fix(agents): honor safe_mode and context.global.unsafe for LLM-agent tool calls #129

Merged

CoreRasurae merged 1 commit from bugfix/m1-llm-tool-unsafe-mode into master 2026-08-20 17:32:34 +00:00

Conversation 6 Commits 1 Files changed 13 +902 -25

CoreRasurae commented 2026-08-09 23:29:13 +00:00

Member

Copy link

Summary

Fixes two compounding bugs that unconditionally blocked every file_write (and other tool) call issued by a type: llm agent's tool-call loop, regardless of actor configuration:

1. LLMAgent._execute_tool_loop() (and its two synthesis-retry mirrors) derived parent_unsafe from a dead unsafe_mode config key that nothing in the codebase ever wrote, instead of the actual safe_mode field an actor author sets (docs/index.md §4.5). Fixed via a new LLMAgent._resolve_parent_unsafe() helper, and by threading the invocation context through _execute_tool_loop() so it can also see a host-propagated _unsafe_mode: true.
2. cleveractors.runtime_dispatch._execute_graph/_execute_graph_stream (the Executor/create_executor() path) had no host-unsafe-flag concept at all, so an actor declaring context.global.unsafe: true (§9.4) was neither refused on an unsafe host nor granted _unsafe_mode propagation when the host was unsafe. Fixed by adding Executor(unsafe=...)/create_executor(unsafe=...) and a new runtime_dispatch._enforce_unsafe_flag() helper mirroring Application._enforce_unsafe_flag.

Default safe_mode: true sandboxing for file_read/shell/directory traversal is unaffected.

Test plan

• features/llm_agent_tool_loop.feature — @tdd_issue_115 scenario now passes without @tdd_expected_fail (LLM agent safe_mode: false path)
• features/runtime_unsafe_host_mode.feature (new) — context.global.unsafe host-mode path: refusal, propagation, and no-regression scenarios, exercised end-to-end through create_executor() → real graph → real LLMAgent/ToolAgent (only the LangChain network boundary is mocked)
• robot/llm_tool_calling.robot (extended) — same three scenarios via Robot integration test, real file I/O
• nox -s unit_tests — 3021/3021 scenarios passing
• nox -s coverage_report — 96.6% (threshold 96.5%)
• nox -s integration_tests — 355/355 passing
• nox -s lint, nox -s format -- --check, nox -s typecheck, nox -s security_scan, nox -s dead_code — all clean

Closes #115

## Summary Fixes two compounding bugs that unconditionally blocked every `file_write` (and other tool) call issued by a `type: llm` agent's tool-call loop, regardless of actor configuration: 1. `LLMAgent._execute_tool_loop()` (and its two synthesis-retry mirrors) derived `parent_unsafe` from a dead `unsafe_mode` config key that nothing in the codebase ever wrote, instead of the actual `safe_mode` field an actor author sets (docs/index.md §4.5). Fixed via a new `LLMAgent._resolve_parent_unsafe()` helper, and by threading the invocation `context` through `_execute_tool_loop()` so it can also see a host-propagated `_unsafe_mode: true`. 2. `cleveractors.runtime_dispatch._execute_graph`/`_execute_graph_stream` (the `Executor`/`create_executor()` path) had no host-unsafe-flag concept at all, so an actor declaring `context.global.unsafe: true` (§9.4) was neither refused on an unsafe host nor granted `_unsafe_mode` propagation when the host was unsafe. Fixed by adding `Executor(unsafe=...)`/`create_executor(unsafe=...)` and a new `runtime_dispatch._enforce_unsafe_flag()` helper mirroring `Application._enforce_unsafe_flag`. Default `safe_mode: true` sandboxing for `file_read`/`shell`/directory traversal is unaffected. ## Test plan - [x] `features/llm_agent_tool_loop.feature` — `@tdd_issue_115` scenario now passes without `@tdd_expected_fail` (LLM agent `safe_mode: false` path) - [x] `features/runtime_unsafe_host_mode.feature` (new) — `context.global.unsafe` host-mode path: refusal, propagation, and no-regression scenarios, exercised end-to-end through `create_executor()` → real graph → real `LLMAgent`/`ToolAgent` (only the LangChain network boundary is mocked) - [x] `robot/llm_tool_calling.robot` (extended) — same three scenarios via Robot integration test, real file I/O - [x] `nox -s unit_tests` — 3021/3021 scenarios passing - [x] `nox -s coverage_report` — 96.6% (threshold 96.5%) - [x] `nox -s integration_tests` — 355/355 passing - [x] `nox -s lint`, `nox -s format -- --check`, `nox -s typecheck`, `nox -s security_scan`, `nox -s dead_code` — all clean Closes #115

CoreRasurae added this to the v2.1.0 milestone 2026-08-09 23:29:13 +00:00

CoreRasurae added the

Type

Bug

label 2026-08-09 23:30:08 +00:00

CoreRasurae added a new dependency 2026-08-09 23:30:46 +00:00

#115 LLM agent file_write calls are always blocked — safe_mode: false / context.global.unsafe: true never reach _unsafe_mode

hurui200320 requested changes 2026-08-10 07:12:39 +00:00

Dismissed

hurui200320 left a comment

Copy link

PR Review: !129 (Ticket #115)

Verdict: Request Changes

The PR correctly fixes the two compounding bugs described in #115 and adds focused BDD/Robot coverage for the graph runtime path and the direct LLMAgent tool-call path. The code is well-typed, the new helpers are clearly documented, and the existing scenarios that relied on the dead unsafe_mode key have been updated appropriately.

One major functional gap remains: the new host unsafe flag is not forwarded to sub-Executors in the multi-actor dispatch path, so a multi-actor bundle containing an unsafe-declared graph actor will still be refused on an unsafe host.

Critical Issues

None.

Major Issues

• Host unsafe flag is lost across multi-actor sub-Executors
  • File: src/cleveractors/runtime_dispatch.py (lines 736–743)
  • Problem: _execute_multi_actor creates a child Executor but does not pass unsafe=executor.unsafe (or registry_api_key=executor.registry_api_key). Consequently, a multi-actor bundle whose selected sub-actor is a graph declaring context.global.unsafe: true will raise UnsafeConfigurationError even when the parent host was placed in unsafe mode. This breaks the host-mode contract that the PR is introducing.
  • Recommendation: Forward unsafe=executor.unsafe and registry_api_key=executor.registry_api_key when constructing the sub-Executor, and add a Behave/Robot scenario that verifies unsafe-mode propagation through a multi-actor bundle.

Minor Issues

• Single LLM actor path ignores context.global.unsafe

  • File: src/cleveractors/runtime_dispatch.py (_execute_llm, _execute_llm_stream)
  • Problem: _enforce_unsafe_flag is only wired into _execute_graph and _execute_graph_stream. A single type: llm actor config that declares context.global.unsafe: true is neither refused on a safe host nor granted _unsafe_mode on an unsafe host, leaving a spec-compliance gap outside the graph path.
  • Recommendation: Apply the same enforcement/propagation in _execute_llm and _execute_llm_stream, or explicitly document that context.global.unsafe is only honored by graph actors.

• Streaming graph path lacks unsafe-mode coverage

  • File: features/runtime_unsafe_host_mode.feature, robot/llm_tool_calling.robot
  • Problem: _execute_graph_stream now calls _enforce_unsafe_flag, but no new scenario exercises refusal or propagation through Executor.execute_stream().
  • Recommendation: Add a Behave or Robot scenario for the streaming graph path.

Nits

• Stale docstring in _execute_multi_actor
  • File: src/cleveractors/runtime_dispatch.py (line 703)
  • Problem: The docstring lists credentials, limits, and pricing as propagated, but now unsafe (and registry_api_key) should also be propagated.
  • Recommendation: Update the docstring to mention the newly forwarded host flags.

Summary

The core fix is sound and the tests directly validate the reported file_write blocking bug. The missing unsafe propagation in _execute_multi_actor is the only blocker; once that is fixed and covered by a test, this PR should be ready to approve.

## PR Review: !129 (Ticket #115) ### Verdict: Request Changes The PR correctly fixes the two compounding bugs described in #115 and adds focused BDD/Robot coverage for the graph runtime path and the direct `LLMAgent` tool-call path. The code is well-typed, the new helpers are clearly documented, and the existing scenarios that relied on the dead `unsafe_mode` key have been updated appropriately. One major functional gap remains: the new host `unsafe` flag is not forwarded to sub-Executors in the multi-actor dispatch path, so a multi-actor bundle containing an unsafe-declared graph actor will still be refused on an unsafe host. ### Critical Issues None. ### Major Issues - **Host `unsafe` flag is lost across multi-actor sub-Executors** - **File:** `src/cleveractors/runtime_dispatch.py` (lines 736–743) - **Problem:** `_execute_multi_actor` creates a child `Executor` but does not pass `unsafe=executor.unsafe` (or `registry_api_key=executor.registry_api_key`). Consequently, a multi-actor bundle whose selected sub-actor is a graph declaring `context.global.unsafe: true` will raise `UnsafeConfigurationError` even when the parent host was placed in unsafe mode. This breaks the host-mode contract that the PR is introducing. - **Recommendation:** Forward `unsafe=executor.unsafe` and `registry_api_key=executor.registry_api_key` when constructing the sub-Executor, and add a Behave/Robot scenario that verifies unsafe-mode propagation through a multi-actor bundle. ### Minor Issues - **Single LLM actor path ignores `context.global.unsafe`** - **File:** `src/cleveractors/runtime_dispatch.py` (`_execute_llm`, `_execute_llm_stream`) - **Problem:** `_enforce_unsafe_flag` is only wired into `_execute_graph` and `_execute_graph_stream`. A single `type: llm` actor config that declares `context.global.unsafe: true` is neither refused on a safe host nor granted `_unsafe_mode` on an unsafe host, leaving a spec-compliance gap outside the graph path. - **Recommendation:** Apply the same enforcement/propagation in `_execute_llm` and `_execute_llm_stream`, or explicitly document that `context.global.unsafe` is only honored by graph actors. - **Streaming graph path lacks unsafe-mode coverage** - **File:** `features/runtime_unsafe_host_mode.feature`, `robot/llm_tool_calling.robot` - **Problem:** `_execute_graph_stream` now calls `_enforce_unsafe_flag`, but no new scenario exercises refusal or propagation through `Executor.execute_stream()`. - **Recommendation:** Add a Behave or Robot scenario for the streaming graph path. ### Nits - **Stale docstring in `_execute_multi_actor`** - **File:** `src/cleveractors/runtime_dispatch.py` (line 703) - **Problem:** The docstring lists credentials, limits, and pricing as propagated, but now `unsafe` (and `registry_api_key`) should also be propagated. - **Recommendation:** Update the docstring to mention the newly forwarded host flags. ### Summary The core fix is sound and the tests directly validate the reported `file_write` blocking bug. The missing `unsafe` propagation in `_execute_multi_actor` is the only blocker; once that is fixed and covered by a test, this PR should be ready to approve.

CoreRasurae added 1 commit 2026-08-10 15:04:55 +00:00

fix(agents): honor safe_mode and context.global.unsafe for LLM-agent tool calls ... 5e27912e59

LLMAgent's tool-call dispatch (_execute_tool_loop and its two synthesis-retry
mirrors) derived parent_unsafe from a dead "unsafe_mode" config key that
nothing in the codebase ever wrote, so file_write (and every other tool)
called by a type: llm agent's tool loop was unconditionally blocked
regardless of the actor's safe_mode setting. Separately, the
Executor/create_executor runtime path (runtime_dispatch._execute_graph /
_execute_graph_stream) had no host-unsafe-flag concept at all, so an actor
declaring context.global.unsafe: true was neither refused on an unsafe host
nor granted _unsafe_mode propagation when the host was actually unsafe.

Added LLMAgent._resolve_parent_unsafe() to derive unsafe status from the
agent's own safe_mode field or from an _unsafe_mode already present in the
invocation context, and threaded that context into _execute_tool_loop() so
all three dispatch sites can see it. Added Executor(unsafe=...) /
create_executor(unsafe=...) and runtime_dispatch._enforce_unsafe_flag(),
mirroring Application._enforce_unsafe_flag, so context.global.unsafe is
checked against the host flag and _unsafe_mode is propagated into every
tool-call context reachable from the graph, including an LLM agent's
internally dispatched tool calls.

Three existing scenarios in llm_agent_tool_calling.feature relied on
directly setting the dead "unsafe_mode" key to unblock a mocked file_read
call; updated them to set safe_mode: false instead.

Extended per PR review (rui.hu): _enforce_unsafe_flag() now also runs in
_execute_llm/_execute_llm_stream, so a top-level type: llm actor declaring
context.global.unsafe: true is enforced and propagated the same way a
graph actor is, not only an LLM agent embedded as a graph node.
_execute_multi_actor now forwards the parent Executor's unsafe flag (and
registry_api_key) to the sub-Executor it constructs for the bundle's
selected actor, closing a gap where a multi-actor bundle wrapping an
unsafe-declaring graph actor was always refused regardless of the host's
actual unsafe state. Added Behave coverage for both gaps plus the
previously-untested streaming graph path, and two Robot scenarios
exercising execute_stream() for the unsafe-host-mode graph actor.

ISSUES CLOSED: #115

CoreRasurae force-pushed bugfix/m1-llm-tool-unsafe-mode from ac50b16b09 to 5e27912e59 2026-08-10 15:04:55 +00:00 Compare

CoreRasurae commented 2026-08-10 15:10:04 +00:00

Author

Member

Copy link

Thanks for the review — all four findings check out against the spec (docs/index.md §9.4/§10.3) and the code, and have been addressed in the amended commit (ac50b16 → 5e27912, force-pushed to this branch).

Critical Issues

None — no action needed.

Major Issues

Host unsafe flag lost across multi-actor sub-Executors — Confirmed, fixed. _execute_multi_actor (src/cleveractors/runtime_dispatch.py) now passes unsafe=executor.unsafe and registry_api_key=executor.registry_api_key when constructing the sub-Executor. Checked against Executor.__init__, which defaults both to False/None — so the sub-Executor previously always ran as if the host were safe, regardless of the parent's actual state, exactly as you described. Added two Behave scenarios in features/runtime_unsafe_host_mode.feature (propagation success + refusal) driving a multi-actor bundle whose selected sub-actor is an unsafe-declaring graph actor.

Minor Issues

Single LLM actor path ignores context.global.unsafe — Confirmed, fixed. _execute_llm and _execute_llm_stream now build global_context from the actor's context.global section and call _enforce_unsafe_flag(), mirroring _execute_graph/_execute_graph_stream exactly. When the host is unsafe and the actor declares it, _unsafe_mode: true is threaded into the llm_context passed to process_message()/stream_message(), where LLMAgent._resolve_parent_unsafe() already knew how to read it. Added 4 Behave scenarios covering refusal and propagation for both execute() and execute_stream() on a top-level type: llm actor.

Streaming graph path lacks unsafe-mode coverage — Confirmed as a test-coverage gap only, not a code bug: _execute_graph_stream already called _enforce_unsafe_flag() before this review. Added 2 Behave scenarios (refusal + propagation via execute_stream()) plus 2 Robot scenarios in robot/llm_tool_calling.robot ("...Via Execute Stream" variants), backed by a new execute_unsafe_host_message_stream keyword in ToolCallingTestLib.py that reuses the existing fixture setup.

Nits

Stale docstring in _execute_multi_actor — Fixed; the docstring now lists registry_api_key and the host unsafe flag among the fields forwarded to the sub-Executor.

Not changed

Nothing from this review was left unaddressed — all four findings map to a fix or test addition above.

Verification run locally (all via nox)

• lint, format -- --check, typecheck, security_scan, dead_code — clean
• unit_tests — 3033/3033 Behave scenarios pass (12 new, 0 regressions)
• coverage_report — 96.6% (≥96.5% threshold; unchanged from the PR description baseline)
• integration_tests — 11/11 Robot scenarios in llm_tool_calling.robot pass, including the 2 new streaming ones

The commit was amended in place rather than a new one added, per this repo's one-issue/one-commit convention.

Separate observation, unrelated to this review and left untouched: the PR currently shows mergeable: false against current master, due to an unrelated CHANGELOG.md conflict with other PRs merged since this branch's base commit. Flagging for visibility only.

Thanks for the review — all four findings check out against the spec (`docs/index.md` §9.4/§10.3) and the code, and have been addressed in the amended commit (`ac50b16` → `5e27912`, force-pushed to this branch). ### Critical Issues None — no action needed. ### Major Issues **Host `unsafe` flag lost across multi-actor sub-Executors** — Confirmed, fixed. `_execute_multi_actor` (`src/cleveractors/runtime_dispatch.py`) now passes `unsafe=executor.unsafe` and `registry_api_key=executor.registry_api_key` when constructing the sub-`Executor`. Checked against `Executor.__init__`, which defaults both to `False`/`None` — so the sub-Executor previously always ran as if the host were safe, regardless of the parent's actual state, exactly as you described. Added two Behave scenarios in `features/runtime_unsafe_host_mode.feature` (propagation success + refusal) driving a multi-actor bundle whose selected sub-actor is an unsafe-declaring graph actor. ### Minor Issues **Single LLM actor path ignores `context.global.unsafe`** — Confirmed, fixed. `_execute_llm` and `_execute_llm_stream` now build `global_context` from the actor's `context.global` section and call `_enforce_unsafe_flag()`, mirroring `_execute_graph`/`_execute_graph_stream` exactly. When the host is unsafe and the actor declares it, `_unsafe_mode: true` is threaded into the `llm_context` passed to `process_message()`/`stream_message()`, where `LLMAgent._resolve_parent_unsafe()` already knew how to read it. Added 4 Behave scenarios covering refusal and propagation for both `execute()` and `execute_stream()` on a top-level `type: llm` actor. **Streaming graph path lacks unsafe-mode coverage** — Confirmed as a test-coverage gap only, not a code bug: `_execute_graph_stream` already called `_enforce_unsafe_flag()` before this review. Added 2 Behave scenarios (refusal + propagation via `execute_stream()`) plus 2 Robot scenarios in `robot/llm_tool_calling.robot` ("...Via Execute Stream" variants), backed by a new `execute_unsafe_host_message_stream` keyword in `ToolCallingTestLib.py` that reuses the existing fixture setup. ### Nits **Stale docstring in `_execute_multi_actor`** — Fixed; the docstring now lists `registry_api_key` and the host `unsafe` flag among the fields forwarded to the sub-Executor. ### Not changed Nothing from this review was left unaddressed — all four findings map to a fix or test addition above. ### Verification run locally (all via `nox`) - `lint`, `format -- --check`, `typecheck`, `security_scan`, `dead_code` — clean - `unit_tests` — 3033/3033 Behave scenarios pass (12 new, 0 regressions) - `coverage_report` — 96.6% (≥96.5% threshold; unchanged from the PR description baseline) - `integration_tests` — 11/11 Robot scenarios in `llm_tool_calling.robot` pass, including the 2 new streaming ones The commit was amended in place rather than a new one added, per this repo's one-issue/one-commit convention. Separate observation, unrelated to this review and left untouched: the PR currently shows `mergeable: false` against current `master`, due to an unrelated `CHANGELOG.md` conflict with other PRs merged since this branch's base commit. Flagging for visibility only.

hurui200320 requested changes 2026-08-10 15:30:29 +00:00

Dismissed

hurui200320 left a comment

Copy link

PR Review: !129 (Ticket #115)

Verdict: Request Changes

The issue #115 fixes are implemented correctly: LLMAgent._resolve_parent_unsafe() now reads safe_mode instead of the dead unsafe_mode key, the invocation context is threaded through _execute_tool_loop() and its synthesis mirrors, and Executor / create_executor() enforce context.global.unsafe across graph, single-LLM, multi-actor, and streaming paths. The four findings from the previous review cycle have all been addressed (multi-actor forwarding of unsafe/registry_api_key, single-LLM enforcement, streaming coverage, and the stale _execute_multi_actor docstring).

However, the branch is currently stale against master and the diff reverts unrelated recently-merged functionality. Merging as-is would regress agent/route package-reference resolution and delete associated docs. This needs a rebase / merge-conflict cleanup before approval.

Critical Issues

• Stale branch reverts unrelated master changes
  • Files: src/cleveractors/runtime_dispatch.py, README.md, docs/index.md, docs/guides/reasoning-aware-llm-agents.md, docs/adr/ADR-2037-agent-and-route-package-reference-resolution.md, CHANGELOG.md
  • Problem: The PR is based on an older master. Relative to current master it removes the AgentReferenceResolver / RouteReferenceResolver wiring added by ADR-2037, strips the credentials documentation from README.md and the reasoning-aware guide, reverts docs/index.md from version 1.5.0 to 1.4.0, deletes ADR-2037, and removes the corresponding changelog entries. The author noted a CHANGELOG.md conflict, but the stale state also affects source code and other docs. Merging would break agent/route package references and leave documentation inconsistent.
  • Recommendation: Rebase this branch onto current origin/master and resolve conflicts so that the new unsafe / registry_api_key forwarding is layered on top of the ADR-2037 code. Preserve AgentReferenceResolver, RouteReferenceResolver, the docs version bump, README credentials docs, the ADR file, and the changelog entries for issues #121 / #123.

Major Issues

None.

Minor Issues

None.

Nits

• Outdated docstring in _enforce_unsafe_flag()
  • File: src/cleveractors/runtime_dispatch.py (line 95)
  • Problem: The docstring says it enforces the contract "for a graph actor", but the helper is now also invoked for single type: llm actors in _execute_llm() and _execute_llm_stream(). Update the wording to reflect that it applies to any actor type.
  • Recommendation: Change the opening sentence to "Enforce the unsafe host contract for an actor (§9.4/§10.3)."

Summary

All of the previous review feedback has been resolved in the amended commit. The implementation correctly closes the two compounding bugs from #115 and adds good BDD/Robot coverage for the runtime host-unsafe contract. The only remaining blocker is cleaning up the stale branch state so the unrelated ADR-2037 and documentation changes are not reverted. After a clean rebase and green CI, this should be ready to approve.

## PR Review: !129 (Ticket #115) ### Verdict: Request Changes The issue #115 fixes are implemented correctly: `LLMAgent._resolve_parent_unsafe()` now reads `safe_mode` instead of the dead `unsafe_mode` key, the invocation `context` is threaded through `_execute_tool_loop()` and its synthesis mirrors, and `Executor` / `create_executor()` enforce `context.global.unsafe` across graph, single-LLM, multi-actor, and streaming paths. The four findings from the previous review cycle have all been addressed (multi-actor forwarding of `unsafe`/`registry_api_key`, single-LLM enforcement, streaming coverage, and the stale `_execute_multi_actor` docstring). However, the branch is currently stale against `master` and the diff reverts unrelated recently-merged functionality. Merging as-is would regress agent/route package-reference resolution and delete associated docs. This needs a rebase / merge-conflict cleanup before approval. ### Critical Issues - **Stale branch reverts unrelated `master` changes** - **Files:** `src/cleveractors/runtime_dispatch.py`, `README.md`, `docs/index.md`, `docs/guides/reasoning-aware-llm-agents.md`, `docs/adr/ADR-2037-agent-and-route-package-reference-resolution.md`, `CHANGELOG.md` - **Problem:** The PR is based on an older `master`. Relative to current `master` it removes the `AgentReferenceResolver` / `RouteReferenceResolver` wiring added by ADR-2037, strips the `credentials` documentation from `README.md` and the reasoning-aware guide, reverts `docs/index.md` from version 1.5.0 to 1.4.0, deletes ADR-2037, and removes the corresponding changelog entries. The author noted a `CHANGELOG.md` conflict, but the stale state also affects source code and other docs. Merging would break agent/route package references and leave documentation inconsistent. - **Recommendation:** Rebase this branch onto current `origin/master` and resolve conflicts so that the new `unsafe` / `registry_api_key` forwarding is layered on top of the ADR-2037 code. Preserve `AgentReferenceResolver`, `RouteReferenceResolver`, the docs version bump, README credentials docs, the ADR file, and the changelog entries for issues #121 / #123. ### Major Issues None. ### Minor Issues None. ### Nits - **Outdated docstring in `_enforce_unsafe_flag()`** - **File:** `src/cleveractors/runtime_dispatch.py` (line 95) - **Problem:** The docstring says it enforces the contract "for a graph actor", but the helper is now also invoked for single `type: llm` actors in `_execute_llm()` and `_execute_llm_stream()`. Update the wording to reflect that it applies to any actor type. - **Recommendation:** Change the opening sentence to "Enforce the `unsafe` host contract for an actor (§9.4/§10.3)." ### Summary All of the previous review feedback has been resolved in the amended commit. The implementation correctly closes the two compounding bugs from #115 and adds good BDD/Robot coverage for the runtime host-unsafe contract. The only remaining blocker is cleaning up the stale branch state so the unrelated ADR-2037 and documentation changes are not reverted. After a clean rebase and green CI, this should be ready to approve.

CoreRasurae force-pushed bugfix/m1-llm-tool-unsafe-mode from 5e27912e59 to e0b20a0165 2026-08-15 11:08:15 +00:00 Compare

CoreRasurae force-pushed bugfix/m1-llm-tool-unsafe-mode from e0b20a0165 to 9b72919720 2026-08-15 11:22:36 +00:00 Compare

CoreRasurae commented 2026-08-15 11:22:45 +00:00

Author

Member

Copy link

Thanks — addressed in the amended commit (5e27912 → 9b72919, force-pushed).

Critical Issues

Stale branch reverts unrelated master changes — Resolved by rebasing onto current master. HEAD~1 is now identical to origin/master (single commit ahead), so AgentReferenceResolver/RouteReferenceResolver (ADR-2037), the README credentials docs, and the CHANGELOG entries for #121/#123/#131 are all preserved untouched — verified git diff origin/master -- README.md docs/index.md docs/adr/ is empty.

Nits

Outdated docstring in _enforce_unsafe_flag() — Fixed per your suggested wording: opening line now reads "Enforce the unsafe host contract for an actor (§9.4/§10.3)." Also updated the body/Args text that still said "graph"/"the graph's seed context" for the same reason, since the helper is exercised by the graph, single-LLM, and multi-actor sub-Executor paths alike.

Verification (via nox)

• lint, typecheck — clean
• unit_tests — 3075/3075 scenarios pass (full suite, 0 regressions)
• Scoped re-run of the pertinent feature files (llm_agent_tool_calling.feature, llm_agent_tool_loop.feature, runtime_unsafe_host_mode.feature) — 88/88 scenarios pass

This is a docstring-only change (no executable lines touched), so no coverage delta and no Robot/benchmark re-run were needed.

Thanks — addressed in the amended commit (`5e27912` → `9b72919`, force-pushed). ### Critical Issues **Stale branch reverts unrelated `master` changes** — Resolved by rebasing onto current `master`. `HEAD~1` is now identical to `origin/master` (single commit ahead), so `AgentReferenceResolver`/`RouteReferenceResolver` (ADR-2037), the README credentials docs, and the CHANGELOG entries for #121/#123/#131 are all preserved untouched — verified `git diff origin/master -- README.md docs/index.md docs/adr/` is empty. ### Nits **Outdated docstring in `_enforce_unsafe_flag()`** — Fixed per your suggested wording: opening line now reads "Enforce the `unsafe` host contract for an actor (§9.4/§10.3)." Also updated the body/Args text that still said "graph"/"the graph's seed context" for the same reason, since the helper is exercised by the graph, single-LLM, and multi-actor sub-Executor paths alike. ### Verification (via `nox`) - `lint`, `typecheck` — clean - `unit_tests` — 3075/3075 scenarios pass (full suite, 0 regressions) - Scoped re-run of the pertinent feature files (`llm_agent_tool_calling.feature`, `llm_agent_tool_loop.feature`, `runtime_unsafe_host_mode.feature`) — 88/88 scenarios pass This is a docstring-only change (no executable lines touched), so no coverage delta and no Robot/benchmark re-run were needed.

hurui200320 approved these changes 2026-08-15 11:55:52 +00:00

hurui200320 left a comment

Copy link

PR Review: !129 (Ticket #115)

Verdict: Approve

The implementation correctly fixes the two compounding bugs described in #115: LLMAgent._execute_tool_loop() now derives parent_unsafe from the real safe_mode field and/or a host-propagated _unsafe_mode, and the Executor/create_executor() runtime path enforces the context.global.unsafe host contract across graph, single-LLM, streaming, and multi-actor dispatch. The test suite is comprehensive (Behave + Robot) and the previous review findings have been addressed. No critical or major issues remain.

Critical Issues

None.

Major Issues

None.

Minor Issues

1. src/cleveractors/runtime_dispatch.py — _execute_tool() does not enforce the host unsafe contract

   • The new Executor.unsafe flag and _enforce_unsafe_flag() helper are wired into _execute_graph, _execute_graph_stream, _execute_llm, _execute_llm_stream, and the _execute_multi_actor sub-Executor, but _execute_tool() (single type: tool actor path) is skipped. A tool actor declaring context.global.unsafe: true will therefore not be refused on a safe host, and _unsafe_mode is never injected into its invocation context (the call is agent.process_message(message) with no context).
   • Recommendation: Apply _enforce_unsafe_flag() in _execute_tool() and pass the resulting context to agent.process_message(message, context=...). This is likely out of scope for the LLM-specific #115, but it leaves a spec §9.4 compliance gap on the Executor path.

2. features/runtime_unsafe_host_mode.feature — missing default single-LLM regression scenario

   • The feature has a strong no-regression scenario for graph actors ("Default graph actor with no unsafe declaration keeps file_write blocked"), but there is no equivalent for a single type: llm actor with no context.global.unsafe declaration. The _execute_llm path is distinct from _execute_graph, so the default safe-mode blocking behavior for that shape is not exercised end-to-end here.
   • Recommendation: Add a scenario using _single_llm_config(declare_unsafe=False) and assert the write is blocked.

3. robot/ToolCallingTestLib.py — temporary directories are not cleaned up

   • _create_graph_executor_for_unsafe_host_test() creates target_dir = Path(tempfile.mkdtemp(...)) but never removes it, so each Robot scenario leaks a temp directory.
   • Recommendation: Register a cleanup/teardown that removes self._unsafe_host_target_path.parent after assertions, or use a context manager / tempfile.TemporaryDirectory.

Nits

• noxfile.py — stale coverage threshold docstring
  • The coverage_report() docstring says "Coverage threshold is enforced at >=97.0%", but COVERAGE_THRESHOLD = 96.5. This is a pre-existing inconsistency, not introduced by this PR, but it adds confusion when the issue DoD mentions 97% while the project gate is 96.5%.
  • Recommendation: Align the docstring with the constant (or vice versa).

Summary

This is a solid, focused fix. The root cause analysis in #115 is accurate, the new _resolve_parent_unsafe() helper is the right seam, and threading the invocation context through _execute_tool_loop() cleanly covers all three dispatch sites (main loop + budget/stuck-model synthesis mirrors). The Executor(unsafe=...) design mirrors Application's --unsafe flag consistently, and forwarding both unsafe and registry_api_key through the multi-actor sub-Executor fixes a real privilege-leak bug. Tests are well-structured, assertions validate actual file I/O outcomes, and the previous review cycle's findings have all been addressed. Approved with the minor notes above.

## PR Review: !129 (Ticket #115) ### Verdict: Approve The implementation correctly fixes the two compounding bugs described in #115: `LLMAgent._execute_tool_loop()` now derives `parent_unsafe` from the real `safe_mode` field and/or a host-propagated `_unsafe_mode`, and the `Executor`/`create_executor()` runtime path enforces the `context.global.unsafe` host contract across graph, single-LLM, streaming, and multi-actor dispatch. The test suite is comprehensive (Behave + Robot) and the previous review findings have been addressed. No critical or major issues remain. ### Critical Issues None. ### Major Issues None. ### Minor Issues 1. **`src/cleveractors/runtime_dispatch.py` — `_execute_tool()` does not enforce the host `unsafe` contract** - The new `Executor.unsafe` flag and `_enforce_unsafe_flag()` helper are wired into `_execute_graph`, `_execute_graph_stream`, `_execute_llm`, `_execute_llm_stream`, and the `_execute_multi_actor` sub-Executor, but `_execute_tool()` (single `type: tool` actor path) is skipped. A tool actor declaring `context.global.unsafe: true` will therefore not be refused on a safe host, and `_unsafe_mode` is never injected into its invocation context (the call is `agent.process_message(message)` with no context). - **Recommendation:** Apply `_enforce_unsafe_flag()` in `_execute_tool()` and pass the resulting context to `agent.process_message(message, context=...)`. This is likely out of scope for the LLM-specific #115, but it leaves a spec §9.4 compliance gap on the Executor path. 2. **`features/runtime_unsafe_host_mode.feature` — missing default single-LLM regression scenario** - The feature has a strong no-regression scenario for graph actors ("Default graph actor with no unsafe declaration keeps file_write blocked"), but there is no equivalent for a single `type: llm` actor with no `context.global.unsafe` declaration. The `_execute_llm` path is distinct from `_execute_graph`, so the default safe-mode blocking behavior for that shape is not exercised end-to-end here. - **Recommendation:** Add a scenario using `_single_llm_config(declare_unsafe=False)` and assert the write is blocked. 3. **`robot/ToolCallingTestLib.py` — temporary directories are not cleaned up** - `_create_graph_executor_for_unsafe_host_test()` creates `target_dir = Path(tempfile.mkdtemp(...))` but never removes it, so each Robot scenario leaks a temp directory. - **Recommendation:** Register a cleanup/teardown that removes `self._unsafe_host_target_path.parent` after assertions, or use a context manager / `tempfile.TemporaryDirectory`. ### Nits - **`noxfile.py` — stale coverage threshold docstring** - The `coverage_report()` docstring says "Coverage threshold is enforced at >=97.0%", but `COVERAGE_THRESHOLD = 96.5`. This is a pre-existing inconsistency, not introduced by this PR, but it adds confusion when the issue DoD mentions 97% while the project gate is 96.5%. - **Recommendation:** Align the docstring with the constant (or vice versa). ### Summary This is a solid, focused fix. The root cause analysis in #115 is accurate, the new `_resolve_parent_unsafe()` helper is the right seam, and threading the invocation `context` through `_execute_tool_loop()` cleanly covers all three dispatch sites (main loop + budget/stuck-model synthesis mirrors). The `Executor(unsafe=...)` design mirrors `Application`'s `--unsafe` flag consistently, and forwarding both `unsafe` and `registry_api_key` through the multi-actor sub-Executor fixes a real privilege-leak bug. Tests are well-structured, assertions validate actual file I/O outcomes, and the previous review cycle's findings have all been addressed. Approved with the minor notes above.

CoreRasurae force-pushed bugfix/m1-llm-tool-unsafe-mode from 9b72919720 to 757a607c81 2026-08-20 14:56:01 +00:00 Compare

CoreRasurae force-pushed bugfix/m1-llm-tool-unsafe-mode from 757a607c81 to de12f778df 2026-08-20 17:11:19 +00:00 Compare

CoreRasurae commented 2026-08-20 17:13:28 +00:00

Author

Member

Copy link

Thanks for the approval and the minor notes — addressed in the amended commit (9b72919 -> de12f77, force-pushed).

Minor Issues

1. _execute_tool() does not enforce the host unsafe contract — Confirmed against spec (§9.4: "An Actor MAY declare that it requires unsafe mode" is not graph/LLM-specific) and fixed. _execute_tool now builds global_context from the actor's context.global section and calls _enforce_unsafe_flag(), mirroring _execute_graph/_execute_llm/_execute_multi_actor exactly: refused with UnsafeConfigurationError on a safe host, and _unsafe_mode: true passed to ToolAgent.process_message(message, context=...) on an unsafe host (only when actually needed, so unrelated test doubles that never expected a second argument are unaffected). You noted this was "likely out of scope for the LLM-specific #115" and might warrant a separate PR/issue — kept it in this PR instead, since it closes the same host-unsafe-mode contract this PR already introduces for every other dispatch path, rather than leaving one path spec-non-compliant in the interim. Added 3 new Behave scenarios (refusal, propagation, no-regression) in features/runtime_unsafe_host_mode.feature for this path.

2. Missing default single-LLM regression scenario — Added: "Default single type:llm actor with no unsafe declaration keeps file_write blocked (no regression)" using _single_llm_config(declare_unsafe=False), mirroring the existing graph-actor no-regression scenario.

3. robot/ToolCallingTestLib.py temp directories not cleaned up — Fixed: added _cleanup_unsafe_host_tempdir(), called from all three Then-keywords (unsafe_configuration_error_should_have_been_raised, file_write_should_have_succeeded, file_write_should_have_been_blocked) after their assertions run, so the robot_unsafe_host_* temp dir is removed once the scenario no longer needs it. Verified no leaked directories after a full nox -s integration_tests run.

Nits

• noxfile.py stale coverage threshold docstring — Not changed. Confirmed with the maintainer: the docstring's "97.0%" is an intentional rounded display value, while COVERAGE_THRESHOLD = 96.5 is the real enforced gate — not the inconsistency it appeared to be, so no edit was made.

Verification (via nox)

• nox -s lint, nox -s typecheck — clean
• nox -s unit_tests — 3079/3079 Behave scenarios pass (0 regressions; includes the new scenarios above)
• nox -s integration_tests — 362/362 Robot scenarios pass (pabot), no leaked temp dirs
• nox -s coverage_report — 96.9% (≥ 96.5% threshold); the only uncovered line touched by this round (runtime_dispatch.py, the elif actor_context: flat-context fallback in the new _execute_tool block) mirrors an already-untested, pre-existing defensive branch duplicated identically in _execute_llm/_execute_graph — not a new gap.

CHANGELOG.md's issue #115 entry was rewritten to a single concise, user-oriented sentence (previously a multi-paragraph technical write-up) to match this project's actual CONTRIBUTING convention rather than the file's pre-existing long-form style.

All fixes were amended into the existing single commit per this repo's one-issue/one-commit convention (no new commit added).

Thanks for the approval and the minor notes — addressed in the amended commit (`9b72919` -> `de12f77`, force-pushed). ### Minor Issues 1. **`_execute_tool()` does not enforce the host `unsafe` contract** — Confirmed against spec (§9.4: "An Actor MAY declare that it requires unsafe mode" is not graph/LLM-specific) and fixed. `_execute_tool` now builds `global_context` from the actor's `context.global` section and calls `_enforce_unsafe_flag()`, mirroring `_execute_graph`/`_execute_llm`/`_execute_multi_actor` exactly: refused with `UnsafeConfigurationError` on a safe host, and `_unsafe_mode: true` passed to `ToolAgent.process_message(message, context=...)` on an unsafe host (only when actually needed, so unrelated test doubles that never expected a second argument are unaffected). You noted this was "likely out of scope for the LLM-specific #115" and might warrant a separate PR/issue — kept it in this PR instead, since it closes the same host-unsafe-mode contract this PR already introduces for every other dispatch path, rather than leaving one path spec-non-compliant in the interim. Added 3 new Behave scenarios (refusal, propagation, no-regression) in `features/runtime_unsafe_host_mode.feature` for this path. 2. **Missing default single-LLM regression scenario** — Added: "Default single type:llm actor with no unsafe declaration keeps file_write blocked (no regression)" using `_single_llm_config(declare_unsafe=False)`, mirroring the existing graph-actor no-regression scenario. 3. **`robot/ToolCallingTestLib.py` temp directories not cleaned up** — Fixed: added `_cleanup_unsafe_host_tempdir()`, called from all three Then-keywords (`unsafe_configuration_error_should_have_been_raised`, `file_write_should_have_succeeded`, `file_write_should_have_been_blocked`) after their assertions run, so the `robot_unsafe_host_*` temp dir is removed once the scenario no longer needs it. Verified no leaked directories after a full `nox -s integration_tests` run. ### Nits - **`noxfile.py` stale coverage threshold docstring** — Not changed. Confirmed with the maintainer: the docstring's "97.0%" is an intentional rounded display value, while `COVERAGE_THRESHOLD = 96.5` is the real enforced gate — not the inconsistency it appeared to be, so no edit was made. ### Verification (via `nox`) - `nox -s lint`, `nox -s typecheck` — clean - `nox -s unit_tests` — 3079/3079 Behave scenarios pass (0 regressions; includes the new scenarios above) - `nox -s integration_tests` — 362/362 Robot scenarios pass (pabot), no leaked temp dirs - `nox -s coverage_report` — 96.9% (≥ 96.5% threshold); the only uncovered line touched by this round (`runtime_dispatch.py`, the `elif actor_context:` flat-context fallback in the new `_execute_tool` block) mirrors an already-untested, pre-existing defensive branch duplicated identically in `_execute_llm`/`_execute_graph` — not a new gap. CHANGELOG.md's issue #115 entry was rewritten to a single concise, user-oriented sentence (previously a multi-paragraph technical write-up) to match this project's actual CONTRIBUTING convention rather than the file's pre-existing long-form style. All fixes were amended into the existing single commit per this repo's one-issue/one-commit convention (no new commit added).

CoreRasurae force-pushed bugfix/m1-llm-tool-unsafe-mode from de12f778df to 3e5cdee737 2026-08-20 17:15:56 +00:00 Compare

CoreRasurae merged commit 3e5cdee737 into master 2026-08-20 17:32:33 +00:00

CoreRasurae deleted branch bugfix/m1-llm-tool-unsafe-mode 2026-08-20 17:32:42 +00:00

Sign in to join this conversation.

Reviewers

No reviewers

hurui200320

Labels

Clear labels

  

auto/blocked-by-deps

PR blocked by an open issue dependency. Operator must close the dep (or remove the dependency link) before the merge driver can act. Auto-cleared by merge_drive when no open deps remain.

  

auto/ci-timeout

Most recent merge cycle hit CI timeout. Driver excludes this PR while last merge_cycle row is < 30 min old; label persists thereafter as visible history.

  

auto/claimed-implementer

Currently being processed by an implementer worker.

  

auto/claimed-merge

Currently being processed by the merge driver.

  

auto/claimed-reviewer

Currently being processed by a reviewer worker.

  

auto/driver-down

Merge driver heartbeat stale; pipeline halted. Closed automatically on next clean tick.

  

auto/invariant-violation

Detected master commit violating the strict merge invariant. Tracked as an issue (not a PR label); kept here for label completeness.

  

auto/last-attempt-tier-0

In-cycle escalation: most recent attempt ran at the Tier 0 slot (`tier-0`). Slot's model defined in .opencode/models/tiers.yaml.

  

auto/last-attempt-tier-1

In-cycle escalation: most recent attempt ran at the Tier 1 slot (`tier-1`). Slot's model defined in .opencode/models/tiers.yaml.

  

auto/last-attempt-tier-2

In-cycle escalation: most recent attempt ran at the Tier 2 slot (`tier-2`). Slot's model defined in .opencode/models/tiers.yaml. Gated behind IMPLEMENTER_ESCALATION_TIER2_ENABLED.

  

auto/last-attempt-tier-min

In-cycle escalation: most recent attempt ran at the Tier -1 slot (`tier-min`). Slot's model defined in .opencode/models/tiers.yaml. Suffix is ``-min`` (not ``--1``) so the Forgejo UI reads naturally.

  

Automation Tracking

Tracking issues used by the AI Automation system for agents to communicate and report.

  

auto/needs-conflict-resolution

Rebase conflict needs LLM conflict-resolver.

  

auto/needs-implementer

Failing CI needs implementer attention.

  

auto/postmortem

Documenting a driver incident or rollback.

  

auto/ready-to-merge

Reviewer has APPROVED this PR and no later REQUEST_CHANGES is outstanding. The merge driver requires this label to even consider a PR for merging. Set by the reviewer worker on APPROVE; cleared on REQUEST_CHANGES.

  

auto/restart-throttled

Train repeatedly lost master-tempo races. Driver excludes via merge_cycle until cooldown elapses; label persists as visible history.

  

auto/revert

Revert PR backing out an invariant violation. Fast-tracked through the merge driver.

  

auto/sentinel

Sentinel PR duplicated from upstream into a personal fork by tools/duplicate_prs_to_fork.py for pipeline testing. Lives only in the fork; the canonical pipeline never sees it.

  

auto/stale-inactivity

No implementer activity for N days. Flagged for human review. Auto-cleared on next push to head branch.

  

auto/unstable

Repeatedly fails on current master (>= 3 ci-fail-on-rebased-sha releases in 12 h). Excluded from driver until human triage.

  

Blocked

A ticket in a blocked state and unable to complete until some other task is completed first.

  

Bounty

$100

A bounty of $100 for any open-source contributor who provides a MR that solves this issue

  

Bounty

$1000

A bounty of $1000 for any open-source contributor who provides a MR that solves this issue

  

Bounty

$10000

A bounty of $10000 for any open-source contributor who provides a MR that solves this issue

  

Bounty

$20

A bounty of $20 for any open-source contributor who provides a MR that solves this issue

  

Bounty

$2000

A bounty of $2000 for any open-source contributor who provides a MR that solves this issue

  

Bounty

$250

A bounty of $250 for any open-source contributor who provides a MR that solves this issue

  

Bounty

$50

A bounty of $50 for any open-source contributor who provides a MR that solves this issue

  

Bounty

$500

A bounty of $500 for any open-source contributor who provides a MR that solves this issue

  

Bounty

$5000

A bounty of $5000 for any open-source contributor who provides a MR that solves this issue

  

Bounty

$750

A bounty of $750 for any open-source contributor who provides a MR that solves this issue

  

MoSCoW

Could have

Could have feature in order to satisfy the epic/legendary.

  

MoSCoW

Must have

Must have feature in order to satisfy the epic/legendary.

  

MoSCoW

Should have

Should have feature in order to satisfy the epic/legendary.

  

Needs Feedback

There are questions in the ticket that can not be completed until the project owner provides clarity.

  

Points

1

1 man-hours worth of work for an expert with no learning curve.

  

Points

13

13 man-hours worth of work for an expert with no learning curve.

  

Points

2

2 man-hours worth of work for an expert with no learning curve.

  

Points

21

21 man-hours worth of work for an expert with no learning curve.

  

Points

3

3 man-hours worth of work for an expert with no learning curve.

  

Points

34

34 man-hours worth of work for an expert with no learning curve.

  

Points

5

5 man-hours worth of work for an expert with no learning curve.

  

Points

55

55 man-hours worth of work for an expert with no learning curve.

  

Points

8

8 man-hours worth of work for an expert with no learning curve.

  

Points

88

88 man-hours worth of work for an expert with no learning curve.

  

Priority

Backlog

This ticket has backlogged priority and is not to be worked on yet

  

Priority

CI Blocker

Critical priority issue that blocks CI/CD pipeline and prevents PR merges

  

Priority

Critical

The priority is critical

  

Priority

High

The priority is high

  

Priority

Low

The priority is low

  

Priority

Medium

The priority is medium

  

Signed-off: Owner

When an epic or legendary is in review it must be signed off by owner, tech lead, and scrum master before being marked as completed.

  

Signed-off: Scrum Master

When an epic or legendary is in review it must be signed off by owner, tech lead, and scrum master before being marked as completed.

  

Signed-off: Tech Lead

When an epic or legendary is in review it must be signed off by owner, tech lead, and scrum master before being marked as completed.

  

Spike

A ticket for learning a tool or technology that is needed to be able to do future planning and design.

  

State

Completed

The ticket has been fully implemented, completed, and merged with the source code. This label should only be applied once a ticket is closed.

  

State

Duplicate

A ticket that represents the same content as an existing ticket.

  

State

In Progress

A ticket that is actively being developed.

  

State

In Review

A ticket that has had some code completed to implement but is waiting to pass peer review and is not yet merged in.

  

State

Paused

This ticket's work started but wasn't finished. It's on hold (likely in a feature branch) and will be resumed later, either due to a blocker or a delay.

  

State

Unverified

All new tickets start in this state. A developer may set it to show the ticket is unverified. This means we haven't agreed to work on it. It will either move to a verified state or be closed as wontdo.

  

State

Verified

The issue has been verified by a developer as legitimate. It will be worked on and verified tickets are now considered part of the backlog.

  

State

Wont Do

This ticket has been decided it wont be done. This may mean the bug has been determined to not be real (cant verify) or the feature is one we have decided we dont want to adopt.

  

Type

Automation

Any edits or discussion about the AI automated coding system.

  

Type

Bug

Something that doesnt work as intended.

  

Type

Discussion

Anytime a ticket represents a discussion about a subject and doesnt fall into one of the other categories.

  

Type

Documentation

An error or improvement needed in the documentation.

  

Type

Epic

Any first tier epic. That is, an epic which contains only issues as children and will not have sub-epics.

  

Type

Feature

Some new functionality not present.

  

Type

Legendary

A type of Epic which will contain other Epics.

  

Type

Refactor

A code change that restructures existing code without changing its external behavior.

  

Type

Support

Someone needs help using the project.

  

Type

Task

A generic task that doesnt fit into the other type categories.

  

Type

Testing

Work exclusively focusing on fixing or expanding testing.

No labels

auto/blocked-by-deps

auto/ci-timeout

auto/claimed-implementer

auto/claimed-merge

auto/claimed-reviewer

auto/driver-down

auto/invariant-violation

auto/last-attempt-tier-0

auto/last-attempt-tier-1

auto/last-attempt-tier-2

auto/last-attempt-tier-min

Automation Tracking

auto/needs-conflict-resolution

auto/needs-implementer

auto/postmortem

auto/ready-to-merge

auto/restart-throttled

auto/revert

auto/sentinel

auto/stale-inactivity

auto/unstable

Blocked

Bounty

$100

Bounty

$1000

Bounty

$10000

Bounty

$20

Bounty

$2000

Bounty

$250

Bounty

$50

Bounty

$500

Bounty

$5000

Bounty

$750

MoSCoW

Could have

MoSCoW

Must have

MoSCoW

Should have

Needs Feedback

Points

1

Points

13

Points

2

Points

21

Points

3

Points

34

Points

5

Points

55

Points

8

Points

88

Priority

Backlog

Priority

CI Blocker

Priority

Critical

Priority

High

Priority

Low

Priority

Medium

Signed-off: Owner

Signed-off: Scrum Master

Signed-off: Tech Lead

Spike

State

Completed

State

Duplicate

State

In Progress

State

In Review

State

Paused

State

Unverified

State

Verified

State

Wont Do

Type

Automation

Type

Bug

Type

Discussion

Type

Documentation

Type

Epic

Type

Feature

Type

Legendary

Type

Refactor

Type

Support

Type

Task

Type

Testing

Milestone

Clear milestone

No items

No milestone

v2.1.0

Projects

Clear projects

No items

No project

Assignees

Clear assignees

No assignees

2 participants

Notifications

Subscribe

Due date

The due date is invalid or out of range. Please use the format "yyyy-mm-dd".

No due date set.

Blocks

#115 LLM agent file_write calls are always blocked — safe_mode: false / context.global.unsafe: true never reach _unsafe_mode

cleveragents/cleveractors-core

Reference

cleveragents/cleveractors-core!129

No description provided.