Register the test_gating and network_guard plugins via
config.pluginmanager.import_plugin() in pytest_configure, which is the
supported mechanism for non-top-level conftest files under pytest 9.1.1+.
This restores test gating and network-guard enforcement that was dropped
when pytest_plugins was removed.
Also fix the core SDK sys.path insertion: praisonaiagents lives at
src/praisonai-agents, so traverse two parents ('..','..') to reach
<root>/src, not three.
Co-authored-by: Mervin Praison <[email protected]>
Addresses two correctness gaps in the crash-recovery labelling (PR #3388):
- CodeRabbit: hand recovery_annotator its own fresh payload copy and only
adopt a dict result, so a mutate-then-raise no longer leaks a half-labelled
payload; validate non-dict returns.
- Greptile P1: keep a crash-recovered entry in the 'recovered' state on a
transient re-send failure (mark_failed keep_recovered) so the duplicate
label survives the retry instead of being demoted to 'failed' and dropped.
Adds 3 regression tests (mutation-then-raise fallback, non-dict return,
sticky-recovered retry). All 69 outbox/delivery/reconcile tests pass.
Co-authored-by: Mervin Praison <[email protected]>
- disarm racing an in-flight wedge no longer triggers os._exit (re-check
_stop before the destructive exit) so a deliberate disarm never causes a
supervisor restart (Greptile P1)
- reject NaN/inf probe_interval_s to avoid a dead watchdog worker or a
hung shutdown via Event.wait/Thread.join (Greptile P2)
- add tests for both cases
Co-authored-by: Mervin Praison <[email protected]>
Add an optional recovery_annotator hook to OutboundQueue.drain, applied only
on the unreconciled 'recovered' branch, so a mid-send-crash message re-sent
at-least-once can be visibly marked as a possible duplicate instead of being
re-delivered silently. DurableDelivery gains a backward-compatible
mark_recovered flag (off by default) that prefixes recovered string replies
via _annotate_recovered_payload, threaded through setup_durable_delivery.
Co-authored-by: MervinPraison <[email protected]>
Add a pure-stdlib, fail-open LoopWatchdog primitive to core
(praisonaiagents.gateway) that runs on a dedicated OS thread, probes the
asyncio loop via call_soon_threadsafe, and after N missed probes dumps
all-thread stacks (faulthandler) and hands the process back to the
supervisor via os._exit(EX_TEMPFAIL). Opt-in and fail-open by design.
Co-authored-by: MervinPraison <[email protected]>
Passive-recorded group messages now land on the same shared per_chat
group key that a subsequent addressed turn reads from, so under
session_scope="per_chat" the bot can actually see the preceding
conversation (Greptile P1 / CodeRabbit Major). record_passive() threads
chat_id/thread_id/account routing and appends directly under a per-key
threading lock instead of the cross-thread mirror bridge, which would
dead-lock when called inline on the inbound handler's own event loop.
Strengthened the observe test to assert routing is threaded and added a
real BotSessionManager per_chat persistence test proving the passive
entry is visible under the group key and does not leak into the sender's
per_user history.
🤖 Generated with [Claude Code](https://claude.ai/code)
Co-authored-by: Mervin Praison <[email protected]>
Address reviewer P1s on the restart-continuation path:
- boot scan requested an explicit large limit so interrupted turns beyond
the store's 50-session default window are no longer silently skipped
- channel_target now falls back to the persisted client_id when it encodes
a channel:target origin, so the field is a live consumer of existing
state rather than dead code for sessions the setter never touched
Adds tests for client_id-derived origin and >50-session scan coverage.
Co-authored-by: Mervin Praison <[email protected]>
A malformed gateway.drain_timeout / gateway.reload_drain_timeout in a
hot-reload edit made _coerce_timeout() return None, which was assigned
unconditionally and silently cleared a previously-configured live drain
window. Distinguish malformed (keep live value via _KEEP sentinel) from
explicit disable (None/absent -> None). Update/extend tests accordingly.
Co-authored-by: Mervin Praison <[email protected]>
On boot, before serving traffic, scan persisted sessions for an in-flight
turn (is_executing / non-empty pending_inbox) lost to a restart. For
channel-originated sessions, deliver an exactly-once "interrupted — resuming"
notice via the existing durable delivery router (idempotency keyed on
(session_id, event_cursor)) and re-drive the turn when the agent is still
registered — independent of any client reconnect.
Reuses the existing durable session store + delivery router rather than
adding new core protocol machinery, keeping the change lightweight and
backward-compatible.
Co-authored-by: MervinPraison <[email protected]>
Add a core HOT_APPLIABLE_KEYS registry + SupportsHotReload protocol in
praisonaiagents/gateway/config.py declaring the closed set of config paths
safe to apply in place. The bot gateway now classifies those paths as HOT in
_build_reload_plan (unknown keys still fall back to a restart plan — fail-safe)
and implements apply_hot_reload to mutate live state (logging level, drain /
reload_drain timeouts) without restarting channels or agents. Applied paths are
already reflected in the existing reload status/health.
Co-authored-by: MervinPraison <[email protected]>
Under the observe policy, group messages that do not mention the bot are
recorded into the session transcript as passive user context (no agent run)
so the bot has memory of the conversation when it is next addressed.
Reuses the existing thread-safe mirror append path via a new
BotSessionManager.record_passive() instead of adding a new core enum /
classifier, keeping the surface area minimal.
Co-authored-by: MervinPraison <[email protected]>
- Reject path-separator/traversal agent names (validate_agent_name) so
writes stay inside .praisonai/agents/ (Greptile Security P1).
- Post-write validation now re-parses the exact file written instead of a
precedence lookup, and warns when a higher-precedence definition shadows
the name so the reported `run --agent` step is never misleading (P1).
- Add tests for traversal rejection and global/project shadow warning.
Co-authored-by: Mervin Praison <[email protected]>
Add a `create` command beside `list`/`show` on the agent Typer app that
authors a valid, permission-scoped .praisonai/agents/<name>.md from a one-line
description. Interactive prompts plus a scriptable
--describe/--model/--permission/--global/--yes form. Permission presets reuse
the existing `mode` shorthands (MODE_RULES) so authored agents round-trip
losslessly through CustomDefinitionsDiscovery and inherit the runtime's
permission grammar. System-prompt drafting reuses the plain Agent path proven
by `init --generate`, degrading to an editable stub on any failure.
Co-authored-by: MervinPraison <[email protected]>
- _shim.py: skip None placeholders in sys.modules when pre-aliasing the
moved subtree so an unloaded optional module can't poison the old dotted
name (import old.sub -> "None in sys.modules"); _AliasFinder still resolves
it on demand.
- codex_cli.py: use an _UNSET sentinel in _build_command so an explicit
output_file=None disables the instance default and parses stdout, instead
of silently falling back to self.output_file (execute_with_schema).
Co-authored-by: Mervin Praison <[email protected]>
Restores eager subtree pre-registration in _shim._register_submodules. The
lazy-only variant broke module identity: a combined `from praisonai.mcp_server.cli
import X` imported the parent shim, then resolved `.cli` through _AliasFinder,
whose nested import_module re-entered mid-import under the import lock and
double-loaded the moved file — producing duplicate classes and breaking the
15 mcp/doctor CI tests that patch the shim path.
Also make the forced-format integration helpers (gemini stream/execute_with_stats,
codex stream, cursor stream) pop caller-supplied override keys before threading
the forced value, matching claude_code.py, so callers can't trigger a duplicate-
keyword TypeError (greptile P2).
Co-authored-by: Mervin Praison <[email protected]>
Address 4 P1 review findings in eval/trials.py:
- Isolation: rebind clone.chat_history to a fresh list and give the clone its
own system-prompt cache so per-attempt state resets never mutate/evict the
caller's live agent state.
- Scorer safety: a raising EvalCase.verify (and any scorer path) is now
recorded as a failed TrialScore instead of propagating through gather and
aborting the whole report.
- Timeout: run attempts on a dedicated ThreadPoolExecutor shut down with
wait=False so a wedged agent call cannot block run_trials past the timeout;
documented that a leaked worker thread cannot be force-killed.
- Duplicate case names: merge/extend attempts (with re-numbered indices)
instead of overwriting, so no attempts are silently discarded.
Add 4 regression tests covering each fix.
🤖 Generated with [Claude Code](https://claude.ai/code)
Co-authored-by: Mervin Praison <[email protected]>
Address three P1 issues in the trials→SFT exporter:
- --all (only_passed=False) now retains unscored attempts instead of
dropping them in _classify.
- Alpaca conversion no longer drops the user prompt when a system message
is present; system→instruction, all user turns→input.
- Score-only pass detection compares against the report's pass_threshold/
threshold/min_score when present, falling back to score>0 only otherwise.
Co-authored-by: Mervin Praison <[email protected]>
Bridge verification to data generation: take a serialised trials report
(K scored attempts per case with captured trajectories), keep the passing
attempts via rejection sampling, and write them as a ShareGPT dataset the
existing trainer consumes unchanged. Adds export_trials() + ExportSummary,
provenance sidecar ({path}.meta.json), and a `from-trials` CLI command
beside generate/dedup/validate. Report is duck-typed over JSON (the contract
with the sibling core trials engine) so there's no hard dependency on unlanded
core code. Pure stdlib, additive, confined to the data namespace.
Co-authored-by: MervinPraison <[email protected]>
Move wrapper-package discovery out of the Tier 2 ui.py CLI into the lazy
_wrapper_bridge via a new wrapper_package_path() helper, per ARCHITECTURE.md
§2 cross-tier boundary. Preserves preset resolution and the None fallback
(install hint) when the wrapper is absent.
Co-authored-by: Mervin Praison <[email protected]>
The UI loader was extracted to praisonai_code but resolved the bundled
default_app.py path via __file__-relative arithmetic anchored in
praisonai_code, which has no ui_* directories. The five bundled UI presets
ship in the praisonai wrapper package. Resolve the path against the installed
wrapper via importlib.util.find_spec("praisonai"), falling back to a clear
'pip install "praisonai[ui]"' hint when the wrapper is absent.
Co-authored-by: MervinPraison <[email protected]>
The executor context manager called shutdown(wait=True) on __exit__, which
re-blocked on a coroutine still hung past PRAISONAI_RUN_SYNC_TIMEOUT — defeating
the timeout (Greptile P1). Manage the pool explicitly and shutdown(wait=False)
in finally so a timed-out tool call returns immediately.
Co-authored-by: Mervin Praison <[email protected]>
The wrapper copy of agent_tools.py had drifted from the canonical
praisonai_code copy (missing the ACP cwd forwarding fix), and the two
resolved via different import spellings. Collapse the wrapper copy to a
thin re-export of the canonical praisonai_code implementation so both
import styles resolve to a single owner.
Reconcile the canonical module-local _run_sync bridge to keep the union
of both historical guarantees: running-loop safety (offload to a worker
thread) AND a timeout, without adding a wrapper dependency (C7 gate).
Co-authored-by: MervinPraison <[email protected]>
Address Greptile P1 review findings on PR #3358:
- _remove_worktree now returns a reason string when `git worktree remove`
fails (nonzero exit or exception) instead of silently returning None.
The caller keeps the tracking entry and records a preservation comment,
preventing orphaned worktrees/branches with dropped tracking.
- create_task persists the resolved top-level repo_path into task metadata
so the repo linkage that drove the worktree auto-upgrade is retained
between create and dispatch (an explicit metadata.repo_path still wins).
Adds tests for failed/exception removal tracking and repo_path persistence.
Co-authored-by: Mervin Praison <[email protected]>
Refinement 1: repo-linked tasks auto-upgrade workspace_kind to "worktree" at
create_task time (linkage-based, not content-inspecting). Explicit
workspace_kind and auto_worktree=false are respected. Derived branch
kanban/<task_id> is persisted for the dispatcher to consume.
Refinement 2: _remove_worktree is lossless by default — refuse to delete a
worktree that is dirty or has commits not in base; preserve it and record a
worktree_preserved comment. Removal with outstanding work is an explicit
force operation only.
Co-authored-by: Mervin Praison <[email protected]>
Addresses two P1 review issues on the workspace-binding change:
- Resolve the UnifiedSession once (before workspace binding + tool load) and
pass it into _bind_resume_workspace, eliminating a second independent
last-session lookup that could race a concurrent CLI process and bind tools
to a different directory than the loaded conversation.
- When a resumed session's saved workspace no longer exists, pin
PRAISONAI_WORKSPACE to cwd instead of leaving it unset, so a stray legacy
PRAISON_WORKSPACE cannot leak through the tool loader and contradict the
"using current directory" warning.
Adds regressions for cwd-pinning over the legacy var and for
provided-session-without-relookup; updates the existing fallback test.
Co-authored-by: Mervin Praison <[email protected]>
The `praisonai code --workspace` flag was dead: it wrote PRAISONAI_WORKSPACE
but ToolConfig.from_env read PRAISON_WORKSPACE, and _load_interactive_tools
unconditionally overrode config.workspace with os.getcwd(). Fix the env-name
mismatch (honour both, prefer canonical PRAISONAI_) and resolve the workspace
as PRAISONAI_WORKSPACE -> PRAISON_WORKSPACE -> cwd instead of hardcoding cwd.
Also bind resumed sessions to their directory: --continue/--session now reads
back the persisted UnifiedSession.workspace and roots the tools there, warning
and falling back to cwd when that directory is gone.
Scope kept minimal per AGENTS.md: no new WorktreeService — worktree isolation
already exists via core GitWorktreeAdapter (wired into `praisonai run`), so a
parallel wrapper service would duplicate it.
Co-authored-by: MervinPraison <[email protected]>
Addresses the Greptile finding that reconnect() restarts the same bot
instance still holding the rejected token. On wake from the parked
credential state the supervisor now opportunistically calls an optional,
duck-typed refresh_credentials() hook so an out-of-band credential repair
(e.g. a rotated env var) driving a bare reconnect() self-heals without a
full restart. No-op for bots without the hook (env-var tokens are already
re-resolved by start()/hot-reload). Best-effort: a failing hook cannot
wedge supervision.
🤖 Generated with [Claude Code](https://claude.ai/code)
Co-authored-by: Mervin Praison <[email protected]>
Address two P1 review findings on the gateway restart-replay change:
- Only persist the start-flags artefact AFTER startup validation passes
(import/config/agent-file), just before the gateway binds, so a failed or
rejected start on the same host:port no longer clobbers the running
gateway's saved posture.
- restart --drain-timeout now defaults to None; an omitted flag replays the
persisted drain window (and falls back to 10s only when none is persisted)
instead of silently forcing a fixed 10s and shortening production drains.
An explicit value still wins for both the OLD-process drain and the relaunch.
Adds 3 regression tests covering omitted/explicit drain replay and the
no-persisted-value fallback.
🤖 Generated with [Claude Code](https://claude.ai/code)
Co-authored-by: Mervin Praison <[email protected]>
Direct (non-service) `gateway restart` silently dropped the runtime flags
the process was started with (--openai-api, --reliability,
--max-concurrent-runs, ...), reverting production settings to defaults.
`start` now persists the non-None runtime flags to
~/.praisonai/gateway.start.<host>.<port>.json (keyed by bind host:port,
matching the PID lock), and `restart` loads and replays them so the
relaunch faithfully reproduces the running gateway. Explicit restart flags
(config / agents / drain-timeout) still win over persisted values.
Co-authored-by: MervinPraison <[email protected]>
Classify runtime 401/403 (and platform text equivalents like invalid_auth/
token_revoked) as a distinct CREDENTIAL_UNAVAILABLE outcome instead of a
generic FAILED. On a runtime credential rejection the supervisor now parks in a
named, redacted degraded state (reusing the boot-time "credential unavailable"
surface from #3159 in health()), stops hammering the invalid token, and
auto-recovers on reconnect/hot-reload without a full process restart.
Co-authored-by: MervinPraison <[email protected]>
Address two P1 review findings on praisonai init --generate:
- Streaming stringified: _generate_agents_md used agent.start(), which
auto-enables streaming in a TTY and returns a generator whose repr was
written to AGENTS.md. Switch to agent.run() (always silent, returns str).
- Global mode wrote ~/AGENTS.md: with --global, base.parent is $HOME so the
agent scanned home and wrote home-tailored rules outside the repo. Resolve
the AGENTS.md target from git root / cwd independently of --global.
Add tests: --global still writes AGENTS.md to the repo root (not home), and
_generate_agents_md calls run() (not start()) returning the string body.
Co-authored-by: Mervin Praison <[email protected]>
Addresses the two Greptile P1 findings on #3346 so the non-zero-on-failure
contract holds across every reachable run path:
- _run_from_file: add `except typer.Exit: raise` so the deliberate failure
exit is not recaught by the broad handler (fixes duplicate reporting).
- _try_attach_runtime: classify the warm-runtime result and report failure so
an empty result exits non-zero instead of silently succeeding.
- _run_prompt_profiled / _run_from_file_profiled: apply the same failure
contract after the profiling report so profiled runs honour it too.
Adds a regression test for the warm-runtime attach path.
Co-authored-by: Mervin Praison <[email protected]>
Adds an opt-in `--generate` flag to `praisonai init` that runs a short,
read-only analysis agent (reusing the existing Agent runtime) to produce a
concise, repository-tailored AGENTS.md at the repo root. Feeds the agent a
cheap pre-scan (top-level tree, detected manifests, README head) to bound cost.
Non-destructive: respects --force, never overwrites an existing AGENTS.md, and
degrades gracefully to the static scaffold when no provider credential is
present or generation fails. The output lands where rules_manager already
discovers it on the next run.
Co-authored-by: Mervin Praison <[email protected]>
praisonai run reported success and exited 0 regardless of whether the
agent produced a result. Wire the existing run result (None/empty ==
failure) through the CLI: emit ok=<success> to the event bridge, a
machine-readable {"status": "failed"} object under --output json, an
actionable remediation, and raise typer.Exit(1) on failure — mirroring
the existing arg-error exit convention. Successful runs are unchanged.
Co-authored-by: Mervin Praison <[email protected]>
The running-loop fallback in _structured_completion/_run_coro_sync spawned a
bare threading.Thread, which does not inherit contextvars. run_sync() inside it
therefore resolved via _default_bridge() -> empty _bridge_var -> shared _BG,
bypassing an active scoped_bridge() and leaking the fallback onto the global
loop. Run the worker under contextvars.copy_context() so both paths resolve the
same AsyncBridge (scoped override, else shared default), preserving per-session
loop isolation.
Co-authored-by: Mervin Praison <[email protected]>
- auto.py: runner threads in _structured_completion and _run_coro_sync now
dispatch to the shared AsyncBridge via run_sync() instead of asyncio.run(),
reusing one process-wide loop/thread and per-loop connection pools.
- auto.py: _completion_impl preserves the original LiteLLM exception and
chains it into the OpenAI fallback failure (raise ... from) so users see
both provider tracebacks instead of only the masking OpenAI one.
- registry.py: _validate_adapter treats a **kwargs catch-all as satisfying
the protocol keywords and also validates arun, unblocking third-party
entry-point adapters that forward **kwargs.
Co-authored-by: MervinPraison <[email protected]>
Embed turn prerequisite errors into the JSON payload instead of printing a
trailing plain-text line, so --json consumers get a single valid document.
Also drop dead critical-tool branch in is_already_approved.
Co-authored-by: Mervin Praison <[email protected]>
1. Circuit breakers: register weakref.finalize at breaker creation so the
process-global registry entry is removed when the Agent is GC'd, closing
the CPython id-reuse window where a new agent could inherit a stale OPEN
breaker. No longer relies on close()/aclose() being called. Uses a
staticmethod callback so the finalizer does not keep the agent alive.
2. Global tool-registry fallback: gate the @tool registry lookup in
execute_tool() behind ToolConfig(allow_global_tools=True) (default False).
By default an agent is scoped strictly to its declared tools=[...] and
cannot execute an @tool-decorated callable it was never given. Plugin-host
style dynamic dispatch can opt in via tool_config.
3. Sync run_task(): re-raise on fail_on_callback_error / fail_on_memory_error,
mirroring arun_task() so the documented fail-fast flags behave identically
on the sync and async surfaces.
Co-authored-by: MervinPraison <[email protected]>
Extract shared probe/shell/turn logic to praisonai_bot.gateway.preflight,
add praisonai gateway test orchestrator, doctor bots shell/probe checks,
fix doctor --json single-document output, and per-channel turn probe gate.
Critical tools honour mark_approved via the approval registry, shell enable
syncs AutoApproveBackend to the registry, @mentions run routing handlers,
and gateway doctor gains an offline --turn probe for inbound simulation.
Addresses two Greptile P1 findings on the allow_shell opt-in:
1. Routed agents lose shell tools: _inject_routing_handler swapped in a raw
agent from self._agents, stripping execute_command applied to the default
channel clone. Re-apply enable_shell_tools to the routed agent (cloned once
per channel+agent so non-shell channels never gain the shell tool).
2. Manual approval retained auto-approval (security): with auto_approve_shell
false and no Slack approval channel, a pre-existing AutoApproveBackend was
left installed, silently auto-approving shell. Now fail-closed with a
deny-by-default backend that rejects shell commands.
Adds tests for both fail-closed approval paths.
Co-authored-by: Mervin Praison <[email protected]>
- Fix missing approval_webhook_url becoming the literal string "None"
which bypassed WebhookApproval's ValueError guard; now falls through
to platform/gateway approval instead of building a broken backend
- Add fail-closed approval_mode validator (channel|gateway|http|webhook)
so config typos are rejected at load time, matching other enum fields
- Add regression tests for webhook-mode with and without a URL
Co-authored-by: Mervin Praison <[email protected]>
Extend allow_shell to route human-in-the-loop approvals through platform
backends (Slack/Telegram/Discord) plus gateway, HTTP, and webhook modes,
with gateway fallback for channels without native wiring.
Address reviewer feedback on the CLI_BACKEND_EXECUTE hook path:
- Security: redact prompt/system argv values (redact_command) at the
payload serialization boundary and in the plugin bridge, so prompts and
system instructions do not leak into downstream log sinks. The live
in-memory command field is preserved for programmatic consumers.
- Observability: emit the hook on failure too (extracted
_emit_cli_backend_hook helper) so subprocess startup errors and timeouts
remain traceable via error=<message>, instead of only on success.
- Tests: cover redaction (serialization + plugin bridge) and failure-path
hook emission.
Co-authored-by: Mervin Praison <[email protected]>
Log subprocess command and confirm PraisonAI skips LLM HTTP when cli_backend delegates a turn.
Enable with PRAISONAI_CLI_BACKEND_DEBUG=1 or LOGLEVEL=DEBUG.
Gemini backend forwarded --resume on every turn including the first,
which would fail resuming a non-existent session. Gate on
CliSessionBinding.is_resume for parity with codex/grok/claude.
Co-authored-by: Mervin Praison <[email protected]>
Codex backend never forwarded the session id to `codex exec`, so
multi-turn conversations lost context (Greptile P1). Now forwards
`resume <session_id>` on resume turns. Gate both codex and grok
session forwarding on CliSessionBinding.is_resume so the first turn
does not attempt to resume a non-existent session.
Co-authored-by: Mervin Praison <[email protected]>
- Isolate ANTHROPIC_TOKEN/CLAUDE_CODE_OAUTH_TOKEN env in the expiring-token
test so the Keychain read-only path is deterministically exercised (was
leaking a real CI OAuth token and returning the wrong access token).
- Assert both anthropic-beta policy changes: fine-grained-tool-streaming is
present and the unsupported context-1m long-context beta stays absent
(per CodeRabbit review).
Co-authored-by: Mervin Praison <[email protected]>
PraisonAI read Keychain OAuth tokens but refreshed without persisting back,
rotating refresh tokens and breaking Claude CLI and other tools on the same
machine. Use read-only credential resolution, skip LLM auth refresh for
claude-code, fix invalid_request_error misclassification, and drop the
context-1m beta header unsupported on many subscriptions.
The async retry loop emitted RETRY events via the sync emit() path, so
consumers registered through the public add_async_callback() API passed
has_callbacks but never received the event. Add an async-aware
_aemit_retry_stream_event helper (using emit_async) and share event
construction via _build_retry_stream_event. Adds tests for the async-only
callback path and the no-callback no-op case.
Co-authored-by: Mervin Praison <[email protected]>
Add StreamEventType.RETRY emitted alongside the existing ON_RETRY hook (guarded by has_callbacks for zero-overhead) in both sync and async retry loops, carrying attempt/max_attempts/delay/reason in metadata. Map it to a canonical run.retry NDJSON event in the CLI stream-json bridge so consumers get a live 'retrying in Ns (attempt k/N)' signal. Non-JSON/silent modes stay silent.
Co-authored-by: Mervin Praison <[email protected]>
A TypeError/ValueError raised inside a successfully-bound tool (domain
validation) no longer receives a misleading required/optional parameter
hint. The hint is now gated on inspect.signature().bind() confirming the
failure is an argument-binding error, so runtime errors surface their
real message and the model fixes the value, not the parameter names.
Co-authored-by: Mervin Praison <[email protected]>
Adds a zero-overhead resilience layer to the core agent tool-dispatch path.
When a tool call cannot be resolved, the runtime now:
1. Self-repairs case/separator drift (e.g. WebSearch/web-search -> web_search)
against the agent's active tools and executes if a unique match exists.
2. Otherwise returns a model-readable error naming the nearest valid tool
(fuzzy) and listing the available tools so the model can self-correct.
3. On an argument-binding failure, echoes the tool's required/optional params.
No new public API or Agent params; stays on the existing dispatch path.
Co-authored-by: MervinPraison <[email protected]>
- Security: refuse non-localhost bind without --api-key/PRAISONAI_SERVE_API_KEY,
mirroring jobs/server.py; enforce the resolved key unconditionally.
- Performance: cache one AgentsGenerator per app via FastAPI lifespan (built at
startup, closed on shutdown) instead of rebuilding YAML parse + framework
resolution + a 32-worker pool per request; falls back to praisonai.arun if
caching fails.
- Correctness: /agents and /agents/{name} now run the same cached generator, so
identical YAML gets identical lowering (ToolResolver, tool_timeout, approval,
guardrails, retry). Removed the hand-rolled PraisonAgent registration that
silently dropped those safety fields.
Co-authored-by: Mervin Praison <[email protected]>
Remove 13 knowledge.db binaries plus managed_ids.json, blog_post.md,
tutorial.md, token_metrics_export.json, market_research_results.json that
were accidentally committed (PR notes explicitly say not to commit local
artifacts). Add .gitignore rules to prevent recurrence.
Co-authored-by: Mervin Praison <[email protected]>
Address reviewer findings on PR #3321:
- Core neutralize_untrusted_text now also collapses U+2028/U+2029/U+0085,
which render as line breaks but bypassed the ASCII control filter.
- Bot _attribute() fallback (core helper unavailable / version skew) now
mirrors the sanitiser: collapses all newline-like separators, strips
control chars, and bounds length.
- Added tests for Unicode separators (core + bot) and the fallback path.
Co-authored-by: Mervin Praison <[email protected]>
Add a core, pure-stdlib helper neutralize_untrusted_text() in
praisonaiagents.session.context and apply it by default to the
third-party-controlled sender display name before it is interpolated into
the [{sender}] attribution prefix. Newlines/CR are collapsed, control chars
stripped and length is bounded, so a hostile name can no longer masquerade
as a fake system directive the model re-reads every turn. Well-behaved names
render byte-identically. On by default, no new config knobs.
Co-authored-by: Mervin Praison <[email protected]>
The durable-reference encoding for long interactive reply/select values only
took effect when a shared store was explicitly passed in tests; every
production renderer/registry used the no-store default and fell back to the
unrecoverable hash. This shares one InMemoryCallbackPayloadStore between the
Telegram render side (render_presentation -> adapt_presentation) and the inbound
registry so long values round-trip losslessly past the 64-byte callback cap.
- create_registry(store=): thread an optional store into the InteractiveRegistry
- TelegramPresentationRenderer.render(callback_store=): pass store to adapt
- TelegramBot: construct one shared store, inject into registry + render
- add regression test asserting production render+dispatch round-trip
Co-authored-by: Mervin Praison <[email protected]>
- POST /agents/{agent_name} now dispatches to the requested registered agent
and returns 404 for an unknown name, instead of silently running the whole
YAML workflow (fixes named-agent regression flagged in review).
- Both routes build a lightweight per-request AgentsGenerator that borrows the
cached generator's shared adapter, tool-timeout pool and resolver while owning
its own cli_config, so concurrent requests overlap without a run-length lock.
Co-authored-by: Mervin Praison <[email protected]>
Addresses two P1 review findings on the tool self-repair path:
- Fold the bind-failure parameter schema into the error string so it survives
conversion to ToolExecutionError (which keeps only the message) and reaches
the model.
- Expand MCP instances into their contained tools via a shared
_iter_active_named_tools() helper so MCP tool names participate in
case/separator repair and appear in the corrective inventory instead of an
empty/opaque container.
Adds 3 tests (public-path param hint, MCP inventory, MCP name repair).
Co-authored-by: Mervin Praison <[email protected]>
Address reviewer feedback on the prompt-injection defence:
- Greptile P1 (security): the bot's _attribute() fallback (taken when the
core helper import fails on an older praisonaiagents still allowed by the
>=1.6.152 range) only replaced CR/LF, leaking control chars and oversized
display names into every shared-chat prompt. It now mirrors the core
helper: collapse newlines, strip control chars, collapse whitespace and
length-bound to 240 — byte-identical to neutralize_untrusted_text.
- Greptile P2: neutralize_untrusted_text with max_chars<=3 returned the
3-char "..." string, exceeding the advertised bound. Small bounds now
truncate hard so len(out) <= max_chars always holds.
- Tests: add small-bounds regression (core) and a fallback-path hostile-name
regression that forces the helper import to fail (bot).
Co-authored-by: Mervin Praison <[email protected]>
An unquoted bare prompt (praisonai build a weather agent) arrives as
multiple argv tokens, but the modern `run` command takes a single
positional target. Join the flag-free tokens into one argument so the
whole prompt reaches `run` intact instead of Typer rejecting the extra
positionals. Addresses Greptile P1 (multi-token prompt).
Also adds a regression test pinning that flagless operational commands
(serve/call/realtime/debug/lsp/diag) stay Typer commands and are never
misrouted to `run` as prompts (Greptile P1 #2 was a false positive).
🤖 Generated with [Claude Code](https://claude.ai/code)
Co-authored-by: Mervin Praison <[email protected]>
Interactive reply/select values that exceed a channel's callback byte-cap
(e.g. Telegram's 64-byte limit) were hashed with no way to recover the
original value, so long option values (URLs, file paths, free-text) made
menus silently non-functional. This generalises the durable-reference
pattern already used by approvals to ordinary interactive callbacks.
- Add CallbackPayloadStoreProtocol (peer of ApprovalStoreProtocol) plus a
bounded, zero-dependency InMemoryCallbackPayloadStore default in core.
- Encoders persist an overflowing value under a short reference and emit
reply:@<ref> / select:<action_id>:@<ref>; inline values are unchanged.
- InteractiveRegistry(store=...) resolves @<ref> back to the exact value on
dispatch, preserving the select action_id prefix, and fails closed on an
unknown/expired reference. Without a store the prior hash behaviour is kept.
Additive and backward compatible: adapt_presentation gains an optional
callback_store kwarg; registry gains an optional store arg.
Co-authored-by: MervinPraison <[email protected]>
- Refuse non-localhost bind without --api-key/PRAISONAI_SERVE_API_KEY,
mirroring jobs/server.py policy (POST /agents can run YAML-defined tools).
- Cache one AgentsGenerator per app via FastAPI lifespan instead of rebuilding
it (YAML re-parse, registry re-resolve, fresh 32-worker pool) per request;
serialise the shared cli_config with an asyncio.Lock and close() on shutdown.
- Converge /agents and /agents/{agent_name} on the same generator path so both
apply identical YAML lowering (ToolResolver, tool_timeout, approval,
guardrails, retry) instead of the safety-stripped hand-rolled PraisonAgent.
Co-authored-by: Mervin Praison <[email protected]>
Add StreamEventType.RETRY and emit it alongside the existing ON_RETRY hook in both the sync and async retry loops (guarded by has_callbacks, so it stays zero-overhead when no consumer is attached). The event carries attempt, max_attempts, delay and reason in metadata.
The CLI stream-event bridge maps it to a canonical run.retry NDJSON event so consumers can render a live 'retrying in Ns (attempt k/N)' status instead of a silent, hung-looking terminal. Non-JSON/silent modes stay silent.
Co-authored-by: MervinPraison <[email protected]>
When the LLM emits an unresolvable tool call, the runtime now:
- attempts a deterministic self-repair by normalising case/separator drift
(e.g. 'WebSearch' -> 'web_search') against the agent's active tools;
- otherwise returns a corrective, model-readable error naming the nearest
valid tool and listing the currently available tools;
- on argument-bind failure, echoes the tool's required/optional parameters.
Reuses existing dispatch path; no new public API or Agent params.
Co-authored-by: Mervin Praison <[email protected]>
Add a pure-stdlib core helper neutralize_untrusted_text() in
praisonaiagents.session.context and apply it by default to the sender
display name in the bot's _attribute() prefix. Collapsing newlines/control
chars and bounding length stops a hostile group display name / title from
masquerading as a fake heading or system directive in the per-turn prompt,
while well-behaved names render byte-identically. Safe-by-default; no new
config knobs.
Co-authored-by: Mervin Praison <[email protected]>
A bare free-text prompt (`praisonai "do something"`) now forwards to the
same modern `run` path as `praisonai run`, inheriting session continuity,
--output modes, the credential gate and permissions. `.yaml` workflows and
flag-bearing legacy invocations remain on the legacy dispatcher.
Co-authored-by: MervinPraison <[email protected]>
Address reviewer feedback on the in-loop continuation guard:
- Emit the just-completed (paid) row before the guard halts, so the
request that trips the sponsorship check is no longer silently dropped.
- Coerce sponsor_check_rows to a safe minimum of 1 so a 0 (divide-by-zero)
or negative (never-poll, guard disabled) value cannot break/disable the
safety check; unset preserves the 5000 default.
Adds tests for trigger-row emission and invalid-interval handling.
🤖 Generated with [Claude Code](https://claude.com/claude-code)
Co-authored-by: Mervin Praison <[email protected]>
Two additive, backward-compatible features for the synthetic-data generator.
1. In-generator continuation guard. generate_dataset gains an optional
should_continue() -> bool callback, re-checked every `sponsor_check_rows`
completed requests (default 5000). When it returns False the run halts
cleanly via the existing shutdown path (touch stop-file + cancel pending
futures + break). If no callback is passed but `subscription_id` (or env
AZ_SUBSCRIPTION) is set, a built-in Azure sponsorship check is wired in
(quotaId == "Sponsored_2016-01-01"). The guard fails OPEN on transient
errors — an external watchdog is the hard fail-closed guard; this is
belt-and-suspenders. New public helper: azure_sponsorship_guard().
2. Per-row provenance stamping. Emitted rows are now self-describing: _call
stamps `model` (deployment) and passes through `task_type`/`topic` from the
prompt spec when present. The axis recipe stamps task_type/topic onto each
spec, so QC (which already reads r["task_type"]) and downstream category
splits key off metadata instead of brittle text. Keys are added only when a
value is available (backward-compatible).
Tests (network stubbed): the guard halts after N rows when the predicate turns
False and never halts while True; rows carry model/task_type/topic when
provided and omit task_type/topic when absent; plus fail-open and
subscription-id wiring.
Co-Authored-By: Claude Opus 4.8 <[email protected]>
The switch to praisonai.arun() dropped the caller's required input on the
native jobs path (job.prompt) and the serve /agents fallback (query), since
arun has no prompt param and the workflow topic is read from YAML input/topic
only. Honour input/topic overrides in AgentsGenerator._merge_cli_config
(reusing the existing merge seam, no new params/exports) and pass
job.prompt / query through cli_config so per-run input drives the workflow
instead of being silently discarded. Static YAML input is preserved when no
prompt is supplied.
Co-authored-by: Mervin Praison <[email protected]>
- serve.py: route the crew fallback through praisonai.arun (owns the
AgentsGenerator lifecycle, runs off the event loop) and replace
asyncio.get_event_loop() with asyncio.to_thread in both agent handlers.
- jobs/executor.py: _run_praisonai_agents now delegates to praisonai.arun
so the caller's agent_file/agent_yaml/framework/config are honoured
instead of silently dropped by a stock Agent.
- scheduler/agent_scheduler.py: reuse an instance-owned ThreadPoolExecutor
with leak accounting and recycling (mirrors AgentsGenerator), unify the
timeout branch into the backoff path, and close() the pool on stop().
Co-authored-by: MervinPraison <[email protected]>
- reuse a supplied empty shared MinHashLSH index (is not None, not `or`)
so cross-batch near-dedup works on the first call
- treat empty MinHash signatures as non-similar (0.0) and reject
minhash_perm/ngram_n < 1, so degenerate config can't collapse a dataset
- global_dedup honours near_dup_method='sliding' and raises a clear
FileNotFoundError for missing paths instead of a cryptic AttributeError
- sliding look-back treats window<=0 as intentional unbounded
- dedup CLI writes to a temp file + atomic replace, so an --out that
aliases an input is never truncated and a malformed line never leaves
a partial output
Co-authored-by: Mervin Praison <[email protected]>
The QC near-dup pass was O(n·3000) sliding-window char-4gram Jaccard, which is
both slow at scale and structurally blind to near-duplicates more than 3000 rows
apart — and score() rebuilds its dedup state per call, so it cannot remove dups
that span multiple batch files (the common case when merging parallel runs).
Add praisonai_train.data.dedup (pure stdlib, no datasketch/numpy):
- MinHasher / MinHashLSH: seeded, deterministic MinHash signatures over the same
normalized char-4gram shingles the sliding engine uses (so near_dup_jaccard
keeps identical semantics), with LSH banding for sub-linear candidate lookup
and no bounded look-back window.
- near_dedup(rows, cfg): one interface, YAML-tunable near_dup_method
("sliding" default = unchanged, or "minhash"). qc.score() now delegates to it
(DRY); default path behaviour is byte-for-byte the same. near_dup_window is now
configurable too.
- global_dedup(sources, cfg): dedup ACROSS many JSONL files / row-lists with ONE
shared exact+LSH index — what per-file score() cannot do.
- CLI: `praisonai-train dedup batch_*.jsonl --out merged.jsonl`.
Backward compatible: no new deps, default engine and outputs unchanged; MinHash
and cross-file dedup are opt-in.
Tests: 13 new (deterministic) — MinHash catches a far near-dup the window misses,
preserves genuinely-distinct rows, cross-file exact+near removal, exact-only, and
default-still-sliding. Full data suite 60 passed.
Co-Authored-By: Claude Opus 4.8 <[email protected]>
- Count completion tokens only for successful (ok) responses so token
throughput and avg_out_tok stay consistent with the success count.
- Thread a json_mode flag through _timed_call/_measure_target/
benchmark_deployments and expose --json-mode/--no-json-mode + yaml key,
so OpenAI-compatible endpoints without JSON mode aren't ranked as failures.
- Validate prompt dicts (require non-empty 'user') and reject n<1 /
concurrency<1 with a clear ValueError instead of KeyError / silent
zero-request runs miscounted as endpoint failures.
- Add 4 tests; existing stubs updated for the new json_mode arg. 102 passed.
Co-authored-by: Mervin Praison <[email protected]>
Add a first-class, protocol-driven speed benchmark to the praisonai-train
data SDK, mirroring the generate/validate idiom. It sends N identical
requests per deployment at a fixed concurrency and reports throughput
(rows/min, completion-tokens/sec), latency (avg/p50/p95), success count
and average output tokens, then ranks deployments fastest-first.
DRY: extract the generator's HTTP call path into shared helpers
(`build_chat_request`, `resolve_cfg`) so benchmarking reuses the exact
same Azure/OpenAI-compatible request construction and credential
resolution as real generation — no duplicated HTTP code. The default
prompt reuses a registered Recipe so benchmarks reflect the real workload.
- data/benchmark.py: `benchmark_deployments(...) -> list[BenchResult]`,
pure `aggregate()`, and network confined to a `_timed_call` seam.
- cli/commands/benchmark.py: `praisonai-train benchmark` (YAML/flags),
ranked table + JSON output.
- setup/benchmark.yaml: documented sample config.
- tests: aggregation math, ranking, success/failure counting, real
threaded path, per-target endpoint override (network stubbed).
Co-Authored-By: Claude Opus 4.8 <[email protected]>
The documented gateway.notify_on_undelivered / undelivered_template YAML
opt-in was a dead path: GatewayServerSchema (extra="forbid") rejected the
keys at load time, and neither start_with_config nor from_config_file
propagated them onto self.config, so the delivery router's getattr always
saw the OFF default.
- Accept both keys in GatewayServerSchema (kept optional, lightweight).
- Stamp them onto self.config in start_with_config (primary CLI path) and
from_config_file, honouring the "no new core GatewayConfig field" design.
- Add a regression test covering schema acceptance + propagation.
Co-authored-by: Mervin Praison <[email protected]>
Connect the previously-inert completion path (Greptile P1): BotSessionManager
now reads the agent's last_stop_reason after each turn and, when the new opt-in
surface_completion_reason flag is set, appends a concise user-safe note for
early stops (max_steps / cancelled / error). Off by default so clean
completions and existing deployments are unchanged.
Co-authored-by: Mervin Praison <[email protected]>
Address greptile P1 review feedback on the inbound media-album coalescer:
when the owning Telegram update is cancelled (e.g. during shutdown) after
sibling updates already folded their media in, the buffered temp files could
leak. Add an on_orphan cleanup hook invoked whenever a flushed album has no
live owner awaiting it, wired to a shared _remove_inbound_media helper in the
Telegram bot. Also clarify the max_items cap semantics (bounded, intentional).
Co-authored-by: Mervin Praison <[email protected]>
On a confirmed-permanent outbound failure the reply was parked in the DLQ /
the target marked dead but the user heard nothing. Add an opt-in last-resort
notice: a short plain-text delivery-failure message on the same channel plus a
MESSAGE_UNDELIVERED lifecycle hook so operators can route the failure.
Core (protocol only): add HookEvent.MESSAGE_UNDELIVERED and
MessageUndeliveredInput. Wrapper (praisonai-bot): DeliveryRouter fires the hook
and best-effort sends the notice on permanent failure; default OFF, wired from
gateway.notify_on_undelivered.
Co-authored-by: MervinPraison <[email protected]>
Telegram delivers a media album (multiple photos/files sent together) as N
separate updates sharing one media_group_id, so the agent saw each part as
its own turn. Add an AlbumCoalescer that buffers media parts by group key
and flushes them as one multimodal turn, mirroring the existing
InboundDebouncer text pattern.
- New praisonai_bot/bots/_album.py: bounded per-group buffer (window +
max_items), operator-tunable via config.metadata (media_group_window_ms /
media_group_max), disabled by default so behaviour is unchanged.
- Wire into TelegramBot.handle_message (sibling updates fold into one turn)
and flush pending buffers on stop().
- Tests in tests/unit/test_bot_album.py.
Co-authored-by: MervinPraison <[email protected]>
Add an optional, backward-compatible TurnCompletion contract to the bot
reply path so a chat user learns *why* a turn stopped early instead of
seeing a silently truncated/empty reply.
- TurnCompletion wraps the existing coarse Agent.last_stop_reason
(completed | max_steps | cancelled | error) into a portable value with
a user-safe, localisable .note(); completed turns surface nothing.
- AgentReply gains an optional `completion` field (defaults None).
- extract_completion() reads it off any result (incl. a bare Agent via
last_stop_reason); append_completion_note() lets a gateway optionally
append the note, off by default so existing deployments are unchanged.
Reuses the existing runtime stop reason rather than adding new detection
or Agent params — no new module, no wrapper knobs.
Co-authored-by: MervinPraison <[email protected]>
Move progress_callback below the yield in generate_dataset so 'kept' reflects
rows the consumer has actually received (Greptile P1). Docstring clarified.
Co-authored-by: Mervin Praison <[email protected]>
generate_dataset() gains an optional progress_callback(done, total, kept)
invoked as teacher requests complete, so a CLI/monitor/notebook can show
completion and remaining work. Backward compatible: default None keeps the
original behaviour (no callback). Fires once up-front with (0, total, 0) then
after each completed request; done is monotonic up to total, kept counts unique
rows emitted.
The `generate` CLI command wires a default reporter (_Progress): a tqdm bar when
tqdm is importable, else throttled done/total prints. tqdm stays optional — the
import is guarded and never becomes a hard dependency.
Adds unit tests asserting the callback fires with monotonic done up to total,
that kept tracks dedup, and that the default (no callback) path is unchanged.
Co-Authored-By: Claude Opus 4.8 <[email protected]>
- intent.translation_intent: an unrecognised non-empty task_type no longer
short-circuits to non-translation; it falls through to the documented text
fallback so an explicit "translate into English" instruction is honoured and
its English output is not dropped as low_script_purity (greptile P1).
- qc.score: the translation `exempted` metric now counts only rows the
English-target exemption ACTUALLY rescued from LowScriptPurity (task-awareness
on AND output below the purity floor), so it reconciles with check behaviour
instead of over-counting failed translations and task_aware_purity=false runs
(greptile P1).
- Add 4 regression tests (29 passed total).
Co-authored-by: Mervin Praison <[email protected]>
Add two dataset-QC features to the praisonai-train data SDK, wired in the
SDK's protocol-driven idiom (registered RowChecks + a corpus metric) and
sharing ONE translation-intent detector so they can never disagree (DRY).
1. Task-aware script purity: the script-purity checks no longer delete an
output that is legitimately in another script because the instruction
asked to translate INTO that language (e.g. a Tamil->English row's
correct output is English). LowScriptPurity/ScriptReview now exempt a
resolved English-target row; a new WrongTargetScript flag catches the
inverse failure (asked for English, output still Tamil). Only a resolved
English target exempts — en->ta / ambiguous / native rows stay strict.
2. Translation-share corpus metric: metrics.translation reports the fraction
of kept rows that are translation tasks, by direction (en_ta/ta_en/unknown)
and an exempted count, computed with the SAME detector so it reconciles
with the per-row exemptions.
New single source of truth: data/intent.translation_intent(); metadata
(task_type) is authoritative, instruction-text detection is the external-row
fallback, ambiguous direction yields no exemption (safe default).
17 new unit tests (mirroring the modeltrain reference); full data suite green.
Co-Authored-By: Claude Opus 4.8 <[email protected]>
Addresses Greptile P1: an issue such as '[QUESTION] How do I handle this
API error?' matched 'error' before the question marker was checked,
receiving 'bug' + auto 'claude'. Guard bug detection on \!title.includes('[question]').
Co-authored-by: Mervin Praison <[email protected]>
Bug issues with a "## How to fix" section were labelled question first,
which blocked autoClaude. Check bug/error/crash before how-to heuristics,
ignore "how to fix" for question detection, and let bug label win over
a stale question label for auto-triage.
- qc: include output in exact-dedup key so distinct answers to the same
prompt are not silently dropped
- _util.fields: pair the final assistant turn with the user turn that
precedes it (never a trailing unanswered user turn)
- cli generate: forward --snapshot-every override; defer output truncation
until the first row is produced (a failed run no longer destroys an
existing file / self-referential dedup source); exit non-zero when 0 rows
are generated instead of reporting empty success
- cli validate: skip malformed JSONL lines (with line numbers) instead of
aborting the whole run
- generate: configurable request_timeout; accurate stop-file docstring
- tests: cover distinct-answer dedup and multi-turn chat pairing
Co-authored-by: Mervin Praison <[email protected]>
Remove the legacy `.praison/config.toml` from PROJECT_CONFIG_NAMES so the
walk-up discovery never mislabels a profile-level legacy global config as a
`project:` source. The legacy file is owned exclusively by
_load_global_config() with a `global:` label. Removes the now-dead
search_dir == home skip guard and adds a regression test.
Co-authored-by: Mervin Praison <[email protected]>
Harbor 0.20's BaseAgent.__init__ requires logs_dir as first positional
arg. The TestPraisonAIExternalAgent tests constructed the agent with zero
args, causing a TypeError. Pass logs_dir via pytest tmp_path to match the
existing pattern used by TestPraisonAICodeAgent and TestPraisonAIInstalledAgent.
Co-authored-by: Mervin Praison <[email protected]>
Harbor 0.20 agents expose extra_env (wired into the container exec
context by Trial), not self.env. Passing env=self.env crashed
PraisonAICodeAgent.run() with AttributeError. Omit the per-exec env=
so Harbor's trial-scoped injection applies. Same fix applied to
praisonai_installed_agent.py.
Co-authored-by: Mervin Praison <[email protected]>
Harbor's BaseInstalledAgent._exec always prefixes commands with
'set -o pipefail; ', so the .startswith assertion failed even when the
adapter command was correct. Switch to a substring match.
Co-authored-by: Mervin Praison <[email protected]>
- PRAISON_DEBUG now parsed as a boolean (0/false/no disable it, not just unset)
- auto-resume filters to real numeric checkpoint-* dirs and passes the validated
path to the trainer (instead of leaving resume=True and re-discovering); also
accepts 1/yes/on/off spellings before treating a string as a literal path
- push_model_gguf honors quantization_method default (q4_k_m) matching save path
- final save/eval cadence reconciliation after training_arguments passthrough so
load_best_model_at_end can't reach SFTConfig with mismatched strategies/steps
Co-authored-by: Mervin Praison <[email protected]>
Reverts the removal of tokenize_function — it is imported and re-exported in
praisonai/train.py __all__, so deleting it breaks that public API. Kept it (now
uncalled by the training path but retained for backward compat) and removed only
the internal, unreferenced tokenize_dataset method.
Co-Authored-By: Claude Fable 5 <[email protected]>
- Remove dead tokenize_function/tokenize_dataset (unused since TRL tokenizes
internally; used padding='max_length' which would waste compute if ever rewired).
- save_model_gguf now honors config quantization_method (was hardcoded 'q4_k_m',
inconsistent with push_model_gguf).
- Gate the per-batch DEBUG prints behind PRAISON_DEBUG so normal runs stay readable.
Co-Authored-By: Claude Fable 5 <[email protected]>
HF requires save_steps to be a multiple of eval_steps when load_best_model_at_end
is on. A config with save_steps=100, eval_steps=500 (both valid-looking) crashed.
Now the trainer aligns eval_steps to save_steps (preserving the user's checkpoint
frequency) with a warning, instead of erroring out.
Co-Authored-By: Claude Fable 5 <[email protected]>
resume_from_checkpoint: true now auto-finds the latest checkpoint in output_dir
and resumes from it; if none exists yet (fresh run) it starts fresh instead of
erroring. So after any crash (OOM, GPU reclaim) you just relaunch the SAME
command and training continues from the last saved step. Validated on GPU:
fresh run -> checkpoint-2/4; relaunch -> resumed from checkpoint-4, continued 5->8.
Co-Authored-By: Claude Fable 5 <[email protected]>
- Remove wrapper import from praisonai-code agent_tools.py (C7 gate
regression): use a module-local _run_sync bridge instead of
`from praisonai._async_bridge import run_sync` so the Tier-2 package
no longer depends on the praisonai wrapper at the hot path.
- suite_runner: start child processes with start_new_session=True on
POSIX so timeout killpg reaps only the child tree, never the runner.
- agent_tools/orchestrator: honor acp_execute_command cwd by forwarding
it into the plan step; orchestrator validates it stays in workspace.
🤖 Generated with [Claude Code](https://claude.ai/code)
Co-authored-by: Mervin Praison <[email protected]>
Gemma 4 E4B (and other multimodal/MoE/elastic models) have LoRA-adapted params
that don't fire every step (vision/audio towers in text-only training), which
crashes DDP with 'Expected to have finished reduction' when find_unused=False.
Default True works universally; set false for a small speedup on dense-text models.
Co-Authored-By: Claude Fable 5 <[email protected]>
- Set UNSLOTH_USE_NEW_MODEL before unsloth import (DDP non-reentrant path)
- Skip LoRA when full_finetuning: true (train base weights directly)
- Reload model with trained precision (dtype/8bit/full_finetuning), not 4bit-only
- Re-align save/eval strategy when early stopping enables load_best_model_at_end
🤖 Generated with [Claude Code](https://claude.ai/code)
Co-authored-by: Mervin Praison <[email protected]>
Only the main process merges/pushes and writes the final adapter, so multi-GPU
runs don't race to write the same repo/dir.
Co-Authored-By: Claude Fable 5 <[email protected]>
Validated end-to-end on 2x A6000 (unsloth 2026.7.4 / trl 0.24 / torch 2.10):
multi-GPU DDP trains (Data Parallel GPUs = 2, no 'marked ready twice'),
checkpointing + eval + best-model + final_model_dir all work.
Multi-GPU (DDP):
- Detect distributed (LOCAL_RANK/WORLD_SIZE); load full model per-rank
(device_map={'':local_rank}); set UNSLOTH_USE_NEW_MODEL=1 under DDP to force
Unsloth's non-reentrant checkpointing (the legacy path hard-codes reentrant,
which double-fires autograd hooks and crashes DDP). Set ddp_find_unused_parameters
=False (LoRA freezes the base). Launch: torchrun --nproc_per_node=N.
Checkpointing / resume:
- Config-driven save_strategy/save_steps/save_total_limit (was only under an env var);
resume_from_checkpoint (bool=latest, or a path); final_model_dir (fixes the old
lora_model vs output_dir mismatch that broke load_model).
Evaluation / early stopping:
- val_split_ratio carves an eval set; eval_strategy/eval_steps/load_best_model_at_end/
metric_for_best_model; EarlyStoppingCallback via early_stopping_patience.
Flexibility:
- Generic training_arguments passthrough (any raw SFTConfig field), auto-filtered to
what the installed TRL accepts (version-proof; e.g. save_safetensors/group_by_length
differ by version) — warns instead of crashing.
- Knobs: max_grad_norm, warmup_ratio, neftune_noise_alpha, dataloader_num_workers,
data_seed, adam betas, report_to/run_name (config, not just env), push_to_hub.
- Quantization: load_in_8bit, dtype, full_finetuning. LoRA: modules_to_save,
rank_pattern, alpha_pattern, use_dora. Data: local files, shuffle, rename, filter.
Correctness fixes:
- load_in_4bit KeyError when omitted; inference().to(cuda) -> self.device;
drop empty-formatted rows instead of silently training on blanks.
All new keys optional with prior defaults; existing configs unchanged.
Co-Authored-By: Claude Fable 5 <[email protected]>
2026-07-22 06:12:38 +01:00
376 changed files with 15585 additions and 1689 deletions
Loaded 100 of 376 files, more files were not shown because too many files have changed in this diff.
Show more
Reference in new issue
Block a user
Blocking a user prevents them from interacting with repositories, such as opening or commenting on pull requests or issues. Learn more about blocking a user.
