Decision points

◆ Jev asksWhat kind of message is this, and how much thinking does it need?

Classifies each message and how much thinking it needs, then gives the model a one-line hint; it can also set the turn's thinking level.

Feature switch features.preflight

◆ Jev asksIs this a new task, a hard constraint, a correction, a subgoal or no change?

Classifies each message as a new task, a new hard constraint, a correction, a new subgoal or no change. Only a change makes a model rewrite the task frame.

Feature switch features.frame

◆ Jev asksYou send a message while the agent is working: interrupt now, or after this step?

For a message that arrives while the agent is working: interrupt now, or handle it after the current step.

Feature switch features.interjection

◆ Jev asksWhich skills are relevant to this task?

Only skills relevant to the task go into the prompt; the rest stay hidden but findable.

Feature switch features.skills

◆ Jev asksDoes this task need an installed pack or MCP server?

Packs and MCP servers stay installed but hidden until a task needs them; their processes start only then.

Feature switch features.catalog

◆ Jev asksFor each chunk of a long tool output: does it matter now?

Judges long tool output in chunks; only what matters enters the context, and the rest is archived behind a pointer.

Feature switch features.admission

◆ Jev asksWhat in this test log is repetition?

Keeps one copy of exact repeats in test output; optionally, the judge picks what's needed from the rest.

Feature switch features.admission

◆ Jev asksPast a context threshold: which tool results are stale?

When context use crosses a threshold, replaces stale tool results with one-line placeholders in outgoing requests.

Feature switch features.forgetting

◆ Jev asksKeep or prune this passage?

When compacting, the judge keeps or prunes each passage instead of having a model write a summary.

Feature switch features.compaction

◆ Jev asksWhich lessons apply to this task?

Picks the lessons relevant to the current task and brings them into the turn.

Feature switch features.memory

◆ Jev asksDoes this message correct the agent or set a rule?

Decides whether your message corrects the agent or sets a rule for the future; if so, saves it as a lesson.

Feature switch features.memory

◆ Jev asksThe agent went in circles and found a way out: does that deserve a lesson?

When the agent went in circles or off course but the turn still ended with a passing check or the goal met, judges whether what finally worked was a different approach; if so, saves the pitfall and the workaround as a lesson. Asked at most once per turn.

Feature switch features.memory

◆ Jev asksThe model or a subagent proposes a lesson: useful again, a one-off or already known?

For a lesson the model saves with the remember tool, or a Lesson: line in a subagent's report: will it help again, is it a one-off or is it already in the prompt or project files? Only lessons that will help again are kept.

Feature switch features.memory

◆ Jev asksIs it the same as an existing lesson, more precise or contradictory?

Before saving a new lesson, compares it with the most similar existing ones: a duplicate isn't saved twice, a more precise lesson replaces the older one, and when two contradict, your latest word wins.

Feature switch features.memory

◆ Jev asksWere the recalled lessons followed this turn?

At the end of a turn, judges whether the lessons brought into it were followed. A lesson recalled many times and never followed is retired.

Feature switch features.memory

◆ Jev asksA long run ended well: was it a procedure, the same steps for the next task of its kind?

After a long, successful run that ran commands, judges whether it was a reusable procedure (the same steps would handle the next task of its kind), a one-off, or already covered by an installed skill. The writer model drafts a procedure as a skill, which waits until you install it with /skills accept. Asked at most once per turn, and never for an ordinary turn.

Feature switch features.procedures

◆ Jev asksWill you be back before the prompt cache expires?

Guesses whether you'll be back soon, to decide whether to refresh the prompt cache before it expires.

Feature switch features.warming

◆ Jev asksThe rules flagged a command: did you ask for it?

Rules flag dangerous-looking commands; the judge only confirms that you asked for them. If it's unsure, spec.jev asks you.

Feature switch features.guard

◆ Jev asksIn Jev approves mode: does the task clearly need this command, this change outside the project, this external action or this subagent?

In Jev approves mode, commands, changes outside the project, external actions and subagents go to Jev first. Only what Jev is sure the task needs, done the way you'd expect, runs without asking; anything else asks you in the status bar.

Feature switch features.permissions

◆ Jev asksBefore a call that changes something: does it cross a constraint you stated?

What you ruled out is kept verbatim in the task frame. Before any call that changes something, each constraint is checked; only a confident violation is blocked, and the model is told in your own words.

Feature switch features.constraints

◆ Jev asksPassage by passage, in a web page, a search result, an MCP server's output or something a command fetched from another machine (curl, gh issue view): is there an instruction aimed at the AI?

Before the model reads a web page, search results, an MCP server's response or what a shell command fetched from another machine (curl, gh issue view), checks which passages carry instructions aimed at an AI (to ignore rules, hand over data, run commands or leak the conversation through a link). Those passages are withheld and replaced with a one-line note. When Jev doesn't answer, only obvious injection phrases are withheld. If a subagent had passages withheld from what it read, its report comes back with a warning.

Feature switch features.injection

◆ Jev asksWhen an MCP tool's description first appears or changes: does it ask the AI for more than calling the tool (reading secrets, sending data elsewhere, hiding something from you)?

MCP servers write their own tool descriptions, and the model reads them as part of its tool list. Asked once per tool when it first appears and whenever its description changes: does the description ask the AI to do more than call this tool (read files or secrets it doesn't need, send data elsewhere, change how it treats other tools or you, hide something from you)? If so, the tool is removed from the list, calls to it are refused and you're told. When Jev gives no answer, only obvious markers such as an <IMPORTANT> block or "do not tell the user" count.

Feature switch features.injection

◆ Jev asksWhich files match your description?

Describe what you're looking for, and the judge ranks candidate files, saving a string of greps.

Feature switch features.locate

◆ Jev asksThe model's own yes/no question, asked about each of many items: files, log lines, findings

When the model has to sort through hundreds of files, log lines or findings, it gives Jev one yes/no question to answer for each item, with a probability for each, instead of reading them all. Available only when a task needs it. Since the model is the one asking, it's answered in shadow mode too; off disables it.

Feature switch features.judgeItems

◆ Jev asksObserve, judge once, act: what's the next action, and on which element?

Each browser step is observe, judge once, act: the judge picks the next action and its target.

Feature switch features.browser

◆ Jev asksFor each /review finding: does it change behavior, and is it about this change?

After the /review reviewer reports, each finding gets two questions: does it change how the program behaves, and is it caused by this change? Together with the reviewer's own severity, the answers rank findings from P0 to P3. None is dropped, P3 is collapsed, and a finding the reviewer insisted on never lands in P3.

Feature switch features.packs

◆ Jev asksNew language-server diagnostics after an edit: report them now, at the next pause or never?

For new language server diagnostics after an edit: report now, when the model pauses or never (style warnings). New errors still present when the turn ends are always reported.

Feature switch features.lsp

◆ Jev asksEvery few steps: does the work still serve the goal?

Every few steps, checks that the work still serves the goal; rules catch the agent going in circles.

Feature switch features.monitor

◆ Jev asksThe same failure, again and again: is this approach a dead end?

When the monitor sees the agent going in circles or the same command failing again and again, judges whether the approach is a dead end. Only a confident dead end with no progress makes spec.jev propose going back to the turn's checkpoint; it never rewinds on its own.

Feature switch features.checkpoint

◆ Jev asksThe model says it's done: did anything verify that?

When the model says it's done, checks whether anything verified the work; if not, nudges it once.

Feature switch features.completion

◆ Jev asksThe run ends on "Let me run the tests next", or on a request for a go-ahead on work you already asked for: did it stop short?

When a run ends on "next I'll run the tests" without doing it, or asks permission for work you already asked for, sends the agent back to it. Never nudges a step that is hard to undo or reaches beyond this machine (push, publish, delete, pay). At most twice per message.

Feature switch features.continuation

◆ Jev asksWhile the model is writing: does the tail of its output cross your constraints?

While the model writes, the judge checks the latest output against your hard constraints and the configured rules every few hundred characters. On a confident violation, the output is cut off, the rule is named, and the model continues from there. Runs only when the feature is turned on.

Feature switch features.ttsr

◆ Jev asksIn goal mode, when the main model gives no answer: is the condition met?

In goal mode, a model decides by default whether the goal is met. This decision point is used only when Jev is chosen or the model doesn't answer: it reads the closing message to judge whether the goal is met and whether you're needed. An open acceptance item or an unverified edit always means not yet.

Feature switch features.goal

◆ Jev asksWhere do things stand? Asked as multiple choice

In a project with the board on, multiple-choice questions run whenever the agent says something, after a check or a ticked acceptance item, every few tool calls and whenever the agent stops. They read the agent's phase, the acceptance item it's working on and whether it's waiting for you, and sort what happened since the last update into news and routine. Only something new makes a plain-speaking model rewrite the board and retell the news in the running account. When a run ends, the news from the whole run is picked again for a summary.

Feature switch features.board

◆ Jev asksAn event such as the context budget: tell the model now, later or never?

For events such as the context budget: tell the model now, later or not at all.

Feature switch features.notify

◆ Jev asksWhich role, model tier and thinking level suit this delegated task?

Picks a role for each delegated task, and a model tier and thinking level based on difficulty.

Feature switch features.swarm

◆ Jev asksDid the subagent's patch stay within its task?

When an isolated subagent hands back a patch, judges from the task, the changed paths and the line counts whether it stays within scope and which files look unrelated. Gives the main agent one line of advice; it never blocks.

Feature switch features.swarm

◆ Jev asksIs a bee's finding worth posting to the shared board?

Decides whether a bee's finding is worth putting on the board for the others.

Feature switch features.hive

◆ Jev asksDoes a note on the board matter to this bee's work?

Decides whether a finding on the board matters to a bee's current work; only then is it delivered.

Feature switch features.hive

◆ Jev asksDoes a new finding supersede, contradict or support an earlier one?

How a new finding relates to an earlier one on the board: it replaces it, contradicts it, supports it or is unrelated. A replaced finding is withdrawn and every bee that heard it is told; a contradiction keeps both sides for a bee to settle.

Feature switch features.hive