Bryan hit this on MobaXterm immediately after v0.6.3 shipped:
C:\Users\...\bin\jq.exe: Bad JSON in --rawfile c /tmp/tmp.AlsIcdX9aO:
Could not open /tmp/tmp.AlsIcdX9aO: No such file or directory
error: OAuth token unavailable; run 'larry-auth.sh login' to re-authenticate
error: empty response from API (timeout or network?)
Root cause is the same one v0.5.4 fixed for the OAuth file reads, but now
biting the new tempfile pipeline introduced for argv-overflow avoidance.
The Windows-native jq.exe shipped to MobaXterm via install-larry.sh
cannot resolve Cygwin paths like /tmp/tmp.X or /home/mobaxterm/.larry/...
when they come in as argv arguments to --rawfile / --slurpfile — it
interprets them as Windows paths and the open() fails.
v0.5.4 fixed this for stdin-read cases by piping via bash redirection
(`< "$file"`), since bash handles the cygwin→windows path open before
jq sees a file descriptor. But --rawfile / --slurpfile DO want a path
argument, so the stdin trick doesn't apply — we must give jq a path it
can actually open.
Fix: new jqpath() helper translates a cygwin path to its real Windows
equivalent via `cygpath -w` when cygpath exists (Cygwin / MobaXterm / MSYS).
On Linux and macOS cygpath is absent and the helper echoes the path
unchanged — true cross-platform no-op outside Cygwin.
Wrapped every --rawfile / --slurpfile argv path in larry.sh:
add_user_text --rawfile c ←
add_assistant_blocks --slurpfile b ←
add_user_tool_results --slurpfile b ←
agent_turn payload --rawfile system, --slurpfile messages,
--slurpfile tools ←
agent_turn tool-result aggregation --rawfile c ←
Verified on macOS: jqpath echoes paths unchanged, the 40KB prompt
smoke test from v0.6.3 still works end-to-end.
Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
v0.6.2 fixed the TOOLS_JSON argv overflow but four other call sites had
the same risk pattern — any of them would have crashed under Cygwin's
~32KB argv cap with large user input, large agent responses, or large
tool results:
add_user_text --arg c "$content" ← multi-paragraph prompts
add_assistant_blocks --argjson b "$blocks" ← long assistant turns
add_user_tool_results --argjson b "$blocks" ← chained tool results
agent_turn loop --arg c "$result" ← tool output (up to 250KB
for read_file, 500 lines
for ssh_exec, etc.)
agent_turn loop --arg system "$system_prompt" ← agents/*.md
total ~25KB
All five are now passed via tempfile + --rawfile (for raw strings) or
--slurpfile (for pre-parsed JSON). Same proven pattern as the v0.6.2
TOOLS_JSON fix. Tempfiles are cleaned at every return path.
Verified by pushing a 60KB user prompt through the pipeline on macOS
(also has the larger 256KB argv cap that masked these bugs locally
before, but the codepath now uses files for the large values regardless
of platform). Messages file stored the full 60025-char prompt with no
warnings.
After this commit, the only --arg / --argjson calls remaining all carry
known-small values (UUIDs, version strings, port numbers, etc.).
Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
Bryan hit this on every chat turn from MobaXterm:
jq: Argument list too long
error: OAuth token unavailable; ...
error: empty response from API
warn: turn ended with error
Root cause: agent_turn was building the API payload with
jq -n ... --argjson tools "$TOOLS_JSON" ...
which puts all 21KB of tool-definition JSON on the jq command line as
argv. Cygwin/Windows argv limit is ~32KB total — combined with the other
args this exceeded it and jq failed with E2BIG. Linux/macOS have a much
larger limit (256KB+) so this never showed in local testing.
The downstream errors were cascade noise: jq fails to write payload →
call_api runs with empty/partial payload → various error paths fire.
Fix: write TOOLS_JSON to a tempfile once per agent_turn and reference
it via --slurpfile (same pattern already used for $MESSAGES_FILE).
Tempfile is cleaned up at every return path.
Verified locally — payload now builds cleanly with no argv warnings.
Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
THREE bugs that Bryan hit in v0.6.0:
(1) TOOLS_JSON: unbound variable crash on any unrecognised input.
Root cause: the TOOLS_JSON assignment was a single-quoted string spanning
~40 lines, and apostrophes in tool descriptions (Anthropic's, 'codametrix',
protocol's, etc.) closed the bash string prematurely. Bash then tried to
execute fragments of the JSON as shell commands ("NAME: command not found"
warnings) and TOOLS_JSON never got assigned. With set -u, the first
reference to $TOOLS_JSON crashed the whole script.
Fix: switch to a quoted-EOF heredoc — TOOLS_JSON=$(cat <<'TOOLS_END' ...
TOOLS_END). Heredoc with single-quoted delimiter preserves content
literally — apostrophes, backslashes, all of it. Verified: all 31 tool
defs now parse as valid JSON (including the previously-broken nc_msgs).
Also fixed the pre-existing \\" → \" escape error in nc_msgs.
(2) Slash command brittleness: /ssh-add\ * pattern matched only when args
were present; /ssh-add alone fell through to the catchall and reported
"unknown command". Also failed silently with no usage message on the
too-few-args path.
Fix: rewrote all SSH slash patterns as /ssh-foo* (matches both with and
without trailing args), with a _slash_args() helper that cleanly extracts
the arg portion. Every handler now validates and prints "usage: ..." on
missing args before continuing. New _run_ssh_helper() wrapper centralises
the installed-check and swallows helper exit codes so they don't propagate
into the main loop.
(3) Backspace not working in MobaXterm/Cygwin terminals.
Root cause: terminal sends ^? (DEL) for backspace but stty erase is often
set to ^H (BS) in MobaXterm, so backspace passes through as a literal
character.
Fix: stty erase '^?' at REPL startup (harmless if already correct), AND
switch read_user_input to use `read -e -r -p` which uses libreadline for
line editing — backspace, arrow keys, history all work via readline now,
bypassing the terminal's stty config entirely. Falls back to plain read
on environments without readline support.
Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
NEW lib/ssh-helper.sh implements the full SSH command surface:
hosts/list show configured remote hosts
add <alias> <user@host[:port]> register a new host
remove <alias> remove + clean cred + socket
pass <alias> set/update password (hidden interactive)
setup <alias> open long-lived ControlMaster
close <alias> close ControlMaster
status [alias] show open masters + cred presence
exec <alias> <command...> run command via master
Architecture:
• $LARRY_HOME/.ssh-hosts.tsv — alias \t user@host \t port (3-col)
• $LARRY_HOME/.ssh-creds/<alias> — raw password, mode 0600
• $LARRY_HOME/.ssh-sockets/<alias>.sock — ControlMaster socket
The password is read from disk by sshpass via -f (file argument), so it
never lands in argv or environment. It is used ONCE to open the master;
all subsequent execs multiplex through the socket with no auth. Daily-
rotating passwords: just overwrite the cred file and re-run setup.
SLASH COMMANDS wired in larry.sh REPL: /ssh-hosts /ssh-add /ssh-remove
/ssh-pass /ssh-setup /ssh-close /ssh-status /ssh <alias> <cmd>.
LARRY TOOLS exposed to the LLM:
ssh_status — list aliases + open-master state
ssh_exec — run command on remote via the master socket
Both tool descriptions explicitly tell Larry the password is unreachable
and to ask Bryan to run /ssh-setup if a master is closed. Tool inputs
and outputs never contain the password. Output capped at max_lines
(default 500) with a "[ssh_exec: exit rc=N]" footer.
Bundle updated: MANIFEST + install-larry.sh both now include
lib/ssh-helper.sh. Auto-update will pull it on next launch.
Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
Bryan asked for an easier-to-remember inline PHI marker than {{phi:VALUE}}
and for name forms like SMITH^JOHN / Smith, John / John Smith / JOHN SMITH
to all collapse to the same hash. Both shipped.
INLINE SYNTAX (in addition to the legacy {{phi:VALUE}} which still works):
@@VALUE unbracketed — VALUE has no whitespace
e.g. @@12345 @@SMITH^JOHN @@V789
@@VALUE@@ bracketed — VALUE may contain spaces
e.g. @@John Smith@@ @@Smith, John@@
Parser is 2-pass to disambiguate mixed forms in the same prompt: bracketed
markers are matched first (via grep -oE with a regex that excludes leading/
trailing whitespace inside the brackets), then the unbracketed pass scans
the remaining text. Verified against:
"look for @@12345 in PID.3 for @@John Smith@@ DOB @@01/15/1985 ..."
extracts 4 markers correctly and routes each to its category.
AUTO-CATEGORY DETECTION (lib/hl7-sanitize.sh: detect_category):
pure digits 4-15 → MRN
9 digits with dashes → SSN
date-shaped → DOB
caret or comma → NAME
2+ alpha tokens → NAME
else → MANUAL
CANONICALIZATION (lib/hl7-sanitize.sh: normalize_value):
NAME: lowercase, replace ',^/' with spaces, sort unique alpha tokens
SMITH^JOHN, Smith John, John Smith, JOHN SMITH → "john smith"
DOB: parse to YYYY-MM-DD (GNU date or BSD date fallback)
SSN: strip dashes/whitespace
MRN/MANUAL: trim outer whitespace only
TABLE SCHEMA bumped to 4 columns (token / category / canonical / original).
Legacy 3-column rows still read fine — lookups key on column 3 which is
"canonical" in new rows and "value" in legacy rows (mismatches just create
a new token, no corruption). Detokenize prefers column 4, falls back to
column 3 for legacy compat.
Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
Symptom: OAuth login succeeded on the work box but cmd_status emitted three
'jq: error: Could not open file' lines and showed empty fields. Same pattern
would have hit every subsequent chat turn via larry.sh's MESSAGES_FILE reads.
Root cause: install-larry.sh fetches a Windows-native jq.exe on cygwin/
mobaxterm platforms. Windows jq can't resolve Cygwin paths like
/home/mobaxterm/.larry/.oauth.json when they come in as argv arguments
(it interprets the leading slash as a Windows root). Bash's `>` redirection
worked because bash itself does the path open and hands jq an fd — the
read-side calls were passing the path string directly.
Fix: every read-side jq call now uses stdin redirection (`jq '...' < file`),
where bash does the open. Universal:
- Linux/macOS native jq: identical behavior (was already file-open-from-bash)
- MobaXterm/Cygwin/Git Bash with Windows jq.exe: now works
- WSL: works (Linux-native jq, same as Linux)
- Native PowerShell/cmd: doesn't apply — larry-anywhere is a bash script
Changes:
- lib/oauth.sh: new jqf() helper; 10 sites converted. Refactored cmd_refresh
to drop --slurpfile (which can only take a path) — pre-reads the previous
refresh_token, then uses --arg.
- larry.sh: add_user_text / add_assistant_blocks / add_user_tool_results
now pipe $MESSAGES_FILE via stdin too.
Verified: cmd_status against a real token file produces clean output, no
jq errors. Syntax check passes both files.
Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
Confirmed against live token endpoint (HTTP/2 200, valid sk-ant-oat01- and
sk-ant-ort01- tokens returned) that the v0.5.2 0.5.2 request body and
URLs were correct — the EXCHANGE itself works fine from my Mac. Bryan's
work-box launches still get 'rate_limit_error' from the same script.
Only meaningful differences in the working curl vs the failing one:
- Working: explicit User-Agent (claude-cli/2.1.85) + Accept: application/json
- Failing: defaults (curl/X.Y.Z, no Accept)
Anthropic's OAuth endpoint apparently checks User-Agent (or the Accept
header) and returns the misleading rate_limit_error for unrecognized
clients. Adding both headers to match what claude-cli and droidrun
send. Patched in cmd_login AND cmd_refresh.
Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
Root cause of every prior 'rate_limit_error' on OAuth login: Anthropic
migrated all the Claude-subscription OAuth endpoints from
console.anthropic.com / claude.ai to platform.claude.com / claude.com.
The old endpoints aren't 404 — they accept the POST and return a generic
'rate_limit_error' for every request, which is what mis-led both me and
several public community implementations.
Confirmed against two current working clients (droidrun/mobilerun and
motiful/cc-gateway, both using the same Claude Code public client_id):
AUTHORIZE_URL: claude.ai/oauth/authorize
→ claude.com/cai/oauth/authorize
TOKEN_URL: console.anthropic.com/v1/oauth/token
→ platform.claude.com/v1/oauth/token
REDIRECT_URI: console.anthropic.com/oauth/code/callback
→ platform.claude.com/oauth/code/callback
SCOPE: org:create_api_key user:profile user:inference
→ ...plus user:sessions:claude_code user:mcp_servers user:file_upload
Also updated the error-hint text to mention the misleading-rate-limit
pattern for both 'malformed code' AND 'dead endpoint' cases, and to cite
the current TOKEN_URL — so if/when these move again, the next person
hitting the same trap finds the answer in the script's own output.
The CODE#STATE parsing from 0.5.0 was correct and stays. State IS sent
in the token-exchange body (verified against droidrun's working flow).
Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
In 0.5.0 (and every prior version), prompt_first_run_auth was called
unconditionally at script load time, BEFORE self_update. On a never-
authenticated box, this meant a broken lib/oauth.sh trapped the user:
1. larry starts
2. no creds → auth prompt fires
3. pick OAuth → old broken oauth.sh runs → rate_limit_error
4. Ctrl-C at the API-key fallback prompt
5. script exits — self_update never ran
6. relaunch → exact same trap, forever
Fix: defer the auth-prompt call to run AFTER self_update. The auth
function DEFINITION stays where it is; only the CALL site moves. Now
on a fresh box:
1. larry starts
2. self_update phase A pulls MANIFEST and refreshes everything,
including a patched lib/oauth.sh
3. THEN the auth prompt fires, using the now-correct OAuth code
Verified: with no ANTHROPIC_API_KEY and no .oauth.json, the manifest
sync log lines appear before the "First-run authentication setup" menu.
Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
Self-update overhaul (no more manual reinstalls when lib/ changes):
- New MANIFEST file at repo root lists every file that should auto-sync
(top-level scripts, agents/, lib/, VERSION, MANUAL.md).
- larry.sh self_update() reworked into two phases:
Phase A — local sync: if $LARRY_HOME/.last-sync-version != $LARRY_VERSION,
fetch MANIFEST and refresh every listed file. Stamps version after.
Phase B — remote check: fetch $LARRY_BASE_URL/VERSION; if newer, pull
larry.sh, self-replace, relaunch with LARRY_JUST_UPDATED=1 so phase B
is skipped on the relaunch (phase A then pulls everything else).
- New LARRY_BASE_URL env var (the legacy LARRY_UPDATE_URL / LARRY_AGENTS_URL
still work as overrides).
- Bumped LARRY_VERSION and VERSION to 0.5.0.
OAuth fix (lib/oauth.sh):
- Anthropic's callback returns the code as 'CODE#STATE' (URL fragment, not
query). Previous prompt told users to copy "between code= and the next &"
which produced the wrong substring; the token endpoint then returned a
misleading 'rate_limit_error' on the malformed code.
- Now splits the pasted input on '#', verifies the returned state matches
the one we generated, sends only CODE to the token endpoint.
- Updated user-facing prompt and error hints to describe the real format
and explain the misleading rate_limit_error symptom.
Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
Each Larry is independent. Bryan's question "how will Larry on Windows
talk to Larry on Linux for regression file transfer" answered: they don't.
File transfer is YOUR responsibility (scp / gh release / shared mount /
USB), but nc-regression now produces and consumes portable bundles that
make the split a one-command-on-each-side workflow.
Changes:
lib/nc-regression.sh
+ --phase env-a convenience for phases 1+2+3 (env-A side)
+ --phase env-b convenience for phases 4+5+6 (env-B side + diff)
+ --bundle-out PATH after env-A phases, tar inputs+outputs/env-a +
manifest.json + README.md + inbounds.txt
+ --bundle-in PATH at start, untar a bundle into $OUT; pulls scope
from the manifest so the env-B side just needs
--env-b and --route-test-cmd
MANUAL.md
+ New "Cross-environment Larry — how the boxes communicate" section
+ Bundle transport table (scp, gh release, NFS, USB, etc.)
+ Notes that the lesson loop uses the same local-capture / manual-
transport / central-merge model
Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
Five small Unix-style loop & format helpers, fully offline:
lib/each.sh
- replaces v1 `each`
- run a CMD per item: args, stdin lines, or {}-placeholder substitution
- example: tbn adt | awk '{print $2}' | each.sh 'route_test {}'
lib/each-site.sh
- replaces v1 each_site / each_site_hdr / each_site_tcl patterns
- iterates every site under $HCIROOT with HCISITE/HCISITEDIR auto-exported
- --filter REGEX limits which sites; --hdr prints a header before each
lib/len2nl.sh
- replaces v1 `len2nl`
- strict superset: handles length-prefixed (digits before MSH),
MLLP (\x0b...\x1c\x0d), and segment CRs (→ LF)
- works as stdin filter or with file arg
lib/csv-to-table.sh
- 2-column CSV → Cloverleaf .tbl format
- emits proper prologue (who, date, bidir, type, version)
- --has-header --default VALUE --bidir 0|1 --in-delim CHAR --user NAME --out PATH
lib/table-to-csv.sh
- reverse: .tbl → CSV
- --with-header --delim CHAR --include-meta
- confirmed clean round-trip: CSV → table → CSV byte-identical for the data rows
All 5 are pipeable, have --help, zero external deps beyond bash+awk+sed.
Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
nc-parse.sh
+ chain <name> [--depth N] [--direction both|up|down]
BFS over sources+destinations from a starting thread; returns the
reachable cluster as TSV (depth, direction, thread).
nc-msgs.sh
+ Filter operator additions:
> >= < <= numeric or lexical (works for HL7 YYYYMMDDHHMMSS timestamps)
>< range "LO..HI" inclusive
+ Filter group additions:
--field AND group (must match; existing behavior)
--or-field OR group (at least one must match)
--not-field NOT group (none may match)
All three groups combine; bug fixed where empty AND group bypassed
OR/NOT checks in the count format.
+ SmatHistory walk:
--include-history also walks $HCISITEDIR/exec/processes/*/SmatHistory/
--all cheat-sheet alias for --include-history
Confirmed working against the real ancout test data:
- chain IB_ADT_muxS finds all 7 downstream destinations
- event=A08 OR event=A03 → 20 (19+1 of 22)
- visit>400000000 → 22 (all numeric in range)
- visit><400000000..400450000 → 22 (range inclusive)
- --include-history → 22 active + 34 history rows = 56 total
Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
Field path improvements (hl7-field.sh + every tool that uses it):
- Accept both `.` and `-` as separators:
PID.3 == PID-3
PV1.3.4 == PV1-3.4 == PV1-3-4 == PV1.3-4
- Field-name aliases (case-insensitive):
mrn → PID.3
account / account_number → PID.18
name / patient_name → PID.5
dob / birthdate → PID.7
ssn → PID.19
visit / encounter / csn → PV1.19
attending → PV1.7
event → MSH.9.2
control_id / msgid → MSH.10
...and ~40 more covering MSH/PID/PV1/EVN/NK1/GT1/IN1/OBR/OBX/DG1/ORC
- Aliases also accept component/subcomponent suffixes:
name.2 → PID.5.2
mrn.1 → PID.3.1
Filter operators (nc-msgs.sh --field):
PATH=VALUE exact equality
PATH!=VALUE not equal
PATH~VALUE contains (case-insensitive)
PATH!~VALUE does not contain (case-insensitive)
PATH=NULL /= null / empty / absent
PATH!=NULL present (any non-empty rep)
PATH=* wildcard — any non-empty value
Multiple --field flags AND; for OR, run two queries.
New output formats for nc-msgs.sh:
text (default) segments per line + metadata header per message
oneline one message per line, segments joined with a ⏎ marker
fields each non-empty field on its own line: "SEG.N: value"
mp alias for fields (matches v1 `mp` semantic)
labeled fields with friendly aliases: "MSH.9 (msg_type): ADT^A08"
raw, json, count — unchanged
MANUAL.md updated with the full operator + format reference.
Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
Bryan's ask: use Larry on prod data without PHI ever leaving the client box.
Added:
lib/hl7-sanitize.sh — tokenize PHI fields in HL7 messages
lib/hl7-desanitize.sh — reverse op (local view-time unmask)
Tokenization model:
- Replace PHI fields with [[CATEGORY_NNNN]] tokens (MRN, NAME, DOB,
ADDR, PHONE, ACCT, SSN, PROV, VISIT, etc.)
- Same value → same token across messages (deterministic via local
lookup table; analysis can still correlate patients).
- Lookup table at $LARRY_HOME/sanitize/lookup.tsv mode 0600 — never
leaves the client.
- Default PHI rule set covers PID, PV1, NK1, GT1, IN1, OBR, OBX,
DG1, ORC; --rules-file to extend.
- --strict also tokenizes unknown Z segments wholesale.
Prompt-side preprocessing in larry.sh:
- {{phi:VALUE}} inline marker, auto-category lookup
- {{phi:CATEGORY:VALUE}} explicit category
- Replaced with the token BEFORE the user input enters conversation
history. The original never reaches the API.
- Local feedback "phi> {{phi:...}} → [[TOKEN]]" printed to terminal only.
New REPL slash commands:
/phi <value> tokenize a single value, print the token
/unmask <token> show original (local terminal only, never API)
/tokens show full PHI ↔ token lookup table
New tools in larry.sh schema:
hl7_sanitize agent can sanitize a file before reading PHI
tokenize-value / detokenize-value (subcommands of hl7-sanitize.sh)
Persona update (agents/larry.md):
- Documented PHI mode and rules for proactive sanitize-first behavior
MANUAL.md updated with the full PHI section including limitations.
Brings total native tools to 29.
Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
Bryan's pivot: until bjnoela.com is back online, transfer learnings via
local file capture on the client + manual paste-back to home-Larry. NO
credentials required on the client box.
Capture flow:
- lib/lessons.sh records lessons to $LARRY_HOME/lessons/<date>.md
- lesson_record tool in larry.sh lets the agent record proactively
- /lesson, /lessons, /export REPL commands
- agents/larry.md updated: capture corrections, conventions, quirks
silently when Bryan teaches them
Export flow:
- lessons.sh export | bundle | --gh-issue (uses gh CLI if available)
- Bryan pastes the bundle to home-Larry on his dev machine
- home-Larry commits the refinement into cloverleaf-larry/agents/
- next launch on any client pulls updated persona via self-update
Brings total native tools to 28.
Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
