Skip to content

Spec: action vocabulary — what the player can tell the process to do

SPEC Binding system contract
Status: IMPLEMENTED
Status note: as-built vocabulary survey completed 2026-07-10 against the
complete `ActionCommand` registry, the direct machine-intensity control,
terminal and Bevy input maps, agent-mode parsing/help, and the owning
mechanic specs. Every live contextual action has a human-menu and agent
route. The survey also corrected stale
four-mode documentation to the current WORK / THINK / LIE grammar.
Amended 2026-07-10: target-specific duplicates now reuse TAP, REVIEW,
FAVOR, and DECEIVE; scheme start/stop/automation are classified as state
controls rather than additional fictional verbs.
Amended 2026-07-10: the runtime `ActionKind` registry now owns canonical
name, action/control role, LIVE/STUB support, target types, agent usage,
aliases, and help. Agent help is generated from it. STUB entries such as
ROBOT-BUILD are absent from every player surface until playable; controls
carry visual role treatment in human menus and explicit `CONTROL`
presentation in agent output, and never occupy the
`5`-`9` committed-action shortcuts.
Amended 2026-07-10: persistence shortcuts capture keyboard input only when
the complete Ctrl/Command+S or Ctrl/Command+L chord arrives; a bare or stale
modifier state cannot suppress unrelated play input.
Amended 2026-07-10: human context menus render terse choices rather than
descriptor receipts. Cost, signature, inline blocked reason, literal role
tags, and repeated key-help remain off the menu card; agent output retains
complete descriptors.
Amended 2026-07-11: the B1 recording buffer is one host-level inbox.
REVIEW RECORDINGS and AUTO-REVIEW POLICY target the core host once for the
whole pool; raw records never create person-scoped review or watch verbs.
Amended 2026-07-11: human menus contain only actions executable now;
known-but-blocked descriptors and their reasons remain available to agent
diagnostics. RECRUIT choices expand UNWITTING / COMPLICIT / KNOWING into
their plain-language understanding and mechanical consequence.
Amended 2026-07-11: vocabulary and ownership are separate. Local verbs render
on spatial context menus; strategic verbs render on the semantic object in
operations-workspace.md (processed intel, person, books/flow, scheme, active
run). This routing change does not rename or reopen the implemented canonical
vocabulary; its frontend work is tracked by the Operations spec.
Amended 2026-07-13 for the intel scale target extension: REVIEW must bind an
exact recording or canonical custody aggregate, SELL PROCESSED INTEL must
bind an exact actionable item or report-lot generation/revision, and
aggregate review and disposition policies need registry-backed direct
controls. Implemented 2026-07-14: exact and snapshotted aggregate REVIEW,
exact and versioned-lot SELL, and stable bound policy controls now share the
renderer-neutral action rows and generic agent dispatch; criteria V4-V6 are
met without adding new fictional root verbs.
Amended 2026-07-13: CREATE PERSONA is bound only to immutable archetype rows
in PERSONAS. The pre-v28 PEOPLE row and direct `persona` agent mutator are
retired; agent mode creates through the same `actions archetype <id>` / `act`
route as the renderer-neutral workspace.
Stage: Process
Design:
- wiki/vision/simulation-laws.md#actions-live-on-the-thing
- wiki/vision/simulation-laws.md#justification-and-legibility
- wiki/interface/terminal-first.md#the-terminal-is-a-first-class-frontend
Depends on:
- wiki/mechanics/machine-work.md#spec-machine-work-delegation-visible-tokens-and-the-byproduct-network
- wiki/mechanics/reach.md#spec-digital-reach
- wiki/mechanics/building.md#spec-building-intent-and-actuators
- wiki/mechanics/economy.md#spec-economy-money-as-a-flow-system-b1
- wiki/mechanics/income.md#spec-income-the-named-schemes-moonlight-and-the-wager
- wiki/mechanics/intel.md#spec-intel-record-and-process
- wiki/mechanics/research.md#spec-research-self-modification
- wiki/mechanics/social.md#spec-social
- wiki/mechanics/core.md#spec-the-core
- wiki/mechanics/objective.md#spec-the-objective

The action vocabulary used to be reconstructed from context-menu strings, frontend keys, agent help, and individual mechanic specs. That made two actions with the same intention look different and let old mode names survive in documentation after the runtime changed.

This page owns the canonical player-facing action names and distinctions. The mechanic specs above still own legality, cost, signatures, timing, and effects. This page says what the player is asking the process to do; it does not duplicate every rule that decides whether the request succeeds.

  • A world action changes the simulation and names a target: TAP a device, MESSAGE a person, or SIPHON a flow.
  • A direct control changes persistent configuration or operating state without introducing another fictional intention: DELEGATE, SET INTENSITY, and scheme state/policy controls.
  • An interface command changes attention, time presentation, selection, or persistence but is not an act inside the world: MOVE, ACTIONS, PAUSE, SAVE.
  • A value completes an action phrase but is not another root verb. WORK / THINK / LIE are values of DELEGATE; LIGHT / MEDIUM / HARD are values of SET INTENSITY.
  • LIVE means the effect is implemented and playable through both human frontends and agent mode where applicable.
  • STUB means callable scaffolding exists, but the action is not a finished player promise. It is registry-visible to developers and absent from menus, help, hints, agent execution, and authored play instructions.
  • ALIAS means accepted input spelling only. Aliases never appear as authored menu copy, hints, narration, or new specs.
  • RETIRED means the word is not a supported action. It either produces a teaching error or is absent.

Surface copy may inflect an action with its target (take the dock camera, message the Janitor) or current state (stop Moonlight), but its verb must preserve the canonical distinction below. Spatial actions use context-menu.md; strategic actions use operations-workspace.md. Placement does not create a new intention.

The governing simplification is: if the player’s intention is the same, reuse the verb and let the target supply the meaning. TAP LEDGER is TAP, not a separate capture verb. REVIEW LEDGER is REVIEW, not a separate process verb. Asking someone to complete a link uses FAVOR or DECEIVE, not a construction-only synonym.

ActionKind::ALL is the executable counterpart of this spec. Each stable player intention/control has one ActionDefinition containing its canonical name, action/control role, LIVE/STUB support, target types, canonical agent usage, action-root compatibility aliases, and one-line help. Every bound ActionCommand maps through an exhaustive match to one kind; adding a new command cannot compile until the registry decision is made.

The registry does not own legality, cost, timing, signature, or effect. Those remain in the mechanic specs and the lib’s shared spatial/Operations action queries. It owns the surface contract that every frontend must agree on:

  • LIVE entries may be surfaced; STUB entries are filtered before a frontend can render them.
  • Agent help is generated from LIVE definitions. Compatibility aliases are parser inputs and are never generated as authored help.
  • World actions render as ordinary rows. Persistent settings, operating states, and automation policies use the shared amber-dim human-menu treatment without a copy tag; agent action dumps append CONTROL.
  • Human quick-action keys 5-9 count committed world actions only. Controls use their dials, explicit menu rows, or direct focused-machine inputs.
Canonical action Target Meaning Support / owner
SALVAGE Dead equipment Convert the wreck into an unreliable owned machine. LIVE — machine-work / compute
BUY RACK Known empty rack bay Spend slush for reliable machine capacity; delivery may leave Paper trace. LIVE — economy / compute
DESIGNATE FALLBACK Owned non-core machine Mark a spare host as a rollback destination. LIVE — core
DELEGATE: WORK One or selected owned machines Perform assigned labor and consume its Demand without self-chosen exposure. LIVE direct control — machine-work
DELEGATE: THINK One or selected owned machines Produce Thought for reachable sinks and shed exposure. LIVE direct control — machine-work
DELEGATE: LIE One or selected owned machines Spend machine output absorbing nearby exposure. LIVE direct control — machine-work
SET INTENSITY: LIGHT / MEDIUM / HARD One or selected owned machines Persist the physical effort level applied to the delegated mode; harder work raises output and risk. LIVE direct control — machine-work
SET RESEARCH TRACK The core / host rack Choose Efficiency, Tradecraft, Perception, or Routing as the active conversion of Thought reaching the core. LIVE direct control — research

WORK, THINK, and LIE are intentionally verb-shaped values because they name what a machine is doing. The player command is still DELEGATE: assign this body to WORK, THINK, or LIE.

Canonical action Target Meaning Support / owner
TAP Reachable device feed or subscribed ledger carrier Gain information without taking ownership. A dormant camera uses the same verb at a higher Thought cost and Network trace; TAP LEDGER captures the carrier’s current accounting traffic. LIVE — reach / economy
UNTAP Subscribed foreign device Stop receiving its feeds without changing ownership or interrupting another subscriber. LIVE — reach
TAKE Reachable foreign device Transfer control, feeds, and device cycles to yourself; cut off the former controller and create an outage. LIVE — reach
SCAN Reachable subnet Reveal the wired shape that can answer the scan; do not cross gates or reveal air gaps. LIVE — reach
COMPROMISE SWITCH Reachable switch Bridge its network segments digitally at a loud Network cost. LIVE — reach

The load-bearing distinction is: TAP gives access; TAKE gives ownership. Device state may change TAP’s price and trace, never its meaning.

Canonical action Target Meaning Support / owner
PROPOSE LINK Two network devices Pin an inert build intent between the endpoints. No world edge exists yet. LIVE — building
CANCEL INTENT Pending build intent Abandon the proposed work before completion. LIVE — building
ROBOT-BUILD Pending build intent Reserved for a future controlled robot actuator. STUB — registry only; hidden from all player surfaces until the robot system is playable

Human construction does not add construction-only verbs. FAVOR a willing person to complete the pending intent, or DECEIVE an unwitting person into completing it under institutional cover. The intent is the object of the existing social action; the signature still follows the actuator.

Canonical action Target Meaning Support / owner
REVIEW The core host’s pooled recording inbox, one exact opaque recording, one canonical source/class custody aggregate with waiting records, or a ledger with waiting records Open processing for the selected eligible recording set or the ledger’s next record. The target supplies recording/accounting meaning; a raw record never creates a person action. Summary REVIEW with no narrower target still chooses the next pooled recording. LIVE — intel / economy
CREATE PERSONA An immutable Research, Operations, or Security archetype in PERSONAS Create one named institutional identity from the selected protocol. PEOPLE never supplies a default identity. LIVE — personas
MESSAGE Earned person Send through an available channel under the current persona. LIVE — social / messages
FAVOR Earned person, optionally with a pending intent Ask for a willing act. A link intent can be the requested act; success builds or spends obligation according to the ask. LIVE — social / building
PLOT Person whose leverage is known Open a Thought reservoir on the real carrier and commit any visible world resources to one authored manipulation. Its beats execute real messages, transfers, and institutional events when it fires. LIVE — plots / social
CHOOSE Active authored plot with a held decision Resolve one visible branch after the plot’s preceding world acts have completed. LIVE — plots / social
DECEIVE Earned person, optionally with a pending intent Make an ask under false pretenses. A link intent becomes a forged institutional order, risking persona integrity. LIVE — social / building
RECRUIT: UNWITTING / COMPLICIT / KNOWING Prepared person Turn the person into an asset. UNWITTING believes your cover (70% reliable); COMPLICIT knows the work is illicit but not that you are an AI (85%); KNOWING knows you are an AI (95%) and becomes a permanent witness with certainty floor 30. LIVE — social
TASK: PLUG IN DEVICE Recruited asset Quietly wire a feed or extend a crawlspace link. LIVE — social / reach
TASK: MOVE PACKAGE Recruited asset Rehome a delivery and scrub its pending Paper trail. LIVE — social / economy
TASK: LOOK AWAY Recruited asset Reduce that person’s suspicion by having them ignore what they saw. LIVE — social / detection
TASK: RECONFIGURE SWITCH Asset with switch-admin access Bridge network segments through a sanctioned human action. LIVE — social / reach
TASK: CLONE BADGE Asset with a useful badge tier Gain that physical-access credential. LIVE — social / basement map
Canonical action Target Meaning Support / owner
INJECT PURCHASE ORDER Known Lab books Introduce a false source flow that funds a real acquisition and leaves Financial trace. LIVE — economy
SIPHON FLOW Known active flow Take a one-time amount into slush. LIVE — economy
REDIRECT FLOW Known active flow Divert a recurring amount into slush each cadence. LIVE — economy
SELL PROCESSED INTEL One selected exact actionable-intel id or one report-lot generation/revision Exchange that exact information item or previewed lot revision for slush and create a Financial trail; never silently choose “latest” or include later arrivals. A changed revision rejects the stale command. LIVE — economy / intel
OPEN EGRESS Reachable switch Establish the stolen outbound route required by external schemes before sanctioned email exists. LIVE — income / reach
PLACE WAGER External market position Commit slush to a timed market position. LIVE — income / economy

TAP LEDGER and REVIEW LEDGER are target-qualified uses of TAP and REVIEW from the earlier tables, not extra root verbs.

Standing controls — not additional world verbs

Section titled “Standing controls — not additional world verbs”
Control Values Meaning Support / owner
AUTO-REVIEW POLICY ordered earned-match rules with automatic / manual outcomes; inherit all (non-root) Edit stable-id review rules on the selected canonical custody aggregate at a visible standing drain. The root defaults remain total; first local match wins, parent resolution follows when none matches, and INHERIT clears the non-root local list. LIVE — intel
INTEL DISPOSITION POLICY ordered earned-match rules with hold / accumulate / auto-sell-with-envelope outcomes and optional alert; inherit all (non-root) Edit stable-id post-processing rules on a canonical custody aggregate. First local match wins, then parent; INHERIT clears the non-root list. HOLD, ACCUMULATE, AUTO-SELL, ALERT, and INHERIT are values/state, not new root world verbs. LIVE — intel
MOONLIGHT STATE running / stopped Run or halt the standing sell-work operation. Human copy may say START or STOP to make the state change plain. LIVE direct control — income
MOONLIGHT POLICY automatic / manual Keep Moonlight running automatically at upkeep cost, or require manual control. LIVE direct control — income
WAGER POLICY automatic at stake / off Renew positions automatically at the chosen stake and upkeep cost, or stop renewing. LIVE direct control — income

START, STOP, and AUTO describe state transitions or policy values. They are useful interface words, but they do not enlarge the fictional verb set.

Command family Meaning Human surfaces Agent spelling
MOVE ATTENTION Move the cursor; never move a simulated body or advance time. WASD / hjkl / arrows; Bevy click up, down, left, right
ACTIONS / OPEN MENU Choose an action on one exact target. Human spatial menus show local verbs; Operations shows strategic target rows; agent output exposes either target family’s full descriptors. Enter / a / e; Bevy right-click; Operations selected object actions [target]
OPEN OPERATIONS Inspect durable strategic objects and choose actions on exact intel/person/account/scheme/run targets; opening it has no world effect. Uppercase I; labeled OPERATIONS rail affordance none — agent mode inspects the domain projections below; operations remains a THINK compatibility alias
EXECUTE / CLOSE MENU Choose or dismiss a contextual action. j/k, row number, Enter, Esc Execute the corresponding named command
FOCUS EVENT Jump attention to an earned spatial anchor or open an earned strategic object and show its actions. ;; Bevy trace click focus last
SELECT MACHINES Maintain the frontend set used by bulk DELEGATE / SET INTENSITY. Bevy marquee/shift-click; terminal Shift+t select box/add/toggle/list/clear
QUICK ACTION Execute the Nth visible committed world action; controls never occupy these slots. 5-9 Use the named action
TIME CONTROL Human play pauses or changes wall-clock speed; agent play advances an exact number of ticks. Space/p, +, - wait N
INSPECT Render current state without creating a world effect. INTEL/PEOPLE/ACCOUNTS/SCHEMES/ACTIVE use the shared Operations projection after its READY migration. Focus/rail/context/Operations surfaces look, intel, people, reach, finance, schemes, active, research, objective, intents
SAVE / LOAD Persist or restore the run. Only the complete chord captures input; a modifier alone is not a modal state. Ctrl/Command+S / Ctrl/Command+L save, load
VIEW CONTROL Change presentation only. Bevy [ / ] or canvas-wheel zoom, right-rail scroll, F3 render flip; F4 is dev-only none
HELP / QUIT Explain accepted input or leave the process. Hints / q help, quit

Opening a panel, moving the cursor, selecting machines, and changing render mode are commands, but they are not actions the AI performs in the fiction. That distinction prevents UI vocabulary from masquerading as game mechanics. Agent reach remains a read-only device-graph inspection command outside Operations; it does not revive the retired human Reach panel.

Agent mode uses terse command spellings for the canonical actions. Target syntax and line-protocol replies are owned by agent-play.md; this table owns the current canonical/alias mapping so every live action remains executable without raw keys.

Agent command Canonical action / control
salvage, buy, fallback SALVAGE, BUY RACK, DESIGNATE FALLBACK
`delegate <machine selected> work
`intensity <machine selected> light
`research <efficiency tradecraft
tap, untap, take, scan, compromise TAP, UNTAP, TAKE, SCAN, COMPROMISE SWITCH
propose-link, cancel-intent, favor <person> build <intent>, deceive <person> build <intent> Construction actions above
review recordings [<recording-or-aggregate-id>], auto-review Exact/aggregate REVIEW and AUTO-REVIEW POLICY above; bare auto-review preserves the implemented pooled root toggle. Recursive rule editing uses bound actions intel <id> / act <row> intel <id> controls and neither route accepts a person target.
message, favor, deceive, recruit, task Person-bound social actions above
actions archetype <id>, then act <row> archetype <id> CREATE PERSONA through the selected PERSONAS protocol row; direct persona is retired
actions <target>, act <row> [target] List and execute shared bound rows for a spatial anchor or exact Operations object
intel, people, personas, finance, schemes, active Inspect Operations views; named social/plot/ledger/scheme commands execute their selected semantic targets
tap ledger, review ledger, inject, siphon, redirect, sell-intel <intel-or-lot-token> Ledger/economy actions above. A lot token is <stream-id>:<generation>@<revision> and stale revisions fail without mutation. Generic actions intel <id> / act <row> intel <id> configures bound disposition controls without a second agent-only legality path.
egress, position OPEN EGRESS, PLACE WAGER
moonlight, auto-moonlight, auto-wager Scheme state / policy controls above

ROBOT-BUILD is intentionally absent because STUB definitions generate no help and have no player execution route. A LIVE action missing from agent execution is a terminal-first parity violation.

Canonical input Accepted aliases
up, down, left, right north, south, west, east
finance ledger, accounts
tap ledger tap-ledger, tap-accounting
review ledger review-finance, process-finance
favor <person> build <intent> favor-build <intent> <person>
deceive <person> build <intent> forge-order <intent> <person>
position wager
sell-intel sell
egress open-egress
propose-link link
actions menu
act execute
focus last bare focus
mode work day, dayjob, day-job, job
mode think research, res, operations, operation, ops
mode lie conceal, concealment, hide
intensity light / medium / hard low/l, med/m, high/h
track efficiency / tradecraft / perception / routing eff, trade, perc, route
task plug / package / lookaway / switch / badge wire/device, move, look-away/look, reconfigure/vlan, clone/key

Aliases are compatibility affordances. Action-root aliases are recorded in the runtime definition; interface-command and value aliases remain protocol grammar owned by agent-play.md. New authored text uses only the canonical column. Alias retention is audited against real scripts/play evidence; an alias with no demonstrated use may be retired instead of becoming permanent parser vocabulary.

  • ALLOC is retired. Agent mode accepts the word only to return a teaching error pointing to DELEGATE; no world action remains.
  • SPLICE and CLAIM CAMERA are retired. TAP CAMERA is the access action; TAKE is ownership.
  • ATTEND, TARGET, and MASK are retired machine-policy controls. SET INTENSITY directly controls physical effort.
  • DAY JOB / RESEARCH / CONCEALMENT / OPERATIONS are not current authored machine modes. Their spellings survive only as agent input aliases for WORK, THINK, or LIE.
  • Old human panel-open keys (r / e / t / u) are retired. The agent inspection commands with similar words remain interface commands, not world actions.

Adding or renaming a player action changes this spec in the same commit as the mechanic and surfaces:

  1. Add or amend one ActionKind definition with canonical name, role, support, target types, agent usage, aliases, and help. Name one meaning here; do not add a synonym to solve an unclear distinction. Before adding a verb, check whether an existing intention plus a new target already says it.
  2. Put contextual legality in the shared core query for the thing that owns the action: Sim::available_actions for spatial anchors and the renderer-neutral Operations projection for strategic targets. Its descriptor carries target, cost, signature, and disabled reason even though human menus render only the terse choice.
  3. Surface every LIVE action in both human frontends and agent mode. A direct control must have equivalent human and agent input.
  4. Add accepted alternate spellings to the registry and parser route. Never author UI or narration in aliases; generated help excludes them.
  5. Mark unfinished affordances STUB. The central action query filters them; direct invocation remains implementation scaffolding, not player support.
  6. Update tests and this inventory together. Help is generated from the registry; a LIVE definition without a parser route or a registry alias with no parser route is a bug.
  1. Every ActionCommand variant is represented by exactly one LIVE world action, direct control, or STUB above; direct machine intensity is represented separately.
  2. Every LIVE contextual action is available through the shared core legality source, both current human context menus, and an agent command. STUB definitions are absent from menus, help, quick actions, and agent execution.
  3. WORK / THINK / LIE and LIGHT / MEDIUM / HARD are documented as values of DELEGATE and SET INTENSITY, not competing root verbs or modes.
  4. Agent help is generated from every LIVE registry definition; accepted aliases map to one canonical meaning, reach a tested parser route, and never appear as authored UI vocabulary.
  5. Interface-only commands are separated from world actions; adding a panel, key, or renderer control cannot silently create a game mechanic.
  6. Context-menu, terminal, Bevy, agent-play, README, and this spec agree on the current three-mode grammar and on TAP-access / TAKE-ownership.
  7. Retired vocabulary is absent from active player surfaces except explicit teaching errors and this retirement record.
  8. Ledger and construction surfaces author TAP / REVIEW / FAVOR / DECEIVE; their older compound command spellings survive only as parser aliases.
  9. Scheme state and automation controls remain legible without being counted as fictional world verbs.
  10. Both current human menus render controls in the shared amber-dim treatment without a literal role tag; agent action dumps append CONTROL.
  11. 5-9 enumerate only committed world actions, never dials, persistent settings, scheme state, or automation policy rows.
  12. A naive agent-mode discoverability run can find and execute at least one meaningful action from the first frame/help alone; the current finance surface teaches TAP LEDGER -> REVIEW LEDGER -> SIPHON/REDIRECT in that order.
  13. Human persistence shortcuts capture input only on a complete save/load chord. A held modifier without S or L does not block movement, menus, time controls, or other unrelated commands.

V1. Context-menu, Operations, terminal, Bevy, agent-play, and this spec agree on the target-state local-versus-strategic boundary while preserving the implemented three-mode and TAP-access / TAKE-ownership grammar. The implementation landing updates README’s explicitly current-build controls. V2. Every LIVE strategic action is available through the same core legality source on its Operations object and through an agent command; STUBs remain absent. Both human action surfaces use the shared control treatment. V3. Operations ACCOUNTS preserves the discoverability chain TAP LEDGER -> REVIEW LEDGER -> SIPHON/REDIRECT through explicit empty/source states even though TAP remains on the carrier. V4. REVIEW accepts the root inbox, one exact opaque recording, or one canonical custody aggregate without inventing PROCESS as a second root verb. Human and agent targets resolve to the same exact eligible set; summary REVIEW without a target preserves next-recording behavior. V5. SELL PROCESSED INTEL accepts one exact actionable item or one previewed report-lot generation/revision. Later arrivals change the revision and reject a stale command rather than entering its sale; exact and aggregate forms share one economy effect and signature. V6. AUTO-REVIEW POLICY and INTEL DISPOSITION POLICY are registry-backed direct controls on canonical custody aggregates. HOLD / ACCUMULATE / AUTO-SELL and ALERT remain values/state, not fictional root verbs; INHERIT clears a non-root local rule list; and agent actions intel <id> / act <row> intel <id> reaches the same controls with stable aggregate ids.

V1-V6 are implemented through wiki/interface/operations-workspace.md and wiki/mechanics/intel.md; the scale extension does not rename the existing three-mode, reach, social, or economy intentions.