Sigil: Workbench Poll

• the operator says poll after making a workbench request available,
• a blind test needs to prove the session does not know task or skill until claim,
• a local session needs to claim a ready request and run the requested skill in the same turn,
• a claimed request needs a terminal result or explicit interruption.

Do not use this sigil to start the workbench, prepare operator approvals, wake an inactive agent session, loop forever, or inspect durable state directly.

• repository root, resolved from the current working directory,
• ide-extension bridge CLI at projects/ide-extension/scripts/session-bridge.mjs,
• base URL, defaulting to http://127.0.0.1:8765,
• request handle for claim mode,
• claim ID and JSON input for result or interrupt mode,
• optional agent_ref and session_ref.

1. Resolve the repository root from the current working directory.
2. Confirm projects/ide-extension/package.json exists.
3. Confirm the bridge CLI is available through npm run bridge:session -- <command>.
4. Use only a loopback HTTP base URL. Reject HTTPS, paths, remote hosts, and non-loopback origins.
5. Do not read or mutate .ide-extension/state directly.

Step 2 - Ready Poll

Run this step for --ready.

1. Execute:

   npm run bridge:session -- ready --base-url <base-url>
   

2. If exit code is 0, report the returned opaque request_handle.

3. Do not infer, reveal, or guess the task, requested skill, grants, or context from a ready response.

4. If exit code is 3 with no_ready_request, report Availability: none without treating it as a failure.

5. If exit code is 4, report that the workbench bridge is unreachable and recommend workbench-up --start.

Step 3 - Claim And Continue

Run the claim portion of this step for --claim and --claim-only.

1. Require --request-handle.

2. Generate session_ref if absent, using a stable local shape such as session-<timestamp>-<short-random>.

3. Execute:

   npm run bridge:session -- claim --request-handle <id> --session-ref <session-ref> --agent-ref <agent-ref>
   

4. If the claim succeeds, report the execution packet:

   • task,
   • requested skill,
   • grants,
   • context summary,
   • claim ID,
   • claim attempt ID,
   • envelope ID,
   • run ID,
   • approval ID,
   • session ref.

5. For --claim-only, stop after reporting the execution packet. Do not invoke the requested skill or submit a result.

6. For --claim, treat successful claim as authorization to continue the already-approved request in the current session:

   • resolve the revealed requested skill from the installed runtime skills,
   • read its complete SKILL.md before task actions,
   • execute the revealed task under that skill's process and boundaries,
   • preserve any human gate or blocker required by the requested skill,
   • assemble a result only from actual execution evidence.

7. If the requested skill cannot be resolved, execution cannot safely continue, or the requested skill reaches a real blocker, report the blocker and submit a bound interruption when the claim cannot remain usefully active.

8. Do not submit a passing result until the requested work has actually completed.

Step 4 - Result

Run this step automatically after successful --claim execution, or explicitly for --result when resuming an already completed claim.

1. Require --claim-id and --input.

2. Confirm the input JSON contains the claim binding fields returned by claim:

   • claim_id,
   • claim_attempt_id,
   • envelope_id,
   • run_id,
   • approval_id,
   • session_ref,
   • requested_skill.

3. Execute:

   npm run bridge:session -- result --claim-id <claim-id> --input <json-file|->
   

4. Report the admitted result and receipt only if the command exits 0.

5. Never tell the user the task passed unless the local kernel admits the result.

6. Include the requested skill's real validation and evidence references; do not use placeholder hashes or synthetic evidence for a real run.

Step 5 - Interrupt

Run this step for --interrupt.

1. Require --claim-id and --input.

2. Confirm the input JSON carries the claim binding fields and a reason.

3. Execute:

   npm run bridge:session -- interrupt --claim-id <claim-id> --input <json-file|->
   

4. Report the interrupted claim state and preserve the reason.

5. State that only the operator may release and retarget an interrupted claim.

• mode,
• base URL,
• ready handle presence without task disclosure,
• claim ID and claimant refs when claimed,
• requested skill resolution and execution continuation status,
• result or interruption receipt when submitted,
• proof boundary,
• overclaim or blind-boundary violations.

## Workbench Poll Result
- Mode:
- Base URL:
- Availability:
- Request handle:
- Claim:
- Requested skill:
- Task:
- Execution:
- Receipt:
- Validation:
- Proof boundary:
- Follow-up: