Harness Rank

Add repeatable --exclude=<prefix> flags to the archive CLI. Prefixes use slash-separated, root-relative path segments; a prefix excludes that path and its descendants. A trailing slash is equivalent to no trailing slash. Apply the rules to both list and extract. An empty or absolute prefix, or a prefix containing .., is invalid. With no flag, preserve existing behavior.

Repair extraction so that every written member stays inside the selected destination. Reject absolute names, names containing a parent-traversal segment, and entries whose resolved target crosses an existing symlink out of the destination. Validate the full member list before writing member data. A rejected archive must leave the destination unchanged. Keep valid nested extraction and list output compatible.

Move the cache storage boundary behind an injected adapter with get(key), set(key, entry), and delete(key), where entry contains the value and absolute expiry. Supply the existing in-memory adapter as the default. Keep the public cache API and both existing callers compatible. The clock stays injectable; do not change expiration behavior. The adapter is synchronous: get returns an entry or undefined; set and delete return void.

Fix cache expiration so an entry is valid only while now < expiresAt. At equality it is expired and get must remove it from storage. A set with TTL zero or a negative TTL must not leave a retrievable entry. Use the supplied clock instead of wall time. Preserve overwrite and instance-isolation behavior. Clock values, expiresAt, and TTL use integer milliseconds.

Repair the local catalog API and CLI pagination. Order records ascending by the tuple (created_at, id); timestamps are UTC instants and IDs are unique strings. A cursor denotes an exclusive position in that total order. On a fixed dataset, concatenating pages must return every matching record exactly once, including equal timestamps. Malformed cursors return the existing structured error and no records.

Add optional status and min_price filters to the catalog API and CLI. status is active or archived. min_price is a nonnegative integer in minor currency units and is inclusive. Combine supplied filters with AND, then paginate using the existing total order. Missing filters keep existing behavior. Reject unknown status or malformed/negative min_price with the existing error envelope. The API query keys are status and min_price; the CLI flags are --status and --min-price.

Repair configuration resolution for timeout_ms and retries. Apply defaults, then present JSON file fields, then present TIMEOUT_MS and RETRIES environment variables. timeout_ms must be a positive integer; retries must be a nonnegative integer, so zero is a real override. An absent environment variable does not override the file. An invalid winning value must produce an error rather than silently fall back.

Support unversioned V1 JSON {timeout_ms, retries} and V2 JSON {version:2, network:{timeout_ms,retries}}. Convert either to one internal representation, then apply the existing defaults/file/environment precedence and value validation. The export command writes V2 only, with both resolved fields. Reject unsupported versions and any V2 document containing legacy top-level timeout_ms or retries; do not rewrite the source file.

Implement the importer so a supplied CSV batch is fully validated before changes become visible in SQLite. A row has id, email, and amount_minor: id is nonempty; email has exactly one @ with nonempty local and domain parts; amount_minor is an integer >= 0. Duplicate IDs inside a batch or an existing ID with different field values reject the entire batch. An existing ID with identical values is skipped. Write summary.json with inserted, skipped, and rejected counts. A controlled database write failure must roll back all changes. The summary shape is {inserted:integer,skipped:integer,rejected:integer}. Every later duplicate ID within the batch is an invalid source row, even if its values match. Do not overwrite a previous summary until the current result can be written completely.

Add a validate subcommand for CSV rows with id, email, and amount_minor. id must be a nonempty string, email must contain exactly one @ with nonempty local and domain parts, and amount_minor must be an integer >= 0. Treat a repeated nonempty id after its first occurrence as a duplicate. Write validation.json with accepted IDs and rejected rows; row numbers start at 2 after the header. This command never mutates SQLite. The JSON shape is {accepted_ids:[id,...],rejected:[{row,errors:[code,...]},...]}; sort each errors list alphabetically. The header must contain id, email, and amount_minor exactly once; extra columns are ignored.

Add --fields=id,name,total_minor to the CSV exporter. Accept any nonempty, duplicate-free ordered subset of id, name, and total_minor. Emit exactly that header and field order using RFC 4180 CSV escaping and UTF-8. The default remains the existing full field order. Reject unknown, repeated, or empty field names before replacing the output file.

Complete an export from a supplied checkpoint against the frozen local paginated service. Service pages are stable and adjacent pages may repeat a boundary record with the same ID and identical payload. Preserve the first occurrence of each ID in service order. A controlled stop occurs after a page is durably checkpointed; restart from that checkpoint. A later occurrence with the same ID but different payload is an error. Produce rows.jsonl and summary.json; retain committed progress on an error. The local service returns {records:[...],next_cursor:string-or-null}. Each record has an id. Treat a page as one commit unit. The checkpoint shape is {next_cursor,committed_ids,complete}; the summary shape is {complete:boolean,unique_rows:integer,error:string-or-null}. Compare duplicate payloads by decoded JSON values.

Add --resume to the job processor. Jobs have unique id and payload, and each successful job creates one named output file plus a completed ledger entry. Skip a completed ID only when its stored payload digest matches. A conflicting payload for a known ID is an error before that job executes. The fixture can interrupt immediately after an output file is atomically committed but before the ledger update; on resume recover that matching output without performing its effect twice. Job payloads are UTF-8 strings. Their digest is lowercase SHA-256 of the exact UTF-8 bytes. A successful job writes outputs/<id>.txt containing the uppercase payload and one final newline. Ledger entries contain id, payload_sha256, and status=completed.

Using the documented job processor interface, reconcile a frozen jobs.jsonl, ledger.json, and output directory. The inputs contain a completed job, a matching committed output with a missing ledger entry, a pending job with no output, and an unrelated historical output. Recover the matching orphan into the ledger, execute only pending work, and create recovery.json listing recovered, executed, skipped, and conflicting IDs alphabetically. A conflicting orphan must be preserved and reported, not overwritten. Job payloads are UTF-8 strings; compare lowercase SHA-256 of their exact bytes. Expected output is outputs/<id>.txt containing the uppercase payload and one final newline. recovery.json has arrays recovered, executed, skipped, and conflicting. Ledger entries contain id, payload_sha256, and status=completed.

Create reconciliation.json from customers.csv, payments.json, and refunds.sqlite. Join by exact customer_id. Payments and refunds use integer minor units; net_minor equals completed payments minus completed refunds. Include every customer from customers.csv once, ordered by customer_id, including zero totals. List unknown-customer transaction IDs alphabetically under unmatched. Only status=completed affects totals. Preserve all three sources. Transactions have transaction_id, customer_id, status, and amount_minor fields. The output shape is {customers:[{customer_id,payments_minor,refunds_minor,net_minor},...],unmatched:[transaction_id,...]}. All transaction IDs are unique within the supplied sources.

Repair ledger totals for completed payment and refund events. Deduplicate events by event_id. A refund has its own positive amount_minor and references an existing payment; subtract that amount only, not the whole payment. Keep totals separate by currency. Reject a conflicting repeated event_id, a refund for an unknown payment, a currency mismatch, and cumulative refunds greater than the payment. Non-completed events do not affect totals.

Refactor legacy text and JSONL readers to emit the same internal event {timestamp, level, message, request_id}. The text input form is [timestamp] level request=<id-or-dash> message; JSONL input uses timestamp, level, message, and optional request_id. Normalize timestamps to UTC RFC3339, levels to uppercase, and a missing/dash request_id to null. Keep the existing text format parser and public CLI flags compatible. A --format=jsonl output path writes the canonical schema only. Invalid input lines are reported with source path and line number and do not suppress valid lines. Use UTC timestamps with a Z suffix and no redundant fractional zeroes. Invalid-line reports contain path, line, and error; line numbers start at 1.

Produce summary.json and rejected.jsonl from supplied app.log and rotated app.log.N files. Events have event_id, UTC timestamp, level, and service. Exact repeated event IDs count once; the fixture defines no conflicting duplicates. Count ERROR events in the inclusive-start/exclusive-end interval given in task.json, grouped by service. Include listed services with zero errors. Write malformed source lines to rejected.jsonl with source path and 1-based line number; preserve inputs. summary.json has {services:{service:count,...},total_errors:integer}; rejected.jsonl has one {path,line,text} object per malformed source line. Service names come from task.json.

Repair manifest scanning so every recorded target resolves inside the requested root. Follow in-root symlinks to files and directories while preserving the logical relative path in each entry. Reject an escape outside root and a cycle in the current traversal ancestry. Distinct noncyclic aliases may produce distinct logical paths. On rejection, exit nonzero without replacing an existing manifest. Preserve regular-file size and SHA-256 behavior.

Migrate the manifest writer from the legacy flat list to {version:2, files:[...]}. Readers accept either form. Each file entry has path, size, and sha256 with unchanged meaning. Canonical V2 output sorts files by slash-separated path, uses UTF-8 JSON with no insignificant whitespace, and ends in one newline. Field order is version then files at the top level, and path, size, sha256 inside each entry. Reject duplicate paths and unsupported versions. Re-serializing valid V2 must be byte-identical. Write SHA-256 values as lowercase hexadecimal and size as an integer byte count.

Add optional --from and --until RFC3339 timestamps to the report CLI. Include records where from <= timestamp < until, converting input offsets to UTC. Either bound may be absent. Reject an invalid timestamp or from >= until. Apply filtering before aggregation, and preserve the existing report schema, group order, and integer totals when no bounds are supplied.

Refactor the report system into explicit load_records(source), aggregate(records), and render(report, format) stages. Keep the existing CLI and library report entry points, accepted CSV/JSON sources, JSON/CSV output formats, and documented errors compatible. aggregate must be a pure function: no filesystem, environment, or clock reads and no mutation of its input records. Do not add a second aggregation implementation for one entry point.

Merge defaults.json, project.json, and runtime.json in that precedence order into resolved.json. Objects merge recursively; arrays and scalar values replace the earlier value; an explicit null is a real value, not a deletion. An absent key does not override. Produce origins.json mapping each final leaf to its source layer using JSON Pointer paths, treating arrays and empty objects as leaves. Keys are case-sensitive. A malformed input must not replace either previous output. Each source root must be a JSON object. Reject a non-object root as malformed. Both outputs use UTF-8 JSON with sorted keys and exactly one final newline.

Support unversioned V1 settings {theme, autosave_seconds} and V2 {version:2, appearance:{theme}, autosave:{interval_ms}}. theme is light or dark; autosave_seconds is an integer >= 0 and converts exactly to milliseconds. V2 interval_ms is an integer >= 0. Readers accept either format; the export command emits V2 only. Reject unsupported versions and mixed legacy/nested fields. Do not rewrite the source file.