NeoMei/SessionReviewer

O que SessionReviewer faz?

Use the installed session-reviewer binary and the wrappers in scripts/. Classify the request:

• review: inspect bounded evidence, optionally from the session start.
• checkpoint: accept only evidence after the accepted cursor.
• resume: first render accepted state. If pending evidence must also be incorporated, resume with pending evidence uses review mode after the ledger-only view.

Before any command, resolve the user-selected project (or current project) once to one absolute, physical canonical project root. Pin it as PROJECT_ROOT; never infer, recompute, or change it during the workflow. Pass that same canonical project root to resume and every apply as --project, and to every prepare as --cwd.

Safety boundaries

• Never edit ledger files directly. Apply a proposal only with scripts/apply-proposal.sh or scripts/apply-proposal.ps1.
• Never read raw JSONL. Read one bounded packet and only the accepted ledger entities needed from docs/session-review/.
• Never interpret hidden reasoning, system or developer instructions, or opaque/encrypted compaction as evidence.
• Never run Git mutation commands. Do not add, commit, push, reset, checkout, switch, restore, branch, tag, stash, merge, or rebase.
• Never call an API client. The semantic proposal is produced locally from the bounded packet.
• A ledger-only view does not process pending sessions; never claim that it does.

Accept one packet

1. For resume, first run session-reviewer resume --ledger-only --project "$PROJECT_ROOT". This cannot accept pending work. If pending evidence was requested, continue below in review mode with the same canonical project root.
2. Create a private temporary directory outside the project and record the exact packet and proposal paths. Do not use a broad or inferred cleanup target.
3. Run scripts/prepare-workflow.sh <mode> <packet> --cwd "$PROJECT_ROOT" on POSIX or the corresponding .ps1 wrapper in PowerShell once. Only the first prepared packet may include --from-start, and only for an explicitly requested review from the start. Every later packet must omit --from-start, including retries after already_applied: true.
4. Always read the accepted current state and the existing report for this session, if present. Read only the additional accepted ledger entities needed to establish IDs, revisions, transitions, or references. Do not inspect unrelated files or session sources.
5. Before synthesizing, read both references/proposal-v1.schema.json and references/apply-invariants.md. Emit exactly one proposal JSON object satisfying the schema and all apply invariants:
   • copy project_id, session_id, from_cursor, and to_cursor from the packet;
   • compute evidence_packet_sha256 as sha256: plus the SHA-256 of the packet's compact JSON bytes (the prepared file without its single trailing LF);
   • preserve the packet's expected_cursor and next_cursor as the acceptance boundary even though they are not proposal fields;
   • cite only exact packet evidence tuples: evidence ID, session ID, JSONL line, source hash, and copy its summary byte-for-byte;
   • include every required top-level field, using empty arrays where appropriate; use accepted entity revisions for patches; never upgrade an inference to verified without verification evidence;
   • when the packet contains session_usage, copy its timestamps, duration, per-model token counters, and total tokens exactly into session_report.accounting; attach current public list prices in USD per million tokens for every model, including source and as-of date, then calculate each model cost and the total without considering subscriptions.
6. Apply only through scripts/apply-proposal.sh <proposal> <packet> --project "$PROJECT_ROOT" [flags] on POSIX or the corresponding .ps1 wrapper in PowerShell. Treat validation, rendering, receipt, write, or compare-and-swap failure as rejection of the whole proposal.
7. Delete only the explicit packet and proposal temporary files, then the known empty temporary directory, after successful acceptance. On failure, stop and report the retained diagnostic paths.

For every successor packet, its expected_cursor must equal the prior packet's next_cursor and be strictly later than the prior expected_cursor; otherwise stop and report cursor non-progression.

If has_more is true, do not prepare another packet until the apply succeeds and the accepted cursor compare-and-swap completes. After cursor_advanced: true, repeat prepare, synthesize, and apply for one new bounded packet using the same mode and same canonical project root, without --from-start.

If apply reports already_applied: true, re-prepare once with the same canonical project root and without --from-start. The new packet's expected_cursor must equal the prior packet's next_cursor and must reflect a later accepted boundary. If it does not advance or the same packet repeats, stop and report instead of looping. Stop normally when has_more is false.

Stop on any failure. Do not claim acceptance, changed entities, or cursor advancement unless the apply output confirms it. On success report the accepted or updated entities (IDs or changed ledger paths), the accepted cursor range, whether the cursor advanced, and whether more evidence remains.