feat: close the governed capability ladder from plan to activate

[chumyin]

Summary

• Problem:
  • LoongClaw already had governed runtime-capability propose/review/index/plan surfaces, but the ladder still stopped short of operator-usable draft artifacts, explicit activation, and rollback-grade activation records.
• Why it matters:
  • Maintainers need a bounded, reviewable path from accepted capability families to concrete managed artifacts and explicit activation outcomes without fake-success paths or environment-dependent verification.
• What changed:
  • Added governed runtime-capability apply support for managed_skill, programmatic_flow, and profile_note_addendum families.
  • Extended runtime-capability plan so operators can inspect structured draft payloads before materializing artifacts.
  • Made managed-skill / programmatic-flow / profile-note drafts self-contained and deterministic.
  • Added governed runtime-capability activate support for managed-skill and profile-note targets, while keeping programmatic-flow activation fail-closed until a real runtime surface exists.
  • Persisted activation records with rollback-oriented evidence so apply/activate outcomes stay auditable.
  • Fixed false-positive activation reporting and hardened local verification so rollback/fail-closed tests do not depend on permission-bit behavior.
  • Kept public repo wording LoongClaw-centric in AGENTS/CLAUDE guidance.
• What did not change (scope boundary):
  • No live runtime mutation path for programmatic flows.
  • No replay executor or trajectory export work in this PR; that surface is already tracked separately on current dev.

Linked Issues

• Closes [Feature]: Establish a governed capability incubator for bounded self-improvement #1092
• Related [Docs]: Archive the 2026-04 governed agent-growth reference study and execution backlog #1091

Change Type

• Bug fix
• Feature
• Refactor
• Documentation
• Security hardening
• CI / workflow / release

Touched Areas

• Kernel / policy / approvals
• Contracts / protocol / spec
• Daemon / CLI / install
• Providers / routing
• Tools
• Browser automation
• Channels / integrations
• ACP / conversation / session runtime
• Memory / context assembly
• Docs / contributor workflow
• CI / release / workflows

Risk Track

• Track A (routine / low-risk)
• Track B (higher-risk / policy-impacting)

Validation

• cargo fmt --all -- --check
• cargo clippy --workspace --all-targets --all-features -- -D warnings
• cargo test --workspace
• cargo test --workspace --all-features
• Relevant architecture / dep-graph / docs checks for touched areas
• Additional scenario, benchmark, or manual checks when behavior changed
• If tests mutate process-global env: document how state is restored or serialized

Commands and evidence:

cargo fmt --all -- --check
cargo clippy --workspace --all-targets --all-features -- -D warnings
cargo test --workspace
cargo test --workspace --all-features
./scripts/check_architecture_boundaries.sh
./scripts/check_dep_graph.sh
cargo test -p loongclaw-daemon runtime_capability --test integration -- --test-threads=1
cargo test -p loongclaw-daemon runtime_trajectory --test integration -- --test-threads=1

User-visible / Operator-visible Changes

• loongclaw runtime-capability apply --root <path> --family-id <id> [--json]
• loongclaw runtime-capability activate --artifact <path> [--config <path>] [--apply] [--replace] [--json]
• runtime-capability plan now surfaces structured draft payload details that match the applied artifact content.
• Activation records now persist rollback-oriented evidence instead of relying on transient console output alone.

Failure Recovery

• Fast rollback or disable path:
  • Revert this PR; applied capability artifacts and activation records are bounded filesystem artifacts and do not require schema migration.
• Observable failure symptoms reviewers should watch for:
  • capability families still materializing only memory_stage_profile payloads instead of the new target kinds
  • managed-skill activation reporting success without installed artifacts actually matching the applied payload
  • profile-note activation failing to record rollback-grade activation evidence
  • config / external-skill / restore tests regressing only on specific filesystem permission models

Reviewer Focus

• crates/daemon/src/runtime_capability_cli.rs
  • target-kind transition, structured payloads, activation flow, and activation record persistence.
• crates/daemon/tests/integration/runtime_capability_cli.rs
  • coverage for apply / activate / rollback and fail-closed paths.
• docs/product-specs/runtime-capability.md, docs/ROADMAP.md
  • product contract updates for the deeper capability ladder.
• crates/app/src/config/runtime.rs, crates/app/src/tools/external_skills.rs, crates/daemon/tests/integration/*
  • deterministic failure injection and rollback verification hardening.

[coderabbitai]

Note

Reviews paused

It looks like this branch is under active development. To avoid overwhelming you with review comments due to an influx of new commits, CodeRabbit has automatically paused this review. You can configure this behavior by changing the reviews.auto_review.auto_pause_after_reviewed_commits setting.

Use the following commands to manage reviews:

• @coderabbitai resume to resume automatic reviews.
• @coderabbitai review to trigger a single review.

Use the checkboxes below for quick actions:

• ▶️ Resume reviews
• 🔍 Trigger review

📝 Walkthrough

Walkthrough

Added workspace tracing deps and observability helpers; instrumented runtime turn/provider/tool flows with structured logs; introduced a read-only runtime-trajectory export/show CLI, artifact schema, session trajectory exporter, session-repo accessors, daemon CLI wiring, tests, and docs. Also added runtime-capability apply CLI and related artifacts.

Changes

Cohort / File(s)	Summary
Docs & Content Guidance AGENTS.md, CLAUDE.md, docs/product-specs/index.md, docs/product-specs/runtime-trajectory.md, docs/ROADMAP.md	Added LoongClaw-centric wording rule; added runtime-trajectory product spec and runtime-capability apply roadmap/spec updates.
Workspace deps Cargo.toml, crates/app/Cargo.toml, crates/daemon/Cargo.toml	Added tracing and tracing-subscriber to workspace deps and enabled tracing.workspace = true for crates.
Observability helpers crates/app/src/observability.rs, crates/daemon/src/observability.rs, crates/app/src/lib.rs	New app/daemon observability modules: JSON/key/error summarizers and daemon tracing init; wired app observability module.
Tracing instrumentation crates/app/src/acp/manager.rs, crates/app/src/channel/mod.rs, crates/app/src/tools/mod.rs, crates/app/src/provider/..., crates/app/src/provider/runtime_binding.rs	Added tracing/timing logs around ACP turns, inbound channel processing, tool execution, and provider failover; added ProviderRuntimeBinding::as_str.
Memory / SQLite accessors crates/app/src/memory/sqlite.rs, crates/app/src/memory/mod.rs	Added PersistedConversationTurnRecord, SQL to fetch full turn rows, session_turn_records_direct, and re-exported types (feature-gated).
Session repo & trajectory gating crates/app/src/session/repository.rs, crates/app/src/session/mod.rs	Added list_all_events, upsert_session_terminal_outcome, ALL_EVENTS_PAGE_LIMIT, and feature-gated trajectory module declaration.
Runtime Trajectory feature crates/app/src/session/trajectory.rs	New large module: artifact schema, export types, export mode, export_runtime_trajectory, mapping helpers, and aggregated statistics.
Daemon CLI wiring & runtime-trajectory CLI crates/daemon/src/lib.rs, crates/daemon/src/main.rs, crates/daemon/src/runtime_trajectory_cli.rs	Init tracing on startup, added runtime-trajectory CLI with export/show, handlers, file I/O, text rendering, and re-exports.
Daemon observability init crates/daemon/src/observability.rs	Tracing initialization: log format parsing, env-var precedence, env-filter resolution, and subscriber setup with tests.
CLI tests & integration crates/daemon/tests/integration/*, crates/daemon/tests/integration/runtime_trajectory_cli.rs	Added integration tests for runtime-trajectory export/show, round-trip, metric/rollup assertions; updated CLI parsing tests and small env isolation fixes.
Runtime-capability apply crates/daemon/src/runtime_capability_cli.rs, crates/daemon/tests/integration/runtime_capability_cli.rs	Added apply subcommand, apply/persistence models, idempotent materialization behavior, schema constants, and integration tests.
Other small edits crates/app/src/channel/registry.rs, crates/app/src/memory/...	Test env isolation changes and minor re-exports; mostly additive logging/timing instrumentation.

Sequence Diagram(s)

sequenceDiagram
    participant CLI as "Daemon CLI (runtime-trajectory)"
    participant Traj as "session::trajectory::export_runtime_trajectory"
    participant Repo as "SessionRepository"
    participant DB as "SQLite DB"

    CLI->>Traj: export_runtime_trajectory(session_id, mode, config, exported_at)
    Traj->>Repo: load_session_summary_with_legacy_fallback(session_id)
    Repo->>DB: SELECT session summary
    DB-->>Repo: session_summary
    Repo-->>Traj: session_summary

    Traj->>Repo: resolve lineage root & collect session IDs (mode)
    Repo->>DB: SELECT sessions (lineage)
    DB-->>Repo: session_list
    Repo-->>Traj: session_list

    loop per session
        Traj->>Repo: session_turn_records_direct(session_id)
        Repo->>DB: SELECT turn rows for session
        DB-->>Repo: PersistedConversationTurnRecord[]
        Repo-->>Traj: turn_records

        Traj->>Repo: list_all_events(session_id)
        Repo->>DB: SELECT events/pages
        DB-->>Repo: SessionEventRecord[]
        Repo-->>Traj: session_events

        Traj->>Repo: load_terminal_outcome(session_id)
        Repo->>DB: SELECT terminal_outcomes
        DB-->>Repo: terminal_outcome?
        Repo-->>Traj: terminal_outcome

        Traj->>DB: SELECT approval_requests for session turns
        DB-->>Traj: approval_requests[]
    end

    Traj->>Traj: aggregate statistics & build artifact
    Traj-->>CLI: RuntimeTrajectoryArtifactDocument (JSON)

Loading

Estimated code review effort

🎯 4 (Complex) | ⏱️ ~50 minutes

Possibly related issues

• [Feature]: Add runtime trajectory artifacts for session replay and evaluation #1090 — Implements runtime-trajectory export/CLI and artifact objectives matching this PR.
• [Feature]: add session search and trajectory artifact CLI surfaces #902 — Related: adds trajectory artifact and CLI surface similar to changes here.
• [Feature]: Add governed runtime-capability apply executor for memory-stage profiles #702 — Related: runtime-capability apply objectives addressed by the new apply subcommand and artifact flow.

Possibly related PRs

• feat: add session search and trajectory artifact CLIs #905 — Closely overlapping runtime-trajectory export and CLI implementation.
• Canonicalize tool retry context across wrapped tool calls #828 — Overlaps observability/tracing additions and instrumentation across same crates.
• feat(daemon): emit dry-run memory stage profile payloads #682 — Related to runtime_capability apply/planned-payload and CLI plumbing.

Suggested labels

docs

Suggested reviewers

• gh-xj

Poem

🐇 I hop through logs with tiny feet,

I gather turns and rows complete,
From root to leaf I stitch the trail,
Export the tale beyond the vale,
A carrot cheer for trajectory neat!

🚥 Pre-merge checks | ✅ 4 | ❌ 1

❌ Failed checks (1 warning)

Check name	Status	Explanation	Resolution
Docstring Coverage	⚠️ Warning	Docstring coverage is 28.35% which is insufficient. The required threshold is 80.00%.	Write docstrings for the functions missing them to satisfy the coverage threshold.

✅ Passed checks (4 passed)

Check name	Status	Explanation
Linked Issues check	✅ Passed	The PR successfully implements all coding requirements from #1089 (runtime trajectory export CLI, stable JSON schema, lineage support, round-trip show command) and #1092 (capability apply command, planned payloads, idempotent apply, target restriction to managed_skill/programmatic_flow/profile_note_addendum).
Out of Scope Changes check	✅ Passed	All changes are directly aligned with stated objectives: trajectory export surface, capability apply implementation, supporting observability/tracing additions, and related CLI/test/docs work. No unrelated refactoring or out-of-scope modifications detected.
Description Check	✅ Passed	Check skipped - CodeRabbit’s high-level summary is enabled.
Title check	✅ Passed	The PR title 'feat: close the governed capability ladder from plan to activate' accurately captures one of the two major features (capability apply), but omits the equally significant runtime trajectory export feature.

_{✏️ Tip: You can configure your own custom pre-merge checks in the settings.}

✨ Finishing Touches

🧪 Generate unit tests (beta)

• Create PR with unit tests
• Commit unit tests in branch work/refstudy-delivery-20260408

---

Thanks for using CodeRabbit! It's free for OSS, and your support helps us grow. If you like it, consider giving us a shout-out.

❤️ Share

• X
• Mastodon
• Reddit
• LinkedIn

_{Comment @coderabbitai help to get the list of available commands and usage tips.}

[coderabbitai]

Actionable comments posted: 3

🧹 Nitpick comments (3)

crates/app/src/tools/mod.rs (1)

651-653: Only materialize payload metadata when this target will emit.

json_value_kind and top_level_json_keys now run on every core-tool call, even when loongclaw.tools logging is disabled. Since this sits on the hot execution path, it’s worth gating that work behind the active log level.

🤖 Prompt for AI Agents

Verify each finding against the current code and only fix it if needed.

In `@crates/app/src/tools/mod.rs` around lines 651 - 653, Only compute payload
metadata eagerly when loongclaw.tools logging is active: guard the calls to
crate::observability::json_value_kind(&request.payload) and
crate::observability::top_level_json_keys(&request.payload) with a check for the
loongclaw.tools log level (e.g. a
crate::observability::is_enabled("loongclaw.tools") or equivalent helper). If
the check is false, avoid calling those functions and set
payload_kind/payload_keys to None/empty (or the existing default) so the hot
path that handles request.payload (referenced as request.payload and
requested_tool_name) does not perform unnecessary work.

crates/app/src/session/repository.rs (1)

1508-1544: Consider consolidating with existing upsert_terminal_outcome to reduce duplication.

This new method is nearly identical to the existing upsert_terminal_outcome (lines 1249-1282), differing only in the session existence check: this method uses load_session_summary_with_legacy_fallback while the existing one uses load_session.

The distinction makes sense for trajectory export (supporting legacy sessions inferred from turn history), but the duplication is notable. Consider extracting a shared private helper that accepts a session-existence-check closure or boolean flag, or document why the split is intentional.

🤖 Prompt for AI Agents

Verify each finding against the current code and only fix it if needed.

In `@crates/app/src/session/repository.rs` around lines 1508 - 1544, The two
functions upsert_session_terminal_outcome and upsert_terminal_outcome are nearly
identical except for how they validate session existence
(load_session_summary_with_legacy_fallback vs load_session); refactor by
extracting a private helper (e.g., fn upsert_terminal_outcome_inner(...,
session_exists_check: impl Fn(&str) -> Result<bool, String>) or a flag) that
performs normalization, payload encoding, DB upsert, and record construction,
and call that helper from both upsert_session_terminal_outcome and
upsert_terminal_outcome passing the appropriate existence-check closure (or
boolean) to remove duplication or alternatively add a short comment explaining
the intentional split if you choose not to refactor.

crates/app/src/session/trajectory.rs (1)

221-247: Performance note: Lineage depth computed per-session via DB traversal.

sort_runtime_trajectory_sessions calls repo.session_lineage_depth() for each session, where each call traverses the parent chain via multiple DB queries. For a lineage with N sessions and max depth D, this results in O(N × D) queries.

This is acceptable for typical session lineages but could become slow for deeply nested or very large lineage trees. If this becomes a bottleneck in practice, consider computing depths in a single pass by building an in-memory parent map from the already-loaded session summaries.

🤖 Prompt for AI Agents

Verify each finding against the current code and only fix it if needed.

In `@crates/app/src/session/trajectory.rs` around lines 221 - 247,
sort_runtime_trajectory_sessions currently calls
SessionRepository::session_lineage_depth for each SessionSummaryRecord which
causes O(N×D) DB traversals; instead, build an in-memory parent map from the
provided sessions (use session.session_id -> session.parent_session_id from the
SessionSummaryRecord list), then compute lineage depths with a single-pass
memoized traversal (or iterative topological/DFS with memoization) into
depth_by_session_id and use that map in the existing sort; keep the sort logic
unchanged and only replace the per-session repo calls with the in-memory depth
computation inside sort_runtime_trajectory_sessions.

🤖 Prompt for all review comments with AI agents

Verify each finding against the current code and only fix it if needed.

Inline comments:
In `@crates/app/src/memory/sqlite.rs`:
- Around line 1923-1925: The export reader currently calls row.get(1) into
session_turn_index as a non-null i64 which will error on NULL DB rows; change
the decode to read an Option<i64> (i.e., row.get::<_, Option<i64>>(1)) and
handle None defensively inside the export path (either treat missing
session_turn_index as a default/sentinel, skip legacy rows, or propagate a more
specific recoverable warning) in the function that performs export reads so
legacy/inconsistent rows do not cause a hard failure; update any subsequent uses
of session_turn_index to account for Option and convert/handle the value before
using it.

In `@crates/app/src/tools/mod.rs`:
- Around line 679-698: The logs for tool invocations always show
canonical_tool_name = "tool.invoke", which hides the actual inner tool (e.g.,
file.read, web.fetch); update both the success (tracing::debug!) and error
(tracing::warn!) blocks around the invoke handling to compute and include the
resolved inner tool name (derive from the canonical_name/payload or the value
used to dispatch the inner tool, e.g., call it resolved_tool_name or
inner_tool_name) and add it as a structured field (inner_tool_name =
%resolved_tool_name) alongside requested_tool_name and canonical_tool_name so
each log entry shows the real invoked tool (references: variables
canonical_name, requested_tool_name, payload_keys, outcome, error).

In `@crates/daemon/src/main.rs`:
- Line 42: The debug log currently serializes and logs the full parsed CLI
payload via tracing::debug!(..., command = ?cli.command), which may include
sensitive user-provided text; update the log to avoid emitting raw payloads by
logging a safe marker instead — for example log only the command variant/name or
a sanitized summary (e.g., cli.command.variant() or cli.command.name()) or
explicitly redact sensitive fields before logging; change the tracing::debug!
invocation to emit that sanitized value (or a fixed string like "CLI command
parsed") rather than ?cli.command.

---

Nitpick comments:
In `@crates/app/src/session/repository.rs`:
- Around line 1508-1544: The two functions upsert_session_terminal_outcome and
upsert_terminal_outcome are nearly identical except for how they validate
session existence (load_session_summary_with_legacy_fallback vs load_session);
refactor by extracting a private helper (e.g., fn
upsert_terminal_outcome_inner(..., session_exists_check: impl Fn(&str) ->
Result<bool, String>) or a flag) that performs normalization, payload encoding,
DB upsert, and record construction, and call that helper from both
upsert_session_terminal_outcome and upsert_terminal_outcome passing the
appropriate existence-check closure (or boolean) to remove duplication or
alternatively add a short comment explaining the intentional split if you choose
not to refactor.

In `@crates/app/src/session/trajectory.rs`:
- Around line 221-247: sort_runtime_trajectory_sessions currently calls
SessionRepository::session_lineage_depth for each SessionSummaryRecord which
causes O(N×D) DB traversals; instead, build an in-memory parent map from the
provided sessions (use session.session_id -> session.parent_session_id from the
SessionSummaryRecord list), then compute lineage depths with a single-pass
memoized traversal (or iterative topological/DFS with memoization) into
depth_by_session_id and use that map in the existing sort; keep the sort logic
unchanged and only replace the per-session repo calls with the in-memory depth
computation inside sort_runtime_trajectory_sessions.

In `@crates/app/src/tools/mod.rs`:
- Around line 651-653: Only compute payload metadata eagerly when
loongclaw.tools logging is active: guard the calls to
crate::observability::json_value_kind(&request.payload) and
crate::observability::top_level_json_keys(&request.payload) with a check for the
loongclaw.tools log level (e.g. a
crate::observability::is_enabled("loongclaw.tools") or equivalent helper). If
the check is false, avoid calling those functions and set
payload_kind/payload_keys to None/empty (or the existing default) so the hot
path that handles request.payload (referenced as request.payload and
requested_tool_name) does not perform unnecessary work.

🪄 Autofix (Beta)

Fix all unresolved CodeRabbit comments on this PR:

• Push a commit to this branch (recommended)
• Create a new PR with the fixes

---

ℹ️ Review info

⚙️ Run configuration

Configuration used: Path: .coderabbit.yaml

Review profile: CHILL

Plan: Pro

Run ID: 4b329228-511f-429c-927e-07b7c802669b

📥 Commits

Reviewing files that changed from the base of the PR and between 90527de and 8d54cfe.

⛔ Files ignored due to path filters (1)

• Cargo.lock is excluded by !**/*.lock

📒 Files selected for processing (29)

• AGENTS.md
• CLAUDE.md
• Cargo.toml
• crates/app/Cargo.toml
• crates/app/src/acp/manager.rs
• crates/app/src/channel/mod.rs
• crates/app/src/channel/registry.rs
• crates/app/src/lib.rs
• crates/app/src/memory/mod.rs
• crates/app/src/memory/sqlite.rs
• crates/app/src/observability.rs
• crates/app/src/provider/request_failover_runtime.rs
• crates/app/src/provider/request_session_runtime.rs
• crates/app/src/provider/runtime_binding.rs
• crates/app/src/session/mod.rs
• crates/app/src/session/repository.rs
• crates/app/src/session/trajectory.rs
• crates/app/src/tools/mod.rs
• crates/daemon/Cargo.toml
• crates/daemon/src/lib.rs
• crates/daemon/src/main.rs
• crates/daemon/src/observability.rs
• crates/daemon/src/runtime_trajectory_cli.rs
• crates/daemon/tests/integration/cli_tests.rs
• crates/daemon/tests/integration/migrate_cli.rs
• crates/daemon/tests/integration/mod.rs
• crates/daemon/tests/integration/runtime_trajectory_cli.rs
• docs/product-specs/index.md
• docs/product-specs/runtime-trajectory.md

[coderabbitai]

Actionable comments posted: 2

🤖 Prompt for all review comments with AI agents

Verify each finding against the current code and only fix it if needed.

Inline comments:
In `@crates/daemon/src/runtime_capability_cli.rs`:
- Around line 1379-1439:
persist_runtime_capability_apply_artifact/write_pretty_json_file_create_new
currently opens the final path with create_new(true) which can make the file
visible before the JSON is fully written, causing races; instead, write the
pretty-serialized bytes to a temporary file in the same directory (e.g., path
with a random/suffixed name), fsync the temp file (and optionally its parent
dir), then atomically rename (fs::rename) the temp file to the final path and
map an AlreadyExists error from rename to the same "already exists" string so
persist_runtime_capability_apply_artifact can still load and compare the
existing artifact; update write_pretty_json_file_create_new to perform these
steps and keep the same error messages and return types so callers like
persist_runtime_capability_apply_artifact continue to work unchanged.

In `@crates/daemon/tests/integration/runtime_capability_cli.rs`:
- Around line 2360-2366: The assertion compares report.output_path using
OS-native separators, which fails on Windows; normalize the path before checking
the suffix by passing report.output_path through normalized_path_text(...) (or
otherwise canonicalizing separators) and then assert that
normalized_path_text(&report.output_path).ends_with(&format!("managed_skills/{}.json",
report.applied_artifact.artifact_id)) so the check is platform independent;
update the assertion that references report.output_path and keep the same
message text.

🪄 Autofix (Beta)

Fix all unresolved CodeRabbit comments on this PR:

• Push a commit to this branch (recommended)
• Create a new PR with the fixes

---

ℹ️ Review info

⚙️ Run configuration

Configuration used: Path: .coderabbit.yaml

Review profile: CHILL

Plan: Pro

Run ID: 8fd3e86e-0c0f-49ce-8b21-f46cc4be47eb

📥 Commits

Reviewing files that changed from the base of the PR and between 8d54cfe and d4d32f6.

📒 Files selected for processing (5)

• crates/daemon/src/runtime_capability_cli.rs
• crates/daemon/tests/integration/cli_tests.rs
• crates/daemon/tests/integration/runtime_capability_cli.rs
• docs/ROADMAP.md
• docs/product-specs/runtime-capability.md

✅ Files skipped from review due to trivial changes (2)

• docs/ROADMAP.md
• docs/product-specs/runtime-capability.md

[coderabbitai]

♻️ Duplicate comments (1)

crates/daemon/src/runtime_capability_cli.rs (1)

1417-1478: ⚠️ Potential issue | 🔴 Critical

Publish the apply artifact atomically to prevent idempotency races.

write_pretty_json_file_create_new exposes the destination path immediately after create_new(true), before JSON is fully written. A concurrent apply can hit “already exists”, then fail parsing partially written content and incorrectly fail idempotent behavior.

💡 Suggested fix direction

- open final path with create_new(true)
- write JSON directly to final path
+ write full JSON to a temp file in the same directory
+ flush + sync temp file
+ publish to final path using an atomic no-clobber step
+ map "already exists" from publish step to current idempotency branch
+ best-effort cleanup of temp file on failure

🤖 Prompt for AI Agents

Verify each finding against the current code and only fix it if needed.

In `@crates/daemon/src/runtime_capability_cli.rs` around lines 1417 - 1478,
persist_runtime_capability_apply_artifact can race because
write_pretty_json_file_create_new creates the target path before the JSON is
fully written; fix by making write_pretty_json_file_create_new write to a
temporary file in the same directory, flush and sync the temp file
(file.sync_all()), then attempt an atomic fs::rename(temp_path, path); if rename
fails because destination already exists, load the existing artifact (use
load_runtime_capability_apply_artifact) and compare to decide AlreadyApplied vs
error; ensure temp file is removed on error and create parent dirs as currently
done.

🤖 Prompt for all review comments with AI agents

Verify each finding against the current code and only fix it if needed.

Duplicate comments:
In `@crates/daemon/src/runtime_capability_cli.rs`:
- Around line 1417-1478: persist_runtime_capability_apply_artifact can race
because write_pretty_json_file_create_new creates the target path before the
JSON is fully written; fix by making write_pretty_json_file_create_new write to
a temporary file in the same directory, flush and sync the temp file
(file.sync_all()), then attempt an atomic fs::rename(temp_path, path); if rename
fails because destination already exists, load the existing artifact (use
load_runtime_capability_apply_artifact) and compare to decide AlreadyApplied vs
error; ensure temp file is removed on error and create parent dirs as currently
done.

---

ℹ️ Review info

⚙️ Run configuration

Configuration used: Path: .coderabbit.yaml

Review profile: CHILL

Plan: Pro

Run ID: 89207569-a044-46ba-b1cc-87b0ae67ffcc

📥 Commits

Reviewing files that changed from the base of the PR and between c320728 and c70246f.

📒 Files selected for processing (2)

• crates/daemon/src/runtime_capability_cli.rs
• crates/daemon/tests/integration/runtime_capability_cli.rs

✅ Files skipped from review due to trivial changes (1)

• crates/daemon/tests/integration/runtime_capability_cli.rs