MSN's feed for a stock is often mostly general market news: of 10 items,
ADRO, BREN and CUAN had none about the company, BBRI one, yet they were
presented as the company's news. Fetch at least 50 items, label each
`relevance: company` (tagged by MSN with the stock's instrument id, or
naming the ticker or the company) or `market`, and return company items
first, newest first, trimmed to --limit.
Company names come from the quote's short and display names, best
effort: MSN tags Telkom articles with its NYSE ADR, not TLKM, so names
catch what tags miss. Classification stays a pure function in parse_news.
Output change, intentional: news JSON gains `relevance`, the table an
ABOUT column (goldens regenerated), and the get_company tool description
explains the field (tools.json and the website tools page regenerated).
News cache keys change so entries without the field are not reused.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01V3syn2fN2LNo5pvTSqc6eh
From the Codex re-gate: the planted-binary check used only ~/.cache, so
the old macOS fallback (~/Library/Caches) would have escaped it.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01V3syn2fN2LNo5pvTSqc6eh
The release-gate Codex review found the relative-path fix incomplete:
with XDG_CACHE_HOME unset and HOME empty or relative, the fallback was
`.cache/...` under the current directory, so a checkout could plant the
binary the launcher runs. Use the home cache only when the home path is
absolute; otherwise the package's bin/ is the only location. npm smoke
plants such a binary and checks it is never run (it was, before this).
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01V3syn2fN2LNo5pvTSqc6eh
From a second Codex review (gpt-6.1-sol, xhigh):
- A relative XDG_CACHE_HOME (invalid per the XDG spec) made the launcher
look for, and run, a binary under the current directory, which an
untrusted checkout could supply. Use it only when absolute.
- The cache dir now includes <platform>-<arch>, so a home directory shared
across machines cannot hand an arm64 host an x64 binary.
- SIGINT sent to the launcher's PID alone during the first-run install is
forwarded to the installer (it was ignored); the launcher exits 130.
- npm smoke runs the first-run install with an empty user cache and checks
the package binary was created, so a cached binary cannot pass for it.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01V3syn2fN2LNo5pvTSqc6eh
- When the package's bin/ is not writable (a root-owned `sudo npm i -g`
whose postinstall npm 12 skipped), install the binary to a per-user
cache ($XDG_CACHE_HOME, ~/.cache, ~/Library/Caches on macOS, under
<package>/<version>/); the launcher looks in both. Previously the first
run failed with EACCES and advice to rerun npm install.
- Run the first-run installer asynchronously with SIGTERM/SIGHUP
forwarded, so a supervisor stopping the launcher stops the download;
the installer removes its partial file and the launcher exits by the
same signal.
- Verify an IDX_BINARY_URL local file's checksum on the copy, not the
source, so the installed bytes are the checked ones.
- npm smoke covers the read-only fallback.
The last three come from a Codex review.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01V3syn2fN2LNo5pvTSqc6eh
From the checkout, agents read its AGENTS.md and docs: Hermes concluded
from docs/IDX_NEWS.md that news "is not implemented" and failed the news
case it had passed before the doc existed. Users' agents never see our
repo, so run every agent from an empty directory under the work dir.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01V3syn2fN2LNo5pvTSqc6eh
The release-gate Codex review: news passed on get_company with only
`sections: ["profile"]`, history on any period, search on any query.
Require `sections: ["news"]`, `period: "1mo"`, `filter: "top-performers"`
and a query containing "Salim"; all 24 calls from the last six-harness
run still pass.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01V3syn2fN2LNo5pvTSqc6eh
The six cases never asked for get_company (where news lives),
get_price_history, screen_stocks or search_investors; across six agents
only one called get_company and none screen_stocks. Cover each.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01V3syn2fN2LNo5pvTSqc6eh
The dev shell's bash has no compgen (no programmable completion), so the
npx-install check failed every agent.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01V3syn2fN2LNo5pvTSqc6eh
From a Codex review of the harness:
- Drop target/debug from PATH in --npx mode (the dev shell adds it), so
a missing package bin cannot fall through to the local build, and fail
an agent whose npm cache holds no installed idx-mcp binary.
- The live `changes` case must diff the two latest releases, read from
the synced snapshot, not any two dates.
- Hermes: `hermes -z` waits ~15s for MCP tools regardless of
connect_timeout, so a slow cold install lost the first case. Warm up
with `hermes mcp test idx`, which installs through Hermes' own npx and
honors connect_timeout (now 300s).
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01V3syn2fN2LNo5pvTSqc6eh
opencode 2 (@opencode/cli) moved MCP servers to `mcp.servers`, turns
Code Mode on by default, replaced `permission` with ordered
`permissions` rules, and dropped `run --pure` for `--standalone` (runs
otherwise go through a shared background service). Pick the config and
flag from `opencode --version`.
Its `execute` tool stays allowed: with it denied, opencode 2.0.30 never
offers the idx tools to the model in `opencode run`, though the server
connects before the first step.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01V3syn2fN2LNo5pvTSqc6eh
opencode kept its built-in tools, so it could route around the server
under test: on published 0.3.1 it answered a quote from web search, and
on the live `changes` case it queried the ownership SQLite file with
bash and timed out. Deny every tool but `idx_*`, as Claude, pi, and
Hermes already are.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01V3syn2fN2LNo5pvTSqc6eh
The `changes` case names the fixture releases (2026-01-30, 2026-02-27),
which the live snapshot does not have, so under --live/--npx it failed
agents that correctly declined to answer about missing releases and
passed ones that called the tool anyway. A case's `live` fields now
replace the mock ones; this one asks about the two most recent releases.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01V3syn2fN2LNo5pvTSqc6eh
All agents shared one npm cache, so only the first ran a cold
`npx -y idx-mcp` install and the others reused its installed binary,
never exercising their own first-run install.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01V3syn2fN2LNo5pvTSqc6eh
- agent-smoke `--npx <spec>` registers `npx -y <spec>` (the published package
or an `npm pack` tarball) with a fresh npm cache, instead of the local build,
so each agent's own npx handling and a cold install are tested.
- `nix develop .#agents-fhs` (Linux) runs agents in an FHS sandbox: Hermes
replaces a bare `npx` with its own generic-Linux Node, which NixOS cannot
load. `agent-fhs <cmd>` runs anything inside it.
- Hermes: only the idx toolset (with its terminal it repaired a broken
server mid-run and passed), the Node toolchain fetched before the first
case (`hermes -z` waits ~15s for MCP tools), JSON config.yaml (as profiles
write it), linked plugins for plugin model providers, and a copy of its
mcp-stderr.log per run.
- A failed run names its cause inline, since output is deleted without --keep.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01WB13omP6h7YeyJx7GNtUnF
npm 12 blocks install scripts of unapproved packages under `npx` and
`npm install -g` (its allow-scripts policy). The postinstall download never
ran, so `npx -y idx-mcp` exited with "installed binary not found" and MCP
clients saw only "Connection closed". Hermes Agent hits this by default: it
runs a bare `npx` with its own npm 12.
The launcher now runs the installer itself when the binary is missing,
logging to stderr so the idx-mcp stdout channel stays clean. This also covers
pnpm 10+ and bun, which skip dependency scripts by default.
npm-smoke installs idx-mcp with --ignore-scripts and checks that the first
run still starts cleanly.
Fixes#41
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01WB13omP6h7YeyJx7GNtUnF
Starlight's pagination cards use a 2xl title, a shadow and 1rem padding,
which outweighs the page content above them. Shrink them to a small label
over a base-size title.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01Ep2qFFvuFEnBvLpoLxxdEi
The site card shows BBCA's KSEI ownership split under the tagline; each docs
page gets a card with its section, title and description. Cards are rendered
at build time with satori and sharp from the bundled Plex and JetBrains Mono
fonts, and a Starlight route middleware points og:image at them.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01Ep2qFFvuFEnBvLpoLxxdEi
- .md output rewrites MDX-only syntax (Tabs, TabItem, asides, the MCP
client component) into plain Markdown.
- MCP client setups live in one data file used by the landing page, the
setup page and agents/install; the smoke-test claim names the five
agents agent-smoke.sh actually drives (not Cursor).
- The landing page reads the version from Cargo.toml; the hand-copied BBCA
data stays dated, and the design doc no longer promises capture.sh.
- MCP overview counts 11 tools plus the opt-in ownership_sync.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Landing page plus docs for install, quickstart, config, the stocks and
ownership CLI, output and exit codes, the MCP server and per-agent setup,
agent usage, and an agent-led install page (agents/install.md) that a
coding agent follows from a one-line prompt.
- Every docs page is also served as Markdown at <path>.md; llms.txt and
llms-full.txt come from starlight-llms-txt; a SKILL.md for skill-based
harnesses is under /skills/idx/.
- The CLI reference and MCP tools pages are generated from idx --help and
the tools golden; a new web-generated flake check fails when they drift.
- nix develop .#web provides bun; Cloudflare Pages builds web/ on pushes.
- Package homepage metadata points at https://idx.rasyidanaf.com.
- .cloudflare/ (the cf CLI's account cache) is git-ignored.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Landing page, docs sitemap, agent features, generated references, and
Cloudflare Pages hosting at idx.rasyidanaf.com. For review before build.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
`idx version` was the only way to get the version; the conventional flag
failed with a usage error. The flag is hidden from help, which keeps
documenting `idx version`. clap's bash completions list hidden flags, so
the completions-bash golden gains -V/--version at the root (intended).
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Patch release for the fixes from the Codex review of 0.3.0: ticker
validation and encoded cache file names (path traversal via a ticker),
cross-process ownership DB locking, history-preserving snapshot sync,
and MCP input limits. MSRV is now 1.89.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Final Codex (gpt-6.1-sol, xhigh) pass over c65982f confirmed every
earlier item fixed and found:
- high: readers skipped the lock on any error, so a lock file owned by
another user (who can still sync) let a reader run unlocked. Readers
now open an existing lock file read-only for the shared lock, and read
unlocked only when no lock file exists and none can be created.
- low: sync resolved the symlinked path for the install target and the
lock separately; it now resolves once and passes that path to the
lock callback, so both always name the same database.
Verified live: a reader with a read-only lock file waits for an
exclusive holder; install waits for readers; a symlinked path keeps its
link. CLI goldens unchanged; live A/B vs 0.2.4: 15 passed, screen hit a
transient MSN 408 on the old build only (retries pass), discover hits
the IDX 403 on both (exit 4 vs 1, intended).
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Fourth Codex (gpt-6.1-sol, xhigh) pass over 5e0831d, each item verified
before fixing:
- sync inspects the local database under the shared lock, so another
sync cannot swap it under that connection.
- symlinks are resolved by reading the link (db::resolve_db_path), not
canonicalize, so a symlinked path finds the same lock and install
target even while the real file is missing (fresh install, mid-swap)
or the link dangles.
- readers go ahead unlocked when no lock file can be created; nothing
can replace the database there. (A read-only data directory still
fails in SQLite's WAL setup, as in 0.2.4.)
- XLSX shared strings are capped at 2,000,000 entries.
Verified live: an install waits for a reader's shared lock; a symlinked
db_path waits on the real path's lock and stays a symlink. CLI goldens
unchanged; live A/B vs 0.2.4: 16 passed, discover hits the IDX 403 on
both builds (exit 4 vs 1, intended).
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Third Codex (gpt-6.1-sol, xhigh) pass over 1e55138, each item verified
before fixing:
- readers now hold the ownership lock shared for the life of their
connection, so a snapshot swap (exclusive) never happens under an
open SQLite connection whose WAL cleanup could hit a later import.
- the lock file name uses the resolved database path, so a symlinked
path shares the real database's lock; sync also installs at the
resolved path, keeping a symlinked path a symlink (it used to be
overwritten with a regular file).
- sync downloads without the lock and takes it only to recheck and
install; snapshot downloads have connect (15 s) and total (10 min)
timeouts. idx-mcp drops its in-process RwLock: idx-core's lock covers
it across processes.
- XLSX parsing caps rows at 1,000,000 and checks a shared string's size
before copying it.
- the temp snapshot's -wal/-shm files are removed instead of left in the
data directory.
Remaining, by design: `changes` loads all rows of two releases from the
local database before filtering; its size is bounded by the snapshot.
CLI goldens unchanged. Live A/B vs 0.2.4: 16 passed; discover hits the
IDX 403 on both builds (exit 4 vs 1, intended).
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Second Codex (gpt-6.1-sol, xhigh) pass over 60634d3, each item verified
before fixing:
- new: the sync recheck hashed an existing empty database before schema
setup, so the first sync falsely reported an external change. Hash
after setup; a published snapshot already has the schema and keeps
its bytes.
- sync, import, and entity edits now hold a cross-process lock
(<db>.lock, std File::lock), so a sync cannot replace the database
under an import running in another process. Import validates and
downloads before taking the lock, keeping its error order. MSRV
1.85 -> 1.89 for File::lock.
- the row-count rebuild branch also never downgrades history.
- cache reads no longer delete corrupt entries (a concurrent writer may
have just replaced them); the next put overwrites atomically.
- idx-mcp validates the changes limit before querying and runs each
requested section or analysis once.
- XLSX parsing bounds total cell text, so one large shared string
referenced from many cells cannot exhaust memory.
- PDF/archive percentages must be 0-100%; HHI uses saturating math.
CLI goldens unchanged. Live A/B vs 0.2.4: 16 passed, discover hits the
IDX 403 on both builds (exit 4 vs 1, intended).
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Findings from a Codex (gpt-6.1-sol, xhigh) review of v0.2.4..HEAD, each
verified in the code before fixing:
- high: ticker symbols reached cache file paths unchecked, so a symbol
like `../../REPORT` read (and deleted, if unparseable) .json files
outside the cache, reachable through MCP get_quotes. resolve_symbol now
accepts only ticker characters, and cache keys are encoded as single
safe path components.
- high: sync decided before downloading and replaced the database
without rechecking; an import that landed meanwhile was erased. It now
rechecks the local state before installing, and idx-mcp runs
ownership_sync exclusively with ownership reads.
- replaces_sqlite_sha256 never downgrades history: the client requires
the rebuild to keep at least the local release count, and the
publisher lists a replaced build only then.
- cache writes are atomic (temp file + rename), so concurrent readers
never see, and delete, a partial entry.
- idx-mcp caps symbols (20, deduplicated) and limits, and runs at most 4
tool calls at once.
- each Yahoo client gets its own cookie jar, paired with its crumb.
- ZIP entries (XLSX, KSEI archives) are read with a 256 MiB bound;
non-ASCII archive dates and huge percentages are parse errors instead
of panics or overflow.
- agent-smoke --keep deletes every copied agent config, not only auth
files.
CLI goldens unchanged; the MCP tool golden changes only to state the
new caps.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
- Require bash 4.4+ up front, and match answer facts with jq regexes
(Oniguruma, (?i) everywhere) instead of GNU-only grep -P.
- Unparseable agent output fails that run instead of aborting the
matrix; check_run now also reports unparseable tool calls, which
previously passed the tool check silently.
- Copy the Hermes model block with awk up to the next top-level key, so
a final or differently ordered block is handled.
- --keep keeps run output but deletes the copied agent credentials.
- docs: full path to the tool golden.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
- tests/agents/agent-smoke.sh runs tests/agents/cases.json through
Claude Code, Codex, pi, Hermes Agent, and opencode headless against a
local idx-mcp build, each with an isolated config, and checks the
tools called, their key arguments, and (mock data) facts in the
answer. Opt-in; never part of the gate.
- crates/idx-mcp/tests/golden/tools.json pins tool names, descriptions,
and input schemas, so changes to what agents see are reviewed.
- nix develop .#agents adds Claude Code, Codex, and opencode from
nixpkgs after the host PATH.
First run: 30/30 (5 agents x 6 cases, mock data).
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
- Sync hashed only the main database file, but local edits can sit in
the -wal sidecar, so an edited database could match an earlier build
and be replaced. Checkpoint the WAL before hashing; if changes remain,
never take the rebuilt-snapshot path.
- MCP tools rejected nothing for empty symbols/sections/kinds and
returned an empty success; they now fail with INVALIDINPUT, as the CLI
requires input.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Bump the workspace and npm packages to 0.3.0. The release workflow
builds and attaches idx-mcp binaries next to idx, the npm launcher and
installer pick their binary from package.json so npm/idx-mcp reuses
them, and the npm smoke test covers both packages.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
A snapshot republished for the same as-of date (the TRUE ticker fix)
kept the date, latest release, and release count, so sync saw nothing
to do and existing installs kept the bad data without --force.
The manifest gains an optional snapshot.replaces_sqlite_sha256 listing
earlier builds for that date. Queries never write to a synced file, so
a local database byte-identical to a listed build is replaced; one that
matches nothing (local imports or edits) still needs --force. The
publisher fills the list on a same-date rebuild. Older clients ignore
the field.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Phase 4 of docs/DESIGN.md. Eleven grouped tools (quotes, history, screen,
company, analysis, compare, shareholders, investor holdings, search,
ownership changes, status); ownership_sync is listed only with
--allow-sync. Live results keep the Fetched envelope, errors use the CLI's
{error, code, message} body, and investor tools on an empty database
point at the sync path.
idx-core gains Period/Interval FromStr, Investor::status, a ticker
filter on changes, and Display for ErrorCode (now used by the CLI too).
CLI output and goldens are unchanged.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
comfy-table counted ANSI escapes as cell width, so colored signal and
change cells pushed the right border out of line on a terminal. Its
custom_styling feature measures the visible width instead. Piped output
has no color and is unchanged.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
The IDX above-1% workbook stores the share code TRUE (TRINITI DINAMIK)
as a boolean cell holding 1, so its holders were imported under ticker
"1". Map t="b" cells to what Excel displays.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
The IDX above-1% workbook stores the share code TRUE (TRINITI DINAMIK)
as a boolean cell holding 1, so its holders were imported under ticker
"1". Map t="b" cells to what Excel displays.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Phase 3c of docs/DESIGN.md. `ownership` handlers are now argument
parsing, one idx-core call, and rendering.
- investor.rs: Investor (also Idx::investor()) with releases, ticker,
entity, search, cross_holders, concentration, flow, changes, graph,
unresolved, map_alias, merge_entities, discover, sync, and import.
Import orchestration (input resolution, download, SHA-256 dedup,
PDF/XLSX/archive parsing, entity resolution, release write) moves
from the CLI unchanged.
- Typed inputs: ImportSource, HolderSource, ConcentrationRank, and
parse_report_family, with the same error messages.
- CoreConfig gains an ownership_db override; db::open_db_at opens an
explicit path. idx-cli no longer depends on rusqlite, sha2, or ureq
at runtime.
- Behavior note: an invalid `ticker --source` or `concentration --by`
is now rejected before the database file is opened (and created);
the error and exit code are unchanged.
- Fix: building idx-cli without the ownership feature no longer warns
about the ownership-only `blue` style.
All 143 goldens unchanged; 170 core + 7 CLI unit tests and 62 CLI tests
pass; clippy clean for default, release, testing, and no-default
feature sets. Live A/B against main: 16/16 stocks identical; ownership
discover got IDX 403 on both (exit 4 vs 1 is the phase 1 change).
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Phase 3b of docs/DESIGN.md. `stocks` handlers are now argument parsing,
one idx-core call, and rendering.
- Idx (client.rs) is the entry point: Idx::new(CoreConfig) with
market(), financial(), and analysis(). Live lookups return
Fetched<T> { data, source, fetched_at, stale } and take a CachePolicy
(Default | Offline | Refresh, from the --offline/--no-cache flags).
- One cached_fetch replaces the CLI's three cache flows; they differed
only in whether a fresh entry is read via get or get_offline when
offline, which yields the same data and diagnostics. Report-building
errors (technical) still skip the stale-cache fallback.
- market.rs: quotes, history (history_source is resolved first so the
CLI's fallback note keeps its position), and a typed screener
(ScreenFilter/ScreenRegion with the same error messages).
- financial.rs: profile, statements, earnings, sentiment, insights,
news, with StatementFilter/EarningsFilter.
- analysis: TechnicalReport moves to analysis::report; Analysis offers
technical, growth, valuation, risk, fundamental, and compare.
- CoreConfig::load() owns provider/exchange/TTL settings from
config.toml and IDX_*; the CLI's IdxConfig adds output and color.
With two different invalid settings at once, the reported one may
differ from before; every single-error message is unchanged.
- Cache gains fresh/any/offline accessors returning fetch metadata.
Unit tests moved with their code (core 157 -> 168, CLI 17 -> 7). All
143 goldens unchanged; live A/B against main: 16/16 stocks identical.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Phase 3a of docs/DESIGN.md. MockProvider moves to api/mock.rs, and it,
the embedded MSN fixture bodies, and the IDX_MOCK_* hooks compile only
with idx-core's new `testing` feature. idx-cli enables it as a
dev-dependency, so `cargo test` keeps the fixture-backed provider while
release binaries no longer embed fixture data or honor
IDX_USE_MOCK_PROVIDER. Parser helpers used only by the mock are gated
the same way.
The packaged-cli flake check no longer relies on the mock; it checks
an offline cache miss (JSON CACHEMISS, exit 3) instead.
Goldens unchanged; clippy is clean with and without `testing`.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
