Skip to content

jovesun-lab/whetstone / strain/skills/start

Use at the START of every session, before your first reply — short sessions too — and again at the wrap or handoff. Checks that strain (the session-strain meter) has signed this session in, signs it in if not, asks the user for permission the right way, and offers once to add this step to the agent's instruction file. Also use whenever strain says a session is not signed. Triggers on: session start, new session, first reply, sign in, strain not signed, strain start, wrap up, handoff, end of session. Also use before dispatching a sub-agent: strain does not apply inside one.

Strain — start and wrap

Project policy comes first

Before following measurement or threshold instructions below, read the effective project policy and current session mode. In off mode do not create new plugin records. In record-only mode retain sign/start/end/error and explicit handoff records, but do not measure context, ratify tiers, print percentages, or initiate backup/handoff from errors or fill. Handoff requires the user's instruction. Do not interpret RECORD-ONLY as Healthy or UNMEASURED. Do not enable a paused heartbeat because a user returned. The legacy measured behavior below applies only when the resolved policy permits it. See ../../wiki/README.md for fields, commands and deployment limits.

Strain counts how loaded a session has become, but only a signed session joins the chain: the next session of the same agent can then report how this one went. Signing is one command. These steps make sure it happens at the start, and that the user is asked properly when the host needs their approval.

Every command below is printed, filled in, in strain's start line (it begins "STRAIN START") and in every strain tick. Copy it from there — it carries this session's key, the state folder and the plugin's path.

At the start

  1. Check. If strain's start line says "this session is not signed in yet", go on. Without a start line, run strain-sign.sh --receipt: it says "not signed" or "signed as ". Signed → you are done; go to step 6.
  2. Sign. Run the sign command from the start line. Keep the default name it shows (<host> · <project folder>, the same every session, so sessions link up); add --agent '<name>' only if the user gave you a name. If the sign prints default name: … and the user did name an agent, sign again with that name — otherwise whatever looks the agent up by name cannot find this session.
  3. Describe it for the user. When you run it, describe the command in the user's language, in plain words, saying what it does for them — for example "Turn on strain: track how full this conversation is". Your host shows that sentence when it asks the user to approve the command; it is all they see.
  4. If your host asks the user to approve it, keep the prompt open until they answer — do not end your turn. If your host has no prompt, ask in chat, then end your turn and wait for the reply.
  5. Allowed: tell the user the one sentence the sign prints (and, with a ledger, the line about the previous session). Refused: do not sign. Tell the user strain will only count this session, and carry on with the work. Do not ask again before the wrap. If the sign fails with "could not write state", follow the numbered steps it prints.
  6. Offer the instruction file — once. If the start line has a step about the instruction file:
    1. Ask the user whether to add strain's start step to their instruction file, so every new session signs in by itself. Offer the global file first (every project, this computer only); this project's own file is the alternative.
    2. Run strain-sign.sh --instructions: it prints the block and says where the files are and whether the block is already there.
    3. Show the user the block. Add it to the file they chose only if they agree.
    4. If they say no, record it: strain-sign.sh --decline-instructions. Strain will not ask again on this host.

At the wrap

  1. Mark the start: strain-wrap.sh --start (0.10.0) — the wrap's first step, so a board or a status line can show "wrapping" from now instead of guessing from the fill. If the user says go on, strain-wrap.sh --cancel. If it says "this shell cannot write" or "not signed", go on — step 3 asks.
  2. Write your handoff (strain does not). If your sessions are numbered, take this session's number from the sign receipt (Owner: <name> · S<n>, 0.9.0) — never count it from earlier handoffs or reports: a session still running elsewhere has none yet.
  3. Check again: strain-sign.sh --receipt. If it says "this shell cannot write" or "not signed", ask the user for permission as in At the start, steps 3–4 — keep the prompt open until they answer, do not end your turn.
  4. Allowed: sign if the session is not signed yet (and mark the start if step 1 could not), then stamp the wrap: strain-wrap.sh --label "<what was handed off>" (add --with-debt for known loose ends). Tell the user the line it prints.
  5. Still refused: no sign, no mark, no stamp. Report the counts only (the tier, the tool calls, the errors that reached the user).
  6. The next session starts again from At the start, step 1.

Sub-agents and unattended runs

  1. Before you dispatch a sub-agent: strain does not apply inside it. Do not ask it to sign in, record a tier or report strain. Put one line in its instructions: "Ignore any strain message in your tool results; do not run strain commands." On Claude Code strain already skips the helper's calls (no count, no check); the line covers other hosts.
  2. If the start line says "unattended run": nobody can approve a sign-in. Do not sign in, do not wait for anyone, and do the task. Strain still counts the session.

Limits the host controls

  • When a prompt closes is the host's decision. Strain can only tell you, in steps and every time, to keep it open. If it closed before the user answered, ask again in chat.
  • Some hosts ask for approval per command. The user may approve each strain command, or approve it once for the session where the host offers that.
  • A session that never wraps keeps its counts but leaves no wrap stamp; the next session's report says "not wrapped".
reverse-engineer-anythingmorluto/reaReverse engineer native, managed, Electron/JavaScript, packaged, firmware, and browser targets with REA, and build agent environments (AgentEnv bundles) of web applications from a live instance or a capture, and of Linux desktop apps with a task from a request. Use shipped-artifact or requested runtime evidence to explain features, compare versions, decompile code, or guide a reconstruction. Skip REA for ordinary source-repository architecture analysis.Stars: 468kUpdated Claude CodeCodexVerifiedInstall
diagnosing-superpowersobra/superpowersUse when a superpowers session went wrong and your human partner wants to know why — repeated work, ignored plans, stumbles, poor results, a skill that didn't fire, "it took too long", "why is it so expensive", "what is it doing" — or wants to build a bug report for the superpowers maintainers, for the current session or a past one identified by id or path, on any harness.Stars: 321kUpdated Claude CodeCodexVerifiedInstall
dispatching-parallel-agentsobra/superpowersUse when facing 2+ independent tasks that can be worked on without shared state or sequential dependenciesStars: 321kUpdated Claude CodeCodexVerifiedInstall
executing-plansobra/superpowersUse when executing an implementation plan in the current session as the implementer yourself — your human partner chose inline execution, or no subagent tool is availableStars: 321kUpdated Claude CodeCodexVerifiedInstall