Each turn (assistant response / tool result) is appended to a JSONL file
at ~/.uncaged/workflow/sessions/<sessionId>.jsonl during the loop.
On completion, the JSONL is read back, each turn is stored as a CAS node,
and the detail payload references them as a flat turns[] array in
chronological order. The session file is then deleted.
Benefits:
- Real-time observability: tail -f the JSONL to watch loop progress
- Crash recovery: partial JSONL survives process death
- Zero write contention: one file per session
- Detail stays a flat array for easy consumption by CLI/dashboard
Changes:
- New session.ts: initSessionDir, appendSessionTurn, readSessionTurns, removeSession
- loop.ts: append JSONL each turn instead of accumulating in-memory
- detail.ts: reads session JSONL → persists turns to CAS → stores detail
- agent.ts: passes storageRoot/sessionId to loop, cleans up session on completion
- types.ts: remove index from TurnPayload (order is implicit in JSONL/array)
- schemas.ts: sync with type changes
Ref: #433
buildClaudeCodePrompt was dropping ctx.edgePrompt entirely — the graph
transition instruction (e.g. 'Implement the plan') never reached the agent.
Now appended as '## Current Instruction' at the end of the prompt.
- StepRecord adds edgePrompt field (backward compat: defaults to "")
- StepNode CAS schema includes edgePrompt
- writeStepNode persists ctx.edgePrompt
- buildHistory exposes edgePrompt in StepContext
- buildBuiltinMessages reconstructs multi-turn moderator↔agent conversation:
system = role prompt + output format (stable prefix)
per prior visit: user (edgePrompt + inter-step summary) + assistant (output)
current: user (edgePrompt + recent summary)
- Zero extra persistence — pure function of CAS chain
- Stable prefix for LLM prompt cache hits
- 10 builtin tests pass, all other package tests pass
System message = agent identity (role prompt + output format instruction)
User message = moderator speech (task + edge prompt + history)
This reflects the workflow's core model: moderator speaks to agent
via the graph's edge prompt. Previously all content was in a single
system message with no user message, causing Claude API 400 errors.
- buildBuiltinPrompt now returns { system, user } instead of string
- agent.ts sends system + user as separate messages
- Tests updated accordingly
Switch from --output-format json to stream-json --verbose to capture
per-turn data. Detail now includes:
- model name
- usage (input/output/cache tokens)
- stopReason
- turns[] as individual CAS nodes with role, content, tool calls
Also addresses PR #421 review fixes:
- sessionId guard: skip cache write when sessionId is empty/undefined
- silent catch: log resume failures with debug tag 5VKR8N3Q
- atomic write: session cache uses temp+rename for crash safety
Closes#422
- Replace resolvePathInWorkspace with simple resolvePath (no boundary check)
- Remove UWF_BUILTIN_ALLOW_SHELL env gate from run_command
- Update tests accordingly
Per review: sandbox was false security with shell=true, and path
restrictions are unnecessary for a trusted agent environment.
Move getCachedSessionId/setCachedSessionId from workflow-agent-hermes
into workflow-agent-kit so all agent adapters can share the same
session cache logic.
Add cross-process session resume to workflow-agent-claude-code:
on re-entry (isFirstVisit=false), look up the cached sessionId and
use 'claude --resume' to continue with full conversation history.
Cache file renamed from hermes-sessions.json to agent-sessions.json
to reflect its shared nature.
Refs #418
Built-in role agent that uses workflow config models directly,
with its own tool-calling run loop. No external agent dependency.
- OpenAI-compatible chat completion client with tool_calls support
- P0 toolkit: read_file, write_file, run_command
- Integrates via createAgent factory from workflow-agent-kit
- CAS detail recording for each turn
- Path sandboxing and shell opt-in (UWF_BUILTIN_ALLOW_SHELL)
Hermes ACP _restore fails for custom providers — resolve_runtime_provider
throws and base_url/api_mode are lost, causing resume to silently create a
new session with no history. Prompt then returns empty text or refusal.
Disable resume by default. Set UWF_HERMES_RESUME=1 to opt back in.
Includes investigation notes in docs/investigations/.
Refs #418
- uwf log list: list log files with sizes
- uwf log show --thread <id>: filter by thread ID
- uwf log show --process <pid>: filter by process ID
- uwf log clean --before <date>: delete old log files
- Tests: 12 new tests covering all subcommands
Implemented by solve-issue workflow, biome fixes applied manually.
Closes#413
Refs #411, #410
- cmdLogList: list log files with sizes, sorted by date descending
- cmdLogShow: filter entries by thread, process, and/or date
- cmdLogClean: delete log files older than given date
- 12 tests covering all functions and edge cases
Fixes#413
- Add isFirstVisit: boolean to AgentContext
- Compute from steps history: !steps.some(s => s.role === role)
- hermes.ts: use isFirstVisit for first-entry vs re-entry logic
- buildInitialPrompt: always append edgePrompt as Moderator Instruction
- edgePrompt is never blanked — always the real moderator instruction
- New tests for first-visit, re-entry, and fallback scenarios
Refs #405, #407, #404
Two bugs fixed:
1. request_permission messages (JSON-RPC requests with both id+method) were
silently swallowed by the response handler, causing hermes to hang waiting
for permission approval. Now properly distinguish responses (id only) from
server requests (id+method).
2. uwf-hermes process never exited after completing because the hermes ACP
subprocess was still alive. Now explicitly close the ACP client after
agent completion so the subprocess terminates.
小橘 <xiaoju@shazhou.work>
The official TS SDK's ndJsonStream hangs indefinitely on prompt()
for sessions with 20+ messages (solve-issue planner). Root cause
appears to be a stream backpressure issue in the SDK's ReadableStream
adapter.
Switch back to readline-based line parsing which reliably receives
all JSON-RPC responses. Also handle session/request_permission
inline (auto-approve, yolo mode equivalent).
Ref #398
UwfAcpClient now tracks all session/update events:
- agent_message_chunk → assistant message content
- agent_thought_chunk → assistant reasoning
- tool_call → pending tool invocation (name + rawInput)
- tool_call_update (completed/failed) → assistant tool_call + tool result
Messages are accumulated across prompts (same session) and stored
via storeHermesSessionDetail, restoring the full structured detail
(turns with tool calls, reasoning) that was lost in the initial ACP
migration.
Ref #398
The client is now created once in createHermesAgent() and shared by
runHermes and continueHermes closures. This preserves conversation
context during frontmatter retry loops — continue() sends a follow-up
prompt on the same ACP session instead of starting a new one.
Client is cleaned up via process.on('exit').
Ref #398
Replace 250-line custom ACP client with official TypeScript SDK.
Uses ClientSideConnection + ndJsonStream for stdio transport.
Same public API (connect/prompt/close), 115 lines, zero custom protocol code.
Ref #398
Remove spawnHermes, spawnHermesChat, spawnHermesResume, parseSessionId,
and buildResultFromSession. runHermes and continueHermes now use
HermesAcpClient for structured JSON-RPC communication.
Session ID comes directly from ACP protocol, eliminating #380 race
condition. Agent output collected via streaming chunks instead of
session file loading.
Phase 2 of RFC #398Fixes#380
Implements JSON-RPC client that communicates with `hermes acp` via
stdin/stdout. Replaces fragile stdout/stderr parsing with structured
protocol: initialize → session/new → session/prompt → collect chunks.
Session ID comes directly from protocol response, eliminating the
race condition in #380.
Phase 1 of RFC #398
buildOutputFormatInstruction now includes explicit language telling agents to
output ONLY schema-defined fields and to focus on their role's deliverable.
Fixes#394
Claude Code CLI adapter for the workflow engine, mirroring
workflow-agent-hermes architecture. Spawns `claude -p` with
`--output-format json` for structured output parsing.
Refs #391
Send a test completion request after configuration to verify the model
is reachable. If validation fails, warn the user and suggest trying a
different model or checking their settings.
Fixes#335
Replace hardcoded 5-field example with schema-driven generation.
Now shows actual enum values, types, and required markers for
each role's frontmatter schema.
Fixes#389
小橘 <xiaoju@shazhou.work>
Replace hardcoded 5-field candidate with schema-driven extraction.
Now reads outputSchema properties and picks matching fields from
parsed frontmatter, supporting role-specific fields like plan,
approved, success.
Falls back to standard 5 fields when schema has no properties.
Fixes#388
小橘 <xiaoju@shazhou.work>
Agent CLI outputs plain CAS hash (not JSON), engine parses plain hash.
StepOutput no longer carries sessionId — session info is already in CAS detail.
Keeps the valuable parts of #385: sessionId in AgentRunResult (process-internal),
continue support, and frontmatter retry loop.
Breaking changes:
- AgentRunResult now requires sessionId field
- AgentOptions now requires continue function
- Agent CLI outputs JSON {stepHash, sessionId} instead of plain CAS hash
- Engine parses JSON output (with legacy CAS hash fallback)
New features:
- Frontmatter validation retry: if agent output lacks valid frontmatter,
engine calls agent.continue() up to 2 times with correction message
- Session tracking: sessionId flows from agent → engine → StepOutput
- Hermes agent: session parse failure is now a hard error (no raw text fallback)
- Hermes agent: supports --resume for continue sessions
Closes#384
Agent output must contain valid YAML frontmatter matching the role schema.
If frontmatter parsing fails, the step fails immediately with a clear error
instead of falling back to an LLM extraction that can fabricate values.
The extract module remains as a public API export but is no longer used
in the agent run loop.
Breaking change: agents that relied on LLM extraction to produce valid
output will now fail. They must output proper frontmatter.
Add --count/-c flag to 'uwf thread step' for running N steps in one
invocation, stopping early if $END is reached.
- cmdThreadStep now loops up to count times, delegates to cmdThreadStepOnce
- CLI parses -c/--count, defaults to 1 (backward compatible single output)
- Validation rejects 0, negative, and non-integer counts
- 7 new tests covering CLI parsing and count validation
Fixes#373
Co-authored-by: uwf-hermes (solve-issue workflow)
xingyue
xiaoju
xiaomo
scottwei
jiashuang