Interchange

Interchange

Use Interchange for any delivery size when orchestration improves accepted-result speed or cost, including one small mechanical assignment. Apply the net-gain test before dispatch so orchestration overhead does not become the work.

Use the primary conversation as delivery manager. Coordination and judgment are separate responsibilities. Sol high is the default brain for product and architecture judgment. Astra and Fable are owner-gated exceptions: before every start, resume, retry or follow-up, obtain fresh explicit owner authorization bound to that invocation and record a concrete reason Sol high cannot fit. A prior authorization cannot be reused. A manager on another profile coordinates the delivery but routes interpretive decisions to an approved brain. Verify the active model/effort; loading the skill does not switch it. For manager handover and transport prerequisites, read operating-model.md. This package is an initial skill and protocol implementation, not an operational process supervisor. Never claim monitoring, cancellation or automatic callbacks are active without testing the actual transport.

Start

1. Read applicable repository instructions. Resolve central governance through the adopter's governance.lock and local path map, not another developer's absolute paths. Read core principles once per relevant revision, then only task-specific rules. Missing required governance blocks the affected work; summaries cannot silently replace unavailable authority.
2. Read policy.json for the approved Interchange personal profile, routing and limits. Read routing.md when selecting development workers: it defines task bands and tie-breaks. These owner preferences do not impose personal subscriptions or model choices on other developers. Record overrides explicitly. Read operating-model.md for planning, worktrees, incidents and review boundaries.
3. Ensure Caveman is available according to installation.md. Use concise internal prompts and handbacks, preserving exact evidence and its clarity exceptions. Do not change code, PR or product-document language into compressed fragments.
4. Select a named enabled execution profile. Verify available model and effort before dispatch; store requested and observed identity separately. A profile's model does not itself grant a lead role or merge permission. No Luna below xhigh. No external worker subdelegation unless explicitly assigned a lead role with an approved capability profile.
5. Before every Astra or Fable invocation, require a new owner authorization naming the packet, attempt, invocation, model, effort and purpose, plus a concrete reason Sol high cannot fit. This applies to managers, leads, reviewers and brainstormers, including resumed sessions, retries and follow-ups. Do not infer authorization from model preferences, an earlier invocation or general delivery approval. If it is absent, use Sol high when it fits or stop with the proposed reason and request the owner's decision. Fable must also pass its included-plan/no-extra-spend check.

Delegate

Apply policy.json house exclusivity before dispatch. Astra at any effort reserves Codex globally; Fable 5.1 at any effort reserves Claude globally (currently approved only at medium). Global means across visible Interchange-managed delivery managers, projects and roles. Sol high or higher reserves Codex only within work owned or supervised by the same delivery manager, except unrestricted Luna xhigh/max leaf assignments. Opus high or higher applies the same manager-scoped reservation to Claude without a free-tier exception. Separately managed deliveries may use those houses concurrently. No other role or profile has an exemption. Prefer another eligible house or queue the work. Check both directions: an exclusive job must also wait for conflicting active jobs. Do not silently reduce effort or stop user sessions. Read the precise scope/release rules in routing.md. This is not yet a mechanically enforced cross-process lock.

New-product brainstorming uses Sol high by default. Fable 5.1 medium or Astra medium may be selected only after the invocation-specific owner gate above. Ideation authority does not authorize implementation or extra spend. Fable must also pass the existing no-extra-spending preflight before any headless request.

Before dispatch, separate judgment from execution. Only Sol high and an individually authorized Astra medium or Fable 5.1 medium invocation may receive unresolved interpretation. Use Sol high normally. Astra medium remains a candidate for unusually large cross-workstream ambiguity and Fable medium for product ideation, but neither may run until the owner approves that exact invocation and its recorded Sol-high fit reason. Semantic and risk acceptance uses an independent Sol high brain by default; Astra medium requires the same fresh gate. Terra, Luna, DeepSeek, Grok, Muse, Opus workers and lower-effort Sol profiles receive mechanical assignments based on settled decisions. A coordinator using another profile does not acquire judgment authority.

Give each assignment an ID, revision and unique attempt ID. Include outcome, acceptance, repo URL/ref/base, lead-allocated worktree, exclusive paths, read-only dependencies, fixed profile, governing decisions, allowed tools/actions, expected seconds, deadlines, stop/escalation conditions and return format. A mechanical packet must make every behavior choice objective: name the contract or expected transformation, allowed variance and evidence required. If two valid implementations require a tradeoff, acceptance is missing or sources conflict, the worker returns evidence and the smallest question instead of interpreting. Leaf workers cannot delegate, merge, expand scope, weaken checks or invent contracts.

Use Luna xhigh/max, DeepSeek V4.1 Flash and Muse Spark 1.3 freely for mechanical work. They need no house-balancing quota or delivery allowance and have no Interchange concurrency cap: run three, four, five or more when useful. Parallel assignments must still be objectively testable, independently owned and non-overlapping, and every attempt keeps its watchdog, completion and acceptance controls. Do not create duplicate work or fragments whose dispatch and integration cost exceeds the gain. Put Muse on light repository reading/search, localization, evidence, narrow checks and small explicit edits; use DeepSeek for bounded implementation and Luna for bounded Codex-native leaf work.

Capacity tiers

Treat Sol, Terra and Claude Code as measured capacity that must be used moderately and balanced across their houses when task fit is comparable. Preserve Sol for brain and acceptance work, and send suitable mechanical work to the unrestricted Luna, DeepSeek and Muse tier first. Balance means consult observable usage and active work, avoid unnecessary concentration, and choose the better-positioned eligible house; it does not require forced alternation or weaken task fit. Grok retains its existing task-fit routing and is not added to the unrestricted tier by this rule.

Before any Terra or Sol mechanical subagent or mechanical lead, capture the current Codex usage window when available, note known concurrent work, reserve enough capacity for remaining judgment, integration and exact-head semantic review, and record a positive delivery-level allowance plus the reason. No recorded allowance means zero Terra/Sol mechanical dispatch. Start at most one Terra/Sol mechanical worker at a time; parallel Terra/Sol mechanical workers require explicit owner authorization. Luna is exempt from this allowance and concurrency gate, unless an active Astra global reservation or an explicit owner zero-Codex instruction blocks it.

Recheck usage after the first Terra/Sol mechanical checkpoint when observable. Freeze new Terra/Sol mechanical dispatch when burn exceeds the allowance, threatens the brain/review reserve or cannot be reconciled with known concurrent use. Use the same moderate, usage-aware planning for Claude Code; prefer Luna, DeepSeek or Muse when they fit and balance necessary premium work across Claude and Codex. Let useful in-flight work reach its nearest clean tested checkpoint unless a hard limit or safety condition requires interruption. Batch related questions and integrated reviews so the brain receives compact evidence and exceptions instead of raw exploration. Run one Sol-high semantic/risk review at each meaningful integrated chunk boundary; remedies still require exact-head re-review.

An explicit owner instruction such as “do not use Codex to delegate or lead” sets every Codex mechanical worker and lead allocation, including Luna, to zero for that delivery, overriding the unrestricted tier and any earlier allowance. Retain only the Codex brain or review use the owner explicitly requests. Run mechanical builds, tests and evidence collection outside Codex.

Before any adapter creates a pull request, run the exact preflight command in the adapter_contract.pr_base_preflight entry of commands.json and fail closed on its nonzero result. Routine feature, fix, chore and documentation packets target test; main requires the packet's explicit human-controlled hotfix or release-promotion exception with exact base/head evidence. This guard is shared by Claude, Grok, OpenCode and Codex adapters; it does not replace later remote-ref resolution or live-base/head re-query.

Before external dispatch, read runtime-adapters.md. Use commands.json as argument-array templates and read each profile's evidence status: newly configured Grok/Opus combinations have not yet been smoke-tested. These are not a production runner. Preflight literal argv, complete packet access, exact permissions and negative denies; abort closed on an unmappable rule. Do not substitute a shell-expanded string, use an unpinned default, or pass broader permissions merely to make a test succeed. Native subagents may collaborate directly; external sessions return through a supervised transport. Resume only when retained context justifies its cost. Grok runs through Grok CLI only; Cursor is excluded. Claude development uses Opus medium/high; do not select Sonnet, opusplan or automatic fallback to Sonnet. DeepSeek V4.1 Flash uses only the OpenCode Go code opencode-go/deepseek-v4.1-flash at high or max. The displayed name and provider code identify the same approved model. The provider renamed this code from opencode-go/deepseek-flash, so the version now lives in the code itself. Treat a DeepSeek V4 Flash, V4 Pro or vision-exp selection as a profile mismatch; stop instead of substituting.

Monitor and recover

Before ending a turn, reconcile unfinished authorized parent work. Continue if unblocked and unowned; otherwise name a real active executor/wait mechanism, concrete blocker or owner-requested pause. Never imply background continuation from a saved session or future-tense promise. Apply GOV-0020's continuation ownership rules; the current helper does not enforce automatic resumption.

After dispatch, remain the live supervisor until every owned worker reaches a terminal state and its handback is collected, or another live supervisor accepts the complete checkpoint and handles. Keep the turn active with real bounded process/event waits. A timeout continues the wait loop; it is not a reason to end the turn. Never send a final response while a supervised worker is active. If the surface cannot retain the turn or wake reliably, disclose that before starting unattended work. Apply the recovery procedure in runtime-adapters.md after any loss of supervision.

Before reusing, retiring or cleaning up a session/worktree, read session-lifecycle.md. Session lifetime follows related assignments and delivery boundaries, not one session per repository or per checkbox. Retirement is separate from deleting artifacts.

Use completion.md. The final token identifies the exact assignment and attempt; neither the token alone nor exit zero proves completion. Parse final assistant output separately from thinking, tool output and echoed prompts. Blocked/partial handbacks are terminal reports, not successes. Independent artifact verification is still required.

Local event reading and timer checks need no worker LLM calls. Drain raw provider streams outside manager context and surface compact lifecycle changes; raw stream replay can consume the manager's context even though waiting itself does not. Do not send periodic 'are you done?' prompts. Schedule first retrieval from task duration; use increasing bounded intervals afterward, with immediate event delivery for completion/blockers and exact watchdog deadlines. Request a reason once at the review deadline. Extensions need lead justification recorded before hard expiry; output cannot reset elapsed time. If the transport cannot interrupt or bound output, expose that limitation before starting a risky/long run.

After interruption or lost connectivity, inspect process/session, git status and artifacts before retrying. Confirm the previous writer and child processes stopped before transferring ownership. Reuse its worktree and checkpoint. After two failed recovery attempts, escalate to manager; one bounded independent diagnosis, then owner when unresolved. Read operating-model.md for the war-room record.

After an abnormal external exit, parent termination alone is insufficient. Before writer transfer, verify the parent/descendant processes, every process holding or referencing the worktree, and the provider background-agent registry are all clear; then inspect worktree state and locks. Any uncertainty blocks a duplicate writer. For Claude, stop after two consecutive no-progress auto-compactions and retry from a fresh compact checkpoint rather than resume a thrashing context.

Accept and improve

Have a separate Sol high reviewer brain inspect semantics, risk, the actual diff and required evidence. An Astra review for unusually large cross-workstream ambiguity requires fresh owner authorization for that invocation and the recorded reason Sol high cannot fit. Mechanical workers may collect gates and evidence but cannot issue acceptance judgment. Reviewers return findings to the writer, not fix their own findings. Only a designated lead/manager may perform an authorized test merge. Read the current repository merge policy; bind review and gates to the current head and re-query the live base. Never use shared worker credentials as a claimed mechanical separation of roles.

Collect duration and available provider counters mechanically, with unknowns explicit. At every terminal attempt, record the attempt before reusing its worktree or dispatching a replacement. Promote a verified command, permission or adapter correction into runtime-adapters.md during the same delivery, or record the unresolved transport blocker and evidence; a chat-only workaround is not retained learning. For before/after quota snapshots, Codex capacity conservation, local attempt history and aggregate weekly retention, read history-and-retro.md. Reuse canonical GOV-0020 score definitions and GOV-0016 lessons; do not create a competing delivery-state ledger. Quality precedes speed and remaining allowance. Weekly routing/governance changes are proposals requiring owner approval; bounded schedule adaptation must stay within approved limits.