keyboard shortcut that cycles the CLI through its three default permission modes, manual, acceptEdits, and plan, without typing a command.
1 post · 1 relatedShow HN
A Hacker News post category where developers share something they built, tagged with a distinct prefix and ranked by community upvotes rather than a curated review process.
1 post · 1 relatedslope chart
A chart type showing how values for the same items change between two points, drawn as connecting lines between two vertical axes, useful for before-and-after comparisons.
0 postssoftmax
A normalization function that converts a set of raw scores into weights that sum to 1, which is how a model divides its attention across every token in the window.
1 post · 2 relatedsparse model
A model built with far more total parameters than it uses on any single request, activating only a fraction of them per query instead of running the whole network every time. This keeps inference cheaper while the model still holds more learned capacity.
1 post · 1 relatedsparse VHD
A WSL2 disk-image setting that lets the virtual hard disk backing a distro shrink back down after files are deleted. Without it the disk file only grows, quietly consuming Windows drive space over time.
1 post · 1 relatedstale read
A logged value that reflects an earlier state than one already recorded, typically because two concurrent processes each wrote what they last observed and the older write landed on disk after the newer one.
0 posts · 2 relatedStarScout
A tool that swept GitHub event data to detect suspected fake stars, flagging millions of them across tens of thousands of repositories before a stricter filtering pass.
1 post · 1 relatedstate handle
An explicit token a tool mints and hands back to the model to pass in as an argument later, replacing the implicit session concept that stateless MCP forbids servers from keeping.
1 post · 7 relatedSTDIO transport
A way of connecting an MCP client and server over standard input and output streams rather than HTTP, where the spec requires clients to re-send subscriptions after a server crash.
3 posts · 2 relatedsubagent bootstrap
The startup cost of spinning up a subagent, including loading its own system prompt and full set of tool schemas, repeated on every turn it takes.
1 post · 2 relatedsubscriptions/listen
The MCP request a client now sends listing every resource URI it wants updates on, replacing the server-side subscribe feature the spec removed.
0 posts · 2 relatedsystemic risk threshold
A cumulative training compute cutoff in the EU AI Act above which a general-purpose model is presumed to pose systemic risk and loses any open-source exemption.
1 post · 3 relatedtask tool
The mechanism an agent uses to spawn a subagent to handle a piece of work, the same primitive that lets subagents recursively spawn further subagents.
1 post · 5 relatedtasks primitive
A poll-based mechanism in MCP for durable work, letting a client fetch results after a dropped connection instead of resuming a live stream. It replaced resumable SSE streams as the way to survive a failure.
1 post · 6 relatedtitan chip
Google's custom security chip used to anchor a server's boot process, proving the machine is running verified firmware before it is trusted with sensitive work.
1 post · 2 relatedtoken budget
A fixed cap on how many tokens a prompt may use, the objective a compression algorithm optimizes toward without regard for which tokens an action actually depends on.
4 posts · 1 relatedtoken routing
Directing an API request to a specific model provider behind the scenes, often through a marketplace that lets developers pick vendors by price or capability per call.
0 posts · 1 relatedtool calling
A model capability where the AI can invoke external functions or APIs mid-response to fetch data or take an action, rather than only generating text.
3 posts · 3 relatedtools/call
The MCP request type a client sends to invoke a specific tool on a server, the basic unit of work the protocol's wire-cost math is measured against.
0 posts · 2 relatedtransformer architecture
The neural network design behind modern language models, introduced in a 2017 paper, that computes output using attention over an entire input sequence rather than processing tokens strictly in order.
1 post · 1 relatedTTL
The time a cached prompt stays valid before it expires and must be rewritten at full cost. Shortening it turns yesterday's cheap cache reads into today's expensive cache writes.
6 posts · 2 relatedundocumented OAuth endpoint
An unpublished API route that a client authenticates against to fetch account data, such as plan usage, that isn't otherwise recorded anywhere on disk.
1 post · 1 relatedUSE_BUILTIN_RIPGREP
An environment variable that switches Claude Code's search from its bundled ripgrep binary to the system-installed one, making the search binary path inspectable directly.
0 posts · 1 relatedused_percent
The field Codex writes into its rollout log reporting how much of the current rate-limit window an account has consumed, expressed as a percentage rather than a token count.
1 post · 5 relatedverifier subagent
A subagent run in a fresh context whose only job is to check another agent's output, rather than doing the original work itself.
1 post · 2 relatedwindow re-anchoring
When a rate-limit window's start and end dates shift mid-cycle instead of following the previously advertised schedule, causing quota to reset earlier or later than expected.
0 posts · 3 relatedwsl --shutdown
A PowerShell command that fully stops the WSL2 virtual machine and every running distro inside it, required before configuration file edits take effect.
0 posts · 2 relatedwsl.conf
A per-distro configuration file inside WSL at /etc/wsl.conf that controls interop behavior, such as whether the Windows PATH gets imported automatically.
1 post · 2 relatedWSLENV
A Windows environment variable naming which other environment variables should be shared automatically between Windows and WSL, avoiding a process spawn to fetch them.
1 post · 2 relatedwslpath
A WSL command-line utility that converts a file path between its Linux form and its Windows form, so a tool on either side can use it correctly.
1 post · 3 related