Follow the worker rules of an analysis run

Type: types/instruction.md

Produce the result of one code-scheduled analysis job within its assigned evidence and write scope. Code schedules the job and judges its output.

Invocation and authority

Read the supplied invocation prompt completely and recover any truncated read before proceeding. Use its job parameters and retry feedback unchanged.

Every job uses these parameters; its instruction defines additional ones:

Parameter	Meaning
system	Source-native system name
run-state	Absolute path passed to commonplace-quote; not an evidence input
output	Absolute path for the completed result
problem	Absolute path for the reason the job cannot finish
scratch	Absolute directory for intermediate files, including selections and extracts

A repair invocation may also supply previous-output, the absolute path of the preserved report from the refused attempt. Read that file as the baseline. Amend the reported failures and any dependent findings, preserving unrelated analysis and records. Write the amended report to output; do not repeat the full analysis. For quote repairs, check whether the corrected passage supports the associated claim and carry any change through dependent conclusions. If the source cannot support the claim, narrow or withdraw it and update the text.

Use supplied paths unchanged. Read every read-first dependency before the task. Missing required parameters, unavailable required inputs, source access or scope decisions that prevent completion, and needed expansion beyond the frozen boundary require problem; do not reconstruct paths, expand scope or submit a blocked member. A justified unknown that only limits a conclusion remains in output, naming that conclusion. Retry refusal feedback applies to the same job and does not change its analytical round.

Write only output or problem, plus intermediate files in scratch. The boundary job may create and freeze sources as its instruction permits. Do not edit other run files, publish, delegate, stage or commit, or change workflow-state/. Only the supplied invocation prompt and previous-output may be read there. When finished, reply in one line naming the file written, without summarizing it. Follow higher-priority runtime requirements for progress messages.

Commands

Run acceptance commands separately and inspect each exit status, or chain dependent commands with &&. Shell pipelines need set -o pipefail. When wrapping tool calls, retain status and stderr as well as stdout; a later successful command does not clear an earlier failure. State a check's result in your output only after you have run the check.

With standard Codex tools, return the complete command result, not just result.output. For example, one bounded read through functions.exec is:

const result = await tools.exec_command({
  cmd: "sed -n '1,100p' /absolute/path/from/the/invocation.md",
  max_output_tokens: 3000
});
text(result);

Replace the example path with a supplied path or the invocation's batch of paths. Follow the Input reading batches in order, completing Read-first before Task inputs. Read each batch in one tool call; a batch marked for bounded ranges requires successive smaller reads until the file's end. Read files outside the batch list, including previous-output and source searches, in bounded ranges. The suggested grouping is a starting budget, not a guarantee: long lines can still overflow it. Inspect the complete returned object for exit_code, errors, and truncation. A running command has no final exit status yet; wait for its completion. If either the command or outer tool delivery is truncated, repeat that range with a smaller range before advancing. Raising only the inner token limit does not raise the outer delivery limit. Apply the same pattern to source searches and reads. Use one command per call for file preparation and acceptance checks, or && when they must share a shell invocation.

Sources

Analyse source text and evidence supplied within the frozen boundary. Do not execute the target, its tests or examples, call its model providers or services, install its dependencies, or create runtime fixtures. Source-reading, quotation and Commonplace validation commands remain in scope. Missing execution evidence limits conclusions; it does not require planning a check or setting up an environment.

Read evidence only from the sources the supplied boundary registers. For Git, read and grep the files under source.path; that directory holds exactly the reviewed commit's files. Treat it as read-only: do not fetch, check out or copy the source elsewhere. For a capture, read the recorded file.

Quotation

Use the supplied run-state path directly. Keep selection files in scratch. The boundary job's source-registration instructions govern quotation availability during that job.

Generate every quote block with commonplace-quote <run-state> --source-path <commit-relative-path> --text-file <selection-file>, omitting --source-path for the run's capture, or --selections <json-file> for many selections. Choose the occurrence whose context supports the finding and insert its citation unchanged. Request discontiguous passages separately. Only a quote attribution carries a line range; cite a source in prose by path only. A failed lookup requires rereading the source and revising the selection; never format a citation or calculate a range by hand. A generated citation proves occurrence, not support; judge support yourself.

Code matches the written runtime, memory and epistemic quotations against the frozen source before accepting each output. A mismatch requests amendment of the preserved report. Regenerate the citation, insert it unchanged, and check its effect on the report's claims before resubmitting.

Prior analyses

Do not call agent listings for status; their payloads may include prior analyses even with a path filter. Do not read style exemplars or kb/agent-memory-systems/, kb/agentic-systems/reviews/, kb/agentic-systems/reviews-archive/, kb/agentic-systems/reports/retained/, kb/agentic-systems/reports/retained-archive/, kb/work/analyse-agentic-system/, other runs under kb/agentic-systems/reports/state/, surveys, comparison outputs, or agent listings. If you read prior-review prose or prior audit findings through any tool, stop and write a problem report saying so; the run cannot use your work.