v0.103.0

Release v0.103.0

Added

• help://quick derived from the adapter registry instead of hand-maintained (BACK-391 M4) — the top command block used to be a static list skewed toward infra adapters (6 of 10 entries were ssl/domain/nginx/cpanel/check) and could never include project-local plugins. New _get_quick_commands() (adapters/help.py) iterates the live adapter registry, pulling each adapter's own get_help() description + first example URI, so the block can't drift from reality and automatically surfaces plugins. A _QUICK_RANK dict controls priority only (not content) to keep code/session/quality adapters ahead of the infra niche. This was the strategy doc's last open item — the M1–M4 onboarding sprint is complete. +4 tests.
• Breadcrumbs teach the outline≠content mental model, not just mechanics (BACK-389 M1) — reveal doc.md now frames its "Next:" hint honestly (Outline only — headings show where, not what.) and suggests --section '<first heading>' with the real heading name, the doc equivalent of the Class.method hint code files already get (_suggest_doc_section_extraction, reveal/utils/breadcrumbs.py). The once-per-install stderr nudge (_show_breadcrumb_hint_once) now leads with New here? reveal --agent-help instead of only advertising how to disable breadcrumbs — this is the one channel with ~100% agent exposure, so it's the best place to land the orientation lesson at first use. +7 tests.

Fixed

• --agent-help / help://agent now obey progressive disclosure (BACK-389) — _FULL_ONLY_TOPICS no longer exempts agent, so help://agent truncates to a ~500-token preview + help://agent/full footer like every other guide. The CLI flag --agent-help was a second, separate implementation that bypassed HelpAdapter entirely (direct file read); rewired to go through the same code path so both doors behave identically. --help epilog now points to help://quick/--agent-help (previously zero pointer). Stale ~40K tokens descriptions corrected across AGENT_HELP.md, README.md, QUICK_START.md, HELP_SYSTEM_GUIDE.md, DUPLICATE_DETECTION_GUIDE.md.

---

Install/Upgrade:

pip install --upgrade reveal-cli

Documentation: https://github.com/Semantic-Infrastructure-Lab/reveal
Full Changelog: https://github.com/Semantic-Infrastructure-Lab/reveal/blob/master/CHANGELOG.md

v0.102.0

Release v0.102.0

Added

• ElementRenderMode registry replaces render-mode if/elif chain (BACK-360) — _render_element in reveal/cli/routing/uri.py dispatched --outline/--links/--frontmatter through an 87-line if/elif chain; replaced with an _ELEMENT_RENDER_MODES registry of _handle_*_mode(result, args, text_field, label) -> bool handlers and a 5-line dispatch loop. No behavior change.
• Heading-aware markdown outline collapse in default text mode (BACK-387) — markdown heading outlines now indent by level (H2/H3 visually distinct, was flush-left) and auto-collapse past 25 headings to the shallowest discriminating level with (+N subheadings) counts, deterministic via _discriminating_level(). --depth N doubles as an explicit collapse override. New _format_markdown_headings()/_discriminating_level() in display/formatting.py. Verified against 9 real docs (e.g. AGENT_HELP.md 107→24 lines); JSON and --outline modes unaffected. +9 tests.

Fixed

• E501 false-flags data/doc files (BACK-386) — long-line rule now skips data/doc file categories via category-aware BaseRule.skip_categories.
• Doc/help counts corrected across 13 docs + 3 code files — README and friends claimed "305+ languages"; a passing integrity test enforced it, but reveal only routes/analyzes 84 languages by extension (305 is the underlying tree-sitter-language-pack grammar count, mostly unmapped). Corrected adapter count (22/23/24 → 25), language count (35/37/50/80/185/305 → 84), and AGENT_HELP token-cost estimate (~12K → ~40K) everywhere they were hardcoded, including CLI --help and help:// output. Rewrote test_registry_integrity.py::test_language_count_in_readme_is_accurate to assert against reveal --languages (the real gate) instead of the raw tree-sitter pack, with floor semantics (actual >= claimed) to tolerate benign registry-pollution leaks from test ordering. Filed BACK-388 to generate/lint these counts from the live registry instead of hand-typing them.

---

Install/Upgrade:

pip install --upgrade reveal-cli

Documentation: https://github.com/Semantic-Infrastructure-Lab/reveal
Full Changelog: https://github.com/Semantic-Infrastructure-Lab/reveal/blob/master/CHANGELOG.md

v0.101.0

Release v0.101.0

Added

• surface --source-only flag (BACK-380) — excludes test files and directories from the surface scan so DD / security reads see only production surface. Prunes test dirs (names starting with test/spec, plus __tests__) and test files (test_*.py, *_test.py, conftest.py, *.test.ts, *.spec.ts, *.test.tsx, *.spec.tsx) during file collection. Single-file invocations are unaffected; _meta.known_limits records the exclusion. Dogfood: 2242 → 286 surface entries on reveal's own tree. +11 tests.
• git://<path>?type=ownership — commit-share ownership query (BACK-383) — aggregates git-log authorship over a file, directory, or whole repo, returning the primary author, per-author commit-share %, contributor count, and last-touch date. Works uniformly for files and directories (a directory resolves to a subtree oid, so one parent-vs-commit comparison detects any change beneath it). Merge commits are excluded by default (?merges=1 to include); ?limit=N caps the history walk. This is straight commit-log aggregation (commit-share, not surviving-line ownership — use ?type=blame for that); the consumer applies the bus-factor / key-person judgment. Carries the BACK-379 shallow_clone warning. New get_ownership() + commit_touches_path() in adapters/git/files.py; git_ownership renderer. +10 tests.

Fixed

• surface Python: mock.patch / mocker.patch / unittest.mock.patch decorators misclassified as HTTP PATCH routes (BACK-377) — Python sibling of BACK-375 (TS supertest). _is_http_route() in nav_surface.py now short-circuits via new _is_mock_patch_decorator() helper when the decorator is any known mock-patch form (mock.patch(, mocker.patch(, patch(, unittest.mock.patch(). Real app.patch('/endpoint') routes are unaffected. Repro: reveal surface saleor/payment was reporting 22 PATCH routes all pointing at test files; now 0 false positives. +10 tests.
• architecture / overview / imports:// flood on repos with committed venv/ or site-packages/ (BACK-378) — ImportsAdapter._build_graph() used bare rglob('*') with no directory filtering. On a repo with a committed venv/, this pulled in all of scipy/joblib/sklearn — producing 239 false entry points and 49 false circular groups, and timing out overview at 90s. Replaced rglob with os.walk + SKIP_DIRECTORIES filter (the same canonical set from reveal.defaults already used by surface, analysis.py, etc.). No behavior change on normal repos; vendor-heavy repos now return clean structural analysis instantly.
• git://?type=blame silently degraded on shallow clones (BACK-379) — Blame on a --depth 1 clone produced empty or limited attribution with no indication of why. get_file_blame() now checks repo.is_shallow (pygit2 native attr) and propagates shallow_clone: true to the result; the renderer prints a one-line warning before the blame output: ⚠ Shallow clone detected — blame attribution is limited to the fetched history. Run git fetch --unshallow for complete bus-factor / key-person data.

---

Install/Upgrade:

pip install --upgrade reveal-cli

Documentation: https://github.com/Semantic-Infrastructure-Lab/reveal
Full Changelog: https://github.com/Semantic-Infrastructure-Lab/reveal/blob/master/CHANGELOG.md

v0.100.2

Release v0.100.2

Added

• Confidence/limitation metadata on inference-heavy outputs (BACK-330) — calls://?uncalled, depends://, and surface now carry explicit confidence caveats. ?uncalled text output gains a "⚠ Static only — runtime dispatch may create hidden callers" line; depends:// adds a _meta dict (analysis_kind, confidence, known_limits) and bumps contract_version to 1.1; surface adds _meta to JSON + a text caveat. AGENT_HELP "Interpreting absence of evidence" notes added.
• contracts for TypeScript (BACK-372) — reveal contracts now classifies TypeScript files: interface → contract (Protocol/ABC analog), abstract class → abstract class analog, type alias → TypedDict analog, class … implements I → implementation edge. Upstream analyzer fixes reusable by all consumers: _extract_class_bases reads class_heritage → extends_clause + implements_clause; abstract_class_declaration added to CLASS_NODE_TYPES; interface extends captured via extends_type_clause. TS-flavored group labels in renderer. Dogfood: 94 contracts found in tia-chess. +11 tests.
• surface for TypeScript (BACK-373) — reveal surface now scans .ts/.tsx files for hidden dependencies. New nav_surface_ts.py tree-sitter walker detects 8 categories: network (axios/fetch/got/…), db (pg/prisma/knex/…), sdk (@anthropic-ai/sdk/openai/stripe/…), env (process.env.), fs (fs.writeFile/Bun.write/Deno.writeTextFile), http (express/fastify/koa route registrations), subprocess (child_process.*/execa — new category), cli (yargs/commander .command()/.option()). Python and TS results merged into one report. Dogfood: 116 surface entries on tia-chess. +23 tests.
• patches:// for Jest/Vitest (BACK-374) — patches:// and the testability patch-pressure signal now work on TypeScript. New scan_patches_ts() detects jest.mock/vi.mock (module path), jest.spyOn/vi.spyOn (object+method), jest.fn/vi.fn, and jest.replaceProperty/vi.replaceProperty. iter_python_test_files generalized to iter_test_files (dispatches by extension; backward-compat alias preserved). Dogfood: 31 vi.fn uses detected in tia-chess. +8 tests.

Fixed

• surface TypeScript supertest false positives (BACK-375) — HTTP-route detection no longer counts supertest request(app).get(...) calls as routes. New _callee_obj_is_call() helper in nav_surface_ts.py skips nodes whose callee object is itself a call expression. tia-chess: 39 → 29 HTTP routes (10 false positives removed). +1 test.
• letsencrypt:// false "certs will expire" on snap-installed certbot (BACK-376) — _check_renewal_timer() previously probed a fixed 5-path list that missed snap's unit (/etc/systemd/system/snap.certbot.renew.timer); since snap is certbot's official install method, this false-alarmed on most modern hosts. Replaced the fixed list with a glob scan of the standard systemd unit dirs (*certbot*.timer) and cron dirs (*certbot*) — catches snap (labelled kind: snap), apt (certbot.timer), and the certbot-renew.timer variant, still no-subprocess. Warning reworded to clarify it reflects a renewal file's presence, not active state. +5 tests.

---

Install/Upgrade:

pip install --upgrade reveal-cli

Documentation: https://github.com/Semantic-Infrastructure-Lab/reveal
Full Changelog: https://github.com/Semantic-Infrastructure-Lab/reveal/blob/master/CHANGELOG.md

v0.100.1

Release v0.100.1

Fixed

• TypeScript class methods not extractable by name — reveal file.ts methodName always failed with "Element not found" for class methods. _get_node_name didn't include property_identifier in its direct-child name lookup; TypeScript method_definition nodes use property_identifier for the method name (vs plain identifier for standalone functions). One-word fix; method_definition was already in FUNCTION_NODE_TYPES. +2 tests (TS + TSX).

---

Install/Upgrade:

pip install --upgrade reveal-cli

Documentation: https://github.com/Semantic-Infrastructure-Lab/reveal
Full Changelog: https://github.com/Semantic-Infrastructure-Lab/reveal/blob/master/CHANGELOG.md

v0.100.0

Release v0.100.0

Added

• D005: cross-file hardcoded literal cluster detection — New duplication rule catching the same list/set/tuple literal duplicated across multiple files — the pattern M104 (per-file) structurally cannot see. On first check per project it scans all .py files under the project root, indexes each literal by its order-independent value set, and flags clusters spanning 3+ distinct files (thresholds: literal ≥5 items, ≥3 files; stable names like __all__/*_format/fixtures exempt). Reuses I002's scan-once-cache-per-root strategy. Dogfooding reveal's own tree surfaced a real 3-file _SECRET_PATTERNS duplication that turned out to be dead code (see Refactored). +21 tests.
• --with-stats for session listing (BACK-348) — reveal claude:// --with-stats reads each displayed session's JSONL (the displayed set only, ≤20) and adds MSGS and DUR columns showing total message count and elapsed duration (1h30m, 45m). Default listing is unchanged; a --with-stats hint appears in the count line. Also fixed a BACK-349 CLI gap: --until was missing from parser.py and defaults.py so the CLI flag had no effect (only ?until= query param worked). +13 tests.
• ?until=DATE date range filtering for claude:// session listings (BACK-349) — ?since=2026-06-10&until=2026-06-12 now scopes session listing results. ?until=today normalizes to today's date. +2 tests.
• Sub-agent navigation from parent sessions (BACK-347) — reveal claude://session/<name>/agents now resolves sub-agent JSOLs: shows step/status/type/duration/tokens/prompt and a → reveal <path> navigation hint for each Agent tool invocation. +4 tests.

Fixed

• Help guides no longer leak YAML front matter — Every help://<topic> guide rendered its raw --- front-matter block (title, help_topic, help_category, help_token_estimate, help_description) ahead of the actual content, in both the default and /full views — noise for every human and agent. _load_static_help now strips the leading front-matter fence before section/truncation, so all render paths are clean; a bare doc or one merely using a --- horizontal rule is untouched. +5 tests.
• D005 wasted ~23s before bailing on oversized trees — The file-count ceiling was checked only after rglob materialized the entire .py list, so a too-large root (e.g. a marker-less parent aggregating many projects) walked and parsed everything before skipping. Collection now aborts the moment the count crosses the ceiling (no materialization, no parsing) — an over-ceiling root bails in ~0s. The 5,000-file ceiling is now overridable via REVEAL_D005_MAX_FILES (mirrors I002's env var). +5 tests.
• "No structure available" dead end now shows fallback hints (BACK-336) — All three sites in display/structure.py that print "No structure available" now follow up with Hint: reveal <path> --grep 'pattern' | reveal <path> --show-ast so users have a clear next step instead of a dead end. +3 test assertions.
• claude:// ?summary renderer reordered for usefulness (BACK-343) — Files-changed list and final assistant snippet now lead the summary; tool success rates and message sizes are secondary. Renderer-only change.
• --until today was a silent no-op; microsecond boundary excluded valid files (BACK-364) — --until today now normalizes to today's ISO date before comparison (was appending T23:59:59 to the literal string 'today', making every file pass). Boundary changed from T23:59:59 to T23:59:59.999999 so files modified in the last second of the day are included. Fixed in both adapter.py and sessions.py. +7 tests.
• tree-sitter Node finalized off-thread no longer emits noise (BACK-368) — A tree-sitter Node (a PyO3 unsendable object) created on the main thread could be swept by a cyclic-GC pass that fired on a ThreadPoolExecutor worker (L002 link checks, claude post-process stats), finalizing it off-thread and printing RuntimeError: ... is unsendable, but is being dropped on another thread. New utils.threadsafe.main_thread_gc keeps cyclic collection on the main thread for the duration of those pools (root-cause fix, not a warning filter). Full test suite now emits zero warnings. +4 tests.

Changed

• M104 detects set/tuple/frozenset collections (BACK-366) — The hardcoded-list rule previously walked only ast.List, missing set/frozenset/tuple-based extension and skip-dir lists (the common de-duped form) — including reveal's own code_exts set. It now resolves list/set/tuple literals plus frozenset()/set()/tuple()/list() call-wrappers, counts set/tuple values in lookup-table dicts, and adds a DIRECTORIES classification so skip-dir lists no longer mislabel as FILE_EXTENSIONS. +8 tests.
• --with-stats JSONL reads parallelized (BACK-362) — Session stats are now fetched concurrently via ThreadPoolExecutor(max_workers=min(8, N)) instead of sequentially, improving reveal claude:// --with-stats speed on large session corpora.

Refactored

• Removed dead _mask_secrets duplication in claude adapter (D005 finding) — D005 flagged _SECRET_PATTERNS duplicated across 3 files; 2 of the 3 copies were dead (claude/adapter.py's method is only self-recursive with no entry point; claude/handlers/system.py's copy is never called by any handler). Git history confirms claude config masking was never wired up — the original _get_config never called it, and the BACK-241 handler split copied the dead code forward. No security regression: claude _get_config returns a curated summary (project paths, MCP names, allowed-tools, flag allowlist), not raw values; only codex (which dumps full parsed TOML) needs masking and keeps its single live copy. Replaced a fossil test (which exercised the dead helper in isolation) with one asserting the real no-leak property of _get_config().
• Registry-derived is_code_file — CATEGORY attribute on analyzers (BACK-369) — Added CATEGORY='code'|'data'|'doc'|'config' to every registered analyzer via @register() (default 'code'); 25 non-code extensions explicitly marked (JSON/YAML/XML/CSV/TOML → data; Markdown/HTML/office docs → doc; INI/TOML/HCL → config). New registry.get_code_extensions() (frozenset, lru_cache) replaces the 35-entry hardcoded set in is_code_file(). Side effects: drops ~12 un-analyzable extensions (.cr, .glsl, .jl, .nim etc) that were listed but had no analyzer; adds ~20 that were missing but are fully supported (.mjs, .cjs, .bat, .m, .mm, .r, .sv, .svh etc). +9 tests.
• Directory skip-list centralized into reveal.defaults (BACK-367) — The set of directories to skip during directory walks was redefined, with drifting contents, in 9 modules (ast/stats/surface/grep/pack/file_checker/patches/nav_*); one walker skipped .ruff_cache/htmlcov while another descended into it. A single canonical SKIP_DIRECTORIES superset now lives in reveal.defaults and every site aliases to it; M104 (BACK-366) guards against reintroducing a scattered literal. Net behavior change: narrower sites (notably --grep) now uniformly skip the full dependency/build/cache set.
• claude/adapter.py split into focused modules (BACK-363) — Extracted schema.py (all _SCHEMA_* constants, 261 lines) and handlers/post_process.py (all _post_process_* functions + _slice_list, 157 lines). adapter.py reduced from 1,383 → 1,011 lines (−27%).
• get_file_blame helpers extracted (BACK-361) — Five sub-concerns extracted from a 137-line cx:30 zero-coverage function into independently testable helpers: _resolve_blame_commit, _read_blob_lines, _format_blame_hunks, _read_blame_ignore_revs, _detect_noise_commits. +17 tests.

---

Install/Upgrade:

pip install --upgrade reveal-cli

Documentation: https://github.com/Semantic-Infrastructure-Lab/reveal
Full Changelog: https://github.com/Semantic-Infrastructure-Lab/reveal/blob/master/CHANGELOG.md

v0.99.0

Release v0.99.0

Added

• claude://session/<name>/prompts — human-typed prompts only (BACK-340) — /user encodes tool-result turns as role: user, so a session with ~15 real prompts shows ~170 "user" messages. The new /prompts resource filters to messages with at least one type: text block, excluding pure tool_result wrappers. New get_human_prompts() in analysis/messages.py, dedicated _render_claude_user_prompts renderer, claude_user_prompts output type in the schema. /user is unchanged for backward compatibility. (risen-scepter-0614)
• ?snippet=N for cross-session search context (BACK-342) — reveal 'claude://sessions/?search=term&snippet=300' controls the characters of context around each match (default 120, clamped 60–500). Threaded window_chars through search_sessions() → search_sessions_for_term() → _extract_first_snippet() → _collect_block_matches() → _search_block() → _find_excerpt(). Removed the hard 200-char cap in the cross-session renderer (within-session search renderer cap unchanged). (risen-scepter-0614)
• claude://session/<name>/timeline renderer (BACK-344) — /timeline was routing correctly to get_timeline() data (events with timestamps, tool names, content previews) but had no renderer — every call fell back to a key/value dump of the list. New _render_claude_timeline shows [idx] LABEL msg=N HH:MM preview per event with distinct formatting for tool_call (shows tool name), tool_result (shows tool name + ok/err status + preview), thinking (shows token estimate), assistant_message/user_message (shows content preview). (prairie-squall-0614)

Fixed

• claude:// session titles: skip harness XML and single-word acks (BACK-341) — UUID-named sessions (Windows / multi-user, no session badge) showed titles like <local-command-caveat>... or aye. _parse_jsonl_line_for_title now skips candidates matching ^<[a-z] (harness-injected XML tags: <local-command-caveat>, <command-name>, <bash-input>, <bash-stdout>) and candidates shorter than 5 chars (aye, ok). A < mid-sentence (Is x < y?) is preserved. Additionally, _scan_jsonl_for_title now scans assistant Bash calls for session badge "..." and returns it when present — the listing path is now badge-aware, matching the detail view. (risen-scepter-0614)
• claude:// /thinking emitted noise for encrypted sessions (BACK-345) — Recent Claude Code versions write thinking blocks with an empty thinking: field (content redacted client-side before JSONL). A 173-turn session produced 141 lines of empty dividers. Fix: if all blocks have empty content, emit one summary line ("N thinking blocks — content unavailable (encrypted)"). If some blocks have content (older pre-encryption sessions), render only non-empty blocks silently skipping empty ones. (prairie-squall-0614)
• claude:// /messages undiscoverable — missing from schema elements and help (BACK-346) — /messages is the best resource for reading narrative turns (assistant text only, strips tool-call-only messages) but was absent from _SCHEMA_ELEMENTS, _SCHEMA_EXAMPLE_QUERIES, and _get_help_examples. Added messages and prompts to _SCHEMA_ELEMENTS with descriptions; added /messages example entry to both query lists. (prairie-squall-0614)

Tests

• +30 tests in tests/test_claude_adapter_gaps.py — TestGetHumanPrompts (7), TestPromptsRoute (3), TestParseJsonlLineForTitleXmlSkip (6), TestParseJsonlLineForTitleMinLength (5), TestScanJsonlForTitleBadge (4), TestSnippetWindowParam (5). Full suite: 8951 passed, 22 skipped. (risen-scepter-0614)
• +18 tests — TestBack344TimelineRenderer (8), TestBack345ThinkingEmptyBlocks (5), TestBack346MessagesDiscoverable (5). Full suite: 8969 passed, 22 skipped. (prairie-squall-0614)

Known gaps (filed, not yet fixed)

• BACK-347 no sub-agent session navigation. BACK-348 listing lacks duration/message-count columns. BACK-349 no ?until=DATE range filter. See internal-docs/feedback/CLAUDE_ADAPTER_RESOURCE_AUDIT_2026-06-14.md.

---

Install/Upgrade:

pip install --upgrade reveal-cli

Documentation: https://github.com/Semantic-Infrastructure-Lab/reveal
Full Changelog: https://github.com/Semantic-Infrastructure-Lab/reveal/blob/master/CHANGELOG.md

v0.98.0

Release v0.98.0

Fixed

• Zig: doubled function names in display — _build_function_signature returned "sin(value)" but the display layer renders f"{name}{signature}", producing sinsin(value). Changed to return params only: "(value)" or "" for zero-param functions. (hehacapo-0613)
• Zig: named extraction failing — reveal file.zig funcname raised an error because base extract_element searches Python/Go/Rust node type names, none of which match Zig's Decl/FnProto tree. Added ZigAnalyzer.extract_element override walking Decl nodes. Verified against /opt/zig/lib/std/math.zig. (hehacapo-0613)
• reveal check/hotspots hang on non-Python projects — BACK-338 — Three layered defects in I002's _find_project_root caused the import-graph build to scan the entire filesystem when a stray ancestor __init__.py existed above a non-Python project. (A) Pass-2 __init__.py walk rewritten to only ascend a contiguous chain rooted at the target's own dir — a non-package directory never ascends. (B) Pass-1 now checks _PROJECT_MARKERS (pyproject.toml, setup.py, package.json, go.mod, Cargo.toml) before falling back to the init-chain walk. (C) _i002_preload was unconditional — now skips the import-graph build when I002 is not in the effective rule set (e.g. --select C901). Hardening: _collect_raw_imports aborts to an empty graph + logged WARNING past REVEAL_I002_MAX_FILES (default 20 000) — future mis-detection degrades to a logged skip, never a hang. (bright-singularity-0613 filed, hehacapo-0613 fixed)

Tests

• +27 tests — TestI002ProjectRootBACK338 (8 tests) in test_rules.py; Defect C + preload-signature tests in test_file_checker.py; Zig display and named-extraction tests in test_zig_analyzer.py. Full suite: 8921 passed, 22 skipped. (hehacapo-0613)

---

---

Install/Upgrade:

pip install --upgrade reveal-cli

Documentation: https://github.com/Semantic-Infrastructure-Lab/reveal
Full Changelog: https://github.com/Semantic-Infrastructure-Lab/reveal/blob/master/CHANGELOG.md

v0.97.0

Release v0.97.0

Added

• codex:// adapter — full Claude Code session explorer — Phase 1+2+3 complete. reveal codex:// lists all sessions with token counts and summaries; reveal 'codex://SESSION' shows full overview (goal, memories, pipeline, config, rules); reveal 'codex://SESSION/messages' browses conversation with role/token/tool breakdowns; reveal 'codex://SESSION/tools' summarizes tool calls with success/failure counts; reveal 'codex://SESSION/shell' shows all bash commands run; reveal 'codex://SESSION/errors' surfaces failures and retries; reveal 'codex://SESSION/workflow' traces task creation and completion; reveal 'codex://SESSION/timeline' shows conversation arc with phase detection; reveal 'codex://SESSION/?goal' extracts session goal; ?search=TERM filters to sessions containing a term with contextual snippets. Adapts to Claude Code's JSONL format with graceful handling of all message types (user, assistant, tool_result, tool_use). (onyx-spectrum-0524)
• Agent discoverability — Sprint 1-3 — claude:// schema gains ?params, ?last, ?tool, ?turns, ?since query parameters; help:// decision_tree and recipes entries guide agents to the right resource for common tasks; MCP_SETUP.md cross-resource workflow examples; README.md AI agents section; AGENT_HELP.md codex row added. 17 new tests across claude adapter schema parity and help registry. (roaring-tide-0524)

Fixed

• codex:// Path(None) crash — dict.get('key', default) returns None (not default) when the key is present with a NULL SQLite value; rollout_path column is sometimes NULL, causing Path(None) TypeError. Fixed with explicit None check. (roaring-tide-0524)
• help:// cli_flags renderer dropping descriptions — Sprint 1-3 changed cli_flags from list to dict; _render_schema_cli_flags was iterating dict keys, silently dropping all flag descriptions. Fixed by iterating .items(). (roaring-tide-0524)
• codex:// grand_total tokens using per-request delta — grand_total was reading last-turn per-request usage delta instead of cumulative session total. Added get_grand_total_tokens() to analysis/messages.py. (roaring-tide-0524)
• codex:// tokens_used == 0 falsy guard — tokens_used or jsonl_total silently replaced a legitimate 0 with the JSONL total. Changed to is not None guard. (roaring-tide-0524)
• codex:// content search false positives — sessions where the search term appeared only in tool payloads (not user/agent turns) were included in results with empty snippet lists. Fixed by only appending sessions with at least one snippet. (roaring-tide-0524)

Chore

• 7 mypy errors fixed in new codex adapter code: str() casts in overview.py, messages.py, timeline.py; explicit last typing in _parse_token_usage; Callable annotation on dynamic dispatch in adapter.py; removed dead # type: ignore in handlers/system.py. (roaring-tide-0524)

Tests

• 13 regression tests added in tests/adapters/test_codex_adapter.py: TestRegressionNullRolloutPath, TestRegressionContentSearchFalsePositive, TestRegressionGrandTotalCumulative, TestRegressionTokensUsedZero, TestRegressionCliFlags, and 8 more covering schema parity and edge cases. (roaring-tide-0524)

---

---

Install/Upgrade:

pip install --upgrade reveal-cli

Documentation: https://github.com/Semantic-Infrastructure-Lab/reveal
Full Changelog: https://github.com/Semantic-Infrastructure-Lab/reveal/blob/master/CHANGELOG.md

v0.96.1

Release v0.96.1

Fixed

• --type <filetype> on a directory now filters files instead of silently returning zero results (crackling-current-0524) — reveal . --type markdown (and --type python, --type yaml, etc.) routed to the AST query handler (ast://.?type=markdown), which found no AST nodes of that name and returned 0 results with no error. Root cause: _handle_directory_path in routing/file.py unconditionally treated --type as an AST node-type filter even when the value was a file-format name. Fix: known file-type names (markdown→.md,.markdown, python→.py, yaml→.yaml,.yml, etc.) are translated to extension filters and applied to the directory display. Unknown values (AST node types like class, function) still route to the AST handler as before. Combination --name Foo --type function also continues to work — --name presence always triggers AST routing.

---

---

Install/Upgrade:

pip install --upgrade reveal-cli

Documentation: https://github.com/Semantic-Infrastructure-Lab/reveal
Full Changelog: https://github.com/Semantic-Infrastructure-Lab/reveal/blob/master/CHANGELOG.md