Skip to content
Success

#73 (Oct 24, 2026, 5:03:09 PM)

Started 25 days ago
Took 2 min 5 sec
Build Artifacts
ROOT.war331.11 MiB view

Started by an SCM change

Revision: 4a607b1d9fc2945fb492f35eec0cb735c86ca1fa
Repository: https://thoughtframe.ai/thoughtframe/repos/thoughtframe-core
  • refs/remotes/origin/master
Initial cleaned up commit
ian at
This will serve as the bases for the plugin loader
ian at
change installer for new thoughtframe project
ian at
Fix the upgrade script to include the jar
ian at
Remove Entermedia depenancies
ian at
cleanup groovy imports
ian at
figure this out later - need to make it all queued.  Add in missing
system components
ian at
fix a bunch of fallbacks
ian at
fix authentication
ian at
Cleanup presents
ian at
fix pdf viewer, fix some other stuff
ian at
Start to cleanup stuff
ian at
cleanup and solidify linkages
ian at
fix some includes for local, and deploy a logo
ian at
fix the folder saving again for now
ian at
correct fallback
ian at
get rid of layhout stuff
ian at
new code for ThoughtFrame - Chain - bitcoin support etc
ian at
prepping TRP
ian at
buid
ian at
need to add manual builds!
ian at
fix permissions
ian at
fixes to package roots
ian at
re-scope some stuff
ian at
speficic leaf categories
ian at
eclispe truoe
ian at
Fixes issues with LLM chat
ian at
cleanup tests page
ian at
permission
ian at
NEED WEB.XML!!
ian at
alert code
ian at
AI State Machine work
ian at
Bringing Frame concept to life !
ian at
First frame
ian at
True elegance- Frame State Machines
ian at
missing symlink
ian at
Gradle compile support - lets see!
ian at
see if this builds
ian at
remove old groovy
ian at
ThoughtFrame DSL
ian at
Thoughtframe  - CORE FMS
ian at
Frame State Machine changes
ian at
check all this out.
ian at
setup script, make sure we build
ian at
Fix the build
ian at
Success!!!!!
ian at
finalizing Frame State Engine
ian at
Frame builder
ian at
current ThoughtFrame code
ian at
Crazy awesome state machine code
ian at
Async code added -
ian at
my brain is hurting - but the event sysytem for chat now connects to the
FSM system
ian at
attach chat to frames and FSM concept
ian at
Fix up the event systems - more more old chaff
ian at
kick off a build
ian at
start removing depracations
ian at
big fixes
ian at
fix fix fix
ian at
update builds to create our WAR fully and build EM directly
ian at
fucking smylinks
ian at
build, don't include a ton of extra jars
ian at
cleanupm a ton of old jars
ian at
build
ian at
fix jar
ian at
fix gradle builds
ian at
EM files
ian at
remove redundant ANT stuff, commit default upgrader
ian at
make fallback work, rename some FSMs
ian at
gather state and change to target TF branch
ian at
remove copy config files
ian at
Handle async in it's own section -
ian at
fix up some classpath issues
ian at
go back from exception
ian at
use MY _site.xconf
ian at
remove crappy _ in groovy
ian at
More updates - much simpler wiring - Ui helper FSM etc
ian at
unneeded code
ian at
Newest layout - cytoscape
ian at
latest TF for Blockchain
ian at
build
ian at
ADded multi-channel support
ian at
bitcoin updates plus plus
ian at
Add CRON support for frame scheduling
ian at
Move to a MESH instead of a tree!  Huge.
ian at
move to a frame mesh based approach
ian at
NESTED for the win baby
ian at
Thoughtframe is now a mesh!
ian at
Mock up a frame for services
ian at
tweak to allow state to be set
ian at
Whoooot programatic logo
ian at
Logo!
ian at
test build
ian at
T4EST
ian at
revert t known working copy
ian at
change to options
ian at
update jdk stuff
ian at
try to fix and deploy tewst
ian at
tweaks
ian at
test a different upgrade system
ian at
don't delete myself!
ian at
upgrade attmpt
ian at
try another way to restart
ian at
tomcat PID attempt
ian at
restart tomcat
ian at
restart tests
ian at
try agian
ian at
newer code
ian at
fixed up?
ian at
one more try
ian at
ordering is important
ian at
argh
ian at
remove duplicates
ian at
AMAZING changes to enable more agency - the frames specificallym handle
LLM functions now
ian at
add more details
ian at
changed and stuff
ian at
fixes and changes to git example and blockchain example
ian at
fix the frames, line everything up
ian at
check if idle detects push and make sure we capture build success
ian at
fix the god damn ordering bug
ian at
bit of logging
ian at
Nice new code for frame events
ian at
Slick new logging
ian at
Newest code
ian at
test new logging system, make sure they are tracking
ian at
add log as visible
ian at
updates to frame descriptions
ian at
Big fixes to edge creation and persistence, clean up some bugs
ian at
new frame code for creation and spawning by hand
ian at
Clean up the layout, graph paper style background
ian at
Simplify exception handling
ian at
add more tranistions info and async
ian at
removed immediate as an option,
ian at
mining code example and exceptions fixed
ian at
Move everything and be more explicit
ian at
newer code!
ian at
building quizzes from scratch, multi-step
ian at
Some extra repo stuff
ian at
use DI
ian at
missed
ian at
add the class
ian at
log it
ian at
attempt tog et it working in 2.4
ian at
hooooop jumping for ancient groovy
ian at
add exception
ian at
fixes
ian at
new code
ian at
fix the props
ian at
try to fix the new details loading
ian at
add much simpler/wiser event handling
ian at
spawn builds
ian at
cleanup stuff related to TRP and don't treat / as special routes anymore
ian at
update some UIs
ian at
next round of code.
ian at
new code
ian at
new bootstrap
ian at
UI cleanup - fairly extensive
ian at
fix the asset picker stuff finally
ian at
updates to UI and frame - reverse logic on approval flow
ian at
add keys
ian at
fix typo
ian at
Fix the classpath up to a standard
ian at
various changes to enfore FSM rules
ian at
fix the images for aemca
ian at
get a build
ian at
fully declarative.  NEed to test capture still
ian at
Some random changes
ian at
Validate eth scan is working
ian at
new mode
ian at
New code - hugely improved logging
ian at
newest code for testing
ian at
fix 'er up
ian at
eth scanned fixed!
ian at
fix the HTML for the notificaiton alos
ian at
send the type
ian at
frame code?
ian at
all code to date - adding json support camel support
ian at
start cleanup - pre-move TF to it's own area outside system
ian at
Bring up to date.
ian at
Add regisrtation code
ian at
newest code examples
ian at
try this with camel and everything
ian at
prevent NPE
ian at
try to exclude jars
ian at
add better logging
ian at
more logs
ian at
API?
ian at
stupid bad build
ian at
Gmail code for camel - more logs
ian at
check tokens
ian at
TF code
ian at
new code
ian at
newer code for specific tokens
ian at
newest updates to TDF
ian at
frame updates
ian at
Fix gradle build
ian at
varios events and changes
ian at
fix the question creator
ian at
fix response editor
ian at
reworked the chat extensively, added multi-function simultaneos support
ian at
update llama server code
ian at
simple survey page
ian at
got this chatting again and added a real agent attached to a test -
which is very interesting.
ian at
new helper
ian at
fix up quiz and a couple other ui bugs.
ian at
fix some bugs with frame memories and everying
ian at
TF law
ian at
roadmaps
ian at
lab and law
ian at
make public
ian at
Channel clear - UI cleanup, etc
ian at
fix commands, add logs
ian at
clean up commands so they make more sense.
ian at
add clear to quiz helper
ian at
redirect to a custom folder
ian at
add reporting
ian at
adding reports
ian at
cheesy reporting stuff
ian at
new code
ian at
Stupid NPE
ian at
some reporting and fix a generate fake data bug
ian at
Local pod only
ian at
Channels per test
ian at
menu link
ian at
change report and aggs
ian at
Fix a ton of stuff around frame sessions, get everythign working great
ian at
new code
ian at
TF is working
ian at
nailed the query
ian at
new code.
ian at
PII deployment options
ian at
new code
ian at
Get DSP streamign from source FLAC files
ian at
Sonar related code.
ian at
Wonder if I can deploy this lol
ian at
get CNN sort of working
ian at
Fix up the code
ian at
sensortech and mock up the slide creator system
ian at
add a bit more for slides.
ian at
slides
ian at
remove symlinks for now
ian at
slide commit
ian at
npuio
ian at
simple empty page
Osprey at
minimal install for sky
ian at
ROund one
ian at
mmoresadfsaf
ian at
banner gradient
Osprey at
Fix checkouts
ian at
Merge branch 'master' of
https://thoughtframe.ai/thoughtframe/classes/AZm_IAz5HKdm91A3zPpO/code/repo/thoughtframe-core/thoughtframe%20core
Osprey at
updates
Osprey at
more updates to getting things started
Osprey at
almost done!
Osprey at
going live!
Osprey at
edits
Osprey at
remove mail update
ian at
Maybe fix email
ian at
try again
ian at
fix?
ian at
debug
ian at
find the rest
ian at
use jakarta angus
ian at
Change profile URL
ian at
latest for contact form
Osprey at
faq
Osprey at
New code for Sky Emails
ian at
try again
ian at
Merge branch 'master' of
https://thoughtframe.ai/thoughtframe/classes/AZm_IAz5HKdm91A3zPpO/code/repo/thoughtframe-core/thoughtframe%20core
Osprey at
tweak.  todo: use replacer.
ian at
extra keys
ian at
updates to contact form & related such
Osprey at
Clean up
ian at
logging for recaptha
ian at
actuallyt set the value
ian at
changes to contact form
ian at
new code
ian at
Merge branch 'master' of
https://thoughtframe.ai/thoughtframe/classes/AZm_IAz5HKdm91A3zPpO/code/repo/thoughtframe-core/thoughtframe%20core
Osprey at
whale image sizing fix
Osprey at
Merge branch 'master' of
https://thoughtframe.ai/thoughtframe/classes/AZm_IAz5HKdm91A3zPpO/code/repo/thoughtframe-core/thoughtframe%20core
Osprey at
add
ian at
Merge branch 'master' of
https://thoughtframe.ai/thoughtframe/classes/AZhhg2hJgniNmOZnUs4e/code/repo/thoughtframe-core/thoughtframe-core
ian at
Starting to rebuild
ian at
Convert ThoughtFrame to a Gradle monorepo

- Create root multi-project Gradle build
- Add shared Gradle wrapper
- Convert engine to subproject
- Convert DAM to project dependency
- Convert core to project dependency
- Remove nested Gradle wrappers and settings
- Reorganize repository into modules
- Preserve engine/core architectural boundaries
ian at
Clean up Gradle artifacts and repository ignores
ian at
Docker protject
ian at
Revamp everytthing!
ian at
remove course.  Start prep for all course materials existing in core
repos
ian at
Got repositories working properly.
ian at
asdf
ian at
AMAZING code updates.  CLI works
ian at
add shell, more CLI tools
ian at
Start adding in the code for room
ian at
Update to opensearch
imiller at
WIP checkpoint: TUI agent tools + Eclipse IDE bridge + FSM event notes

Pre-refactor baseline before moving the agent turn into the FSM. Includes the tool surface (open/peek/glob/grep/write/highlight/replaceSelection/console/problems/build/runTests/run), content-addressed edits, per-file problems, the Eclipse plugin (bridge/FrameViewer/EngineIsland) and eclipse-build staging, and the TUI engine (TuiModule/IdeModule/DeepSeek) + frames/tui. Builds green (engine assemble + plugin javac).
ian at
FSM event delivery: trace logging, pure transition, self-sourced observations

- TestBench.fireFrameEvent + FrameEventHandler.eventFired log origin/source/fsm/asyncframe and which path (A tick/async, B, interested-frames count)
- FrameDsl.transition is a pure state swap; it no longer fires frame:transition (which re-entered the held frame lock and was usually dropped)
- IdeModule fires observations with no request, so the event has no fsm -> FrameEventHandler path B (handleAsync) -> the frame hears its own thoughts. Previously the action request's fsm sent ide:* down path A (handleEvent), skipping async handlers and preempting the listener scan.
ian at
Move the agent turn into the frame's thinking state

- agentStep: one model step per tick; all turn state (prompt/model/step/nudge/fired/ran/lasttext) lives on the frame, so several agents (frames) can be mid-turn
- session.groovy thinking state runs actions/agentstep then transitions to frame.nextstate
- converse sets the turn on the frame, fires thought:ask, and drives the existing tick loop (tickUntilDone) to done; no Java while-loop of its own
- tickUntilDone does not hold the frame lock, because a tool in the state body fires a self-sourced observation that needs the lock (path B)
ian at
Agent turn transitions on outcome events, not a computed nextstate

- agentStep fires thought:agentturn (loop) or thought:agentdone (finish); the session frame's async handlers own the transition
- the thinking state body is now just action(actions/agentstep); it does not read frame.nextstate or decide its own next state
ian at
Turn holds the frame lock: in-memory frame/fsm, inline events

- tickUntilDone acquires the frame lock once, loads the frame and machine once, and ticks to a terminal state in place (state shared, not reconstituted per tick). A running turn is exclusive: the mesh's lock attempt skips it.
- Because the turn holds the lock, events fired within it are inline: IdeModule observations and agentStep's thought:agentturn/agentdone fire with the request (carrying the ticking fsm) and asyncframe=true, so FrameEventHandler path A calls handleAsync. Path B would be refused.
ian at
Nudge on announce-without-acting; prompt: don't promise, act

- agentStep treats 'I'll do X' with no tool call like a blank reply: it nudges once or twice (agentturn) with a corrective prompt before accepting text as the answer
- framemessage.html: do not announce an action and stop; treat your own earlier claims as unverified unless a tool result backs them (self-imitation was reinforcing fabricated edits)
ian at
Fix conversation order: record the user turn before the machine runs

- converse recorded the user 'chat' row AFTER the turn, so by dateUp the tool calls and responses preceded the user message that caused them - the history read assistant-tool-calls -> user, inverting the conversation and twisting the model about what had been done
- now the user row is recorded first; agentStep no longer re-sends the prompt as $prompt (it is in $recent)
- deepseek.html appends the final user message only when $prompt is set, which is now just the ephemeral nudge (nudgemessage) for a single step
ian at
Eclipse agent surface: engine web stack in the plugin + tf-eclipse/tf-engine verbs

The plugin hosts the full ThoughtFrame webapp on a loopback port, so the same
engine is reachable over HTTP by the UI, the TUI, and any external agent.

- EmbeddedWebStack boots the webapp on Jetty in-process; DefaultEngineRuntime
  gains a WebServer-unified core with two boot modes (fixture or web stack) and
  writes ~/.thoughtframe/instances/<hash>.json for discovery (jdt-style).
- EngineConfig.withPort; EngineIsland passes a port from thoughtframe.eclipse.port
  (defaults to 0 = ephemeral, no launch args needed).
- tf-eclipse: 51 verbs over the live IDE - read (projects/editors/selection/
  stat/markers/tasks/bookmarks/classpath/console/tests/git/debug/status...),
  edit (open/close/write/save/delete/mkdir/moveFile/copy/move/rename/format...),
  build/run/test, and a debugger (breakpoints incl. exception, eval, recursive
  variables, runToLine, step filters, dropToFrame, debugRecipe, attach).
  Self-documented at .../tf-eclipse/help.html; errors surface as $error, not 500s.
- tf-engine: status/frame/tick over the FSM.
- IdeModule: read arguments by request parameter first, so OpenEdit's reserved
  "path" page value can no longer shadow a caller's file; replaceSelection
  reports errors instead of throwing a 500.
ian at
tf-eclipse git: push/fetch/pull via EGit credentials (+ creds check, help)

Adds org.eclipse.egit.core and push/fetch/pull ops to the git verb. Credentials
are read from EGit's secure store (the ones Eclipse uses) and wrapped in the
engine's own JGit provider, since the plugin and the island load different JGit
classes. Read verbs keep working when egit.core is absent (lazy resolution).
ian at
Spike test
ian at
tf-eclipse: semantic search verbs via Eclipse's JDT index

find / references / implementors / hierarchy / type-info / source, backed by the
JDT SearchEngine (the precise, always-fresh Java index) rather than a code
embedding index. Hot Groovy pages over the live workspace.
ian at
Untrack module compiler output (bin/) and the throwaway spike

The Eclipse JDT output dirs (core/bin, dam/bin, shell/bin) were tracked,
so every build dirtied the tree (1361 files, mostly .class). Removed them
from the index and added explicit .gitignore rules. webapp/WEB-INF/bin is
left tracked - that is real content, not build output.
ian at
Ignore module bin/ dirs; prefer source types in search findType

collect the .gitignore rules for core/dam/.../bin, and make findType pick a
source type over a binary duplicate so implementors/search resolve to the
project sources.
ian at
tf-eclipse: split the plugin webapp from core; organize verbs by category; Groovy-first verbs

- Plugin gets its own standalone webapp (own node.xml, base/tf-eclipse) + engine-side
  module (eclipse-build) with plugin.xml/depends; no Eclipse bits in core.
- Organize the ~56 verb pages into search/ read/ edit/ run/ debug/ agent/ engine/;
  the tf client maps verb names to folders.
- Port the IdeModule-backed verbs to Groovy Script.run pages (delete IdeModule.java,
  trim EclipseWorkspaceBridge to captureContext); webapp Groovy is live, no restart.
- rename.groovy: pick the LTK refactoring by element kind (type/method/field/local
  variable) — previously always RENAME_TYPE.
- Add WebSearchModule (WebSearcher interface + TavilyWebSearcher: search + OkHttp fetch)
  with tf + base-frame actions; tf-eclipse/errors reads the Eclipse error log.
- write.groovy reveals the written file; AGENTS.md: prefer tf over shell.
ian at
Untrack runtime OpenSearch data; gitignore engine data/scratch dirs

git add -A in the previous commit swept in the plugin's regenerated OpenSearch node
data (eclipse/webapp/WEB-INF/elastic), which must never be tracked.
ian at
tf git push: support force=true (for history rewrites)
ian at
tf-eclipse: prune unused frame sets from the plugin webapp (keep base + tui)
ian at
tf-eclipse: agent memory via the thought searcher (topics + tags/status taxonomy)
ian at
tf-eclipse: engine/frames - connect + traverse the frame graph; document search/ASK rules
ian at
docs: make checking the graph (working memory) a habit, not an afterthought
ian at
tf-eclipse: package for install - dropins zip, p2 update site (feature + publisher), lander page
ian at
rename the OpenSearch data folder elastic -> opensearch (node.xml, lock guard, ignore)
ian at
tf-eclipse lander: drop html/head/body - the engine supplies the document wrapper
ian at
tf-eclipse packaging: stamp a build version (VERSION= or <base>.<UTC timestamp>) so publishes are upgrades
ian at
tf-eclipse webapp: fix the chat loop - add the velocity macro library (#jesc), tui actions/exception, and a chat verb to drive a turn
ian at
llmusage: per-call DeepSeek prompt/cache-hit/miss tokens in a table + a usage verb (aggregatable)
ian at
tf-eclipse: point the ask at the plugin's own base (EngineConfig.askPage) + jsonUtil on the chat verb; fix the pane's NPE and the empty-message renders
ian at
chat goes through the velocity ask page (cli subtree has jsonUtil+testbenchid); drop the agent/chat verb
ian at
tf-eclipse view: don't NPE when a turn is submitted before the engine is ready
ian at
chat tools: expose the JDT search verbs (find/references/implementors/hierarchy/type-info/source) + rename/move as tools (22 total)
ian at
chat: add a bash tool (runs shell on the host, so the agent can drive the tf client and reach the full verb surface)
ian at
chat: retain full tool results (resultfull) + a 'more' tool (grep/window by id), and a paged 'search' tool over any searcher (OpenSearch pagination)
ian at
p2 site: ship the bundle as a jar (Eclipse-BundleShape: dir drives unpack on install) - the folder artifact broke the .jar mapping
ian at
p2 site: add p2.index (pin content.jar/artifacts.jar factories so Collect finds the artifact repo) + a category
ian at
tf-eclipse: add a Restart Engine command (bounce the engine island -> fresh classloader, no Eclipse restart)
ian at
engine/restart: bounce the engine island via the host bridge (WorkspaceBridge.restartEngine) + a tf engine restart verb
ian at
engine/marker: a build marker verb, so a tf engine restart can be confirmed to have loaded a fresh jar
ian at
agent loop: let a turn run to the tool-step budget, then answer

- TuiModule: raise max tool steps 5 -> 12 (a real edit loop needs more) and, when the
  budget is spent mid-turn, take one final tool-less step so the model summarises in
  words instead of the harness emitting turnFallback's bare 'Done: <tool names>'.
  Bug found while driving the agent: the first cut of this dropped the continue branch,
  so any tool call ended the turn after one step.
- framemessage.html: the prompt listed only the original 14 tools; add the JDT search
  family (find/references/implementors/hierarchy/type-info/source), the refactorings
  (rename/move), bash+more+search, and the fact that bash reaches the full tf verb
  surface (run 'tf help'). The agent now discovers verbs instead of guessing.
- transcript: a tf verb to dump a session frame's thought rows (the conversation the
  model sees) - the tool that made these bugs visible.
ian at
StringHitTracker: page a stored record by lines, the same idiom as search

A HitTracker whose hits are the lines of a String. Pass in the text, hitsPerPage is lines
per page, page N is the Nth chunk - so a long tool response or any stored record can be
read a page at a time with 'page N of M' / next / prev, exactly like search.

No new persistence and no session id: the record already exists as a Data row (a tool
response is thoughts.function_response -> resultfull), so its id is the address and the
tracker is rebuilt from the text on each call. 'more' now windows through it instead of
slicing a string by offset.
ian at
StringHitTracker: implement HitTracker directly (hits are Strings)

Drop the ListHitTracker base - its get(int) casts hits to Data, which a String tracker
does not want. HitTracker already gives concrete Collection bodies, so only size/iterator/
contains/get plus an overridden getPageOfHits are needed.
ian at
StringHitTracker: fixed-size sections, each hit a StringSegmentData

A hit is one section of the document (chunkSize chars), not a line. Each hit is a
StringSegmentData - a Data that knows its own text, start and length - so get(n) is the
list lookup and a hit maps back onto the document.

more is now dumb: resolve the record's text, hand it to the tracker, expose the page, and
let more.html render it (hits / page / total pages come off the tracker; no StringBuilder,
no paging math in the verb).
ian at
TuiModule: compaction floor in recentThoughts (the summary is the thread's first thought)

If the frame has a summarydate, recentThoughts() floors at it: thought query for the
framepath, .after(date, summarydate), sort dateUp. after() is inclusive (date >= floor), so
the summary is returned as the first thought and older thoughts simply stop being sent -
never deleted, still searchable. Inert until a summary sets the date; no schema change
(an existing date carries it). The date must come from the summary thought's own value:
the compare is on millis, so a rounded/reformatted date can push the first message out.
ian at
TuiModule: a repeated identical tool call returns the earlier result, not a meta-message

The dedupe answered a repeat with 'already done this turn with those arguments'. That is
not a tool result, and the model only treats a real result as evidence, so it retried the
same call step after step until the budget ran out, then answered from memory. Instead
hand back the earlier result, cached per turn in a frame value (keyed by call signature) -
turn state has to live on the frame because the module is not the same instance from one
tick to the next. Cleared at turn start and after any mutating tool, so later reads still
see the new revision.
ian at
grep/glob: don't double-list files under a Gradle root project

Buildship imports the Gradle root project and each subproject, and the subprojects live
inside the root, so walking the root re-walked every subproject and each file was found
twice (/core/... and /thoughtframe-core/core/...). The agent read that as two copies of the
source. Skip any directory that is itself an accessible project root - it is walked in its
own right, and the root still contributes its own files (build.gradle, AGENTS.md, ...).
ian at
search verbs: accept a simple class name, not just an FQN

find returns a fully-qualified name and matches simple names/patterns, but type-info,
source, references, hierarchy and implementors all resolved via project.findType(fqn),
which needs an FQN - so the natural 'find then type-info StringHitTracker' pair failed on
the second call and cost the agent a step. Fall back to the JDT index for a bare simple
name, the same way find resolves it.
ian at
edit verbs: accept a JSON request body (prime _jsonRequest)

Calling context.getJsonRequest() first parses a POSTed JSON body into _jsonRequest, and
BaseWebPageRequest.getRequestParameter then falls back to it - so every existing parameter
read works unchanged, and a client can send multi-line text as JSON with no URL encoding or
shell quoting. This is the ergonomics fix: the verbs were already content-addressed, but
quoting multi-line text through a k=v HTTP call was the reason edits drifted to raw FS
writes. Client side: 'tf --json <verb> <file.json>' POSTs the file (client is the local
install at ~/.thoughtframe/tf, not in this repo).
ian at
TuiModule: remove the tool-step cap

The cap (and its tool-less wrap-up step) is gone: a turn now keeps stepping until the model
answers, with FrameManager.tickUntilDone's tick budget as the only outer bound. The budget
was engine policy in a Java field, which made tuning it a rebuild + bounce; where the limit
should live instead is still open (parked).
ian at
framemessage: two unblocks for the agent - path spaces, and read economy

- Paths: state that workspace paths (/project/...) are what peek/grep/glob/replaceSelection
  take, while bash sees the filesystem (the project's location from 'tf project'). The agent
  burned a whole turn ls/sed-ing a /project path in bash and then scanning '/' with find.
- Be economical with reads: reading is not progress; peek a window, not a whole file. The
  META prompt's 'more calls, not fewer' had it spend every step reading and never editing.
ian at
tf verbs: prime the JSON body in every verb, not just the edit two

context.getJsonRequest() at the top of each verb page means getRequestParameter falls back
to a POSTed JSON body everywhere, so any verb can be driven with 'tf --json <verb> file.json'
and multi-line values need no URL encoding or shell quoting. Applied across all 72 verbs.
ian at
TUI turn: prune orphaned tool rows, and add manual /compact

- pruneOrphanToolRows: the list sent to the model is now always API-valid. An interrupted
  turn could leave an assistant tool_calls row with no function_response, and every later
  request then failed with DeepSeek 400 'must be followed by tool messages responding to
  each tool_call_id' until /clear. Pair function_call.id against function_response.command
  and drop unpaired rows both ways, date order preserved. Covered by PruneOrphanToolRowsTest.
- /compact: summarise the conversation the model currently sees, write it as a
  messagetype=summary thought, and move the frame's summarydate floor to it. recentThoughts
  floors there, so the next turn starts at the summary; older rows stay searchable. Mirrors
  /clear's shape in runCommand.
ian at
frames: declare summarydate, and add tf reindex

recentThoughts floors at the frame's summarydate, but the value round-tripped through the
frame's _state overlay as Date.toString() and getDate() could not parse it - so the floor
never applied and /compact changed nothing. Declaring summarydate datatype=date makes the
searcher store it in storage format; tf reindex re-saves a searcher's rows so the new
mapping takes effect on rows already written.
ian at
agent tools: one real tool per verb, schema + router, generated

Every tf verb is now a typed tool call instead of something the model must reach through
bash. genTools (itself a tf verb) walks the verb pages and writes both halves - the schema
(frames/tui/functions/<verb>.json: description from the verb's own header comment, params
from its getRequestParameter reads) and the router (frames/tui/actions/<verb>.xconf) that
runTool executes - then asserts every tool named in llm-functions has both, so a schema can
never be advertised without a router. framemessage.html drops its hand-written tool
catalogue (the schemas describe every tool now), halving the prompt.

Fixed cost rises ~5.7k -> ~8.9k tokens but stays ~97% prefix-cache hit, and the model stops
emitting shell command lines for verbs that are now typed calls.
ian at
chore: drop the stray ${sys:es.logs} log files and ignore them

The OpenSearch node logging config leaves ${sys:es.logs} unresolved, so logs are written to
files literally named that - and nine had been committed under dam/ and thoughtframe/, so
git status was never clean. Remove them, ignore the pattern, and ignore eclipse-build/bin.
ian at
git: fix BaseGitRepository.getVersions, and add a fixture-free unit test

getVersions never returned what it claimed - three defects in one loop:
- the RevWalk was never markStart'd, so the for-each body never ran and the method always
   returned an empty list;
- the TreeWalk was built with forPath(repository, inPath) - no tree - so it resolved
   against HEAD: had the loop run, every version would have carried the HEAD blob (metadata
   per commit, content identical), making the version list useless;
- forPath returns a walk already positioned on the entry, so a following next() advanced
   off it and the body never ran even once the walk was correct.

Now: markStart(HEAD) with a null guard, forPath(repository, inPath, commit.getTree()), no
next(), and skip a commit whose blobId equals the previous one's - 'one entry per commit
that touched the path', not one per commit that merely still contains it.

Adds BaseGitRepositoryTest - 19 tests against a real JGit repo in a temp dir, no engine
fixture (no BaseEnterMediaTest/TestBench), so the suite runs in ~2s. Built by driving the
embedded agent; each failure pinned a real defect, and its initial review claim (that the
path was being used as a revision) was wrong - its own assertion disproved it.
ian at
docker: port the Damsel deploy model (thoughtframe-os + runtime, compose, scripts)

- docker/thoughtframe-os: JDK 25 base (temurin 25) + apt tooling + python libs
- docker/runtime: Jetty 12.0.35 + baked ROOT.war + overlay entrypoint,
  compose.yml/network/ports overlays, site.sh.example, scripts/
  (publish, publish-os, install-local, install-testsite, upgrade, start/stop/...)
- docker/build.gradle: dockerOsBuild/Push, stageRuntime, dockerBuild/Push
  (numbered + :latest), dockerUp/Down; registry-relative OS base
- thoughtframe/build.gradle: createWar -> fixed ROOT.war name + include
  WEB-INF/base/** (was only base/thoughtframe/**, so base/system was missing
  and the app failed with 'Could not resolve system'); setupLinks emfrontend2
  and drop dead emshare link
ian at
cms/openedit: require administrators for all editor view permissions

The editor subtree downgraded view from administrators (openedit/_site.xconf)
to <user/> (components/html + toolbar), so any authenticated user could reach
the page editor / filemanager. Restore administrators on all view blocks
(authentication/_site.xconf stays public for login).
ian at
thoughtframe: untrack the setupLinks-generated symlinks

base/{emfrontend,emshare,entermedia,manager,mediadb,openedit,system} were committed
as symlinks with absolute /home/ian/... targets, so a checkout on any other host
(e.g. Jenkins) had a dead emfrontend link and createWar failed with 'Couldn't follow
symbolic link'. Untrack + gitignore them (setupLinks recreates them per host), and
have setupLinks delete stale emfrontend/emshare strays.
ian at
dam: keep base/entermedia (empty scaffold) in git

/WEB-INF/base/entermedia is a fallbackdirectory target but the dir was empty, so
git never tracked it and the setupLinks symlink dangled on any non-laptop checkout
(createWar: 'Couldn't follow symbolic link .../base/entermedia'). Add a .gitkeep so
the directory exists off-host.
ian at
thoughtframe: createWar copies platform bases directly (no symlink dependency)

Mirror the :damsel createWar: copy WEB-INF/base/{entermedia,emfrontend2,system,
manager,mediadb} from :dam and WEB-INF/base/openedit from :cms directly, instead of
following the setupLinks symlinks (host-specific absolute targets that dangle on a
fresh checkout / CI). Keeps setupLinks for the local run. WAR now assembles on any host.
ian at
thoughtframe: createWar ships product code in WEB-INF/lib only (classpath = files())

The War task was routing the classpath dirs into WEB-INF/classes (579 files) where they
shadow the jar and are never refreshed by the runtime's overlay upgrade. Mirror the
:damsel createWar: classpath = files(), so only WEB-INF/lib carries product code.
ian at
build: JDK 25 toolchain + foojay auto-provisioning

:thoughtframe now declares a 25 toolchain and settings.gradle enables the foojay
resolver, so Gradle provisions JDK 25 itself instead of depending on the host JDK
being 25. Mirrors :damsel. Build is JDK-agnostic on any agent.
ian at
runtime: ThoughtFrame-clean entrypoint

- drop the SensorTech quick-upgrade zip block (ThoughtFrame never writes that zip)
- drop the forced jdk.incubator.vector module (SensorTech DSP); keep
  --enable-native-access=ALL-UNNAMED (silences real JNA/Jetty warnings, JDK-25 future-proof)
- supervisor now traps TERM/INT and forwards to Jetty (clean 'docker stop') and exits on a
  clean Jetty exit instead of always looping
ian at
build: drop foojay-resolver (auto-provision) — every build host has JDK 25

foojay-resolver-convention 0.8.0 is incompatible with Gradle 9.7.1
(NoSuchField JvmVendorSpec.IBM_SEMERU) and only fires when Gradle has to download a
toolchain. With JDK 25 present on laptop/Jenkins/the box, the :thoughtframe 25 toolchain
resolves locally; keep that, drop the resolver.
ian at
cms/openedit: relax toolbar view permissions (embedded in the core app)

toolbar/admintoolbarselector is included in ordinary app pages for every signed-in user,
so requiring administrators made those includes redirect (broken pages / redirect loops).
The toolbar is read-only, and admintoolbarselector already self-gates on
canView(/openedit/_site.xconf). Keep the write paths (components/html editor/filemanager)
administrators-only.
ian at
livetest: restore LiveTestModule + LiveTest (deleted in cb777b278)

Brings back the live-voting controller (loadLiveTest/loadTestStatus/updateActiveNode/
validateVote/recordResponse/triggerPageReload/muteVideoRooms) and the LiveTest plan
wrapper, so the plugin.xml bean and the livetest/livepage .xconf actions resolve again.

LiveTest now extends BaseData: SearchHitData no longer has a no-arg constructor (it
requires a Searcher), and LiveTest is just a TestPlan holder.
ian at
ship a default admin user + administrators group

WEB-INF/data is gitignored, so the workspace admin.xml/groups never reached a fresh
checkout and new instances booted with no login. Force-track the two defaults; createWar
already packages WEB-INF/data/system/** into the WAR.
ian at
livetest: guard availability/tally/recording and parse votes safely

- getTotalAvailableVotes/getTotalsForResponse: exact id matches, null-check the question/
  profile/aggregation, parse proxy counts tolerantly ($, commas, '-', blank).
- recordResponse: guard a missing live test/question, use exact deletes, compute available
  votes once, skip invalid or over-count split values, fix the responsetext NPE.
- validateVote: null guards and tolerant parsing.
ian at
importbmusers: persist the per-voteclass uservoteclass rows

The BMVA-style import computed each user's voteclass proxycount but only saved the
'general' row, so classed allocations were always 0. Add the voteclass row to the save
batch, guard failed user creation, and parse VOTES tolerantly.
ian at
runtime: drop the sensortech datadisk /data mount

ThoughtFrame has no bulk /data volume (that was SensorTech Edge Vault). Remove the mount
from compose and DATADISK_HOST from site.sh.example/install scripts.
ian at
testbench: restore live/AGM activity types and plan layouts
ian at
testbench: restore document review type and residential vote class
ian at
testbench: add proxyvote and AGM activity types
ian at
Revert "testbench: add proxyvote and AGM activity types"

This reverts commit ad52161181f791c3462b1005116b438bbef760fc.
ian at
testbench: drop dead course/lemon beans and guard missing videoroom
ian at
testbench: wire websocket live slide updates and jtmx ajax
ian at
benchconnection: compare planid as strings for slidechange
ian at
testbench: submit live responses via jtmx ajax
ian at
livepage: clean placeholder from AGM sidebar layout
ian at
livepage: restore Show Live Status offcanvas trigger
ian at
auth: send login codes through Postmark and simplify code email
ian at
mail: migrate the stack to Jakarta Mail/Angus
ian at
auth: use the code email template and OneLiveWeb sender
ian at
auth: keep PasswordHelper unchanged; use the page emaillayout
ian at
auth: default app mail sender to OneLiveWeb
ian at
auth: revert global sender to ThoughtFrame default
ian at
auth: brand the login code email like OneLiveWeb
ian at
AGM Status: full-page meeting dashboard (vote counts, participants, responses, login links)

New class/agmstatus page renders the important live-meeting info as tabs
on one screen: vote counts, participant/quorum status, per-participant
response details and login links. Linked from the class menu and from the
meeting room sidebar (Full Page).
ian at
AGM Status link: show for instructors too, outside the class-admin block
ian at
AGM status: rebuild as a self-contained admin dashboard

Drop the $pages.include() assembly of the meeting fragments and render the
whole tool here: KPI strip (quorum, roster, proxies, on-screen node), vote
counts with shares, participant status, per-participant responses and login
links. Sections refresh in place over ?oemaxlevel=1&only=SECTION. Full-width
inner layout (no pod sidebar/menu); view restricted to admins/instructors.
ian at
AGM status: drop the on-screen node tile and vote badge

The liveteststatus currentnode does not track what is actually on the
meeting screen, so showing it was misleading.
ian at
jtmx: data-refresh auto-refresh; use it to wrap the AGM status page

Any element with data-refresh="10s" now reloads its own contents on that
interval (ms/s/m/h), fetching data-url (default: the current page) and
swapping in the matching element from the response; fires jtmx:refreshed and
can be run on demand with jQuery(el).trigger("jtmx:refresh"). The AGM status
tool is wrapped in such a div at 10s instead of its own per-section timers.
ian at
Sanitize default admin; fix AGM status KPI counts

- admin.xml: remove Google OAuth clientid/refreshtoken and Stripe id;
  move the personal email to support@thoughtframe.ai
- TestBench: scope online/offline counts to the class roster (excluding
  instructors and archived members) and clamp offline at zero
- TestBench: add getOnlineProxyCounts(classId), reading uservoteclass per
  vote class for online members with a legacy studentclasses.proxycount
  fallback
- LiveTestManager: make parseCount public so the KPI code reuses it
- agmstatus: show a total plus per-voteclass online proxy tiles
ian at
Migrate tests to JUnit 6 and prune obsolete suites

- Upgrade to JUnit 6.1.3 via a shared gradle.properties version
- Migrate core/dam fixture tests from JUnit 3/4 to JUnit 5
  (assertions, annotations, constructors, imports) and add a shared
  BaseJUnit5Test / BaseEnterMediaJUnit5Test
- Tag fixture tests and make them opt-in via :core:fixtureTest and
  :dam:fixtureTest; the default test task runs the modern suite only
- Remove legacy JUnit 3 aggregators and the obsolete thoughtframe
  storefront legacyTest suite
- Drop obsolete/high-maintenance tests (ancient browser UAs, external
  services, timing tests) and trim stale URL/category assertions
- Remove committed OpenSearch indexes and gitignore them
- Add FrameDsl.getAnalysisModel(), required by JsonFrame tests

Default ./gradlew test: 277 tests, 0 failures.
ian at
agmstatus: full-width layout; fix BMVA roster importer

- Add class/agmstatus/innerlayout.xconf so the operator status page chains
  straight to the app theme instead of rendering inside the pod sidebar.
- importbmusers.groovy: fix Groovy parse error ($ in a GString), create the
  userprofile (roster/participants/login-links) and set activepod, and give
  every uservoteclass row a non-null <name> so re-runs don't fail in saveAllData.
ian at
ui: migrate all Font Awesome / Glyphicon usages to Bootstrap Icons

Only Bootstrap Icons (v1.11.3) is loaded site-wide; the legacy fa-*/glyphicon-*
classes were dead markup (no Font Awesome assets ship in the webapp). Converted
106 framework templates across status pages, live/voting UI, asset editor,
stores, settings and shared table headers. Handles Velocity $sorticon cases and
the fa-spinner -> Bootstrap .spinner-border swap.
ian at
ui: migrate Font Awesome / Glyphicon to Bootstrap Icons in dam/cms webapps

Completes the site-wide icon migration for the bases symlinked into the
thoughtframe webapp (emfrontend2, mediadb, entermedia, manager, system +
cms/openedit) plus the emfrontend2 theme CSS selectors that targeted the old
glyphicon classes. 150 templates.
ian at
ui: retarget remaining icon CSS selectors to Bootstrap Icons

- thoughtframe theme .star-rating .glyphicon-star -> .bi-star
- emfrontend2 theme .emtree/..fa-caret-right -> .bi-caret-right, .collectionrow ..fa-star -> .bi-star
Final stragglers after the Font Awesome/Glyphicon -> Bootstrap Icons markup migration.
ian at
class: add AGM delivery type with immediate launch

- new deliverytype 'agm' (AGM / Live Meeting)
- class home for agm auto-launches the first activity (liveroom -> livepage),
  with a join link fallback when no activity is attached
- livepage falls back to the AGM layout for agm pods (planlayout still wins)
- AGM class menu shows an 'Enter AGM Meeting' button
- AGM class layout hides course/module nav and course admin actions
ian at
livepage: AGM layout is slide/vote only, no video room

The AGM has no video room; remove the video panel entirely. The active
slide/vote now fills the page for members, with the admin/instructor tabs
(Participant Status, Manage Program, Response Summary, Response Details)
alongside for staff.
ian at
class layout: full-width AGM (hide pod sidebar)

The AGM experience is the livepage only; drop the class/pod sidebar for
deliverytype=agm so the slide/vote fills the width. Staff controls remain in
the livepage tabs/offcanvas.
ian at
livepage: AGM uses card-slot tiles (no pagination)

Use the banner/card1/card2/card3 slot layout for deliverytype=agm, matching
the member view: current slide plus slide tiles. Replaces the vote.html view
with its admin pagination.
ian at
class: AGM inner layout driven by deliverytype

- LearningGroupSearcher.validateFolder now sets layouttype from the group's
  deliverytype when a matching class/layouts/<type> exists (it used to strip
  layouttype entirely), else default.
- add class/layouts/agm inner layout: pod sidebar with the AGM menu + full
  card-slot content.
- AGM sidebar menu trimmed to the relevant links (Enter Meeting, AGM Status,
  Resources, Participants/Import/Email, Manage Program, Edit Pod, Add Activities).
- default inner layout made deliverytype-agnostic again.
ian at
fix: PageProperty(name,value) ctor does not exist; set value explicitly
ian at
agm layout: no sidebar; move AGM links into top-right admin dropdown
ian at
class layout: AGM include (fallback page uses default inner-layout)

The class instance _site.xconf layouttype/inner-layout is ignored for
fallback pages, so branch in the default inner layout: deliverytype=agm
includes class/layouts/agm/innerlayout.html (no sidebar, top-right admin
dropdown), everything else uses the existing default layout.
ian at
agm: shared toolbar (breadcrumb + admin dropdown) on meeting room and status page

Extract the AGM navbar into class/components/agm/toolbar.html and include it
from both the AGM inner layout and the agmstatus inner layout, so the status
page has the same top-right admin menu (Enter Meeting, AGM Status, Resources,
Participants, Manage Program, Edit Pod, Add Activities).
ian at
agmstatus: inherit the class/AGM inner layout

Drop the agmstatus-specific inner-layout pin (and its innerlayout files) so the
status page inherits class/layouts/default -> AGM branch: full width, no pod
sidebar, shared AGM toolbar.
ian at
agmstatus: fix login-link copy, pause refresh on links/responses, proxy class names

- agmCopy looked at btn.previousElementSibling (null); use the row input and
  navigator.clipboard with execCommand fallback.
- stop the 10s auto-refresh while the Login Links / Responses tabs are open so
  the list (and copy action) is not rebuilt underneath the user.
- KPI proxy tiles resolve voteclass id -> name via mediaarchive.getData(),
  falling back to 'Unassigned' instead of showing raw ids.
ian at
agmstatus: avoid $el in JS (Velocity parses $ refs in script)
ian at
WIP: tf native tools, genTools fixes, git sourcepath key, groovy 6, chatbot

- OpenCode plugin: tf verbs as native tools (tools.json manifest + @opencode/plugin shim)
- genTools: write to the tracked source (not a stray webapp/tf-eclipse copy); op-dispatched verbs no longer over-require params; find/runMain/launches duplicate-output fixed; 78 schemas incl semantic/reindex/runMain/launches
- gitrepository: sourcepath indexed, name non-multilanguage; findRepo falls back to sourcepath (/repo/<key>)
- VelocityGenerator: build+init before publishing the engine (startup race)
- Groovy 6 pins, duplicatesStrategy, Buildship auto.sync, warm-up page
ian at
git: resolve repo dirs by sourcepath; plural /repos mount; fix validateRepository NPE

- RepoResolver.open: use externalpath -> sourcepath -> id instead of the generated row id, and 404 when findRepo misses
- rename the top-level git mount repo/ -> repos/ (gitprojectroot=//repos)
- GitModule.validateRepository: set the GitManager before doesExist (was an NPE on save)
ian at
cleanup: untrack colony42 lucene index and *.js.map; ignore both

The colony42 lucene index is machine state (regenerates on boot), and the *.js.map files are ~70MB of dev-only source maps. Files stay on disk; git stops carrying them. .swarm-libs/libtorrent4j.so left tracked for now - the Java sync code imports libtorrent4j and we have not confirmed whether that native lib is loaded at runtime.
ian at
cleanup: untrack remaining *.map source maps and ignore *.map

Adds the *.css.map / *.mjs.map files the previous *.js.map pattern missed (217 files, 30.7 MB in HEAD). Files stay on disk.
ian at